mm: vmscan: rename lumpy_mode to reclaim_mode
[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/compaction.h>
36 #include <linux/notifier.h>
37 #include <linux/rwsem.h>
38 #include <linux/delay.h>
39 #include <linux/kthread.h>
40 #include <linux/freezer.h>
41 #include <linux/memcontrol.h>
42 #include <linux/delayacct.h>
43 #include <linux/sysctl.h>
44
45 #include <asm/tlbflush.h>
46 #include <asm/div64.h>
47
48 #include <linux/swapops.h>
49
50 #include "internal.h"
51
52 #define CREATE_TRACE_POINTS
53 #include <trace/events/vmscan.h>
54
55 /*
56  * reclaim_mode determines how the inactive list is shrunk
57  * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
58  * RECLAIM_MODE_ASYNC:  Do not block
59  * RECLAIM_MODE_SYNC:   Allow blocking e.g. call wait_on_page_writeback
60  * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
61  *                      page from the LRU and reclaim all pages within a
62  *                      naturally aligned range
63  * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
64  *                      order-0 pages and then compact the zone
65  */
66 typedef unsigned __bitwise__ reclaim_mode_t;
67 #define RECLAIM_MODE_SINGLE             ((__force reclaim_mode_t)0x01u)
68 #define RECLAIM_MODE_ASYNC              ((__force reclaim_mode_t)0x02u)
69 #define RECLAIM_MODE_SYNC               ((__force reclaim_mode_t)0x04u)
70 #define RECLAIM_MODE_LUMPYRECLAIM       ((__force reclaim_mode_t)0x08u)
71 #define RECLAIM_MODE_COMPACTION         ((__force reclaim_mode_t)0x10u)
72
73 struct scan_control {
74         /* Incremented by the number of inactive pages that were scanned */
75         unsigned long nr_scanned;
76
77         /* Number of pages freed so far during a call to shrink_zones() */
78         unsigned long nr_reclaimed;
79
80         /* How many pages shrink_list() should reclaim */
81         unsigned long nr_to_reclaim;
82
83         unsigned long hibernation_mode;
84
85         /* This context's GFP mask */
86         gfp_t gfp_mask;
87
88         int may_writepage;
89
90         /* Can mapped pages be reclaimed? */
91         int may_unmap;
92
93         /* Can pages be swapped as part of reclaim? */
94         int may_swap;
95
96         int swappiness;
97
98         int order;
99
100         /*
101          * Intend to reclaim enough continuous memory rather than reclaim
102          * enough amount of memory. i.e, mode for high order allocation.
103          */
104         reclaim_mode_t reclaim_mode;
105
106         /* Which cgroup do we reclaim from */
107         struct mem_cgroup *mem_cgroup;
108
109         /*
110          * Nodemask of nodes allowed by the caller. If NULL, all nodes
111          * are scanned.
112          */
113         nodemask_t      *nodemask;
114 };
115
116 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
117
118 #ifdef ARCH_HAS_PREFETCH
119 #define prefetch_prev_lru_page(_page, _base, _field)                    \
120         do {                                                            \
121                 if ((_page)->lru.prev != _base) {                       \
122                         struct page *prev;                              \
123                                                                         \
124                         prev = lru_to_page(&(_page->lru));              \
125                         prefetch(&prev->_field);                        \
126                 }                                                       \
127         } while (0)
128 #else
129 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
130 #endif
131
132 #ifdef ARCH_HAS_PREFETCHW
133 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
134         do {                                                            \
135                 if ((_page)->lru.prev != _base) {                       \
136                         struct page *prev;                              \
137                                                                         \
138                         prev = lru_to_page(&(_page->lru));              \
139                         prefetchw(&prev->_field);                       \
140                 }                                                       \
141         } while (0)
142 #else
143 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
144 #endif
145
146 /*
147  * From 0 .. 100.  Higher means more swappy.
148  */
149 int vm_swappiness = 60;
150 long vm_total_pages;    /* The total number of pages which the VM controls */
151
152 static LIST_HEAD(shrinker_list);
153 static DECLARE_RWSEM(shrinker_rwsem);
154
155 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
156 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
157 #else
158 #define scanning_global_lru(sc) (1)
159 #endif
160
161 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
162                                                   struct scan_control *sc)
163 {
164         if (!scanning_global_lru(sc))
165                 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
166
167         return &zone->reclaim_stat;
168 }
169
170 static unsigned long zone_nr_lru_pages(struct zone *zone,
171                                 struct scan_control *sc, enum lru_list lru)
172 {
173         if (!scanning_global_lru(sc))
174                 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
175
176         return zone_page_state(zone, NR_LRU_BASE + lru);
177 }
178
179
180 /*
181  * Add a shrinker callback to be called from the vm
182  */
183 void register_shrinker(struct shrinker *shrinker)
184 {
185         shrinker->nr = 0;
186         down_write(&shrinker_rwsem);
187         list_add_tail(&shrinker->list, &shrinker_list);
188         up_write(&shrinker_rwsem);
189 }
190 EXPORT_SYMBOL(register_shrinker);
191
192 /*
193  * Remove one
194  */
195 void unregister_shrinker(struct shrinker *shrinker)
196 {
197         down_write(&shrinker_rwsem);
198         list_del(&shrinker->list);
199         up_write(&shrinker_rwsem);
200 }
201 EXPORT_SYMBOL(unregister_shrinker);
202
203 #define SHRINK_BATCH 128
204 /*
205  * Call the shrink functions to age shrinkable caches
206  *
207  * Here we assume it costs one seek to replace a lru page and that it also
208  * takes a seek to recreate a cache object.  With this in mind we age equal
209  * percentages of the lru and ageable caches.  This should balance the seeks
210  * generated by these structures.
211  *
212  * If the vm encountered mapped pages on the LRU it increase the pressure on
213  * slab to avoid swapping.
214  *
215  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
216  *
217  * `lru_pages' represents the number of on-LRU pages in all the zones which
218  * are eligible for the caller's allocation attempt.  It is used for balancing
219  * slab reclaim versus page reclaim.
220  *
221  * Returns the number of slab objects which we shrunk.
222  */
223 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
224                         unsigned long lru_pages)
225 {
226         struct shrinker *shrinker;
227         unsigned long ret = 0;
228
229         if (scanned == 0)
230                 scanned = SWAP_CLUSTER_MAX;
231
232         if (!down_read_trylock(&shrinker_rwsem))
233                 return 1;       /* Assume we'll be able to shrink next time */
234
235         list_for_each_entry(shrinker, &shrinker_list, list) {
236                 unsigned long long delta;
237                 unsigned long total_scan;
238                 unsigned long max_pass;
239
240                 max_pass = (*shrinker->shrink)(shrinker, 0, gfp_mask);
241                 delta = (4 * scanned) / shrinker->seeks;
242                 delta *= max_pass;
243                 do_div(delta, lru_pages + 1);
244                 shrinker->nr += delta;
245                 if (shrinker->nr < 0) {
246                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
247                                "delete nr=%ld\n",
248                                shrinker->shrink, shrinker->nr);
249                         shrinker->nr = max_pass;
250                 }
251
252                 /*
253                  * Avoid risking looping forever due to too large nr value:
254                  * never try to free more than twice the estimate number of
255                  * freeable entries.
256                  */
257                 if (shrinker->nr > max_pass * 2)
258                         shrinker->nr = max_pass * 2;
259
260                 total_scan = shrinker->nr;
261                 shrinker->nr = 0;
262
263                 while (total_scan >= SHRINK_BATCH) {
264                         long this_scan = SHRINK_BATCH;
265                         int shrink_ret;
266                         int nr_before;
267
268                         nr_before = (*shrinker->shrink)(shrinker, 0, gfp_mask);
269                         shrink_ret = (*shrinker->shrink)(shrinker, this_scan,
270                                                                 gfp_mask);
271                         if (shrink_ret == -1)
272                                 break;
273                         if (shrink_ret < nr_before)
274                                 ret += nr_before - shrink_ret;
275                         count_vm_events(SLABS_SCANNED, this_scan);
276                         total_scan -= this_scan;
277
278                         cond_resched();
279                 }
280
281                 shrinker->nr += total_scan;
282         }
283         up_read(&shrinker_rwsem);
284         return ret;
285 }
286
287 static void set_reclaim_mode(int priority, struct scan_control *sc,
288                                    bool sync)
289 {
290         reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
291
292         /*
293          * Initially assume we are entering either lumpy reclaim or
294          * reclaim/compaction.Depending on the order, we will either set the
295          * sync mode or just reclaim order-0 pages later.
296          */
297         if (COMPACTION_BUILD)
298                 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
299         else
300                 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
301
302         /*
303          * Avoid using lumpy reclaim or reclaim/compaction if possible by
304          * restricting when its set to either costly allocations or when
305          * under memory pressure
306          */
307         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
308                 sc->reclaim_mode |= syncmode;
309         else if (sc->order && priority < DEF_PRIORITY - 2)
310                 sc->reclaim_mode |= syncmode;
311         else
312                 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
313 }
314
315 static void reset_reclaim_mode(struct scan_control *sc)
316 {
317         sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
318 }
319
320 static inline int is_page_cache_freeable(struct page *page)
321 {
322         /*
323          * A freeable page cache page is referenced only by the caller
324          * that isolated the page, the page cache radix tree and
325          * optional buffer heads at page->private.
326          */
327         return page_count(page) - page_has_private(page) == 2;
328 }
329
330 static int may_write_to_queue(struct backing_dev_info *bdi,
331                               struct scan_control *sc)
332 {
333         if (current->flags & PF_SWAPWRITE)
334                 return 1;
335         if (!bdi_write_congested(bdi))
336                 return 1;
337         if (bdi == current->backing_dev_info)
338                 return 1;
339
340         /* lumpy reclaim for hugepage often need a lot of write */
341         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
342                 return 1;
343         return 0;
344 }
345
346 /*
347  * We detected a synchronous write error writing a page out.  Probably
348  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
349  * fsync(), msync() or close().
350  *
351  * The tricky part is that after writepage we cannot touch the mapping: nothing
352  * prevents it from being freed up.  But we have a ref on the page and once
353  * that page is locked, the mapping is pinned.
354  *
355  * We're allowed to run sleeping lock_page() here because we know the caller has
356  * __GFP_FS.
357  */
358 static void handle_write_error(struct address_space *mapping,
359                                 struct page *page, int error)
360 {
361         lock_page_nosync(page);
362         if (page_mapping(page) == mapping)
363                 mapping_set_error(mapping, error);
364         unlock_page(page);
365 }
366
367 /* possible outcome of pageout() */
368 typedef enum {
369         /* failed to write page out, page is locked */
370         PAGE_KEEP,
371         /* move page to the active list, page is locked */
372         PAGE_ACTIVATE,
373         /* page has been sent to the disk successfully, page is unlocked */
374         PAGE_SUCCESS,
375         /* page is clean and locked */
376         PAGE_CLEAN,
377 } pageout_t;
378
379 /*
380  * pageout is called by shrink_page_list() for each dirty page.
381  * Calls ->writepage().
382  */
383 static pageout_t pageout(struct page *page, struct address_space *mapping,
384                          struct scan_control *sc)
385 {
386         /*
387          * If the page is dirty, only perform writeback if that write
388          * will be non-blocking.  To prevent this allocation from being
389          * stalled by pagecache activity.  But note that there may be
390          * stalls if we need to run get_block().  We could test
391          * PagePrivate for that.
392          *
393          * If this process is currently in __generic_file_aio_write() against
394          * this page's queue, we can perform writeback even if that
395          * will block.
396          *
397          * If the page is swapcache, write it back even if that would
398          * block, for some throttling. This happens by accident, because
399          * swap_backing_dev_info is bust: it doesn't reflect the
400          * congestion state of the swapdevs.  Easy to fix, if needed.
401          */
402         if (!is_page_cache_freeable(page))
403                 return PAGE_KEEP;
404         if (!mapping) {
405                 /*
406                  * Some data journaling orphaned pages can have
407                  * page->mapping == NULL while being dirty with clean buffers.
