mm: vmscan: do not writeback filesystem pages in direct reclaim
[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         shrinker->nr = 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                 unsigned long total_scan;
251                 unsigned 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                 /*
259                  * copy the current shrinker scan count into a local variable
260                  * and zero it so that other concurrent shrinker invocations
261                  * don't also do this scanning work.
262                  */
263                 do {
264                         nr = shrinker->nr;
265                 } while (cmpxchg(&shrinker->nr, nr, 0) != nr);
266
267                 total_scan = nr;
268                 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
269                 delta = (4 * nr_pages_scanned) / shrinker->seeks;
270                 delta *= max_pass;
271                 do_div(delta, lru_pages + 1);
272                 total_scan += delta;
273                 if (total_scan < 0) {
274                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
275                                "delete nr=%ld\n",
276                                shrinker->shrink, total_scan);
277                         total_scan = max_pass;
278                 }
279
280                 /*
281                  * We need to avoid excessive windup on filesystem shrinkers
282                  * due to large numbers of GFP_NOFS allocations causing the
283                  * shrinkers to return -1 all the time. This results in a large
284                  * nr being built up so when a shrink that can do some work
285                  * comes along it empties the entire cache due to nr >>>
286                  * max_pass.  This is bad for sustaining a working set in
287                  * memory.
288                  *
289                  * Hence only allow the shrinker to scan the entire cache when
290                  * a large delta change is calculated directly.
291                  */
292                 if (delta < max_pass / 4)
293                         total_scan = min(total_scan, max_pass / 2);
294
295                 /*
296                  * Avoid risking looping forever due to too large nr value:
297                  * never try to free more than twice the estimate number of
298                  * freeable entries.
299                  */
300                 if (total_scan > max_pass * 2)
301                         total_scan = max_pass * 2;
302
303                 trace_mm_shrink_slab_start(shrinker, shrink, nr,
304                                         nr_pages_scanned, lru_pages,
305                                         max_pass, delta, total_scan);
306
307                 while (total_scan >= batch_size) {
308                         int nr_before;
309
310                         nr_before = do_shrinker_shrink(shrinker, shrink, 0);
311                         shrink_ret = do_shrinker_shrink(shrinker, shrink,
312                                                         batch_size);
313                         if (shrink_ret == -1)
314                                 break;
315                         if (shrink_ret < nr_before)
316                                 ret += nr_before - shrink_ret;
317                         count_vm_events(SLABS_SCANNED, batch_size);
318                         total_scan -= batch_size;
319
320                         cond_resched();
321                 }
322
323                 /*
324                  * move the unused scan count back into the shrinker in a
325                  * manner that handles concurrent updates. If we exhausted the
326                  * scan, there is no need to do an update.
327                  */
328                 do {
329                         nr = shrinker->nr;
330                         new_nr = total_scan + nr;
331                         if (total_scan <= 0)
332                                 break;
333                 } while (cmpxchg(&shrinker->nr, nr, new_nr) != nr);
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                 /*
499                  * Wait on writeback if requested to. This happens when
500                  * direct reclaiming a large contiguous area and the
501                  * first attempt to free a range of pages fails.
502                  */
503                 if (PageWriteback(page) &&
504                     (sc->reclaim_mode & RECLAIM_MODE_SYNC))
505                         wait_on_page_writeback(page);
506
507                 if (!PageWriteback(page)) {
508                         /* synchronous write or broken a_ops? */
509                         ClearPageReclaim(page);
510                 }
511                 trace_mm_vmscan_writepage(page,
512                         trace_reclaim_flags(page, sc->reclaim_mode));
513                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
514                 return PAGE_SUCCESS;
515         }
516
517         return PAGE_CLEAN;
518 }
519
520 /*
521  * Same as remove_mapping, but if the page is removed from the mapping, it
522  * gets returned with a refcount of 0.
523  */
524 static int __remove_mapping(struct address_space *mapping, struct page *page)
525 {
526         BUG_ON(!PageLocked(page));
527         BUG_ON(mapping != page_mapping(page));
528
529         spin_lock_irq(&mapping->tree_lock);
530         /*
531          * The non racy check for a busy page.
532          *
533          * Must be careful with the order of the tests. When someone has
534          * a ref to the page, it may be possible that they dirty it then
535          * drop the reference. So if PageDirty is tested before page_count
536          * here, then the following race may occur:
537          *
538          * get_user_pages(&page);
539          * [user mapping goes away]
540          * write_to(page);
541          *                              !PageDirty(page)    [good]
542          * SetPageDirty(page);
543          * put_page(page);
544          *                              !page_count(page)   [good, discard it]
545          *
546          * [oops, our write_to data is lost]
547          *
548          * Reversing the order of the tests ensures such a situation cannot
549          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
550          * load is not satisfied before that of page->_count.
551          *
552          * Note that if SetPageDirty is always performed via set_page_dirty,
553          * and thus under tree_lock, then this ordering is not required.
554          */
555         if (!page_freeze_refs(page, 2))
556                 goto cannot_free;
557         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
558         if (unlikely(PageDirty(page))) {
559                 page_unfreeze_refs(page, 2);
560                 goto cannot_free;
561         }
562
563         if (PageSwapCache(page)) {
564                 swp_entry_t swap = { .val = page_private(page) };
565                 __delete_from_swap_cache(page);
566                 spin_unlock_irq(&mapping->tree_lock);
567                 swapcache_free(swap, page);
568         } else {
569                 void (*freepage)(struct page *);
570
571                 freepage = mapping->a_ops->freepage;
572
573                 __delete_from_page_cache(page);
574                 spin_unlock_irq(&mapping->tree_lock);
575                 mem_cgroup_uncharge_cache_page(page);
576
577                 if (freepage != NULL)
578                         freepage(page);
579         }
580
581         return 1;
582
583 cannot_free:
584         spin_unlock_irq(&mapping->tree_lock);
585         return 0;
586 }
587
588 /*
589  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
590  * someone else has a ref on the page, abort and return 0.  If it was
591  * successfully detached, return 1.  Assumes the caller has a single ref on
592  * this page.
593  */
594 int remove_mapping(struct address_space *mapping, struct page *page)
595 {
596         if (__remove_mapping(mapping, page)) {
597                 /*
598                  * Unfreezing the refcount with 1 rather than 2 effectively
599                  * drops the pagecache ref for us without requiring another
600                  * atomic operation.
601                  */
602                 page_unfreeze_refs(page, 1);
603                 return 1;
604         }
605         return 0;
606 }
607
608 /**
609  * putback_lru_page - put previously isolated page onto appropriate LRU list
610  * @page: page to be put back to appropriate lru list
611  *
612  * Add previously isolated @page to appropriate LRU list.
613  * Page may still be unevictable for other reasons.
614  *
615  * lru_lock must not be held, interrupts must be enabled.
616  */
617 void putback_lru_page(struct page *page)
618 {
619         int lru;
620         int active = !!TestClearPageActive(page);
621         int was_unevictable = PageUnevictable(page);
622
623         VM_BUG_ON(PageLRU(page));
624
625 redo:
626         ClearPageUnevictable(page);
627
628         if (page_evictable(page, NULL)) {
629                 /*
630                  * For evictable pages, we can use the cache.
631                  * In event of a race, worst case is we end up with an
632                  * unevictable page on [in]active list.
633                  * We know how to handle that.
634                  */
635                 lru = active + page_lru_base_type(page);
636                 lru_cache_add_lru(page, lru);
637         } else {
638                 /*
639                  * Put unevictable pages directly on zone's unevictable
640                  * list.
641                  */
642                 lru = LRU_UNEVICTABLE;
643                 add_page_to_unevictable_list(page);
644                 /*
645                  * When racing with an mlock clearing (page is
646                  * unlocked), make sure that if the other thread does
647                  * not observe our setting of PG_lru and fails
648                  * isolation, we see PG_mlocked cleared below and move
649                  * the page back to the evictable list.
650                  *
651                  * The other side is TestClearPageMlocked().
652                  */
653                 smp_mb();
654         }
655
656         /*
657          * page's status can change while we move it among lru. If an evictable
658          * page is on unevictable list, it never be freed. To avoid that,
659          * check after we added it to the list, again.
660          */
661         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
662                 if (!isolate_lru_page(page)) {
663                         put_page(page);
664                         goto redo;
665                 }
666                 /* This means someone else dropped this page from LRU
667                  * So, it will be freed or putback to LRU again. There is
668                  * nothing to do here.
669                  */
670         }
671
672         if (was_unevictable && lru != LRU_UNEVICTABLE)
673                 count_vm_event(UNEVICTABLE_PGRESCUED);
674         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
675                 count_vm_event(UNEVICTABLE_PGCULLED);
676
677         put_page(page);         /* drop ref from isolate */
678 }
679
680 enum page_references {
681         PAGEREF_RECLAIM,
682         PAGEREF_RECLAIM_CLEAN,
683         PAGEREF_KEEP,
684         PAGEREF_ACTIVATE,
685 };
686
687 static enum page_references page_check_references(struct page *page,
688                                                   struct scan_control *sc)
689 {
690         int referenced_ptes, referenced_page;
691         unsigned long vm_flags;
692
693         referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
694         referenced_page = TestClearPageReferenced(page);
695
696         /* Lumpy reclaim - ignore references */
697         if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
698                 return PAGEREF_RECLAIM;
699
700         /*
701          * Mlock lost the isolation race with us.  Let try_to_unmap()
702          * move the page to the unevictable list.
703          */
704         if (vm_flags & VM_LOCKED)
705                 return PAGEREF_RECLAIM;
706
707         if (referenced_ptes) {
708                 if (PageAnon(page))
709                         return PAGEREF_ACTIVATE;
710                 /*
711                  * All mapped pages start out with page table
712                  * references from the instantiating fault, so we need
713                  * to look twice if a mapped file page is used more
714                  * than once.
715                  *
716                  * Mark it and spare it for another trip around the
717                  * inactive list.  Another page table reference will
718                  * lead to its activation.
719                  *
720                  * Note: the mark is set for activated pages as well
721                  * so that recently deactivated but used pages are
722                  * quickly recovered.
