Merge branch 'writeback-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git...
[linux-3.10.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                 if (!PageWriteback(page)) {
499                         /* synchronous write or broken a_ops? */
500                         ClearPageReclaim(page);
501                 }
502                 trace_mm_vmscan_writepage(page,
503                         trace_reclaim_flags(page, sc->reclaim_mode));
504                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
505                 return PAGE_SUCCESS;
506         }
507
508         return PAGE_CLEAN;
509 }
510
511 /*
512  * Same as remove_mapping, but if the page is removed from the mapping, it
513  * gets returned with a refcount of 0.
514  */
515 static int __remove_mapping(struct address_space *mapping, struct page *page)
516 {
517         BUG_ON(!PageLocked(page));
518         BUG_ON(mapping != page_mapping(page));
519
520         spin_lock_irq(&mapping->tree_lock);
521         /*
522          * The non racy check for a busy page.
523          *
524          * Must be careful with the order of the tests. When someone has
525          * a ref to the page, it may be possible that they dirty it then
526          * drop the reference. So if PageDirty is tested before page_count
527          * here, then the following race may occur:
528          *
529          * get_user_pages(&page);
530          * [user mapping goes away]
531          * write_to(page);
532          *                              !PageDirty(page)    [good]
533          * SetPageDirty(page);
534          * put_page(page);
535          *                              !page_count(page)   [good, discard it]
536          *
537          * [oops, our write_to data is lost]
538          *
539          * Reversing the order of the tests ensures such a situation cannot
540          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
541          * load is not satisfied before that of page->_count.
542          *
543          * Note that if SetPageDirty is always performed via set_page_dirty,
544          * and thus under tree_lock, then this ordering is not required.
545          */
546         if (!page_freeze_refs(page, 2))
547                 goto cannot_free;
548         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
549         if (unlikely(PageDirty(page))) {
550                 page_unfreeze_refs(page, 2);
551                 goto cannot_free;
552         }
553
554         if (PageSwapCache(page)) {
555                 swp_entry_t swap = { .val = page_private(page) };
556                 __delete_from_swap_cache(page);
557                 spin_unlock_irq(&mapping->tree_lock);
558                 swapcache_free(swap, page);
559         } else {
560                 void (*freepage)(struct page *);
561
562                 freepage = mapping->a_ops->freepage;
563
564                 __delete_from_page_cache(page);
565                 spin_unlock_irq(&mapping->tree_lock);
566                 mem_cgroup_uncharge_cache_page(page);
567
568                 if (freepage != NULL)
569                         freepage(page);
570         }
571
572         return 1;
573
574 cannot_free:
575         spin_unlock_irq(&mapping->tree_lock);
576         return 0;
577 }
578
579 /*
580  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
581  * someone else has a ref on the page, abort and return 0.  If it was
582  * successfully detached, return 1.  Assumes the caller has a single ref on
583  * this page.
584  */
585 int remove_mapping(struct address_space *mapping, struct page *page)
586 {
587         if (__remove_mapping(mapping, page)) {
588                 /*
589                  * Unfreezing the refcount with 1 rather than 2 effectively
590                  * drops the pagecache ref for us without requiring another
591                  * atomic operation.
592                  */
593                 page_unfreeze_refs(page, 1);
594                 return 1;
595         }
596         return 0;
597 }
598
599 /**
600  * putback_lru_page - put previously isolated page onto appropriate LRU list
601  * @page: page to be put back to appropriate lru list
602  *
603  * Add previously isolated @page to appropriate LRU list.
604  * Page may still be unevictable for other reasons.
605  *
606  * lru_lock must not be held, interrupts must be enabled.
607  */
608 void putback_lru_page(struct page *page)
609 {
610         int lru;
611         int active = !!TestClearPageActive(page);
612         int was_unevictable = PageUnevictable(page);
613
614         VM_BUG_ON(PageLRU(page));
615
616 redo:
617         ClearPageUnevictable(page);
618
619         if (page_evictable(page, NULL)) {
620                 /*
621                  * For evictable pages, we can use the cache.
622                  * In event of a race, worst case is we end up with an
623                  * unevictable page on [in]active list.
624                  * We know how to handle that.
625                  */
626                 lru = active + page_lru_base_type(page);
627                 lru_cache_add_lru(page, lru);
628         } else {
629                 /*
630                  * Put unevictable pages directly on zone's unevictable
631                  * list.
632                  */
633                 lru = LRU_UNEVICTABLE;
634                 add_page_to_unevictable_list(page);
635                 /*
636                  * When racing with an mlock or AS_UNEVICTABLE clearing
637                  * (page is unlocked) make sure that if the other thread
638                  * does not observe our setting of PG_lru and fails
639                  * isolation/check_move_unevictable_page,
640                  * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
641                  * the page back to the evictable list.
642                  *
643                  * The other side is TestClearPageMlocked() or shmem_lock().
644                  */
645                 smp_mb();
646         }
647
648         /*
649          * page's status can change while we move it among lru. If an evictable
650          * page is on unevictable list, it never be freed. To avoid that,
651          * check after we added it to the list, again.
652          */
653         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
654                 if (!isolate_lru_page(page)) {
655                         put_page(page);
656                         goto redo;
657                 }
658                 /* This means someone else dropped this page from LRU
659                  * So, it will be freed or putback to LRU again. There is
660                  * nothing to do here.
661                  */
662         }
663
664         if (was_unevictable && lru != LRU_UNEVICTABLE)
665                 count_vm_event(UNEVICTABLE_PGRESCUED);
666         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
667                 count_vm_event(UNEVICTABLE_PGCULLED);
668
669         put_page(page);         /* drop ref from isolate */
670 }
671
672 enum page_references {
673         PAGEREF_RECLAIM,
674         PAGEREF_RECLAIM_CLEAN,
675         PAGEREF_KEEP,
676         PAGEREF_ACTIVATE,
677 };
678
679 static enum page_references page_check_references(struct page *page,
680                                                   struct scan_control *sc)
681 {
682         int referenced_ptes, referenced_page;
683         unsigned long vm_flags;
684
685         referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
686         referenced_page = TestClearPageReferenced(page);
687
688         /* Lumpy reclaim - ignore references */
689         if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
690                 return PAGEREF_RECLAIM;
691
692         /*
693          * Mlock lost the isolation race with us.  Let try_to_unmap()
694          * move the page to the unevictable list.
695          */
696         if (vm_flags & VM_LOCKED)
697                 return PAGEREF_RECLAIM;
698
699         if (referenced_ptes) {
700                 if (PageAnon(page))
701                         return PAGEREF_ACTIVATE;
702                 /*
703                  * All mapped pages start out with page table
704                  * references from the instantiating fault, so we need
705                  * to look twice if a mapped file page is used more
706                  * than once.
707                  *
708                  * Mark it and spare it for another trip around the
709                  * inactive list.  Another page table reference will
710                  * lead to its activation.
711                  *
712                  * Note: the mark is set for activated pages as well
713                  * so that recently deactivated but used pages are
714                  * quickly recovered.
715                  */
716                 SetPageReferenced(page);
717
718                 if (referenced_page)
719                         return PAGEREF_ACTIVATE;
720
721                 return PAGEREF_KEEP;
722         }
723
724         /* Reclaim if clean, defer dirty pages to writeback */
725         if (referenced_page && !PageSwapBacked(page))
726                 return PAGEREF_RECLAIM_CLEAN;
727
728         return PAGEREF_RECLAIM;
729 }
730
731 static noinline_for_stack void free_page_list(struct list_head *free_pages)
732 {
733         struct pagevec freed_pvec;
734         struct page *page, *tmp;
735
736         pagevec_init(&freed_pvec, 1);
737
738         list_for_each_entry_safe(page, tmp, free_pages, lru) {
739                 list_del(&page->lru);
740                 if (!pagevec_add(&freed_pvec, page)) {
741                         __pagevec_free(&freed_pvec);
742                         pagevec_reinit(&freed_pvec);
743                 }
744         }
745
746         pagevec_free(&freed_pvec);
747 }
748
749 /*
750  * shrink_page_list() returns the number of reclaimed pages
751  */
752 static unsigned long shrink_page_list(struct list_head *page_list,
753                                       struct zone *zone,
754                                       struct scan_control *sc,
755                                       int priority,
756                                       unsigned long *ret_nr_dirty,
757                                       unsigned long *ret_nr_writeback)
758 {
759         LIST_HEAD(ret_pages);
760         LIST_HEAD(free_pages);
761         int pgactivate = 0;
762         unsigned long nr_dirty = 0;
763         unsigned long nr_congested = 0;
764         unsigned long nr_reclaimed = 0;
765         unsigned long nr_writeback = 0;
766
767         cond_resched();
768
769         while (!list_empty(page_list)) {
770                 enum page_references references;
771                 struct address_space *mapping;
772                 struct page *page;
773                 int may_enter_fs;
774
775                 cond_resched();
776
777                 page = lru_to_page(page_list);
778                 list_del(&page->lru);
779
780                 if (!trylock_page(page))
781                         goto keep;
782
783                 VM_BUG_ON(PageActive(page));
784                 VM_BUG_ON(page_zone(page) != zone);
785
786                 sc->nr_scanned++;
787
788                 if (unlikely(!page_evictable(page, NULL)))
789                         goto cull_mlocked;
790
791                 if (!sc->may_unmap && page_mapped(page))
792                         goto keep_locked;
793
794                 /* Double the slab pressure for mapped and swapcache pages */
795                 if (page_mapped(page) || PageSwapCache(page))
796                         sc->nr_scanned++;
797
798                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
799                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
800
801                 if (PageWriteback(page)) {
802                         nr_writeback++;
803                         /*
804                          * Synchronous reclaim cannot queue pages for
805                          * writeback due to the possibility of stack overflow
806                          * but if it encounters a page under writeback, wait
807                          * for the IO to complete.
808                          */
809                         if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
810                             may_enter_fs)
811                                 wait_on_page_writeback(page);
812                         else {
813                                 unlock_page(page);
814                                 goto keep_lumpy;
815                         }
816                 }
817
818                 references = page_check_references(page, sc);
819                 switch (references) {
820                 case PAGEREF_ACTIVATE:
821                         goto activate_locked;
822                 case PAGEREF_KEEP:
823                         goto keep_locked;
824                 case PAGEREF_RECLAIM:
825                 case PAGEREF_RECLAIM_CLEAN:
826                         ; /* try to reclaim the page below */
827                 }
828
829                 /*
830                  * Anonymous process memory has backing store?
831                  * Try to allocate it some swap space here.
832                  */
833                 if (PageAnon(page) && !PageSwapCache(page)) {
834                         if (!(sc->gfp_mask & __GFP_IO))
835                                 goto keep_locked;
836                         if (!add_to_swap(page))
837                                 goto activate_locked;
838                         may_enter_fs = 1;
839                 }
840
841                 mapping = page_mapping(page);
842
843                 /*
844                  * The page is mapped into the page tables of one or more
845                  * processes. Try to unmap it here.
846                  */
847                 if (page_mapped(page) && mapping) {
848                         switch (try_to_unmap(page, TTU_UNMAP)) {
849                         case SWAP_FAIL:
850                                 goto activate_locked;
851                         case SWAP_AGAIN:
852                                 goto keep_locked;
853                         case SWAP_MLOCK:
854                                 goto cull_mlocked;
855                         case SWAP_SUCCESS:
856                                 ; /* try to free the page below */
857                         }
858                 }
859
860                 if (PageDirty(page)) {
861                         nr_dirty++;
862
863                         /*
864                          * Only kswapd can writeback filesystem pages to
865                          * avoid risk of stack overflow but do not writeback
866                          * unless under significant pressure.
867                          */
868                         if (page_is_file_cache(page) &&
869                                         (!current_is_kswapd() || priority >= DEF_PRIORITY - 2)) {
870                                 /*
871                                  * Immediately reclaim when written back.
