vmscan: synchronous lumpy reclaim should not call congestion_wait()
[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/gfp.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 #define CREATE_TRACE_POINTS
52 #include <trace/events/vmscan.h>
53
54 struct scan_control {
55         /* Incremented by the number of inactive pages that were scanned */
56         unsigned long nr_scanned;
57
58         /* Number of pages freed so far during a call to shrink_zones() */
59         unsigned long nr_reclaimed;
60
61         /* How many pages shrink_list() should reclaim */
62         unsigned long nr_to_reclaim;
63
64         unsigned long hibernation_mode;
65
66         /* This context's GFP mask */
67         gfp_t gfp_mask;
68
69         int may_writepage;
70
71         /* Can mapped pages be reclaimed? */
72         int may_unmap;
73
74         /* Can pages be swapped as part of reclaim? */
75         int may_swap;
76
77         int swappiness;
78
79         int order;
80
81         /*
82          * Intend to reclaim enough continuous memory rather than reclaim
83          * enough amount of memory. i.e, mode for high order allocation.
84          */
85         bool lumpy_reclaim_mode;
86
87         /* Which cgroup do we reclaim from */
88         struct mem_cgroup *mem_cgroup;
89
90         /*
91          * Nodemask of nodes allowed by the caller. If NULL, all nodes
92          * are scanned.
93          */
94         nodemask_t      *nodemask;
95 };
96
97 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
98
99 #ifdef ARCH_HAS_PREFETCH
100 #define prefetch_prev_lru_page(_page, _base, _field)                    \
101         do {                                                            \
102                 if ((_page)->lru.prev != _base) {                       \
103                         struct page *prev;                              \
104                                                                         \
105                         prev = lru_to_page(&(_page->lru));              \
106                         prefetch(&prev->_field);                        \
107                 }                                                       \
108         } while (0)
109 #else
110 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
111 #endif
112
113 #ifdef ARCH_HAS_PREFETCHW
114 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
115         do {                                                            \
116                 if ((_page)->lru.prev != _base) {                       \
117                         struct page *prev;                              \
118                                                                         \
119                         prev = lru_to_page(&(_page->lru));              \
120                         prefetchw(&prev->_field);                       \
121                 }                                                       \
122         } while (0)
123 #else
124 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
125 #endif
126
127 /*
128  * From 0 .. 100.  Higher means more swappy.
129  */
130 int vm_swappiness = 60;
131 long vm_total_pages;    /* The total number of pages which the VM controls */
132
133 static LIST_HEAD(shrinker_list);
134 static DECLARE_RWSEM(shrinker_rwsem);
135
136 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
137 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
138 #else
139 #define scanning_global_lru(sc) (1)
140 #endif
141
142 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
143                                                   struct scan_control *sc)
144 {
145         if (!scanning_global_lru(sc))
146                 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
147
148         return &zone->reclaim_stat;
149 }
150
151 static unsigned long zone_nr_lru_pages(struct zone *zone,
152                                 struct scan_control *sc, enum lru_list lru)
153 {
154         if (!scanning_global_lru(sc))
155                 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
156
157         return zone_page_state(zone, NR_LRU_BASE + lru);
158 }
159
160
161 /*
162  * Add a shrinker callback to be called from the vm
163  */
164 void register_shrinker(struct shrinker *shrinker)
165 {
166         shrinker->nr = 0;
167         down_write(&shrinker_rwsem);
168         list_add_tail(&shrinker->list, &shrinker_list);
169         up_write(&shrinker_rwsem);
170 }
171 EXPORT_SYMBOL(register_shrinker);
172
173 /*
174  * Remove one
175  */
176 void unregister_shrinker(struct shrinker *shrinker)
177 {
178         down_write(&shrinker_rwsem);
179         list_del(&shrinker->list);
180         up_write(&shrinker_rwsem);
181 }
182 EXPORT_SYMBOL(unregister_shrinker);
183
184 #define SHRINK_BATCH 128
185 /*
186  * Call the shrink functions to age shrinkable caches
187  *
188  * Here we assume it costs one seek to replace a lru page and that it also
189  * takes a seek to recreate a cache object.  With this in mind we age equal
190  * percentages of the lru and ageable caches.  This should balance the seeks
191  * generated by these structures.
192  *
193  * If the vm encountered mapped pages on the LRU it increase the pressure on
194  * slab to avoid swapping.
195  *
196  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
197  *
198  * `lru_pages' represents the number of on-LRU pages in all the zones which
199  * are eligible for the caller's allocation attempt.  It is used for balancing
200  * slab reclaim versus page reclaim.
201  *
202  * Returns the number of slab objects which we shrunk.
203  */
204 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
205                         unsigned long lru_pages)
206 {
207         struct shrinker *shrinker;
208         unsigned long ret = 0;
209
210         if (scanned == 0)
211                 scanned = SWAP_CLUSTER_MAX;
212
213         if (!down_read_trylock(&shrinker_rwsem))
214                 return 1;       /* Assume we'll be able to shrink next time */
215
216         list_for_each_entry(shrinker, &shrinker_list, list) {
217                 unsigned long long delta;
218                 unsigned long total_scan;
219                 unsigned long max_pass;
220
221                 max_pass = (*shrinker->shrink)(shrinker, 0, gfp_mask);
222                 delta = (4 * scanned) / shrinker->seeks;
223                 delta *= max_pass;
224                 do_div(delta, lru_pages + 1);
225                 shrinker->nr += delta;
226                 if (shrinker->nr < 0) {
227                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
228                                "delete nr=%ld\n",
229                                shrinker->shrink, shrinker->nr);
230                         shrinker->nr = max_pass;
231                 }
232
233                 /*
234                  * Avoid risking looping forever due to too large nr value:
235                  * never try to free more than twice the estimate number of
236                  * freeable entries.
237                  */
238                 if (shrinker->nr > max_pass * 2)
239                         shrinker->nr = max_pass * 2;
240
241                 total_scan = shrinker->nr;
242                 shrinker->nr = 0;
243
244                 while (total_scan >= SHRINK_BATCH) {
245                         long this_scan = SHRINK_BATCH;
246                         int shrink_ret;
247                         int nr_before;
248
249                         nr_before = (*shrinker->shrink)(shrinker, 0, gfp_mask);
250                         shrink_ret = (*shrinker->shrink)(shrinker, this_scan,
251                                                                 gfp_mask);
252                         if (shrink_ret == -1)
253                                 break;
254                         if (shrink_ret < nr_before)
255                                 ret += nr_before - shrink_ret;
256                         count_vm_events(SLABS_SCANNED, this_scan);
257                         total_scan -= this_scan;
258
259                         cond_resched();
260                 }
261
262                 shrinker->nr += total_scan;
263         }
264         up_read(&shrinker_rwsem);
265         return ret;
266 }
267
268 static inline int is_page_cache_freeable(struct page *page)
269 {
270         /*
271          * A freeable page cache page is referenced only by the caller
272          * that isolated the page, the page cache radix tree and
273          * optional buffer heads at page->private.
274          */
275         return page_count(page) - page_has_private(page) == 2;
276 }
277
278 static int may_write_to_queue(struct backing_dev_info *bdi)
279 {
280         if (current->flags & PF_SWAPWRITE)
281                 return 1;
282         if (!bdi_write_congested(bdi))
283                 return 1;
284         if (bdi == current->backing_dev_info)
285                 return 1;
286         return 0;
287 }
288
289 /*
290  * We detected a synchronous write error writing a page out.  Probably
291  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
292  * fsync(), msync() or close().
293  *
294  * The tricky part is that after writepage we cannot touch the mapping: nothing
295  * prevents it from being freed up.  But we have a ref on the page and once
296  * that page is locked, the mapping is pinned.
297  *
298  * We're allowed to run sleeping lock_page() here because we know the caller has
299  * __GFP_FS.
300  */
301 static void handle_write_error(struct address_space *mapping,
302                                 struct page *page, int error)
303 {
304         lock_page_nosync(page);
305         if (page_mapping(page) == mapping)
306                 mapping_set_error(mapping, error);
307         unlock_page(page);
308 }
309
310 /* Request for sync pageout. */
311 enum pageout_io {
312         PAGEOUT_IO_ASYNC,
313         PAGEOUT_IO_SYNC,
314 };
315
316 /* possible outcome of pageout() */
317 typedef enum {
318         /* failed to write page out, page is locked */
319         PAGE_KEEP,
320         /* move page to the active list, page is locked */
321         PAGE_ACTIVATE,
322         /* page has been sent to the disk successfully, page is unlocked */
323         PAGE_SUCCESS,
324         /* page is clean and locked */
325         PAGE_CLEAN,
326 } pageout_t;
327
328 /*
329  * pageout is called by shrink_page_list() for each dirty page.
330  * Calls ->writepage().
331  */
332 static pageout_t pageout(struct page *page, struct address_space *mapping,
333                                                 enum pageout_io sync_writeback)
334 {
335         /*
336          * If the page is dirty, only perform writeback if that write
337          * will be non-blocking.  To prevent this allocation from being
338          * stalled by pagecache activity.  But note that there may be
339          * stalls if we need to run get_block().  We could test
340          * PagePrivate for that.
341          *
342          * If this process is currently in __generic_file_aio_write() against
343          * this page's queue, we can perform writeback even if that
344          * will block.
345          *
346          * If the page is swapcache, write it back even if that would
347          * block, for some throttling. This happens by accident, because
348          * swap_backing_dev_info is bust: it doesn't reflect the
349          * congestion state of the swapdevs.  Easy to fix, if needed.
350          */
351         if (!is_page_cache_freeable(page))
352                 return PAGE_KEEP;
353         if (!mapping) {
354                 /*
355                  * Some data journaling orphaned pages can have
356                  * page->mapping == NULL while being dirty with clean buffers.
357                  */
358                 if (page_has_private(page)) {
359                         if (try_to_free_buffers(page)) {
360                                 ClearPageDirty(page);
361                                 printk("%s: orphaned page\n", __func__);
362                                 return PAGE_CLEAN;
363                         }
364                 }
365                 return PAGE_KEEP;
366         }
367         if (mapping->a_ops->writepage == NULL)
368                 return PAGE_ACTIVATE;
369         if (!may_write_to_queue(mapping->backing_dev_info))
370                 return PAGE_KEEP;
371
372         if (clear_page_dirty_for_io(page)) {
373                 int res;
374                 struct writeback_control wbc = {
375                         .sync_mode = WB_SYNC_NONE,
376                         .nr_to_write = SWAP_CLUSTER_MAX,
377                         .range_start = 0,
378                         .range_end = LLONG_MAX,
379                         .for_reclaim = 1,
380                 };
381
382                 SetPageReclaim(page);
383                 res = mapping->a_ops->writepage(page, &wbc);
384                 if (res < 0)
385                         handle_write_error(mapping, page, res);
386                 if (res == AOP_WRITEPAGE_ACTIVATE) {
387                         ClearPageReclaim(page);
388                         return PAGE_ACTIVATE;
389                 }
390
391                 /*
392                  * Wait on writeback if requested to. This happens when
393                  * direct reclaiming a large contiguous area and the
394                  * first attempt to free a range of pages fails.
395                  */
396                 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
397                         wait_on_page_writeback(page);
398
399                 if (!PageWriteback(page)) {
400                         /* synchronous write or broken a_ops? */
401                         ClearPageReclaim(page);
402                 }
403                 trace_mm_vmscan_writepage(page,
404                         trace_reclaim_flags(page, sync_writeback));
405                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
406                 return PAGE_SUCCESS;
407         }
408
409         return PAGE_CLEAN;
410 }
411
412 /*
413  * Same as remove_mapping, but if the page is removed from the mapping, it
414  * gets returned with a refcount of 0.
