mm: vmscan: throttle reclaim if encountering too many dirty pages under writeback
[pandora-kernel.git] / mm / vmscan.c
1 /*
2  *  linux/mm/vmscan.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *
6  *  Swap reorganised 29.12.95, Stephen Tweedie.
7  *  kswapd added: 7.1.96  sct
8  *  Removed kswapd_ctl limits, and swap out as many pages as needed
9  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11  *  Multiqueue VM started 5.8.00, Rik van Riel.
12  */
13
14 #include <linux/mm.h>
15 #include <linux/module.h>
16 #include <linux/gfp.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h>  /* for try_to_release_page(),
27                                         buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/pagevec.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/compaction.h>
36 #include <linux/notifier.h>
37 #include <linux/rwsem.h>
38 #include <linux/delay.h>
39 #include <linux/kthread.h>
40 #include <linux/freezer.h>
41 #include <linux/memcontrol.h>
42 #include <linux/delayacct.h>
43 #include <linux/sysctl.h>
44 #include <linux/oom.h>
45 #include <linux/prefetch.h>
46
47 #include <asm/tlbflush.h>
48 #include <asm/div64.h>
49
50 #include <linux/swapops.h>
51
52 #include "internal.h"
53
54 #define CREATE_TRACE_POINTS
55 #include <trace/events/vmscan.h>
56
57 /*
58  * reclaim_mode determines how the inactive list is shrunk
59  * RECLAIM_MODE_SINGLE: Reclaim only order-0 pages
60  * RECLAIM_MODE_ASYNC:  Do not block
61  * RECLAIM_MODE_SYNC:   Allow blocking e.g. call wait_on_page_writeback
62  * RECLAIM_MODE_LUMPYRECLAIM: For high-order allocations, take a reference
63  *                      page from the LRU and reclaim all pages within a
64  *                      naturally aligned range
65  * RECLAIM_MODE_COMPACTION: For high-order allocations, reclaim a number of
66  *                      order-0 pages and then compact the zone
67  */
68 typedef unsigned __bitwise__ reclaim_mode_t;
69 #define RECLAIM_MODE_SINGLE             ((__force reclaim_mode_t)0x01u)
70 #define RECLAIM_MODE_ASYNC              ((__force reclaim_mode_t)0x02u)
71 #define RECLAIM_MODE_SYNC               ((__force reclaim_mode_t)0x04u)
72 #define RECLAIM_MODE_LUMPYRECLAIM       ((__force reclaim_mode_t)0x08u)
73 #define RECLAIM_MODE_COMPACTION         ((__force reclaim_mode_t)0x10u)
74
75 struct scan_control {
76         /* Incremented by the number of inactive pages that were scanned */
77         unsigned long nr_scanned;
78
79         /* Number of pages freed so far during a call to shrink_zones() */
80         unsigned long nr_reclaimed;
81
82         /* How many pages shrink_list() should reclaim */
83         unsigned long nr_to_reclaim;
84
85         unsigned long hibernation_mode;
86
87         /* This context's GFP mask */
88         gfp_t gfp_mask;
89
90         int may_writepage;
91
92         /* Can mapped pages be reclaimed? */
93         int may_unmap;
94
95         /* Can pages be swapped as part of reclaim? */
96         int may_swap;
97
98         int order;
99
100         /*
101          * Intend to reclaim enough continuous memory rather than reclaim
102          * enough amount of memory. i.e, mode for high order allocation.
103          */
104         reclaim_mode_t reclaim_mode;
105
106         /* Which cgroup do we reclaim from */
107         struct mem_cgroup *mem_cgroup;
108
109         /*
110          * Nodemask of nodes allowed by the caller. If NULL, all nodes
111          * are scanned.
