ARM: OMAP3: PM: remove access to PRM_VOLTCTRL register
[pandora-kernel.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 #include <linux/oom.h>
45 #include <linux/prefetch.h>
46
47 #include <asm/tlbflush.h>
48 #include <asm/div64.h>
49
50 #include <linux/swapops.h>
51
52 #include "internal.h"
53
54 #define CREATE_TRACE_POINTS
55 #include <trace/events/vmscan.h>
56
57 /*
58  * reclaim_mode determines how the inactive list is shrunk
59  * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
60  * RECLAIM_MODE_ASYNC:  Do not block
61  * RECLAIM_MODE_SYNC:   Allow blocking e.g. call wait_on_page_writeback
62  * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
63  *                      page from the LRU and reclaim all pages within a
64  *                      naturally aligned range
65  * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
66  *                      order-0 pages and then compact the zone
67  */
68 typedef unsigned __bitwise__ reclaim_mode_t;
69 #define RECLAIM_MODE_SINGLE             ((__force reclaim_mode_t)0x01u)
70 #define RECLAIM_MODE_ASYNC              ((__force reclaim_mode_t)0x02u)
71 #define RECLAIM_MODE_SYNC               ((__force reclaim_mode_t)0x04u)
72 #define RECLAIM_MODE_LUMPYRECLAIM       ((__force reclaim_mode_t)0x08u)
73 #define RECLAIM_MODE_COMPACTION         ((__force reclaim_mode_t)0x10u)
74
75 struct scan_control {
76         /* Incremented by the number of inactive pages that were scanned */
77         unsigned long nr_scanned;
78
79         /* Number of pages freed so far during a call to shrink_zones() */
80         unsigned long nr_reclaimed;
81
82         /* How many pages shrink_list() should reclaim */
83         unsigned long nr_to_reclaim;
84
85         unsigned long hibernation_mode;
86
87         /* This context's GFP mask */
88         gfp_t gfp_mask;
89
90         int may_writepage;
91
92         /* Can mapped pages be reclaimed? */
93         int may_unmap;
94
95         /* Can pages be swapped as part of reclaim? */
96         int may_swap;
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_lru_pages(sc->mem_cgroup,
175                                 zone_to_nid(zone), zone_idx(zone), BIT(lru));
176
177         return zone_page_state(zone, NR_LRU_BASE + lru);
178 }
179
180
181 /*
182  * Add a shrinker callback to be called from the vm
183  */
184 void register_shrinker(struct shrinker *shrinker)
185 {
186         atomic_long_set(&shrinker->nr_in_batch, 0);
187         down_write(&shrinker_rwsem);
188         list_add_tail(&shrinker->list, &shrinker_list);
189         up_write(&shrinker_rwsem);
190 }
191 EXPORT_SYMBOL(register_shrinker);
192
193 /*
194  * Remove one
195  */
196 void unregister_shrinker(struct shrinker *shrinker)
197 {
198         down_write(&shrinker_rwsem);
199         list_del(&shrinker->list);
200         up_write(&shrinker_rwsem);
201 }
202 EXPORT_SYMBOL(unregister_shrinker);
203
204 static inline int do_shrinker_shrink(struct shrinker *shrinker,
205                                      struct shrink_control *sc,
206                                      unsigned long nr_to_scan)
207 {
208         sc->nr_to_scan = nr_to_scan;
209         return (*shrinker->shrink)(shrinker, sc);
210 }
211
212 #define SHRINK_BATCH 128
213 /*
214  * Call the shrink functions to age shrinkable caches
215  *
216  * Here we assume it costs one seek to replace a lru page and that it also
217  * takes a seek to recreate a cache object.  With this in mind we age equal
218  * percentages of the lru and ageable caches.  This should balance the seeks
219  * generated by these structures.
220  *
221  * If the vm encountered mapped pages on the LRU it increase the pressure on
222  * slab to avoid swapping.
223  *
224  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
225  *
226  * `lru_pages' represents the number of on-LRU pages in all the zones which
227  * are eligible for the caller's allocation attempt.  It is used for balancing
228  * slab reclaim versus page reclaim.
229  *
230  * Returns the number of slab objects which we shrunk.
231  */
232 unsigned long shrink_slab(struct shrink_control *shrink,
233                           unsigned long nr_pages_scanned,
234                           unsigned long lru_pages)
235 {
236         struct shrinker *shrinker;
237         unsigned long ret = 0;
238
239         if (nr_pages_scanned == 0)
240                 nr_pages_scanned = SWAP_CLUSTER_MAX;
241
242         if (!down_read_trylock(&shrinker_rwsem)) {
243                 /* Assume we'll be able to shrink next time */
244                 ret = 1;
245                 goto out;
246         }
247
248         list_for_each_entry(shrinker, &shrinker_list, list) {
249                 unsigned long long delta;
250                 long total_scan;
251                 long max_pass;
252                 int shrink_ret = 0;
253                 long nr;
254                 long new_nr;
255                 long batch_size = shrinker->batch ? shrinker->batch
256                                                   : SHRINK_BATCH;
257
258                 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
259                 if (max_pass <= 0)
260                         continue;
261
262                 /*
263                  * copy the current shrinker scan count into a local variable
264                  * and zero it so that other concurrent shrinker invocations
265                  * don't also do this scanning work.
266                  */
267                 nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
268
269                 total_scan = nr;
270                 delta = (4 * nr_pages_scanned) / shrinker->seeks;
271                 delta *= max_pass;
272                 do_div(delta, lru_pages + 1);
273                 total_scan += delta;
274                 if (total_scan < 0) {
275                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
276                                "delete nr=%ld\n",
277                                shrinker->shrink, total_scan);
278                         total_scan = max_pass;
279                 }
280
281                 /*
282                  * We need to avoid excessive windup on filesystem shrinkers
283                  * due to large numbers of GFP_NOFS allocations causing the
284                  * shrinkers to return -1 all the time. This results in a large
285                  * nr being built up so when a shrink that can do some work
286                  * comes along it empties the entire cache due to nr >>>
287                  * max_pass.  This is bad for sustaining a working set in
288                  * memory.
289                  *
290                  * Hence only allow the shrinker to scan the entire cache when
291                  * a large delta change is calculated directly.
292                  */
293                 if (delta < max_pass / 4)
294                         total_scan = min(total_scan, max_pass / 2);
295
296                 /*
297                  * Avoid risking looping forever due to too large nr value:
298                  * never try to free more than twice the estimate number of
299                  * freeable entries.
300                  */
301                 if (total_scan > max_pass * 2)
302                         total_scan = max_pass * 2;
303
304                 trace_mm_shrink_slab_start(shrinker, shrink, nr,
305                                         nr_pages_scanned, lru_pages,
306                                         max_pass, delta, total_scan);
307
308                 while (total_scan >= batch_size) {
309                         int nr_before;
310
311                         nr_before = do_shrinker_shrink(shrinker, shrink, 0);
312                         shrink_ret = do_shrinker_shrink(shrinker, shrink,
313                                                         batch_size);
314                         if (shrink_ret == -1)
315                                 break;
316                         if (shrink_ret < nr_before)
317                                 ret += nr_before - shrink_ret;
318                         count_vm_events(SLABS_SCANNED, batch_size);
319                         total_scan -= batch_size;
320
321                         cond_resched();
322                 }
323
324                 /*
325                  * move the unused scan count back into the shrinker in a
326                  * manner that handles concurrent updates. If we exhausted the
327                  * scan, there is no need to do an update.
328                  */
329                 if (total_scan > 0)
330                         new_nr = atomic_long_add_return(total_scan,
331                                         &shrinker->nr_in_batch);
332                 else
333                         new_nr = atomic_long_read(&shrinker->nr_in_batch);
334
335                 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
336         }
337         up_read(&shrinker_rwsem);
338 out:
339         cond_resched();
340         return ret;
341 }
342
343 static void set_reclaim_mode(int priority, struct scan_control *sc,
344                                    bool sync)
345 {
346         reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
347
348         /*
349          * Initially assume we are entering either lumpy reclaim or
350          * reclaim/compaction.Depending on the order, we will either set the
351          * sync mode or just reclaim order-0 pages later.
352          */
353         if (COMPACTION_BUILD)
354                 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
355         else
356                 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
357
358         /*
359          * Avoid using lumpy reclaim or reclaim/compaction if possible by
360          * restricting when its set to either costly allocations or when
361          * under memory pressure
362          */
363         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
364                 sc->reclaim_mode |= syncmode;
365         else if (sc->order && priority < DEF_PRIORITY - 2)
366                 sc->reclaim_mode |= syncmode;
367         else
368                 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
369 }
370
371 static void reset_reclaim_mode(struct scan_control *sc)
372 {
373         sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
374 }
375
376 static inline int is_page_cache_freeable(struct page *page)
377 {
378         /*
379          * A freeable page cache page is referenced only by the caller
380          * that isolated the page, the page cache radix tree and
381          * optional buffer heads at page->private.
382          */
383         return page_count(page) - page_has_private(page) == 2;
384 }
385
386 static int may_write_to_queue(struct backing_dev_info *bdi,
387                               struct scan_control *sc)
388 {
389         if (current->flags & PF_SWAPWRITE)
390                 return 1;
391         if (!bdi_write_congested(bdi))
392                 return 1;
393         if (bdi == current->backing_dev_info)
394                 return 1;
395
396         /* lumpy reclaim for hugepage often need a lot of write */
397         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
398                 return 1;
399         return 0;
400 }
401
402 /*
403  * We detected a synchronous write error writing a page out.  Probably
404  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
405  * fsync(), msync() or close().
406  *
407  * The tricky part is that after writepage we cannot touch the mapping: nothing
408  * prevents it from being freed up.  But we have a ref on the page and once
409  * that page is locked, the mapping is pinned.
410  *
411  * We're allowed to run sleeping lock_page() here because we know the caller has
412  * __GFP_FS.
413  */
414 static void handle_write_error(struct address_space *mapping,
415                                 struct page *page, int error)
416 {
417         lock_page(page);
418         if (page_mapping(page) == mapping)
419                 mapping_set_error(mapping, error);
420         unlock_page(page);
421 }
422
423 /* possible outcome of pageout() */
424 typedef enum {
425         /* failed to write page out, page is locked */
426         PAGE_KEEP,
427         /* move page to the active list, page is locked */
428         PAGE_ACTIVATE,
429         /* page has been sent to the disk successfully, page is unlocked */
430         PAGE_SUCCESS,
431         /* page is clean and locked */
432         PAGE_CLEAN,
433 } pageout_t;
434
435 /*
436  * pageout is called by shrink_page_list() for each dirty page.
437  * Calls ->writepage().
438  */
439 static pageout_t pageout(struct page *page, struct address_space *mapping,
440                          struct scan_control *sc)
441 {
442         /*
443          * If the page is dirty, only perform writeback if that write
444          * will be non-blocking.  To prevent this allocation from being
445          * stalled by pagecache activity.  But note that there may be
446          * stalls if we need to run get_block().  We could test
447          * PagePrivate for that.
448          *
449          * If this process is currently in __generic_file_aio_write() against
450          * this page's queue, we can perform writeback even if that
451          * will block.
452          *
453          * If the page is swapcache, write it back even if that would
454          * block, for some throttling. This happens by accident, because
455          * swap_backing_dev_info is bust: it doesn't reflect the
456          * congestion state of the swapdevs.  Easy to fix, if needed.
457          */
458         if (!is_page_cache_freeable(page))
459                 return PAGE_KEEP;
460         if (!mapping) {
461                 /*
462                  * Some data journaling orphaned pages can have
463                  * page->mapping == NULL while being dirty with clean buffers.
464                  */
465                 if (page_has_private(page)) {
466                         if (try_to_free_buffers(page)) {
467                                 ClearPageDirty(page);
468                                 printk("%s: orphaned page\n", __func__);
469                                 return PAGE_CLEAN;
470                         }
471                 }
472                 return PAGE_KEEP;
473         }
474         if (mapping->a_ops->writepage == NULL)
475                 return PAGE_ACTIVATE;
476         if (!may_write_to_queue(mapping->backing_dev_info, sc))
477                 return PAGE_KEEP;
478
479         if (clear_page_dirty_for_io(page)) {
480                 int res;
481                 struct writeback_control wbc = {
482                         .sync_mode = WB_SYNC_NONE,
483                         .nr_to_write = SWAP_CLUSTER_MAX,
484                         .range_start = 0,
485                         .range_end = LLONG_MAX,
486                         .for_reclaim = 1,
487                 };
488
489                 SetPageReclaim(page);
490                 res = mapping->a_ops->writepage(page, &wbc);
491                 if (res < 0)
492                         handle_write_error(mapping, page, res);
493                 if (res == AOP_WRITEPAGE_ACTIVATE) {
494                         ClearPageReclaim(page);
495                         return PAGE_ACTIVATE;
496                 }
497
498                 if (!PageWriteback(page)) {
499                         /* synchronous write or broken a_ops? */
500                         ClearPageReclaim(page);
501                 }
502                 trace_mm_vmscan_writepage(page,
503                         trace_reclaim_flags(page, sc->reclaim_mode));
504                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
505                 return PAGE_SUCCESS;
506         }
507
508         return PAGE_CLEAN;
509 }
510
511 /*
512  * Same as remove_mapping, but if the page is removed from the mapping, it
513  * gets returned with a refcount of 0.