415  */
416 static int __remove_mapping(struct address_space *mapping, struct page *page)
417 {
418         BUG_ON(!PageLocked(page));
419         BUG_ON(mapping != page_mapping(page));
420
421         spin_lock_irq(&mapping->tree_lock);
422         /*
423          * The non racy check for a busy page.
424          *
425          * Must be careful with the order of the tests. When someone has
426          * a ref to the page, it may be possible that they dirty it then
427          * drop the reference. So if PageDirty is tested before page_count
428          * here, then the following race may occur:
429          *
430          * get_user_pages(&page);
431          * [user mapping goes away]
432          * write_to(page);
433          *                              !PageDirty(page)    [good]
434          * SetPageDirty(page);
435          * put_page(page);
436          *                              !page_count(page)   [good, discard it]
437          *
438          * [oops, our write_to data is lost]
439          *
440          * Reversing the order of the tests ensures such a situation cannot
441          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
442          * load is not satisfied before that of page->_count.
443          *
444          * Note that if SetPageDirty is always performed via set_page_dirty,
445          * and thus under tree_lock, then this ordering is not required.
446          */
447         if (!page_freeze_refs(page, 2))
448                 goto cannot_free;
449         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
450         if (unlikely(PageDirty(page))) {
451                 page_unfreeze_refs(page, 2);
452                 goto cannot_free;
453         }
454
455         if (PageSwapCache(page)) {
456                 swp_entry_t swap = { .val = page_private(page) };
457                 __delete_from_swap_cache(page);
458                 spin_unlock_irq(&mapping->tree_lock);
459                 swapcache_free(swap, page);
460         } else {
461                 __remove_from_page_cache(page);
462                 spin_unlock_irq(&mapping->tree_lock);
463                 mem_cgroup_uncharge_cache_page(page);
464         }
465
466         return 1;
467
468 cannot_free:
469         spin_unlock_irq(&mapping->tree_lock);
470         return 0;
471 }
472
473 /*
474  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
475  * someone else has a ref on the page, abort and return 0.  If it was
476  * successfully detached, return 1.  Assumes the caller has a single ref on
477  * this page.
478  */
479 int remove_mapping(struct address_space *mapping, struct page *page)
480 {
481         if (__remove_mapping(mapping, page)) {
482                 /*
483                  * Unfreezing the refcount with 1 rather than 2 effectively
484                  * drops the pagecache ref for us without requiring another
485                  * atomic operation.
486                  */
487                 page_unfreeze_refs(page, 1);
488                 return 1;
489         }
490         return 0;
491 }
492
493 /**
494  * putback_lru_page - put previously isolated page onto appropriate LRU list
495  * @page: page to be put back to appropriate lru list
496  *
497  * Add previously isolated @page to appropriate LRU list.
498  * Page may still be unevictable for other reasons.
499  *
500  * lru_lock must not be held, interrupts must be enabled.
501  */
502 void putback_lru_page(struct page *page)
503 {
504         int lru;
505         int active = !!TestClearPageActive(page);
506         int was_unevictable = PageUnevictable(page);
507
508         VM_BUG_ON(PageLRU(page));
509
510 redo:
511         ClearPageUnevictable(page);
512
513         if (page_evictable(page, NULL)) {
514                 /*
515                  * For evictable pages, we can use the cache.
516                  * In event of a race, worst case is we end up with an
517                  * unevictable page on [in]active list.
518                  * We know how to handle that.
519                  */
520                 lru = active + page_lru_base_type(page);
521                 lru_cache_add_lru(page, lru);
522         } else {
523                 /*
524                  * Put unevictable pages directly on zone's unevictable
525                  * list.
526                  */
527                 lru = LRU_UNEVICTABLE;
528                 add_page_to_unevictable_list(page);
529                 /*
530                  * When racing with an mlock clearing (page is
531                  * unlocked), make sure that if the other thread does
532                  * not observe our setting of PG_lru and fails
533                  * isolation, we see PG_mlocked cleared below and move
534                  * the page back to the evictable list.
535                  *
536                  * The other side is TestClearPageMlocked().
537                  */
538                 smp_mb();
539         }
540
541         /*
542          * page's status can change while we move it among lru. If an evictable
543          * page is on unevictable list, it never be freed. To avoid that,
544          * check after we added it to the list, again.
545          */
546         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
547                 if (!isolate_lru_page(page)) {
548                         put_page(page);
549                         goto redo;
550                 }
551                 /* This means someone else dropped this page from LRU
552                  * So, it will be freed or putback to LRU again. There is
553                  * nothing to do here.
554                  */
555         }
556
557         if (was_unevictable && lru != LRU_UNEVICTABLE)
558                 count_vm_event(UNEVICTABLE_PGRESCUED);
559         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
560                 count_vm_event(UNEVICTABLE_PGCULLED);
561
562         put_page(page);         /* drop ref from isolate */
563 }
564
565 enum page_references {
566         PAGEREF_RECLAIM,
567         PAGEREF_RECLAIM_CLEAN,
568         PAGEREF_KEEP,
569         PAGEREF_ACTIVATE,
570 };
571
572 static enum page_references page_check_references(struct page *page,
573                                                   struct scan_control *sc)
574 {
575         int referenced_ptes, referenced_page;
576         unsigned long vm_flags;
577
578         referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
579         referenced_page = TestClearPageReferenced(page);
580
581         /* Lumpy reclaim - ignore references */
582         if (sc->lumpy_reclaim_mode)
583                 return PAGEREF_RECLAIM;
584
585         /*
586          * Mlock lost the isolation race with us.  Let try_to_unmap()
587          * move the page to the unevictable list.
588          */
589         if (vm_flags & VM_LOCKED)
590                 return PAGEREF_RECLAIM;
591
592         if (referenced_ptes) {
593                 if (PageAnon(page))
594                         return PAGEREF_ACTIVATE;
595                 /*
596                  * All mapped pages start out with page table
597                  * references from the instantiating fault, so we need
598                  * to look twice if a mapped file page is used more
599                  * than once.
600                  *
601                  * Mark it and spare it for another trip around the
602                  * inactive list.  Another page table reference will
603                  * lead to its activation.
604                  *
605                  * Note: the mark is set for activated pages as well
606                  * so that recently deactivated but used pages are
607                  * quickly recovered.
608                  */
609                 SetPageReferenced(page);
610
611                 if (referenced_page)
612                         return PAGEREF_ACTIVATE;
613
614                 return PAGEREF_KEEP;
615         }
616
617         /* Reclaim if clean, defer dirty pages to writeback */
618         if (referenced_page)
619                 return PAGEREF_RECLAIM_CLEAN;
620
621         return PAGEREF_RECLAIM;
622 }
623
624 static noinline_for_stack void free_page_list(struct list_head *free_pages)
625 {
626         struct pagevec freed_pvec;
627         struct page *page, *tmp;
628
629         pagevec_init(&freed_pvec, 1);
630
631         list_for_each_entry_safe(page, tmp, free_pages, lru) {
632                 list_del(&page->lru);
633                 if (!pagevec_add(&freed_pvec, page)) {
634                         __pagevec_free(&freed_pvec);
635                         pagevec_reinit(&freed_pvec);
636                 }
637         }
638
639         pagevec_free(&freed_pvec);
640 }
641
642 /*
643  * shrink_page_list() returns the number of reclaimed pages
644  */
645 static unsigned long shrink_page_list(struct list_head *page_list,
646                                         struct scan_control *sc,
647                                         enum pageout_io sync_writeback)
648 {
649         LIST_HEAD(ret_pages);
650         LIST_HEAD(free_pages);
651         int pgactivate = 0;
652         unsigned long nr_reclaimed = 0;
653
654         cond_resched();
655
656         while (!list_empty(page_list)) {
657                 enum page_references references;
658                 struct address_space *mapping;
659                 struct page *page;
660                 int may_enter_fs;
661
662                 cond_resched();
663
664                 page = lru_to_page(page_list);
665                 list_del(&page->lru);
666
667                 if (!trylock_page(page))
668                         goto keep;
669
670                 VM_BUG_ON(PageActive(page));
671
672                 sc->nr_scanned++;
673
674                 if (unlikely(!page_evictable(page, NULL)))
675                         goto cull_mlocked;
676
677                 if (!sc->may_unmap && page_mapped(page))
678                         goto keep_locked;
679
680                 /* Double the slab pressure for mapped and swapcache pages */
681                 if (page_mapped(page) || PageSwapCache(page))
682                         sc->nr_scanned++;
683
684                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
685                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
686
687                 if (PageWriteback(page)) {
688                         /*
689                          * Synchronous reclaim is performed in two passes,
690                          * first an asynchronous pass over the list to
691                          * start parallel writeback, and a second synchronous
692                          * pass to wait for the IO to complete.  Wait here
693                          * for any page for which writeback has already
694                          * started.
695                          */
696                         if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
697                                 wait_on_page_writeback(page);
698                         else
699                                 goto keep_locked;
700                 }
701
702                 references = page_check_references(page, sc);
703                 switch (references) {
704                 case PAGEREF_ACTIVATE:
705                         goto activate_locked;
706                 case PAGEREF_KEEP:
707                         goto keep_locked;
708                 case PAGEREF_RECLAIM:
709                 case PAGEREF_RECLAIM_CLEAN:
710                         ; /* try to reclaim the page below */
711                 }
712
713                 /*
714                  * Anonymous process memory has backing store?
715                  * Try to allocate it some swap space here.
716                  */
717                 if (PageAnon(page) && !PageSwapCache(page)) {
718                         if (!(sc->gfp_mask & __GFP_IO))
719                                 goto keep_locked;
720                         if (!add_to_swap(page))
721                                 goto activate_locked;
722                         may_enter_fs = 1;
723                 }
724
725                 mapping = page_mapping(page);
726
727                 /*
728                  * The page is mapped into the page tables of one or more
729                  * processes. Try to unmap it here.
730                  */
731                 if (page_mapped(page) && mapping) {
732                         switch (try_to_unmap(page, TTU_UNMAP)) {
733                         case SWAP_FAIL:
734                                 goto activate_locked;
735                         case SWAP_AGAIN:
736                                 goto keep_locked;
737                         case SWAP_MLOCK:
738                                 goto cull_mlocked;
739                         case SWAP_SUCCESS:
740                                 ; /* try to free the page below */
741                         }
742                 }
743
744                 if (PageDirty(page)) {
745                         if (references == PAGEREF_RECLAIM_CLEAN)
746                                 goto keep_locked;
747                         if (!may_enter_fs)
748                                 goto keep_locked;
749                         if (!sc->may_writepage)
750                                 goto keep_locked;
751
752                         /* Page is dirty, try to write it out here */
753                         switch (pageout(page, mapping, sync_writeback)) {
754                         case PAGE_KEEP:
755                                 goto keep_locked;
756                         case PAGE_ACTIVATE:
757                                 goto activate_locked;
758                         case PAGE_SUCCESS:
759                                 if (PageWriteback(page) || PageDirty(page))
760                                         goto keep;
761                                 /*
762                                  * A synchronous write - probably a ramdisk.  Go
763                                  * ahead and try to reclaim the page.
764                                  */
765                                 if (!trylock_page(page))
766                                         goto keep;
767                                 if (PageDirty(page) || PageWriteback(page))
768                                         goto keep_locked;
769                                 mapping = page_mapping(page);
770                         case PAGE_CLEAN:
771                                 ; /* try to free the page below */
772                         }
773                 }
774
775                 /*
776                  * If the page has buffers, try to free the buffer mappings
777                  * associated with this page. If we succeed we try to free
778                  * the page as well.
779                  *
780                  * We do this even if the page is PageDirty().
781                  * try_to_release_page() does not perform I/O, but it is
782                  * possible for a page to have PageDirty set, but it is actually
783                  * clean (all its buffers are clean).  This happens if the
784                  * buffers were written out directly, with submit_bh(). ext3
785                  * will do this, as well as the blockdev mapping.