112          */
113         nodemask_t      *nodemask;
114 };
115
116 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
117
118 #ifdef ARCH_HAS_PREFETCH
119 #define prefetch_prev_lru_page(_page, _base, _field)                    \
120         do {                                                            \
121                 if ((_page)->lru.prev != _base) {                       \
122                         struct page *prev;                              \
123                                                                         \
124                         prev = lru_to_page(&(_page->lru));              \
125                         prefetch(&prev->_field);                        \
126                 }                                                       \
127         } while (0)
128 #else
129 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
130 #endif
131
132 #ifdef ARCH_HAS_PREFETCHW
133 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
134         do {                                                            \
135                 if ((_page)->lru.prev != _base) {                       \
136                         struct page *prev;                              \
137                                                                         \
138                         prev = lru_to_page(&(_page->lru));              \
139                         prefetchw(&prev->_field);                       \
140                 }                                                       \
141         } while (0)
142 #else
143 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
144 #endif
145
146 /*
147  * From 0 .. 100.  Higher means more swappy.
148  */
149 int vm_swappiness = 60;
150 long vm_total_pages;    /* The total number of pages which the VM controls */
151
152 static LIST_HEAD(shrinker_list);
153 static DECLARE_RWSEM(shrinker_rwsem);
154
155 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
156 #define scanning_global_lru(sc) (!(sc)->mem_cgroup)
157 #else
158 #define scanning_global_lru(sc) (1)
159 #endif
160
161 static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
162                                                   struct scan_control *sc)
163 {
164         if (!scanning_global_lru(sc))
165                 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
166
167         return &zone->reclaim_stat;
168 }
169
170 static unsigned long zone_nr_lru_pages(struct zone *zone,
171                                 struct scan_control *sc, enum lru_list lru)
172 {
173         if (!scanning_global_lru(sc))
174                 return mem_cgroup_zone_nr_lru_pages(sc->mem_cgroup,
175                                 zone_to_nid(zone), zone_idx(zone), BIT(lru));
176
177         return zone_page_state(zone, NR_LRU_BASE + lru);
178 }
179
180
181 /*
182  * Add a shrinker callback to be called from the vm
183  */
184 void register_shrinker(struct shrinker *shrinker)
185 {
186         shrinker->nr = 0;
187         down_write(&shrinker_rwsem);
188         list_add_tail(&shrinker->list, &shrinker_list);
189         up_write(&shrinker_rwsem);
190 }
191 EXPORT_SYMBOL(register_shrinker);
192
193 /*
194  * Remove one
195  */
196 void unregister_shrinker(struct shrinker *shrinker)
197 {
198         down_write(&shrinker_rwsem);
199         list_del(&shrinker->list);
200         up_write(&shrinker_rwsem);
201 }
202 EXPORT_SYMBOL(unregister_shrinker);
203
204 static inline int do_shrinker_shrink(struct shrinker *shrinker,
205                                      struct shrink_control *sc,
206                                      unsigned long nr_to_scan)
207 {
208         sc->nr_to_scan = nr_to_scan;
209         return (*shrinker->shrink)(shrinker, sc);
210 }
211
212 #define SHRINK_BATCH 128
213 /*
214  * Call the shrink functions to age shrinkable caches
215  *
216  * Here we assume it costs one seek to replace a lru page and that it also
217  * takes a seek to recreate a cache object.  With this in mind we age equal
218  * percentages of the lru and ageable caches.  This should balance the seeks
219  * generated by these structures.
220  *
221  * If the vm encountered mapped pages on the LRU it increase the pressure on
222  * slab to avoid swapping.
223  *
224  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
225  *
226  * `lru_pages' represents the number of on-LRU pages in all the zones which
227  * are eligible for the caller's allocation attempt.  It is used for balancing
228  * slab reclaim versus page reclaim.
229  *
230  * Returns the number of slab objects which we shrunk.