514  */
515 static int __remove_mapping(struct address_space *mapping, struct page *page)
516 {
517         BUG_ON(!PageLocked(page));
518         BUG_ON(mapping != page_mapping(page));
519
520         spin_lock_irq(&mapping->tree_lock);
521         /*
522          * The non racy check for a busy page.
523          *
524          * Must be careful with the order of the tests. When someone has
525          * a ref to the page, it may be possible that they dirty it then
526          * drop the reference. So if PageDirty is tested before page_count
527          * here, then the following race may occur:
528          *
529          * get_user_pages(&page);
530          * [user mapping goes away]
531          * write_to(page);
532          *                              !PageDirty(page)    [good]
533          * SetPageDirty(page);
534          * put_page(page);
535          *                              !page_count(page)   [good, discard it]
536          *
537          * [oops, our write_to data is lost]
538          *
539          * Reversing the order of the tests ensures such a situation cannot
540          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
541          * load is not satisfied before that of page->_count.
542          *
543          * Note that if SetPageDirty is always performed via set_page_dirty,
544          * and thus under tree_lock, then this ordering is not required.
545          */
546         if (!page_freeze_refs(page, 2))
547                 goto cannot_free;
548         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
549         if (unlikely(PageDirty(page))) {
550                 page_unfreeze_refs(page, 2);
551                 goto cannot_free;
552         }
553
554         if (PageSwapCache(page)) {
555                 swp_entry_t swap = { .val = page_private(page) };
556                 __delete_from_swap_cache(page);
557                 spin_unlock_irq(&mapping->tree_lock);
558                 swapcache_free(swap, page);
559         } else {
560                 void (*freepage)(struct page *);
561
562                 freepage = mapping->a_ops->freepage;
563
564                 __delete_from_page_cache(page);
565                 spin_unlock_irq(&mapping->tree_lock);
566                 mem_cgroup_uncharge_cache_page(page);
567
568                 if (freepage != NULL)
569                         freepage(page);
570         }
571
572         return 1;
573
574 cannot_free:
575         spin_unlock_irq(&mapping->tree_lock);
576         return 0;
577 }
578
579 /*
580  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
581  * someone else has a ref on the page, abort and return 0.  If it was
582  * successfully detached, return 1.  Assumes the caller has a single ref on
583  * this page.
584  */
585 int remove_mapping(struct address_space *mapping, struct page *page)
586 {
587         if (__remove_mapping(mapping, page)) {
588                 /*
589                  * Unfreezing the refcount with 1 rather than 2 effectively
590                  * drops the pagecache ref for us without requiring another
591                  * atomic operation.
592                  */
593                 page_unfreeze_refs(page, 1);
594                 return 1;
595         }
596         return 0;
597 }
598
599 /**
600  * putback_lru_page - put previously isolated page onto appropriate LRU list
601  * @page: page to be put back to appropriate lru list
602  *
603  * Add previously isolated @page to appropriate LRU list.
604  * Page may still be unevictable for other reasons.
605  *
606  * lru_lock must not be held, interrupts must be enabled.
607  */
608 void putback_lru_page(struct page *page)
609 {
610         int lru;
611         int active = !!TestClearPageActive(page);
612         int was_unevictable = PageUnevictable(page);
613
614         VM_BUG_ON(PageLRU(page));
615
616 redo:
617         ClearPageUnevictable(page);
618
619         if (page_evictable(page, NULL)) {
620                 /*
621                  * For evictable pages, we can use the cache.
622                  * In event of a race, worst case is we end up with an
623                  * unevictable page on [in]active list.
624                  * We know how to handle that.
625                  */
626                 lru = active + page_lru_base_type(page);
627                 lru_cache_add_lru(page, lru);
628         } else {
629                 /*
630                  * Put unevictable pages directly on zone's unevictable
631                  * list.
632                  */
633                 lru = LRU_UNEVICTABLE;
634                 add_page_to_unevictable_list(page);
635                 /*
636                  * When racing with an mlock or AS_UNEVICTABLE clearing
637                  * (page is unlocked) make sure that if the other thread
638                  * does not observe our setting of PG_lru and fails
639                  * isolation/check_move_unevictable_pages,
640                  * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
641                  * the page back to the evictable list.
642                  *
643                  * The other side is TestClearPageMlocked() or shmem_lock().
644                  */
645                 smp_mb();
646         }
647
648         /*
649          * page's status can change while we move it among lru. If an evictable
650          * page is on unevictable list, it never be freed. To avoid that,
651          * check after we added it to the list, again.
652          */
653         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
654                 if (!isolate_lru_page(page)) {
655                         put_page(page);
656                         goto redo;
657                 }
658                 /* This means someone else dropped this page from LRU
659                  * So, it will be freed or putback to LRU again. There is
660                  * nothing to do here.
661                  */
662         }
663
664         if (was_unevictable && lru != LRU_UNEVICTABLE)
665                 count_vm_event(UNEVICTABLE_PGRESCUED);
666         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
667                 count_vm_event(UNEVICTABLE_PGCULLED);
668
669         put_page(page);         /* drop ref from isolate */
670 }
671
672 enum page_references {
673         PAGEREF_RECLAIM,
674         PAGEREF_RECLAIM_CLEAN,
675         PAGEREF_KEEP,
676         PAGEREF_ACTIVATE,
677 };
678
679 static enum page_references page_check_references(struct page *page,
680                                                   struct scan_control *sc)
681 {
682         int referenced_ptes, referenced_page;
683         unsigned long vm_flags;
684
685         referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
686         referenced_page = TestClearPageReferenced(page);
687
688         /* Lumpy reclaim - ignore references */
689         if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
690                 return PAGEREF_RECLAIM;
691
692         /*
693          * Mlock lost the isolation race with us.  Let try_to_unmap()
694          * move the page to the unevictable list.
695          */
696         if (vm_flags & VM_LOCKED)
697                 return PAGEREF_RECLAIM;
698
699         if (referenced_ptes) {
700                 if (PageSwapBacked(page))
701                         return PAGEREF_ACTIVATE;
702                 /*
703                  * All mapped pages start out with page table
704                  * references from the instantiating fault, so we need
705                  * to look twice if a mapped file page is used more
706                  * than once.
707                  *
708                  * Mark it and spare it for another trip around the
709                  * inactive list.  Another page table reference will
710                  * lead to its activation.
711                  *
712                  * Note: the mark is set for activated pages as well
713                  * so that recently deactivated but used pages are
714                  * quickly recovered.
715                  */
716                 SetPageReferenced(page);
717
718                 if (referenced_page || referenced_ptes > 1)
719                         return PAGEREF_ACTIVATE;
720
721                 /*
722                  * Activate file-backed executable pages after first usage.
723                  */
724                 if (vm_flags & VM_EXEC)
725                         return PAGEREF_ACTIVATE;
726
727                 return PAGEREF_KEEP;
728         }
729
730         /* Reclaim if clean, defer dirty pages to writeback */
731         if (referenced_page && !PageSwapBacked(page))
732                 return PAGEREF_RECLAIM_CLEAN;
733
734         return PAGEREF_RECLAIM;
735 }
736
737 /*
738  * shrink_page_list() returns the number of reclaimed pages
739  */
740 static unsigned long shrink_page_list(struct list_head *page_list,
741                                       struct zone *zone,
742                                       struct scan_control *sc,
743                                       int priority,
744                                       unsigned long *ret_nr_dirty,
745                                       unsigned long *ret_nr_writeback)
746 {
747         LIST_HEAD(ret_pages);
748         LIST_HEAD(free_pages);
749         int pgactivate = 0;
750         unsigned long nr_dirty = 0;
751         unsigned long nr_congested = 0;
752         unsigned long nr_reclaimed = 0;
753         unsigned long nr_writeback = 0;
754
755         cond_resched();
756
757         while (!list_empty(page_list)) {
758                 enum page_references references;
759                 struct address_space *mapping;
760                 struct page *page;
761                 int may_enter_fs;
762
763                 cond_resched();
764
765                 page = lru_to_page(page_list);
766                 list_del(&page->lru);
767
768                 if (!trylock_page(page))
769                         goto keep;
770
771                 VM_BUG_ON(PageActive(page));
772                 VM_BUG_ON(page_zone(page) != zone);
773
774                 sc->nr_scanned++;
775
776                 if (unlikely(!page_evictable(page, NULL)))
777                         goto cull_mlocked;
778
779                 if (!sc->may_unmap && page_mapped(page))
780                         goto keep_locked;
781
782                 /* Double the slab pressure for mapped and swapcache pages */
783                 if (page_mapped(page) || PageSwapCache(page))
784                         sc->nr_scanned++;
785
786                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
787                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
788
789                 if (PageWriteback(page)) {
790                         nr_writeback++;
791                         /*
792                          * Synchronous reclaim cannot queue pages for
793                          * writeback due to the possibility of stack overflow
794                          * but if it encounters a page under writeback, wait
795                          * for the IO to complete.
796                          */
797                         if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
798                             may_enter_fs)
799                                 wait_on_page_writeback(page);
800                         else {
801                                 unlock_page(page);
802                                 goto keep_lumpy;
803                         }
804                 }
805
806                 references = page_check_references(page, sc);
807                 switch (references) {
808                 case PAGEREF_ACTIVATE:
809                         goto activate_locked;
810                 case PAGEREF_KEEP:
811                         goto keep_locked;
812                 case PAGEREF_RECLAIM:
813                 case PAGEREF_RECLAIM_CLEAN:
814                         ; /* try to reclaim the page below */
815                 }
816
817                 /*
818                  * Anonymous process memory has backing store?
819                  * Try to allocate it some swap space here.
820                  */
821                 if (PageAnon(page) && !PageSwapCache(page)) {
822                         if (!(sc->gfp_mask & __GFP_IO))
823                                 goto keep_locked;
824                         if (!add_to_swap(page))
825                                 goto activate_locked;
826                         may_enter_fs = 1;
827                 }
828
829                 mapping = page_mapping(page);
830
831                 /*
832                  * The page is mapped into the page tables of one or more
833                  * processes. Try to unmap it here.
834                  */
835                 if (page_mapped(page) && mapping) {
836                         switch (try_to_unmap(page, TTU_UNMAP)) {
837                         case SWAP_FAIL:
838                                 goto activate_locked;
839                         case SWAP_AGAIN:
840                                 goto keep_locked;
841                         case SWAP_MLOCK:
842                                 goto cull_mlocked;
843                         case SWAP_SUCCESS:
844                                 ; /* try to free the page below */
845                         }
846                 }
847
848                 if (PageDirty(page)) {
849                         nr_dirty++;
850
851                         /*
852                          * Only kswapd can writeback filesystem pages to
853                          * avoid risk of stack overflow but do not writeback
854                          * unless under significant pressure.
855                          */
856                         if (page_is_file_cache(page) &&
857                                         (!current_is_kswapd() || priority >= DEF_PRIORITY - 2)) {
858                                 /*
859                                  * Immediately reclaim when written back.
860                                  * Similar in principal to deactivate_page()
861                                  * except we already have the page isolated
862                                  * and know it's dirty
863                                  */
864                                 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
865                                 SetPageReclaim(page);
866
867                                 goto keep_locked;
868                         }
869
870                         if (references == PAGEREF_RECLAIM_CLEAN)
871                                 goto keep_locked;
872                         if (!may_enter_fs)
873                                 goto keep_locked;
874                         if (!sc->may_writepage)
875                                 goto keep_locked;
876
877                         /* Page is dirty, try to write it out here */
878                         switch (pageout(page, mapping, sc)) {
879                         case PAGE_KEEP:
880                                 nr_congested++;
881                                 goto keep_locked;
882                         case PAGE_ACTIVATE:
883                                 goto activate_locked;
884                         case PAGE_SUCCESS:
885                                 if (PageWriteback(page))
886                                         goto keep_lumpy;
887                                 if (PageDirty(page))
888                                         goto keep;
889
890                                 /*
891                                  * A synchronous write - probably a ramdisk.  Go
892                                  * ahead and try to reclaim the page.