872                                  * Similar in principal to deactivate_page()
873                                  * except we already have the page isolated
874                                  * and know it's dirty
875                                  */
876                                 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
877                                 SetPageReclaim(page);
878
879                                 goto keep_locked;
880                         }
881
882                         if (references == PAGEREF_RECLAIM_CLEAN)
883                                 goto keep_locked;
884                         if (!may_enter_fs)
885                                 goto keep_locked;
886                         if (!sc->may_writepage)
887                                 goto keep_locked;
888
889                         /* Page is dirty, try to write it out here */
890                         switch (pageout(page, mapping, sc)) {
891                         case PAGE_KEEP:
892                                 nr_congested++;
893                                 goto keep_locked;
894                         case PAGE_ACTIVATE:
895                                 goto activate_locked;
896                         case PAGE_SUCCESS:
897                                 if (PageWriteback(page))
898                                         goto keep_lumpy;
899                                 if (PageDirty(page))
900                                         goto keep;
901
902                                 /*
903                                  * A synchronous write - probably a ramdisk.  Go
904                                  * ahead and try to reclaim the page.
905                                  */
906                                 if (!trylock_page(page))
907                                         goto keep;
908                                 if (PageDirty(page) || PageWriteback(page))
909                                         goto keep_locked;
910                                 mapping = page_mapping(page);
911                         case PAGE_CLEAN:
912                                 ; /* try to free the page below */
913                         }
914                 }
915
916                 /*
917                  * If the page has buffers, try to free the buffer mappings
918                  * associated with this page. If we succeed we try to free
919                  * the page as well.
920                  *
921                  * We do this even if the page is PageDirty().
922                  * try_to_release_page() does not perform I/O, but it is
923                  * possible for a page to have PageDirty set, but it is actually
924                  * clean (all its buffers are clean).  This happens if the
925                  * buffers were written out directly, with submit_bh(). ext3
926                  * will do this, as well as the blockdev mapping.
927                  * try_to_release_page() will discover that cleanness and will
928                  * drop the buffers and mark the page clean - it can be freed.
929                  *
930                  * Rarely, pages can have buffers and no ->mapping.  These are
931                  * the pages which were not successfully invalidated in
932                  * truncate_complete_page().  We try to drop those buffers here
933                  * and if that worked, and the page is no longer mapped into
934                  * process address space (page_count == 1) it can be freed.
935                  * Otherwise, leave the page on the LRU so it is swappable.
936                  */
937                 if (page_has_private(page)) {
938                         if (!try_to_release_page(page, sc->gfp_mask))
939                                 goto activate_locked;
940                         if (!mapping && page_count(page) == 1) {
941                                 unlock_page(page);
942                                 if (put_page_testzero(page))
943                                         goto free_it;
944                                 else {
945                                         /*
946                                          * rare race with speculative reference.
947                                          * the speculative reference will free
948                                          * this page shortly, so we may
949                                          * increment nr_reclaimed here (and
950                                          * leave it off the LRU).
951                                          */
952                                         nr_reclaimed++;
953                                         continue;
954                                 }
955                         }
956                 }
957
958                 if (!mapping || !__remove_mapping(mapping, page))
959                         goto keep_locked;
960
961                 /*
962                  * At this point, we have no other references and there is
963                  * no way to pick any more up (removed from LRU, removed
964                  * from pagecache). Can use non-atomic bitops now (and
965                  * we obviously don't have to worry about waking up a process
966                  * waiting on the page lock, because there are no references.
967                  */
968                 __clear_page_locked(page);
969 free_it:
970                 nr_reclaimed++;
971
972                 /*
973                  * Is there need to periodically free_page_list? It would
974                  * appear not as the counts should be low
975                  */
976                 list_add(&page->lru, &free_pages);
977                 continue;
978
979 cull_mlocked:
980                 if (PageSwapCache(page))
981                         try_to_free_swap(page);
982                 unlock_page(page);
983                 putback_lru_page(page);
984                 reset_reclaim_mode(sc);
985                 continue;
986
987 activate_locked:
988                 /* Not a candidate for swapping, so reclaim swap space. */
989                 if (PageSwapCache(page) && vm_swap_full())
990                         try_to_free_swap(page);
991                 VM_BUG_ON(PageActive(page));
992                 SetPageActive(page);
993                 pgactivate++;
994 keep_locked:
995                 unlock_page(page);
996 keep:
997                 reset_reclaim_mode(sc);
998 keep_lumpy:
999                 list_add(&page->lru, &ret_pages);
1000                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
1001         }
1002
1003         /*
1004          * Tag a zone as congested if all the dirty pages encountered were
1005          * backed by a congested BDI. In this case, reclaimers should just
1006          * back off and wait for congestion to clear because further reclaim
1007          * will encounter the same problem
1008          */
1009         if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc))
1010                 zone_set_flag(zone, ZONE_CONGESTED);
1011
1012         free_page_list(&free_pages);
1013
1014         list_splice(&ret_pages, page_list);
1015         count_vm_events(PGACTIVATE, pgactivate);
1016         *ret_nr_dirty += nr_dirty;
1017         *ret_nr_writeback += nr_writeback;
1018         return nr_reclaimed;
1019 }
1020
1021 /*
1022  * Attempt to remove the specified page from its LRU.  Only take this page
1023  * if it is of the appropriate PageActive status.  Pages which are being
1024  * freed elsewhere are also ignored.
1025  *
1026  * page:        page to consider
1027  * mode:        one of the LRU isolation modes defined above
1028  *
1029  * returns 0 on success, -ve errno on failure.
1030  */
1031 int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file)
1032 {
1033         bool all_lru_mode;
1034         int ret = -EINVAL;
1035
1036         /* Only take pages on the LRU. */
1037         if (!PageLRU(page))
1038                 return ret;
1039
1040         all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) ==
1041                 (ISOLATE_ACTIVE|ISOLATE_INACTIVE);
1042
1043         /*
1044          * When checking the active state, we need to be sure we are
1045          * dealing with comparible boolean values.  Take the logical not
1046          * of each.
1047          */
1048         if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE))
1049                 return ret;
1050
1051         if (!all_lru_mode && !!page_is_file_cache(page) != file)
1052                 return ret;
1053
1054         /*
1055          * When this function is being called for lumpy reclaim, we
1056          * initially look into all LRU pages, active, inactive and
1057          * unevictable; only give shrink_page_list evictable pages.
1058          */
1059         if (PageUnevictable(page))
1060                 return ret;
1061
1062         ret = -EBUSY;
1063
1064         if ((mode & ISOLATE_CLEAN) && (PageDirty(page) || PageWriteback(page)))
1065                 return ret;
1066
1067         if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1068                 return ret;
1069
1070         if (likely(get_page_unless_zero(page))) {
1071                 /*
1072                  * Be careful not to clear PageLRU until after we're
1073                  * sure the page is not being freed elsewhere -- the
1074                  * page release code relies on it.
1075                  */
1076                 ClearPageLRU(page);
1077                 ret = 0;
1078         }
1079
1080         return ret;
1081 }
1082
1083 /*
1084  * zone->lru_lock is heavily contended.  Some of the functions that
1085  * shrink the lists perform better by taking out a batch of pages
1086  * and working on them outside the LRU lock.
1087  *
1088  * For pagecache intensive workloads, this function is the hottest
1089  * spot in the kernel (apart from copy_*_user functions).
1090  *
1091  * Appropriate locks must be held before calling this function.
1092  *
1093  * @nr_to_scan: The number of pages to look through on the list.
1094  * @src:        The LRU list to pull pages off.
1095  * @dst:        The temp list to put pages on to.
1096  * @scanned:    The number of pages that were scanned.
1097  * @order:      The caller's attempted allocation order
1098  * @mode:       One of the LRU isolation modes
1099  * @file:       True [1] if isolating file [!anon] pages
1100  *
1101  * returns how many pages were moved onto *@dst.
1102  */
1103 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1104                 struct list_head *src, struct list_head *dst,
1105                 unsigned long *scanned, int order, isolate_mode_t mode,
1106                 int file)
1107 {
1108         unsigned long nr_taken = 0;
1109         unsigned long nr_lumpy_taken = 0;
1110         unsigned long nr_lumpy_dirty = 0;
1111         unsigned long nr_lumpy_failed = 0;
1112         unsigned long scan;
1113
1114         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1115                 struct page *page;
1116                 unsigned long pfn;
1117                 unsigned long end_pfn;
1118                 unsigned long page_pfn;
1119                 int zone_id;
1120
1121                 page = lru_to_page(src);
1122                 prefetchw_prev_lru_page(page, src, flags);
1123
1124                 VM_BUG_ON(!PageLRU(page));
1125
1126                 switch (__isolate_lru_page(page, mode, file)) {
1127                 case 0:
1128                         list_move(&page->lru, dst);
1129                         mem_cgroup_del_lru(page);
1130                         nr_taken += hpage_nr_pages(page);
1131                         break;
1132
1133                 case -EBUSY:
1134                         /* else it is being freed elsewhere */
1135                         list_move(&page->lru, src);
1136                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1137                         continue;
1138
1139                 default:
1140                         BUG();
1141                 }
1142
1143                 if (!order)
1144                         continue;
1145
1146                 /*
1147                  * Attempt to take all pages in the order aligned region
1148                  * surrounding the tag page.  Only take those pages of
1149                  * the same active state as that tag page.  We may safely
1150                  * round the target page pfn down to the requested order
1151                  * as the mem_map is guaranteed valid out to MAX_ORDER,
1152                  * where that page is in a different zone we will detect
1153                  * it from its zone id and abort this block scan.
1154                  */
1155                 zone_id = page_zone_id(page);
1156                 page_pfn = page_to_pfn(page);
1157                 pfn = page_pfn & ~((1 << order) - 1);
1158                 end_pfn = pfn + (1 << order);
1159                 for (; pfn < end_pfn; pfn++) {
1160                         struct page *cursor_page;
1161
1162                         /* The target page is in the block, ignore it. */
1163                         if (unlikely(pfn == page_pfn))
1164                                 continue;
1165
1166                         /* Avoid holes within the zone. */
1167                         if (unlikely(!pfn_valid_within(pfn)))
1168                                 break;
1169
1170                         cursor_page = pfn_to_page(pfn);
1171
1172                         /* Check that we have not crossed a zone boundary. */
1173                         if (unlikely(page_zone_id(cursor_page) != zone_id))
1174                                 break;
1175
1176                         /*
1177                          * If we don't have enough swap space, reclaiming of
1178                          * anon page which don't already have a swap slot is
1179                          * pointless.
1180                          */
1181                         if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1182                             !PageSwapCache(cursor_page))
1183                                 break;
1184
1185                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1186                                 list_move(&cursor_page->lru, dst);
1187                                 mem_cgroup_del_lru(cursor_page);
1188                                 nr_taken += hpage_nr_pages(page);
1189                                 nr_lumpy_taken++;
1190                                 if (PageDirty(cursor_page))
1191                                         nr_lumpy_dirty++;
1192                                 scan++;
1193                         } else {
1194                                 /*
1195                                  * Check if the page is freed already.
1196                                  *
1197                                  * We can't use page_count() as that
1198                                  * requires compound_head and we don't
1199                                  * have a pin on the page here. If a
1200                                  * page is tail, we may or may not
1201                                  * have isolated the head, so assume
1202                                  * it's not free, it'd be tricky to
1203                                  * track the head status without a
1204                                  * page pin.
1205                                  */
1206                                 if (!PageTail(cursor_page) &&
1207                                     !atomic_read(&cursor_page->_count))
1208                                         continue;
1209                                 break;
1210                         }
1211                 }
1212
1213                 /* If we break out of the loop above, lumpy reclaim failed */
1214                 if (pfn < end_pfn)
1215                         nr_lumpy_failed++;
1216         }
1217
1218         *scanned = scan;
1219
1220         trace_mm_vmscan_lru_isolate(order,
1221                         nr_to_scan, scan,
1222                         nr_taken,
1223                         nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1224                         mode);
1225         return nr_taken;
1226 }
1227
1228 static unsigned long isolate_pages_global(unsigned long nr,
1229                                         struct list_head *dst,
1230                                         unsigned long *scanned, int order,
1231                                         isolate_mode_t mode,
1232                                         struct zone *z, int active, int file)
1233 {
1234         int lru = LRU_BASE;
1235         if (active)
1236                 lru += LRU_ACTIVE;
1237         if (file)
1238                 lru += LRU_FILE;
1239         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1240                                                                 mode, file);
1241 }
1242
1243 /*
1244  * clear_active_flags() is a helper for shrink_active_list(), clearing
1245  * any active bits from the pages in the list.
