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