723                  */
724                 SetPageReferenced(page);
725
726                 if (referenced_page)
727                         return PAGEREF_ACTIVATE;
728
729                 return PAGEREF_KEEP;
730         }
731
732         /* Reclaim if clean, defer dirty pages to writeback */
733         if (referenced_page && !PageSwapBacked(page))
734                 return PAGEREF_RECLAIM_CLEAN;
735
736         return PAGEREF_RECLAIM;
737 }
738
739 static noinline_for_stack void free_page_list(struct list_head *free_pages)
740 {
741         struct pagevec freed_pvec;
742         struct page *page, *tmp;
743
744         pagevec_init(&freed_pvec, 1);
745
746         list_for_each_entry_safe(page, tmp, free_pages, lru) {
747                 list_del(&page->lru);
748                 if (!pagevec_add(&freed_pvec, page)) {
749                         __pagevec_free(&freed_pvec);
750                         pagevec_reinit(&freed_pvec);
751                 }
752         }
753
754         pagevec_free(&freed_pvec);
755 }
756
757 /*
758  * shrink_page_list() returns the number of reclaimed pages
759  */
760 static unsigned long shrink_page_list(struct list_head *page_list,
761                                       struct zone *zone,
762                                       struct scan_control *sc)
763 {
764         LIST_HEAD(ret_pages);
765         LIST_HEAD(free_pages);
766         int pgactivate = 0;
767         unsigned long nr_dirty = 0;
768         unsigned long nr_congested = 0;
769         unsigned long nr_reclaimed = 0;
770
771         cond_resched();
772
773         while (!list_empty(page_list)) {
774                 enum page_references references;
775                 struct address_space *mapping;
776                 struct page *page;
777                 int may_enter_fs;
778
779                 cond_resched();
780
781                 page = lru_to_page(page_list);
782                 list_del(&page->lru);
783
784                 if (!trylock_page(page))
785                         goto keep;
786
787                 VM_BUG_ON(PageActive(page));
788                 VM_BUG_ON(page_zone(page) != zone);
789
790                 sc->nr_scanned++;
791
792                 if (unlikely(!page_evictable(page, NULL)))
793                         goto cull_mlocked;
794
795                 if (!sc->may_unmap && page_mapped(page))
796                         goto keep_locked;
797
798                 /* Double the slab pressure for mapped and swapcache pages */
799                 if (page_mapped(page) || PageSwapCache(page))
800                         sc->nr_scanned++;
801
802                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
803                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
804
805                 if (PageWriteback(page)) {
806                         /*
807                          * Synchronous reclaim is performed in two passes,
808                          * first an asynchronous pass over the list to
809                          * start parallel writeback, and a second synchronous
810                          * pass to wait for the IO to complete.  Wait here
811                          * for any page for which writeback has already
812                          * started.
813                          */
814                         if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
815                             may_enter_fs)
816                                 wait_on_page_writeback(page);
817                         else {
818                                 unlock_page(page);
819                                 goto keep_lumpy;
820                         }
821                 }
822
823                 references = page_check_references(page, sc);
824                 switch (references) {
825                 case PAGEREF_ACTIVATE:
826                         goto activate_locked;
827                 case PAGEREF_KEEP:
828                         goto keep_locked;
829                 case PAGEREF_RECLAIM:
830                 case PAGEREF_RECLAIM_CLEAN:
831                         ; /* try to reclaim the page below */
832                 }
833
834                 /*
835                  * Anonymous process memory has backing store?
836                  * Try to allocate it some swap space here.
837                  */
838                 if (PageAnon(page) && !PageSwapCache(page)) {
839                         if (!(sc->gfp_mask & __GFP_IO))
840                                 goto keep_locked;
841                         if (!add_to_swap(page))
842                                 goto activate_locked;
843                         may_enter_fs = 1;
844                 }
845
846                 mapping = page_mapping(page);
847
848                 /*
849                  * The page is mapped into the page tables of one or more
850                  * processes. Try to unmap it here.
851                  */
852                 if (page_mapped(page) && mapping) {
853                         switch (try_to_unmap(page, TTU_UNMAP)) {
854                         case SWAP_FAIL:
855                                 goto activate_locked;
856                         case SWAP_AGAIN:
857                                 goto keep_locked;
858                         case SWAP_MLOCK:
859                                 goto cull_mlocked;
860                         case SWAP_SUCCESS:
861                                 ; /* try to free the page below */
862                         }
863                 }
864
865                 if (PageDirty(page)) {
866                         nr_dirty++;
867
868                         /*
869                          * Only kswapd can writeback filesystem pages to
870                          * avoid risk of stack overflow
871                          */
872                         if (page_is_file_cache(page) && !current_is_kswapd()) {
873                                 inc_zone_page_state(page, NR_VMSCAN_WRITE_SKIP);
874                                 goto keep_locked;
875                         }
876
877                         if (references == PAGEREF_RECLAIM_CLEAN)
878                                 goto keep_locked;
879                         if (!may_enter_fs)
880                                 goto keep_locked;
881                         if (!sc->may_writepage)
882                                 goto keep_locked;
883
884                         /* Page is dirty, try to write it out here */
885                         switch (pageout(page, mapping, sc)) {
886                         case PAGE_KEEP:
887                                 nr_congested++;
888                                 goto keep_locked;
889                         case PAGE_ACTIVATE:
890                                 goto activate_locked;
891                         case PAGE_SUCCESS:
892                                 if (PageWriteback(page))
893                                         goto keep_lumpy;
894                                 if (PageDirty(page))
895                                         goto keep;
896
897                                 /*
898                                  * A synchronous write - probably a ramdisk.  Go
899                                  * ahead and try to reclaim the page.
900                                  */
901                                 if (!trylock_page(page))
902                                         goto keep;
903                                 if (PageDirty(page) || PageWriteback(page))
904                                         goto keep_locked;
905                                 mapping = page_mapping(page);
906                         case PAGE_CLEAN:
907                                 ; /* try to free the page below */
908                         }
909                 }
910
911                 /*
912                  * If the page has buffers, try to free the buffer mappings
913                  * associated with this page. If we succeed we try to free
914                  * the page as well.
915                  *
916                  * We do this even if the page is PageDirty().
917                  * try_to_release_page() does not perform I/O, but it is
918                  * possible for a page to have PageDirty set, but it is actually
919                  * clean (all its buffers are clean).  This happens if the
920                  * buffers were written out directly, with submit_bh(). ext3
921                  * will do this, as well as the blockdev mapping.
922                  * try_to_release_page() will discover that cleanness and will
923                  * drop the buffers and mark the page clean - it can be freed.
924                  *
925                  * Rarely, pages can have buffers and no ->mapping.  These are
926                  * the pages which were not successfully invalidated in
927                  * truncate_complete_page().  We try to drop those buffers here
928                  * and if that worked, and the page is no longer mapped into
929                  * process address space (page_count == 1) it can be freed.
930                  * Otherwise, leave the page on the LRU so it is swappable.
931                  */
932                 if (page_has_private(page)) {
933                         if (!try_to_release_page(page, sc->gfp_mask))
934                                 goto activate_locked;
935                         if (!mapping && page_count(page) == 1) {
936                                 unlock_page(page);
937                                 if (put_page_testzero(page))
938                                         goto free_it;
939                                 else {
940                                         /*
941                                          * rare race with speculative reference.
942                                          * the speculative reference will free
943                                          * this page shortly, so we may
944                                          * increment nr_reclaimed here (and
945                                          * leave it off the LRU).
946                                          */
947                                         nr_reclaimed++;
948                                         continue;
949                                 }
950                         }
951                 }
952
953                 if (!mapping || !__remove_mapping(mapping, page))
954                         goto keep_locked;
955
956                 /*
957                  * At this point, we have no other references and there is
958                  * no way to pick any more up (removed from LRU, removed
959                  * from pagecache). Can use non-atomic bitops now (and
960                  * we obviously don't have to worry about waking up a process
961                  * waiting on the page lock, because there are no references.
962                  */
963                 __clear_page_locked(page);
964 free_it:
965                 nr_reclaimed++;
966
967                 /*
968                  * Is there need to periodically free_page_list? It would
969                  * appear not as the counts should be low
970                  */
971                 list_add(&page->lru, &free_pages);
972                 continue;
973
974 cull_mlocked:
975                 if (PageSwapCache(page))
976                         try_to_free_swap(page);
977                 unlock_page(page);
978                 putback_lru_page(page);
979                 reset_reclaim_mode(sc);
980                 continue;
981
982 activate_locked:
983                 /* Not a candidate for swapping, so reclaim swap space. */
984                 if (PageSwapCache(page) && vm_swap_full())
985                         try_to_free_swap(page);
986                 VM_BUG_ON(PageActive(page));
987                 SetPageActive(page);
988                 pgactivate++;
989 keep_locked:
990                 unlock_page(page);
991 keep:
992                 reset_reclaim_mode(sc);
993 keep_lumpy:
994                 list_add(&page->lru, &ret_pages);
995                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
996         }
997
998         /*
999          * Tag a zone as congested if all the dirty pages encountered were
1000          * backed by a congested BDI. In this case, reclaimers should just
1001          * back off and wait for congestion to clear because further reclaim
1002          * will encounter the same problem
1003          */
1004         if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc))
1005                 zone_set_flag(zone, ZONE_CONGESTED);
1006
1007         free_page_list(&free_pages);
1008
1009         list_splice(&ret_pages, page_list);
1010         count_vm_events(PGACTIVATE, pgactivate);
1011         return nr_reclaimed;
1012 }
1013
1014 /*
1015  * Attempt to remove the specified page from its LRU.  Only take this page
1016  * if it is of the appropriate PageActive status.  Pages which are being
1017  * freed elsewhere are also ignored.
1018  *
1019  * page:        page to consider
1020  * mode:        one of the LRU isolation modes defined above
1021  *
1022  * returns 0 on success, -ve errno on failure.
1023  */
1024 int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file)
1025 {
1026         bool all_lru_mode;
1027         int ret = -EINVAL;
1028
1029         /* Only take pages on the LRU. */
1030         if (!PageLRU(page))
1031                 return ret;
1032
1033         all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) ==
1034                 (ISOLATE_ACTIVE|ISOLATE_INACTIVE);
1035
1036         /*
1037          * When checking the active state, we need to be sure we are
1038          * dealing with comparible boolean values.  Take the logical not
1039          * of each.
1040          */
1041         if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE))
1042                 return ret;
1043
1044         if (!all_lru_mode && !!page_is_file_cache(page) != file)
1045                 return ret;
1046
1047         /*
1048          * When this function is being called for lumpy reclaim, we
1049          * initially look into all LRU pages, active, inactive and
1050          * unevictable; only give shrink_page_list evictable pages.
1051          */
1052         if (PageUnevictable(page))
1053                 return ret;
1054
1055         ret = -EBUSY;
1056
1057         if ((mode & ISOLATE_CLEAN) && (PageDirty(page) || PageWriteback(page)))
1058                 return ret;
1059
1060         if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1061                 return ret;
1062
1063         if (likely(get_page_unless_zero(page))) {
1064                 /*
1065                  * Be careful not to clear PageLRU until after we're
1066                  * sure the page is not being freed elsewhere -- the
1067                  * page release code relies on it.
1068                  */
1069                 ClearPageLRU(page);
1070                 ret = 0;
1071         }
1072
1073         return ret;
1074 }
1075
1076 /*
1077  * zone->lru_lock is heavily contended.  Some of the functions that
1078  * shrink the lists perform better by taking out a batch of pages
1079  * and working on them outside the LRU lock.
1080  *
1081  * For pagecache intensive workloads, this function is the hottest
1082  * spot in the kernel (apart from copy_*_user functions).
1083  *
1084  * Appropriate locks must be held before calling this function.
1085  *
1086  * @nr_to_scan: The number of pages to look through on the list.
1087  * @src:        The LRU list to pull pages off.
1088  * @dst:        The temp list to put pages on to.
1089  * @scanned:    The number of pages that were scanned.