1246  */
1247 static unsigned long clear_active_flags(struct list_head *page_list,
1248                                         unsigned int *count)
1249 {
1250         int nr_active = 0;
1251         int lru;
1252         struct page *page;
1253
1254         list_for_each_entry(page, page_list, lru) {
1255                 int numpages = hpage_nr_pages(page);
1256                 lru = page_lru_base_type(page);
1257                 if (PageActive(page)) {
1258                         lru += LRU_ACTIVE;
1259                         ClearPageActive(page);
1260                         nr_active += numpages;
1261                 }
1262                 if (count)
1263                         count[lru] += numpages;
1264         }
1265
1266         return nr_active;
1267 }
1268
1269 /**
1270  * isolate_lru_page - tries to isolate a page from its LRU list
1271  * @page: page to isolate from its LRU list
1272  *
1273  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1274  * vmstat statistic corresponding to whatever LRU list the page was on.
1275  *
1276  * Returns 0 if the page was removed from an LRU list.
1277  * Returns -EBUSY if the page was not on an LRU list.
1278  *
1279  * The returned page will have PageLRU() cleared.  If it was found on
1280  * the active list, it will have PageActive set.  If it was found on
1281  * the unevictable list, it will have the PageUnevictable bit set. That flag
1282  * may need to be cleared by the caller before letting the page go.
1283  *
1284  * The vmstat statistic corresponding to the list on which the page was
1285  * found will be decremented.
1286  *
1287  * Restrictions:
1288  * (1) Must be called with an elevated refcount on the page. This is a
1289  *     fundamentnal difference from isolate_lru_pages (which is called
1290  *     without a stable reference).
1291  * (2) the lru_lock must not be held.
1292  * (3) interrupts must be enabled.
1293  */
1294 int isolate_lru_page(struct page *page)
1295 {
1296         int ret = -EBUSY;
1297
1298         VM_BUG_ON(!page_count(page));
1299
1300         if (PageLRU(page)) {
1301                 struct zone *zone = page_zone(page);
1302
1303                 spin_lock_irq(&zone->lru_lock);
1304                 if (PageLRU(page)) {
1305                         int lru = page_lru(page);
1306                         ret = 0;
1307                         get_page(page);
1308                         ClearPageLRU(page);
1309
1310                         del_page_from_lru_list(zone, page, lru);
1311                 }
1312                 spin_unlock_irq(&zone->lru_lock);
1313         }
1314         return ret;
1315 }
1316
1317 /*
1318  * Are there way too many processes in the direct reclaim path already?
1319  */
1320 static int too_many_isolated(struct zone *zone, int file,
1321                 struct scan_control *sc)
1322 {
1323         unsigned long inactive, isolated;
1324
1325         if (current_is_kswapd())
1326                 return 0;
1327
1328         if (!scanning_global_lru(sc))
1329                 return 0;
1330
1331         if (file) {
1332                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1333                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1334         } else {
1335                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1336                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1337         }
1338
1339         return isolated > inactive;
1340 }
1341
1342 /*
1343  * TODO: Try merging with migrations version of putback_lru_pages
1344  */
1345 static noinline_for_stack void
1346 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1347                                 unsigned long nr_anon, unsigned long nr_file,
1348                                 struct list_head *page_list)
1349 {
1350         struct page *page;
1351         struct pagevec pvec;
1352         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1353
1354         pagevec_init(&pvec, 1);
1355
1356         /*
1357          * Put back any unfreeable pages.
1358          */
1359         spin_lock(&zone->lru_lock);
1360         while (!list_empty(page_list)) {
1361                 int lru;
1362                 page = lru_to_page(page_list);
1363                 VM_BUG_ON(PageLRU(page));
1364                 list_del(&page->lru);
1365                 if (unlikely(!page_evictable(page, NULL))) {
1366                         spin_unlock_irq(&zone->lru_lock);
1367                         putback_lru_page(page);
1368                         spin_lock_irq(&zone->lru_lock);
1369                         continue;
1370                 }
1371                 SetPageLRU(page);
1372                 lru = page_lru(page);
1373                 add_page_to_lru_list(zone, page, lru);
1374                 if (is_active_lru(lru)) {
1375                         int file = is_file_lru(lru);
1376                         int numpages = hpage_nr_pages(page);
1377                         reclaim_stat->recent_rotated[file] += numpages;
1378                 }
1379                 if (!pagevec_add(&pvec, page)) {
1380                         spin_unlock_irq(&zone->lru_lock);
1381                         __pagevec_release(&pvec);
1382                         spin_lock_irq(&zone->lru_lock);
1383                 }
1384         }
1385         __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1386         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1387
1388         spin_unlock_irq(&zone->lru_lock);
1389         pagevec_release(&pvec);
1390 }
1391
1392 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1393                                         struct scan_control *sc,
1394                                         unsigned long *nr_anon,
1395                                         unsigned long *nr_file,
1396                                         struct list_head *isolated_list)
1397 {
1398         unsigned long nr_active;
1399         unsigned int count[NR_LRU_LISTS] = { 0, };
1400         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1401
1402         nr_active = clear_active_flags(isolated_list, count);
1403         __count_vm_events(PGDEACTIVATE, nr_active);
1404
1405         __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1406                               -count[LRU_ACTIVE_FILE]);
1407         __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1408                               -count[LRU_INACTIVE_FILE]);
1409         __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1410                               -count[LRU_ACTIVE_ANON]);
1411         __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1412                               -count[LRU_INACTIVE_ANON]);
1413
1414         *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1415         *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1416         __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1417         __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1418
1419         reclaim_stat->recent_scanned[0] += *nr_anon;
1420         reclaim_stat->recent_scanned[1] += *nr_file;
1421 }
1422
1423 /*
1424  * Returns true if a direct reclaim should wait on pages under writeback.
1425  *
1426  * If we are direct reclaiming for contiguous pages and we do not reclaim
1427  * everything in the list, try again and wait for writeback IO to complete.
1428  * This will stall high-order allocations noticeably. Only do that when really
1429  * need to free the pages under high memory pressure.
1430  */
1431 static inline bool should_reclaim_stall(unsigned long nr_taken,
1432                                         unsigned long nr_freed,
1433                                         int priority,
1434                                         struct scan_control *sc)
1435 {
1436         int lumpy_stall_priority;
1437
1438         /* kswapd should not stall on sync IO */
1439         if (current_is_kswapd())
1440                 return false;
1441
1442         /* Only stall on lumpy reclaim */
1443         if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1444                 return false;
1445
1446         /* If we have reclaimed everything on the isolated list, no stall */
1447         if (nr_freed == nr_taken)
1448                 return false;
1449
1450         /*
1451          * For high-order allocations, there are two stall thresholds.
1452          * High-cost allocations stall immediately where as lower
1453          * order allocations such as stacks require the scanning
1454          * priority to be much higher before stalling.
1455          */
1456         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1457                 lumpy_stall_priority = DEF_PRIORITY;
1458         else
1459                 lumpy_stall_priority = DEF_PRIORITY / 3;
1460
1461         return priority <= lumpy_stall_priority;
1462 }
1463
1464 /*
1465  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1466  * of reclaimed pages
1467  */
1468 static noinline_for_stack unsigned long
1469 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1470                         struct scan_control *sc, int priority, int file)
1471 {
1472         LIST_HEAD(page_list);
1473         unsigned long nr_scanned;
1474         unsigned long nr_reclaimed = 0;
1475         unsigned long nr_taken;
1476         unsigned long nr_anon;
1477         unsigned long nr_file;
1478         unsigned long nr_dirty = 0;
1479         unsigned long nr_writeback = 0;
1480         isolate_mode_t reclaim_mode = ISOLATE_INACTIVE;
1481
1482         while (unlikely(too_many_isolated(zone, file, sc))) {
1483                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1484
1485                 /* We are about to die and free our memory. Return now. */
1486                 if (fatal_signal_pending(current))
1487                         return SWAP_CLUSTER_MAX;
1488         }
1489
1490         set_reclaim_mode(priority, sc, false);
1491         if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
1492                 reclaim_mode |= ISOLATE_ACTIVE;
1493
1494         lru_add_drain();
1495
1496         if (!sc->may_unmap)
1497                 reclaim_mode |= ISOLATE_UNMAPPED;
1498         if (!sc->may_writepage)
1499                 reclaim_mode |= ISOLATE_CLEAN;
1500
1501         spin_lock_irq(&zone->lru_lock);
1502
1503         if (scanning_global_lru(sc)) {
1504                 nr_taken = isolate_pages_global(nr_to_scan, &page_list,
1505                         &nr_scanned, sc->order, reclaim_mode, zone, 0, file);
1506                 zone->pages_scanned += nr_scanned;
1507                 if (current_is_kswapd())
1508                         __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1509                                                nr_scanned);
1510                 else
1511                         __count_zone_vm_events(PGSCAN_DIRECT, zone,
1512                                                nr_scanned);
1513         } else {
1514                 nr_taken = mem_cgroup_isolate_pages(nr_to_scan, &page_list,
1515                         &nr_scanned, sc->order, reclaim_mode, zone,
1516                         sc->mem_cgroup, 0, file);
1517                 /*
1518                  * mem_cgroup_isolate_pages() keeps track of
1519                  * scanned pages on its own.
1520                  */
1521         }
1522
1523         if (nr_taken == 0) {
1524                 spin_unlock_irq(&zone->lru_lock);
1525                 return 0;
1526         }
1527
1528         update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1529
1530         spin_unlock_irq(&zone->lru_lock);
1531
1532         nr_reclaimed = shrink_page_list(&page_list, zone, sc, priority,
1533                                                 &nr_dirty, &nr_writeback);
1534
1535         /* Check if we should syncronously wait for writeback */
1536         if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1537                 set_reclaim_mode(priority, sc, true);
1538                 nr_reclaimed += shrink_page_list(&page_list, zone, sc,
1539                                         priority, &nr_dirty, &nr_writeback);
1540         }
1541
1542         local_irq_disable();
1543         if (current_is_kswapd())
1544                 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1545         __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1546
1547         putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1548
1549         /*
1550          * If reclaim is isolating dirty pages under writeback, it implies
1551          * that the long-lived page allocation rate is exceeding the page
1552          * laundering rate. Either the global limits are not being effective
1553          * at throttling processes due to the page distribution throughout
1554          * zones or there is heavy usage of a slow backing device. The
1555          * only option is to throttle from reclaim context which is not ideal
1556          * as there is no guarantee the dirtying process is throttled in the
1557          * same way balance_dirty_pages() manages.
1558          *
1559          * This scales the number of dirty pages that must be under writeback
1560          * before throttling depending on priority. It is a simple backoff
1561          * function that has the most effect in the range DEF_PRIORITY to
1562          * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1563          * in trouble and reclaim is considered to be in trouble.
1564          *
1565          * DEF_PRIORITY   100% isolated pages must be PageWriteback to throttle
1566          * DEF_PRIORITY-1  50% must be PageWriteback
1567          * DEF_PRIORITY-2  25% must be PageWriteback, kswapd in trouble
1568          * ...
1569          * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1570          *                     isolated page is PageWriteback
1571          */
1572         if (nr_writeback && nr_writeback >= (nr_taken >> (DEF_PRIORITY-priority)))
1573                 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1574
1575         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1576                 zone_idx(zone),
1577                 nr_scanned, nr_reclaimed,
1578                 priority,
1579                 trace_shrink_flags(file, sc->reclaim_mode));
1580         return nr_reclaimed;
1581 }
1582
1583 /*
1584  * This moves pages from the active list to the inactive list.
1585  *
1586  * We move them the other way if the page is referenced by one or more
1587  * processes, from rmap.
1588  *
1589  * If the pages are mostly unmapped, the processing is fast and it is
1590  * appropriate to hold zone->lru_lock across the whole operation.  But if
1591  * the pages are mapped, the processing is slow (page_referenced()) so we
1592  * should drop zone->lru_lock around each page.  It's impossible to balance
1593  * this, so instead we remove the pages from the LRU while processing them.
1594  * It is safe to rely on PG_active against the non-LRU pages in here because
1595  * nobody will play with that bit on a non-LRU page.
1596  *
1597  * The downside is that we have to touch page->_count against each page.
1598  * But we had to alter page->flags anyway.