408                  */
409                 if (page_has_private(page)) {
410                         if (try_to_free_buffers(page)) {
411                                 ClearPageDirty(page);
412                                 printk("%s: orphaned page\n", __func__);
413                                 return PAGE_CLEAN;
414                         }
415                 }
416                 return PAGE_KEEP;
417         }
418         if (mapping->a_ops->writepage == NULL)
419                 return PAGE_ACTIVATE;
420         if (!may_write_to_queue(mapping->backing_dev_info, sc))
421                 return PAGE_KEEP;
422
423         if (clear_page_dirty_for_io(page)) {
424                 int res;
425                 struct writeback_control wbc = {
426                         .sync_mode = WB_SYNC_NONE,
427                         .nr_to_write = SWAP_CLUSTER_MAX,
428                         .range_start = 0,
429                         .range_end = LLONG_MAX,
430                         .for_reclaim = 1,
431                 };
432
433                 SetPageReclaim(page);
434                 res = mapping->a_ops->writepage(page, &wbc);
435                 if (res < 0)
436                         handle_write_error(mapping, page, res);
437                 if (res == AOP_WRITEPAGE_ACTIVATE) {
438                         ClearPageReclaim(page);
439                         return PAGE_ACTIVATE;
440                 }
441
442                 /*
443                  * Wait on writeback if requested to. This happens when
444                  * direct reclaiming a large contiguous area and the
445                  * first attempt to free a range of pages fails.
446                  */
447                 if (PageWriteback(page) &&
448                     (sc->reclaim_mode & RECLAIM_MODE_SYNC))
449                         wait_on_page_writeback(page);
450
451                 if (!PageWriteback(page)) {
452                         /* synchronous write or broken a_ops? */
453                         ClearPageReclaim(page);
454                 }
455                 trace_mm_vmscan_writepage(page,
456                         trace_reclaim_flags(page, sc->reclaim_mode));
457                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
458                 return PAGE_SUCCESS;
459         }
460
461         return PAGE_CLEAN;
462 }
463
464 /*
465  * Same as remove_mapping, but if the page is removed from the mapping, it
466  * gets returned with a refcount of 0.
467  */
468 static int __remove_mapping(struct address_space *mapping, struct page *page)
469 {
470         BUG_ON(!PageLocked(page));
471         BUG_ON(mapping != page_mapping(page));
472
473         spin_lock_irq(&mapping->tree_lock);
474         /*
475          * The non racy check for a busy page.
476          *
477          * Must be careful with the order of the tests. When someone has
478          * a ref to the page, it may be possible that they dirty it then
479          * drop the reference. So if PageDirty is tested before page_count
480          * here, then the following race may occur:
481          *
482          * get_user_pages(&page);
483          * [user mapping goes away]
484          * write_to(page);
485          *                              !PageDirty(page)    [good]
486          * SetPageDirty(page);
487          * put_page(page);
488          *                              !page_count(page)   [good, discard it]
489          *
490          * [oops, our write_to data is lost]
491          *
492          * Reversing the order of the tests ensures such a situation cannot
493          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
494          * load is not satisfied before that of page->_count.
495          *
496          * Note that if SetPageDirty is always performed via set_page_dirty,
497          * and thus under tree_lock, then this ordering is not required.
498          */
499         if (!page_freeze_refs(page, 2))
500                 goto cannot_free;
501         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
502         if (unlikely(PageDirty(page))) {
503                 page_unfreeze_refs(page, 2);
504                 goto cannot_free;
505         }
506
507         if (PageSwapCache(page)) {
508                 swp_entry_t swap = { .val = page_private(page) };
509                 __delete_from_swap_cache(page);
510                 spin_unlock_irq(&mapping->tree_lock);
511                 swapcache_free(swap, page);
512         } else {
513                 void (*freepage)(struct page *);
514
515                 freepage = mapping->a_ops->freepage;
516
517                 __remove_from_page_cache(page);
518                 spin_unlock_irq(&mapping->tree_lock);
519                 mem_cgroup_uncharge_cache_page(page);
520
521                 if (freepage != NULL)
522                         freepage(page);
523         }
524
525         return 1;
526
527 cannot_free:
528         spin_unlock_irq(&mapping->tree_lock);
529         return 0;
530 }
531
532 /*
533  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
534  * someone else has a ref on the page, abort and return 0.  If it was
535  * successfully detached, return 1.  Assumes the caller has a single ref on
536  * this page.
537  */
538 int remove_mapping(struct address_space *mapping, struct page *page)
539 {
540         if (__remove_mapping(mapping, page)) {
541                 /*
542                  * Unfreezing the refcount with 1 rather than 2 effectively
543                  * drops the pagecache ref for us without requiring another
544                  * atomic operation.
545                  */
546                 page_unfreeze_refs(page, 1);
547                 return 1;
548         }
549         return 0;
550 }
551
552 /**
553  * putback_lru_page - put previously isolated page onto appropriate LRU list
554  * @page: page to be put back to appropriate lru list
555  *
556  * Add previously isolated @page to appropriate LRU list.
557  * Page may still be unevictable for other reasons.
558  *
559  * lru_lock must not be held, interrupts must be enabled.
560  */
561 void putback_lru_page(struct page *page)
562 {
563         int lru;
564         int active = !!TestClearPageActive(page);
565         int was_unevictable = PageUnevictable(page);
566
567         VM_BUG_ON(PageLRU(page));
568
569 redo:
570         ClearPageUnevictable(page);
571
572         if (page_evictable(page, NULL)) {
573                 /*
574                  * For evictable pages, we can use the cache.
575                  * In event of a race, worst case is we end up with an
576                  * unevictable page on [in]active list.
577                  * We know how to handle that.
578                  */
579                 lru = active + page_lru_base_type(page);
580                 lru_cache_add_lru(page, lru);
581         } else {
582                 /*
583                  * Put unevictable pages directly on zone's unevictable
584                  * list.
585                  */
586                 lru = LRU_UNEVICTABLE;
587                 add_page_to_unevictable_list(page);
588                 /*
589                  * When racing with an mlock clearing (page is
590                  * unlocked), make sure that if the other thread does
591                  * not observe our setting of PG_lru and fails
592                  * isolation, we see PG_mlocked cleared below and move
593                  * the page back to the evictable list.
594                  *
595                  * The other side is TestClearPageMlocked().
596                  */
597                 smp_mb();
598         }
599
600         /*
601          * page's status can change while we move it among lru. If an evictable
602          * page is on unevictable list, it never be freed. To avoid that,
603          * check after we added it to the list, again.
604          */
605         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
606                 if (!isolate_lru_page(page)) {
607                         put_page(page);
608                         goto redo;
609                 }
610                 /* This means someone else dropped this page from LRU
611                  * So, it will be freed or putback to LRU again. There is
612                  * nothing to do here.
613                  */
614         }
615
616         if (was_unevictable && lru != LRU_UNEVICTABLE)
617                 count_vm_event(UNEVICTABLE_PGRESCUED);
618         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
619                 count_vm_event(UNEVICTABLE_PGCULLED);
620
621         put_page(page);         /* drop ref from isolate */
622 }
623
624 enum page_references {
625         PAGEREF_RECLAIM,
626         PAGEREF_RECLAIM_CLEAN,
627         PAGEREF_KEEP,
628         PAGEREF_ACTIVATE,
629 };
630
631 static enum page_references page_check_references(struct page *page,
632                                                   struct scan_control *sc)
633 {
634         int referenced_ptes, referenced_page;
635         unsigned long vm_flags;
636
637         referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
638         referenced_page = TestClearPageReferenced(page);
639
640         /* Lumpy reclaim - ignore references */
641         if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
642                 return PAGEREF_RECLAIM;
643
644         /*
645          * Mlock lost the isolation race with us.  Let try_to_unmap()
646          * move the page to the unevictable list.
647          */
648         if (vm_flags & VM_LOCKED)
649                 return PAGEREF_RECLAIM;
650
651         if (referenced_ptes) {
652                 if (PageAnon(page))
653                         return PAGEREF_ACTIVATE;
654                 /*
655                  * All mapped pages start out with page table
656                  * references from the instantiating fault, so we need
657                  * to look twice if a mapped file page is used more
658                  * than once.
659                  *
660                  * Mark it and spare it for another trip around the
661                  * inactive list.  Another page table reference will
662                  * lead to its activation.
663                  *
664                  * Note: the mark is set for activated pages as well
665                  * so that recently deactivated but used pages are
666                  * quickly recovered.
667                  */
668                 SetPageReferenced(page);
669
670                 if (referenced_page)
671                         return PAGEREF_ACTIVATE;
672
673                 return PAGEREF_KEEP;
674         }
675
676         /* Reclaim if clean, defer dirty pages to writeback */
677         if (referenced_page && !PageSwapBacked(page))
678                 return PAGEREF_RECLAIM_CLEAN;
679
680         return PAGEREF_RECLAIM;
681 }
682
683 static noinline_for_stack void free_page_list(struct list_head *free_pages)
684 {
685         struct pagevec freed_pvec;
686         struct page *page, *tmp;
687
688         pagevec_init(&freed_pvec, 1);
689
690         list_for_each_entry_safe(page, tmp, free_pages, lru) {
691                 list_del(&page->lru);
692                 if (!pagevec_add(&freed_pvec, page)) {
693                         __pagevec_free(&freed_pvec);
694                         pagevec_reinit(&freed_pvec);
695                 }
696         }
697
698         pagevec_free(&freed_pvec);
699 }
700
701 /*
702  * shrink_page_list() returns the number of reclaimed pages
703  */
704 static unsigned long shrink_page_list(struct list_head *page_list,
705                                       struct zone *zone,
706                                       struct scan_control *sc)
707 {
708         LIST_HEAD(ret_pages);
709         LIST_HEAD(free_pages);
710         int pgactivate = 0;
711         unsigned long nr_dirty = 0;
712         unsigned long nr_congested = 0;
713         unsigned long nr_reclaimed = 0;
714
715         cond_resched();
716
717         while (!list_empty(page_list)) {
718                 enum page_references references;
719                 struct address_space *mapping;
720                 struct page *page;
721                 int may_enter_fs;
722
723                 cond_resched();
724
725                 page = lru_to_page(page_list);
726                 list_del(&page->lru);
727
728                 if (!trylock_page(page))
729                         goto keep;
730
731                 VM_BUG_ON(PageActive(page));
732                 VM_BUG_ON(page_zone(page) != zone);
733
734                 sc->nr_scanned++;
735
736                 if (unlikely(!page_evictable(page, NULL)))
737                         goto cull_mlocked;
738
739                 if (!sc->may_unmap && page_mapped(page))
740                         goto keep_locked;
741
742                 /* Double the slab pressure for mapped and swapcache pages */
743                 if (page_mapped(page) || PageSwapCache(page))
744                         sc->nr_scanned++;
745
746                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
747                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
748
749                 if (PageWriteback(page)) {
750                         /*
751                          * Synchronous reclaim is performed in two passes,
752                          * first an asynchronous pass over the list to
753                          * start parallel writeback, and a second synchronous
754                          * pass to wait for the IO to complete.  Wait here
755                          * for any page for which writeback has already
756                          * started.
757                          */
758                         if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
759                             may_enter_fs)
760                                 wait_on_page_writeback(page);
761                         else {
762                                 unlock_page(page);
763                                 goto keep_lumpy;
764                         }
765                 }
766
767                 references = page_check_references(page, sc);
768                 switch (references) {
769                 case PAGEREF_ACTIVATE:
770                         goto activate_locked;
771                 case PAGEREF_KEEP:
772                         goto keep_locked;
773                 case PAGEREF_RECLAIM:
774                 case PAGEREF_RECLAIM_CLEAN:
775                         ; /* try to reclaim the page below */
776                 }
777
778                 /*
779                  * Anonymous process memory has backing store?
780                  * Try to allocate it some swap space here.
781                  */
782                 if (PageAnon(page) && !PageSwapCache(page)) {
783                         if (!(sc->gfp_mask & __GFP_IO))
784                                 goto keep_locked;
785                         if (!add_to_swap(page))
786                                 goto activate_locked;
787                         may_enter_fs = 1;
788                 }
789
790                 mapping = page_mapping(page);
791
792                 /*
793                  * The page is mapped into the page tables of one or more
794                  * processes. Try to unmap it here.
795                  */
796                 if (page_mapped(page) && mapping) {
797                         switch (try_to_unmap(page, TTU_UNMAP)) {
798                         case SWAP_FAIL:
799                                 goto activate_locked;
800                         case SWAP_AGAIN:
801                                 goto keep_locked;
802                         case SWAP_MLOCK:
803                                 goto cull_mlocked;
804                         case SWAP_SUCCESS:
805                                 ; /* try to free the page below */
806                         }
807                 }
808
809                 if (PageDirty(page)) {
810                         nr_dirty++;
811
812                         if (references == PAGEREF_RECLAIM_CLEAN)
813                                 goto keep_locked;
814                         if (!may_enter_fs)
815                                 goto keep_locked;
816                         if (!sc->may_writepage)
817                                 goto keep_locked;
818
819                         /* Page is dirty, try to write it out here */
820                         switch (pageout(page, mapping, sc)) {
821                         case PAGE_KEEP:
822                                 nr_congested++;
823                                 goto keep_locked;
824                         case PAGE_ACTIVATE:
825                                 goto activate_locked;
826                         case PAGE_SUCCESS:
827                                 if (PageWriteback(page))
828                                         goto keep_lumpy;
829                                 if (PageDirty(page))
830                                         goto keep;
831
832                                 /*
833                                  * A synchronous write - probably a ramdisk.  Go
834                                  * ahead and try to reclaim the page.