1090  * @order:      The caller's attempted allocation order
1091  * @mode:       One of the LRU isolation modes
1092  * @file:       True [1] if isolating file [!anon] pages
1093  *
1094  * returns how many pages were moved onto *@dst.
1095  */
1096 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1097                 struct list_head *src, struct list_head *dst,
1098                 unsigned long *scanned, int order, isolate_mode_t mode,
1099                 int file)
1100 {
1101         unsigned long nr_taken = 0;
1102         unsigned long nr_lumpy_taken = 0;
1103         unsigned long nr_lumpy_dirty = 0;
1104         unsigned long nr_lumpy_failed = 0;
1105         unsigned long scan;
1106
1107         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1108                 struct page *page;
1109                 unsigned long pfn;
1110                 unsigned long end_pfn;
1111                 unsigned long page_pfn;
1112                 int zone_id;
1113
1114                 page = lru_to_page(src);
1115                 prefetchw_prev_lru_page(page, src, flags);
1116
1117                 VM_BUG_ON(!PageLRU(page));
1118
1119                 switch (__isolate_lru_page(page, mode, file)) {
1120                 case 0:
1121                         list_move(&page->lru, dst);
1122                         mem_cgroup_del_lru(page);
1123                         nr_taken += hpage_nr_pages(page);
1124                         break;
1125
1126                 case -EBUSY:
1127                         /* else it is being freed elsewhere */
1128                         list_move(&page->lru, src);
1129                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1130                         continue;
1131
1132                 default:
1133                         BUG();
1134                 }
1135
1136                 if (!order)
1137                         continue;
1138
1139                 /*
1140                  * Attempt to take all pages in the order aligned region
1141                  * surrounding the tag page.  Only take those pages of
1142                  * the same active state as that tag page.  We may safely
1143                  * round the target page pfn down to the requested order
1144                  * as the mem_map is guaranteed valid out to MAX_ORDER,
1145                  * where that page is in a different zone we will detect
1146                  * it from its zone id and abort this block scan.
1147                  */
1148                 zone_id = page_zone_id(page);
1149                 page_pfn = page_to_pfn(page);
1150                 pfn = page_pfn & ~((1 << order) - 1);
1151                 end_pfn = pfn + (1 << order);
1152                 for (; pfn < end_pfn; pfn++) {
1153                         struct page *cursor_page;
1154
1155                         /* The target page is in the block, ignore it. */
1156                         if (unlikely(pfn == page_pfn))
1157                                 continue;
1158
1159                         /* Avoid holes within the zone. */
1160                         if (unlikely(!pfn_valid_within(pfn)))
1161                                 break;
1162
1163                         cursor_page = pfn_to_page(pfn);
1164
1165                         /* Check that we have not crossed a zone boundary. */
1166                         if (unlikely(page_zone_id(cursor_page) != zone_id))
1167                                 break;
1168
1169                         /*
1170                          * If we don't have enough swap space, reclaiming of
1171                          * anon page which don't already have a swap slot is
1172                          * pointless.
1173                          */
1174                         if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1175                             !PageSwapCache(cursor_page))
1176                                 break;
1177
1178                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1179                                 list_move(&cursor_page->lru, dst);
1180                                 mem_cgroup_del_lru(cursor_page);
1181                                 nr_taken += hpage_nr_pages(page);
1182                                 nr_lumpy_taken++;
1183                                 if (PageDirty(cursor_page))
1184                                         nr_lumpy_dirty++;
1185                                 scan++;
1186                         } else {
1187                                 /*
1188                                  * Check if the page is freed already.
1189                                  *
1190                                  * We can't use page_count() as that
1191                                  * requires compound_head and we don't
1192                                  * have a pin on the page here. If a
1193                                  * page is tail, we may or may not
1194                                  * have isolated the head, so assume
1195                                  * it's not free, it'd be tricky to
1196                                  * track the head status without a
1197                                  * page pin.
1198                                  */
1199                                 if (!PageTail(cursor_page) &&
1200                                     !atomic_read(&cursor_page->_count))
1201                                         continue;
1202                                 break;
1203                         }
1204                 }
1205
1206                 /* If we break out of the loop above, lumpy reclaim failed */
1207                 if (pfn < end_pfn)
1208                         nr_lumpy_failed++;
1209         }
1210
1211         *scanned = scan;
1212
1213         trace_mm_vmscan_lru_isolate(order,
1214                         nr_to_scan, scan,
1215                         nr_taken,
1216                         nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1217                         mode);
1218         return nr_taken;
1219 }
1220
1221 static unsigned long isolate_pages_global(unsigned long nr,
1222                                         struct list_head *dst,
1223                                         unsigned long *scanned, int order,
1224                                         isolate_mode_t mode,
1225                                         struct zone *z, int active, int file)
1226 {
1227         int lru = LRU_BASE;
1228         if (active)
1229                 lru += LRU_ACTIVE;
1230         if (file)
1231                 lru += LRU_FILE;
1232         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1233                                                                 mode, file);
1234 }
1235
1236 /*
1237  * clear_active_flags() is a helper for shrink_active_list(), clearing
1238  * any active bits from the pages in the list.
1239  */
1240 static unsigned long clear_active_flags(struct list_head *page_list,
1241                                         unsigned int *count)
1242 {
1243         int nr_active = 0;
1244         int lru;
1245         struct page *page;
1246
1247         list_for_each_entry(page, page_list, lru) {
1248                 int numpages = hpage_nr_pages(page);
1249                 lru = page_lru_base_type(page);
1250                 if (PageActive(page)) {
1251                         lru += LRU_ACTIVE;
1252                         ClearPageActive(page);
1253                         nr_active += numpages;
1254                 }
1255                 if (count)
1256                         count[lru] += numpages;
1257         }
1258
1259         return nr_active;
1260 }
1261
1262 /**
1263  * isolate_lru_page - tries to isolate a page from its LRU list
1264  * @page: page to isolate from its LRU list
1265  *
1266  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1267  * vmstat statistic corresponding to whatever LRU list the page was on.
1268  *
1269  * Returns 0 if the page was removed from an LRU list.
1270  * Returns -EBUSY if the page was not on an LRU list.
1271  *
1272  * The returned page will have PageLRU() cleared.  If it was found on
1273  * the active list, it will have PageActive set.  If it was found on
1274  * the unevictable list, it will have the PageUnevictable bit set. That flag
1275  * may need to be cleared by the caller before letting the page go.
1276  *
1277  * The vmstat statistic corresponding to the list on which the page was
1278  * found will be decremented.
1279  *
1280  * Restrictions:
1281  * (1) Must be called with an elevated refcount on the page. This is a
1282  *     fundamentnal difference from isolate_lru_pages (which is called
1283  *     without a stable reference).
1284  * (2) the lru_lock must not be held.
1285  * (3) interrupts must be enabled.
1286  */
1287 int isolate_lru_page(struct page *page)
1288 {
1289         int ret = -EBUSY;
1290
1291         VM_BUG_ON(!page_count(page));
1292
1293         if (PageLRU(page)) {
1294                 struct zone *zone = page_zone(page);
1295
1296                 spin_lock_irq(&zone->lru_lock);
1297                 if (PageLRU(page)) {
1298                         int lru = page_lru(page);
1299                         ret = 0;
1300                         get_page(page);
1301                         ClearPageLRU(page);
1302
1303                         del_page_from_lru_list(zone, page, lru);
1304                 }
1305                 spin_unlock_irq(&zone->lru_lock);
1306         }
1307         return ret;
1308 }
1309
1310 /*
1311  * Are there way too many processes in the direct reclaim path already?
1312  */
1313 static int too_many_isolated(struct zone *zone, int file,
1314                 struct scan_control *sc)
1315 {
1316         unsigned long inactive, isolated;
1317
1318         if (current_is_kswapd())
1319                 return 0;
1320
1321         if (!scanning_global_lru(sc))
1322                 return 0;
1323
1324         if (file) {
1325                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1326                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1327         } else {
1328                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1329                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1330         }
1331
1332         return isolated > inactive;
1333 }
1334
1335 /*
1336  * TODO: Try merging with migrations version of putback_lru_pages
1337  */
1338 static noinline_for_stack void
1339 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1340                                 unsigned long nr_anon, unsigned long nr_file,
1341                                 struct list_head *page_list)
1342 {
1343         struct page *page;
1344         struct pagevec pvec;
1345         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1346
1347         pagevec_init(&pvec, 1);
1348
1349         /*
1350          * Put back any unfreeable pages.
1351          */
1352         spin_lock(&zone->lru_lock);
1353         while (!list_empty(page_list)) {
1354                 int lru;
1355                 page = lru_to_page(page_list);
1356                 VM_BUG_ON(PageLRU(page));
1357                 list_del(&page->lru);
1358                 if (unlikely(!page_evictable(page, NULL))) {
1359                         spin_unlock_irq(&zone->lru_lock);
1360                         putback_lru_page(page);
1361                         spin_lock_irq(&zone->lru_lock);
1362                         continue;
1363                 }
1364                 SetPageLRU(page);
1365                 lru = page_lru(page);
1366                 add_page_to_lru_list(zone, page, lru);
1367                 if (is_active_lru(lru)) {
1368                         int file = is_file_lru(lru);
1369                         int numpages = hpage_nr_pages(page);
1370                         reclaim_stat->recent_rotated[file] += numpages;
1371                 }
1372                 if (!pagevec_add(&pvec, page)) {
1373                         spin_unlock_irq(&zone->lru_lock);
1374                         __pagevec_release(&pvec);
1375                         spin_lock_irq(&zone->lru_lock);
1376                 }
1377         }
1378         __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1379         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1380
1381         spin_unlock_irq(&zone->lru_lock);
1382         pagevec_release(&pvec);
1383 }
1384
1385 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1386                                         struct scan_control *sc,
1387                                         unsigned long *nr_anon,
1388                                         unsigned long *nr_file,
1389                                         struct list_head *isolated_list)
1390 {
1391         unsigned long nr_active;
1392         unsigned int count[NR_LRU_LISTS] = { 0, };
1393         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1394
1395         nr_active = clear_active_flags(isolated_list, count);
1396         __count_vm_events(PGDEACTIVATE, nr_active);
1397
1398         __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1399                               -count[LRU_ACTIVE_FILE]);
1400         __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1401                               -count[LRU_INACTIVE_FILE]);
1402         __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1403                               -count[LRU_ACTIVE_ANON]);
1404         __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1405                               -count[LRU_INACTIVE_ANON]);
1406
1407         *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1408         *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1409         __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1410         __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1411
1412         reclaim_stat->recent_scanned[0] += *nr_anon;
1413         reclaim_stat->recent_scanned[1] += *nr_file;
1414 }
1415
1416 /*
1417  * Returns true if the caller should wait to clean dirty/writeback pages.
1418  *
1419  * If we are direct reclaiming for contiguous pages and we do not reclaim
1420  * everything in the list, try again and wait for writeback IO to complete.
1421  * This will stall high-order allocations noticeably. Only do that when really
1422  * need to free the pages under high memory pressure.