1599  */
1600
1601 static void move_active_pages_to_lru(struct zone *zone,
1602                                      struct list_head *list,
1603                                      enum lru_list lru)
1604 {
1605         unsigned long pgmoved = 0;
1606         struct pagevec pvec;
1607         struct page *page;
1608
1609         pagevec_init(&pvec, 1);
1610
1611         while (!list_empty(list)) {
1612                 page = lru_to_page(list);
1613
1614                 VM_BUG_ON(PageLRU(page));
1615                 SetPageLRU(page);
1616
1617                 list_move(&page->lru, &zone->lru[lru].list);
1618                 mem_cgroup_add_lru_list(page, lru);
1619                 pgmoved += hpage_nr_pages(page);
1620
1621                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1622                         spin_unlock_irq(&zone->lru_lock);
1623                         if (buffer_heads_over_limit)
1624                                 pagevec_strip(&pvec);
1625                         __pagevec_release(&pvec);
1626                         spin_lock_irq(&zone->lru_lock);
1627                 }
1628         }
1629         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1630         if (!is_active_lru(lru))
1631                 __count_vm_events(PGDEACTIVATE, pgmoved);
1632 }
1633
1634 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1635                         struct scan_control *sc, int priority, int file)
1636 {
1637         unsigned long nr_taken;
1638         unsigned long pgscanned;
1639         unsigned long vm_flags;
1640         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1641         LIST_HEAD(l_active);
1642         LIST_HEAD(l_inactive);
1643         struct page *page;
1644         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1645         unsigned long nr_rotated = 0;
1646         isolate_mode_t reclaim_mode = ISOLATE_ACTIVE;
1647
1648         lru_add_drain();
1649
1650         if (!sc->may_unmap)
1651                 reclaim_mode |= ISOLATE_UNMAPPED;
1652         if (!sc->may_writepage)
1653                 reclaim_mode |= ISOLATE_CLEAN;
1654
1655         spin_lock_irq(&zone->lru_lock);
1656         if (scanning_global_lru(sc)) {
1657                 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1658                                                 &pgscanned, sc->order,
1659                                                 reclaim_mode, zone,
1660                                                 1, file);
1661                 zone->pages_scanned += pgscanned;
1662         } else {
1663                 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1664                                                 &pgscanned, sc->order,
1665                                                 reclaim_mode, zone,
1666                                                 sc->mem_cgroup, 1, file);
1667                 /*
1668                  * mem_cgroup_isolate_pages() keeps track of
1669                  * scanned pages on its own.
1670                  */
1671         }
1672
1673         reclaim_stat->recent_scanned[file] += nr_taken;
1674
1675         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1676         if (file)
1677                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1678         else
1679                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1680         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1681         spin_unlock_irq(&zone->lru_lock);
1682
1683         while (!list_empty(&l_hold)) {
1684                 cond_resched();
1685                 page = lru_to_page(&l_hold);
1686                 list_del(&page->lru);
1687
1688                 if (unlikely(!page_evictable(page, NULL))) {
1689                         putback_lru_page(page);
1690                         continue;
1691                 }
1692
1693                 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1694                         nr_rotated += hpage_nr_pages(page);
1695                         /*
1696                          * Identify referenced, file-backed active pages and
1697                          * give them one more trip around the active list. So
1698                          * that executable code get better chances to stay in
1699                          * memory under moderate memory pressure.  Anon pages
1700                          * are not likely to be evicted by use-once streaming
1701                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1702                          * so we ignore them here.
1703                          */
1704                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1705                                 list_add(&page->lru, &l_active);
1706                                 continue;
1707                         }
1708                 }
1709
1710                 ClearPageActive(page);  /* we are de-activating */
1711                 list_add(&page->lru, &l_inactive);
1712         }
1713
1714         /*
1715          * Move pages back to the lru list.
1716          */
1717         spin_lock_irq(&zone->lru_lock);
1718         /*
1719          * Count referenced pages from currently used mappings as rotated,
1720          * even though only some of them are actually re-activated.  This
1721          * helps balance scan pressure between file and anonymous pages in
1722          * get_scan_ratio.
1723          */
1724         reclaim_stat->recent_rotated[file] += nr_rotated;
1725
1726         move_active_pages_to_lru(zone, &l_active,
1727                                                 LRU_ACTIVE + file * LRU_FILE);
1728         move_active_pages_to_lru(zone, &l_inactive,
1729                                                 LRU_BASE   + file * LRU_FILE);
1730         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1731         spin_unlock_irq(&zone->lru_lock);
1732 }
1733
1734 #ifdef CONFIG_SWAP
1735 static int inactive_anon_is_low_global(struct zone *zone)
1736 {
1737         unsigned long active, inactive;
1738
1739         active = zone_page_state(zone, NR_ACTIVE_ANON);
1740         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1741
1742         if (inactive * zone->inactive_ratio < active)
1743                 return 1;
1744
1745         return 0;
1746 }
1747
1748 /**
1749  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1750  * @zone: zone to check
1751  * @sc:   scan control of this context
1752  *
1753  * Returns true if the zone does not have enough inactive anon pages,
1754  * meaning some active anon pages need to be deactivated.
1755  */
1756 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1757 {
1758         int low;
1759
1760         /*
1761          * If we don't have swap space, anonymous page deactivation
1762          * is pointless.
1763          */
1764         if (!total_swap_pages)
1765                 return 0;
1766
1767         if (scanning_global_lru(sc))
1768                 low = inactive_anon_is_low_global(zone);
1769         else
1770                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup, zone);
1771         return low;
1772 }
1773 #else
1774 static inline int inactive_anon_is_low(struct zone *zone,
1775                                         struct scan_control *sc)
1776 {
1777         return 0;
1778 }
1779 #endif
1780
1781 static int inactive_file_is_low_global(struct zone *zone)
1782 {
1783         unsigned long active, inactive;
1784
1785         active = zone_page_state(zone, NR_ACTIVE_FILE);
1786         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1787
1788         return (active > inactive);
1789 }
1790
1791 /**
1792  * inactive_file_is_low - check if file pages need to be deactivated
1793  * @zone: zone to check
1794  * @sc:   scan control of this context
1795  *
1796  * When the system is doing streaming IO, memory pressure here
1797  * ensures that active file pages get deactivated, until more
1798  * than half of the file pages are on the inactive list.
1799  *
1800  * Once we get to that situation, protect the system's working
1801  * set from being evicted by disabling active file page aging.
1802  *
1803  * This uses a different ratio than the anonymous pages, because
1804  * the page cache uses a use-once replacement algorithm.
1805  */
1806 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1807 {
1808         int low;
1809
1810         if (scanning_global_lru(sc))
1811                 low = inactive_file_is_low_global(zone);
1812         else
1813                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup, zone);
1814         return low;
1815 }
1816
1817 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1818                                 int file)
1819 {
1820         if (file)
1821                 return inactive_file_is_low(zone, sc);
1822         else
1823                 return inactive_anon_is_low(zone, sc);
1824 }
1825
1826 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1827         struct zone *zone, struct scan_control *sc, int priority)
1828 {
1829         int file = is_file_lru(lru);
1830
1831         if (is_active_lru(lru)) {
1832                 if (inactive_list_is_low(zone, sc, file))
1833                     shrink_active_list(nr_to_scan, zone, sc, priority, file);
1834                 return 0;
1835         }
1836
1837         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1838 }
1839
1840 static int vmscan_swappiness(struct scan_control *sc)
1841 {
1842         if (scanning_global_lru(sc))
1843                 return vm_swappiness;
1844         return mem_cgroup_swappiness(sc->mem_cgroup);
1845 }
1846
1847 /*
1848  * Determine how aggressively the anon and file LRU lists should be
1849  * scanned.  The relative value of each set of LRU lists is determined
1850  * by looking at the fraction of the pages scanned we did rotate back
1851  * onto the active list instead of evict.
1852  *
1853  * nr[0] = anon pages to scan; nr[1] = file pages to scan
1854  */
1855 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1856                                         unsigned long *nr, int priority)
1857 {
1858         unsigned long anon, file, free;
1859         unsigned long anon_prio, file_prio;
1860         unsigned long ap, fp;
1861         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1862         u64 fraction[2], denominator;
1863         enum lru_list l;
1864         int noswap = 0;
1865         bool force_scan = false;
1866
1867         /*
1868          * If the zone or memcg is small, nr[l] can be 0.  This
1869          * results in no scanning on this priority and a potential
1870          * priority drop.  Global direct reclaim can go to the next
1871          * zone and tends to have no problems. Global kswapd is for
1872          * zone balancing and it needs to scan a minimum amount. When
1873          * reclaiming for a memcg, a priority drop can cause high
1874          * latencies, so it's better to scan a minimum amount there as
1875          * well.
1876          */
1877         if (scanning_global_lru(sc) && current_is_kswapd())
1878                 force_scan = true;
1879         if (!scanning_global_lru(sc))
1880                 force_scan = true;
1881
1882         /* If we have no swap space, do not bother scanning anon pages. */
1883         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1884                 noswap = 1;
1885                 fraction[0] = 0;
1886                 fraction[1] = 1;
1887                 denominator = 1;
1888                 goto out;
1889         }
1890
1891         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1892                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1893         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1894                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1895
1896         if (scanning_global_lru(sc)) {
1897                 free  = zone_page_state(zone, NR_FREE_PAGES);
1898                 /* If we have very few page cache pages,
1899                    force-scan anon pages. */
1900                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1901                         fraction[0] = 1;
1902                         fraction[1] = 0;
1903                         denominator = 1;
1904                         goto out;
1905                 }
1906         }
1907
1908         /*
1909          * With swappiness at 100, anonymous and file have the same priority.
1910          * This scanning priority is essentially the inverse of IO cost.
1911          */
1912         anon_prio = vmscan_swappiness(sc);
1913         file_prio = 200 - vmscan_swappiness(sc);
1914
1915         /*
1916          * OK, so we have swap space and a fair amount of page cache
1917          * pages.  We use the recently rotated / recently scanned
1918          * ratios to determine how valuable each cache is.
1919          *
1920          * Because workloads change over time (and to avoid overflow)
1921          * we keep these statistics as a floating average, which ends
1922          * up weighing recent references more than old ones.
1923          *
1924          * anon in [0], file in [1]
1925          */
1926         spin_lock_irq(&zone->lru_lock);
1927         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1928                 reclaim_stat->recent_scanned[0] /= 2;
1929                 reclaim_stat->recent_rotated[0] /= 2;
1930         }
1931
1932         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1933                 reclaim_stat->recent_scanned[1] /= 2;
1934                 reclaim_stat->recent_rotated[1] /= 2;
1935         }
1936
1937         /*
1938          * The amount of pressure on anon vs file pages is inversely
1939          * proportional to the fraction of recently scanned pages on
1940          * each list that were recently referenced and in active use.