835                                  */
836                                 if (!trylock_page(page))
837                                         goto keep;
838                                 if (PageDirty(page) || PageWriteback(page))
839                                         goto keep_locked;
840                                 mapping = page_mapping(page);
841                         case PAGE_CLEAN:
842                                 ; /* try to free the page below */
843                         }
844                 }
845
846                 /*
847                  * If the page has buffers, try to free the buffer mappings
848                  * associated with this page. If we succeed we try to free
849                  * the page as well.
850                  *
851                  * We do this even if the page is PageDirty().
852                  * try_to_release_page() does not perform I/O, but it is
853                  * possible for a page to have PageDirty set, but it is actually
854                  * clean (all its buffers are clean).  This happens if the
855                  * buffers were written out directly, with submit_bh(). ext3
856                  * will do this, as well as the blockdev mapping.
857                  * try_to_release_page() will discover that cleanness and will
858                  * drop the buffers and mark the page clean - it can be freed.
859                  *
860                  * Rarely, pages can have buffers and no ->mapping.  These are
861                  * the pages which were not successfully invalidated in
862                  * truncate_complete_page().  We try to drop those buffers here
863                  * and if that worked, and the page is no longer mapped into
864                  * process address space (page_count == 1) it can be freed.
865                  * Otherwise, leave the page on the LRU so it is swappable.
866                  */
867                 if (page_has_private(page)) {
868                         if (!try_to_release_page(page, sc->gfp_mask))
869                                 goto activate_locked;
870                         if (!mapping && page_count(page) == 1) {
871                                 unlock_page(page);
872                                 if (put_page_testzero(page))
873                                         goto free_it;
874                                 else {
875                                         /*
876                                          * rare race with speculative reference.
877                                          * the speculative reference will free
878                                          * this page shortly, so we may
879                                          * increment nr_reclaimed here (and
880                                          * leave it off the LRU).
881                                          */
882                                         nr_reclaimed++;
883                                         continue;
884                                 }
885                         }
886                 }
887
888                 if (!mapping || !__remove_mapping(mapping, page))
889                         goto keep_locked;
890
891                 /*
892                  * At this point, we have no other references and there is
893                  * no way to pick any more up (removed from LRU, removed
894                  * from pagecache). Can use non-atomic bitops now (and
895                  * we obviously don't have to worry about waking up a process
896                  * waiting on the page lock, because there are no references.
897                  */
898                 __clear_page_locked(page);
899 free_it:
900                 nr_reclaimed++;
901
902                 /*
903                  * Is there need to periodically free_page_list? It would
904                  * appear not as the counts should be low
905                  */
906                 list_add(&page->lru, &free_pages);
907                 continue;
908
909 cull_mlocked:
910                 if (PageSwapCache(page))
911                         try_to_free_swap(page);
912                 unlock_page(page);
913                 putback_lru_page(page);
914                 reset_reclaim_mode(sc);
915                 continue;
916
917 activate_locked:
918                 /* Not a candidate for swapping, so reclaim swap space. */
919                 if (PageSwapCache(page) && vm_swap_full())
920                         try_to_free_swap(page);
921                 VM_BUG_ON(PageActive(page));
922                 SetPageActive(page);
923                 pgactivate++;
924 keep_locked:
925                 unlock_page(page);
926 keep:
927                 reset_reclaim_mode(sc);
928 keep_lumpy:
929                 list_add(&page->lru, &ret_pages);
930                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
931         }
932
933         /*
934          * Tag a zone as congested if all the dirty pages encountered were
935          * backed by a congested BDI. In this case, reclaimers should just
936          * back off and wait for congestion to clear because further reclaim
937          * will encounter the same problem
938          */
939         if (nr_dirty == nr_congested && nr_dirty != 0)
940                 zone_set_flag(zone, ZONE_CONGESTED);
941
942         free_page_list(&free_pages);
943
944         list_splice(&ret_pages, page_list);
945         count_vm_events(PGACTIVATE, pgactivate);
946         return nr_reclaimed;
947 }
948
949 /*
950  * Attempt to remove the specified page from its LRU.  Only take this page
951  * if it is of the appropriate PageActive status.  Pages which are being
952  * freed elsewhere are also ignored.
953  *
954  * page:        page to consider
955  * mode:        one of the LRU isolation modes defined above
956  *
957  * returns 0 on success, -ve errno on failure.
958  */
959 int __isolate_lru_page(struct page *page, int mode, int file)
960 {
961         int ret = -EINVAL;
962
963         /* Only take pages on the LRU. */
964         if (!PageLRU(page))
965                 return ret;
966
967         /*
968          * When checking the active state, we need to be sure we are
969          * dealing with comparible boolean values.  Take the logical not
970          * of each.
971          */
972         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
973                 return ret;
974
975         if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
976                 return ret;
977
978         /*
979          * When this function is being called for lumpy reclaim, we
980          * initially look into all LRU pages, active, inactive and
981          * unevictable; only give shrink_page_list evictable pages.
982          */
983         if (PageUnevictable(page))
984                 return ret;
985
986         ret = -EBUSY;
987
988         if (likely(get_page_unless_zero(page))) {
989                 /*
990                  * Be careful not to clear PageLRU until after we're
991                  * sure the page is not being freed elsewhere -- the
992                  * page release code relies on it.
993                  */
994                 ClearPageLRU(page);
995                 ret = 0;
996         }
997
998         return ret;
999 }
1000
1001 /*
1002  * zone->lru_lock is heavily contended.  Some of the functions that
1003  * shrink the lists perform better by taking out a batch of pages
1004  * and working on them outside the LRU lock.
1005  *
1006  * For pagecache intensive workloads, this function is the hottest
1007  * spot in the kernel (apart from copy_*_user functions).
1008  *
1009  * Appropriate locks must be held before calling this function.
1010  *
1011  * @nr_to_scan: The number of pages to look through on the list.
1012  * @src:        The LRU list to pull pages off.
1013  * @dst:        The temp list to put pages on to.
1014  * @scanned:    The number of pages that were scanned.
1015  * @order:      The caller's attempted allocation order
1016  * @mode:       One of the LRU isolation modes
1017  * @file:       True [1] if isolating file [!anon] pages
1018  *
1019  * returns how many pages were moved onto *@dst.
1020  */
1021 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1022                 struct list_head *src, struct list_head *dst,
1023                 unsigned long *scanned, int order, int mode, int file)
1024 {
1025         unsigned long nr_taken = 0;
1026         unsigned long nr_lumpy_taken = 0;
1027         unsigned long nr_lumpy_dirty = 0;
1028         unsigned long nr_lumpy_failed = 0;
1029         unsigned long scan;
1030
1031         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1032                 struct page *page;
1033                 unsigned long pfn;
1034                 unsigned long end_pfn;
1035                 unsigned long page_pfn;
1036                 int zone_id;
1037
1038                 page = lru_to_page(src);
1039                 prefetchw_prev_lru_page(page, src, flags);
1040
1041                 VM_BUG_ON(!PageLRU(page));
1042
1043                 switch (__isolate_lru_page(page, mode, file)) {
1044                 case 0:
1045                         list_move(&page->lru, dst);
1046                         mem_cgroup_del_lru(page);
1047                         nr_taken++;
1048                         break;
1049
1050                 case -EBUSY:
1051                         /* else it is being freed elsewhere */
1052                         list_move(&page->lru, src);
1053                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1054                         continue;
1055
1056                 default:
1057                         BUG();
1058                 }
1059
1060                 if (!order)
1061                         continue;
1062
1063                 /*
1064                  * Attempt to take all pages in the order aligned region
1065                  * surrounding the tag page.  Only take those pages of
1066                  * the same active state as that tag page.  We may safely
1067                  * round the target page pfn down to the requested order
1068                  * as the mem_map is guarenteed valid out to MAX_ORDER,
1069                  * where that page is in a different zone we will detect
1070                  * it from its zone id and abort this block scan.
1071                  */
1072                 zone_id = page_zone_id(page);
1073                 page_pfn = page_to_pfn(page);
1074                 pfn = page_pfn & ~((1 << order) - 1);
1075                 end_pfn = pfn + (1 << order);
1076                 for (; pfn < end_pfn; pfn++) {
1077                         struct page *cursor_page;
1078
1079                         /* The target page is in the block, ignore it. */
1080                         if (unlikely(pfn == page_pfn))
1081                                 continue;
1082
1083                         /* Avoid holes within the zone. */
1084                         if (unlikely(!pfn_valid_within(pfn)))
1085                                 break;
1086
1087                         cursor_page = pfn_to_page(pfn);
1088
1089                         /* Check that we have not crossed a zone boundary. */
1090                         if (unlikely(page_zone_id(cursor_page) != zone_id))
1091                                 break;
1092
1093                         /*
1094                          * If we don't have enough swap space, reclaiming of
1095                          * anon page which don't already have a swap slot is
1096                          * pointless.
1097                          */
1098                         if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1099                             !PageSwapCache(cursor_page))
1100                                 break;
1101
1102                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1103                                 list_move(&cursor_page->lru, dst);
1104                                 mem_cgroup_del_lru(cursor_page);
1105                                 nr_taken++;
1106                                 nr_lumpy_taken++;
1107                                 if (PageDirty(cursor_page))
1108                                         nr_lumpy_dirty++;
1109                                 scan++;
1110                         } else {
1111                                 /* the page is freed already. */
1112                                 if (!page_count(cursor_page))
1113                                         continue;
1114                                 break;
1115                         }
1116                 }
1117
1118                 /* If we break out of the loop above, lumpy reclaim failed */
1119                 if (pfn < end_pfn)
1120                         nr_lumpy_failed++;
1121         }
1122
1123         *scanned = scan;
1124
1125         trace_mm_vmscan_lru_isolate(order,
1126                         nr_to_scan, scan,
1127                         nr_taken,
1128                         nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1129                         mode);
1130         return nr_taken;
1131 }
1132
1133 static unsigned long isolate_pages_global(unsigned long nr,
1134                                         struct list_head *dst,
1135                                         unsigned long *scanned, int order,
1136                                         int mode, struct zone *z,
1137                                         int active, int file)
1138 {
1139         int lru = LRU_BASE;
1140         if (active)
1141                 lru += LRU_ACTIVE;
1142         if (file)
1143                 lru += LRU_FILE;
1144         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1145                                                                 mode, file);
1146 }
1147
1148 /*
1149  * clear_active_flags() is a helper for shrink_active_list(), clearing
1150  * any active bits from the pages in the list.
1151  */
1152 static unsigned long clear_active_flags(struct list_head *page_list,
1153                                         unsigned int *count)
1154 {
1155         int nr_active = 0;
1156         int lru;
1157         struct page *page;
1158
1159         list_for_each_entry(page, page_list, lru) {
1160                 lru = page_lru_base_type(page);
1161                 if (PageActive(page)) {
1162                         lru += LRU_ACTIVE;
1163                         ClearPageActive(page);
1164                         nr_active++;
1165                 }
1166                 if (count)
1167                         count[lru]++;
1168         }
1169
1170         return nr_active;
1171 }
1172
1173 /**
1174  * isolate_lru_page - tries to isolate a page from its LRU list
1175  * @page: page to isolate from its LRU list
1176  *
1177  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1178  * vmstat statistic corresponding to whatever LRU list the page was on.
1179  *
1180  * Returns 0 if the page was removed from an LRU list.
1181  * Returns -EBUSY if the page was not on an LRU list.
1182  *
1183  * The returned page will have PageLRU() cleared.  If it was found on
1184  * the active list, it will have PageActive set.  If it was found on
1185  * the unevictable list, it will have the PageUnevictable bit set. That flag
1186  * may need to be cleared by the caller before letting the page go.
1187  *
1188  * The vmstat statistic corresponding to the list on which the page was
1189  * found will be decremented.
1190  *
1191  * Restrictions:
1192  * (1) Must be called with an elevated refcount on the page. This is a
1193  *     fundamentnal difference from isolate_lru_pages (which is called
1194  *     without a stable reference).
1195  * (2) the lru_lock must not be held.
1196  * (3) interrupts must be enabled.