893                                  */
894                                 if (!trylock_page(page))
895                                         goto keep;
896                                 if (PageDirty(page) || PageWriteback(page))
897                                         goto keep_locked;
898                                 mapping = page_mapping(page);
899                         case PAGE_CLEAN:
900                                 ; /* try to free the page below */
901                         }
902                 }
903
904                 /*
905                  * If the page has buffers, try to free the buffer mappings
906                  * associated with this page. If we succeed we try to free
907                  * the page as well.
908                  *
909                  * We do this even if the page is PageDirty().
910                  * try_to_release_page() does not perform I/O, but it is
911                  * possible for a page to have PageDirty set, but it is actually
912                  * clean (all its buffers are clean).  This happens if the
913                  * buffers were written out directly, with submit_bh(). ext3
914                  * will do this, as well as the blockdev mapping.
915                  * try_to_release_page() will discover that cleanness and will
916                  * drop the buffers and mark the page clean - it can be freed.
917                  *
918                  * Rarely, pages can have buffers and no ->mapping.  These are
919                  * the pages which were not successfully invalidated in
920                  * truncate_complete_page().  We try to drop those buffers here
921                  * and if that worked, and the page is no longer mapped into
922                  * process address space (page_count == 1) it can be freed.
923                  * Otherwise, leave the page on the LRU so it is swappable.
924                  */
925                 if (page_has_private(page)) {
926                         if (!try_to_release_page(page, sc->gfp_mask))
927                                 goto activate_locked;
928                         if (!mapping && page_count(page) == 1) {
929                                 unlock_page(page);
930                                 if (put_page_testzero(page))
931                                         goto free_it;
932                                 else {
933                                         /*
934                                          * rare race with speculative reference.
935                                          * the speculative reference will free
936                                          * this page shortly, so we may
937                                          * increment nr_reclaimed here (and
938                                          * leave it off the LRU).
939                                          */
940                                         nr_reclaimed++;
941                                         continue;
942                                 }
943                         }
944                 }
945
946                 if (!mapping || !__remove_mapping(mapping, page))
947                         goto keep_locked;
948
949                 /*
950                  * At this point, we have no other references and there is
951                  * no way to pick any more up (removed from LRU, removed
952                  * from pagecache). Can use non-atomic bitops now (and
953                  * we obviously don't have to worry about waking up a process
954                  * waiting on the page lock, because there are no references.
955                  */
956                 __clear_page_locked(page);
957 free_it:
958                 nr_reclaimed++;
959
960                 /*
961                  * Is there need to periodically free_page_list? It would
962                  * appear not as the counts should be low
963                  */
964                 list_add(&page->lru, &free_pages);
965                 continue;
966
967 cull_mlocked:
968                 if (PageSwapCache(page))
969                         try_to_free_swap(page);
970                 unlock_page(page);
971                 putback_lru_page(page);
972                 reset_reclaim_mode(sc);
973                 continue;
974
975 activate_locked:
976                 /* Not a candidate for swapping, so reclaim swap space. */
977                 if (PageSwapCache(page) && vm_swap_full())
978                         try_to_free_swap(page);
979                 VM_BUG_ON(PageActive(page));
980                 SetPageActive(page);
981                 pgactivate++;
982 keep_locked:
983                 unlock_page(page);
984 keep:
985                 reset_reclaim_mode(sc);
986 keep_lumpy:
987                 list_add(&page->lru, &ret_pages);
988                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
989         }
990
991         /*
992          * Tag a zone as congested if all the dirty pages encountered were
993          * backed by a congested BDI. In this case, reclaimers should just
994          * back off and wait for congestion to clear because further reclaim
995          * will encounter the same problem
996          */
997         if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc))
998                 zone_set_flag(zone, ZONE_CONGESTED);
999
1000         free_hot_cold_page_list(&free_pages, 1);
1001
1002         list_splice(&ret_pages, page_list);
1003         count_vm_events(PGACTIVATE, pgactivate);
1004         *ret_nr_dirty += nr_dirty;
1005         *ret_nr_writeback += nr_writeback;
1006         return nr_reclaimed;
1007 }
1008
1009 /*
1010  * Attempt to remove the specified page from its LRU.  Only take this page
1011  * if it is of the appropriate PageActive status.  Pages which are being
1012  * freed elsewhere are also ignored.
1013  *
1014  * page:        page to consider
1015  * mode:        one of the LRU isolation modes defined above
1016  *
1017  * returns 0 on success, -ve errno on failure.
1018  */
1019 int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file)
1020 {
1021         bool all_lru_mode;
1022         int ret = -EINVAL;
1023
1024         /* Only take pages on the LRU. */
1025         if (!PageLRU(page))
1026                 return ret;
1027
1028         all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) ==
1029                 (ISOLATE_ACTIVE|ISOLATE_INACTIVE);
1030
1031         /*
1032          * When checking the active state, we need to be sure we are
1033          * dealing with comparible boolean values.  Take the logical not
1034          * of each.
1035          */
1036         if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE))
1037                 return ret;
1038
1039         if (!all_lru_mode && !!page_is_file_cache(page) != file)
1040                 return ret;
1041
1042         /*
1043          * When this function is being called for lumpy reclaim, we
1044          * initially look into all LRU pages, active, inactive and
1045          * unevictable; only give shrink_page_list evictable pages.
1046          */
1047         if (PageUnevictable(page))
1048                 return ret;
1049
1050         ret = -EBUSY;
1051
1052         /*
1053          * To minimise LRU disruption, the caller can indicate that it only
1054          * wants to isolate pages it will be able to operate on without
1055          * blocking - clean pages for the most part.
1056          *
1057          * ISOLATE_CLEAN means that only clean pages should be isolated. This
1058          * is used by reclaim when it is cannot write to backing storage
1059          *
1060          * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1061          * that it is possible to migrate without blocking
1062          */
1063         if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1064                 /* All the caller can do on PageWriteback is block */
1065                 if (PageWriteback(page))
1066                         return ret;
1067
1068                 if (PageDirty(page)) {
1069                         struct address_space *mapping;
1070
1071                         /* ISOLATE_CLEAN means only clean pages */
1072                         if (mode & ISOLATE_CLEAN)
1073                                 return ret;
1074
1075                         /*
1076                          * Only pages without mappings or that have a
1077                          * ->migratepage callback are possible to migrate
1078                          * without blocking
1079                          */
1080                         mapping = page_mapping(page);
1081                         if (mapping && !mapping->a_ops->migratepage)
1082                                 return ret;
1083                 }
1084         }
1085
1086         if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1087                 return ret;
1088
1089         if (likely(get_page_unless_zero(page))) {
1090                 /*
1091                  * Be careful not to clear PageLRU until after we're
1092                  * sure the page is not being freed elsewhere -- the
1093                  * page release code relies on it.
1094                  */
1095                 ClearPageLRU(page);
1096                 ret = 0;
1097         }
1098
1099         return ret;
1100 }
1101
1102 /*
1103  * zone->lru_lock is heavily contended.  Some of the functions that
1104  * shrink the lists perform better by taking out a batch of pages
1105  * and working on them outside the LRU lock.
1106  *
1107  * For pagecache intensive workloads, this function is the hottest
1108  * spot in the kernel (apart from copy_*_user functions).
1109  *
1110  * Appropriate locks must be held before calling this function.
1111  *
1112  * @nr_to_scan: The number of pages to look through on the list.
1113  * @src:        The LRU list to pull pages off.
1114  * @dst:        The temp list to put pages on to.
1115  * @scanned:    The number of pages that were scanned.
1116  * @order:      The caller's attempted allocation order
1117  * @mode:       One of the LRU isolation modes
1118  * @file:       True [1] if isolating file [!anon] pages
1119  *
1120  * returns how many pages were moved onto *@dst.
1121  */
1122 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1123                 struct list_head *src, struct list_head *dst,
1124                 unsigned long *scanned, int order, isolate_mode_t mode,
1125                 int file)
1126 {
1127         unsigned long nr_taken = 0;
1128         unsigned long nr_lumpy_taken = 0;
1129         unsigned long nr_lumpy_dirty = 0;
1130         unsigned long nr_lumpy_failed = 0;
1131         unsigned long scan;
1132
1133         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1134                 struct page *page;
1135                 unsigned long pfn;
1136                 unsigned long end_pfn;
1137                 unsigned long page_pfn;
1138                 int zone_id;
1139
1140                 page = lru_to_page(src);
1141                 prefetchw_prev_lru_page(page, src, flags);
1142
1143                 VM_BUG_ON(!PageLRU(page));
1144
1145                 switch (__isolate_lru_page(page, mode, file)) {
1146                 case 0:
1147                         list_move(&page->lru, dst);
1148                         mem_cgroup_del_lru(page);
1149                         nr_taken += hpage_nr_pages(page);
1150                         break;
1151
1152                 case -EBUSY:
1153                         /* else it is being freed elsewhere */
1154                         list_move(&page->lru, src);
1155                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1156                         continue;
1157
1158                 default:
1159                         BUG();
1160                 }
1161
1162                 if (!order)
1163                         continue;
1164
1165                 /*
1166                  * Attempt to take all pages in the order aligned region
1167                  * surrounding the tag page.  Only take those pages of
1168                  * the same active state as that tag page.  We may safely
1169                  * round the target page pfn down to the requested order
1170                  * as the mem_map is guaranteed valid out to MAX_ORDER,
1171                  * where that page is in a different zone we will detect
1172                  * it from its zone id and abort this block scan.
1173                  */
1174                 zone_id = page_zone_id(page);
1175                 page_pfn = page_to_pfn(page);
1176                 pfn = page_pfn & ~((1 << order) - 1);
1177                 end_pfn = pfn + (1 << order);
1178                 for (; pfn < end_pfn; pfn++) {
1179                         struct page *cursor_page;
1180
1181                         /* The target page is in the block, ignore it. */
1182                         if (unlikely(pfn == page_pfn))
1183                                 continue;
1184
1185                         /* Avoid holes within the zone. */
1186                         if (unlikely(!pfn_valid_within(pfn)))
1187                                 break;
1188
1189                         cursor_page = pfn_to_page(pfn);
1190
1191                         /* Check that we have not crossed a zone boundary. */
1192                         if (unlikely(page_zone_id(cursor_page) != zone_id))
1193                                 break;
1194
1195                         /*
1196                          * If we don't have enough swap space, reclaiming of
1197                          * anon page which don't already have a swap slot is
1198                          * pointless.
1199                          */
1200                         if (nr_swap_pages <= 0 && PageSwapBacked(cursor_page) &&
1201                             !PageSwapCache(cursor_page))
1202                                 break;
1203
1204                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1205                                 list_move(&cursor_page->lru, dst);
1206                                 mem_cgroup_del_lru(cursor_page);
1207                                 nr_taken += hpage_nr_pages(page);
1208                                 nr_lumpy_taken++;
1209                                 if (PageDirty(cursor_page))
1210                                         nr_lumpy_dirty++;
1211                                 scan++;
1212                         } else {
1213                                 /*
1214                                  * Check if the page is freed already.
1215                                  *
1216                                  * We can't use page_count() as that
1217                                  * requires compound_head and we don't
1218                                  * have a pin on the page here. If a
1219                                  * page is tail, we may or may not
1220                                  * have isolated the head, so assume
1221                                  * it's not free, it'd be tricky to
1222                                  * track the head status without a
1223                                  * page pin.
1224                                  */
1225                                 if (!PageTail(cursor_page) &&
1226                                     !atomic_read(&cursor_page->_count))
1227                                         continue;
1228                                 break;
1229                         }
1230                 }
1231
1232                 /* If we break out of the loop above, lumpy reclaim failed */
1233                 if (pfn < end_pfn)
1234                         nr_lumpy_failed++;
1235         }
1236
1237         *scanned = scan;
1238
1239         trace_mm_vmscan_lru_isolate(order,
1240                         nr_to_scan, scan,
1241                         nr_taken,
1242                         nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1243                         mode);
1244         return nr_taken;
1245 }
1246
1247 static unsigned long isolate_pages_global(unsigned long nr,
1248                                         struct list_head *dst,
1249                                         unsigned long *scanned, int order,
1250                                         isolate_mode_t mode,
1251                                         struct zone *z, int active, int file)
1252 {
1253         int lru = LRU_BASE;
1254         if (active)
1255                 lru += LRU_ACTIVE;
1256         if (file)
1257                 lru += LRU_FILE;
1258         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1259                                                                 mode, file);
1260 }
1261
1262 /*
1263  * clear_active_flags() is a helper for shrink_active_list(), clearing
1264  * any active bits from the pages in the list.