231  */
232 unsigned long shrink_slab(struct shrink_control *shrink,
233                           unsigned long nr_pages_scanned,
234                           unsigned long lru_pages)
235 {
236         struct shrinker *shrinker;
237         unsigned long ret = 0;
238
239         if (nr_pages_scanned == 0)
240                 nr_pages_scanned = SWAP_CLUSTER_MAX;
241
242         if (!down_read_trylock(&shrinker_rwsem)) {
243                 /* Assume we'll be able to shrink next time */
244                 ret = 1;
245                 goto out;
246         }
247
248         list_for_each_entry(shrinker, &shrinker_list, list) {
249                 unsigned long long delta;
250                 unsigned long total_scan;
251                 unsigned long max_pass;
252                 int shrink_ret = 0;
253                 long nr;
254                 long new_nr;
255                 long batch_size = shrinker->batch ? shrinker->batch
256                                                   : SHRINK_BATCH;
257
258                 /*
259                  * copy the current shrinker scan count into a local variable
260                  * and zero it so that other concurrent shrinker invocations
261                  * don't also do this scanning work.
262                  */
263                 do {
264                         nr = shrinker->nr;
265                 } while (cmpxchg(&shrinker->nr, nr, 0) != nr);
266
267                 total_scan = nr;
268                 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
269                 delta = (4 * nr_pages_scanned) / shrinker->seeks;
270                 delta *= max_pass;
271                 do_div(delta, lru_pages + 1);
272                 total_scan += delta;
273                 if (total_scan < 0) {
274                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
275                                "delete nr=%ld\n",
276                                shrinker->shrink, total_scan);
277                         total_scan = max_pass;
278                 }
279
280                 /*
281                  * We need to avoid excessive windup on filesystem shrinkers
282                  * due to large numbers of GFP_NOFS allocations causing the
283                  * shrinkers to return -1 all the time. This results in a large
284                  * nr being built up so when a shrink that can do some work
285                  * comes along it empties the entire cache due to nr >>>
286                  * max_pass.  This is bad for sustaining a working set in
287                  * memory.
288                  *
289                  * Hence only allow the shrinker to scan the entire cache when
290                  * a large delta change is calculated directly.
291                  */
292                 if (delta < max_pass / 4)
293                         total_scan = min(total_scan, max_pass / 2);
294
295                 /*
296                  * Avoid risking looping forever due to too large nr value:
297                  * never try to free more than twice the estimate number of
298                  * freeable entries.
299                  */
300                 if (total_scan > max_pass * 2)
301                         total_scan = max_pass * 2;
302
303                 trace_mm_shrink_slab_start(shrinker, shrink, nr,
304                                         nr_pages_scanned, lru_pages,
305                                         max_pass, delta, total_scan);
306
307                 while (total_scan >= batch_size) {
308                         int nr_before;
309
310                         nr_before = do_shrinker_shrink(shrinker, shrink, 0);
311                         shrink_ret = do_shrinker_shrink(shrinker, shrink,
312                                                         batch_size);
313                         if (shrink_ret == -1)
314                                 break;
315                         if (shrink_ret < nr_before)
316                                 ret += nr_before - shrink_ret;
317                         count_vm_events(SLABS_SCANNED, batch_size);
318                         total_scan -= batch_size;
319
320                         cond_resched();
321                 }
322
323                 /*
324                  * move the unused scan count back into the shrinker in a
325                  * manner that handles concurrent updates. If we exhausted the
326                  * scan, there is no need to do an update.
327                  */
328                 do {
329                         nr = shrinker->nr;
330                         new_nr = total_scan + nr;
331                         if (total_scan <= 0)
332                                 break;
333                 } while (cmpxchg(&shrinker->nr, nr, new_nr) != nr);
334
335                 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
336         }
337         up_read(&shrinker_rwsem);
338 out:
339         cond_resched();
340         return ret;
341 }
342
343 static void set_reclaim_mode(int priority, struct scan_control *sc,
344                                    bool sync)
345 {
346         reclaim_mode_t syncmode = sync ? RECLAIM_MODE_SYNC : RECLAIM_MODE_ASYNC;
347
348         /*
349          * Initially assume we are entering either lumpy reclaim or
350          * reclaim/compaction.Depending on the order, we will either set the
351          * sync mode or just reclaim order-0 pages later.