1197  */
1198 int isolate_lru_page(struct page *page)
1199 {
1200         int ret = -EBUSY;
1201
1202         if (PageLRU(page)) {
1203                 struct zone *zone = page_zone(page);
1204
1205                 spin_lock_irq(&zone->lru_lock);
1206                 if (PageLRU(page) && get_page_unless_zero(page)) {
1207                         int lru = page_lru(page);
1208                         ret = 0;
1209                         ClearPageLRU(page);
1210
1211                         del_page_from_lru_list(zone, page, lru);
1212                 }
1213                 spin_unlock_irq(&zone->lru_lock);
1214         }
1215         return ret;
1216 }
1217
1218 /*
1219  * Are there way too many processes in the direct reclaim path already?
1220  */
1221 static int too_many_isolated(struct zone *zone, int file,
1222                 struct scan_control *sc)
1223 {
1224         unsigned long inactive, isolated;
1225
1226         if (current_is_kswapd())
1227                 return 0;
1228
1229         if (!scanning_global_lru(sc))
1230                 return 0;
1231
1232         if (file) {
1233                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1234                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1235         } else {
1236                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1237                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1238         }
1239
1240         return isolated > inactive;
1241 }
1242
1243 /*
1244  * TODO: Try merging with migrations version of putback_lru_pages
1245  */
1246 static noinline_for_stack void
1247 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1248                                 unsigned long nr_anon, unsigned long nr_file,
1249                                 struct list_head *page_list)
1250 {
1251         struct page *page;
1252         struct pagevec pvec;
1253         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1254
1255         pagevec_init(&pvec, 1);
1256
1257         /*
1258          * Put back any unfreeable pages.
1259          */
1260         spin_lock(&zone->lru_lock);
1261         while (!list_empty(page_list)) {
1262                 int lru;
1263                 page = lru_to_page(page_list);
1264                 VM_BUG_ON(PageLRU(page));
1265                 list_del(&page->lru);
1266                 if (unlikely(!page_evictable(page, NULL))) {
1267                         spin_unlock_irq(&zone->lru_lock);
1268                         putback_lru_page(page);
1269                         spin_lock_irq(&zone->lru_lock);
1270                         continue;
1271                 }
1272                 SetPageLRU(page);
1273                 lru = page_lru(page);
1274                 add_page_to_lru_list(zone, page, lru);
1275                 if (is_active_lru(lru)) {
1276                         int file = is_file_lru(lru);
1277                         reclaim_stat->recent_rotated[file]++;
1278                 }
1279                 if (!pagevec_add(&pvec, page)) {
1280                         spin_unlock_irq(&zone->lru_lock);
1281                         __pagevec_release(&pvec);
1282                         spin_lock_irq(&zone->lru_lock);
1283                 }
1284         }
1285         __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1286         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1287
1288         spin_unlock_irq(&zone->lru_lock);
1289         pagevec_release(&pvec);
1290 }
1291
1292 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1293                                         struct scan_control *sc,
1294                                         unsigned long *nr_anon,
1295                                         unsigned long *nr_file,
1296                                         struct list_head *isolated_list)
1297 {
1298         unsigned long nr_active;
1299         unsigned int count[NR_LRU_LISTS] = { 0, };
1300         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1301
1302         nr_active = clear_active_flags(isolated_list, count);
1303         __count_vm_events(PGDEACTIVATE, nr_active);
1304
1305         __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1306                               -count[LRU_ACTIVE_FILE]);
1307         __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1308                               -count[LRU_INACTIVE_FILE]);
1309         __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1310                               -count[LRU_ACTIVE_ANON]);
1311         __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1312                               -count[LRU_INACTIVE_ANON]);
1313
1314         *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1315         *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1316         __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1317         __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1318
1319         reclaim_stat->recent_scanned[0] += *nr_anon;
1320         reclaim_stat->recent_scanned[1] += *nr_file;
1321 }
1322
1323 /*
1324  * Returns true if the caller should wait to clean dirty/writeback pages.
1325  *
1326  * If we are direct reclaiming for contiguous pages and we do not reclaim
1327  * everything in the list, try again and wait for writeback IO to complete.
1328  * This will stall high-order allocations noticeably. Only do that when really
1329  * need to free the pages under high memory pressure.
1330  */
1331 static inline bool should_reclaim_stall(unsigned long nr_taken,
1332                                         unsigned long nr_freed,
1333                                         int priority,
1334                                         struct scan_control *sc)
1335 {
1336         int lumpy_stall_priority;
1337
1338         /* kswapd should not stall on sync IO */
1339         if (current_is_kswapd())
1340                 return false;
1341
1342         /* Only stall on lumpy reclaim */
1343         if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1344                 return false;
1345
1346         /* If we have relaimed everything on the isolated list, no stall */
1347         if (nr_freed == nr_taken)
1348                 return false;
1349
1350         /*
1351          * For high-order allocations, there are two stall thresholds.
1352          * High-cost allocations stall immediately where as lower
1353          * order allocations such as stacks require the scanning
1354          * priority to be much higher before stalling.
1355          */
1356         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1357                 lumpy_stall_priority = DEF_PRIORITY;
1358         else
1359                 lumpy_stall_priority = DEF_PRIORITY / 3;
1360
1361         return priority <= lumpy_stall_priority;
1362 }
1363
1364 /*
1365  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1366  * of reclaimed pages
1367  */
1368 static noinline_for_stack unsigned long
1369 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1370                         struct scan_control *sc, int priority, int file)
1371 {
1372         LIST_HEAD(page_list);
1373         unsigned long nr_scanned;
1374         unsigned long nr_reclaimed = 0;
1375         unsigned long nr_taken;
1376         unsigned long nr_anon;
1377         unsigned long nr_file;
1378
1379         while (unlikely(too_many_isolated(zone, file, sc))) {
1380                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1381
1382                 /* We are about to die and free our memory. Return now. */
1383                 if (fatal_signal_pending(current))
1384                         return SWAP_CLUSTER_MAX;
1385         }
1386
1387         set_reclaim_mode(priority, sc, false);
1388         lru_add_drain();
1389         spin_lock_irq(&zone->lru_lock);
1390
1391         if (scanning_global_lru(sc)) {
1392                 nr_taken = isolate_pages_global(nr_to_scan,
1393                         &page_list, &nr_scanned, sc->order,
1394                         sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
1395                                         ISOLATE_BOTH : ISOLATE_INACTIVE,
1396                         zone, 0, file);
1397                 zone->pages_scanned += nr_scanned;
1398                 if (current_is_kswapd())
1399                         __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1400                                                nr_scanned);
1401                 else
1402                         __count_zone_vm_events(PGSCAN_DIRECT, zone,
1403                                                nr_scanned);
1404         } else {
1405                 nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1406                         &page_list, &nr_scanned, sc->order,
1407                         sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
1408                                         ISOLATE_BOTH : ISOLATE_INACTIVE,
1409                         zone, sc->mem_cgroup,
1410                         0, file);
1411                 /*
1412                  * mem_cgroup_isolate_pages() keeps track of
1413                  * scanned pages on its own.
1414                  */
1415         }
1416
1417         if (nr_taken == 0) {
1418                 spin_unlock_irq(&zone->lru_lock);
1419                 return 0;
1420         }
1421
1422         update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1423
1424         spin_unlock_irq(&zone->lru_lock);
1425
1426         nr_reclaimed = shrink_page_list(&page_list, zone, sc);
1427
1428         /* Check if we should syncronously wait for writeback */
1429         if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1430                 set_reclaim_mode(priority, sc, true);
1431                 nr_reclaimed += shrink_page_list(&page_list, zone, sc);
1432         }
1433
1434         local_irq_disable();
1435         if (current_is_kswapd())
1436                 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1437         __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1438
1439         putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1440
1441         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1442                 zone_idx(zone),
1443                 nr_scanned, nr_reclaimed,
1444                 priority,
1445                 trace_shrink_flags(file, sc->reclaim_mode));
1446         return nr_reclaimed;
1447 }
1448
1449 /*
1450  * This moves pages from the active list to the inactive list.
1451  *
1452  * We move them the other way if the page is referenced by one or more
1453  * processes, from rmap.
1454  *
1455  * If the pages are mostly unmapped, the processing is fast and it is
1456  * appropriate to hold zone->lru_lock across the whole operation.  But if
1457  * the pages are mapped, the processing is slow (page_referenced()) so we
1458  * should drop zone->lru_lock around each page.  It's impossible to balance
1459  * this, so instead we remove the pages from the LRU while processing them.
1460  * It is safe to rely on PG_active against the non-LRU pages in here because
1461  * nobody will play with that bit on a non-LRU page.
1462  *
1463  * The downside is that we have to touch page->_count against each page.
1464  * But we had to alter page->flags anyway.
1465  */
1466
1467 static void move_active_pages_to_lru(struct zone *zone,
1468                                      struct list_head *list,
1469                                      enum lru_list lru)
1470 {
1471         unsigned long pgmoved = 0;
1472         struct pagevec pvec;
1473         struct page *page;
1474
1475         pagevec_init(&pvec, 1);
1476
1477         while (!list_empty(list)) {
1478                 page = lru_to_page(list);
1479
1480                 VM_BUG_ON(PageLRU(page));
1481                 SetPageLRU(page);
1482
1483                 list_move(&page->lru, &zone->lru[lru].list);
1484                 mem_cgroup_add_lru_list(page, lru);
1485                 pgmoved++;
1486
1487                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1488                         spin_unlock_irq(&zone->lru_lock);
1489                         if (buffer_heads_over_limit)
1490                                 pagevec_strip(&pvec);
1491                         __pagevec_release(&pvec);
1492                         spin_lock_irq(&zone->lru_lock);
1493                 }
1494         }
1495         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1496         if (!is_active_lru(lru))
1497                 __count_vm_events(PGDEACTIVATE, pgmoved);
1498 }
1499
1500 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1501                         struct scan_control *sc, int priority, int file)
1502 {
1503         unsigned long nr_taken;
1504         unsigned long pgscanned;
1505         unsigned long vm_flags;
1506         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1507         LIST_HEAD(l_active);
1508         LIST_HEAD(l_inactive);
1509         struct page *page;
1510         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1511         unsigned long nr_rotated = 0;
1512
1513         lru_add_drain();
1514         spin_lock_irq(&zone->lru_lock);
1515         if (scanning_global_lru(sc)) {
1516                 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1517                                                 &pgscanned, sc->order,
1518                                                 ISOLATE_ACTIVE, zone,
1519                                                 1, file);
1520                 zone->pages_scanned += pgscanned;
1521         } else {
1522                 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1523                                                 &pgscanned, sc->order,
1524                                                 ISOLATE_ACTIVE, zone,
1525                                                 sc->mem_cgroup, 1, file);
1526                 /*
1527                  * mem_cgroup_isolate_pages() keeps track of
1528                  * scanned pages on its own.
1529                  */
1530         }
1531
1532         reclaim_stat->recent_scanned[file] += nr_taken;
1533
1534         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1535         if (file)
1536                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1537         else
1538                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1539         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1540         spin_unlock_irq(&zone->lru_lock);
1541
1542         while (!list_empty(&l_hold)) {
1543                 cond_resched();
1544                 page = lru_to_page(&l_hold);
1545                 list_del(&page->lru);
1546
1547                 if (unlikely(!page_evictable(page, NULL))) {
1548                         putback_lru_page(page);
1549                         continue;
1550                 }
1551
1552                 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1553                         nr_rotated++;
1554                         /*
1555                          * Identify referenced, file-backed active pages and
1556                          * give them one more trip around the active list. So
1557                          * that executable code get better chances to stay in
1558                          * memory under moderate memory pressure.  Anon pages
1559                          * are not likely to be evicted by use-once streaming
1560                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1561                          * so we ignore them here.
1562                          */
1563                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1564                                 list_add(&page->lru, &l_active);
1565                                 continue;
1566                         }
1567                 }
1568
1569                 ClearPageActive(page);  /* we are de-activating */
1570                 list_add(&page->lru, &l_inactive);
1571         }
1572
1573         /*
1574          * Move pages back to the lru list.
1575          */
1576         spin_lock_irq(&zone->lru_lock);
1577         /*
1578          * Count referenced pages from currently used mappings as rotated,
1579          * even though only some of them are actually re-activated.  This
1580          * helps balance scan pressure between file and anonymous pages in
1581          * get_scan_ratio.