1423  */
1424 static inline bool should_reclaim_stall(unsigned long nr_taken,
1425                                         unsigned long nr_freed,
1426                                         int priority,
1427                                         struct scan_control *sc)
1428 {
1429         int lumpy_stall_priority;
1430
1431         /* kswapd should not stall on sync IO */
1432         if (current_is_kswapd())
1433                 return false;
1434
1435         /* Only stall on lumpy reclaim */
1436         if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1437                 return false;
1438
1439         /* If we have reclaimed everything on the isolated list, no stall */
1440         if (nr_freed == nr_taken)
1441                 return false;
1442
1443         /*
1444          * For high-order allocations, there are two stall thresholds.
1445          * High-cost allocations stall immediately where as lower
1446          * order allocations such as stacks require the scanning
1447          * priority to be much higher before stalling.
1448          */
1449         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1450                 lumpy_stall_priority = DEF_PRIORITY;
1451         else
1452                 lumpy_stall_priority = DEF_PRIORITY / 3;
1453
1454         return priority <= lumpy_stall_priority;
1455 }
1456
1457 /*
1458  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1459  * of reclaimed pages
1460  */
1461 static noinline_for_stack unsigned long
1462 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1463                         struct scan_control *sc, int priority, int file)
1464 {
1465         LIST_HEAD(page_list);
1466         unsigned long nr_scanned;
1467         unsigned long nr_reclaimed = 0;
1468         unsigned long nr_taken;
1469         unsigned long nr_anon;
1470         unsigned long nr_file;
1471         isolate_mode_t reclaim_mode = ISOLATE_INACTIVE;
1472
1473         while (unlikely(too_many_isolated(zone, file, sc))) {
1474                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1475
1476                 /* We are about to die and free our memory. Return now. */
1477                 if (fatal_signal_pending(current))
1478                         return SWAP_CLUSTER_MAX;
1479         }
1480
1481         set_reclaim_mode(priority, sc, false);
1482         if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
1483                 reclaim_mode |= ISOLATE_ACTIVE;
1484
1485         lru_add_drain();
1486
1487         if (!sc->may_unmap)
1488                 reclaim_mode |= ISOLATE_UNMAPPED;
1489         if (!sc->may_writepage)
1490                 reclaim_mode |= ISOLATE_CLEAN;
1491
1492         spin_lock_irq(&zone->lru_lock);
1493
1494         if (scanning_global_lru(sc)) {
1495                 nr_taken = isolate_pages_global(nr_to_scan, &page_list,
1496                         &nr_scanned, sc->order, reclaim_mode, zone, 0, file);
1497                 zone->pages_scanned += nr_scanned;
1498                 if (current_is_kswapd())
1499                         __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1500                                                nr_scanned);
1501                 else
1502                         __count_zone_vm_events(PGSCAN_DIRECT, zone,
1503                                                nr_scanned);
1504         } else {
1505                 nr_taken = mem_cgroup_isolate_pages(nr_to_scan, &page_list,
1506                         &nr_scanned, sc->order, reclaim_mode, zone,
1507                         sc->mem_cgroup, 0, file);
1508                 /*
1509                  * mem_cgroup_isolate_pages() keeps track of
1510                  * scanned pages on its own.
1511                  */
1512         }
1513
1514         if (nr_taken == 0) {
1515                 spin_unlock_irq(&zone->lru_lock);
1516                 return 0;
1517         }
1518
1519         update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1520
1521         spin_unlock_irq(&zone->lru_lock);
1522
1523         nr_reclaimed = shrink_page_list(&page_list, zone, sc);
1524
1525         /* Check if we should syncronously wait for writeback */
1526         if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1527                 set_reclaim_mode(priority, sc, true);
1528                 nr_reclaimed += shrink_page_list(&page_list, zone, sc);
1529         }
1530
1531         local_irq_disable();
1532         if (current_is_kswapd())
1533                 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1534         __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1535
1536         putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1537
1538         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1539                 zone_idx(zone),
1540                 nr_scanned, nr_reclaimed,
1541                 priority,
1542                 trace_shrink_flags(file, sc->reclaim_mode));
1543         return nr_reclaimed;
1544 }
1545
1546 /*
1547  * This moves pages from the active list to the inactive list.
1548  *
1549  * We move them the other way if the page is referenced by one or more
1550  * processes, from rmap.
1551  *
1552  * If the pages are mostly unmapped, the processing is fast and it is
1553  * appropriate to hold zone->lru_lock across the whole operation.  But if
1554  * the pages are mapped, the processing is slow (page_referenced()) so we
1555  * should drop zone->lru_lock around each page.  It's impossible to balance
1556  * this, so instead we remove the pages from the LRU while processing them.
1557  * It is safe to rely on PG_active against the non-LRU pages in here because
1558  * nobody will play with that bit on a non-LRU page.
1559  *
1560  * The downside is that we have to touch page->_count against each page.
1561  * But we had to alter page->flags anyway.
1562  */
1563
1564 static void move_active_pages_to_lru(struct zone *zone,
1565                                      struct list_head *list,
1566                                      enum lru_list lru)
1567 {
1568         unsigned long pgmoved = 0;
1569         struct pagevec pvec;
1570         struct page *page;
1571
1572         pagevec_init(&pvec, 1);
1573
1574         while (!list_empty(list)) {
1575                 page = lru_to_page(list);
1576
1577                 VM_BUG_ON(PageLRU(page));
1578                 SetPageLRU(page);
1579
1580                 list_move(&page->lru, &zone->lru[lru].list);
1581                 mem_cgroup_add_lru_list(page, lru);
1582                 pgmoved += hpage_nr_pages(page);
1583
1584                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1585                         spin_unlock_irq(&zone->lru_lock);
1586                         if (buffer_heads_over_limit)
1587                                 pagevec_strip(&pvec);
1588                         __pagevec_release(&pvec);
1589                         spin_lock_irq(&zone->lru_lock);
1590                 }
1591         }
1592         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1593         if (!is_active_lru(lru))
1594                 __count_vm_events(PGDEACTIVATE, pgmoved);
1595 }
1596
1597 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1598                         struct scan_control *sc, int priority, int file)
1599 {
1600         unsigned long nr_taken;
1601         unsigned long pgscanned;
1602         unsigned long vm_flags;
1603         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1604         LIST_HEAD(l_active);
1605         LIST_HEAD(l_inactive);
1606         struct page *page;
1607         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1608         unsigned long nr_rotated = 0;
1609         isolate_mode_t reclaim_mode = ISOLATE_ACTIVE;
1610
1611         lru_add_drain();
1612
1613         if (!sc->may_unmap)
1614                 reclaim_mode |= ISOLATE_UNMAPPED;
1615         if (!sc->may_writepage)
1616                 reclaim_mode |= ISOLATE_CLEAN;
1617
1618         spin_lock_irq(&zone->lru_lock);
1619         if (scanning_global_lru(sc)) {
1620                 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1621                                                 &pgscanned, sc->order,
1622                                                 reclaim_mode, zone,
1623                                                 1, file);
1624                 zone->pages_scanned += pgscanned;
1625         } else {
1626                 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1627                                                 &pgscanned, sc->order,
1628                                                 reclaim_mode, zone,
1629                                                 sc->mem_cgroup, 1, file);
1630                 /*
1631                  * mem_cgroup_isolate_pages() keeps track of
1632                  * scanned pages on its own.
1633                  */
1634         }
1635
1636         reclaim_stat->recent_scanned[file] += nr_taken;
1637
1638         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1639         if (file)
1640                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1641         else
1642                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1643         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1644         spin_unlock_irq(&zone->lru_lock);
1645
1646         while (!list_empty(&l_hold)) {
1647                 cond_resched();
1648                 page = lru_to_page(&l_hold);
1649                 list_del(&page->lru);
1650
1651                 if (unlikely(!page_evictable(page, NULL))) {
1652                         putback_lru_page(page);
1653                         continue;
1654                 }
1655
1656                 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1657                         nr_rotated += hpage_nr_pages(page);
1658                         /*
1659                          * Identify referenced, file-backed active pages and
1660                          * give them one more trip around the active list. So
1661                          * that executable code get better chances to stay in
1662                          * memory under moderate memory pressure.  Anon pages
1663                          * are not likely to be evicted by use-once streaming
1664                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1665                          * so we ignore them here.
1666                          */
1667                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1668                                 list_add(&page->lru, &l_active);
1669                                 continue;
1670                         }
1671                 }
1672
1673                 ClearPageActive(page);  /* we are de-activating */
1674                 list_add(&page->lru, &l_inactive);
1675         }
1676
1677         /*
1678          * Move pages back to the lru list.
1679          */
1680         spin_lock_irq(&zone->lru_lock);
1681         /*
1682          * Count referenced pages from currently used mappings as rotated,
1683          * even though only some of them are actually re-activated.  This
1684          * helps balance scan pressure between file and anonymous pages in
1685          * get_scan_ratio.
1686          */
1687         reclaim_stat->recent_rotated[file] += nr_rotated;
1688
1689         move_active_pages_to_lru(zone, &l_active,
1690                                                 LRU_ACTIVE + file * LRU_FILE);
1691         move_active_pages_to_lru(zone, &l_inactive,
1692                                                 LRU_BASE   + file * LRU_FILE);
1693         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1694         spin_unlock_irq(&zone->lru_lock);
1695 }
1696
1697 #ifdef CONFIG_SWAP
1698 static int inactive_anon_is_low_global(struct zone *zone)
1699 {
1700         unsigned long active, inactive;
1701
1702         active = zone_page_state(zone, NR_ACTIVE_ANON);
1703         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1704
1705         if (inactive * zone->inactive_ratio < active)
1706                 return 1;
1707
1708         return 0;
1709 }
1710
1711 /**
1712  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1713  * @zone: zone to check
1714  * @sc:   scan control of this context
1715  *
1716  * Returns true if the zone does not have enough inactive anon pages,
1717  * meaning some active anon pages need to be deactivated.
1718  */
1719 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1720 {
1721         int low;
1722
1723         /*
1724          * If we don't have swap space, anonymous page deactivation
1725          * is pointless.
1726          */
1727         if (!total_swap_pages)
1728                 return 0;
1729
1730         if (scanning_global_lru(sc))
1731                 low = inactive_anon_is_low_global(zone);
1732         else
1733                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1734         return low;
1735 }
1736 #else
1737 static inline int inactive_anon_is_low(struct zone *zone,
1738                                         struct scan_control *sc)
1739 {
1740         return 0;
1741 }
1742 #endif
1743
1744 static int inactive_file_is_low_global(struct zone *zone)
1745 {
1746         unsigned long active, inactive;
1747
1748         active = zone_page_state(zone, NR_ACTIVE_FILE);
1749         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1750
1751         return (active > inactive);
1752 }
1753
1754 /**
1755  * inactive_file_is_low - check if file pages need to be deactivated
1756  * @zone: zone to check
1757  * @sc:   scan control of this context
1758  *
1759  * When the system is doing streaming IO, memory pressure here
1760  * ensures that active file pages get deactivated, until more
1761  * than half of the file pages are on the inactive list.