1941          */
1942         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1943         ap /= reclaim_stat->recent_rotated[0] + 1;
1944
1945         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1946         fp /= reclaim_stat->recent_rotated[1] + 1;
1947         spin_unlock_irq(&zone->lru_lock);
1948
1949         fraction[0] = ap;
1950         fraction[1] = fp;
1951         denominator = ap + fp + 1;
1952 out:
1953         for_each_evictable_lru(l) {
1954                 int file = is_file_lru(l);
1955                 unsigned long scan;
1956
1957                 scan = zone_nr_lru_pages(zone, sc, l);
1958                 if (priority || noswap) {
1959                         scan >>= priority;
1960                         if (!scan && force_scan)
1961                                 scan = SWAP_CLUSTER_MAX;
1962                         scan = div64_u64(scan * fraction[file], denominator);
1963                 }
1964                 nr[l] = scan;
1965         }
1966 }
1967
1968 /*
1969  * Reclaim/compaction depends on a number of pages being freed. To avoid
1970  * disruption to the system, a small number of order-0 pages continue to be
1971  * rotated and reclaimed in the normal fashion. However, by the time we get
1972  * back to the allocator and call try_to_compact_zone(), we ensure that
1973  * there are enough free pages for it to be likely successful
1974  */
1975 static inline bool should_continue_reclaim(struct zone *zone,
1976                                         unsigned long nr_reclaimed,
1977                                         unsigned long nr_scanned,
1978                                         struct scan_control *sc)
1979 {
1980         unsigned long pages_for_compaction;
1981         unsigned long inactive_lru_pages;
1982
1983         /* If not in reclaim/compaction mode, stop */
1984         if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1985                 return false;
1986
1987         /* Consider stopping depending on scan and reclaim activity */
1988         if (sc->gfp_mask & __GFP_REPEAT) {
1989                 /*
1990                  * For __GFP_REPEAT allocations, stop reclaiming if the
1991                  * full LRU list has been scanned and we are still failing
1992                  * to reclaim pages. This full LRU scan is potentially
1993                  * expensive but a __GFP_REPEAT caller really wants to succeed
1994                  */
1995                 if (!nr_reclaimed && !nr_scanned)
1996                         return false;
1997         } else {
1998                 /*
1999                  * For non-__GFP_REPEAT allocations which can presumably
2000                  * fail without consequence, stop if we failed to reclaim
2001                  * any pages from the last SWAP_CLUSTER_MAX number of
2002                  * pages that were scanned. This will return to the
2003                  * caller faster at the risk reclaim/compaction and
2004                  * the resulting allocation attempt fails
2005                  */
2006                 if (!nr_reclaimed)
2007                         return false;
2008         }
2009
2010         /*
2011          * If we have not reclaimed enough pages for compaction and the
2012          * inactive lists are large enough, continue reclaiming
2013          */
2014         pages_for_compaction = (2UL << sc->order);
2015         inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
2016                                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
2017         if (sc->nr_reclaimed < pages_for_compaction &&
2018                         inactive_lru_pages > pages_for_compaction)
2019                 return true;
2020
2021         /* If compaction would go ahead or the allocation would succeed, stop */
2022         switch (compaction_suitable(zone, sc->order)) {
2023         case COMPACT_PARTIAL:
2024         case COMPACT_CONTINUE:
2025                 return false;
2026         default:
2027                 return true;
2028         }
2029 }
2030
2031 /*
2032  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
2033  */
2034 static void shrink_zone(int priority, struct zone *zone,
2035                                 struct scan_control *sc)
2036 {
2037         unsigned long nr[NR_LRU_LISTS];
2038         unsigned long nr_to_scan;
2039         enum lru_list l;
2040         unsigned long nr_reclaimed, nr_scanned;
2041         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2042         struct blk_plug plug;
2043
2044 restart:
2045         nr_reclaimed = 0;
2046         nr_scanned = sc->nr_scanned;
2047         get_scan_count(zone, sc, nr, priority);
2048
2049         blk_start_plug(&plug);
2050         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2051                                         nr[LRU_INACTIVE_FILE]) {
2052                 for_each_evictable_lru(l) {
2053                         if (nr[l]) {
2054                                 nr_to_scan = min_t(unsigned long,
2055                                                    nr[l], SWAP_CLUSTER_MAX);
2056                                 nr[l] -= nr_to_scan;
2057
2058                                 nr_reclaimed += shrink_list(l, nr_to_scan,
2059                                                             zone, sc, priority);
2060                         }
2061                 }
2062                 /*
2063                  * On large memory systems, scan >> priority can become
2064                  * really large. This is fine for the starting priority;
2065                  * we want to put equal scanning pressure on each zone.
2066                  * However, if the VM has a harder time of freeing pages,
2067                  * with multiple processes reclaiming pages, the total
2068                  * freeing target can get unreasonably large.
2069                  */
2070                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
2071                         break;
2072         }
2073         blk_finish_plug(&plug);
2074         sc->nr_reclaimed += nr_reclaimed;
2075
2076         /*
2077          * Even if we did not try to evict anon pages at all, we want to
2078          * rebalance the anon lru active/inactive ratio.
2079          */
2080         if (inactive_anon_is_low(zone, sc))
2081                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
2082
2083         /* reclaim/compaction might need reclaim to continue */
2084         if (should_continue_reclaim(zone, nr_reclaimed,
2085                                         sc->nr_scanned - nr_scanned, sc))
2086                 goto restart;
2087
2088         throttle_vm_writeout(sc->gfp_mask);
2089 }
2090
2091 /*
2092  * This is the direct reclaim path, for page-allocating processes.  We only
2093  * try to reclaim pages from zones which will satisfy the caller's allocation
2094  * request.
2095  *
2096  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2097  * Because:
2098  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2099  *    allocation or
2100  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2101  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2102  *    zone defense algorithm.
2103  *
2104  * If a zone is deemed to be full of pinned pages then just give it a light
2105  * scan then give up on it.
2106  *
2107  * This function returns true if a zone is being reclaimed for a costly
2108  * high-order allocation and compaction is either ready to begin or deferred.
2109  * This indicates to the caller that it should retry the allocation or fail.
2110  */
2111 static bool shrink_zones(int priority, struct zonelist *zonelist,
2112                                         struct scan_control *sc)
2113 {
2114         struct zoneref *z;
2115         struct zone *zone;
2116         unsigned long nr_soft_reclaimed;
2117         unsigned long nr_soft_scanned;
2118         bool should_abort_reclaim = false;
2119
2120         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2121                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2122                 if (!populated_zone(zone))
2123                         continue;
2124                 /*
2125                  * Take care memory controller reclaiming has small influence
2126                  * to global LRU.
2127                  */
2128                 if (scanning_global_lru(sc)) {
2129                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2130                                 continue;
2131                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2132                                 continue;       /* Let kswapd poll it */
2133                         if (COMPACTION_BUILD) {
2134                                 /*
2135                                  * If we already have plenty of memory free for
2136                                  * compaction in this zone, don't free any more.
2137                                  * Even though compaction is invoked for any
2138                                  * non-zero order, only frequent costly order
2139                                  * reclamation is disruptive enough to become a
2140                                  * noticable problem, like transparent huge page
2141                                  * allocations.
2142                                  */
2143                                 if (sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2144                                         (compaction_suitable(zone, sc->order) ||
2145                                          compaction_deferred(zone))) {
2146                                         should_abort_reclaim = true;
2147                                         continue;
2148                                 }
2149                         }
2150                         /*
2151                          * This steals pages from memory cgroups over softlimit
2152                          * and returns the number of reclaimed pages and
2153                          * scanned pages. This works for global memory pressure
2154                          * and balancing, not for a memcg's limit.
2155                          */
2156                         nr_soft_scanned = 0;
2157                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2158                                                 sc->order, sc->gfp_mask,
2159                                                 &nr_soft_scanned);
2160                         sc->nr_reclaimed += nr_soft_reclaimed;
2161                         sc->nr_scanned += nr_soft_scanned;
2162                         /* need some check for avoid more shrink_zone() */
2163                 }
2164
2165                 shrink_zone(priority, zone, sc);
2166         }
2167
2168         return should_abort_reclaim;
2169 }
2170
2171 static bool zone_reclaimable(struct zone *zone)
2172 {
2173         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2174 }
2175
2176 /* All zones in zonelist are unreclaimable? */
2177 static bool all_unreclaimable(struct zonelist *zonelist,
2178                 struct scan_control *sc)
2179 {
2180         struct zoneref *z;
2181         struct zone *zone;
2182
2183         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2184                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2185                 if (!populated_zone(zone))
2186                         continue;
2187                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2188                         continue;
2189                 if (!zone->all_unreclaimable)
2190                         return false;
2191         }
2192
2193         return true;
2194 }
2195
2196 /*
2197  * This is the main entry point to direct page reclaim.
2198  *
2199  * If a full scan of the inactive list fails to free enough memory then we
2200  * are "out of memory" and something needs to be killed.
2201  *
2202  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2203  * high - the zone may be full of dirty or under-writeback pages, which this
2204  * caller can't do much about.  We kick the writeback threads and take explicit
2205  * naps in the hope that some of these pages can be written.  But if the
2206  * allocating task holds filesystem locks which prevent writeout this might not
2207  * work, and the allocation attempt will fail.
2208  *
2209  * returns:     0, if no pages reclaimed
2210  *              else, the number of pages reclaimed
2211  */
2212 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2213                                         struct scan_control *sc,
2214                                         struct shrink_control *shrink)
2215 {
2216         int priority;
2217         unsigned long total_scanned = 0;
2218         struct reclaim_state *reclaim_state = current->reclaim_state;
2219         struct zoneref *z;
2220         struct zone *zone;
2221         unsigned long writeback_threshold;
2222
2223         get_mems_allowed();
2224         delayacct_freepages_start();
2225
2226         if (scanning_global_lru(sc))
2227                 count_vm_event(ALLOCSTALL);
2228
2229         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2230                 sc->nr_scanned = 0;
2231                 if (!priority)
2232                         disable_swap_token(sc->mem_cgroup);
2233                 if (shrink_zones(priority, zonelist, sc))
2234                         break;
2235
2236                 /*
2237                  * Don't shrink slabs when reclaiming memory from
2238                  * over limit cgroups
2239                  */
2240                 if (scanning_global_lru(sc)) {
2241                         unsigned long lru_pages = 0;
2242                         for_each_zone_zonelist(zone, z, zonelist,
2243                                         gfp_zone(sc->gfp_mask)) {
2244                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2245                                         continue;
2246
2247                                 lru_pages += zone_reclaimable_pages(zone);
2248                         }
2249
2250                         shrink_slab(shrink, sc->nr_scanned, lru_pages);
2251                         if (reclaim_state) {
2252                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2253                                 reclaim_state->reclaimed_slab = 0;
2254                         }
2255                 }
2256                 total_scanned += sc->nr_scanned;
2257                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2258                         goto out;
2259
2260                 /*
2261                  * Try to write back as many pages as we just scanned.  This
2262                  * tends to cause slow streaming writers to write data to the
2263                  * disk smoothly, at the dirtying rate, which is nice.   But
2264                  * that's undesirable in laptop mode, where we *want* lumpy
2265                  * writeout.  So in laptop mode, write out the whole world.
2266                  */
2267                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2268                 if (total_scanned > writeback_threshold) {
2269                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2270                                                 WB_REASON_TRY_TO_FREE_PAGES);
2271                         sc->may_writepage = 1;
2272                 }
2273
2274                 /* Take a nap, wait for some writeback to complete */
2275                 if (!sc->hibernation_mode && sc->nr_scanned &&
2276                     priority < DEF_PRIORITY - 2) {
2277                         struct zone *preferred_zone;
2278
2279                         first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2280                                                 &cpuset_current_mems_allowed,
2281                                                 &preferred_zone);
2282                         wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2283                 }
2284         }
2285
2286 out:
2287         delayacct_freepages_end();
2288         put_mems_allowed();
2289
2290         if (sc->nr_reclaimed)
2291                 return sc->nr_reclaimed;
2292
2293         /*
2294          * As hibernation is going on, kswapd is freezed so that it can't mark
2295          * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2296          * check.
2297          */
2298         if (oom_killer_disabled)
2299                 return 0;
2300
2301         /* top priority shrink_zones still had more to do? don't OOM, then */
2302         if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2303                 return 1;
2304
2305         return 0;
2306 }
2307
2308 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2309                                 gfp_t gfp_mask, nodemask_t *nodemask)
2310 {
2311         unsigned long nr_reclaimed;
2312         struct scan_control sc = {
2313                 .gfp_mask = gfp_mask,
2314                 .may_writepage = !laptop_mode,
2315                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2316                 .may_unmap = 1,
2317                 .may_swap = 1,
2318                 .order = order,
2319                 .mem_cgroup = NULL,
2320                 .nodemask = nodemask,
2321         };
2322         struct shrink_control shrink = {
2323                 .gfp_mask = sc.gfp_mask,
2324         };
2325
2326         trace_mm_vmscan_direct_reclaim_begin(order,
2327                                 sc.may_writepage,
2328                                 gfp_mask);
2329
2330         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2331
2332         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2333
2334         return nr_reclaimed;
2335 }
2336
2337 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2338
2339 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2340                                                 gfp_t gfp_mask, bool noswap,
2341                                                 struct zone *zone,
2342                                                 unsigned long *nr_scanned)
2343 {
2344         struct scan_control sc = {
2345                 .nr_scanned = 0,
2346                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2347                 .may_writepage = !laptop_mode,
2348                 .may_unmap = 1,
2349                 .may_swap = !noswap,
2350                 .order = 0,
2351                 .mem_cgroup = mem,
2352         };
2353
2354         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2355                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2356
2357         trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2358                                                       sc.may_writepage,
2359                                                       sc.gfp_mask);
2360
2361         /*
2362          * NOTE: Although we can get the priority field, using it
2363          * here is not a good idea, since it limits the pages we can scan.