786                  * try_to_release_page() will discover that cleanness and will
787                  * drop the buffers and mark the page clean - it can be freed.
788                  *
789                  * Rarely, pages can have buffers and no ->mapping.  These are
790                  * the pages which were not successfully invalidated in
791                  * truncate_complete_page().  We try to drop those buffers here
792                  * and if that worked, and the page is no longer mapped into
793                  * process address space (page_count == 1) it can be freed.
794                  * Otherwise, leave the page on the LRU so it is swappable.
795                  */
796                 if (page_has_private(page)) {
797                         if (!try_to_release_page(page, sc->gfp_mask))
798                                 goto activate_locked;
799                         if (!mapping && page_count(page) == 1) {
800                                 unlock_page(page);
801                                 if (put_page_testzero(page))
802                                         goto free_it;
803                                 else {
804                                         /*
805                                          * rare race with speculative reference.
806                                          * the speculative reference will free
807                                          * this page shortly, so we may
808                                          * increment nr_reclaimed here (and
809                                          * leave it off the LRU).
810                                          */
811                                         nr_reclaimed++;
812                                         continue;
813                                 }
814                         }
815                 }
816
817                 if (!mapping || !__remove_mapping(mapping, page))
818                         goto keep_locked;
819
820                 /*
821                  * At this point, we have no other references and there is
822                  * no way to pick any more up (removed from LRU, removed
823                  * from pagecache). Can use non-atomic bitops now (and
824                  * we obviously don't have to worry about waking up a process
825                  * waiting on the page lock, because there are no references.
826                  */
827                 __clear_page_locked(page);
828 free_it:
829                 nr_reclaimed++;
830
831                 /*
832                  * Is there need to periodically free_page_list? It would
833                  * appear not as the counts should be low
834                  */
835                 list_add(&page->lru, &free_pages);
836                 continue;
837
838 cull_mlocked:
839                 if (PageSwapCache(page))
840                         try_to_free_swap(page);
841                 unlock_page(page);
842                 putback_lru_page(page);
843                 continue;
844
845 activate_locked:
846                 /* Not a candidate for swapping, so reclaim swap space. */
847                 if (PageSwapCache(page) && vm_swap_full())
848                         try_to_free_swap(page);
849                 VM_BUG_ON(PageActive(page));
850                 SetPageActive(page);
851                 pgactivate++;
852 keep_locked:
853                 unlock_page(page);
854 keep:
855                 list_add(&page->lru, &ret_pages);
856                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
857         }
858
859         free_page_list(&free_pages);
860
861         list_splice(&ret_pages, page_list);
862         count_vm_events(PGACTIVATE, pgactivate);
863         return nr_reclaimed;
864 }
865
866 /*
867  * Attempt to remove the specified page from its LRU.  Only take this page
868  * if it is of the appropriate PageActive status.  Pages which are being
869  * freed elsewhere are also ignored.
870  *
871  * page:        page to consider
872  * mode:        one of the LRU isolation modes defined above
873  *
874  * returns 0 on success, -ve errno on failure.
875  */
876 int __isolate_lru_page(struct page *page, int mode, int file)
877 {
878         int ret = -EINVAL;
879
880         /* Only take pages on the LRU. */
881         if (!PageLRU(page))
882                 return ret;
883
884         /*
885          * When checking the active state, we need to be sure we are
886          * dealing with comparible boolean values.  Take the logical not
887          * of each.
888          */
889         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
890                 return ret;
891
892         if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
893                 return ret;
894
895         /*
896          * When this function is being called for lumpy reclaim, we
897          * initially look into all LRU pages, active, inactive and
898          * unevictable; only give shrink_page_list evictable pages.
899          */
900         if (PageUnevictable(page))
901                 return ret;
902
903         ret = -EBUSY;
904
905         if (likely(get_page_unless_zero(page))) {
906                 /*
907                  * Be careful not to clear PageLRU until after we're
908                  * sure the page is not being freed elsewhere -- the
909                  * page release code relies on it.
910                  */
911                 ClearPageLRU(page);
912                 ret = 0;
913         }
914
915         return ret;
916 }
917
918 /*
919  * zone->lru_lock is heavily contended.  Some of the functions that
920  * shrink the lists perform better by taking out a batch of pages
921  * and working on them outside the LRU lock.
922  *
923  * For pagecache intensive workloads, this function is the hottest
924  * spot in the kernel (apart from copy_*_user functions).
925  *
926  * Appropriate locks must be held before calling this function.
927  *
928  * @nr_to_scan: The number of pages to look through on the list.
929  * @src:        The LRU list to pull pages off.
930  * @dst:        The temp list to put pages on to.
931  * @scanned:    The number of pages that were scanned.
932  * @order:      The caller's attempted allocation order
933  * @mode:       One of the LRU isolation modes
934  * @file:       True [1] if isolating file [!anon] pages
935  *
936  * returns how many pages were moved onto *@dst.
937  */
938 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
939                 struct list_head *src, struct list_head *dst,
940                 unsigned long *scanned, int order, int mode, int file)
941 {
942         unsigned long nr_taken = 0;
943         unsigned long nr_lumpy_taken = 0;
944         unsigned long nr_lumpy_dirty = 0;
945         unsigned long nr_lumpy_failed = 0;
946         unsigned long scan;
947
948         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
949                 struct page *page;
950                 unsigned long pfn;
951                 unsigned long end_pfn;
952                 unsigned long page_pfn;
953                 int zone_id;
954
955                 page = lru_to_page(src);
956                 prefetchw_prev_lru_page(page, src, flags);
957
958                 VM_BUG_ON(!PageLRU(page));
959
960                 switch (__isolate_lru_page(page, mode, file)) {
961                 case 0:
962                         list_move(&page->lru, dst);
963                         mem_cgroup_del_lru(page);
964                         nr_taken++;
965                         break;
966
967                 case -EBUSY:
968                         /* else it is being freed elsewhere */
969                         list_move(&page->lru, src);
970                         mem_cgroup_rotate_lru_list(page, page_lru(page));
971                         continue;
972
973                 default:
974                         BUG();
975                 }
976
977                 if (!order)
978                         continue;
979
980                 /*
981                  * Attempt to take all pages in the order aligned region
982                  * surrounding the tag page.  Only take those pages of
983                  * the same active state as that tag page.  We may safely
984                  * round the target page pfn down to the requested order
985                  * as the mem_map is guarenteed valid out to MAX_ORDER,
986                  * where that page is in a different zone we will detect
987                  * it from its zone id and abort this block scan.
988                  */
989                 zone_id = page_zone_id(page);
990                 page_pfn = page_to_pfn(page);
991                 pfn = page_pfn & ~((1 << order) - 1);
992                 end_pfn = pfn + (1 << order);
993                 for (; pfn < end_pfn; pfn++) {
994                         struct page *cursor_page;
995
996                         /* The target page is in the block, ignore it. */
997                         if (unlikely(pfn == page_pfn))
998                                 continue;
999
1000                         /* Avoid holes within the zone. */
1001                         if (unlikely(!pfn_valid_within(pfn)))
1002                                 break;
1003
1004                         cursor_page = pfn_to_page(pfn);
1005
1006                         /* Check that we have not crossed a zone boundary. */
1007                         if (unlikely(page_zone_id(cursor_page) != zone_id))
1008                                 continue;
1009
1010                         /*
1011                          * If we don't have enough swap space, reclaiming of
1012                          * anon page which don't already have a swap slot is
1013                          * pointless.
1014                          */
1015                         if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1016                                         !PageSwapCache(cursor_page))
1017                                 continue;
1018
1019                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1020                                 list_move(&cursor_page->lru, dst);
1021                                 mem_cgroup_del_lru(cursor_page);
1022                                 nr_taken++;
1023                                 nr_lumpy_taken++;
1024                                 if (PageDirty(cursor_page))
1025                                         nr_lumpy_dirty++;
1026                                 scan++;
1027                         } else {
1028                                 if (mode == ISOLATE_BOTH &&
1029                                                 page_count(cursor_page))
1030                                         nr_lumpy_failed++;
1031                         }
1032                 }
1033         }
1034
1035         *scanned = scan;
1036
1037         trace_mm_vmscan_lru_isolate(order,
1038                         nr_to_scan, scan,
1039                         nr_taken,
1040                         nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1041                         mode);
1042         return nr_taken;
1043 }
1044
1045 static unsigned long isolate_pages_global(unsigned long nr,
1046                                         struct list_head *dst,
1047                                         unsigned long *scanned, int order,
1048                                         int mode, struct zone *z,
1049                                         int active, int file)
1050 {
1051         int lru = LRU_BASE;
1052         if (active)
1053                 lru += LRU_ACTIVE;
1054         if (file)
1055                 lru += LRU_FILE;
1056         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1057                                                                 mode, file);
1058 }
1059
1060 /*
1061  * clear_active_flags() is a helper for shrink_active_list(), clearing
1062  * any active bits from the pages in the list.
1063  */
1064 static unsigned long clear_active_flags(struct list_head *page_list,
1065                                         unsigned int *count)
1066 {
1067         int nr_active = 0;
1068         int lru;
1069         struct page *page;
1070
1071         list_for_each_entry(page, page_list, lru) {
1072                 lru = page_lru_base_type(page);
1073                 if (PageActive(page)) {
1074                         lru += LRU_ACTIVE;
1075                         ClearPageActive(page);
1076                         nr_active++;
1077                 }
1078                 if (count)
1079                         count[lru]++;
1080         }
1081
1082         return nr_active;
1083 }
1084
1085 /**
1086  * isolate_lru_page - tries to isolate a page from its LRU list
1087  * @page: page to isolate from its LRU list
1088  *
1089  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1090  * vmstat statistic corresponding to whatever LRU list the page was on.
1091  *
1092  * Returns 0 if the page was removed from an LRU list.
1093  * Returns -EBUSY if the page was not on an LRU list.
1094  *
1095  * The returned page will have PageLRU() cleared.  If it was found on
1096  * the active list, it will have PageActive set.  If it was found on
1097  * the unevictable list, it will have the PageUnevictable bit set. That flag
1098  * may need to be cleared by the caller before letting the page go.
1099  *
1100  * The vmstat statistic corresponding to the list on which the page was
1101  * found will be decremented.
1102  *
1103  * Restrictions:
1104  * (1) Must be called with an elevated refcount on the page. This is a
1105  *     fundamentnal difference from isolate_lru_pages (which is called
1106  *     without a stable reference).
1107  * (2) the lru_lock must not be held.
1108  * (3) interrupts must be enabled.
1109  */
1110 int isolate_lru_page(struct page *page)
1111 {
1112         int ret = -EBUSY;
1113
1114         if (PageLRU(page)) {
1115                 struct zone *zone = page_zone(page);
1116
1117                 spin_lock_irq(&zone->lru_lock);
1118                 if (PageLRU(page) && get_page_unless_zero(page)) {
1119                         int lru = page_lru(page);
1120                         ret = 0;
1121                         ClearPageLRU(page);
1122
1123                         del_page_from_lru_list(zone, page, lru);
1124                 }
1125                 spin_unlock_irq(&zone->lru_lock);
1126         }
1127         return ret;
1128 }
1129
1130 /*
1131  * Are there way too many processes in the direct reclaim path already?
1132  */
1133 static int too_many_isolated(struct zone *zone, int file,
1134                 struct scan_control *sc)
1135 {
1136         unsigned long inactive, isolated;
1137
1138         if (current_is_kswapd())
1139                 return 0;
1140
1141         if (!scanning_global_lru(sc))
1142                 return 0;
1143
1144         if (file) {
1145                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1146                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1147         } else {
1148                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1149                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1150         }
1151
1152         return isolated > inactive;
1153 }
1154
1155 /*
1156  * TODO: Try merging with migrations version of putback_lru_pages
1157  */
1158 static noinline_for_stack void
1159 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1160                                 unsigned long nr_anon, unsigned long nr_file,
1161                                 struct list_head *page_list)
1162 {
1163         struct page *page;
1164         struct pagevec pvec;
1165         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1166
1167         pagevec_init(&pvec, 1);
1168
1169         /*
1170          * Put back any unfreeable pages.