1762  *
1763  * Once we get to that situation, protect the system's working
1764  * set from being evicted by disabling active file page aging.
1765  *
1766  * This uses a different ratio than the anonymous pages, because
1767  * the page cache uses a use-once replacement algorithm.
1768  */
1769 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1770 {
1771         int low;
1772
1773         if (scanning_global_lru(sc))
1774                 low = inactive_file_is_low_global(zone);
1775         else
1776                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1777         return low;
1778 }
1779
1780 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1781                                 int file)
1782 {
1783         if (file)
1784                 return inactive_file_is_low(zone, sc);
1785         else
1786                 return inactive_anon_is_low(zone, sc);
1787 }
1788
1789 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1790         struct zone *zone, struct scan_control *sc, int priority)
1791 {
1792         int file = is_file_lru(lru);
1793
1794         if (is_active_lru(lru)) {
1795                 if (inactive_list_is_low(zone, sc, file))
1796                     shrink_active_list(nr_to_scan, zone, sc, priority, file);
1797                 return 0;
1798         }
1799
1800         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1801 }
1802
1803 static int vmscan_swappiness(struct scan_control *sc)
1804 {
1805         if (scanning_global_lru(sc))
1806                 return vm_swappiness;
1807         return mem_cgroup_swappiness(sc->mem_cgroup);
1808 }
1809
1810 /*
1811  * Determine how aggressively the anon and file LRU lists should be
1812  * scanned.  The relative value of each set of LRU lists is determined
1813  * by looking at the fraction of the pages scanned we did rotate back
1814  * onto the active list instead of evict.
1815  *
1816  * nr[0] = anon pages to scan; nr[1] = file pages to scan
1817  */
1818 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1819                                         unsigned long *nr, int priority)
1820 {
1821         unsigned long anon, file, free;
1822         unsigned long anon_prio, file_prio;
1823         unsigned long ap, fp;
1824         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1825         u64 fraction[2], denominator;
1826         enum lru_list l;
1827         int noswap = 0;
1828         bool force_scan = false;
1829
1830         /*
1831          * If the zone or memcg is small, nr[l] can be 0.  This
1832          * results in no scanning on this priority and a potential
1833          * priority drop.  Global direct reclaim can go to the next
1834          * zone and tends to have no problems. Global kswapd is for
1835          * zone balancing and it needs to scan a minimum amount. When
1836          * reclaiming for a memcg, a priority drop can cause high
1837          * latencies, so it's better to scan a minimum amount there as
1838          * well.
1839          */
1840         if (scanning_global_lru(sc) && current_is_kswapd())
1841                 force_scan = true;
1842         if (!scanning_global_lru(sc))
1843                 force_scan = true;
1844
1845         /* If we have no swap space, do not bother scanning anon pages. */
1846         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1847                 noswap = 1;
1848                 fraction[0] = 0;
1849                 fraction[1] = 1;
1850                 denominator = 1;
1851                 goto out;
1852         }
1853
1854         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1855                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1856         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1857                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1858
1859         if (scanning_global_lru(sc)) {
1860                 free  = zone_page_state(zone, NR_FREE_PAGES);
1861                 /* If we have very few page cache pages,
1862                    force-scan anon pages. */
1863                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1864                         fraction[0] = 1;
1865                         fraction[1] = 0;
1866                         denominator = 1;
1867                         goto out;
1868                 }
1869         }
1870
1871         /*
1872          * With swappiness at 100, anonymous and file have the same priority.
1873          * This scanning priority is essentially the inverse of IO cost.
1874          */
1875         anon_prio = vmscan_swappiness(sc);
1876         file_prio = 200 - vmscan_swappiness(sc);
1877
1878         /*
1879          * OK, so we have swap space and a fair amount of page cache
1880          * pages.  We use the recently rotated / recently scanned
1881          * ratios to determine how valuable each cache is.
1882          *
1883          * Because workloads change over time (and to avoid overflow)
1884          * we keep these statistics as a floating average, which ends
1885          * up weighing recent references more than old ones.
1886          *
1887          * anon in [0], file in [1]
1888          */
1889         spin_lock_irq(&zone->lru_lock);
1890         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1891                 reclaim_stat->recent_scanned[0] /= 2;
1892                 reclaim_stat->recent_rotated[0] /= 2;
1893         }
1894
1895         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1896                 reclaim_stat->recent_scanned[1] /= 2;
1897                 reclaim_stat->recent_rotated[1] /= 2;
1898         }
1899
1900         /*
1901          * The amount of pressure on anon vs file pages is inversely
1902          * proportional to the fraction of recently scanned pages on
1903          * each list that were recently referenced and in active use.
1904          */
1905         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1906         ap /= reclaim_stat->recent_rotated[0] + 1;
1907
1908         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1909         fp /= reclaim_stat->recent_rotated[1] + 1;
1910         spin_unlock_irq(&zone->lru_lock);
1911
1912         fraction[0] = ap;
1913         fraction[1] = fp;
1914         denominator = ap + fp + 1;
1915 out:
1916         for_each_evictable_lru(l) {
1917                 int file = is_file_lru(l);
1918                 unsigned long scan;
1919
1920                 scan = zone_nr_lru_pages(zone, sc, l);
1921                 if (priority || noswap) {
1922                         scan >>= priority;
1923                         if (!scan && force_scan)
1924                                 scan = SWAP_CLUSTER_MAX;
1925                         scan = div64_u64(scan * fraction[file], denominator);
1926                 }
1927                 nr[l] = scan;
1928         }
1929 }
1930
1931 /*
1932  * Reclaim/compaction depends on a number of pages being freed. To avoid
1933  * disruption to the system, a small number of order-0 pages continue to be
1934  * rotated and reclaimed in the normal fashion. However, by the time we get
1935  * back to the allocator and call try_to_compact_zone(), we ensure that
1936  * there are enough free pages for it to be likely successful
1937  */
1938 static inline bool should_continue_reclaim(struct zone *zone,
1939                                         unsigned long nr_reclaimed,
1940                                         unsigned long nr_scanned,
1941                                         struct scan_control *sc)
1942 {
1943         unsigned long pages_for_compaction;
1944         unsigned long inactive_lru_pages;
1945
1946         /* If not in reclaim/compaction mode, stop */
1947         if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1948                 return false;
1949
1950         /* Consider stopping depending on scan and reclaim activity */
1951         if (sc->gfp_mask & __GFP_REPEAT) {
1952                 /*
1953                  * For __GFP_REPEAT allocations, stop reclaiming if the
1954                  * full LRU list has been scanned and we are still failing
1955                  * to reclaim pages. This full LRU scan is potentially
1956                  * expensive but a __GFP_REPEAT caller really wants to succeed
1957                  */
1958                 if (!nr_reclaimed && !nr_scanned)
1959                         return false;
1960         } else {
1961                 /*
1962                  * For non-__GFP_REPEAT allocations which can presumably
1963                  * fail without consequence, stop if we failed to reclaim
1964                  * any pages from the last SWAP_CLUSTER_MAX number of
1965                  * pages that were scanned. This will return to the
1966                  * caller faster at the risk reclaim/compaction and
1967                  * the resulting allocation attempt fails
1968                  */
1969                 if (!nr_reclaimed)
1970                         return false;
1971         }
1972
1973         /*
1974          * If we have not reclaimed enough pages for compaction and the
1975          * inactive lists are large enough, continue reclaiming
1976          */
1977         pages_for_compaction = (2UL << sc->order);
1978         inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
1979                                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1980         if (sc->nr_reclaimed < pages_for_compaction &&
1981                         inactive_lru_pages > pages_for_compaction)
1982                 return true;
1983
1984         /* If compaction would go ahead or the allocation would succeed, stop */
1985         switch (compaction_suitable(zone, sc->order)) {
1986         case COMPACT_PARTIAL:
1987         case COMPACT_CONTINUE:
1988                 return false;
1989         default:
1990                 return true;
1991         }
1992 }
1993
1994 /*
1995  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1996  */
1997 static void shrink_zone(int priority, struct zone *zone,
1998                                 struct scan_control *sc)
1999 {
2000         unsigned long nr[NR_LRU_LISTS];
2001         unsigned long nr_to_scan;
2002         enum lru_list l;
2003         unsigned long nr_reclaimed, nr_scanned;
2004         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2005         struct blk_plug plug;
2006
2007 restart:
2008         nr_reclaimed = 0;
2009         nr_scanned = sc->nr_scanned;
2010         get_scan_count(zone, sc, nr, priority);
2011
2012         blk_start_plug(&plug);
2013         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2014                                         nr[LRU_INACTIVE_FILE]) {
2015                 for_each_evictable_lru(l) {
2016                         if (nr[l]) {
2017                                 nr_to_scan = min_t(unsigned long,
2018                                                    nr[l], SWAP_CLUSTER_MAX);
2019                                 nr[l] -= nr_to_scan;
2020
2021                                 nr_reclaimed += shrink_list(l, nr_to_scan,
2022                                                             zone, sc, priority);
2023                         }
2024                 }
2025                 /*
2026                  * On large memory systems, scan >> priority can become
2027                  * really large. This is fine for the starting priority;
2028                  * we want to put equal scanning pressure on each zone.
2029                  * However, if the VM has a harder time of freeing pages,
2030                  * with multiple processes reclaiming pages, the total
2031                  * freeing target can get unreasonably large.
2032                  */
2033                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
2034                         break;
2035         }
2036         blk_finish_plug(&plug);
2037         sc->nr_reclaimed += nr_reclaimed;
2038
2039         /*
2040          * Even if we did not try to evict anon pages at all, we want to
2041          * rebalance the anon lru active/inactive ratio.
2042          */
2043         if (inactive_anon_is_low(zone, sc))
2044                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
2045
2046         /* reclaim/compaction might need reclaim to continue */
2047         if (should_continue_reclaim(zone, nr_reclaimed,
2048                                         sc->nr_scanned - nr_scanned, sc))
2049                 goto restart;
2050
2051         throttle_vm_writeout(sc->gfp_mask);
2052 }
2053
2054 /*
2055  * This is the direct reclaim path, for page-allocating processes.  We only
2056  * try to reclaim pages from zones which will satisfy the caller's allocation
2057  * request.
2058  *
2059  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2060  * Because:
2061  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2062  *    allocation or
2063  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2064  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2065  *    zone defense algorithm.
2066  *
2067  * If a zone is deemed to be full of pinned pages then just give it a light
2068  * scan then give up on it.
2069  */
2070 static void shrink_zones(int priority, struct zonelist *zonelist,
2071                                         struct scan_control *sc)
2072 {
2073         struct zoneref *z;
2074         struct zone *zone;
2075         unsigned long nr_soft_reclaimed;
2076         unsigned long nr_soft_scanned;
2077
2078         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2079                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2080                 if (!populated_zone(zone))
2081                         continue;
2082                 /*
2083                  * Take care memory controller reclaiming has small influence
2084                  * to global LRU.