352          */
353         if (COMPACTION_BUILD)
354                 sc->reclaim_mode = RECLAIM_MODE_COMPACTION;
355         else
356                 sc->reclaim_mode = RECLAIM_MODE_LUMPYRECLAIM;
357
358         /*
359          * Avoid using lumpy reclaim or reclaim/compaction if possible by
360          * restricting when its set to either costly allocations or when
361          * under memory pressure
362          */
363         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
364                 sc->reclaim_mode |= syncmode;
365         else if (sc->order && priority < DEF_PRIORITY - 2)
366                 sc->reclaim_mode |= syncmode;
367         else
368                 sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
369 }
370
371 static void reset_reclaim_mode(struct scan_control *sc)
372 {
373         sc->reclaim_mode = RECLAIM_MODE_SINGLE | RECLAIM_MODE_ASYNC;
374 }
375
376 static inline int is_page_cache_freeable(struct page *page)
377 {
378         /*
379          * A freeable page cache page is referenced only by the caller
380          * that isolated the page, the page cache radix tree and
381          * optional buffer heads at page->private.
382          */
383         return page_count(page) - page_has_private(page) == 2;
384 }
385
386 static int may_write_to_queue(struct backing_dev_info *bdi,
387                               struct scan_control *sc)
388 {
389         if (current->flags & PF_SWAPWRITE)
390                 return 1;
391         if (!bdi_write_congested(bdi))
392                 return 1;
393         if (bdi == current->backing_dev_info)
394                 return 1;
395
396         /* lumpy reclaim for hugepage often need a lot of write */
397         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
398                 return 1;
399         return 0;
400 }
401
402 /*
403  * We detected a synchronous write error writing a page out.  Probably
404  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
405  * fsync(), msync() or close().
406  *
407  * The tricky part is that after writepage we cannot touch the mapping: nothing
408  * prevents it from being freed up.  But we have a ref on the page and once
409  * that page is locked, the mapping is pinned.
410  *
411  * We're allowed to run sleeping lock_page() here because we know the caller has
412  * __GFP_FS.
413  */
414 static void handle_write_error(struct address_space *mapping,
415                                 struct page *page, int error)
416 {
417         lock_page(page);
418         if (page_mapping(page) == mapping)
419                 mapping_set_error(mapping, error);
420         unlock_page(page);
421 }
422
423 /* possible outcome of pageout() */
424 typedef enum {
425         /* failed to write page out, page is locked */
426         PAGE_KEEP,
427         /* move page to the active list, page is locked */
428         PAGE_ACTIVATE,
429         /* page has been sent to the disk successfully, page is unlocked */
430         PAGE_SUCCESS,
431         /* page is clean and locked */
432         PAGE_CLEAN,
433 } pageout_t;
434
435 /*
436  * pageout is called by shrink_page_list() for each dirty page.
437  * Calls ->writepage().
438  */
439 static pageout_t pageout(struct page *page, struct address_space *mapping,
440                          struct scan_control *sc)
441 {
442         /*
443          * If the page is dirty, only perform writeback if that write
444          * will be non-blocking.  To prevent this allocation from being
445          * stalled by pagecache activity.  But note that there may be
446          * stalls if we need to run get_block().  We could test
447          * PagePrivate for that.
448          *
449          * If this process is currently in __generic_file_aio_write() against
450          * this page's queue, we can perform writeback even if that
451          * will block.
452          *
453          * If the page is swapcache, write it back even if that would
454          * block, for some throttling. This happens by accident, because
455          * swap_backing_dev_info is bust: it doesn't reflect the
456          * congestion state of the swapdevs.  Easy to fix, if needed.
457          */
458         if (!is_page_cache_freeable(page))
459                 return PAGE_KEEP;
460         if (!mapping) {
461                 /*
462                  * Some data journaling orphaned pages can have
463                  * page->mapping == NULL while being dirty with clean buffers.