1171          */
1172         spin_lock(&zone->lru_lock);
1173         while (!list_empty(page_list)) {
1174                 int lru;
1175                 page = lru_to_page(page_list);
1176                 VM_BUG_ON(PageLRU(page));
1177                 list_del(&page->lru);
1178                 if (unlikely(!page_evictable(page, NULL))) {
1179                         spin_unlock_irq(&zone->lru_lock);
1180                         putback_lru_page(page);
1181                         spin_lock_irq(&zone->lru_lock);
1182                         continue;
1183                 }
1184                 SetPageLRU(page);
1185                 lru = page_lru(page);
1186                 add_page_to_lru_list(zone, page, lru);
1187                 if (is_active_lru(lru)) {
1188                         int file = is_file_lru(lru);
1189                         reclaim_stat->recent_rotated[file]++;
1190                 }
1191                 if (!pagevec_add(&pvec, page)) {
1192                         spin_unlock_irq(&zone->lru_lock);
1193                         __pagevec_release(&pvec);
1194                         spin_lock_irq(&zone->lru_lock);
1195                 }
1196         }
1197         __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1198         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1199
1200         spin_unlock_irq(&zone->lru_lock);
1201         pagevec_release(&pvec);
1202 }
1203
1204 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1205                                         struct scan_control *sc,
1206                                         unsigned long *nr_anon,
1207                                         unsigned long *nr_file,
1208                                         struct list_head *isolated_list)
1209 {
1210         unsigned long nr_active;
1211         unsigned int count[NR_LRU_LISTS] = { 0, };
1212         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1213
1214         nr_active = clear_active_flags(isolated_list, count);
1215         __count_vm_events(PGDEACTIVATE, nr_active);
1216
1217         __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1218                               -count[LRU_ACTIVE_FILE]);
1219         __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1220                               -count[LRU_INACTIVE_FILE]);
1221         __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1222                               -count[LRU_ACTIVE_ANON]);
1223         __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1224                               -count[LRU_INACTIVE_ANON]);
1225
1226         *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1227         *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1228         __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1229         __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1230
1231         reclaim_stat->recent_scanned[0] += *nr_anon;
1232         reclaim_stat->recent_scanned[1] += *nr_file;
1233 }
1234
1235 /*
1236  * Returns true if the caller should wait to clean dirty/writeback pages.
1237  *
1238  * If we are direct reclaiming for contiguous pages and we do not reclaim
1239  * everything in the list, try again and wait for writeback IO to complete.
1240  * This will stall high-order allocations noticeably. Only do that when really
1241  * need to free the pages under high memory pressure.
1242  */
1243 static inline bool should_reclaim_stall(unsigned long nr_taken,
1244                                         unsigned long nr_freed,
1245                                         int priority,
1246                                         struct scan_control *sc)
1247 {
1248         int lumpy_stall_priority;
1249
1250         /* kswapd should not stall on sync IO */
1251         if (current_is_kswapd())
1252                 return false;
1253
1254         /* Only stall on lumpy reclaim */
1255         if (!sc->lumpy_reclaim_mode)
1256                 return false;
1257
1258         /* If we have relaimed everything on the isolated list, no stall */
1259         if (nr_freed == nr_taken)
1260                 return false;
1261
1262         /*
1263          * For high-order allocations, there are two stall thresholds.
1264          * High-cost allocations stall immediately where as lower
1265          * order allocations such as stacks require the scanning
1266          * priority to be much higher before stalling.
1267          */
1268         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1269                 lumpy_stall_priority = DEF_PRIORITY;
1270         else
1271                 lumpy_stall_priority = DEF_PRIORITY / 3;
1272
1273         return priority <= lumpy_stall_priority;
1274 }
1275
1276 /*
1277  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1278  * of reclaimed pages
1279  */
1280 static noinline_for_stack unsigned long
1281 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1282                         struct scan_control *sc, int priority, int file)
1283 {
1284         LIST_HEAD(page_list);
1285         unsigned long nr_scanned;
1286         unsigned long nr_reclaimed = 0;
1287         unsigned long nr_taken;
1288         unsigned long nr_active;
1289         unsigned long nr_anon;
1290         unsigned long nr_file;
1291
1292         while (unlikely(too_many_isolated(zone, file, sc))) {
1293                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1294
1295                 /* We are about to die and free our memory. Return now. */
1296                 if (fatal_signal_pending(current))
1297                         return SWAP_CLUSTER_MAX;
1298         }
1299
1300
1301         lru_add_drain();
1302         spin_lock_irq(&zone->lru_lock);
1303
1304         if (scanning_global_lru(sc)) {
1305                 nr_taken = isolate_pages_global(nr_to_scan,
1306                         &page_list, &nr_scanned, sc->order,
1307                         sc->lumpy_reclaim_mode ?
1308                                 ISOLATE_BOTH : ISOLATE_INACTIVE,
1309                         zone, 0, file);
1310                 zone->pages_scanned += nr_scanned;
1311                 if (current_is_kswapd())
1312                         __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1313                                                nr_scanned);
1314                 else
1315                         __count_zone_vm_events(PGSCAN_DIRECT, zone,
1316                                                nr_scanned);
1317         } else {
1318                 nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1319                         &page_list, &nr_scanned, sc->order,
1320                         sc->lumpy_reclaim_mode ?
1321                                 ISOLATE_BOTH : ISOLATE_INACTIVE,
1322                         zone, sc->mem_cgroup,
1323                         0, file);
1324                 /*
1325                  * mem_cgroup_isolate_pages() keeps track of
1326                  * scanned pages on its own.
1327                  */
1328         }
1329
1330         if (nr_taken == 0) {
1331                 spin_unlock_irq(&zone->lru_lock);
1332                 return 0;
1333         }
1334
1335         update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1336
1337         spin_unlock_irq(&zone->lru_lock);
1338
1339         nr_reclaimed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1340
1341         /* Check if we should syncronously wait for writeback */
1342         if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1343                 /*
1344                  * The attempt at page out may have made some
1345                  * of the pages active, mark them inactive again.
1346                  */
1347                 nr_active = clear_active_flags(&page_list, NULL);
1348                 count_vm_events(PGDEACTIVATE, nr_active);
1349
1350                 nr_reclaimed += shrink_page_list(&page_list, sc, PAGEOUT_IO_SYNC);
1351         }
1352
1353         local_irq_disable();
1354         if (current_is_kswapd())
1355                 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1356         __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1357
1358         putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1359
1360         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1361                 zone_idx(zone),
1362                 nr_scanned, nr_reclaimed,
1363                 priority,
1364                 trace_shrink_flags(file, sc->lumpy_reclaim_mode));
1365         return nr_reclaimed;
1366 }
1367
1368 /*
1369  * This moves pages from the active list to the inactive list.
1370  *
1371  * We move them the other way if the page is referenced by one or more
1372  * processes, from rmap.
1373  *
1374  * If the pages are mostly unmapped, the processing is fast and it is
1375  * appropriate to hold zone->lru_lock across the whole operation.  But if
1376  * the pages are mapped, the processing is slow (page_referenced()) so we
1377  * should drop zone->lru_lock around each page.  It's impossible to balance
1378  * this, so instead we remove the pages from the LRU while processing them.
1379  * It is safe to rely on PG_active against the non-LRU pages in here because
1380  * nobody will play with that bit on a non-LRU page.
1381  *
1382  * The downside is that we have to touch page->_count against each page.
1383  * But we had to alter page->flags anyway.
1384  */
1385
1386 static void move_active_pages_to_lru(struct zone *zone,
1387                                      struct list_head *list,
1388                                      enum lru_list lru)
1389 {
1390         unsigned long pgmoved = 0;
1391         struct pagevec pvec;
1392         struct page *page;
1393
1394         pagevec_init(&pvec, 1);
1395
1396         while (!list_empty(list)) {
1397                 page = lru_to_page(list);
1398
1399                 VM_BUG_ON(PageLRU(page));
1400                 SetPageLRU(page);
1401
1402                 list_move(&page->lru, &zone->lru[lru].list);
1403                 mem_cgroup_add_lru_list(page, lru);
1404                 pgmoved++;
1405
1406                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1407                         spin_unlock_irq(&zone->lru_lock);
1408                         if (buffer_heads_over_limit)
1409                                 pagevec_strip(&pvec);
1410                         __pagevec_release(&pvec);
1411                         spin_lock_irq(&zone->lru_lock);
1412                 }
1413         }
1414         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1415         if (!is_active_lru(lru))
1416                 __count_vm_events(PGDEACTIVATE, pgmoved);
1417 }
1418
1419 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1420                         struct scan_control *sc, int priority, int file)
1421 {
1422         unsigned long nr_taken;
1423         unsigned long pgscanned;
1424         unsigned long vm_flags;
1425         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1426         LIST_HEAD(l_active);
1427         LIST_HEAD(l_inactive);
1428         struct page *page;
1429         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1430         unsigned long nr_rotated = 0;
1431
1432         lru_add_drain();
1433         spin_lock_irq(&zone->lru_lock);
1434         if (scanning_global_lru(sc)) {
1435                 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1436                                                 &pgscanned, sc->order,
1437                                                 ISOLATE_ACTIVE, zone,
1438                                                 1, file);
1439                 zone->pages_scanned += pgscanned;
1440         } else {
1441                 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1442                                                 &pgscanned, sc->order,
1443                                                 ISOLATE_ACTIVE, zone,
1444                                                 sc->mem_cgroup, 1, file);
1445                 /*
1446                  * mem_cgroup_isolate_pages() keeps track of
1447                  * scanned pages on its own.
1448                  */
1449         }
1450
1451         reclaim_stat->recent_scanned[file] += nr_taken;
1452
1453         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1454         if (file)
1455                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1456         else
1457                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1458         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1459         spin_unlock_irq(&zone->lru_lock);
1460
1461         while (!list_empty(&l_hold)) {
1462                 cond_resched();
1463                 page = lru_to_page(&l_hold);
1464                 list_del(&page->lru);
1465
1466                 if (unlikely(!page_evictable(page, NULL))) {
1467                         putback_lru_page(page);
1468                         continue;
1469                 }
1470
1471                 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1472                         nr_rotated++;
1473                         /*
1474                          * Identify referenced, file-backed active pages and
1475                          * give them one more trip around the active list. So
1476                          * that executable code get better chances to stay in
1477                          * memory under moderate memory pressure.  Anon pages
1478                          * are not likely to be evicted by use-once streaming
1479                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1480                          * so we ignore them here.
1481                          */
1482                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1483                                 list_add(&page->lru, &l_active);
1484                                 continue;
1485                         }
1486                 }
1487
1488                 ClearPageActive(page);  /* we are de-activating */
1489                 list_add(&page->lru, &l_inactive);
1490         }
1491
1492         /*
1493          * Move pages back to the lru list.
1494          */
1495         spin_lock_irq(&zone->lru_lock);
1496         /*
1497          * Count referenced pages from currently used mappings as rotated,
1498          * even though only some of them are actually re-activated.  This
1499          * helps balance scan pressure between file and anonymous pages in
1500          * get_scan_ratio.