1265  */
1266 static unsigned long clear_active_flags(struct list_head *page_list,
1267                                         unsigned int *count)
1268 {
1269         int nr_active = 0;
1270         int lru;
1271         struct page *page;
1272
1273         list_for_each_entry(page, page_list, lru) {
1274                 int numpages = hpage_nr_pages(page);
1275                 lru = page_lru_base_type(page);
1276                 if (PageActive(page)) {
1277                         lru += LRU_ACTIVE;
1278                         ClearPageActive(page);
1279                         nr_active += numpages;
1280                 }
1281                 if (count)
1282                         count[lru] += numpages;
1283         }
1284
1285         return nr_active;
1286 }
1287
1288 /**
1289  * isolate_lru_page - tries to isolate a page from its LRU list
1290  * @page: page to isolate from its LRU list
1291  *
1292  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1293  * vmstat statistic corresponding to whatever LRU list the page was on.
1294  *
1295  * Returns 0 if the page was removed from an LRU list.
1296  * Returns -EBUSY if the page was not on an LRU list.
1297  *
1298  * The returned page will have PageLRU() cleared.  If it was found on
1299  * the active list, it will have PageActive set.  If it was found on
1300  * the unevictable list, it will have the PageUnevictable bit set. That flag
1301  * may need to be cleared by the caller before letting the page go.
1302  *
1303  * The vmstat statistic corresponding to the list on which the page was
1304  * found will be decremented.
1305  *
1306  * Restrictions:
1307  * (1) Must be called with an elevated refcount on the page. This is a
1308  *     fundamentnal difference from isolate_lru_pages (which is called
1309  *     without a stable reference).
1310  * (2) the lru_lock must not be held.
1311  * (3) interrupts must be enabled.
1312  */
1313 int isolate_lru_page(struct page *page)
1314 {
1315         int ret = -EBUSY;
1316
1317         VM_BUG_ON(!page_count(page));
1318
1319         if (PageLRU(page)) {
1320                 struct zone *zone = page_zone(page);
1321
1322                 spin_lock_irq(&zone->lru_lock);
1323                 if (PageLRU(page)) {
1324                         int lru = page_lru(page);
1325                         ret = 0;
1326                         get_page(page);
1327                         ClearPageLRU(page);
1328
1329                         del_page_from_lru_list(zone, page, lru);
1330                 }
1331                 spin_unlock_irq(&zone->lru_lock);
1332         }
1333         return ret;
1334 }
1335
1336 /*
1337  * Are there way too many processes in the direct reclaim path already?
1338  */
1339 static int too_many_isolated(struct zone *zone, int file,
1340                 struct scan_control *sc)
1341 {
1342         unsigned long inactive, isolated;
1343
1344         if (current_is_kswapd())
1345                 return 0;
1346
1347         if (!scanning_global_lru(sc))
1348                 return 0;
1349
1350         if (file) {
1351                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1352                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1353         } else {
1354                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1355                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1356         }
1357
1358         return isolated > inactive;
1359 }
1360
1361 /*
1362  * TODO: Try merging with migrations version of putback_lru_pages
1363  */
1364 static noinline_for_stack void
1365 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1366                                 unsigned long nr_anon, unsigned long nr_file,
1367                                 struct list_head *page_list)
1368 {
1369         struct page *page;
1370         struct pagevec pvec;
1371         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1372
1373         pagevec_init(&pvec, 1);
1374
1375         /*
1376          * Put back any unfreeable pages.
1377          */
1378         spin_lock(&zone->lru_lock);
1379         while (!list_empty(page_list)) {
1380                 int lru;
1381                 page = lru_to_page(page_list);
1382                 VM_BUG_ON(PageLRU(page));
1383                 list_del(&page->lru);
1384                 if (unlikely(!page_evictable(page, NULL))) {
1385                         spin_unlock_irq(&zone->lru_lock);
1386                         putback_lru_page(page);
1387                         spin_lock_irq(&zone->lru_lock);
1388                         continue;
1389                 }
1390                 SetPageLRU(page);
1391                 lru = page_lru(page);
1392                 add_page_to_lru_list(zone, page, lru);
1393                 if (is_active_lru(lru)) {
1394                         int file = is_file_lru(lru);
1395                         int numpages = hpage_nr_pages(page);
1396                         reclaim_stat->recent_rotated[file] += numpages;
1397                 }
1398                 if (!pagevec_add(&pvec, page)) {
1399                         spin_unlock_irq(&zone->lru_lock);
1400                         __pagevec_release(&pvec);
1401                         spin_lock_irq(&zone->lru_lock);
1402                 }
1403         }
1404         __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1405         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1406
1407         spin_unlock_irq(&zone->lru_lock);
1408         pagevec_release(&pvec);
1409 }
1410
1411 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1412                                         struct scan_control *sc,
1413                                         unsigned long *nr_anon,
1414                                         unsigned long *nr_file,
1415                                         struct list_head *isolated_list)
1416 {
1417         unsigned long nr_active;
1418         unsigned int count[NR_LRU_LISTS] = { 0, };
1419         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1420
1421         nr_active = clear_active_flags(isolated_list, count);
1422         __count_vm_events(PGDEACTIVATE, nr_active);
1423
1424         __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1425                               -count[LRU_ACTIVE_FILE]);
1426         __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1427                               -count[LRU_INACTIVE_FILE]);
1428         __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1429                               -count[LRU_ACTIVE_ANON]);
1430         __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1431                               -count[LRU_INACTIVE_ANON]);
1432
1433         *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1434         *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1435         __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1436         __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1437
1438         reclaim_stat->recent_scanned[0] += *nr_anon;
1439         reclaim_stat->recent_scanned[1] += *nr_file;
1440 }
1441
1442 /*
1443  * Returns true if a direct reclaim should wait on pages under writeback.
1444  *
1445  * If we are direct reclaiming for contiguous pages and we do not reclaim
1446  * everything in the list, try again and wait for writeback IO to complete.
1447  * This will stall high-order allocations noticeably. Only do that when really
1448  * need to free the pages under high memory pressure.
1449  */
1450 static inline bool should_reclaim_stall(unsigned long nr_taken,
1451                                         unsigned long nr_freed,
1452                                         int priority,
1453                                         struct scan_control *sc)
1454 {
1455         int lumpy_stall_priority;
1456
1457         /* kswapd should not stall on sync IO */
1458         if (current_is_kswapd())
1459                 return false;
1460
1461         /* Only stall on lumpy reclaim */
1462         if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1463                 return false;
1464
1465         /* If we have reclaimed everything on the isolated list, no stall */
1466         if (nr_freed == nr_taken)
1467                 return false;
1468
1469         /*
1470          * For high-order allocations, there are two stall thresholds.
1471          * High-cost allocations stall immediately where as lower
1472          * order allocations such as stacks require the scanning
1473          * priority to be much higher before stalling.
1474          */
1475         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1476                 lumpy_stall_priority = DEF_PRIORITY;
1477         else
1478                 lumpy_stall_priority = DEF_PRIORITY / 3;
1479
1480         return priority <= lumpy_stall_priority;
1481 }
1482
1483 /*
1484  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1485  * of reclaimed pages
1486  */
1487 static noinline_for_stack unsigned long
1488 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1489                         struct scan_control *sc, int priority, int file)
1490 {
1491         LIST_HEAD(page_list);
1492         unsigned long nr_scanned;
1493         unsigned long nr_reclaimed = 0;
1494         unsigned long nr_taken;
1495         unsigned long nr_anon;
1496         unsigned long nr_file;
1497         unsigned long nr_dirty = 0;
1498         unsigned long nr_writeback = 0;
1499         isolate_mode_t reclaim_mode = ISOLATE_INACTIVE;
1500
1501         while (unlikely(too_many_isolated(zone, file, sc))) {
1502                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1503
1504                 /* We are about to die and free our memory. Return now. */
1505                 if (fatal_signal_pending(current))
1506                         return SWAP_CLUSTER_MAX;
1507         }
1508
1509         set_reclaim_mode(priority, sc, false);
1510         if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
1511                 reclaim_mode |= ISOLATE_ACTIVE;
1512
1513         lru_add_drain();
1514
1515         if (!sc->may_unmap)
1516                 reclaim_mode |= ISOLATE_UNMAPPED;
1517         if (!sc->may_writepage)
1518                 reclaim_mode |= ISOLATE_CLEAN;
1519
1520         spin_lock_irq(&zone->lru_lock);
1521
1522         if (scanning_global_lru(sc)) {
1523                 nr_taken = isolate_pages_global(nr_to_scan, &page_list,
1524                         &nr_scanned, sc->order, reclaim_mode, zone, 0, file);
1525                 zone->pages_scanned += nr_scanned;
1526                 if (current_is_kswapd())
1527                         __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1528                                                nr_scanned);
1529                 else
1530                         __count_zone_vm_events(PGSCAN_DIRECT, zone,
1531                                                nr_scanned);
1532         } else {
1533                 nr_taken = mem_cgroup_isolate_pages(nr_to_scan, &page_list,
1534                         &nr_scanned, sc->order, reclaim_mode, zone,
1535                         sc->mem_cgroup, 0, file);
1536                 /*
1537                  * mem_cgroup_isolate_pages() keeps track of
1538                  * scanned pages on its own.
1539                  */
1540         }
1541
1542         if (nr_taken == 0) {
1543                 spin_unlock_irq(&zone->lru_lock);
1544                 return 0;
1545         }
1546
1547         update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1548
1549         spin_unlock_irq(&zone->lru_lock);
1550
1551         nr_reclaimed = shrink_page_list(&page_list, zone, sc, priority,
1552                                                 &nr_dirty, &nr_writeback);
1553
1554         /* Check if we should syncronously wait for writeback */
1555         if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1556                 set_reclaim_mode(priority, sc, true);
1557                 nr_reclaimed += shrink_page_list(&page_list, zone, sc,
1558                                         priority, &nr_dirty, &nr_writeback);
1559         }
1560
1561         local_irq_disable();
1562         if (current_is_kswapd())
1563                 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1564         __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1565
1566         putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1567
1568         /*
1569          * If reclaim is isolating dirty pages under writeback, it implies
1570          * that the long-lived page allocation rate is exceeding the page
1571          * laundering rate. Either the global limits are not being effective
1572          * at throttling processes due to the page distribution throughout
1573          * zones or there is heavy usage of a slow backing device. The
1574          * only option is to throttle from reclaim context which is not ideal
1575          * as there is no guarantee the dirtying process is throttled in the
1576          * same way balance_dirty_pages() manages.
1577          *
1578          * This scales the number of dirty pages that must be under writeback
1579          * before throttling depending on priority. It is a simple backoff
1580          * function that has the most effect in the range DEF_PRIORITY to
1581          * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1582          * in trouble and reclaim is considered to be in trouble.
1583          *
1584          * DEF_PRIORITY   100% isolated pages must be PageWriteback to throttle
1585          * DEF_PRIORITY-1  50% must be PageWriteback
1586          * DEF_PRIORITY-2  25% must be PageWriteback, kswapd in trouble
1587          * ...
1588          * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1589          *                     isolated page is PageWriteback
1590          */
1591         if (nr_writeback && nr_writeback >= (nr_taken >> (DEF_PRIORITY-priority)))
1592                 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1593
1594         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1595                 zone_idx(zone),
1596                 nr_scanned, nr_reclaimed,
1597                 priority,
1598                 trace_shrink_flags(file, sc->reclaim_mode));
1599         return nr_reclaimed;
1600 }
1601
1602 /*
1603  * This moves pages from the active list to the inactive list.
1604  *
1605  * We move them the other way if the page is referenced by one or more
1606  * processes, from rmap.
1607  *
1608  * If the pages are mostly unmapped, the processing is fast and it is
1609  * appropriate to hold zone->lru_lock across the whole operation.  But if
1610  * the pages are mapped, the processing is slow (page_referenced()) so we
1611  * should drop zone->lru_lock around each page.  It's impossible to balance
1612  * this, so instead we remove the pages from the LRU while processing them.
1613  * It is safe to rely on PG_active against the non-LRU pages in here because
1614  * nobody will play with that bit on a non-LRU page.
1615  *
1616  * The downside is that we have to touch page->_count against each page.
1617  * But we had to alter page->flags anyway.