464                  */
465                 if (page_has_private(page)) {
466                         if (try_to_free_buffers(page)) {
467                                 ClearPageDirty(page);
468                                 printk("%s: orphaned page\n", __func__);
469                                 return PAGE_CLEAN;
470                         }
471                 }
472                 return PAGE_KEEP;
473         }
474         if (mapping->a_ops->writepage == NULL)
475                 return PAGE_ACTIVATE;
476         if (!may_write_to_queue(mapping->backing_dev_info, sc))
477                 return PAGE_KEEP;
478
479         if (clear_page_dirty_for_io(page)) {
480                 int res;
481                 struct writeback_control wbc = {
482                         .sync_mode = WB_SYNC_NONE,
483                         .nr_to_write = SWAP_CLUSTER_MAX,
484                         .range_start = 0,
485                         .range_end = LLONG_MAX,
486                         .for_reclaim = 1,
487                 };
488
489                 SetPageReclaim(page);
490                 res = mapping->a_ops->writepage(page, &wbc);
491                 if (res < 0)
492                         handle_write_error(mapping, page, res);
493                 if (res == AOP_WRITEPAGE_ACTIVATE) {
494                         ClearPageReclaim(page);
495                         return PAGE_ACTIVATE;
496                 }
497
498                 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 clearing (page is
637                  * unlocked), make sure that if the other thread does
638                  * not observe our setting of PG_lru and fails
639                  * isolation, we see PG_mlocked cleared below and move
640                  * the page back to the evictable list.
641                  *
642                  * The other side is TestClearPageMlocked().
643                  */
644                 smp_mb();
645         }
646
647         /*
648          * page's status can change while we move it among lru. If an evictable
649          * page is on unevictable list, it never be freed. To avoid that,
650          * check after we added it to the list, again.
651          */
652         if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
653                 if (!isolate_lru_page(page)) {
654                         put_page(page);
655                         goto redo;
656                 }
657                 /* This means someone else dropped this page from LRU
658                  * So, it will be freed or putback to LRU again. There is
659                  * nothing to do here.
660                  */
661         }
662
663         if (was_unevictable && lru != LRU_UNEVICTABLE)
664                 count_vm_event(UNEVICTABLE_PGRESCUED);
665         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
666                 count_vm_event(UNEVICTABLE_PGCULLED);
667
668         put_page(page);         /* drop ref from isolate */
669 }
670
671 enum page_references {
672         PAGEREF_RECLAIM,
673         PAGEREF_RECLAIM_CLEAN,
674         PAGEREF_KEEP,
675         PAGEREF_ACTIVATE,
676 };
677
678 static enum page_references page_check_references(struct page *page,
679                                                   struct scan_control *sc)
680 {
681         int referenced_ptes, referenced_page;
682         unsigned long vm_flags;
683
684         referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
685         referenced_page = TestClearPageReferenced(page);
686
687         /* Lumpy reclaim - ignore references */
688         if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
689                 return PAGEREF_RECLAIM;
690
691         /*
692          * Mlock lost the isolation race with us.  Let try_to_unmap()
693          * move the page to the unevictable list.
694          */
695         if (vm_flags & VM_LOCKED)
696                 return PAGEREF_RECLAIM;
697
698         if (referenced_ptes) {
699                 if (PageAnon(page))
700                         return PAGEREF_ACTIVATE;
701                 /*
702                  * All mapped pages start out with page table
703                  * references from the instantiating fault, so we need
704                  * to look twice if a mapped file page is used more
705                  * than once.
706                  *
707                  * Mark it and spare it for another trip around the
708                  * inactive list.  Another page table reference will
709                  * lead to its activation.
710                  *
711                  * Note: the mark is set for activated pages as well
712                  * so that recently deactivated but used pages are
713                  * quickly recovered.