1501          */
1502         reclaim_stat->recent_rotated[file] += nr_rotated;
1503
1504         move_active_pages_to_lru(zone, &l_active,
1505                                                 LRU_ACTIVE + file * LRU_FILE);
1506         move_active_pages_to_lru(zone, &l_inactive,
1507                                                 LRU_BASE   + file * LRU_FILE);
1508         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1509         spin_unlock_irq(&zone->lru_lock);
1510 }
1511
1512 #ifdef CONFIG_SWAP
1513 static int inactive_anon_is_low_global(struct zone *zone)
1514 {
1515         unsigned long active, inactive;
1516
1517         active = zone_page_state(zone, NR_ACTIVE_ANON);
1518         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1519
1520         if (inactive * zone->inactive_ratio < active)
1521                 return 1;
1522
1523         return 0;
1524 }
1525
1526 /**
1527  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1528  * @zone: zone to check
1529  * @sc:   scan control of this context
1530  *
1531  * Returns true if the zone does not have enough inactive anon pages,
1532  * meaning some active anon pages need to be deactivated.
1533  */
1534 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1535 {
1536         int low;
1537
1538         /*
1539          * If we don't have swap space, anonymous page deactivation
1540          * is pointless.
1541          */
1542         if (!total_swap_pages)
1543                 return 0;
1544
1545         if (scanning_global_lru(sc))
1546                 low = inactive_anon_is_low_global(zone);
1547         else
1548                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1549         return low;
1550 }
1551 #else
1552 static inline int inactive_anon_is_low(struct zone *zone,
1553                                         struct scan_control *sc)
1554 {
1555         return 0;
1556 }
1557 #endif
1558
1559 static int inactive_file_is_low_global(struct zone *zone)
1560 {
1561         unsigned long active, inactive;
1562
1563         active = zone_page_state(zone, NR_ACTIVE_FILE);
1564         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1565
1566         return (active > inactive);
1567 }
1568
1569 /**
1570  * inactive_file_is_low - check if file pages need to be deactivated
1571  * @zone: zone to check
1572  * @sc:   scan control of this context
1573  *
1574  * When the system is doing streaming IO, memory pressure here
1575  * ensures that active file pages get deactivated, until more
1576  * than half of the file pages are on the inactive list.
1577  *
1578  * Once we get to that situation, protect the system's working
1579  * set from being evicted by disabling active file page aging.
1580  *
1581  * This uses a different ratio than the anonymous pages, because
1582  * the page cache uses a use-once replacement algorithm.
1583  */
1584 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1585 {
1586         int low;
1587
1588         if (scanning_global_lru(sc))
1589                 low = inactive_file_is_low_global(zone);
1590         else
1591                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1592         return low;
1593 }
1594
1595 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1596                                 int file)
1597 {
1598         if (file)
1599                 return inactive_file_is_low(zone, sc);
1600         else
1601                 return inactive_anon_is_low(zone, sc);
1602 }
1603
1604 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1605         struct zone *zone, struct scan_control *sc, int priority)
1606 {
1607         int file = is_file_lru(lru);
1608
1609         if (is_active_lru(lru)) {
1610                 if (inactive_list_is_low(zone, sc, file))
1611                     shrink_active_list(nr_to_scan, zone, sc, priority, file);
1612                 return 0;
1613         }
1614
1615         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1616 }
1617
1618 /*
1619  * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1620  * until we collected @swap_cluster_max pages to scan.
1621  */
1622 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1623                                        unsigned long *nr_saved_scan)
1624 {
1625         unsigned long nr;
1626
1627         *nr_saved_scan += nr_to_scan;
1628         nr = *nr_saved_scan;
1629
1630         if (nr >= SWAP_CLUSTER_MAX)
1631                 *nr_saved_scan = 0;
1632         else
1633                 nr = 0;
1634
1635         return nr;
1636 }
1637
1638 /*
1639  * Determine how aggressively the anon and file LRU lists should be
1640  * scanned.  The relative value of each set of LRU lists is determined
1641  * by looking at the fraction of the pages scanned we did rotate back
1642  * onto the active list instead of evict.
1643  *
1644  * nr[0] = anon pages to scan; nr[1] = file pages to scan
1645  */
1646 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1647                                         unsigned long *nr, int priority)
1648 {
1649         unsigned long anon, file, free;
1650         unsigned long anon_prio, file_prio;
1651         unsigned long ap, fp;
1652         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1653         u64 fraction[2], denominator;
1654         enum lru_list l;
1655         int noswap = 0;
1656
1657         /* If we have no swap space, do not bother scanning anon pages. */
1658         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1659                 noswap = 1;
1660                 fraction[0] = 0;
1661                 fraction[1] = 1;
1662                 denominator = 1;
1663                 goto out;
1664         }
1665
1666         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1667                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1668         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1669                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1670
1671         if (scanning_global_lru(sc)) {
1672                 free  = zone_page_state(zone, NR_FREE_PAGES);
1673                 /* If we have very few page cache pages,
1674                    force-scan anon pages. */
1675                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1676                         fraction[0] = 1;
1677                         fraction[1] = 0;
1678                         denominator = 1;
1679                         goto out;
1680                 }
1681         }
1682
1683         /*
1684          * With swappiness at 100, anonymous and file have the same priority.
1685          * This scanning priority is essentially the inverse of IO cost.
1686          */
1687         anon_prio = sc->swappiness;
1688         file_prio = 200 - sc->swappiness;
1689
1690         /*
1691          * OK, so we have swap space and a fair amount of page cache
1692          * pages.  We use the recently rotated / recently scanned
1693          * ratios to determine how valuable each cache is.
1694          *
1695          * Because workloads change over time (and to avoid overflow)
1696          * we keep these statistics as a floating average, which ends
1697          * up weighing recent references more than old ones.
1698          *
1699          * anon in [0], file in [1]
1700          */
1701         spin_lock_irq(&zone->lru_lock);
1702         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1703                 reclaim_stat->recent_scanned[0] /= 2;
1704                 reclaim_stat->recent_rotated[0] /= 2;
1705         }
1706
1707         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1708                 reclaim_stat->recent_scanned[1] /= 2;
1709                 reclaim_stat->recent_rotated[1] /= 2;
1710         }
1711
1712         /*
1713          * The amount of pressure on anon vs file pages is inversely
1714          * proportional to the fraction of recently scanned pages on
1715          * each list that were recently referenced and in active use.
1716          */
1717         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1718         ap /= reclaim_stat->recent_rotated[0] + 1;
1719
1720         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1721         fp /= reclaim_stat->recent_rotated[1] + 1;
1722         spin_unlock_irq(&zone->lru_lock);
1723
1724         fraction[0] = ap;
1725         fraction[1] = fp;
1726         denominator = ap + fp + 1;
1727 out:
1728         for_each_evictable_lru(l) {
1729                 int file = is_file_lru(l);
1730                 unsigned long scan;
1731
1732                 scan = zone_nr_lru_pages(zone, sc, l);
1733                 if (priority || noswap) {
1734                         scan >>= priority;
1735                         scan = div64_u64(scan * fraction[file], denominator);
1736                 }
1737                 nr[l] = nr_scan_try_batch(scan,
1738                                           &reclaim_stat->nr_saved_scan[l]);
1739         }
1740 }
1741
1742 static void set_lumpy_reclaim_mode(int priority, struct scan_control *sc)
1743 {
1744         /*
1745          * If we need a large contiguous chunk of memory, or have
1746          * trouble getting a small set of contiguous pages, we
1747          * will reclaim both active and inactive pages.
1748          */
1749         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1750                 sc->lumpy_reclaim_mode = 1;
1751         else if (sc->order && priority < DEF_PRIORITY - 2)
1752                 sc->lumpy_reclaim_mode = 1;
1753         else
1754                 sc->lumpy_reclaim_mode = 0;
1755 }
1756
1757 /*
1758  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1759  */
1760 static void shrink_zone(int priority, struct zone *zone,
1761                                 struct scan_control *sc)
1762 {
1763         unsigned long nr[NR_LRU_LISTS];
1764         unsigned long nr_to_scan;
1765         enum lru_list l;
1766         unsigned long nr_reclaimed = sc->nr_reclaimed;
1767         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1768
1769         get_scan_count(zone, sc, nr, priority);
1770
1771         set_lumpy_reclaim_mode(priority, sc);
1772
1773         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1774                                         nr[LRU_INACTIVE_FILE]) {
1775                 for_each_evictable_lru(l) {
1776                         if (nr[l]) {
1777                                 nr_to_scan = min_t(unsigned long,
1778                                                    nr[l], SWAP_CLUSTER_MAX);
1779                                 nr[l] -= nr_to_scan;
1780
1781                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1782                                                             zone, sc, priority);
1783                         }
1784                 }
1785                 /*
1786                  * On large memory systems, scan >> priority can become
1787                  * really large. This is fine for the starting priority;
1788                  * we want to put equal scanning pressure on each zone.
1789                  * However, if the VM has a harder time of freeing pages,
1790                  * with multiple processes reclaiming pages, the total
1791                  * freeing target can get unreasonably large.
1792                  */
1793                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
1794                         break;
1795         }
1796
1797         sc->nr_reclaimed = nr_reclaimed;
1798
1799         /*
1800          * Even if we did not try to evict anon pages at all, we want to
1801          * rebalance the anon lru active/inactive ratio.
1802          */
1803         if (inactive_anon_is_low(zone, sc))
1804                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1805
1806         throttle_vm_writeout(sc->gfp_mask);
1807 }
1808
1809 /*
1810  * This is the direct reclaim path, for page-allocating processes.  We only
1811  * try to reclaim pages from zones which will satisfy the caller's allocation
1812  * request.
1813  *
1814  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1815  * Because:
1816  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1817  *    allocation or
1818  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1819  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1820  *    zone defense algorithm.
1821  *
1822  * If a zone is deemed to be full of pinned pages then just give it a light
1823  * scan then give up on it.
1824  */
1825 static void shrink_zones(int priority, struct zonelist *zonelist,
1826                                         struct scan_control *sc)
1827 {
1828         struct zoneref *z;
1829         struct zone *zone;
1830
1831         for_each_zone_zonelist_nodemask(zone, z, zonelist,
1832                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
1833                 if (!populated_zone(zone))
1834                         continue;
1835                 /*
1836                  * Take care memory controller reclaiming has small influence
1837                  * to global LRU.
1838                  */
1839                 if (scanning_global_lru(sc)) {
1840                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1841                                 continue;
1842                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1843                                 continue;       /* Let kswapd poll it */
1844                 }
1845
1846                 shrink_zone(priority, zone, sc);
1847         }
1848 }
1849
1850 static bool zone_reclaimable(struct zone *zone)
1851 {
1852         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
1853 }
1854
1855 /*
1856  * As hibernation is going on, kswapd is freezed so that it can't mark
1857  * the zone into all_unreclaimable. It can't handle OOM during hibernation.
1858  * So let's check zone's unreclaimable in direct reclaim as well as kswapd.
1859  */
1860 static bool all_unreclaimable(struct zonelist *zonelist,
1861                 struct scan_control *sc)
1862 {
1863         struct zoneref *z;
1864         struct zone *zone;
1865         bool all_unreclaimable = true;
1866
1867         for_each_zone_zonelist_nodemask(zone, z, zonelist,
1868                         gfp_zone(sc->gfp_mask), sc->nodemask) {
1869                 if (!populated_zone(zone))
1870                         continue;
1871                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1872                         continue;
1873                 if (zone_reclaimable(zone)) {
1874                         all_unreclaimable = false;
1875                         break;
1876                 }
1877         }
1878
1879         return all_unreclaimable;
1880 }
1881
1882 /*
1883  * This is the main entry point to direct page reclaim.
1884  *
1885  * If a full scan of the inactive list fails to free enough memory then we
1886  * are "out of memory" and something needs to be killed.
1887  *
1888  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1889  * high - the zone may be full of dirty or under-writeback pages, which this
1890  * caller can't do much about.  We kick the writeback threads and take explicit
1891  * naps in the hope that some of these pages can be written.  But if the
1892  * allocating task holds filesystem locks which prevent writeout this might not
1893  * work, and the allocation attempt will fail.