1582          */
1583         reclaim_stat->recent_rotated[file] += nr_rotated;
1584
1585         move_active_pages_to_lru(zone, &l_active,
1586                                                 LRU_ACTIVE + file * LRU_FILE);
1587         move_active_pages_to_lru(zone, &l_inactive,
1588                                                 LRU_BASE   + file * LRU_FILE);
1589         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1590         spin_unlock_irq(&zone->lru_lock);
1591 }
1592
1593 #ifdef CONFIG_SWAP
1594 static int inactive_anon_is_low_global(struct zone *zone)
1595 {
1596         unsigned long active, inactive;
1597
1598         active = zone_page_state(zone, NR_ACTIVE_ANON);
1599         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1600
1601         if (inactive * zone->inactive_ratio < active)
1602                 return 1;
1603
1604         return 0;
1605 }
1606
1607 /**
1608  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1609  * @zone: zone to check
1610  * @sc:   scan control of this context
1611  *
1612  * Returns true if the zone does not have enough inactive anon pages,
1613  * meaning some active anon pages need to be deactivated.
1614  */
1615 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1616 {
1617         int low;
1618
1619         /*
1620          * If we don't have swap space, anonymous page deactivation
1621          * is pointless.
1622          */
1623         if (!total_swap_pages)
1624                 return 0;
1625
1626         if (scanning_global_lru(sc))
1627                 low = inactive_anon_is_low_global(zone);
1628         else
1629                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1630         return low;
1631 }
1632 #else
1633 static inline int inactive_anon_is_low(struct zone *zone,
1634                                         struct scan_control *sc)
1635 {
1636         return 0;
1637 }
1638 #endif
1639
1640 static int inactive_file_is_low_global(struct zone *zone)
1641 {
1642         unsigned long active, inactive;
1643
1644         active = zone_page_state(zone, NR_ACTIVE_FILE);
1645         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1646
1647         return (active > inactive);
1648 }
1649
1650 /**
1651  * inactive_file_is_low - check if file pages need to be deactivated
1652  * @zone: zone to check
1653  * @sc:   scan control of this context
1654  *
1655  * When the system is doing streaming IO, memory pressure here
1656  * ensures that active file pages get deactivated, until more
1657  * than half of the file pages are on the inactive list.
1658  *
1659  * Once we get to that situation, protect the system's working
1660  * set from being evicted by disabling active file page aging.
1661  *
1662  * This uses a different ratio than the anonymous pages, because
1663  * the page cache uses a use-once replacement algorithm.
1664  */
1665 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1666 {
1667         int low;
1668
1669         if (scanning_global_lru(sc))
1670                 low = inactive_file_is_low_global(zone);
1671         else
1672                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1673         return low;
1674 }
1675
1676 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1677                                 int file)
1678 {
1679         if (file)
1680                 return inactive_file_is_low(zone, sc);
1681         else
1682                 return inactive_anon_is_low(zone, sc);
1683 }
1684
1685 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1686         struct zone *zone, struct scan_control *sc, int priority)
1687 {
1688         int file = is_file_lru(lru);
1689
1690         if (is_active_lru(lru)) {
1691                 if (inactive_list_is_low(zone, sc, file))
1692                     shrink_active_list(nr_to_scan, zone, sc, priority, file);
1693                 return 0;
1694         }
1695
1696         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1697 }
1698
1699 /*
1700  * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1701  * until we collected @swap_cluster_max pages to scan.
1702  */
1703 static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1704                                        unsigned long *nr_saved_scan)
1705 {
1706         unsigned long nr;
1707
1708         *nr_saved_scan += nr_to_scan;
1709         nr = *nr_saved_scan;
1710
1711         if (nr >= SWAP_CLUSTER_MAX)
1712                 *nr_saved_scan = 0;
1713         else
1714                 nr = 0;
1715
1716         return nr;
1717 }
1718
1719 /*
1720  * Determine how aggressively the anon and file LRU lists should be
1721  * scanned.  The relative value of each set of LRU lists is determined
1722  * by looking at the fraction of the pages scanned we did rotate back
1723  * onto the active list instead of evict.
1724  *
1725  * nr[0] = anon pages to scan; nr[1] = file pages to scan
1726  */
1727 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1728                                         unsigned long *nr, int priority)
1729 {
1730         unsigned long anon, file, free;
1731         unsigned long anon_prio, file_prio;
1732         unsigned long ap, fp;
1733         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1734         u64 fraction[2], denominator;
1735         enum lru_list l;
1736         int noswap = 0;
1737
1738         /* If we have no swap space, do not bother scanning anon pages. */
1739         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1740                 noswap = 1;
1741                 fraction[0] = 0;
1742                 fraction[1] = 1;
1743                 denominator = 1;
1744                 goto out;
1745         }
1746
1747         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1748                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1749         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1750                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1751
1752         if (scanning_global_lru(sc)) {
1753                 free  = zone_page_state(zone, NR_FREE_PAGES);
1754                 /* If we have very few page cache pages,
1755                    force-scan anon pages. */
1756                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1757                         fraction[0] = 1;
1758                         fraction[1] = 0;
1759                         denominator = 1;
1760                         goto out;
1761                 }
1762         }
1763
1764         /*
1765          * With swappiness at 100, anonymous and file have the same priority.
1766          * This scanning priority is essentially the inverse of IO cost.
1767          */
1768         anon_prio = sc->swappiness;
1769         file_prio = 200 - sc->swappiness;
1770
1771         /*
1772          * OK, so we have swap space and a fair amount of page cache
1773          * pages.  We use the recently rotated / recently scanned
1774          * ratios to determine how valuable each cache is.
1775          *
1776          * Because workloads change over time (and to avoid overflow)
1777          * we keep these statistics as a floating average, which ends
1778          * up weighing recent references more than old ones.
1779          *
1780          * anon in [0], file in [1]
1781          */
1782         spin_lock_irq(&zone->lru_lock);
1783         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1784                 reclaim_stat->recent_scanned[0] /= 2;
1785                 reclaim_stat->recent_rotated[0] /= 2;
1786         }
1787
1788         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1789                 reclaim_stat->recent_scanned[1] /= 2;
1790                 reclaim_stat->recent_rotated[1] /= 2;
1791         }
1792
1793         /*
1794          * The amount of pressure on anon vs file pages is inversely
1795          * proportional to the fraction of recently scanned pages on
1796          * each list that were recently referenced and in active use.
1797          */
1798         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1799         ap /= reclaim_stat->recent_rotated[0] + 1;
1800
1801         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1802         fp /= reclaim_stat->recent_rotated[1] + 1;
1803         spin_unlock_irq(&zone->lru_lock);
1804
1805         fraction[0] = ap;
1806         fraction[1] = fp;
1807         denominator = ap + fp + 1;
1808 out:
1809         for_each_evictable_lru(l) {
1810                 int file = is_file_lru(l);
1811                 unsigned long scan;
1812
1813                 scan = zone_nr_lru_pages(zone, sc, l);
1814                 if (priority || noswap) {
1815                         scan >>= priority;
1816                         scan = div64_u64(scan * fraction[file], denominator);
1817                 }
1818                 nr[l] = nr_scan_try_batch(scan,
1819                                           &reclaim_stat->nr_saved_scan[l]);
1820         }
1821 }
1822
1823 /*
1824  * Reclaim/compaction depends on a number of pages being freed. To avoid
1825  * disruption to the system, a small number of order-0 pages continue to be
1826  * rotated and reclaimed in the normal fashion. However, by the time we get
1827  * back to the allocator and call try_to_compact_zone(), we ensure that
1828  * there are enough free pages for it to be likely successful
1829  */
1830 static inline bool should_continue_reclaim(struct zone *zone,
1831                                         unsigned long nr_reclaimed,
1832                                         unsigned long nr_scanned,
1833                                         struct scan_control *sc)
1834 {
1835         unsigned long pages_for_compaction;
1836         unsigned long inactive_lru_pages;
1837
1838         /* If not in reclaim/compaction mode, stop */
1839         if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1840                 return false;
1841
1842         /*
1843          * If we failed to reclaim and have scanned the full list, stop.
1844          * NOTE: Checking just nr_reclaimed would exit reclaim/compaction far
1845          *       faster but obviously would be less likely to succeed
1846          *       allocation. If this is desirable, use GFP_REPEAT to decide
1847          *       if both reclaimed and scanned should be checked or just
1848          *       reclaimed
1849          */
1850         if (!nr_reclaimed && !nr_scanned)
1851                 return false;
1852
1853         /*
1854          * If we have not reclaimed enough pages for compaction and the
1855          * inactive lists are large enough, continue reclaiming
1856          */
1857         pages_for_compaction = (2UL << sc->order);
1858         inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
1859                                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1860         if (sc->nr_reclaimed < pages_for_compaction &&
1861                         inactive_lru_pages > pages_for_compaction)
1862                 return true;
1863
1864         /* If compaction would go ahead or the allocation would succeed, stop */
1865         switch (compaction_suitable(zone, sc->order)) {
1866         case COMPACT_PARTIAL:
1867         case COMPACT_CONTINUE:
1868                 return false;
1869         default:
1870                 return true;
1871         }
1872 }
1873
1874 /*
1875  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1876  */
1877 static void shrink_zone(int priority, struct zone *zone,
1878                                 struct scan_control *sc)
1879 {
1880         unsigned long nr[NR_LRU_LISTS];
1881         unsigned long nr_to_scan;
1882         enum lru_list l;
1883         unsigned long nr_reclaimed;
1884         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1885         unsigned long nr_scanned = sc->nr_scanned;
1886
1887 restart:
1888         nr_reclaimed = 0;
1889         get_scan_count(zone, sc, nr, priority);
1890
1891         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1892                                         nr[LRU_INACTIVE_FILE]) {
1893                 for_each_evictable_lru(l) {
1894                         if (nr[l]) {
1895                                 nr_to_scan = min_t(unsigned long,
1896                                                    nr[l], SWAP_CLUSTER_MAX);
1897                                 nr[l] -= nr_to_scan;
1898
1899                                 nr_reclaimed += shrink_list(l, nr_to_scan,
1900                                                             zone, sc, priority);
1901                         }
1902                 }
1903                 /*
1904                  * On large memory systems, scan >> priority can become
1905                  * really large. This is fine for the starting priority;
1906                  * we want to put equal scanning pressure on each zone.
1907                  * However, if the VM has a harder time of freeing pages,
1908                  * with multiple processes reclaiming pages, the total
1909                  * freeing target can get unreasonably large.
1910                  */
1911                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
1912                         break;
1913         }
1914         sc->nr_reclaimed += nr_reclaimed;
1915
1916         /*
1917          * Even if we did not try to evict anon pages at all, we want to
1918          * rebalance the anon lru active/inactive ratio.
1919          */
1920         if (inactive_anon_is_low(zone, sc))
1921                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1922
1923         /* reclaim/compaction might need reclaim to continue */
1924         if (should_continue_reclaim(zone, nr_reclaimed,
1925                                         sc->nr_scanned - nr_scanned, sc))
1926                 goto restart;
1927
1928         throttle_vm_writeout(sc->gfp_mask);
1929 }
1930
1931 /*
1932  * This is the direct reclaim path, for page-allocating processes.  We only
1933  * try to reclaim pages from zones which will satisfy the caller's allocation
1934  * request.
1935  *
1936  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1937  * Because:
1938  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1939  *    allocation or
1940  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1941  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1942  *    zone defense algorithm.
1943  *
1944  * If a zone is deemed to be full of pinned pages then just give it a light
1945  * scan then give up on it.
1946  */
1947 static void shrink_zones(int priority, struct zonelist *zonelist,
1948                                         struct scan_control *sc)
1949 {
1950         struct zoneref *z;
1951         struct zone *zone;
1952
1953         for_each_zone_zonelist_nodemask(zone, z, zonelist,
1954                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
1955                 if (!populated_zone(zone))
1956                         continue;
1957                 /*
1958                  * Take care memory controller reclaiming has small influence
1959                  * to global LRU.
1960                  */
1961                 if (scanning_global_lru(sc)) {
1962                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1963                                 continue;
1964                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
1965                                 continue;       /* Let kswapd poll it */
1966                 }
1967
1968                 shrink_zone(priority, zone, sc);
1969         }
1970 }
1971
1972 static bool zone_reclaimable(struct zone *zone)
1973 {
1974         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
1975 }
1976
1977 /*
1978  * As hibernation is going on, kswapd is freezed so that it can't mark
1979  * the zone into all_unreclaimable. It can't handle OOM during hibernation.
1980  * So let's check zone's unreclaimable in direct reclaim as well as kswapd.
1981  */
1982 static bool all_unreclaimable(struct zonelist *zonelist,
1983                 struct scan_control *sc)
1984 {
1985         struct zoneref *z;
1986         struct zone *zone;
1987         bool all_unreclaimable = true;
1988
1989         for_each_zone_zonelist_nodemask(zone, z, zonelist,
1990                         gfp_zone(sc->gfp_mask), sc->nodemask) {
1991                 if (!populated_zone(zone))
1992                         continue;
1993                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1994                         continue;
1995                 if (zone_reclaimable(zone)) {
1996                         all_unreclaimable = false;
1997                         break;
1998                 }
1999         }
2000
2001         return all_unreclaimable;
2002 }
2003
2004 /*
2005  * This is the main entry point to direct page reclaim.