1618  */
1619
1620 static void move_active_pages_to_lru(struct zone *zone,
1621                                      struct list_head *list,
1622                                      enum lru_list lru)
1623 {
1624         unsigned long pgmoved = 0;
1625         struct pagevec pvec;
1626         struct page *page;
1627
1628         pagevec_init(&pvec, 1);
1629
1630         while (!list_empty(list)) {
1631                 page = lru_to_page(list);
1632
1633                 VM_BUG_ON(PageLRU(page));
1634                 SetPageLRU(page);
1635
1636                 list_move(&page->lru, &zone->lru[lru].list);
1637                 mem_cgroup_add_lru_list(page, lru);
1638                 pgmoved += hpage_nr_pages(page);
1639
1640                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1641                         spin_unlock_irq(&zone->lru_lock);
1642                         if (buffer_heads_over_limit)
1643                                 pagevec_strip(&pvec);
1644                         __pagevec_release(&pvec);
1645                         spin_lock_irq(&zone->lru_lock);
1646                 }
1647         }
1648         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1649         if (!is_active_lru(lru))
1650                 __count_vm_events(PGDEACTIVATE, pgmoved);
1651 }
1652
1653 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1654                         struct scan_control *sc, int priority, int file)
1655 {
1656         unsigned long nr_taken;
1657         unsigned long pgscanned;
1658         unsigned long vm_flags;
1659         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1660         LIST_HEAD(l_active);
1661         LIST_HEAD(l_inactive);
1662         struct page *page;
1663         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1664         unsigned long nr_rotated = 0;
1665         isolate_mode_t reclaim_mode = ISOLATE_ACTIVE;
1666
1667         lru_add_drain();
1668
1669         if (!sc->may_unmap)
1670                 reclaim_mode |= ISOLATE_UNMAPPED;
1671         if (!sc->may_writepage)
1672                 reclaim_mode |= ISOLATE_CLEAN;
1673
1674         spin_lock_irq(&zone->lru_lock);
1675         if (scanning_global_lru(sc)) {
1676                 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1677                                                 &pgscanned, sc->order,
1678                                                 reclaim_mode, zone,
1679                                                 1, file);
1680                 zone->pages_scanned += pgscanned;
1681         } else {
1682                 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1683                                                 &pgscanned, sc->order,
1684                                                 reclaim_mode, zone,
1685                                                 sc->mem_cgroup, 1, file);
1686                 /*
1687                  * mem_cgroup_isolate_pages() keeps track of
1688                  * scanned pages on its own.
1689                  */
1690         }
1691
1692         reclaim_stat->recent_scanned[file] += nr_taken;
1693
1694         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1695         if (file)
1696                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1697         else
1698                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1699         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1700         spin_unlock_irq(&zone->lru_lock);
1701
1702         while (!list_empty(&l_hold)) {
1703                 cond_resched();
1704                 page = lru_to_page(&l_hold);
1705                 list_del(&page->lru);
1706
1707                 if (unlikely(!page_evictable(page, NULL))) {
1708                         putback_lru_page(page);
1709                         continue;
1710                 }
1711
1712                 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1713                         nr_rotated += hpage_nr_pages(page);
1714                         /*
1715                          * Identify referenced, file-backed active pages and
1716                          * give them one more trip around the active list. So
1717                          * that executable code get better chances to stay in
1718                          * memory under moderate memory pressure.  Anon pages
1719                          * are not likely to be evicted by use-once streaming
1720                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1721                          * so we ignore them here.
1722                          */
1723                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1724                                 list_add(&page->lru, &l_active);
1725                                 continue;
1726                         }
1727                 }
1728
1729                 ClearPageActive(page);  /* we are de-activating */
1730                 list_add(&page->lru, &l_inactive);
1731         }
1732
1733         /*
1734          * Move pages back to the lru list.
1735          */
1736         spin_lock_irq(&zone->lru_lock);
1737         /*
1738          * Count referenced pages from currently used mappings as rotated,
1739          * even though only some of them are actually re-activated.  This
1740          * helps balance scan pressure between file and anonymous pages in
1741          * get_scan_ratio.
1742          */
1743         reclaim_stat->recent_rotated[file] += nr_rotated;
1744
1745         move_active_pages_to_lru(zone, &l_active,
1746                                                 LRU_ACTIVE + file * LRU_FILE);
1747         move_active_pages_to_lru(zone, &l_inactive,
1748                                                 LRU_BASE   + file * LRU_FILE);
1749         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1750         spin_unlock_irq(&zone->lru_lock);
1751 }
1752
1753 #ifdef CONFIG_SWAP
1754 static int inactive_anon_is_low_global(struct zone *zone)
1755 {
1756         unsigned long active, inactive;
1757
1758         active = zone_page_state(zone, NR_ACTIVE_ANON);
1759         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1760
1761         if (inactive * zone->inactive_ratio < active)
1762                 return 1;
1763
1764         return 0;
1765 }
1766
1767 /**
1768  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1769  * @zone: zone to check
1770  * @sc:   scan control of this context
1771  *
1772  * Returns true if the zone does not have enough inactive anon pages,
1773  * meaning some active anon pages need to be deactivated.
1774  */
1775 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1776 {
1777         int low;
1778
1779         /*
1780          * If we don't have swap space, anonymous page deactivation
1781          * is pointless.
1782          */
1783         if (!total_swap_pages)
1784                 return 0;
1785
1786         if (scanning_global_lru(sc))
1787                 low = inactive_anon_is_low_global(zone);
1788         else
1789                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup, zone);
1790         return low;
1791 }
1792 #else
1793 static inline int inactive_anon_is_low(struct zone *zone,
1794                                         struct scan_control *sc)
1795 {
1796         return 0;
1797 }
1798 #endif
1799
1800 static int inactive_file_is_low_global(struct zone *zone)
1801 {
1802         unsigned long active, inactive;
1803
1804         active = zone_page_state(zone, NR_ACTIVE_FILE);
1805         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1806
1807         return (active > inactive);
1808 }
1809
1810 /**
1811  * inactive_file_is_low - check if file pages need to be deactivated
1812  * @zone: zone to check
1813  * @sc:   scan control of this context
1814  *
1815  * When the system is doing streaming IO, memory pressure here
1816  * ensures that active file pages get deactivated, until more
1817  * than half of the file pages are on the inactive list.
1818  *
1819  * Once we get to that situation, protect the system's working
1820  * set from being evicted by disabling active file page aging.
1821  *
1822  * This uses a different ratio than the anonymous pages, because
1823  * the page cache uses a use-once replacement algorithm.
1824  */
1825 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1826 {
1827         int low;
1828
1829         if (scanning_global_lru(sc))
1830                 low = inactive_file_is_low_global(zone);
1831         else
1832                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup, zone);
1833         return low;
1834 }
1835
1836 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1837                                 int file)
1838 {
1839         if (file)
1840                 return inactive_file_is_low(zone, sc);
1841         else
1842                 return inactive_anon_is_low(zone, sc);
1843 }
1844
1845 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1846         struct zone *zone, struct scan_control *sc, int priority)
1847 {
1848         int file = is_file_lru(lru);
1849
1850         if (is_active_lru(lru)) {
1851                 if (inactive_list_is_low(zone, sc, file))
1852                     shrink_active_list(nr_to_scan, zone, sc, priority, file);
1853                 return 0;
1854         }
1855
1856         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1857 }
1858
1859 static int vmscan_swappiness(struct scan_control *sc)
1860 {
1861         if (scanning_global_lru(sc))
1862                 return vm_swappiness;
1863         return mem_cgroup_swappiness(sc->mem_cgroup);
1864 }
1865
1866 /*
1867  * Determine how aggressively the anon and file LRU lists should be
1868  * scanned.  The relative value of each set of LRU lists is determined
1869  * by looking at the fraction of the pages scanned we did rotate back
1870  * onto the active list instead of evict.
1871  *
1872  * nr[0] = anon pages to scan; nr[1] = file pages to scan
1873  */
1874 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1875                                         unsigned long *nr, int priority)
1876 {
1877         unsigned long anon, file, free;
1878         unsigned long anon_prio, file_prio;
1879         unsigned long ap, fp;
1880         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1881         u64 fraction[2], denominator;
1882         enum lru_list l;
1883         int noswap = 0;
1884         bool force_scan = false;
1885
1886         /*
1887          * If the zone or memcg is small, nr[l] can be 0.  This
1888          * results in no scanning on this priority and a potential
1889          * priority drop.  Global direct reclaim can go to the next
1890          * zone and tends to have no problems. Global kswapd is for
1891          * zone balancing and it needs to scan a minimum amount. When
1892          * reclaiming for a memcg, a priority drop can cause high
1893          * latencies, so it's better to scan a minimum amount there as
1894          * well.
1895          */
1896         if (scanning_global_lru(sc) && current_is_kswapd() &&
1897             zone->all_unreclaimable)
1898                 force_scan = true;
1899         if (!scanning_global_lru(sc))
1900                 force_scan = true;
1901
1902         /* If we have no swap space, do not bother scanning anon pages. */
1903         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1904                 noswap = 1;
1905                 fraction[0] = 0;
1906                 fraction[1] = 1;
1907                 denominator = 1;
1908                 goto out;
1909         }
1910
1911         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1912                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1913         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1914                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1915
1916         if (scanning_global_lru(sc)) {
1917                 free  = zone_page_state(zone, NR_FREE_PAGES);
1918                 /* If we have very few page cache pages,
1919                    force-scan anon pages. */
1920                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1921                         fraction[0] = 1;
1922                         fraction[1] = 0;
1923                         denominator = 1;
1924                         goto out;
1925                 }
1926         }
1927
1928         /*
1929          * With swappiness at 100, anonymous and file have the same priority.
1930          * This scanning priority is essentially the inverse of IO cost.
1931          */
1932         anon_prio = vmscan_swappiness(sc);
1933         file_prio = 200 - vmscan_swappiness(sc);
1934
1935         /*
1936          * OK, so we have swap space and a fair amount of page cache
1937          * pages.  We use the recently rotated / recently scanned
1938          * ratios to determine how valuable each cache is.
1939          *
1940          * Because workloads change over time (and to avoid overflow)
1941          * we keep these statistics as a floating average, which ends
1942          * up weighing recent references more than old ones.
1943          *
1944          * anon in [0], file in [1]
1945          */
1946         spin_lock_irq(&zone->lru_lock);
1947         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1948                 reclaim_stat->recent_scanned[0] /= 2;
1949                 reclaim_stat->recent_rotated[0] /= 2;
1950         }
1951
1952         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1953                 reclaim_stat->recent_scanned[1] /= 2;
1954                 reclaim_stat->recent_rotated[1] /= 2;
1955         }
1956
1957         /*
1958          * The amount of pressure on anon vs file pages is inversely
1959          * proportional to the fraction of recently scanned pages on
1960          * each list that were recently referenced and in active use.
1961          */
1962         ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
1963         ap /= reclaim_stat->recent_rotated[0] + 1;
1964
1965         fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
1966         fp /= reclaim_stat->recent_rotated[1] + 1;
1967         spin_unlock_irq(&zone->lru_lock);
1968
1969         fraction[0] = ap;
1970         fraction[1] = fp;
1971         denominator = ap + fp + 1;
1972 out:
1973         for_each_evictable_lru(l) {
1974                 int file = is_file_lru(l);
1975                 unsigned long scan;
1976
1977                 scan = zone_nr_lru_pages(zone, sc, l);
1978                 if (priority || noswap || !vmscan_swappiness(sc)) {
1979                         scan >>= priority;
1980                         if (!scan && force_scan)
1981                                 scan = SWAP_CLUSTER_MAX;
1982                         scan = div64_u64(scan * fraction[file], denominator);
1983                 }
1984                 nr[l] = scan;
1985         }
1986 }
1987
1988 /*
1989  * Reclaim/compaction depends on a number of pages being freed. To avoid
1990  * disruption to the system, a small number of order-0 pages continue to be
1991  * rotated and reclaimed in the normal fashion. However, by the time we get
1992  * back to the allocator and call try_to_compact_zone(), we ensure that
1993  * there are enough free pages for it to be likely successful
1994  */
1995 static inline bool should_continue_reclaim(struct zone *zone,
1996                                         unsigned long nr_reclaimed,
1997                                         unsigned long nr_scanned,
1998                                         struct scan_control *sc)
1999 {
2000         unsigned long pages_for_compaction;
2001         unsigned long inactive_lru_pages;
2002
2003         /* If not in reclaim/compaction mode, stop */
2004         if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
2005                 return false;
2006
2007         /* Consider stopping depending on scan and reclaim activity */
2008         if (sc->gfp_mask & __GFP_REPEAT) {
2009                 /*
2010                  * For __GFP_REPEAT allocations, stop reclaiming if the
2011                  * full LRU list has been scanned and we are still failing
2012                  * to reclaim pages. This full LRU scan is potentially
2013                  * expensive but a __GFP_REPEAT caller really wants to succeed
2014                  */
2015                 if (!nr_reclaimed && !nr_scanned)
2016                         return false;
2017         } else {
2018                 /*
2019                  * For non-__GFP_REPEAT allocations which can presumably
2020                  * fail without consequence, stop if we failed to reclaim
2021                  * any pages from the last SWAP_CLUSTER_MAX number of
2022                  * pages that were scanned. This will return to the
2023                  * caller faster at the risk reclaim/compaction and
2024                  * the resulting allocation attempt fails
2025                  */
2026                 if (!nr_reclaimed)
2027                         return false;
2028         }
2029
2030         /*
2031          * If we have not reclaimed enough pages for compaction and the
2032          * inactive lists are large enough, continue reclaiming
2033          */
2034         pages_for_compaction = (2UL << sc->order);
2035         inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
2036         if (nr_swap_pages > 0)
2037                 inactive_lru_pages += zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
2038         if (sc->nr_reclaimed < pages_for_compaction &&
2039                         inactive_lru_pages > pages_for_compaction)
2040                 return true;
2041
2042         /* If compaction would go ahead or the allocation would succeed, stop */
2043         switch (compaction_suitable(zone, sc->order)) {
2044         case COMPACT_PARTIAL:
2045         case COMPACT_CONTINUE:
2046                 return false;
2047         default:
2048                 return true;
2049         }
2050 }
2051
2052 /*
2053  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
2054  */
2055 static void shrink_zone(int priority, struct zone *zone,
2056                                 struct scan_control *sc)
2057 {
2058         unsigned long nr[NR_LRU_LISTS];
2059         unsigned long nr_to_scan;
2060         enum lru_list l;
2061         unsigned long nr_reclaimed, nr_scanned;
2062         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2063         struct blk_plug plug;
2064
2065 restart:
2066         nr_reclaimed = 0;
2067         nr_scanned = sc->nr_scanned;
2068         get_scan_count(zone, sc, nr, priority);
2069
2070         blk_start_plug(&plug);
2071         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2072                                         nr[LRU_INACTIVE_FILE]) {
2073                 for_each_evictable_lru(l) {
2074                         if (nr[l]) {
2075                                 nr_to_scan = min_t(unsigned long,
2076                                                    nr[l], SWAP_CLUSTER_MAX);
2077                                 nr[l] -= nr_to_scan;
2078
2079                                 nr_reclaimed += shrink_list(l, nr_to_scan,
2080                                                             zone, sc, priority);
2081                         }
2082                 }
2083                 /*
2084                  * On large memory systems, scan >> priority can become
2085                  * really large. This is fine for the starting priority;
2086                  * we want to put equal scanning pressure on each zone.
