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