2085                  */
2086                 if (scanning_global_lru(sc)) {
2087                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2088                                 continue;
2089                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2090                                 continue;       /* Let kswapd poll it */
2091                         /*
2092                          * This steals pages from memory cgroups over softlimit
2093                          * and returns the number of reclaimed pages and
2094                          * scanned pages. This works for global memory pressure
2095                          * and balancing, not for a memcg's limit.
2096                          */
2097                         nr_soft_scanned = 0;
2098                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2099                                                 sc->order, sc->gfp_mask,
2100                                                 &nr_soft_scanned);
2101                         sc->nr_reclaimed += nr_soft_reclaimed;
2102                         sc->nr_scanned += nr_soft_scanned;
2103                         /* need some check for avoid more shrink_zone() */
2104                 }
2105
2106                 shrink_zone(priority, zone, sc);
2107         }
2108 }
2109
2110 static bool zone_reclaimable(struct zone *zone)
2111 {
2112         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2113 }
2114
2115 /* All zones in zonelist are unreclaimable? */
2116 static bool all_unreclaimable(struct zonelist *zonelist,
2117                 struct scan_control *sc)
2118 {
2119         struct zoneref *z;
2120         struct zone *zone;
2121
2122         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2123                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2124                 if (!populated_zone(zone))
2125                         continue;
2126                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2127                         continue;
2128                 if (!zone->all_unreclaimable)
2129                         return false;
2130         }
2131
2132         return true;
2133 }
2134
2135 /*
2136  * This is the main entry point to direct page reclaim.
2137  *
2138  * If a full scan of the inactive list fails to free enough memory then we
2139  * are "out of memory" and something needs to be killed.
2140  *
2141  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2142  * high - the zone may be full of dirty or under-writeback pages, which this
2143  * caller can't do much about.  We kick the writeback threads and take explicit
2144  * naps in the hope that some of these pages can be written.  But if the
2145  * allocating task holds filesystem locks which prevent writeout this might not
2146  * work, and the allocation attempt will fail.
2147  *
2148  * returns:     0, if no pages reclaimed
2149  *              else, the number of pages reclaimed
2150  */
2151 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2152                                         struct scan_control *sc,
2153                                         struct shrink_control *shrink)
2154 {
2155         int priority;
2156         unsigned long total_scanned = 0;
2157         struct reclaim_state *reclaim_state = current->reclaim_state;
2158         struct zoneref *z;
2159         struct zone *zone;
2160         unsigned long writeback_threshold;
2161
2162         get_mems_allowed();
2163         delayacct_freepages_start();
2164
2165         if (scanning_global_lru(sc))
2166                 count_vm_event(ALLOCSTALL);
2167
2168         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2169                 sc->nr_scanned = 0;
2170                 if (!priority)
2171                         disable_swap_token(sc->mem_cgroup);
2172                 shrink_zones(priority, zonelist, sc);
2173                 /*
2174                  * Don't shrink slabs when reclaiming memory from
2175                  * over limit cgroups
2176                  */
2177                 if (scanning_global_lru(sc)) {
2178                         unsigned long lru_pages = 0;
2179                         for_each_zone_zonelist(zone, z, zonelist,
2180                                         gfp_zone(sc->gfp_mask)) {
2181                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2182                                         continue;
2183
2184                                 lru_pages += zone_reclaimable_pages(zone);
2185                         }
2186
2187                         shrink_slab(shrink, sc->nr_scanned, lru_pages);
2188                         if (reclaim_state) {
2189                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2190                                 reclaim_state->reclaimed_slab = 0;
2191                         }
2192                 }
2193                 total_scanned += sc->nr_scanned;
2194                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2195                         goto out;
2196
2197                 /*
2198                  * Try to write back as many pages as we just scanned.  This
2199                  * tends to cause slow streaming writers to write data to the
2200                  * disk smoothly, at the dirtying rate, which is nice.   But
2201                  * that's undesirable in laptop mode, where we *want* lumpy
2202                  * writeout.  So in laptop mode, write out the whole world.
2203                  */
2204                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2205                 if (total_scanned > writeback_threshold) {
2206                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
2207                         sc->may_writepage = 1;
2208                 }
2209
2210                 /* Take a nap, wait for some writeback to complete */
2211                 if (!sc->hibernation_mode && sc->nr_scanned &&
2212                     priority < DEF_PRIORITY - 2) {
2213                         struct zone *preferred_zone;
2214
2215                         first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2216                                                 &cpuset_current_mems_allowed,
2217                                                 &preferred_zone);
2218                         wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2219                 }
2220         }
2221
2222 out:
2223         delayacct_freepages_end();
2224         put_mems_allowed();
2225
2226         if (sc->nr_reclaimed)
2227                 return sc->nr_reclaimed;
2228
2229         /*
2230          * As hibernation is going on, kswapd is freezed so that it can't mark
2231          * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2232          * check.
2233          */
2234         if (oom_killer_disabled)
2235                 return 0;
2236
2237         /* top priority shrink_zones still had more to do? don't OOM, then */
2238         if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2239                 return 1;
2240
2241         return 0;
2242 }
2243
2244 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2245                                 gfp_t gfp_mask, nodemask_t *nodemask)
2246 {
2247         unsigned long nr_reclaimed;
2248         struct scan_control sc = {
2249                 .gfp_mask = gfp_mask,
2250                 .may_writepage = !laptop_mode,
2251                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2252                 .may_unmap = 1,
2253                 .may_swap = 1,
2254                 .order = order,
2255                 .mem_cgroup = NULL,
2256                 .nodemask = nodemask,
2257         };
2258         struct shrink_control shrink = {
2259                 .gfp_mask = sc.gfp_mask,
2260         };
2261
2262         trace_mm_vmscan_direct_reclaim_begin(order,
2263                                 sc.may_writepage,
2264                                 gfp_mask);
2265
2266         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2267
2268         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2269
2270         return nr_reclaimed;
2271 }
2272
2273 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2274
2275 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2276                                                 gfp_t gfp_mask, bool noswap,
2277                                                 struct zone *zone,
2278                                                 unsigned long *nr_scanned)
2279 {
2280         struct scan_control sc = {
2281                 .nr_scanned = 0,
2282                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2283                 .may_writepage = !laptop_mode,
2284                 .may_unmap = 1,
2285                 .may_swap = !noswap,
2286                 .order = 0,
2287                 .mem_cgroup = mem,
2288         };
2289
2290         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2291                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2292
2293         trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2294                                                       sc.may_writepage,
2295                                                       sc.gfp_mask);
2296
2297         /*
2298          * NOTE: Although we can get the priority field, using it
2299          * here is not a good idea, since it limits the pages we can scan.
2300          * if we don't reclaim here, the shrink_zone from balance_pgdat
2301          * will pick up pages from other mem cgroup's as well. We hack
2302          * the priority and make it zero.
2303          */
2304         shrink_zone(0, zone, &sc);
2305
2306         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2307
2308         *nr_scanned = sc.nr_scanned;
2309         return sc.nr_reclaimed;
2310 }
2311
2312 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2313                                            gfp_t gfp_mask,
2314                                            bool noswap)
2315 {
2316         struct zonelist *zonelist;
2317         unsigned long nr_reclaimed;
2318         int nid;
2319         struct scan_control sc = {
2320                 .may_writepage = !laptop_mode,
2321                 .may_unmap = 1,
2322                 .may_swap = !noswap,
2323                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2324                 .order = 0,
2325                 .mem_cgroup = mem_cont,
2326                 .nodemask = NULL, /* we don't care the placement */
2327                 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2328                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2329         };
2330         struct shrink_control shrink = {
2331                 .gfp_mask = sc.gfp_mask,
2332         };
2333
2334         /*
2335          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2336          * take care of from where we get pages. So the node where we start the
2337          * scan does not need to be the current node.
2338          */
2339         nid = mem_cgroup_select_victim_node(mem_cont);
2340
2341         zonelist = NODE_DATA(nid)->node_zonelists;
2342
2343         trace_mm_vmscan_memcg_reclaim_begin(0,
2344                                             sc.may_writepage,
2345                                             sc.gfp_mask);
2346
2347         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2348
2349         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2350
2351         return nr_reclaimed;
2352 }
2353 #endif
2354
2355 /*
2356  * pgdat_balanced is used when checking if a node is balanced for high-order
2357  * allocations. Only zones that meet watermarks and are in a zone allowed
2358  * by the callers classzone_idx are added to balanced_pages. The total of
2359  * balanced pages must be at least 25% of the zones allowed by classzone_idx
2360  * for the node to be considered balanced. Forcing all zones to be balanced
2361  * for high orders can cause excessive reclaim when there are imbalanced zones.
2362  * The choice of 25% is due to
2363  *   o a 16M DMA zone that is balanced will not balance a zone on any
2364  *     reasonable sized machine
2365  *   o On all other machines, the top zone must be at least a reasonable
2366  *     percentage of the middle zones. For example, on 32-bit x86, highmem
2367  *     would need to be at least 256M for it to be balance a whole node.
2368  *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2369  *     to balance a node on its own. These seemed like reasonable ratios.
2370  */
2371 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2372                                                 int classzone_idx)
2373 {
2374         unsigned long present_pages = 0;
2375         int i;
2376
2377         for (i = 0; i <= classzone_idx; i++)
2378                 present_pages += pgdat->node_zones[i].present_pages;
2379
2380         /* A special case here: if zone has no page, we think it's balanced */
2381         return balanced_pages >= (present_pages >> 2);
2382 }
2383
2384 /* is kswapd sleeping prematurely? */
2385 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2386                                         int classzone_idx)
2387 {
2388         int i;
2389         unsigned long balanced = 0;
2390         bool all_zones_ok = true;
2391
2392         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2393         if (remaining)
2394                 return true;
2395
2396         /* Check the watermark levels */
2397         for (i = 0; i <= classzone_idx; i++) {
2398                 struct zone *zone = pgdat->node_zones + i;
2399
2400                 if (!populated_zone(zone))
2401                         continue;
2402
2403                 /*
2404                  * balance_pgdat() skips over all_unreclaimable after
2405                  * DEF_PRIORITY. Effectively, it considers them balanced so
2406                  * they must be considered balanced here as well if kswapd
2407                  * is to sleep
2408                  */
2409                 if (zone->all_unreclaimable) {
2410                         balanced += zone->present_pages;
2411                         continue;
2412                 }
2413
2414                 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2415                                                         i, 0))
2416                         all_zones_ok = false;
2417                 else
2418                         balanced += zone->present_pages;
2419         }
2420
2421         /*
2422          * For high-order requests, the balanced zones must contain at least
2423          * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2424          * must be balanced
2425          */
2426         if (order)
2427                 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2428         else
2429                 return !all_zones_ok;
2430 }
2431
2432 /*
2433  * For kswapd, balance_pgdat() will work across all this node's zones until
2434  * they are all at high_wmark_pages(zone).
2435  *
2436  * Returns the final order kswapd was reclaiming at
2437  *
2438  * There is special handling here for zones which are full of pinned pages.
2439  * This can happen if the pages are all mlocked, or if they are all used by
2440  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2441  * What we do is to detect the case where all pages in the zone have been
2442  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2443  * dead and from now on, only perform a short scan.  Basically we're polling
2444  * the zone for when the problem goes away.