2364          * if we don't reclaim here, the shrink_zone from balance_pgdat
2365          * will pick up pages from other mem cgroup's as well. We hack
2366          * the priority and make it zero.
2367          */
2368         shrink_zone(0, zone, &sc);
2369
2370         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2371
2372         *nr_scanned = sc.nr_scanned;
2373         return sc.nr_reclaimed;
2374 }
2375
2376 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2377                                            gfp_t gfp_mask,
2378                                            bool noswap)
2379 {
2380         struct zonelist *zonelist;
2381         unsigned long nr_reclaimed;
2382         int nid;
2383         struct scan_control sc = {
2384                 .may_writepage = !laptop_mode,
2385                 .may_unmap = 1,
2386                 .may_swap = !noswap,
2387                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2388                 .order = 0,
2389                 .mem_cgroup = mem_cont,
2390                 .nodemask = NULL, /* we don't care the placement */
2391                 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2392                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2393         };
2394         struct shrink_control shrink = {
2395                 .gfp_mask = sc.gfp_mask,
2396         };
2397
2398         /*
2399          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2400          * take care of from where we get pages. So the node where we start the
2401          * scan does not need to be the current node.
2402          */
2403         nid = mem_cgroup_select_victim_node(mem_cont);
2404
2405         zonelist = NODE_DATA(nid)->node_zonelists;
2406
2407         trace_mm_vmscan_memcg_reclaim_begin(0,
2408                                             sc.may_writepage,
2409                                             sc.gfp_mask);
2410
2411         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2412
2413         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2414
2415         return nr_reclaimed;
2416 }
2417 #endif
2418
2419 /*
2420  * pgdat_balanced is used when checking if a node is balanced for high-order
2421  * allocations. Only zones that meet watermarks and are in a zone allowed
2422  * by the callers classzone_idx are added to balanced_pages. The total of
2423  * balanced pages must be at least 25% of the zones allowed by classzone_idx
2424  * for the node to be considered balanced. Forcing all zones to be balanced
2425  * for high orders can cause excessive reclaim when there are imbalanced zones.
2426  * The choice of 25% is due to
2427  *   o a 16M DMA zone that is balanced will not balance a zone on any
2428  *     reasonable sized machine
2429  *   o On all other machines, the top zone must be at least a reasonable
2430  *     percentage of the middle zones. For example, on 32-bit x86, highmem
2431  *     would need to be at least 256M for it to be balance a whole node.
2432  *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2433  *     to balance a node on its own. These seemed like reasonable ratios.
2434  */
2435 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2436                                                 int classzone_idx)
2437 {
2438         unsigned long present_pages = 0;
2439         int i;
2440
2441         for (i = 0; i <= classzone_idx; i++)
2442                 present_pages += pgdat->node_zones[i].present_pages;
2443
2444         /* A special case here: if zone has no page, we think it's balanced */
2445         return balanced_pages >= (present_pages >> 2);
2446 }
2447
2448 /* is kswapd sleeping prematurely? */
2449 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2450                                         int classzone_idx)
2451 {
2452         int i;
2453         unsigned long balanced = 0;
2454         bool all_zones_ok = true;
2455
2456         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2457         if (remaining)
2458                 return true;
2459
2460         /* Check the watermark levels */
2461         for (i = 0; i <= classzone_idx; i++) {
2462                 struct zone *zone = pgdat->node_zones + i;
2463
2464                 if (!populated_zone(zone))
2465                         continue;
2466
2467                 /*
2468                  * balance_pgdat() skips over all_unreclaimable after
2469                  * DEF_PRIORITY. Effectively, it considers them balanced so
2470                  * they must be considered balanced here as well if kswapd
2471                  * is to sleep
2472                  */
2473                 if (zone->all_unreclaimable) {
2474                         balanced += zone->present_pages;
2475                         continue;
2476                 }
2477
2478                 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2479                                                         i, 0))
2480                         all_zones_ok = false;
2481                 else
2482                         balanced += zone->present_pages;
2483         }
2484
2485         /*
2486          * For high-order requests, the balanced zones must contain at least
2487          * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2488          * must be balanced
2489          */
2490         if (order)
2491                 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2492         else
2493                 return !all_zones_ok;
2494 }
2495
2496 /*
2497  * For kswapd, balance_pgdat() will work across all this node's zones until
2498  * they are all at high_wmark_pages(zone).
2499  *
2500  * Returns the final order kswapd was reclaiming at
2501  *
2502  * There is special handling here for zones which are full of pinned pages.
2503  * This can happen if the pages are all mlocked, or if they are all used by
2504  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2505  * What we do is to detect the case where all pages in the zone have been
2506  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2507  * dead and from now on, only perform a short scan.  Basically we're polling
2508  * the zone for when the problem goes away.
2509  *
2510  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2511  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2512  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2513  * lower zones regardless of the number of free pages in the lower zones. This
2514  * interoperates with the page allocator fallback scheme to ensure that aging
2515  * of pages is balanced across the zones.
2516  */
2517 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2518                                                         int *classzone_idx)
2519 {
2520         int all_zones_ok;
2521         unsigned long balanced;
2522         int priority;
2523         int i;
2524         int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2525         unsigned long total_scanned;
2526         struct reclaim_state *reclaim_state = current->reclaim_state;
2527         unsigned long nr_soft_reclaimed;
2528         unsigned long nr_soft_scanned;
2529         struct scan_control sc = {
2530                 .gfp_mask = GFP_KERNEL,
2531                 .may_unmap = 1,
2532                 .may_swap = 1,
2533                 /*
2534                  * kswapd doesn't want to be bailed out while reclaim. because
2535                  * we want to put equal scanning pressure on each zone.
2536                  */
2537                 .nr_to_reclaim = ULONG_MAX,
2538                 .order = order,
2539                 .mem_cgroup = NULL,
2540         };
2541         struct shrink_control shrink = {
2542                 .gfp_mask = sc.gfp_mask,
2543         };
2544 loop_again:
2545         total_scanned = 0;
2546         sc.nr_reclaimed = 0;
2547         sc.may_writepage = !laptop_mode;
2548         count_vm_event(PAGEOUTRUN);
2549
2550         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2551                 unsigned long lru_pages = 0;
2552                 int has_under_min_watermark_zone = 0;
2553
2554                 /* The swap token gets in the way of swapout... */
2555                 if (!priority)
2556                         disable_swap_token(NULL);
2557
2558                 all_zones_ok = 1;
2559                 balanced = 0;
2560
2561                 /*
2562                  * Scan in the highmem->dma direction for the highest
2563                  * zone which needs scanning
2564                  */
2565                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2566                         struct zone *zone = pgdat->node_zones + i;
2567
2568                         if (!populated_zone(zone))
2569                                 continue;
2570
2571                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2572                                 continue;
2573
2574                         /*
2575                          * Do some background aging of the anon list, to give
2576                          * pages a chance to be referenced before reclaiming.
2577                          */
2578                         if (inactive_anon_is_low(zone, &sc))
2579                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2580                                                         &sc, priority, 0);
2581
2582                         if (!zone_watermark_ok_safe(zone, order,
2583                                         high_wmark_pages(zone), 0, 0)) {
2584                                 end_zone = i;
2585                                 break;
2586                         } else {
2587                                 /* If balanced, clear the congested flag */
2588                                 zone_clear_flag(zone, ZONE_CONGESTED);
2589                         }
2590                 }
2591                 if (i < 0)
2592                         goto out;
2593
2594                 for (i = 0; i <= end_zone; i++) {
2595                         struct zone *zone = pgdat->node_zones + i;
2596
2597                         lru_pages += zone_reclaimable_pages(zone);
2598                 }
2599
2600                 /*
2601                  * Now scan the zone in the dma->highmem direction, stopping
2602                  * at the last zone which needs scanning.
2603                  *
2604                  * We do this because the page allocator works in the opposite
2605                  * direction.  This prevents the page allocator from allocating
2606                  * pages behind kswapd's direction of progress, which would
2607                  * cause too much scanning of the lower zones.
2608                  */
2609                 for (i = 0; i <= end_zone; i++) {
2610                         struct zone *zone = pgdat->node_zones + i;
2611                         int nr_slab;
2612                         unsigned long balance_gap;
2613
2614                         if (!populated_zone(zone))
2615                                 continue;
2616
2617                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2618                                 continue;
2619
2620                         sc.nr_scanned = 0;
2621
2622                         nr_soft_scanned = 0;
2623                         /*
2624                          * Call soft limit reclaim before calling shrink_zone.
2625                          */
2626                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2627                                                         order, sc.gfp_mask,
2628                                                         &nr_soft_scanned);
2629                         sc.nr_reclaimed += nr_soft_reclaimed;
2630                         total_scanned += nr_soft_scanned;
2631
2632                         /*
2633                          * We put equal pressure on every zone, unless
2634                          * one zone has way too many pages free
2635                          * already. The "too many pages" is defined
2636                          * as the high wmark plus a "gap" where the
2637                          * gap is either the low watermark or 1%
2638                          * of the zone, whichever is smaller.
2639                          */
2640                         balance_gap = min(low_wmark_pages(zone),
2641                                 (zone->present_pages +
2642                                         KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2643                                 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2644                         if (!zone_watermark_ok_safe(zone, order,
2645                                         high_wmark_pages(zone) + balance_gap,
2646                                         end_zone, 0)) {
2647                                 shrink_zone(priority, zone, &sc);
2648
2649                                 reclaim_state->reclaimed_slab = 0;
2650                                 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2651                                 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2652                                 total_scanned += sc.nr_scanned;
2653
2654                                 if (nr_slab == 0 && !zone_reclaimable(zone))
2655                                         zone->all_unreclaimable = 1;
2656                         }
2657
2658                         /*
2659                          * If we've done a decent amount of scanning and
2660                          * the reclaim ratio is low, start doing writepage
2661                          * even in laptop mode
2662                          */
2663                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2664                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2665                                 sc.may_writepage = 1;
2666
2667                         if (zone->all_unreclaimable) {
2668                                 if (end_zone && end_zone == i)
2669                                         end_zone--;
2670                                 continue;
2671                         }
2672
2673                         if (!zone_watermark_ok_safe(zone, order,
2674                                         high_wmark_pages(zone), end_zone, 0)) {
2675                                 all_zones_ok = 0;
2676                                 /*
2677                                  * We are still under min water mark.  This
2678                                  * means that we have a GFP_ATOMIC allocation
2679                                  * failure risk. Hurry up!
2680                                  */
2681                                 if (!zone_watermark_ok_safe(zone, order,
2682                                             min_wmark_pages(zone), end_zone, 0))
2683                                         has_under_min_watermark_zone = 1;
2684                         } else {
2685                                 /*
2686                                  * If a zone reaches its high watermark,
2687                                  * consider it to be no longer congested. It's
2688                                  * possible there are dirty pages backed by
2689                                  * congested BDIs but as pressure is relieved,
2690                                  * spectulatively avoid congestion waits
2691                                  */
2692                                 zone_clear_flag(zone, ZONE_CONGESTED);
2693                                 if (i <= *classzone_idx)
2694                                         balanced += zone->present_pages;
2695                         }
2696
2697                 }
2698                 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2699                         break;          /* kswapd: all done */
2700                 /*
2701                  * OK, kswapd is getting into trouble.  Take a nap, then take
2702                  * another pass across the zones.
2703                  */
2704                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2705                         if (has_under_min_watermark_zone)
2706                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2707                         else
2708                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2709                 }
2710
2711                 /*
2712                  * We do this so kswapd doesn't build up large priorities for
2713                  * example when it is freeing in parallel with allocators. It
2714                  * matches the direct reclaim path behaviour in terms of impact
2715                  * on zone->*_priority.
2716                  */
2717                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2718                         break;
2719         }
2720 out:
2721
2722         /*
2723          * order-0: All zones must meet high watermark for a balanced node
2724          * high-order: Balanced zones must make up at least 25% of the node
2725          *             for the node to be balanced
2726          */
2727         if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2728                 cond_resched();
2729
2730                 try_to_freeze();
2731
2732                 /*
2733                  * Fragmentation may mean that the system cannot be
2734                  * rebalanced for high-order allocations in all zones.