2006  *
2007  * If a full scan of the inactive list fails to free enough memory then we
2008  * are "out of memory" and something needs to be killed.
2009  *
2010  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2011  * high - the zone may be full of dirty or under-writeback pages, which this
2012  * caller can't do much about.  We kick the writeback threads and take explicit
2013  * naps in the hope that some of these pages can be written.  But if the
2014  * allocating task holds filesystem locks which prevent writeout this might not
2015  * work, and the allocation attempt will fail.
2016  *
2017  * returns:     0, if no pages reclaimed
2018  *              else, the number of pages reclaimed
2019  */
2020 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2021                                         struct scan_control *sc)
2022 {
2023         int priority;
2024         unsigned long total_scanned = 0;
2025         struct reclaim_state *reclaim_state = current->reclaim_state;
2026         struct zoneref *z;
2027         struct zone *zone;
2028         unsigned long writeback_threshold;
2029
2030         get_mems_allowed();
2031         delayacct_freepages_start();
2032
2033         if (scanning_global_lru(sc))
2034                 count_vm_event(ALLOCSTALL);
2035
2036         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2037                 sc->nr_scanned = 0;
2038                 if (!priority)
2039                         disable_swap_token();
2040                 shrink_zones(priority, zonelist, sc);
2041                 /*
2042                  * Don't shrink slabs when reclaiming memory from
2043                  * over limit cgroups
2044                  */
2045                 if (scanning_global_lru(sc)) {
2046                         unsigned long lru_pages = 0;
2047                         for_each_zone_zonelist(zone, z, zonelist,
2048                                         gfp_zone(sc->gfp_mask)) {
2049                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2050                                         continue;
2051
2052                                 lru_pages += zone_reclaimable_pages(zone);
2053                         }
2054
2055                         shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
2056                         if (reclaim_state) {
2057                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2058                                 reclaim_state->reclaimed_slab = 0;
2059                         }
2060                 }
2061                 total_scanned += sc->nr_scanned;
2062                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2063                         goto out;
2064
2065                 /*
2066                  * Try to write back as many pages as we just scanned.  This
2067                  * tends to cause slow streaming writers to write data to the
2068                  * disk smoothly, at the dirtying rate, which is nice.   But
2069                  * that's undesirable in laptop mode, where we *want* lumpy
2070                  * writeout.  So in laptop mode, write out the whole world.
2071                  */
2072                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2073                 if (total_scanned > writeback_threshold) {
2074                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
2075                         sc->may_writepage = 1;
2076                 }
2077
2078                 /* Take a nap, wait for some writeback to complete */
2079                 if (!sc->hibernation_mode && sc->nr_scanned &&
2080                     priority < DEF_PRIORITY - 2) {
2081                         struct zone *preferred_zone;
2082
2083                         first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2084                                                         NULL, &preferred_zone);
2085                         wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2086                 }
2087         }
2088
2089 out:
2090         delayacct_freepages_end();
2091         put_mems_allowed();
2092
2093         if (sc->nr_reclaimed)
2094                 return sc->nr_reclaimed;
2095
2096         /* top priority shrink_zones still had more to do? don't OOM, then */
2097         if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2098                 return 1;
2099
2100         return 0;
2101 }
2102
2103 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2104                                 gfp_t gfp_mask, nodemask_t *nodemask)
2105 {
2106         unsigned long nr_reclaimed;
2107         struct scan_control sc = {
2108                 .gfp_mask = gfp_mask,
2109                 .may_writepage = !laptop_mode,
2110                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2111                 .may_unmap = 1,
2112                 .may_swap = 1,
2113                 .swappiness = vm_swappiness,
2114                 .order = order,
2115                 .mem_cgroup = NULL,
2116                 .nodemask = nodemask,
2117         };
2118
2119         trace_mm_vmscan_direct_reclaim_begin(order,
2120                                 sc.may_writepage,
2121                                 gfp_mask);
2122
2123         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2124
2125         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2126
2127         return nr_reclaimed;
2128 }
2129
2130 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2131
2132 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2133                                                 gfp_t gfp_mask, bool noswap,
2134                                                 unsigned int swappiness,
2135                                                 struct zone *zone)
2136 {
2137         struct scan_control sc = {
2138                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2139                 .may_writepage = !laptop_mode,
2140                 .may_unmap = 1,
2141                 .may_swap = !noswap,
2142                 .swappiness = swappiness,
2143                 .order = 0,
2144                 .mem_cgroup = mem,
2145         };
2146         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2147                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2148
2149         trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2150                                                       sc.may_writepage,
2151                                                       sc.gfp_mask);
2152
2153         /*
2154          * NOTE: Although we can get the priority field, using it
2155          * here is not a good idea, since it limits the pages we can scan.
2156          * if we don't reclaim here, the shrink_zone from balance_pgdat
2157          * will pick up pages from other mem cgroup's as well. We hack
2158          * the priority and make it zero.
2159          */
2160         shrink_zone(0, zone, &sc);
2161
2162         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2163
2164         return sc.nr_reclaimed;
2165 }
2166
2167 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2168                                            gfp_t gfp_mask,
2169                                            bool noswap,
2170                                            unsigned int swappiness)
2171 {
2172         struct zonelist *zonelist;
2173         unsigned long nr_reclaimed;
2174         struct scan_control sc = {
2175                 .may_writepage = !laptop_mode,
2176                 .may_unmap = 1,
2177                 .may_swap = !noswap,
2178                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2179                 .swappiness = swappiness,
2180                 .order = 0,
2181                 .mem_cgroup = mem_cont,
2182                 .nodemask = NULL, /* we don't care the placement */
2183         };
2184
2185         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2186                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2187         zonelist = NODE_DATA(numa_node_id())->node_zonelists;
2188
2189         trace_mm_vmscan_memcg_reclaim_begin(0,
2190                                             sc.may_writepage,
2191                                             sc.gfp_mask);
2192
2193         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2194
2195         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2196
2197         return nr_reclaimed;
2198 }
2199 #endif
2200
2201 /* is kswapd sleeping prematurely? */
2202 static int sleeping_prematurely(pg_data_t *pgdat, int order, long remaining)
2203 {
2204         int i;
2205
2206         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2207         if (remaining)
2208                 return 1;
2209
2210         /* If after HZ/10, a zone is below the high mark, it's premature */
2211         for (i = 0; i < pgdat->nr_zones; i++) {
2212                 struct zone *zone = pgdat->node_zones + i;
2213
2214                 if (!populated_zone(zone))
2215                         continue;
2216
2217                 if (zone->all_unreclaimable)
2218                         continue;
2219
2220                 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2221                                                                 0, 0))
2222                         return 1;
2223         }
2224
2225         return 0;
2226 }
2227
2228 /*
2229  * For kswapd, balance_pgdat() will work across all this node's zones until
2230  * they are all at high_wmark_pages(zone).
2231  *
2232  * Returns the number of pages which were actually freed.
2233  *
2234  * There is special handling here for zones which are full of pinned pages.
2235  * This can happen if the pages are all mlocked, or if they are all used by
2236  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2237  * What we do is to detect the case where all pages in the zone have been
2238  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2239  * dead and from now on, only perform a short scan.  Basically we're polling
2240  * the zone for when the problem goes away.
2241  *
2242  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2243  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2244  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2245  * lower zones regardless of the number of free pages in the lower zones. This
2246  * interoperates with the page allocator fallback scheme to ensure that aging
2247  * of pages is balanced across the zones.
2248  */
2249 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
2250 {
2251         int all_zones_ok;
2252         int priority;
2253         int i;
2254         unsigned long total_scanned;
2255         struct reclaim_state *reclaim_state = current->reclaim_state;
2256         struct scan_control sc = {
2257                 .gfp_mask = GFP_KERNEL,
2258                 .may_unmap = 1,
2259                 .may_swap = 1,
2260                 /*
2261                  * kswapd doesn't want to be bailed out while reclaim. because
2262                  * we want to put equal scanning pressure on each zone.
2263                  */
2264                 .nr_to_reclaim = ULONG_MAX,
2265                 .swappiness = vm_swappiness,
2266                 .order = order,
2267                 .mem_cgroup = NULL,
2268         };
2269 loop_again:
2270         total_scanned = 0;
2271         sc.nr_reclaimed = 0;
2272         sc.may_writepage = !laptop_mode;
2273         count_vm_event(PAGEOUTRUN);
2274
2275         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2276                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2277                 unsigned long lru_pages = 0;
2278                 int has_under_min_watermark_zone = 0;
2279
2280                 /* The swap token gets in the way of swapout... */
2281                 if (!priority)
2282                         disable_swap_token();
2283
2284                 all_zones_ok = 1;
2285
2286                 /*
2287                  * Scan in the highmem->dma direction for the highest
2288                  * zone which needs scanning
2289                  */
2290                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2291                         struct zone *zone = pgdat->node_zones + i;
2292
2293                         if (!populated_zone(zone))
2294                                 continue;
2295
2296                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2297                                 continue;
2298
2299                         /*
2300                          * Do some background aging of the anon list, to give
2301                          * pages a chance to be referenced before reclaiming.
2302                          */
2303                         if (inactive_anon_is_low(zone, &sc))
2304                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2305                                                         &sc, priority, 0);
2306
2307                         if (!zone_watermark_ok_safe(zone, order,
2308                                         high_wmark_pages(zone), 0, 0)) {
2309                                 end_zone = i;
2310                                 break;
2311                         }
2312                 }
2313                 if (i < 0)
2314                         goto out;
2315
2316                 for (i = 0; i <= end_zone; i++) {
2317                         struct zone *zone = pgdat->node_zones + i;
2318
2319                         lru_pages += zone_reclaimable_pages(zone);
2320                 }
2321
2322                 /*
2323                  * Now scan the zone in the dma->highmem direction, stopping
2324                  * at the last zone which needs scanning.
2325                  *
2326                  * We do this because the page allocator works in the opposite
2327                  * direction.  This prevents the page allocator from allocating
2328                  * pages behind kswapd's direction of progress, which would
2329                  * cause too much scanning of the lower zones.
2330                  */
2331                 for (i = 0; i <= end_zone; i++) {
2332                         struct zone *zone = pgdat->node_zones + i;
2333                         int nr_slab;
2334
2335                         if (!populated_zone(zone))
2336                                 continue;
2337
2338                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2339                                 continue;
2340
2341                         sc.nr_scanned = 0;
2342
2343                         /*
2344                          * Call soft limit reclaim before calling shrink_zone.
2345                          * For now we ignore the return value
2346                          */
2347                         mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask);
2348
2349                         /*
2350                          * We put equal pressure on every zone, unless one
2351                          * zone has way too many pages free already.
2352                          */
2353                         if (!zone_watermark_ok_safe(zone, order,
2354                                         8*high_wmark_pages(zone), end_zone, 0))
2355                                 shrink_zone(priority, zone, &sc);
2356                         reclaim_state->reclaimed_slab = 0;
2357                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2358                                                 lru_pages);
2359                         sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2360                         total_scanned += sc.nr_scanned;
2361                         if (zone->all_unreclaimable)
2362                                 continue;
2363                         if (nr_slab == 0 && !zone_reclaimable(zone))
2364                                 zone->all_unreclaimable = 1;
2365                         /*
2366                          * If we've done a decent amount of scanning and
2367                          * the reclaim ratio is low, start doing writepage
2368                          * even in laptop mode
2369                          */
2370                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2371                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2372                                 sc.may_writepage = 1;
2373
2374                         /*
2375                          * Compact the zone for higher orders to reduce
2376                          * latencies for higher-order allocations that
2377                          * would ordinarily call try_to_compact_pages()
2378                          */
2379                         if (sc.order > PAGE_ALLOC_COSTLY_ORDER)
2380                                 compact_zone_order(zone, sc.order, sc.gfp_mask,
2381                                                         false);
2382
2383                         if (!zone_watermark_ok_safe(zone, order,
2384                                         high_wmark_pages(zone), end_zone, 0)) {
2385                                 all_zones_ok = 0;
2386                                 /*
2387                                  * We are still under min water mark.  This
2388                                  * means that we have a GFP_ATOMIC allocation
2389                                  * failure risk. Hurry up!