2445  *
2446  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2447  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2448  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2449  * lower zones regardless of the number of free pages in the lower zones. This
2450  * interoperates with the page allocator fallback scheme to ensure that aging
2451  * of pages is balanced across the zones.
2452  */
2453 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2454                                                         int *classzone_idx)
2455 {
2456         int all_zones_ok;
2457         unsigned long balanced;
2458         int priority;
2459         int i;
2460         int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2461         unsigned long total_scanned;
2462         struct reclaim_state *reclaim_state = current->reclaim_state;
2463         unsigned long nr_soft_reclaimed;
2464         unsigned long nr_soft_scanned;
2465         struct scan_control sc = {
2466                 .gfp_mask = GFP_KERNEL,
2467                 .may_unmap = 1,
2468                 .may_swap = 1,
2469                 /*
2470                  * kswapd doesn't want to be bailed out while reclaim. because
2471                  * we want to put equal scanning pressure on each zone.
2472                  */
2473                 .nr_to_reclaim = ULONG_MAX,
2474                 .order = order,
2475                 .mem_cgroup = NULL,
2476         };
2477         struct shrink_control shrink = {
2478                 .gfp_mask = sc.gfp_mask,
2479         };
2480 loop_again:
2481         total_scanned = 0;
2482         sc.nr_reclaimed = 0;
2483         sc.may_writepage = !laptop_mode;
2484         count_vm_event(PAGEOUTRUN);
2485
2486         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2487                 unsigned long lru_pages = 0;
2488                 int has_under_min_watermark_zone = 0;
2489
2490                 /* The swap token gets in the way of swapout... */
2491                 if (!priority)
2492                         disable_swap_token(NULL);
2493
2494                 all_zones_ok = 1;
2495                 balanced = 0;
2496
2497                 /*
2498                  * Scan in the highmem->dma direction for the highest
2499                  * zone which needs scanning
2500                  */
2501                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2502                         struct zone *zone = pgdat->node_zones + i;
2503
2504                         if (!populated_zone(zone))
2505                                 continue;
2506
2507                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2508                                 continue;
2509
2510                         /*
2511                          * Do some background aging of the anon list, to give
2512                          * pages a chance to be referenced before reclaiming.
2513                          */
2514                         if (inactive_anon_is_low(zone, &sc))
2515                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2516                                                         &sc, priority, 0);
2517
2518                         if (!zone_watermark_ok_safe(zone, order,
2519                                         high_wmark_pages(zone), 0, 0)) {
2520                                 end_zone = i;
2521                                 break;
2522                         } else {
2523                                 /* If balanced, clear the congested flag */
2524                                 zone_clear_flag(zone, ZONE_CONGESTED);
2525                         }
2526                 }
2527                 if (i < 0)
2528                         goto out;
2529
2530                 for (i = 0; i <= end_zone; i++) {
2531                         struct zone *zone = pgdat->node_zones + i;
2532
2533                         lru_pages += zone_reclaimable_pages(zone);
2534                 }
2535
2536                 /*
2537                  * Now scan the zone in the dma->highmem direction, stopping
2538                  * at the last zone which needs scanning.
2539                  *
2540                  * We do this because the page allocator works in the opposite
2541                  * direction.  This prevents the page allocator from allocating
2542                  * pages behind kswapd's direction of progress, which would
2543                  * cause too much scanning of the lower zones.
2544                  */
2545                 for (i = 0; i <= end_zone; i++) {
2546                         struct zone *zone = pgdat->node_zones + i;
2547                         int nr_slab;
2548                         unsigned long balance_gap;
2549
2550                         if (!populated_zone(zone))
2551                                 continue;
2552
2553                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2554                                 continue;
2555
2556                         sc.nr_scanned = 0;
2557
2558                         nr_soft_scanned = 0;
2559                         /*
2560                          * Call soft limit reclaim before calling shrink_zone.
2561                          */
2562                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2563                                                         order, sc.gfp_mask,
2564                                                         &nr_soft_scanned);
2565                         sc.nr_reclaimed += nr_soft_reclaimed;
2566                         total_scanned += nr_soft_scanned;
2567
2568                         /*
2569                          * We put equal pressure on every zone, unless
2570                          * one zone has way too many pages free
2571                          * already. The "too many pages" is defined
2572                          * as the high wmark plus a "gap" where the
2573                          * gap is either the low watermark or 1%
2574                          * of the zone, whichever is smaller.
2575                          */
2576                         balance_gap = min(low_wmark_pages(zone),
2577                                 (zone->present_pages +
2578                                         KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2579                                 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2580                         if (!zone_watermark_ok_safe(zone, order,
2581                                         high_wmark_pages(zone) + balance_gap,
2582                                         end_zone, 0)) {
2583                                 shrink_zone(priority, zone, &sc);
2584
2585                                 reclaim_state->reclaimed_slab = 0;
2586                                 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2587                                 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2588                                 total_scanned += sc.nr_scanned;
2589
2590                                 if (nr_slab == 0 && !zone_reclaimable(zone))
2591                                         zone->all_unreclaimable = 1;
2592                         }
2593
2594                         /*
2595                          * If we've done a decent amount of scanning and
2596                          * the reclaim ratio is low, start doing writepage
2597                          * even in laptop mode
2598                          */
2599                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2600                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2601                                 sc.may_writepage = 1;
2602
2603                         if (zone->all_unreclaimable) {
2604                                 if (end_zone && end_zone == i)
2605                                         end_zone--;
2606                                 continue;
2607                         }
2608
2609                         if (!zone_watermark_ok_safe(zone, order,
2610                                         high_wmark_pages(zone), end_zone, 0)) {
2611                                 all_zones_ok = 0;
2612                                 /*
2613                                  * We are still under min water mark.  This
2614                                  * means that we have a GFP_ATOMIC allocation
2615                                  * failure risk. Hurry up!
2616                                  */
2617                                 if (!zone_watermark_ok_safe(zone, order,
2618                                             min_wmark_pages(zone), end_zone, 0))
2619                                         has_under_min_watermark_zone = 1;
2620                         } else {
2621                                 /*
2622                                  * If a zone reaches its high watermark,
2623                                  * consider it to be no longer congested. It's
2624                                  * possible there are dirty pages backed by
2625                                  * congested BDIs but as pressure is relieved,
2626                                  * spectulatively avoid congestion waits
2627                                  */
2628                                 zone_clear_flag(zone, ZONE_CONGESTED);
2629                                 if (i <= *classzone_idx)
2630                                         balanced += zone->present_pages;
2631                         }
2632
2633                 }
2634                 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2635                         break;          /* kswapd: all done */
2636                 /*
2637                  * OK, kswapd is getting into trouble.  Take a nap, then take
2638                  * another pass across the zones.
2639                  */
2640                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2641                         if (has_under_min_watermark_zone)
2642                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2643                         else
2644                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2645                 }
2646
2647                 /*
2648                  * We do this so kswapd doesn't build up large priorities for
2649                  * example when it is freeing in parallel with allocators. It
2650                  * matches the direct reclaim path behaviour in terms of impact
2651                  * on zone->*_priority.
2652                  */
2653                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2654                         break;
2655         }
2656 out:
2657
2658         /*
2659          * order-0: All zones must meet high watermark for a balanced node
2660          * high-order: Balanced zones must make up at least 25% of the node
2661          *             for the node to be balanced
2662          */
2663         if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2664                 cond_resched();
2665
2666                 try_to_freeze();
2667
2668                 /*
2669                  * Fragmentation may mean that the system cannot be
2670                  * rebalanced for high-order allocations in all zones.
2671                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2672                  * it means the zones have been fully scanned and are still
2673                  * not balanced. For high-order allocations, there is
2674                  * little point trying all over again as kswapd may
2675                  * infinite loop.
2676                  *
2677                  * Instead, recheck all watermarks at order-0 as they
2678                  * are the most important. If watermarks are ok, kswapd will go
2679                  * back to sleep. High-order users can still perform direct
2680                  * reclaim if they wish.
2681                  */
2682                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2683                         order = sc.order = 0;
2684
2685                 goto loop_again;
2686         }
2687
2688         /*
2689          * If kswapd was reclaiming at a higher order, it has the option of
2690          * sleeping without all zones being balanced. Before it does, it must
2691          * ensure that the watermarks for order-0 on *all* zones are met and
2692          * that the congestion flags are cleared. The congestion flag must
2693          * be cleared as kswapd is the only mechanism that clears the flag
2694          * and it is potentially going to sleep here.
2695          */
2696         if (order) {
2697                 for (i = 0; i <= end_zone; i++) {
2698                         struct zone *zone = pgdat->node_zones + i;
2699
2700                         if (!populated_zone(zone))
2701                                 continue;
2702
2703                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2704                                 continue;
2705
2706                         /* Confirm the zone is balanced for order-0 */
2707                         if (!zone_watermark_ok(zone, 0,
2708                                         high_wmark_pages(zone), 0, 0)) {
2709                                 order = sc.order = 0;
2710                                 goto loop_again;
2711                         }
2712
2713                         /* If balanced, clear the congested flag */
2714                         zone_clear_flag(zone, ZONE_CONGESTED);
2715                 }
2716         }
2717
2718         /*
2719          * Return the order we were reclaiming at so sleeping_prematurely()
2720          * makes a decision on the order we were last reclaiming at. However,
2721          * if another caller entered the allocator slow path while kswapd
2722          * was awake, order will remain at the higher level
2723          */
2724         *classzone_idx = end_zone;
2725         return order;
2726 }
2727
2728 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2729 {
2730         long remaining = 0;
2731         DEFINE_WAIT(wait);
2732
2733         if (freezing(current) || kthread_should_stop())
2734                 return;
2735
2736         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2737
2738         /* Try to sleep for a short interval */
2739         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2740                 remaining = schedule_timeout(HZ/10);
2741                 finish_wait(&pgdat->kswapd_wait, &wait);
2742                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2743         }
2744
2745         /*
2746          * After a short sleep, check if it was a premature sleep. If not, then
2747          * go fully to sleep until explicitly woken up.
2748          */
2749         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2750                 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2751
2752                 /*
2753                  * vmstat counters are not perfectly accurate and the estimated
2754                  * value for counters such as NR_FREE_PAGES can deviate from the
2755                  * true value by nr_online_cpus * threshold. To avoid the zone
2756                  * watermarks being breached while under pressure, we reduce the
2757                  * per-cpu vmstat threshold while kswapd is awake and restore
2758                  * them before going back to sleep.
2759                  */
2760                 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2761                 schedule();
2762                 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2763         } else {
2764                 if (remaining)
2765                         count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2766                 else
2767                         count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2768         }
2769         finish_wait(&pgdat->kswapd_wait, &wait);
2770 }
2771
2772 /*
2773  * The background pageout daemon, started as a kernel thread
2774  * from the init process.
2775  *
2776  * This basically trickles out pages so that we have _some_
2777  * free memory available even if there is no other activity
2778  * that frees anything up. This is needed for things like routing
2779  * etc, where we otherwise might have all activity going on in
2780  * asynchronous contexts that cannot page things out.