2087                  * However, if the VM has a harder time of freeing pages,
2088                  * with multiple processes reclaiming pages, the total
2089                  * freeing target can get unreasonably large.
2090                  */
2091                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
2092                         break;
2093         }
2094         blk_finish_plug(&plug);
2095         sc->nr_reclaimed += nr_reclaimed;
2096
2097         /*
2098          * Even if we did not try to evict anon pages at all, we want to
2099          * rebalance the anon lru active/inactive ratio.
2100          */
2101         if (inactive_anon_is_low(zone, sc))
2102                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
2103
2104         /* reclaim/compaction might need reclaim to continue */
2105         if (should_continue_reclaim(zone, nr_reclaimed,
2106                                         sc->nr_scanned - nr_scanned, sc))
2107                 goto restart;
2108
2109         throttle_vm_writeout(sc->gfp_mask);
2110 }
2111
2112 /* Returns true if compaction should go ahead for a high-order request */
2113 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2114 {
2115         unsigned long balance_gap, watermark;
2116         bool watermark_ok;
2117
2118         /* Do not consider compaction for orders reclaim is meant to satisfy */
2119         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2120                 return false;
2121
2122         /*
2123          * Compaction takes time to run and there are potentially other
2124          * callers using the pages just freed. Continue reclaiming until
2125          * there is a buffer of free pages available to give compaction
2126          * a reasonable chance of completing and allocating the page
2127          */
2128         balance_gap = min(low_wmark_pages(zone),
2129                 (zone->present_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2130                         KSWAPD_ZONE_BALANCE_GAP_RATIO);
2131         watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2132         watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2133
2134         /*
2135          * If compaction is deferred, reclaim up to a point where
2136          * compaction will have a chance of success when re-enabled
2137          */
2138         if (compaction_deferred(zone))
2139                 return watermark_ok;
2140
2141         /* If compaction is not ready to start, keep reclaiming */
2142         if (!compaction_suitable(zone, sc->order))
2143                 return false;
2144
2145         return watermark_ok;
2146 }
2147
2148 /*
2149  * This is the direct reclaim path, for page-allocating processes.  We only
2150  * try to reclaim pages from zones which will satisfy the caller's allocation
2151  * request.
2152  *
2153  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2154  * Because:
2155  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2156  *    allocation or
2157  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2158  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2159  *    zone defense algorithm.
2160  *
2161  * If a zone is deemed to be full of pinned pages then just give it a light
2162  * scan then give up on it.
2163  *
2164  * This function returns true if a zone is being reclaimed for a costly
2165  * high-order allocation and compaction is ready to begin. This indicates to
2166  * the caller that it should consider retrying the allocation instead of
2167  * further reclaim.
2168  */
2169 static bool shrink_zones(int priority, struct zonelist *zonelist,
2170                                         struct scan_control *sc)
2171 {
2172         struct zoneref *z;
2173         struct zone *zone;
2174         unsigned long nr_soft_reclaimed;
2175         unsigned long nr_soft_scanned;
2176         bool aborted_reclaim = false;
2177
2178         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2179                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2180                 if (!populated_zone(zone))
2181                         continue;
2182                 /*
2183                  * Take care memory controller reclaiming has small influence
2184                  * to global LRU.
2185                  */
2186                 if (scanning_global_lru(sc)) {
2187                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2188                                 continue;
2189                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2190                                 continue;       /* Let kswapd poll it */
2191                         if (COMPACTION_BUILD) {
2192                                 /*
2193                                  * If we already have plenty of memory free for
2194                                  * compaction in this zone, don't free any more.
2195                                  * Even though compaction is invoked for any
2196                                  * non-zero order, only frequent costly order
2197                                  * reclamation is disruptive enough to become a
2198                                  * noticable problem, like transparent huge page
2199                                  * allocations.
2200                                  */
2201                                 if (compaction_ready(zone, sc)) {
2202                                         aborted_reclaim = true;
2203                                         continue;
2204                                 }
2205                         }
2206                         /*
2207                          * This steals pages from memory cgroups over softlimit
2208                          * and returns the number of reclaimed pages and
2209                          * scanned pages. This works for global memory pressure
2210                          * and balancing, not for a memcg's limit.
2211                          */
2212                         nr_soft_scanned = 0;
2213                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2214                                                 sc->order, sc->gfp_mask,
2215                                                 &nr_soft_scanned);
2216                         sc->nr_reclaimed += nr_soft_reclaimed;
2217                         sc->nr_scanned += nr_soft_scanned;
2218                         /* need some check for avoid more shrink_zone() */
2219                 }
2220
2221                 shrink_zone(priority, zone, sc);
2222         }
2223
2224         return aborted_reclaim;
2225 }
2226
2227 static bool zone_reclaimable(struct zone *zone)
2228 {
2229         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2230 }
2231
2232 /* All zones in zonelist are unreclaimable? */
2233 static bool all_unreclaimable(struct zonelist *zonelist,
2234                 struct scan_control *sc)
2235 {
2236         struct zoneref *z;
2237         struct zone *zone;
2238
2239         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2240                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2241                 if (!populated_zone(zone))
2242                         continue;
2243                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2244                         continue;
2245                 if (!zone->all_unreclaimable)
2246                         return false;
2247         }
2248
2249         return true;
2250 }
2251
2252 /*
2253  * This is the main entry point to direct page reclaim.
2254  *
2255  * If a full scan of the inactive list fails to free enough memory then we
2256  * are "out of memory" and something needs to be killed.
2257  *
2258  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2259  * high - the zone may be full of dirty or under-writeback pages, which this
2260  * caller can't do much about.  We kick the writeback threads and take explicit
2261  * naps in the hope that some of these pages can be written.  But if the
2262  * allocating task holds filesystem locks which prevent writeout this might not
2263  * work, and the allocation attempt will fail.
2264  *
2265  * returns:     0, if no pages reclaimed
2266  *              else, the number of pages reclaimed
2267  */
2268 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2269                                         struct scan_control *sc,
2270                                         struct shrink_control *shrink)
2271 {
2272         int priority;
2273         unsigned long total_scanned = 0;
2274         struct reclaim_state *reclaim_state = current->reclaim_state;
2275         struct zoneref *z;
2276         struct zone *zone;
2277         unsigned long writeback_threshold;
2278         bool aborted_reclaim;
2279
2280         delayacct_freepages_start();
2281
2282         if (scanning_global_lru(sc))
2283                 count_vm_event(ALLOCSTALL);
2284
2285         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2286                 sc->nr_scanned = 0;
2287                 if (!priority)
2288                         disable_swap_token(sc->mem_cgroup);
2289                 aborted_reclaim = shrink_zones(priority, zonelist, sc);
2290
2291                 /*
2292                  * Don't shrink slabs when reclaiming memory from
2293                  * over limit cgroups
2294                  */
2295                 if (scanning_global_lru(sc)) {
2296                         unsigned long lru_pages = 0;
2297                         for_each_zone_zonelist(zone, z, zonelist,
2298                                         gfp_zone(sc->gfp_mask)) {
2299                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2300                                         continue;
2301
2302                                 lru_pages += zone_reclaimable_pages(zone);
2303                         }
2304
2305                         shrink_slab(shrink, sc->nr_scanned, lru_pages);
2306                         if (reclaim_state) {
2307                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2308                                 reclaim_state->reclaimed_slab = 0;
2309                         }
2310                 }
2311                 total_scanned += sc->nr_scanned;
2312                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2313                         goto out;
2314
2315                 /*
2316                  * Try to write back as many pages as we just scanned.  This
2317                  * tends to cause slow streaming writers to write data to the
2318                  * disk smoothly, at the dirtying rate, which is nice.   But
2319                  * that's undesirable in laptop mode, where we *want* lumpy
2320                  * writeout.  So in laptop mode, write out the whole world.
2321                  */
2322                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2323                 if (total_scanned > writeback_threshold) {
2324                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2325                                                 WB_REASON_TRY_TO_FREE_PAGES);
2326                         sc->may_writepage = 1;
2327                 }
2328
2329                 /* Take a nap, wait for some writeback to complete */
2330                 if (!sc->hibernation_mode && sc->nr_scanned &&
2331                     priority < DEF_PRIORITY - 2) {
2332                         struct zone *preferred_zone;
2333
2334                         first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2335                                                 &cpuset_current_mems_allowed,
2336                                                 &preferred_zone);
2337                         wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2338                 }
2339         }
2340
2341 out:
2342         delayacct_freepages_end();
2343
2344         if (sc->nr_reclaimed)
2345                 return sc->nr_reclaimed;
2346
2347         /*
2348          * As hibernation is going on, kswapd is freezed so that it can't mark
2349          * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2350          * check.
2351          */
2352         if (oom_killer_disabled)
2353                 return 0;
2354
2355         /* Aborted reclaim to try compaction? don't OOM, then */
2356         if (aborted_reclaim)
2357                 return 1;
2358
2359         /* top priority shrink_zones still had more to do? don't OOM, then */
2360         if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2361                 return 1;
2362
2363         return 0;
2364 }
2365
2366 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2367                                 gfp_t gfp_mask, nodemask_t *nodemask)
2368 {
2369         unsigned long nr_reclaimed;
2370         struct scan_control sc = {
2371                 .gfp_mask = gfp_mask,
2372                 .may_writepage = !laptop_mode,
2373                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2374                 .may_unmap = 1,
2375                 .may_swap = 1,
2376                 .order = order,
2377                 .mem_cgroup = NULL,
2378                 .nodemask = nodemask,
2379         };
2380         struct shrink_control shrink = {
2381                 .gfp_mask = sc.gfp_mask,
2382         };
2383
2384         trace_mm_vmscan_direct_reclaim_begin(order,
2385                                 sc.may_writepage,
2386                                 gfp_mask);
2387
2388         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2389
2390         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2391
2392         return nr_reclaimed;
2393 }
2394
2395 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2396
2397 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2398                                                 gfp_t gfp_mask, bool noswap,
2399                                                 struct zone *zone,
2400                                                 unsigned long *nr_scanned)
2401 {
2402         struct scan_control sc = {
2403                 .nr_scanned = 0,
2404                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2405                 .may_writepage = !laptop_mode,
2406                 .may_unmap = 1,
2407                 .may_swap = !noswap,
2408                 .order = 0,
2409                 .mem_cgroup = mem,
2410         };
2411
2412         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2413                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2414
2415         trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2416                                                       sc.may_writepage,
2417                                                       sc.gfp_mask);
2418
2419         /*
2420          * NOTE: Although we can get the priority field, using it
2421          * here is not a good idea, since it limits the pages we can scan.