2390                                  */
2391                                 if (!zone_watermark_ok_safe(zone, order,
2392                                             min_wmark_pages(zone), end_zone, 0))
2393                                         has_under_min_watermark_zone = 1;
2394                         } else {
2395                                 /*
2396                                  * If a zone reaches its high watermark,
2397                                  * consider it to be no longer congested. It's
2398                                  * possible there are dirty pages backed by
2399                                  * congested BDIs but as pressure is relieved,
2400                                  * spectulatively avoid congestion waits
2401                                  */
2402                                 zone_clear_flag(zone, ZONE_CONGESTED);
2403                         }
2404
2405                 }
2406                 if (all_zones_ok)
2407                         break;          /* kswapd: all done */
2408                 /*
2409                  * OK, kswapd is getting into trouble.  Take a nap, then take
2410                  * another pass across the zones.
2411                  */
2412                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2413                         if (has_under_min_watermark_zone)
2414                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2415                         else
2416                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2417                 }
2418
2419                 /*
2420                  * We do this so kswapd doesn't build up large priorities for
2421                  * example when it is freeing in parallel with allocators. It
2422                  * matches the direct reclaim path behaviour in terms of impact
2423                  * on zone->*_priority.
2424                  */
2425                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2426                         break;
2427         }
2428 out:
2429         if (!all_zones_ok) {
2430                 cond_resched();
2431
2432                 try_to_freeze();
2433
2434                 /*
2435                  * Fragmentation may mean that the system cannot be
2436                  * rebalanced for high-order allocations in all zones.
2437                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2438                  * it means the zones have been fully scanned and are still
2439                  * not balanced. For high-order allocations, there is
2440                  * little point trying all over again as kswapd may
2441                  * infinite loop.
2442                  *
2443                  * Instead, recheck all watermarks at order-0 as they
2444                  * are the most important. If watermarks are ok, kswapd will go
2445                  * back to sleep. High-order users can still perform direct
2446                  * reclaim if they wish.
2447                  */
2448                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2449                         order = sc.order = 0;
2450
2451                 goto loop_again;
2452         }
2453
2454         return sc.nr_reclaimed;
2455 }
2456
2457 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order)
2458 {
2459         long remaining = 0;
2460         DEFINE_WAIT(wait);
2461
2462         if (freezing(current) || kthread_should_stop())
2463                 return;
2464
2465         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2466
2467         /* Try to sleep for a short interval */
2468         if (!sleeping_prematurely(pgdat, order, remaining)) {
2469                 remaining = schedule_timeout(HZ/10);
2470                 finish_wait(&pgdat->kswapd_wait, &wait);
2471                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2472         }
2473
2474         /*
2475          * After a short sleep, check if it was a premature sleep. If not, then
2476          * go fully to sleep until explicitly woken up.
2477          */
2478         if (!sleeping_prematurely(pgdat, order, remaining)) {
2479                 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2480
2481                 /*
2482                  * vmstat counters are not perfectly accurate and the estimated
2483                  * value for counters such as NR_FREE_PAGES can deviate from the
2484                  * true value by nr_online_cpus * threshold. To avoid the zone
2485                  * watermarks being breached while under pressure, we reduce the
2486                  * per-cpu vmstat threshold while kswapd is awake and restore
2487                  * them before going back to sleep.
2488                  */
2489                 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2490                 schedule();
2491                 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2492         } else {
2493                 if (remaining)
2494                         count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2495                 else
2496                         count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2497         }
2498         finish_wait(&pgdat->kswapd_wait, &wait);
2499 }
2500
2501 /*
2502  * The background pageout daemon, started as a kernel thread
2503  * from the init process.
2504  *
2505  * This basically trickles out pages so that we have _some_
2506  * free memory available even if there is no other activity
2507  * that frees anything up. This is needed for things like routing
2508  * etc, where we otherwise might have all activity going on in
2509  * asynchronous contexts that cannot page things out.
2510  *
2511  * If there are applications that are active memory-allocators
2512  * (most normal use), this basically shouldn't matter.
2513  */
2514 static int kswapd(void *p)
2515 {
2516         unsigned long order;
2517         pg_data_t *pgdat = (pg_data_t*)p;
2518         struct task_struct *tsk = current;
2519
2520         struct reclaim_state reclaim_state = {
2521                 .reclaimed_slab = 0,
2522         };
2523         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2524
2525         lockdep_set_current_reclaim_state(GFP_KERNEL);
2526
2527         if (!cpumask_empty(cpumask))
2528                 set_cpus_allowed_ptr(tsk, cpumask);
2529         current->reclaim_state = &reclaim_state;
2530
2531         /*
2532          * Tell the memory management that we're a "memory allocator",
2533          * and that if we need more memory we should get access to it
2534          * regardless (see "__alloc_pages()"). "kswapd" should
2535          * never get caught in the normal page freeing logic.
2536          *
2537          * (Kswapd normally doesn't need memory anyway, but sometimes
2538          * you need a small amount of memory in order to be able to
2539          * page out something else, and this flag essentially protects
2540          * us from recursively trying to free more memory as we're
2541          * trying to free the first piece of memory in the first place).
2542          */
2543         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2544         set_freezable();
2545
2546         order = 0;
2547         for ( ; ; ) {
2548                 unsigned long new_order;
2549                 int ret;
2550
2551                 new_order = pgdat->kswapd_max_order;
2552                 pgdat->kswapd_max_order = 0;
2553                 if (order < new_order) {
2554                         /*
2555                          * Don't sleep if someone wants a larger 'order'
2556                          * allocation
2557                          */
2558                         order = new_order;
2559                 } else {
2560                         kswapd_try_to_sleep(pgdat, order);
2561                         order = pgdat->kswapd_max_order;
2562                 }
2563
2564                 ret = try_to_freeze();
2565                 if (kthread_should_stop())
2566                         break;
2567
2568                 /*
2569                  * We can speed up thawing tasks if we don't call balance_pgdat
2570                  * after returning from the refrigerator
2571                  */
2572                 if (!ret) {
2573                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2574                         balance_pgdat(pgdat, order);
2575                 }
2576         }
2577         return 0;
2578 }
2579
2580 /*
2581  * A zone is low on free memory, so wake its kswapd task to service it.
2582  */
2583 void wakeup_kswapd(struct zone *zone, int order)
2584 {
2585         pg_data_t *pgdat;
2586
2587         if (!populated_zone(zone))
2588                 return;
2589
2590         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2591                 return;
2592         pgdat = zone->zone_pgdat;
2593         if (pgdat->kswapd_max_order < order)
2594                 pgdat->kswapd_max_order = order;
2595         if (!waitqueue_active(&pgdat->kswapd_wait))
2596                 return;
2597         if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2598                 return;
2599
2600         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2601         wake_up_interruptible(&pgdat->kswapd_wait);
2602 }
2603
2604 /*
2605  * The reclaimable count would be mostly accurate.
2606  * The less reclaimable pages may be
2607  * - mlocked pages, which will be moved to unevictable list when encountered
2608  * - mapped pages, which may require several travels to be reclaimed
2609  * - dirty pages, which is not "instantly" reclaimable
2610  */
2611 unsigned long global_reclaimable_pages(void)
2612 {
2613         int nr;
2614
2615         nr = global_page_state(NR_ACTIVE_FILE) +
2616              global_page_state(NR_INACTIVE_FILE);
2617
2618         if (nr_swap_pages > 0)
2619                 nr += global_page_state(NR_ACTIVE_ANON) +
2620                       global_page_state(NR_INACTIVE_ANON);
2621
2622         return nr;
2623 }
2624
2625 unsigned long zone_reclaimable_pages(struct zone *zone)
2626 {
2627         int nr;
2628
2629         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2630              zone_page_state(zone, NR_INACTIVE_FILE);
2631
2632         if (nr_swap_pages > 0)
2633                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2634                       zone_page_state(zone, NR_INACTIVE_ANON);
2635
2636         return nr;
2637 }
2638
2639 #ifdef CONFIG_HIBERNATION
2640 /*
2641  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2642  * freed pages.
2643  *
2644  * Rather than trying to age LRUs the aim is to preserve the overall
2645  * LRU order by reclaiming preferentially
2646  * inactive > active > active referenced > active mapped
2647  */
2648 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2649 {
2650         struct reclaim_state reclaim_state;
2651         struct scan_control sc = {
2652                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2653                 .may_swap = 1,
2654                 .may_unmap = 1,
2655                 .may_writepage = 1,
2656                 .nr_to_reclaim = nr_to_reclaim,
2657                 .hibernation_mode = 1,
2658                 .swappiness = vm_swappiness,
2659                 .order = 0,
2660         };
2661         struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2662         struct task_struct *p = current;
2663         unsigned long nr_reclaimed;
2664
2665         p->flags |= PF_MEMALLOC;
2666         lockdep_set_current_reclaim_state(sc.gfp_mask);
2667         reclaim_state.reclaimed_slab = 0;
2668         p->reclaim_state = &reclaim_state;
2669
2670         nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2671
2672         p->reclaim_state = NULL;
2673         lockdep_clear_current_reclaim_state();
2674         p->flags &= ~PF_MEMALLOC;
2675
2676         return nr_reclaimed;
2677 }
2678 #endif /* CONFIG_HIBERNATION */
2679
2680 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2681    not required for correctness.  So if the last cpu in a node goes
2682    away, we get changed to run anywhere: as the first one comes back,
2683    restore their cpu bindings. */
2684 static int __devinit cpu_callback(struct notifier_block *nfb,
2685                                   unsigned long action, void *hcpu)
2686 {
2687         int nid;
2688
2689         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2690                 for_each_node_state(nid, N_HIGH_MEMORY) {
2691                         pg_data_t *pgdat = NODE_DATA(nid);
2692                         const struct cpumask *mask;
2693
2694                         mask = cpumask_of_node(pgdat->node_id);
2695
2696                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2697                                 /* One of our CPUs online: restore mask */
2698                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2699                 }
2700         }
2701         return NOTIFY_OK;
2702 }
2703
2704 /*
2705  * This kswapd start function will be called by init and node-hot-add.
2706  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2707  */
2708 int kswapd_run(int nid)
2709 {
2710         pg_data_t *pgdat = NODE_DATA(nid);
2711         int ret = 0;
2712
2713         if (pgdat->kswapd)
2714                 return 0;
2715
2716         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2717         if (IS_ERR(pgdat->kswapd)) {
2718                 /* failure at boot is fatal */
2719                 BUG_ON(system_state == SYSTEM_BOOTING);
2720                 printk("Failed to start kswapd on node %d\n",nid);
2721                 ret = -1;
2722         }
2723         return ret;
2724 }
2725
2726 /*
2727  * Called by memory hotplug when all memory in a node is offlined.
2728  */
2729 void kswapd_stop(int nid)
2730 {
2731         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2732
2733         if (kswapd)
2734                 kthread_stop(kswapd);
2735 }
2736
2737 static int __init kswapd_init(void)
2738 {
2739         int nid;
2740
2741         swap_setup();
2742         for_each_node_state(nid, N_HIGH_MEMORY)
2743                 kswapd_run(nid);
2744         hotcpu_notifier(cpu_callback, 0);
2745         return 0;
2746 }
2747
2748 module_init(kswapd_init)
2749
2750 #ifdef CONFIG_NUMA
2751 /*
2752  * Zone reclaim mode
2753  *
2754  * If non-zero call zone_reclaim when the number of free pages falls below
2755  * the watermarks.
2756  */
2757 int zone_reclaim_mode __read_mostly;
2758
2759 #define RECLAIM_OFF 0
2760 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
2761 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
2762 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
2763
2764 /*
2765  * Priority for ZONE_RECLAIM. This determines the fraction of pages
2766  * of a node considered for each zone_reclaim. 4 scans 1/16th of
2767  * a zone.
2768  */
2769 #define ZONE_RECLAIM_PRIORITY 4
2770
2771 /*
2772  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2773  * occur.
2774  */
2775 int sysctl_min_unmapped_ratio = 1;
2776
2777 /*
2778  * If the number of slab pages in a zone grows beyond this percentage then
2779  * slab reclaim needs to occur.
