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