2735                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2736                  * it means the zones have been fully scanned and are still
2737                  * not balanced. For high-order allocations, there is
2738                  * little point trying all over again as kswapd may
2739                  * infinite loop.
2740                  *
2741                  * Instead, recheck all watermarks at order-0 as they
2742                  * are the most important. If watermarks are ok, kswapd will go
2743                  * back to sleep. High-order users can still perform direct
2744                  * reclaim if they wish.
2745                  */
2746                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2747                         order = sc.order = 0;
2748
2749                 goto loop_again;
2750         }
2751
2752         /*
2753          * If kswapd was reclaiming at a higher order, it has the option of
2754          * sleeping without all zones being balanced. Before it does, it must
2755          * ensure that the watermarks for order-0 on *all* zones are met and
2756          * that the congestion flags are cleared. The congestion flag must
2757          * be cleared as kswapd is the only mechanism that clears the flag
2758          * and it is potentially going to sleep here.
2759          */
2760         if (order) {
2761                 for (i = 0; i <= end_zone; i++) {
2762                         struct zone *zone = pgdat->node_zones + i;
2763
2764                         if (!populated_zone(zone))
2765                                 continue;
2766
2767                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2768                                 continue;
2769
2770                         /* Confirm the zone is balanced for order-0 */
2771                         if (!zone_watermark_ok(zone, 0,
2772                                         high_wmark_pages(zone), 0, 0)) {
2773                                 order = sc.order = 0;
2774                                 goto loop_again;
2775                         }
2776
2777                         /* If balanced, clear the congested flag */
2778                         zone_clear_flag(zone, ZONE_CONGESTED);
2779                         if (i <= *classzone_idx)
2780                                 balanced += zone->present_pages;
2781                 }
2782         }
2783
2784         /*
2785          * Return the order we were reclaiming at so sleeping_prematurely()
2786          * makes a decision on the order we were last reclaiming at. However,
2787          * if another caller entered the allocator slow path while kswapd
2788          * was awake, order will remain at the higher level
2789          */
2790         *classzone_idx = end_zone;
2791         return order;
2792 }
2793
2794 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2795 {
2796         long remaining = 0;
2797         DEFINE_WAIT(wait);
2798
2799         if (freezing(current) || kthread_should_stop())
2800                 return;
2801
2802         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2803
2804         /* Try to sleep for a short interval */
2805         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2806                 remaining = schedule_timeout(HZ/10);
2807                 finish_wait(&pgdat->kswapd_wait, &wait);
2808                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2809         }
2810
2811         /*
2812          * After a short sleep, check if it was a premature sleep. If not, then
2813          * go fully to sleep until explicitly woken up.
2814          */
2815         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2816                 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2817
2818                 /*
2819                  * vmstat counters are not perfectly accurate and the estimated
2820                  * value for counters such as NR_FREE_PAGES can deviate from the
2821                  * true value by nr_online_cpus * threshold. To avoid the zone
2822                  * watermarks being breached while under pressure, we reduce the
2823                  * per-cpu vmstat threshold while kswapd is awake and restore
2824                  * them before going back to sleep.
2825                  */
2826                 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2827                 schedule();
2828                 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2829         } else {
2830                 if (remaining)
2831                         count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2832                 else
2833                         count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2834         }
2835         finish_wait(&pgdat->kswapd_wait, &wait);
2836 }
2837
2838 /*
2839  * The background pageout daemon, started as a kernel thread
2840  * from the init process.
2841  *
2842  * This basically trickles out pages so that we have _some_
2843  * free memory available even if there is no other activity
2844  * that frees anything up. This is needed for things like routing
2845  * etc, where we otherwise might have all activity going on in
2846  * asynchronous contexts that cannot page things out.
2847  *
2848  * If there are applications that are active memory-allocators
2849  * (most normal use), this basically shouldn't matter.
2850  */
2851 static int kswapd(void *p)
2852 {
2853         unsigned long order, new_order;
2854         unsigned balanced_order;
2855         int classzone_idx, new_classzone_idx;
2856         int balanced_classzone_idx;
2857         pg_data_t *pgdat = (pg_data_t*)p;
2858         struct task_struct *tsk = current;
2859
2860         struct reclaim_state reclaim_state = {
2861                 .reclaimed_slab = 0,
2862         };
2863         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2864
2865         lockdep_set_current_reclaim_state(GFP_KERNEL);
2866
2867         if (!cpumask_empty(cpumask))
2868                 set_cpus_allowed_ptr(tsk, cpumask);
2869         current->reclaim_state = &reclaim_state;
2870
2871         /*
2872          * Tell the memory management that we're a "memory allocator",
2873          * and that if we need more memory we should get access to it
2874          * regardless (see "__alloc_pages()"). "kswapd" should
2875          * never get caught in the normal page freeing logic.
2876          *
2877          * (Kswapd normally doesn't need memory anyway, but sometimes
2878          * you need a small amount of memory in order to be able to
2879          * page out something else, and this flag essentially protects
2880          * us from recursively trying to free more memory as we're
2881          * trying to free the first piece of memory in the first place).
2882          */
2883         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2884         set_freezable();
2885
2886         order = new_order = 0;
2887         balanced_order = 0;
2888         classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2889         balanced_classzone_idx = classzone_idx;
2890         for ( ; ; ) {
2891                 int ret;
2892
2893                 /*
2894                  * If the last balance_pgdat was unsuccessful it's unlikely a
2895                  * new request of a similar or harder type will succeed soon
2896                  * so consider going to sleep on the basis we reclaimed at
2897                  */
2898                 if (balanced_classzone_idx >= new_classzone_idx &&
2899                                         balanced_order == new_order) {
2900                         new_order = pgdat->kswapd_max_order;
2901                         new_classzone_idx = pgdat->classzone_idx;
2902                         pgdat->kswapd_max_order =  0;
2903                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2904                 }
2905
2906                 if (order < new_order || classzone_idx > new_classzone_idx) {
2907                         /*
2908                          * Don't sleep if someone wants a larger 'order'
2909                          * allocation or has tigher zone constraints
2910                          */
2911                         order = new_order;
2912                         classzone_idx = new_classzone_idx;
2913                 } else {
2914                         kswapd_try_to_sleep(pgdat, balanced_order,
2915                                                 balanced_classzone_idx);
2916                         order = pgdat->kswapd_max_order;
2917                         classzone_idx = pgdat->classzone_idx;
2918                         new_order = order;
2919                         new_classzone_idx = classzone_idx;
2920                         pgdat->kswapd_max_order = 0;
2921                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2922                 }
2923
2924                 ret = try_to_freeze();
2925                 if (kthread_should_stop())
2926                         break;
2927
2928                 /*
2929                  * We can speed up thawing tasks if we don't call balance_pgdat
2930                  * after returning from the refrigerator
2931                  */
2932                 if (!ret) {
2933                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2934                         balanced_classzone_idx = classzone_idx;
2935                         balanced_order = balance_pgdat(pgdat, order,
2936                                                 &balanced_classzone_idx);
2937                 }
2938         }
2939         return 0;
2940 }
2941
2942 /*
2943  * A zone is low on free memory, so wake its kswapd task to service it.
2944  */
2945 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2946 {
2947         pg_data_t *pgdat;
2948
2949         if (!populated_zone(zone))
2950                 return;
2951
2952         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2953                 return;
2954         pgdat = zone->zone_pgdat;
2955         if (pgdat->kswapd_max_order < order) {
2956                 pgdat->kswapd_max_order = order;
2957                 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2958         }
2959         if (!waitqueue_active(&pgdat->kswapd_wait))
2960                 return;
2961         if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2962                 return;
2963
2964         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2965         wake_up_interruptible(&pgdat->kswapd_wait);
2966 }
2967
2968 /*
2969  * The reclaimable count would be mostly accurate.
2970  * The less reclaimable pages may be
2971  * - mlocked pages, which will be moved to unevictable list when encountered
2972  * - mapped pages, which may require several travels to be reclaimed
2973  * - dirty pages, which is not "instantly" reclaimable
2974  */
2975 unsigned long global_reclaimable_pages(void)
2976 {
2977         int nr;
2978
2979         nr = global_page_state(NR_ACTIVE_FILE) +
2980              global_page_state(NR_INACTIVE_FILE);
2981
2982         if (nr_swap_pages > 0)
2983                 nr += global_page_state(NR_ACTIVE_ANON) +
2984                       global_page_state(NR_INACTIVE_ANON);
2985
2986         return nr;
2987 }
2988
2989 unsigned long zone_reclaimable_pages(struct zone *zone)
2990 {
2991         int nr;
2992
2993         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2994              zone_page_state(zone, NR_INACTIVE_FILE);
2995
2996         if (nr_swap_pages > 0)
2997                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2998                       zone_page_state(zone, NR_INACTIVE_ANON);
2999
3000         return nr;
3001 }
3002
3003 #ifdef CONFIG_HIBERNATION
3004 /*
3005  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3006  * freed pages.
3007  *
3008  * Rather than trying to age LRUs the aim is to preserve the overall
3009  * LRU order by reclaiming preferentially
3010  * inactive > active > active referenced > active mapped
3011  */
3012 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3013 {
3014         struct reclaim_state reclaim_state;
3015         struct scan_control sc = {
3016                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3017                 .may_swap = 1,
3018                 .may_unmap = 1,
3019                 .may_writepage = 1,
3020                 .nr_to_reclaim = nr_to_reclaim,
3021                 .hibernation_mode = 1,
3022                 .order = 0,
3023         };
3024         struct shrink_control shrink = {
3025                 .gfp_mask = sc.gfp_mask,
3026         };
3027         struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3028         struct task_struct *p = current;
3029         unsigned long nr_reclaimed;
3030
3031         p->flags |= PF_MEMALLOC;
3032         lockdep_set_current_reclaim_state(sc.gfp_mask);
3033         reclaim_state.reclaimed_slab = 0;
3034         p->reclaim_state = &reclaim_state;
3035
3036         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3037
3038         p->reclaim_state = NULL;
3039         lockdep_clear_current_reclaim_state();
3040         p->flags &= ~PF_MEMALLOC;
3041
3042         return nr_reclaimed;
3043 }
3044 #endif /* CONFIG_HIBERNATION */
3045
3046 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3047    not required for correctness.  So if the last cpu in a node goes
3048    away, we get changed to run anywhere: as the first one comes back,
3049    restore their cpu bindings. */
3050 static int __devinit cpu_callback(struct notifier_block *nfb,
3051                                   unsigned long action, void *hcpu)
3052 {
3053         int nid;
3054
3055         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3056                 for_each_node_state(nid, N_HIGH_MEMORY) {
3057                         pg_data_t *pgdat = NODE_DATA(nid);
3058                         const struct cpumask *mask;
3059
3060                         mask = cpumask_of_node(pgdat->node_id);
3061
3062                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3063                                 /* One of our CPUs online: restore mask */
3064                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3065                 }
3066         }
3067         return NOTIFY_OK;
3068 }
3069
3070 /*
3071  * This kswapd start function will be called by init and node-hot-add.
3072  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3073  */
3074 int kswapd_run(int nid)
3075 {
3076         pg_data_t *pgdat = NODE_DATA(nid);
3077         int ret = 0;
3078
3079         if (pgdat->kswapd)
3080                 return 0;
3081
3082         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3083         if (IS_ERR(pgdat->kswapd)) {
3084                 /* failure at boot is fatal */
3085                 BUG_ON(system_state == SYSTEM_BOOTING);
3086                 printk("Failed to start kswapd on node %d\n",nid);
3087                 ret = -1;
3088         }
3089         return ret;
3090 }
3091
3092 /*
3093  * Called by memory hotplug when all memory in a node is offlined.
3094  */
3095 void kswapd_stop(int nid)
3096 {
3097         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3098
3099         if (kswapd)
3100                 kthread_stop(kswapd);
3101 }
3102
3103 static int __init kswapd_init(void)
3104 {
3105         int nid;
3106
3107         swap_setup();
3108         for_each_node_state(nid, N_HIGH_MEMORY)
3109                 kswapd_run(nid);
3110         hotcpu_notifier(cpu_callback, 0);
3111         return 0;
3112 }
3113
3114 module_init(kswapd_init)
3115
3116 #ifdef CONFIG_NUMA
3117 /*
3118  * Zone reclaim mode
3119  *
3120  * If non-zero call zone_reclaim when the number of free pages falls below
3121  * the watermarks.