2781  *
2782  * If there are applications that are active memory-allocators
2783  * (most normal use), this basically shouldn't matter.
2784  */
2785 static int kswapd(void *p)
2786 {
2787         unsigned long order, new_order;
2788         int classzone_idx, new_classzone_idx;
2789         pg_data_t *pgdat = (pg_data_t*)p;
2790         struct task_struct *tsk = current;
2791
2792         struct reclaim_state reclaim_state = {
2793                 .reclaimed_slab = 0,
2794         };
2795         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2796
2797         lockdep_set_current_reclaim_state(GFP_KERNEL);
2798
2799         if (!cpumask_empty(cpumask))
2800                 set_cpus_allowed_ptr(tsk, cpumask);
2801         current->reclaim_state = &reclaim_state;
2802
2803         /*
2804          * Tell the memory management that we're a "memory allocator",
2805          * and that if we need more memory we should get access to it
2806          * regardless (see "__alloc_pages()"). "kswapd" should
2807          * never get caught in the normal page freeing logic.
2808          *
2809          * (Kswapd normally doesn't need memory anyway, but sometimes
2810          * you need a small amount of memory in order to be able to
2811          * page out something else, and this flag essentially protects
2812          * us from recursively trying to free more memory as we're
2813          * trying to free the first piece of memory in the first place).
2814          */
2815         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2816         set_freezable();
2817
2818         order = new_order = 0;
2819         classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2820         for ( ; ; ) {
2821                 int ret;
2822
2823                 /*
2824                  * If the last balance_pgdat was unsuccessful it's unlikely a
2825                  * new request of a similar or harder type will succeed soon
2826                  * so consider going to sleep on the basis we reclaimed at
2827                  */
2828                 if (classzone_idx >= new_classzone_idx && order == new_order) {
2829                         new_order = pgdat->kswapd_max_order;
2830                         new_classzone_idx = pgdat->classzone_idx;
2831                         pgdat->kswapd_max_order =  0;
2832                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2833                 }
2834
2835                 if (order < new_order || classzone_idx > new_classzone_idx) {
2836                         /*
2837                          * Don't sleep if someone wants a larger 'order'
2838                          * allocation or has tigher zone constraints
2839                          */
2840                         order = new_order;
2841                         classzone_idx = new_classzone_idx;
2842                 } else {
2843                         kswapd_try_to_sleep(pgdat, order, classzone_idx);
2844                         order = pgdat->kswapd_max_order;
2845                         classzone_idx = pgdat->classzone_idx;
2846                         pgdat->kswapd_max_order = 0;
2847                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2848                 }
2849
2850                 ret = try_to_freeze();
2851                 if (kthread_should_stop())
2852                         break;
2853
2854                 /*
2855                  * We can speed up thawing tasks if we don't call balance_pgdat
2856                  * after returning from the refrigerator
2857                  */
2858                 if (!ret) {
2859                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2860                         order = balance_pgdat(pgdat, order, &classzone_idx);
2861                 }
2862         }
2863         return 0;
2864 }
2865
2866 /*
2867  * A zone is low on free memory, so wake its kswapd task to service it.
2868  */
2869 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2870 {
2871         pg_data_t *pgdat;
2872
2873         if (!populated_zone(zone))
2874                 return;
2875
2876         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2877                 return;
2878         pgdat = zone->zone_pgdat;
2879         if (pgdat->kswapd_max_order < order) {
2880                 pgdat->kswapd_max_order = order;
2881                 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2882         }
2883         if (!waitqueue_active(&pgdat->kswapd_wait))
2884                 return;
2885         if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2886                 return;
2887
2888         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2889         wake_up_interruptible(&pgdat->kswapd_wait);
2890 }
2891
2892 /*
2893  * The reclaimable count would be mostly accurate.
2894  * The less reclaimable pages may be
2895  * - mlocked pages, which will be moved to unevictable list when encountered
2896  * - mapped pages, which may require several travels to be reclaimed
2897  * - dirty pages, which is not "instantly" reclaimable
2898  */
2899 unsigned long global_reclaimable_pages(void)
2900 {
2901         int nr;
2902
2903         nr = global_page_state(NR_ACTIVE_FILE) +
2904              global_page_state(NR_INACTIVE_FILE);
2905
2906         if (nr_swap_pages > 0)
2907                 nr += global_page_state(NR_ACTIVE_ANON) +
2908                       global_page_state(NR_INACTIVE_ANON);
2909
2910         return nr;
2911 }
2912
2913 unsigned long zone_reclaimable_pages(struct zone *zone)
2914 {
2915         int nr;
2916
2917         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2918              zone_page_state(zone, NR_INACTIVE_FILE);
2919
2920         if (nr_swap_pages > 0)
2921                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2922                       zone_page_state(zone, NR_INACTIVE_ANON);
2923
2924         return nr;
2925 }
2926
2927 #ifdef CONFIG_HIBERNATION
2928 /*
2929  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2930  * freed pages.
2931  *
2932  * Rather than trying to age LRUs the aim is to preserve the overall
2933  * LRU order by reclaiming preferentially
2934  * inactive > active > active referenced > active mapped
2935  */
2936 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2937 {
2938         struct reclaim_state reclaim_state;
2939         struct scan_control sc = {
2940                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2941                 .may_swap = 1,
2942                 .may_unmap = 1,
2943                 .may_writepage = 1,
2944                 .nr_to_reclaim = nr_to_reclaim,
2945                 .hibernation_mode = 1,
2946                 .order = 0,
2947         };
2948         struct shrink_control shrink = {
2949                 .gfp_mask = sc.gfp_mask,
2950         };
2951         struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2952         struct task_struct *p = current;
2953         unsigned long nr_reclaimed;
2954
2955         p->flags |= PF_MEMALLOC;
2956         lockdep_set_current_reclaim_state(sc.gfp_mask);
2957         reclaim_state.reclaimed_slab = 0;
2958         p->reclaim_state = &reclaim_state;
2959
2960         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2961
2962         p->reclaim_state = NULL;
2963         lockdep_clear_current_reclaim_state();
2964         p->flags &= ~PF_MEMALLOC;
2965
2966         return nr_reclaimed;
2967 }
2968 #endif /* CONFIG_HIBERNATION */
2969
2970 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2971    not required for correctness.  So if the last cpu in a node goes
2972    away, we get changed to run anywhere: as the first one comes back,
2973    restore their cpu bindings. */
2974 static int __devinit cpu_callback(struct notifier_block *nfb,
2975                                   unsigned long action, void *hcpu)
2976 {
2977         int nid;
2978
2979         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2980                 for_each_node_state(nid, N_HIGH_MEMORY) {
2981                         pg_data_t *pgdat = NODE_DATA(nid);
2982                         const struct cpumask *mask;
2983
2984                         mask = cpumask_of_node(pgdat->node_id);
2985
2986                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2987                                 /* One of our CPUs online: restore mask */
2988                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
2989                 }
2990         }
2991         return NOTIFY_OK;
2992 }
2993
2994 /*
2995  * This kswapd start function will be called by init and node-hot-add.
2996  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2997  */
2998 int kswapd_run(int nid)
2999 {
3000         pg_data_t *pgdat = NODE_DATA(nid);
3001         int ret = 0;
3002
3003         if (pgdat->kswapd)
3004                 return 0;
3005
3006         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3007         if (IS_ERR(pgdat->kswapd)) {
3008                 /* failure at boot is fatal */
3009                 BUG_ON(system_state == SYSTEM_BOOTING);
3010                 printk("Failed to start kswapd on node %d\n",nid);
3011                 ret = -1;
3012         }
3013         return ret;
3014 }
3015
3016 /*
3017  * Called by memory hotplug when all memory in a node is offlined.
3018  */
3019 void kswapd_stop(int nid)
3020 {
3021         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3022
3023         if (kswapd)
3024                 kthread_stop(kswapd);
3025 }
3026
3027 static int __init kswapd_init(void)
3028 {
3029         int nid;
3030
3031         swap_setup();
3032         for_each_node_state(nid, N_HIGH_MEMORY)
3033                 kswapd_run(nid);
3034         hotcpu_notifier(cpu_callback, 0);
3035         return 0;
3036 }
3037
3038 module_init(kswapd_init)
3039
3040 #ifdef CONFIG_NUMA
3041 /*
3042  * Zone reclaim mode
3043  *
3044  * If non-zero call zone_reclaim when the number of free pages falls below
3045  * the watermarks.
3046  */
3047 int zone_reclaim_mode __read_mostly;
3048
3049 #define RECLAIM_OFF 0
3050 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
3051 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
3052 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
3053
3054 /*
3055  * Priority for ZONE_RECLAIM. This determines the fraction of pages
3056  * of a node considered for each zone_reclaim. 4 scans 1/16th of
3057  * a zone.
3058  */
3059 #define ZONE_RECLAIM_PRIORITY 4
3060
3061 /*
3062  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3063  * occur.
3064  */
3065 int sysctl_min_unmapped_ratio = 1;
3066
3067 /*
3068  * If the number of slab pages in a zone grows beyond this percentage then
3069  * slab reclaim needs to occur.
3070  */
3071 int sysctl_min_slab_ratio = 5;
3072
3073 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3074 {
3075         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3076         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3077                 zone_page_state(zone, NR_ACTIVE_FILE);
3078
3079         /*
3080          * It's possible for there to be more file mapped pages than
3081          * accounted for by the pages on the file LRU lists because
3082          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3083          */
3084         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3085 }
3086
3087 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3088 static long zone_pagecache_reclaimable(struct zone *zone)
3089 {
3090         long nr_pagecache_reclaimable;
3091         long delta = 0;
3092
3093         /*
3094          * If RECLAIM_SWAP is set, then all file pages are considered
3095          * potentially reclaimable. Otherwise, we have to worry about
3096          * pages like swapcache and zone_unmapped_file_pages() provides
3097          * a better estimate
3098          */
3099         if (zone_reclaim_mode & RECLAIM_SWAP)
3100                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3101         else
3102                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3103
3104         /* If we can't clean pages, remove dirty pages from consideration */
3105         if (!(zone_reclaim_mode & RECLAIM_WRITE))
3106                 delta += zone_page_state(zone, NR_FILE_DIRTY);
3107
3108         /* Watch for any possible underflows due to delta */
3109         if (unlikely(delta > nr_pagecache_reclaimable))
3110                 delta = nr_pagecache_reclaimable;
3111
3112         return nr_pagecache_reclaimable - delta;
3113 }
3114
3115 /*
3116  * Try to free up some pages from this zone through reclaim.
3117  */
3118 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3119 {
3120         /* Minimum pages needed in order to stay on node */
3121         const unsigned long nr_pages = 1 << order;
3122         struct task_struct *p = current;
3123         struct reclaim_state reclaim_state;
3124         int priority;
3125         struct scan_control sc = {
3126                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3127                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3128                 .may_swap = 1,
3129                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3130                                        SWAP_CLUSTER_MAX),
3131                 .gfp_mask = gfp_mask,
3132                 .order = order,
3133         };
3134         struct shrink_control shrink = {