1894  *
1895  * returns:     0, if no pages reclaimed
1896  *              else, the number of pages reclaimed
1897  */
1898 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
1899                                         struct scan_control *sc)
1900 {
1901         int priority;
1902         unsigned long total_scanned = 0;
1903         struct reclaim_state *reclaim_state = current->reclaim_state;
1904         struct zoneref *z;
1905         struct zone *zone;
1906         unsigned long writeback_threshold;
1907
1908         get_mems_allowed();
1909         delayacct_freepages_start();
1910
1911         if (scanning_global_lru(sc))
1912                 count_vm_event(ALLOCSTALL);
1913
1914         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1915                 sc->nr_scanned = 0;
1916                 if (!priority)
1917                         disable_swap_token();
1918                 shrink_zones(priority, zonelist, sc);
1919                 /*
1920                  * Don't shrink slabs when reclaiming memory from
1921                  * over limit cgroups
1922                  */
1923                 if (scanning_global_lru(sc)) {
1924                         unsigned long lru_pages = 0;
1925                         for_each_zone_zonelist(zone, z, zonelist,
1926                                         gfp_zone(sc->gfp_mask)) {
1927                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1928                                         continue;
1929
1930                                 lru_pages += zone_reclaimable_pages(zone);
1931                         }
1932
1933                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
1934                         if (reclaim_state) {
1935                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
1936                                 reclaim_state->reclaimed_slab = 0;
1937                         }
1938                 }
1939                 total_scanned += sc->nr_scanned;
1940                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
1941                         goto out;
1942
1943                 /*
1944                  * Try to write back as many pages as we just scanned.  This
1945                  * tends to cause slow streaming writers to write data to the
1946                  * disk smoothly, at the dirtying rate, which is nice.   But
1947                  * that's undesirable in laptop mode, where we *want* lumpy
1948                  * writeout.  So in laptop mode, write out the whole world.
1949                  */
1950                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
1951                 if (total_scanned > writeback_threshold) {
1952                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
1953                         sc->may_writepage = 1;
1954                 }
1955
1956                 /* Take a nap, wait for some writeback to complete */
1957                 if (!sc->hibernation_mode && sc->nr_scanned &&
1958                     priority < DEF_PRIORITY - 2)
1959                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1960         }
1961
1962 out:
1963         delayacct_freepages_end();
1964         put_mems_allowed();
1965
1966         if (sc->nr_reclaimed)
1967                 return sc->nr_reclaimed;
1968
1969         /* top priority shrink_zones still had more to do? don't OOM, then */
1970         if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
1971                 return 1;
1972
1973         return 0;
1974 }
1975
1976 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1977                                 gfp_t gfp_mask, nodemask_t *nodemask)
1978 {
1979         unsigned long nr_reclaimed;
1980         struct scan_control sc = {
1981                 .gfp_mask = gfp_mask,
1982                 .may_writepage = !laptop_mode,
1983                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
1984                 .may_unmap = 1,
1985                 .may_swap = 1,
1986                 .swappiness = vm_swappiness,
1987                 .order = order,
1988                 .mem_cgroup = NULL,
1989                 .nodemask = nodemask,
1990         };
1991
1992         trace_mm_vmscan_direct_reclaim_begin(order,
1993                                 sc.may_writepage,
1994                                 gfp_mask);
1995
1996         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
1997
1998         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
1999
2000         return nr_reclaimed;
2001 }
2002
2003 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2004
2005 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2006                                                 gfp_t gfp_mask, bool noswap,
2007                                                 unsigned int swappiness,
2008                                                 struct zone *zone)
2009 {
2010         struct scan_control sc = {
2011                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2012                 .may_writepage = !laptop_mode,
2013                 .may_unmap = 1,
2014                 .may_swap = !noswap,
2015                 .swappiness = swappiness,
2016                 .order = 0,
2017                 .mem_cgroup = mem,
2018         };
2019         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2020                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2021
2022         trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2023                                                       sc.may_writepage,
2024                                                       sc.gfp_mask);
2025
2026         /*
2027          * NOTE: Although we can get the priority field, using it
2028          * here is not a good idea, since it limits the pages we can scan.
2029          * if we don't reclaim here, the shrink_zone from balance_pgdat
2030          * will pick up pages from other mem cgroup's as well. We hack
2031          * the priority and make it zero.
2032          */
2033         shrink_zone(0, zone, &sc);
2034
2035         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2036
2037         return sc.nr_reclaimed;
2038 }
2039
2040 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2041                                            gfp_t gfp_mask,
2042                                            bool noswap,
2043                                            unsigned int swappiness)
2044 {
2045         struct zonelist *zonelist;
2046         unsigned long nr_reclaimed;
2047         struct scan_control sc = {
2048                 .may_writepage = !laptop_mode,
2049                 .may_unmap = 1,
2050                 .may_swap = !noswap,
2051                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2052                 .swappiness = swappiness,
2053                 .order = 0,
2054                 .mem_cgroup = mem_cont,
2055                 .nodemask = NULL, /* we don't care the placement */
2056         };
2057
2058         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2059                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2060         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
2061
2062         trace_mm_vmscan_memcg_reclaim_begin(0,
2063                                             sc.may_writepage,
2064                                             sc.gfp_mask);
2065
2066         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2067
2068         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2069
2070         return nr_reclaimed;
2071 }
2072 #endif
2073
2074 /* is kswapd sleeping prematurely? */
2075 static int sleeping_prematurely(pg_data_t *pgdat, int order, long remaining)
2076 {
2077         int i;
2078
2079         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2080         if (remaining)
2081                 return 1;
2082
2083         /* If after HZ/10, a zone is below the high mark, it's premature */
2084         for (i = 0; i < pgdat->nr_zones; i++) {
2085                 struct zone *zone = pgdat->node_zones + i;
2086
2087                 if (!populated_zone(zone))
2088                         continue;
2089
2090                 if (zone->all_unreclaimable)
2091                         continue;
2092
2093                 if (!zone_watermark_ok(zone, order, high_wmark_pages(zone),
2094                                                                 0, 0))
2095                         return 1;
2096         }
2097
2098         return 0;
2099 }
2100
2101 /*
2102  * For kswapd, balance_pgdat() will work across all this node's zones until
2103  * they are all at high_wmark_pages(zone).
2104  *
2105  * Returns the number of pages which were actually freed.
2106  *
2107  * There is special handling here for zones which are full of pinned pages.
2108  * This can happen if the pages are all mlocked, or if they are all used by
2109  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2110  * What we do is to detect the case where all pages in the zone have been
2111  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2112  * dead and from now on, only perform a short scan.  Basically we're polling
2113  * the zone for when the problem goes away.
2114  *
2115  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2116  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2117  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2118  * lower zones regardless of the number of free pages in the lower zones. This
2119  * interoperates with the page allocator fallback scheme to ensure that aging
2120  * of pages is balanced across the zones.
2121  */
2122 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
2123 {
2124         int all_zones_ok;
2125         int priority;
2126         int i;
2127         unsigned long total_scanned;
2128         struct reclaim_state *reclaim_state = current->reclaim_state;
2129         struct scan_control sc = {
2130                 .gfp_mask = GFP_KERNEL,
2131                 .may_unmap = 1,
2132                 .may_swap = 1,
2133                 /*
2134                  * kswapd doesn't want to be bailed out while reclaim. because
2135                  * we want to put equal scanning pressure on each zone.
2136                  */
2137                 .nr_to_reclaim = ULONG_MAX,
2138                 .swappiness = vm_swappiness,
2139                 .order = order,
2140                 .mem_cgroup = NULL,
2141         };
2142 loop_again:
2143         total_scanned = 0;
2144         sc.nr_reclaimed = 0;
2145         sc.may_writepage = !laptop_mode;
2146         count_vm_event(PAGEOUTRUN);
2147
2148         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2149                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2150                 unsigned long lru_pages = 0;
2151                 int has_under_min_watermark_zone = 0;
2152
2153                 /* The swap token gets in the way of swapout... */
2154                 if (!priority)
2155                         disable_swap_token();
2156
2157                 all_zones_ok = 1;
2158
2159                 /*
2160                  * Scan in the highmem->dma direction for the highest
2161                  * zone which needs scanning
2162                  */
2163                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2164                         struct zone *zone = pgdat->node_zones + i;
2165
2166                         if (!populated_zone(zone))
2167                                 continue;
2168
2169                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2170                                 continue;
2171
2172                         /*
2173                          * Do some background aging of the anon list, to give
2174                          * pages a chance to be referenced before reclaiming.
2175                          */
2176                         if (inactive_anon_is_low(zone, &sc))
2177                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2178                                                         &sc, priority, 0);
2179
2180                         if (!zone_watermark_ok(zone, order,
2181                                         high_wmark_pages(zone), 0, 0)) {
2182                                 end_zone = i;
2183                                 break;
2184                         }
2185                 }
2186                 if (i < 0)
2187                         goto out;
2188
2189                 for (i = 0; i <= end_zone; i++) {
2190                         struct zone *zone = pgdat->node_zones + i;
2191
2192                         lru_pages += zone_reclaimable_pages(zone);
2193                 }
2194
2195                 /*
2196                  * Now scan the zone in the dma->highmem direction, stopping
2197                  * at the last zone which needs scanning.
2198                  *
2199                  * We do this because the page allocator works in the opposite
2200                  * direction.  This prevents the page allocator from allocating
2201                  * pages behind kswapd's direction of progress, which would
2202                  * cause too much scanning of the lower zones.
2203                  */
2204                 for (i = 0; i <= end_zone; i++) {
2205                         struct zone *zone = pgdat->node_zones + i;
2206                         int nr_slab;
2207
2208                         if (!populated_zone(zone))
2209                                 continue;
2210
2211                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2212                                 continue;
2213
2214                         sc.nr_scanned = 0;
2215
2216                         /*
2217                          * Call soft limit reclaim before calling shrink_zone.
2218                          * For now we ignore the return value
2219                          */
2220                         mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask);
2221
2222                         /*
2223                          * We put equal pressure on every zone, unless one
2224                          * zone has way too many pages free already.
2225                          */
2226                         if (!zone_watermark_ok(zone, order,
2227                                         8*high_wmark_pages(zone), end_zone, 0))
2228                                 shrink_zone(priority, zone, &sc);
2229                         reclaim_state->reclaimed_slab = 0;
2230                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2231                                                 lru_pages);
2232                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2233                         total_scanned += sc.nr_scanned;
2234                         if (zone->all_unreclaimable)
2235                                 continue;
2236                         if (nr_slab == 0 && !zone_reclaimable(zone))
2237                                 zone->all_unreclaimable = 1;
2238                         /*
2239                          * If we've done a decent amount of scanning and
2240                          * the reclaim ratio is low, start doing writepage
2241                          * even in laptop mode
2242                          */
2243                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2244                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2245                                 sc.may_writepage = 1;
2246
2247                         if (!zone_watermark_ok(zone, order,
2248                                         high_wmark_pages(zone), end_zone, 0)) {
2249                                 all_zones_ok = 0;
2250                                 /*
2251                                  * We are still under min water mark.  This
2252                                  * means that we have a GFP_ATOMIC allocation
2253                                  * failure risk. Hurry up!
2254                                  */
2255                                 if (!zone_watermark_ok(zone, order,
2256                                             min_wmark_pages(zone), end_zone, 0))
2257                                         has_under_min_watermark_zone = 1;
2258                         }
2259
2260                 }
2261                 if (all_zones_ok)
2262                         break;          /* kswapd: all done */
2263                 /*
2264                  * OK, kswapd is getting into trouble.  Take a nap, then take
2265                  * another pass across the zones.
2266                  */
2267                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2268                         if (has_under_min_watermark_zone)
2269                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2270                         else
2271                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2272                 }
2273
2274                 /*
2275                  * We do this so kswapd doesn't build up large priorities for
2276                  * example when it is freeing in parallel with allocators. It
2277                  * matches the direct reclaim path behaviour in terms of impact
2278                  * on zone->*_priority.