2422          * if we don't reclaim here, the shrink_zone from balance_pgdat
2423          * will pick up pages from other mem cgroup's as well. We hack
2424          * the priority and make it zero.
2425          */
2426         shrink_zone(0, zone, &sc);
2427
2428         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2429
2430         *nr_scanned = sc.nr_scanned;
2431         return sc.nr_reclaimed;
2432 }
2433
2434 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2435                                            gfp_t gfp_mask,
2436                                            bool noswap)
2437 {
2438         struct zonelist *zonelist;
2439         unsigned long nr_reclaimed;
2440         int nid;
2441         struct scan_control sc = {
2442                 .may_writepage = !laptop_mode,
2443                 .may_unmap = 1,
2444                 .may_swap = !noswap,
2445                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2446                 .order = 0,
2447                 .mem_cgroup = mem_cont,
2448                 .nodemask = NULL, /* we don't care the placement */
2449                 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2450                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2451         };
2452         struct shrink_control shrink = {
2453                 .gfp_mask = sc.gfp_mask,
2454         };
2455
2456         /*
2457          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2458          * take care of from where we get pages. So the node where we start the
2459          * scan does not need to be the current node.
2460          */
2461         nid = mem_cgroup_select_victim_node(mem_cont);
2462
2463         zonelist = NODE_DATA(nid)->node_zonelists;
2464
2465         trace_mm_vmscan_memcg_reclaim_begin(0,
2466                                             sc.may_writepage,
2467                                             sc.gfp_mask);
2468
2469         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2470
2471         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2472
2473         return nr_reclaimed;
2474 }
2475 #endif
2476
2477 /*
2478  * pgdat_balanced is used when checking if a node is balanced for high-order
2479  * allocations. Only zones that meet watermarks and are in a zone allowed
2480  * by the callers classzone_idx are added to balanced_pages. The total of
2481  * balanced pages must be at least 25% of the zones allowed by classzone_idx
2482  * for the node to be considered balanced. Forcing all zones to be balanced
2483  * for high orders can cause excessive reclaim when there are imbalanced zones.
2484  * The choice of 25% is due to
2485  *   o a 16M DMA zone that is balanced will not balance a zone on any
2486  *     reasonable sized machine
2487  *   o On all other machines, the top zone must be at least a reasonable
2488  *     percentage of the middle zones. For example, on 32-bit x86, highmem
2489  *     would need to be at least 256M for it to be balance a whole node.
2490  *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2491  *     to balance a node on its own. These seemed like reasonable ratios.
2492  */
2493 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2494                                                 int classzone_idx)
2495 {
2496         unsigned long present_pages = 0;
2497         int i;
2498
2499         for (i = 0; i <= classzone_idx; i++)
2500                 present_pages += pgdat->node_zones[i].present_pages;
2501
2502         /* A special case here: if zone has no page, we think it's balanced */
2503         return balanced_pages >= (present_pages >> 2);
2504 }
2505
2506 /* is kswapd sleeping prematurely? */
2507 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2508                                         int classzone_idx)
2509 {
2510         int i;
2511         unsigned long balanced = 0;
2512         bool all_zones_ok = true;
2513
2514         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2515         if (remaining)
2516                 return true;
2517
2518         /* Check the watermark levels */
2519         for (i = 0; i <= classzone_idx; i++) {
2520                 struct zone *zone = pgdat->node_zones + i;
2521
2522                 if (!populated_zone(zone))
2523                         continue;
2524
2525                 /*
2526                  * balance_pgdat() skips over all_unreclaimable after
2527                  * DEF_PRIORITY. Effectively, it considers them balanced so
2528                  * they must be considered balanced here as well if kswapd
2529                  * is to sleep
2530                  */
2531                 if (zone->all_unreclaimable) {
2532                         balanced += zone->present_pages;
2533                         continue;
2534                 }
2535
2536                 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2537                                                         i, 0))
2538                         all_zones_ok = false;
2539                 else
2540                         balanced += zone->present_pages;
2541         }
2542
2543         /*
2544          * For high-order requests, the balanced zones must contain at least
2545          * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2546          * must be balanced
2547          */
2548         if (order)
2549                 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2550         else
2551                 return !all_zones_ok;
2552 }
2553
2554 /*
2555  * For kswapd, balance_pgdat() will work across all this node's zones until
2556  * they are all at high_wmark_pages(zone).
2557  *
2558  * Returns the final order kswapd was reclaiming at
2559  *
2560  * There is special handling here for zones which are full of pinned pages.
2561  * This can happen if the pages are all mlocked, or if they are all used by
2562  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2563  * What we do is to detect the case where all pages in the zone have been
2564  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2565  * dead and from now on, only perform a short scan.  Basically we're polling
2566  * the zone for when the problem goes away.
2567  *
2568  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2569  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2570  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2571  * lower zones regardless of the number of free pages in the lower zones. This
2572  * interoperates with the page allocator fallback scheme to ensure that aging
2573  * of pages is balanced across the zones.
2574  */
2575 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2576                                                         int *classzone_idx)
2577 {
2578         int all_zones_ok;
2579         unsigned long balanced;
2580         int priority;
2581         int i;
2582         int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2583         unsigned long total_scanned;
2584         struct reclaim_state *reclaim_state = current->reclaim_state;
2585         unsigned long nr_soft_reclaimed;
2586         unsigned long nr_soft_scanned;
2587         struct scan_control sc = {
2588                 .gfp_mask = GFP_KERNEL,
2589                 .may_unmap = 1,
2590                 .may_swap = 1,
2591                 /*
2592                  * kswapd doesn't want to be bailed out while reclaim. because
2593                  * we want to put equal scanning pressure on each zone.
2594                  */
2595                 .nr_to_reclaim = ULONG_MAX,
2596                 .order = order,
2597                 .mem_cgroup = NULL,
2598         };
2599         struct shrink_control shrink = {
2600                 .gfp_mask = sc.gfp_mask,
2601         };
2602 loop_again:
2603         total_scanned = 0;
2604         sc.nr_reclaimed = 0;
2605         sc.may_writepage = !laptop_mode;
2606         count_vm_event(PAGEOUTRUN);
2607
2608         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2609                 unsigned long lru_pages = 0;
2610                 int has_under_min_watermark_zone = 0;
2611
2612                 /* The swap token gets in the way of swapout... */
2613                 if (!priority)
2614                         disable_swap_token(NULL);
2615
2616                 all_zones_ok = 1;
2617                 balanced = 0;
2618
2619                 /*
2620                  * Scan in the highmem->dma direction for the highest
2621                  * zone which needs scanning
2622                  */
2623                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2624                         struct zone *zone = pgdat->node_zones + i;
2625
2626                         if (!populated_zone(zone))
2627                                 continue;
2628
2629                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2630                                 continue;
2631
2632                         /*
2633                          * Do some background aging of the anon list, to give
2634                          * pages a chance to be referenced before reclaiming.
2635                          */
2636                         if (inactive_anon_is_low(zone, &sc))
2637                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2638                                                         &sc, priority, 0);
2639
2640                         if (!zone_watermark_ok_safe(zone, order,
2641                                         high_wmark_pages(zone), 0, 0)) {
2642                                 end_zone = i;
2643                                 break;
2644                         } else {
2645                                 /* If balanced, clear the congested flag */
2646                                 zone_clear_flag(zone, ZONE_CONGESTED);
2647                         }
2648                 }
2649                 if (i < 0)
2650                         goto out;
2651
2652                 for (i = 0; i <= end_zone; i++) {
2653                         struct zone *zone = pgdat->node_zones + i;
2654
2655                         lru_pages += zone_reclaimable_pages(zone);
2656                 }
2657
2658                 /*
2659                  * Now scan the zone in the dma->highmem direction, stopping
2660                  * at the last zone which needs scanning.
2661                  *
2662                  * We do this because the page allocator works in the opposite
2663                  * direction.  This prevents the page allocator from allocating
2664                  * pages behind kswapd's direction of progress, which would
2665                  * cause too much scanning of the lower zones.
2666                  */
2667                 for (i = 0; i <= end_zone; i++) {
2668                         struct zone *zone = pgdat->node_zones + i;
2669                         int nr_slab;
2670                         unsigned long balance_gap;
2671
2672                         if (!populated_zone(zone))
2673                                 continue;
2674
2675                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2676                                 continue;
2677
2678                         sc.nr_scanned = 0;
2679
2680                         nr_soft_scanned = 0;
2681                         /*
2682                          * Call soft limit reclaim before calling shrink_zone.
2683                          */
2684                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2685                                                         order, sc.gfp_mask,
2686                                                         &nr_soft_scanned);
2687                         sc.nr_reclaimed += nr_soft_reclaimed;
2688                         total_scanned += nr_soft_scanned;
2689
2690                         /*
2691                          * We put equal pressure on every zone, unless
2692                          * one zone has way too many pages free
2693                          * already. The "too many pages" is defined
2694                          * as the high wmark plus a "gap" where the
2695                          * gap is either the low watermark or 1%
2696                          * of the zone, whichever is smaller.
2697                          */
2698                         balance_gap = min(low_wmark_pages(zone),
2699                                 (zone->present_pages +
2700                                         KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2701                                 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2702                         if (!zone_watermark_ok_safe(zone, order,
2703                                         high_wmark_pages(zone) + balance_gap,
2704                                         end_zone, 0)) {
2705                                 shrink_zone(priority, zone, &sc);
2706
2707                                 reclaim_state->reclaimed_slab = 0;
2708                                 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2709                                 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2710                                 total_scanned += sc.nr_scanned;
2711
2712                                 if (nr_slab == 0 && !zone_reclaimable(zone))
2713                                         zone->all_unreclaimable = 1;
2714                         }
2715
2716                         /*
2717                          * If we've done a decent amount of scanning and
2718                          * the reclaim ratio is low, start doing writepage
2719                          * even in laptop mode
2720                          */
2721                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2722                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2723                                 sc.may_writepage = 1;
2724
2725                         if (zone->all_unreclaimable) {
2726                                 if (end_zone && end_zone == i)
2727                                         end_zone--;
2728                                 continue;
2729                         }
2730
2731                         if (!zone_watermark_ok_safe(zone, order,
2732                                         high_wmark_pages(zone), end_zone, 0)) {
2733                                 all_zones_ok = 0;
2734                                 /*
2735                                  * We are still under min water mark.  This
2736                                  * means that we have a GFP_ATOMIC allocation
2737                                  * failure risk. Hurry up!
2738                                  */
2739                                 if (!zone_watermark_ok_safe(zone, order,
2740                                             min_wmark_pages(zone), end_zone, 0))
2741                                         has_under_min_watermark_zone = 1;
2742                         } else {
2743                                 /*
2744                                  * If a zone reaches its high watermark,
2745                                  * consider it to be no longer congested. It's
2746                                  * possible there are dirty pages backed by
2747                                  * congested BDIs but as pressure is relieved,
2748                                  * spectulatively avoid congestion waits
2749                                  */
2750                                 zone_clear_flag(zone, ZONE_CONGESTED);
2751                                 if (i <= *classzone_idx)
2752                                         balanced += zone->present_pages;
2753                         }
2754
2755                 }
2756                 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2757                         break;          /* kswapd: all done */
2758                 /*
2759                  * OK, kswapd is getting into trouble.  Take a nap, then take
2760                  * another pass across the zones.
2761                  */
2762                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2763                         if (has_under_min_watermark_zone)
2764                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2765                         else
2766                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2767                 }
2768
2769                 /*
2770                  * We do this so kswapd doesn't build up large priorities for
2771                  * example when it is freeing in parallel with allocators. It
2772                  * matches the direct reclaim path behaviour in terms of impact
2773                  * on zone->*_priority.
2774                  */
2775                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2776                         break;
2777         }
2778 out:
2779
2780         /*
2781          * order-0: All zones must meet high watermark for a balanced node
2782          * high-order: Balanced zones must make up at least 25% of the node
2783          *             for the node to be balanced
2784          */
2785         if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2786                 cond_resched();
2787
2788                 try_to_freeze();
2789
2790                 /*
2791                  * Fragmentation may mean that the system cannot be
2792                  * rebalanced for high-order allocations in all zones.
2793                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2794                  * it means the zones have been fully scanned and are still
2795                  * not balanced. For high-order allocations, there is
2796                  * little point trying all over again as kswapd may
2797                  * infinite loop.
2798                  *
2799                  * Instead, recheck all watermarks at order-0 as they
2800                  * are the most important. If watermarks are ok, kswapd will go
2801                  * back to sleep. High-order users can still perform direct
2802                  * reclaim if they wish.