3122  */
3123 int zone_reclaim_mode __read_mostly;
3124
3125 #define RECLAIM_OFF 0
3126 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
3127 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
3128 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
3129
3130 /*
3131  * Priority for ZONE_RECLAIM. This determines the fraction of pages
3132  * of a node considered for each zone_reclaim. 4 scans 1/16th of
3133  * a zone.
3134  */
3135 #define ZONE_RECLAIM_PRIORITY 4
3136
3137 /*
3138  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3139  * occur.
3140  */
3141 int sysctl_min_unmapped_ratio = 1;
3142
3143 /*
3144  * If the number of slab pages in a zone grows beyond this percentage then
3145  * slab reclaim needs to occur.
3146  */
3147 int sysctl_min_slab_ratio = 5;
3148
3149 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3150 {
3151         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3152         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3153                 zone_page_state(zone, NR_ACTIVE_FILE);
3154
3155         /*
3156          * It's possible for there to be more file mapped pages than
3157          * accounted for by the pages on the file LRU lists because
3158          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3159          */
3160         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3161 }
3162
3163 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3164 static long zone_pagecache_reclaimable(struct zone *zone)
3165 {
3166         long nr_pagecache_reclaimable;
3167         long delta = 0;
3168
3169         /*
3170          * If RECLAIM_SWAP is set, then all file pages are considered
3171          * potentially reclaimable. Otherwise, we have to worry about
3172          * pages like swapcache and zone_unmapped_file_pages() provides
3173          * a better estimate
3174          */
3175         if (zone_reclaim_mode & RECLAIM_SWAP)
3176                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3177         else
3178                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3179
3180         /* If we can't clean pages, remove dirty pages from consideration */
3181         if (!(zone_reclaim_mode & RECLAIM_WRITE))
3182                 delta += zone_page_state(zone, NR_FILE_DIRTY);
3183
3184         /* Watch for any possible underflows due to delta */
3185         if (unlikely(delta > nr_pagecache_reclaimable))
3186                 delta = nr_pagecache_reclaimable;
3187
3188         return nr_pagecache_reclaimable - delta;
3189 }
3190
3191 /*
3192  * Try to free up some pages from this zone through reclaim.
3193  */
3194 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3195 {
3196         /* Minimum pages needed in order to stay on node */
3197         const unsigned long nr_pages = 1 << order;
3198         struct task_struct *p = current;
3199         struct reclaim_state reclaim_state;
3200         int priority;
3201         struct scan_control sc = {
3202                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3203                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3204                 .may_swap = 1,
3205                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3206                                        SWAP_CLUSTER_MAX),
3207                 .gfp_mask = gfp_mask,
3208                 .order = order,
3209         };
3210         struct shrink_control shrink = {
3211                 .gfp_mask = sc.gfp_mask,
3212         };
3213         unsigned long nr_slab_pages0, nr_slab_pages1;
3214
3215         cond_resched();
3216         /*
3217          * We need to be able to allocate from the reserves for RECLAIM_SWAP
3218          * and we also need to be able to write out pages for RECLAIM_WRITE
3219          * and RECLAIM_SWAP.
3220          */
3221         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3222         lockdep_set_current_reclaim_state(gfp_mask);
3223         reclaim_state.reclaimed_slab = 0;
3224         p->reclaim_state = &reclaim_state;
3225
3226         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3227                 /*
3228                  * Free memory by calling shrink zone with increasing
3229                  * priorities until we have enough memory freed.
3230                  */
3231                 priority = ZONE_RECLAIM_PRIORITY;
3232                 do {
3233                         shrink_zone(priority, zone, &sc);
3234                         priority--;
3235                 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
3236         }
3237
3238         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3239         if (nr_slab_pages0 > zone->min_slab_pages) {
3240                 /*
3241                  * shrink_slab() does not currently allow us to determine how
3242                  * many pages were freed in this zone. So we take the current
3243                  * number of slab pages and shake the slab until it is reduced
3244                  * by the same nr_pages that we used for reclaiming unmapped
3245                  * pages.
3246                  *
3247                  * Note that shrink_slab will free memory on all zones and may
3248                  * take a long time.
3249                  */
3250                 for (;;) {
3251                         unsigned long lru_pages = zone_reclaimable_pages(zone);
3252
3253                         /* No reclaimable slab or very low memory pressure */
3254                         if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3255                                 break;
3256
3257                         /* Freed enough memory */
3258                         nr_slab_pages1 = zone_page_state(zone,
3259                                                         NR_SLAB_RECLAIMABLE);
3260                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3261                                 break;
3262                 }
3263
3264                 /*
3265                  * Update nr_reclaimed by the number of slab pages we
3266                  * reclaimed from this zone.
3267                  */
3268                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3269                 if (nr_slab_pages1 < nr_slab_pages0)
3270                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3271         }
3272
3273         p->reclaim_state = NULL;
3274         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3275         lockdep_clear_current_reclaim_state();
3276         return sc.nr_reclaimed >= nr_pages;
3277 }
3278
3279 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3280 {
3281         int node_id;
3282         int ret;
3283
3284         /*
3285          * Zone reclaim reclaims unmapped file backed pages and
3286          * slab pages if we are over the defined limits.
3287          *
3288          * A small portion of unmapped file backed pages is needed for
3289          * file I/O otherwise pages read by file I/O will be immediately
3290          * thrown out if the zone is overallocated. So we do not reclaim
3291          * if less than a specified percentage of the zone is used by
3292          * unmapped file backed pages.
3293          */
3294         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3295             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3296                 return ZONE_RECLAIM_FULL;
3297
3298         if (zone->all_unreclaimable)
3299                 return ZONE_RECLAIM_FULL;
3300
3301         /*
3302          * Do not scan if the allocation should not be delayed.
3303          */
3304         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3305                 return ZONE_RECLAIM_NOSCAN;
3306
3307         /*
3308          * Only run zone reclaim on the local zone or on zones that do not
3309          * have associated processors. This will favor the local processor
3310          * over remote processors and spread off node memory allocations
3311          * as wide as possible.
3312          */
3313         node_id = zone_to_nid(zone);
3314         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3315                 return ZONE_RECLAIM_NOSCAN;
3316
3317         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3318                 return ZONE_RECLAIM_NOSCAN;
3319
3320         ret = __zone_reclaim(zone, gfp_mask, order);
3321         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3322
3323         if (!ret)
3324                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3325
3326         return ret;
3327 }
3328 #endif
3329
3330 /*
3331  * page_evictable - test whether a page is evictable
3332  * @page: the page to test
3333  * @vma: the VMA in which the page is or will be mapped, may be NULL
3334  *
3335  * Test whether page is evictable--i.e., should be placed on active/inactive
3336  * lists vs unevictable list.  The vma argument is !NULL when called from the
3337  * fault path to determine how to instantate a new page.
3338  *
3339  * Reasons page might not be evictable:
3340  * (1) page's mapping marked unevictable
3341  * (2) page is part of an mlocked VMA
3342  *
3343  */
3344 int page_evictable(struct page *page, struct vm_area_struct *vma)
3345 {
3346
3347         if (mapping_unevictable(page_mapping(page)))
3348                 return 0;
3349
3350         if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
3351                 return 0;
3352
3353         return 1;
3354 }
3355
3356 /**
3357  * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
3358  * @page: page to check evictability and move to appropriate lru list
3359  * @zone: zone page is in
3360  *
3361  * Checks a page for evictability and moves the page to the appropriate
3362  * zone lru list.
3363  *
3364  * Restrictions: zone->lru_lock must be held, page must be on LRU and must
3365  * have PageUnevictable set.
3366  */
3367 static void check_move_unevictable_page(struct page *page, struct zone *zone)
3368 {
3369         VM_BUG_ON(PageActive(page));
3370
3371 retry:
3372         ClearPageUnevictable(page);
3373         if (page_evictable(page, NULL)) {
3374                 enum lru_list l = page_lru_base_type(page);
3375
3376                 __dec_zone_state(zone, NR_UNEVICTABLE);
3377                 list_move(&page->lru, &zone->lru[l].list);
3378                 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
3379                 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
3380                 __count_vm_event(UNEVICTABLE_PGRESCUED);
3381         } else {
3382                 /*
3383                  * rotate unevictable list
3384                  */
3385                 SetPageUnevictable(page);
3386                 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
3387                 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
3388                 if (page_evictable(page, NULL))
3389                         goto retry;
3390         }
3391 }
3392
3393 /**
3394  * scan_mapping_unevictable_pages - scan an address space for evictable pages
3395  * @mapping: struct address_space to scan for evictable pages
3396  *
3397  * Scan all pages in mapping.  Check unevictable pages for
3398  * evictability and move them to the appropriate zone lru list.
3399  */
3400 void scan_mapping_unevictable_pages(struct address_space *mapping)
3401 {
3402         pgoff_t next = 0;
3403         pgoff_t end   = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
3404                          PAGE_CACHE_SHIFT;
3405         struct zone *zone;
3406         struct pagevec pvec;
3407
3408         if (mapping->nrpages == 0)
3409                 return;
3410
3411         pagevec_init(&pvec, 0);
3412         while (next < end &&
3413                 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
3414                 int i;
3415                 int pg_scanned = 0;
3416
3417                 zone = NULL;
3418
3419                 for (i = 0; i < pagevec_count(&pvec); i++) {
3420                         struct page *page = pvec.pages[i];
3421                         pgoff_t page_index = page->index;
3422                         struct zone *pagezone = page_zone(page);
3423
3424                         pg_scanned++;
3425                         if (page_index > next)
3426                                 next = page_index;
3427                         next++;
3428
3429                         if (pagezone != zone) {
3430                                 if (zone)
3431                                         spin_unlock_irq(&zone->lru_lock);
3432                                 zone = pagezone;
3433                                 spin_lock_irq(&zone->lru_lock);
3434                         }
3435
3436                         if (PageLRU(page) && PageUnevictable(page))
3437                                 check_move_unevictable_page(page, zone);
3438                 }
3439                 if (zone)
3440                         spin_unlock_irq(&zone->lru_lock);
3441                 pagevec_release(&pvec);
3442
3443                 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
3444         }
3445
3446 }
3447
3448 static void warn_scan_unevictable_pages(void)
3449 {
3450         printk_once(KERN_WARNING
3451                     "The scan_unevictable_pages sysctl/node-interface has been "
3452                     "disabled for lack of a legitimate use case.  If you have "
3453                     "one, please send an email to linux-mm@kvack.org.\n");
3454 }
3455
3456 /*
3457  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3458  * all nodes' unevictable lists for evictable pages
3459  */
3460 unsigned long scan_unevictable_pages;
3461
3462 int scan_unevictable_handler(struct ctl_table *table, int write,
3463                            void __user *buffer,
3464                            size_t *length, loff_t *ppos)
3465 {
3466         warn_scan_unevictable_pages();
3467         proc_doulongvec_minmax(table, write, buffer, length, ppos);
3468         scan_unevictable_pages = 0;
3469         return 0;
3470 }
3471
3472 #ifdef CONFIG_NUMA
3473 /*
3474  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3475  * a specified node's per zone unevictable lists for evictable pages.
3476  */
3477
3478 static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3479                                           struct sysdev_attribute *attr,
3480                                           char *buf)
3481 {
3482         warn_scan_unevictable_pages();
3483         return sprintf(buf, "0\n");     /* always zero; should fit... */
3484 }
3485
3486 static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3487                                            struct sysdev_attribute *attr,
3488                                         const char *buf, size_t count)
3489 {
3490         warn_scan_unevictable_pages();
3491         return 1;
3492 }
3493
3494
3495 static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3496                         read_scan_unevictable_node,
3497                         write_scan_unevictable_node);
3498
3499 int scan_unevictable_register_node(struct node *node)
3500 {
3501         return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3502 }
3503
3504 void scan_unevictable_unregister_node(struct node *node)
3505 {
3506         sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
3507 }
3508 #endif