2279                  */
2280                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2281                         break;
2282         }
2283 out:
2284         if (!all_zones_ok) {
2285                 cond_resched();
2286
2287                 try_to_freeze();
2288
2289                 /*
2290                  * Fragmentation may mean that the system cannot be
2291                  * rebalanced for high-order allocations in all zones.
2292                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2293                  * it means the zones have been fully scanned and are still
2294                  * not balanced. For high-order allocations, there is
2295                  * little point trying all over again as kswapd may
2296                  * infinite loop.
2297                  *
2298                  * Instead, recheck all watermarks at order-0 as they
2299                  * are the most important. If watermarks are ok, kswapd will go
2300                  * back to sleep. High-order users can still perform direct
2301                  * reclaim if they wish.
2302                  */
2303                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2304                         order = sc.order = 0;
2305
2306                 goto loop_again;
2307         }
2308
2309         return sc.nr_reclaimed;
2310 }
2311
2312 /*
2313  * The background pageout daemon, started as a kernel thread
2314  * from the init process.
2315  *
2316  * This basically trickles out pages so that we have _some_
2317  * free memory available even if there is no other activity
2318  * that frees anything up. This is needed for things like routing
2319  * etc, where we otherwise might have all activity going on in
2320  * asynchronous contexts that cannot page things out.
2321  *
2322  * If there are applications that are active memory-allocators
2323  * (most normal use), this basically shouldn't matter.
2324  */
2325 static int kswapd(void *p)
2326 {
2327         unsigned long order;
2328         pg_data_t *pgdat = (pg_data_t*)p;
2329         struct task_struct *tsk = current;
2330         DEFINE_WAIT(wait);
2331         struct reclaim_state reclaim_state = {
2332                 .reclaimed_slab = 0,
2333         };
2334         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2335
2336         lockdep_set_current_reclaim_state(GFP_KERNEL);
2337
2338         if (!cpumask_empty(cpumask))
2339                 set_cpus_allowed_ptr(tsk, cpumask);
2340         current->reclaim_state = &reclaim_state;
2341
2342         /*
2343          * Tell the memory management that we're a "memory allocator",
2344          * and that if we need more memory we should get access to it
2345          * regardless (see "__alloc_pages()"). "kswapd" should
2346          * never get caught in the normal page freeing logic.
2347          *
2348          * (Kswapd normally doesn't need memory anyway, but sometimes
2349          * you need a small amount of memory in order to be able to
2350          * page out something else, and this flag essentially protects
2351          * us from recursively trying to free more memory as we're
2352          * trying to free the first piece of memory in the first place).
2353          */
2354         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2355         set_freezable();
2356
2357         order = 0;
2358         for ( ; ; ) {
2359                 unsigned long new_order;
2360                 int ret;
2361
2362                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2363                 new_order = pgdat->kswapd_max_order;
2364                 pgdat->kswapd_max_order = 0;
2365                 if (order < new_order) {
2366                         /*
2367                          * Don't sleep if someone wants a larger 'order'
2368                          * allocation
2369                          */
2370                         order = new_order;
2371                 } else {
2372                         if (!freezing(current) && !kthread_should_stop()) {
2373                                 long remaining = 0;
2374
2375                                 /* Try to sleep for a short interval */
2376                                 if (!sleeping_prematurely(pgdat, order, remaining)) {
2377                                         remaining = schedule_timeout(HZ/10);
2378                                         finish_wait(&pgdat->kswapd_wait, &wait);
2379                                         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2380                                 }
2381
2382                                 /*
2383                                  * After a short sleep, check if it was a
2384                                  * premature sleep. If not, then go fully
2385                                  * to sleep until explicitly woken up
2386                                  */
2387                                 if (!sleeping_prematurely(pgdat, order, remaining)) {
2388                                         trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2389                                         schedule();
2390                                 } else {
2391                                         if (remaining)
2392                                                 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2393                                         else
2394                                                 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2395                                 }
2396                         }
2397
2398                         order = pgdat->kswapd_max_order;
2399                 }
2400                 finish_wait(&pgdat->kswapd_wait, &wait);
2401
2402                 ret = try_to_freeze();
2403                 if (kthread_should_stop())
2404                         break;
2405
2406                 /*
2407                  * We can speed up thawing tasks if we don't call balance_pgdat
2408                  * after returning from the refrigerator
2409                  */
2410                 if (!ret) {
2411                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2412                         balance_pgdat(pgdat, order);
2413                 }
2414         }
2415         return 0;
2416 }
2417
2418 /*
2419  * A zone is low on free memory, so wake its kswapd task to service it.
2420  */
2421 void wakeup_kswapd(struct zone *zone, int order)
2422 {
2423         pg_data_t *pgdat;
2424
2425         if (!populated_zone(zone))
2426                 return;
2427
2428         pgdat = zone->zone_pgdat;
2429         if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
2430                 return;
2431         if (pgdat->kswapd_max_order < order)
2432                 pgdat->kswapd_max_order = order;
2433         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2434         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2435                 return;
2436         if (!waitqueue_active(&pgdat->kswapd_wait))
2437                 return;
2438         wake_up_interruptible(&pgdat->kswapd_wait);
2439 }
2440
2441 /*
2442  * The reclaimable count would be mostly accurate.
2443  * The less reclaimable pages may be
2444  * - mlocked pages, which will be moved to unevictable list when encountered
2445  * - mapped pages, which may require several travels to be reclaimed
2446  * - dirty pages, which is not "instantly" reclaimable
2447  */
2448 unsigned long global_reclaimable_pages(void)
2449 {
2450         int nr;
2451
2452         nr = global_page_state(NR_ACTIVE_FILE) +
2453              global_page_state(NR_INACTIVE_FILE);
2454
2455         if (nr_swap_pages > 0)
2456                 nr += global_page_state(NR_ACTIVE_ANON) +
2457                       global_page_state(NR_INACTIVE_ANON);
2458
2459         return nr;
2460 }
2461
2462 unsigned long zone_reclaimable_pages(struct zone *zone)
2463 {
2464         int nr;
2465
2466         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2467              zone_page_state(zone, NR_INACTIVE_FILE);
2468
2469         if (nr_swap_pages > 0)
2470                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2471                       zone_page_state(zone, NR_INACTIVE_ANON);
2472
2473         return nr;
2474 }
2475
2476 #ifdef CONFIG_HIBERNATION
2477 /*
2478  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2479  * freed pages.
2480  *
2481  * Rather than trying to age LRUs the aim is to preserve the overall
2482  * LRU order by reclaiming preferentially
2483  * inactive > active > active referenced > active mapped
2484  */
2485 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2486 {
2487         struct reclaim_state reclaim_state;
2488         struct scan_control sc = {
2489                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2490                 .may_swap = 1,
2491                 .may_unmap = 1,
2492                 .may_writepage = 1,
2493                 .nr_to_reclaim = nr_to_reclaim,
2494                 .hibernation_mode = 1,
2495                 .swappiness = vm_swappiness,
2496                 .order = 0,
2497         };
2498         struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2499         struct task_struct *p = current;
2500         unsigned long nr_reclaimed;
2501
2502         p->flags |= PF_MEMALLOC;
2503         lockdep_set_current_reclaim_state(sc.gfp_mask);
2504         reclaim_state.reclaimed_slab = 0;
2505         p->reclaim_state = &reclaim_state;
2506
2507         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2508
2509         p->reclaim_state = NULL;
2510         lockdep_clear_current_reclaim_state();
2511         p->flags &= ~PF_MEMALLOC;
2512
2513         return nr_reclaimed;
2514 }
2515 #endif /* CONFIG_HIBERNATION */
2516
2517 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2518    not required for correctness.  So if the last cpu in a node goes
2519    away, we get changed to run anywhere: as the first one comes back,
2520    restore their cpu bindings. */
2521 static int __devinit cpu_callback(struct notifier_block *nfb,
2522                                   unsigned long action, void *hcpu)
2523 {
2524         int nid;
2525
2526         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2527                 for_each_node_state(nid, N_HIGH_MEMORY) {
2528                         pg_data_t *pgdat = NODE_DATA(nid);
2529                         const struct cpumask *mask;
2530
2531                         mask = cpumask_of_node(pgdat->node_id);
2532
2533                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2534                                 /* One of our CPUs online: restore mask */
2535                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2536                 }
2537         }
2538         return NOTIFY_OK;
2539 }
2540
2541 /*
2542  * This kswapd start function will be called by init and node-hot-add.
2543  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2544  */
2545 int kswapd_run(int nid)
2546 {
2547         pg_data_t *pgdat = NODE_DATA(nid);
2548         int ret = 0;
2549
2550         if (pgdat->kswapd)
2551                 return 0;
2552
2553         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2554         if (IS_ERR(pgdat->kswapd)) {
2555                 /* failure at boot is fatal */
2556                 BUG_ON(system_state == SYSTEM_BOOTING);
2557                 printk("Failed to start kswapd on node %d\n",nid);
2558                 ret = -1;
2559         }
2560         return ret;
2561 }
2562
2563 /*
2564  * Called by memory hotplug when all memory in a node is offlined.
2565  */
2566 void kswapd_stop(int nid)
2567 {
2568         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2569
2570         if (kswapd)
2571                 kthread_stop(kswapd);
2572 }
2573
2574 static int __init kswapd_init(void)
2575 {
2576         int nid;
2577
2578         swap_setup();
2579         for_each_node_state(nid, N_HIGH_MEMORY)
2580                 kswapd_run(nid);
2581         hotcpu_notifier(cpu_callback, 0);
2582         return 0;
2583 }
2584
2585 module_init(kswapd_init)
2586
2587 #ifdef CONFIG_NUMA
2588 /*
2589  * Zone reclaim mode
2590  *
2591  * If non-zero call zone_reclaim when the number of free pages falls below
2592  * the watermarks.
2593  */
2594 int zone_reclaim_mode __read_mostly;
2595
2596 #define RECLAIM_OFF 0
2597 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2598 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2599 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2600
2601 /*
2602  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2603  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2604  * a zone.
2605  */
2606 #define ZONE_RECLAIM_PRIORITY 4
2607
2608 /*
2609  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2610  * occur.
2611  */
2612 int sysctl_min_unmapped_ratio = 1;
2613
2614 /*
2615  * If the number of slab pages in a zone grows beyond this percentage then
2616  * slab reclaim needs to occur.
2617  */
2618 int sysctl_min_slab_ratio = 5;
2619
2620 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2621 {
2622         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2623         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2624                 zone_page_state(zone, NR_ACTIVE_FILE);
2625
2626         /*
2627          * It's possible for there to be more file mapped pages than
2628          * accounted for by the pages on the file LRU lists because
2629          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2630          */
2631         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2632 }
2633
2634 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2635 static long zone_pagecache_reclaimable(struct zone *zone)
2636 {
2637         long nr_pagecache_reclaimable;
2638         long delta = 0;
2639
2640         /*
2641          * If RECLAIM_SWAP is set, then all file pages are considered
2642          * potentially reclaimable. Otherwise, we have to worry about
2643          * pages like swapcache and zone_unmapped_file_pages() provides
2644          * a better estimate
2645          */
2646         if (zone_reclaim_mode & RECLAIM_SWAP)
2647                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2648         else
2649                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2650
2651         /* If we can't clean pages, remove dirty pages from consideration */
2652         if (!(zone_reclaim_mode & RECLAIM_WRITE))
2653                 delta += zone_page_state(zone, NR_FILE_DIRTY);
2654
2655         /* Watch for any possible underflows due to delta */
2656         if (unlikely(delta > nr_pagecache_reclaimable))
2657                 delta = nr_pagecache_reclaimable;
2658
2659         return nr_pagecache_reclaimable - delta;
2660 }
2661
2662 /*
2663  * Try to free up some pages from this zone through reclaim.