2803                  */
2804                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2805                         order = sc.order = 0;
2806
2807                 goto loop_again;
2808         }
2809
2810         /*
2811          * If kswapd was reclaiming at a higher order, it has the option of
2812          * sleeping without all zones being balanced. Before it does, it must
2813          * ensure that the watermarks for order-0 on *all* zones are met and
2814          * that the congestion flags are cleared. The congestion flag must
2815          * be cleared as kswapd is the only mechanism that clears the flag
2816          * and it is potentially going to sleep here.
2817          */
2818         if (order) {
2819                 for (i = 0; i <= end_zone; i++) {
2820                         struct zone *zone = pgdat->node_zones + i;
2821
2822                         if (!populated_zone(zone))
2823                                 continue;
2824
2825                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2826                                 continue;
2827
2828                         /* Confirm the zone is balanced for order-0 */
2829                         if (!zone_watermark_ok(zone, 0,
2830                                         high_wmark_pages(zone), 0, 0)) {
2831                                 order = sc.order = 0;
2832                                 goto loop_again;
2833                         }
2834
2835                         /* If balanced, clear the congested flag */
2836                         zone_clear_flag(zone, ZONE_CONGESTED);
2837                         if (i <= *classzone_idx)
2838                                 balanced += zone->present_pages;
2839                 }
2840         }
2841
2842         /*
2843          * Return the order we were reclaiming at so sleeping_prematurely()
2844          * makes a decision on the order we were last reclaiming at. However,
2845          * if another caller entered the allocator slow path while kswapd
2846          * was awake, order will remain at the higher level
2847          */
2848         *classzone_idx = end_zone;
2849         return order;
2850 }
2851
2852 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2853 {
2854         long remaining = 0;
2855         DEFINE_WAIT(wait);
2856
2857         if (freezing(current) || kthread_should_stop())
2858                 return;
2859
2860         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2861
2862         /* Try to sleep for a short interval */
2863         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2864                 remaining = schedule_timeout(HZ/10);
2865                 finish_wait(&pgdat->kswapd_wait, &wait);
2866                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2867         }
2868
2869         /*
2870          * After a short sleep, check if it was a premature sleep. If not, then
2871          * go fully to sleep until explicitly woken up.
2872          */
2873         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2874                 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2875
2876                 /*
2877                  * vmstat counters are not perfectly accurate and the estimated
2878                  * value for counters such as NR_FREE_PAGES can deviate from the
2879                  * true value by nr_online_cpus * threshold. To avoid the zone
2880                  * watermarks being breached while under pressure, we reduce the
2881                  * per-cpu vmstat threshold while kswapd is awake and restore
2882                  * them before going back to sleep.
2883                  */
2884                 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2885
2886                 if (!kthread_should_stop())
2887                         schedule();
2888
2889                 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2890         } else {
2891                 if (remaining)
2892                         count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2893                 else
2894                         count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2895         }
2896         finish_wait(&pgdat->kswapd_wait, &wait);
2897 }
2898
2899 /*
2900  * The background pageout daemon, started as a kernel thread
2901  * from the init process.
2902  *
2903  * This basically trickles out pages so that we have _some_
2904  * free memory available even if there is no other activity
2905  * that frees anything up. This is needed for things like routing
2906  * etc, where we otherwise might have all activity going on in
2907  * asynchronous contexts that cannot page things out.
2908  *
2909  * If there are applications that are active memory-allocators
2910  * (most normal use), this basically shouldn't matter.
2911  */
2912 static int kswapd(void *p)
2913 {
2914         unsigned long order, new_order;
2915         unsigned balanced_order;
2916         int classzone_idx, new_classzone_idx;
2917         int balanced_classzone_idx;
2918         pg_data_t *pgdat = (pg_data_t*)p;
2919         struct task_struct *tsk = current;
2920
2921         struct reclaim_state reclaim_state = {
2922                 .reclaimed_slab = 0,
2923         };
2924         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2925
2926         lockdep_set_current_reclaim_state(GFP_KERNEL);
2927
2928         if (!cpumask_empty(cpumask))
2929                 set_cpus_allowed_ptr(tsk, cpumask);
2930         current->reclaim_state = &reclaim_state;
2931
2932         /*
2933          * Tell the memory management that we're a "memory allocator",
2934          * and that if we need more memory we should get access to it
2935          * regardless (see "__alloc_pages()"). "kswapd" should
2936          * never get caught in the normal page freeing logic.
2937          *
2938          * (Kswapd normally doesn't need memory anyway, but sometimes
2939          * you need a small amount of memory in order to be able to
2940          * page out something else, and this flag essentially protects
2941          * us from recursively trying to free more memory as we're
2942          * trying to free the first piece of memory in the first place).
2943          */
2944         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2945         set_freezable();
2946
2947         order = new_order = 0;
2948         balanced_order = 0;
2949         classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2950         balanced_classzone_idx = classzone_idx;
2951         for ( ; ; ) {
2952                 int ret;
2953
2954                 /*
2955                  * If the last balance_pgdat was unsuccessful it's unlikely a
2956                  * new request of a similar or harder type will succeed soon
2957                  * so consider going to sleep on the basis we reclaimed at
2958                  */
2959                 if (balanced_classzone_idx >= new_classzone_idx &&
2960                                         balanced_order == new_order) {
2961                         new_order = pgdat->kswapd_max_order;
2962                         new_classzone_idx = pgdat->classzone_idx;
2963                         pgdat->kswapd_max_order =  0;
2964                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2965                 }
2966
2967                 if (order < new_order || classzone_idx > new_classzone_idx) {
2968                         /*
2969                          * Don't sleep if someone wants a larger 'order'
2970                          * allocation or has tigher zone constraints
2971                          */
2972                         order = new_order;
2973                         classzone_idx = new_classzone_idx;
2974                 } else {
2975                         kswapd_try_to_sleep(pgdat, balanced_order,
2976                                                 balanced_classzone_idx);
2977                         order = pgdat->kswapd_max_order;
2978                         classzone_idx = pgdat->classzone_idx;
2979                         new_order = order;
2980                         new_classzone_idx = classzone_idx;
2981                         pgdat->kswapd_max_order = 0;
2982                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2983                 }
2984
2985                 ret = try_to_freeze();
2986                 if (kthread_should_stop())
2987                         break;
2988
2989                 /*
2990                  * We can speed up thawing tasks if we don't call balance_pgdat
2991                  * after returning from the refrigerator
2992                  */
2993                 if (!ret) {
2994                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2995                         balanced_classzone_idx = classzone_idx;
2996                         balanced_order = balance_pgdat(pgdat, order,
2997                                                 &balanced_classzone_idx);
2998                 }
2999         }
3000
3001         current->reclaim_state = NULL;
3002         return 0;
3003 }
3004
3005 /*
3006  * A zone is low on free memory, so wake its kswapd task to service it.
3007  */
3008 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3009 {
3010         pg_data_t *pgdat;
3011
3012         if (!populated_zone(zone))
3013                 return;
3014
3015         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3016                 return;
3017         pgdat = zone->zone_pgdat;
3018         if (pgdat->kswapd_max_order < order) {
3019                 pgdat->kswapd_max_order = order;
3020                 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3021         }
3022         if (!waitqueue_active(&pgdat->kswapd_wait))
3023                 return;
3024         if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
3025                 return;
3026
3027         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3028         wake_up_interruptible(&pgdat->kswapd_wait);
3029 }
3030
3031 /*
3032  * The reclaimable count would be mostly accurate.
3033  * The less reclaimable pages may be
3034  * - mlocked pages, which will be moved to unevictable list when encountered
3035  * - mapped pages, which may require several travels to be reclaimed
3036  * - dirty pages, which is not "instantly" reclaimable
3037  */
3038 unsigned long global_reclaimable_pages(void)
3039 {
3040         int nr;
3041
3042         nr = global_page_state(NR_ACTIVE_FILE) +
3043              global_page_state(NR_INACTIVE_FILE);
3044
3045         if (nr_swap_pages > 0)
3046                 nr += global_page_state(NR_ACTIVE_ANON) +
3047                       global_page_state(NR_INACTIVE_ANON);
3048
3049         return nr;
3050 }
3051
3052 unsigned long zone_reclaimable_pages(struct zone *zone)
3053 {
3054         int nr;
3055
3056         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3057              zone_page_state(zone, NR_INACTIVE_FILE);
3058
3059         if (nr_swap_pages > 0)
3060                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3061                       zone_page_state(zone, NR_INACTIVE_ANON);
3062
3063         return nr;
3064 }
3065
3066 #ifdef CONFIG_HIBERNATION
3067 /*
3068  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3069  * freed pages.
3070  *
3071  * Rather than trying to age LRUs the aim is to preserve the overall
3072  * LRU order by reclaiming preferentially
3073  * inactive > active > active referenced > active mapped
3074  */
3075 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3076 {
3077         struct reclaim_state reclaim_state;
3078         struct scan_control sc = {
3079                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3080                 .may_swap = 1,
3081                 .may_unmap = 1,
3082                 .may_writepage = 1,
3083                 .nr_to_reclaim = nr_to_reclaim,
3084                 .hibernation_mode = 1,
3085                 .order = 0,
3086         };
3087         struct shrink_control shrink = {
3088                 .gfp_mask = sc.gfp_mask,
3089         };
3090         struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3091         struct task_struct *p = current;
3092         unsigned long nr_reclaimed;
3093
3094         p->flags |= PF_MEMALLOC;
3095         lockdep_set_current_reclaim_state(sc.gfp_mask);
3096         reclaim_state.reclaimed_slab = 0;
3097         p->reclaim_state = &reclaim_state;
3098
3099         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3100
3101         p->reclaim_state = NULL;
3102         lockdep_clear_current_reclaim_state();
3103         p->flags &= ~PF_MEMALLOC;
3104
3105         return nr_reclaimed;
3106 }
3107 #endif /* CONFIG_HIBERNATION */
3108
3109 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3110    not required for correctness.  So if the last cpu in a node goes
3111    away, we get changed to run anywhere: as the first one comes back,
3112    restore their cpu bindings. */
3113 static int __devinit cpu_callback(struct notifier_block *nfb,
3114                                   unsigned long action, void *hcpu)
3115 {
3116         int nid;
3117
3118         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3119                 for_each_node_state(nid, N_HIGH_MEMORY) {
3120                         pg_data_t *pgdat = NODE_DATA(nid);
3121                         const struct cpumask *mask;
3122
3123                         mask = cpumask_of_node(pgdat->node_id);
3124
3125                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3126                                 /* One of our CPUs online: restore mask */
3127                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3128                 }
3129         }
3130         return NOTIFY_OK;
3131 }
3132
3133 /*
3134  * This kswapd start function will be called by init and node-hot-add.
3135  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3136  */
3137 int kswapd_run(int nid)
3138 {
3139         pg_data_t *pgdat = NODE_DATA(nid);
3140         int ret = 0;
3141
3142         if (pgdat->kswapd)
3143                 return 0;
3144
3145         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3146         if (IS_ERR(pgdat->kswapd)) {
3147                 /* failure at boot is fatal */
3148                 BUG_ON(system_state == SYSTEM_BOOTING);
3149                 printk("Failed to start kswapd on node %d\n",nid);
3150                 ret = -1;
3151         }
3152         return ret;
3153 }
3154
3155 /*
3156  * Called by memory hotplug when all memory in a node is offlined.  Caller must
3157  * hold lock_memory_hotplug().
3158  */
3159 void kswapd_stop(int nid)
3160 {
3161         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3162
3163         if (kswapd) {
3164                 kthread_stop(kswapd);
3165                 NODE_DATA(nid)->kswapd = NULL;
3166         }
3167 }
3168
3169 static int __init kswapd_init(void)
3170 {
3171         int nid;
3172
3173         swap_setup();
3174         for_each_node_state(nid, N_HIGH_MEMORY)
3175                 kswapd_run(nid);
3176         hotcpu_notifier(cpu_callback, 0);
3177         return 0;
3178 }
3179
3180 module_init(kswapd_init)
3181
3182 #ifdef CONFIG_NUMA
3183 /*
3184  * Zone reclaim mode
3185  *
3186  * If non-zero call zone_reclaim when the number of free pages falls below
3187  * the watermarks.
3188  */
3189 int zone_reclaim_mode __read_mostly;
3190
3191 #define RECLAIM_OFF 0
3192 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
3193 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
3194 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
3195
3196 /*
3197  * Priority for ZONE_RECLAIM. This determines the fraction of pages
3198  * of a node considered for each zone_reclaim. 4 scans 1/16th of
3199  * a zone.
3200  */
3201 #define ZONE_RECLAIM_PRIORITY 4
3202
3203 /*
3204  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3205  * occur.
3206  */
3207 int sysctl_min_unmapped_ratio = 1;
3208
3209 /*
3210  * If the number of slab pages in a zone grows beyond this percentage then