2780  */
2781 int sysctl_min_slab_ratio = 5;
2782
2783 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2784 {
2785         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2786         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2787                 zone_page_state(zone, NR_ACTIVE_FILE);
2788
2789         /*
2790          * It's possible for there to be more file mapped pages than
2791          * accounted for by the pages on the file LRU lists because
2792          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2793          */
2794         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2795 }
2796
2797 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
2798 static long zone_pagecache_reclaimable(struct zone *zone)
2799 {
2800         long nr_pagecache_reclaimable;
2801         long delta = 0;
2802
2803         /*
2804          * If RECLAIM_SWAP is set, then all file pages are considered
2805          * potentially reclaimable. Otherwise, we have to worry about
2806          * pages like swapcache and zone_unmapped_file_pages() provides
2807          * a better estimate
2808          */
2809         if (zone_reclaim_mode & RECLAIM_SWAP)
2810                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2811         else
2812                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2813
2814         /* If we can't clean pages, remove dirty pages from consideration */
2815         if (!(zone_reclaim_mode & RECLAIM_WRITE))
2816                 delta += zone_page_state(zone, NR_FILE_DIRTY);
2817
2818         /* Watch for any possible underflows due to delta */
2819         if (unlikely(delta > nr_pagecache_reclaimable))
2820                 delta = nr_pagecache_reclaimable;
2821
2822         return nr_pagecache_reclaimable - delta;
2823 }
2824
2825 /*
2826  * Try to free up some pages from this zone through reclaim.
2827  */
2828 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2829 {
2830         /* Minimum pages needed in order to stay on node */
2831         const unsigned long nr_pages = 1 << order;
2832         struct task_struct *p = current;
2833         struct reclaim_state reclaim_state;
2834         int priority;
2835         struct scan_control sc = {
2836                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2837                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
2838                 .may_swap = 1,
2839                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
2840                                        SWAP_CLUSTER_MAX),
2841                 .gfp_mask = gfp_mask,
2842                 .swappiness = vm_swappiness,
2843                 .order = order,
2844         };
2845         unsigned long nr_slab_pages0, nr_slab_pages1;
2846
2847         cond_resched();
2848         /*
2849          * We need to be able to allocate from the reserves for RECLAIM_SWAP
2850          * and we also need to be able to write out pages for RECLAIM_WRITE
2851          * and RECLAIM_SWAP.
2852          */
2853         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
2854         lockdep_set_current_reclaim_state(gfp_mask);
2855         reclaim_state.reclaimed_slab = 0;
2856         p->reclaim_state = &reclaim_state;
2857
2858         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
2859                 /*
2860                  * Free memory by calling shrink zone with increasing
2861                  * priorities until we have enough memory freed.
2862                  */
2863                 priority = ZONE_RECLAIM_PRIORITY;
2864                 do {
2865                         shrink_zone(priority, zone, &sc);
2866                         priority--;
2867                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
2868         }
2869
2870         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2871         if (nr_slab_pages0 > zone->min_slab_pages) {
2872                 /*
2873                  * shrink_slab() does not currently allow us to determine how
2874                  * many pages were freed in this zone. So we take the current
2875                  * number of slab pages and shake the slab until it is reduced
2876                  * by the same nr_pages that we used for reclaiming unmapped
2877                  * pages.
2878                  *
2879                  * Note that shrink_slab will free memory on all zones and may
2880                  * take a long time.
2881                  */
2882                 for (;;) {
2883                         unsigned long lru_pages = zone_reclaimable_pages(zone);
2884
2885                         /* No reclaimable slab or very low memory pressure */
2886                         if (!shrink_slab(sc.nr_scanned, gfp_mask, lru_pages))
2887                                 break;
2888
2889                         /* Freed enough memory */
2890                         nr_slab_pages1 = zone_page_state(zone,
2891                                                         NR_SLAB_RECLAIMABLE);
2892                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
2893                                 break;
2894                 }
2895
2896                 /*
2897                  * Update nr_reclaimed by the number of slab pages we
2898                  * reclaimed from this zone.
2899                  */
2900                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2901                 if (nr_slab_pages1 < nr_slab_pages0)
2902                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
2903         }
2904
2905         p->reclaim_state = NULL;
2906         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
2907         lockdep_clear_current_reclaim_state();
2908         return sc.nr_reclaimed >= nr_pages;
2909 }
2910
2911 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2912 {
2913         int node_id;
2914         int ret;
2915
2916         /*
2917          * Zone reclaim reclaims unmapped file backed pages and
2918          * slab pages if we are over the defined limits.
2919          *
2920          * A small portion of unmapped file backed pages is needed for
2921          * file I/O otherwise pages read by file I/O will be immediately
2922          * thrown out if the zone is overallocated. So we do not reclaim
2923          * if less than a specified percentage of the zone is used by
2924          * unmapped file backed pages.
2925          */
2926         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2927             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
2928                 return ZONE_RECLAIM_FULL;
2929
2930         if (zone->all_unreclaimable)
2931                 return ZONE_RECLAIM_FULL;
2932
2933         /*
2934          * Do not scan if the allocation should not be delayed.
2935          */
2936         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
2937                 return ZONE_RECLAIM_NOSCAN;
2938
2939         /*
2940          * Only run zone reclaim on the local zone or on zones that do not
2941          * have associated processors. This will favor the local processor
2942          * over remote processors and spread off node memory allocations
2943          * as wide as possible.
2944          */
2945         node_id = zone_to_nid(zone);
2946         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
2947                 return ZONE_RECLAIM_NOSCAN;
2948
2949         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2950                 return ZONE_RECLAIM_NOSCAN;
2951
2952         ret = __zone_reclaim(zone, gfp_mask, order);
2953         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2954
2955         if (!ret)
2956                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2957
2958         return ret;
2959 }
2960 #endif
2961
2962 /*
2963  * page_evictable - test whether a page is evictable
2964  * @page: the page to test
2965  * @vma: the VMA in which the page is or will be mapped, may be NULL
2966  *
2967  * Test whether page is evictable--i.e., should be placed on active/inactive
2968  * lists vs unevictable list.  The vma argument is !NULL when called from the
2969  * fault path to determine how to instantate a new page.
2970  *
2971  * Reasons page might not be evictable:
2972  * (1) page's mapping marked unevictable
2973  * (2) page is part of an mlocked VMA
2974  *
2975  */
2976 int page_evictable(struct page *page, struct vm_area_struct *vma)
2977 {
2978
2979         if (mapping_unevictable(page_mapping(page)))
2980                 return 0;
2981
2982         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2983                 return 0;
2984
2985         return 1;
2986 }
2987
2988 /**
2989  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2990  * @page: page to check evictability and move to appropriate lru list
2991  * @zone: zone page is in
2992  *
2993  * Checks a page for evictability and moves the page to the appropriate
2994  * zone lru list.
2995  *
2996  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2997  * have PageUnevictable set.
2998  */
2999 static void check_move_unevictable_page(struct page *page, struct zone *zone)
3000 {
3001         VM_BUG_ON(PageActive(page));
3002
3003 retry:
3004         ClearPageUnevictable(page);
3005         if (page_evictable(page, NULL)) {
3006                 enum lru_list l = page_lru_base_type(page);
3007
3008                 __dec_zone_state(zone, NR_UNEVICTABLE);
3009                 list_move(&page->lru, &zone->lru[l].list);
3010                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
3011                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3012                 __count_vm_event(UNEVICTABLE_PGRESCUED);
3013         } else {
3014                 /*
3015                  * rotate unevictable list
3016                  */
3017                 SetPageUnevictable(page);
3018                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
3019                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
3020                 if (page_evictable(page, NULL))
3021                         goto retry;
3022         }
3023 }
3024
3025 /**
3026  * scan_mapping_unevictable_pages - scan an address space for evictable pages
3027  * @mapping: struct address_space to scan for evictable pages
3028  *
3029  * Scan all pages in mapping.  Check unevictable pages for
3030  * evictability and move them to the appropriate zone lru list.
3031  */
3032 void scan_mapping_unevictable_pages(struct address_space *mapping)
3033 {
3034         pgoff_t next = 0;
3035         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3036                          PAGE_CACHE_SHIFT;
3037         struct zone *zone;
3038         struct pagevec pvec;
3039
3040         if (mapping->nrpages == 0)
3041                 return;
3042
3043         pagevec_init(&pvec, 0);
3044         while (next < end &&
3045                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3046                 int i;
3047                 int pg_scanned = 0;
3048
3049                 zone = NULL;
3050
3051                 for (i = 0; i < pagevec_count(&pvec); i++) {
3052                         struct page *page = pvec.pages[i];
3053                         pgoff_t page_index = page->index;
3054                         struct zone *pagezone = page_zone(page);
3055
3056                         pg_scanned++;
3057                         if (page_index > next)
3058                                 next = page_index;
3059                         next++;
3060
3061                         if (pagezone != zone) {
3062                                 if (zone)
3063                                         spin_unlock_irq(&zone->lru_lock);
3064                                 zone = pagezone;
3065                                 spin_lock_irq(&zone->lru_lock);
3066                         }
3067
3068                         if (PageLRU(page) && PageUnevictable(page))
3069                                 check_move_unevictable_page(page, zone);
3070                 }
3071                 if (zone)
3072                         spin_unlock_irq(&zone->lru_lock);
3073                 pagevec_release(&pvec);
3074
3075                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3076         }
3077
3078 }
3079
3080 /**
3081  * scan_zone_unevictable_pages - check unevictable list for evictable pages
3082  * @zone - zone of which to scan the unevictable list
3083  *
3084  * Scan @zone's unevictable LRU lists to check for pages that have become
3085  * evictable.  Move those that have to @zone's inactive list where they
3086  * become candidates for reclaim, unless shrink_inactive_zone() decides
3087  * to reactivate them.  Pages that are still unevictable are rotated
3088  * back onto @zone's unevictable list.
3089  */
3090 #define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
3091 static void scan_zone_unevictable_pages(struct zone *zone)
3092 {
3093         struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
3094         unsigned long scan;
3095         unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
3096
3097         while (nr_to_scan > 0) {
3098                 unsigned long batch_size = min(nr_to_scan,
3099                                                 SCAN_UNEVICTABLE_BATCH_SIZE);
3100
3101                 spin_lock_irq(&zone->lru_lock);
3102                 for (scan = 0;  scan < batch_size; scan++) {
3103                         struct page *page = lru_to_page(l_unevictable);
3104
3105                         if (!trylock_page(page))
3106                                 continue;
3107
3108                         prefetchw_prev_lru_page(page, l_unevictable, flags);
3109
3110                         if (likely(PageLRU(page) && PageUnevictable(page)))
3111                                 check_move_unevictable_page(page, zone);
3112
3113                         unlock_page(page);
3114                 }
3115                 spin_unlock_irq(&zone->lru_lock);
3116
3117                 nr_to_scan -= batch_size;
3118         }
3119 }
3120
3121
3122 /**
3123  * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3124  *
3125  * A really big hammer:  scan all zones' unevictable LRU lists to check for
3126  * pages that have become evictable.  Move those back to the zones'
3127  * inactive list where they become candidates for reclaim.
3128  * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3129  * and we add swap to the system.  As such, it runs in the context of a task
3130  * that has possibly/probably made some previously unevictable pages
3131  * evictable.
3132  */
3133 static void scan_all_zones_unevictable_pages(void)
3134 {
3135         struct zone *zone;
3136
3137         for_each_zone(zone) {
3138                 scan_zone_unevictable_pages(zone);
3139         }
3140 }
3141
3142 /*
3143  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3144  * all nodes' unevictable lists for evictable pages
3145  */
3146 unsigned long scan_unevictable_pages;
3147
3148 int scan_unevictable_handler(struct ctl_table *table, int write,
3149                            void __user *buffer,
3150                            size_t *length, loff_t *ppos)
3151 {
3152         proc_doulongvec_minmax(table, write, buffer, length, ppos);
3153
3154         if (write && *(unsigned long *)table->data)
3155                 scan_all_zones_unevictable_pages();
3156
3157         scan_unevictable_pages = 0;
3158         return 0;
3159 }
3160
3161 #ifdef CONFIG_NUMA
3162 /*
3163  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3164  * a specified node's per zone unevictable lists for evictable pages.
3165  */
3166
3167 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3168                                           struct sysdev_attribute *attr,
3169                                           char *buf)
3170 {
3171         return sprintf(buf, "0\n");     /* always zero; should fit... */
3172 }
3173
3174 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3175                                            struct sysdev_attribute *attr,
3176                                         const char *buf, size_t count)
3177 {
3178         struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3179         struct zone *zone;
3180         unsigned long res;
3181         unsigned long req = strict_strtoul(buf, 10, &res);
3182
3183         if (!req)
3184                 return 1;       /* zero is no-op */
3185
3186         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3187                 if (!populated_zone(zone))
3188                         continue;
3189                 scan_zone_unevictable_pages(zone);
3190         }
3191         return 1;
3192 }
3193
3194
3195 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3196                         read_scan_unevictable_node,
3197                         write_scan_unevictable_node);
3198
3199 int scan_unevictable_register_node(struct node *node)
3200 {
3201         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3202 }
3203
3204 void scan_unevictable_unregister_node(struct node *node)
3205 {
3206         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
3207 }
3208 #endif