2664  */
2665 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2666 {
2667         /* Minimum pages needed in order to stay on node */
2668         const unsigned long nr_pages = 1 << order;
2669         struct task_struct *p = current;
2670         struct reclaim_state reclaim_state;
2671         int priority;
2672         struct scan_control sc = {
2673                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2674                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2675                 .may_swap = 1,
2676                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
2677                                        SWAP_CLUSTER_MAX),
2678                 .gfp_mask = gfp_mask,
2679                 .swappiness = vm_swappiness,
2680                 .order = order,
2681         };
2682         unsigned long nr_slab_pages0, nr_slab_pages1;
2683
2684         cond_resched();
2685         /*
2686          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2687          * and we also need to be able to write out pages for RECLAIM_WRITE
2688          * and RECLAIM_SWAP.
2689          */
2690         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2691         lockdep_set_current_reclaim_state(gfp_mask);
2692         reclaim_state.reclaimed_slab = 0;
2693         p->reclaim_state = &reclaim_state;
2694
2695         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
2696                 /*
2697                  * Free memory by calling shrink zone with increasing
2698                  * priorities until we have enough memory freed.
2699                  */
2700                 priority = ZONE_RECLAIM_PRIORITY;
2701                 do {
2702                         shrink_zone(priority, zone, &sc);
2703                         priority--;
2704                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2705         }
2706
2707         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2708         if (nr_slab_pages0 > zone->min_slab_pages) {
2709                 /*
2710                  * shrink_slab() does not currently allow us to determine how
2711                  * many pages were freed in this zone. So we take the current
2712                  * number of slab pages and shake the slab until it is reduced
2713                  * by the same nr_pages that we used for reclaiming unmapped
2714                  * pages.
2715                  *
2716                  * Note that shrink_slab will free memory on all zones and may
2717                  * take a long time.
2718                  */
2719                 for (;;) {
2720                         unsigned long lru_pages = zone_reclaimable_pages(zone);
2721
2722                         /* No reclaimable slab or very low memory pressure */
2723                         if (!shrink_slab(sc.nr_scanned, gfp_mask, lru_pages))
2724                                 break;
2725
2726                         /* Freed enough memory */
2727                         nr_slab_pages1 = zone_page_state(zone,
2728                                                         NR_SLAB_RECLAIMABLE);
2729                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
2730                                 break;
2731                 }
2732
2733                 /*
2734                  * Update nr_reclaimed by the number of slab pages we
2735                  * reclaimed from this zone.
2736                  */
2737                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2738                 if (nr_slab_pages1 < nr_slab_pages0)
2739                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
2740         }
2741
2742         p->reclaim_state = NULL;
2743         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2744         lockdep_clear_current_reclaim_state();
2745         return sc.nr_reclaimed >= nr_pages;
2746 }
2747
2748 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2749 {
2750         int node_id;
2751         int ret;
2752
2753         /*
2754          * Zone reclaim reclaims unmapped file backed pages and
2755          * slab pages if we are over the defined limits.
2756          *
2757          * A small portion of unmapped file backed pages is needed for
2758          * file I/O otherwise pages read by file I/O will be immediately
2759          * thrown out if the zone is overallocated. So we do not reclaim
2760          * if less than a specified percentage of the zone is used by
2761          * unmapped file backed pages.
2762          */
2763         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2764             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
2765                 return ZONE_RECLAIM_FULL;
2766
2767         if (zone->all_unreclaimable)
2768                 return ZONE_RECLAIM_FULL;
2769
2770         /*
2771          * Do not scan if the allocation should not be delayed.
2772          */
2773         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2774                 return ZONE_RECLAIM_NOSCAN;
2775
2776         /*
2777          * Only run zone reclaim on the local zone or on zones that do not
2778          * have associated processors. This will favor the local processor
2779          * over remote processors and spread off node memory allocations
2780          * as wide as possible.
2781          */
2782         node_id = zone_to_nid(zone);
2783         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2784                 return ZONE_RECLAIM_NOSCAN;
2785
2786         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2787                 return ZONE_RECLAIM_NOSCAN;
2788
2789         ret = __zone_reclaim(zone, gfp_mask, order);
2790         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2791
2792         if (!ret)
2793                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2794
2795         return ret;
2796 }
2797 #endif
2798
2799 /*
2800  * page_evictable - test whether a page is evictable
2801  * @page: the page to test
2802  * @vma: the VMA in which the page is or will be mapped, may be NULL
2803  *
2804  * Test whether page is evictable--i.e., should be placed on active/inactive
2805  * lists vs unevictable list.  The vma argument is !NULL when called from the
2806  * fault path to determine how to instantate a new page.
2807  *
2808  * Reasons page might not be evictable:
2809  * (1) page's mapping marked unevictable
2810  * (2) page is part of an mlocked VMA
2811  *
2812  */
2813 int page_evictable(struct page *page, struct vm_area_struct *vma)
2814 {
2815
2816         if (mapping_unevictable(page_mapping(page)))
2817                 return 0;
2818
2819         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2820                 return 0;
2821
2822         return 1;
2823 }
2824
2825 /**
2826  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2827  * @page: page to check evictability and move to appropriate lru list
2828  * @zone: zone page is in
2829  *
2830  * Checks a page for evictability and moves the page to the appropriate
2831  * zone lru list.
2832  *
2833  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2834  * have PageUnevictable set.
2835  */
2836 static void check_move_unevictable_page(struct page *page, struct zone *zone)
2837 {
2838         VM_BUG_ON(PageActive(page));
2839
2840 retry:
2841         ClearPageUnevictable(page);
2842         if (page_evictable(page, NULL)) {
2843                 enum lru_list l = page_lru_base_type(page);
2844
2845                 __dec_zone_state(zone, NR_UNEVICTABLE);
2846                 list_move(&page->lru, &zone->lru[l].list);
2847                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
2848                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2849                 __count_vm_event(UNEVICTABLE_PGRESCUED);
2850         } else {
2851                 /*
2852                  * rotate unevictable list
2853                  */
2854                 SetPageUnevictable(page);
2855                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2856                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
2857                 if (page_evictable(page, NULL))
2858                         goto retry;
2859         }
2860 }
2861
2862 /**
2863  * scan_mapping_unevictable_pages - scan an address space for evictable pages
2864  * @mapping: struct address_space to scan for evictable pages
2865  *
2866  * Scan all pages in mapping.  Check unevictable pages for
2867  * evictability and move them to the appropriate zone lru list.
2868  */
2869 void scan_mapping_unevictable_pages(struct address_space *mapping)
2870 {
2871         pgoff_t next = 0;
2872         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2873                          PAGE_CACHE_SHIFT;
2874         struct zone *zone;
2875         struct pagevec pvec;
2876
2877         if (mapping->nrpages == 0)
2878                 return;
2879
2880         pagevec_init(&pvec, 0);
2881         while (next < end &&
2882                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2883                 int i;
2884                 int pg_scanned = 0;
2885
2886                 zone = NULL;
2887
2888                 for (i = 0; i < pagevec_count(&pvec); i++) {
2889                         struct page *page = pvec.pages[i];
2890                         pgoff_t page_index = page->index;
2891                         struct zone *pagezone = page_zone(page);
2892
2893                         pg_scanned++;
2894                         if (page_index > next)
2895                                 next = page_index;
2896                         next++;
2897
2898                         if (pagezone != zone) {
2899                                 if (zone)
2900                                         spin_unlock_irq(&zone->lru_lock);
2901                                 zone = pagezone;
2902                                 spin_lock_irq(&zone->lru_lock);
2903                         }
2904
2905                         if (PageLRU(page) && PageUnevictable(page))
2906                                 check_move_unevictable_page(page, zone);
2907                 }
2908                 if (zone)
2909                         spin_unlock_irq(&zone->lru_lock);
2910                 pagevec_release(&pvec);
2911
2912                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2913         }
2914
2915 }
2916
2917 /**
2918  * scan_zone_unevictable_pages - check unevictable list for evictable pages
2919  * @zone - zone of which to scan the unevictable list
2920  *
2921  * Scan @zone's unevictable LRU lists to check for pages that have become
2922  * evictable.  Move those that have to @zone's inactive list where they
2923  * become candidates for reclaim, unless shrink_inactive_zone() decides
2924  * to reactivate them.  Pages that are still unevictable are rotated
2925  * back onto @zone's unevictable list.
2926  */
2927 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
2928 static void scan_zone_unevictable_pages(struct zone *zone)
2929 {
2930         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2931         unsigned long scan;
2932         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2933
2934         while (nr_to_scan > 0) {
2935                 unsigned long batch_size = min(nr_to_scan,
2936                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
2937
2938                 spin_lock_irq(&zone->lru_lock);
2939                 for (scan = 0;  scan < batch_size; scan++) {
2940                         struct page *page = lru_to_page(l_unevictable);
2941
2942                         if (!trylock_page(page))
2943                                 continue;
2944
2945                         prefetchw_prev_lru_page(page, l_unevictable, flags);
2946
2947                         if (likely(PageLRU(page) && PageUnevictable(page)))
2948                                 check_move_unevictable_page(page, zone);
2949
2950                         unlock_page(page);
2951                 }
2952                 spin_unlock_irq(&zone->lru_lock);
2953
2954                 nr_to_scan -= batch_size;
2955         }
2956 }
2957
2958
2959 /**
2960  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2961  *
2962  * A really big hammer:  scan all zones' unevictable LRU lists to check for
2963  * pages that have become evictable.  Move those back to the zones'
2964  * inactive list where they become candidates for reclaim.
2965  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2966  * and we add swap to the system.  As such, it runs in the context of a task
2967  * that has possibly/probably made some previously unevictable pages
2968  * evictable.
2969  */
2970 static void scan_all_zones_unevictable_pages(void)
2971 {
2972         struct zone *zone;
2973
2974         for_each_zone(zone) {
2975                 scan_zone_unevictable_pages(zone);
2976         }
2977 }
2978
2979 /*
2980  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
2981  * all nodes' unevictable lists for evictable pages
2982  */
2983 unsigned long scan_unevictable_pages;
2984
2985 int scan_unevictable_handler(struct ctl_table *table, int write,
2986                            void __user *buffer,
2987                            size_t *length, loff_t *ppos)
2988 {
2989         proc_doulongvec_minmax(table, write, buffer, length, ppos);
2990
2991         if (write && *(unsigned long *)table->data)
2992                 scan_all_zones_unevictable_pages();
2993
2994         scan_unevictable_pages = 0;
2995         return 0;
2996 }
2997
2998 #ifdef CONFIG_NUMA
2999 /*
3000  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3001  * a specified node's per zone unevictable lists for evictable pages.
3002  */
3003
3004 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3005                                           struct sysdev_attribute *attr,
3006                                           char *buf)
3007 {
3008         return sprintf(buf, "0\n");     /* always zero; should fit... */
3009 }
3010
3011 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3012                                            struct sysdev_attribute *attr,
3013                                         const char *buf, size_t count)
3014 {
3015         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3016         struct zone *zone;
3017         unsigned long res;
3018         unsigned long req = strict_strtoul(buf, 10, &res);
3019
3020         if (!req)
3021                 return 1;       /* zero is no-op */
3022
3023         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3024                 if (!populated_zone(zone))
3025                         continue;
3026                 scan_zone_unevictable_pages(zone);
3027         }
3028         return 1;
3029 }
3030
3031
3032 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3033                         read_scan_unevictable_node,
3034                         write_scan_unevictable_node);
3035
3036 int scan_unevictable_register_node(struct node *node)
3037 {
3038         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3039 }
3040
3041 void scan_unevictable_unregister_node(struct node *node)
3042 {
3043         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
3044 }
3045 #endif