714                  */
715                 SetPageReferenced(page);
716
717                 if (referenced_page)
718                         return PAGEREF_ACTIVATE;
719
720                 return PAGEREF_KEEP;
721         }
722
723         /* Reclaim if clean, defer dirty pages to writeback */
724         if (referenced_page && !PageSwapBacked(page))
725                 return PAGEREF_RECLAIM_CLEAN;
726
727         return PAGEREF_RECLAIM;
728 }
729
730 static noinline_for_stack void free_page_list(struct list_head *free_pages)
731 {
732         struct pagevec freed_pvec;
733         struct page *page, *tmp;
734
735         pagevec_init(&freed_pvec, 1);
736
737         list_for_each_entry_safe(page, tmp, free_pages, lru) {
738                 list_del(&page->lru);
739                 if (!pagevec_add(&freed_pvec, page)) {
740                         __pagevec_free(&freed_pvec);
741                         pagevec_reinit(&freed_pvec);
742                 }
743         }
744
745         pagevec_free(&freed_pvec);
746 }
747
748 /*
749  * shrink_page_list() returns the number of reclaimed pages
750  */
751 static unsigned long shrink_page_list(struct list_head *page_list,
752                                       struct zone *zone,
753                                       struct scan_control *sc,
754                                       int priority,
755                                       unsigned long *ret_nr_dirty,
756                                       unsigned long *ret_nr_writeback)
757 {
758         LIST_HEAD(ret_pages);
759         LIST_HEAD(free_pages);
760         int pgactivate = 0;
761         unsigned long nr_dirty = 0;
762         unsigned long nr_congested = 0;
763         unsigned long nr_reclaimed = 0;
764         unsigned long nr_writeback = 0;
765
766         cond_resched();
767
768         while (!list_empty(page_list)) {
769                 enum page_references references;
770                 struct address_space *mapping;
771                 struct page *page;
772                 int may_enter_fs;
773
774                 cond_resched();
775
776                 page = lru_to_page(page_list);
777                 list_del(&page->lru);
778
779                 if (!trylock_page(page))
780                         goto keep;
781
782                 VM_BUG_ON(PageActive(page));
783                 VM_BUG_ON(page_zone(page) != zone);
784
785                 sc->nr_scanned++;
786
787                 if (unlikely(!page_evictable(page, NULL)))
788                         goto cull_mlocked;
789
790                 if (!sc->may_unmap && page_mapped(page))
791                         goto keep_locked;
792
793                 /* Double the slab pressure for mapped and swapcache pages */
794                 if (page_mapped(page) || PageSwapCache(page))
795                         sc->nr_scanned++;
796
797                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
798                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
799
800                 if (PageWriteback(page)) {
801                         nr_writeback++;
802                         /*
803                          * Synchronous reclaim cannot queue pages for
804                          * writeback due to the possibility of stack overflow
805                          * but if it encounters a page under writeback, wait
806                          * for the IO to complete.
807                          */
808                         if ((sc->reclaim_mode & RECLAIM_MODE_SYNC) &&
809                             may_enter_fs)
810                                 wait_on_page_writeback(page);
811                         else {
812                                 unlock_page(page);
813                                 goto keep_lumpy;
814                         }
815                 }
816
817                 references = page_check_references(page, sc);
818                 switch (references) {
819                 case PAGEREF_ACTIVATE:
820                         goto activate_locked;
821                 case PAGEREF_KEEP:
822                         goto keep_locked;
823                 case PAGEREF_RECLAIM:
824                 case PAGEREF_RECLAIM_CLEAN:
825                         ; /* try to reclaim the page below */
826                 }
827
828                 /*
829                  * Anonymous process memory has backing store?
830                  * Try to allocate it some swap space here.
831                  */
832                 if (PageAnon(page) && !PageSwapCache(page)) {
833                         if (!(sc->gfp_mask & __GFP_IO))
834                                 goto keep_locked;
835                         if (!add_to_swap(page))
836                                 goto activate_locked;
837                         may_enter_fs = 1;
838                 }
839
840                 mapping = page_mapping(page);
841
842                 /*
843                  * The page is mapped into the page tables of one or more
844                  * processes. Try to unmap it here.
845                  */
846                 if (page_mapped(page) && mapping) {
847                         switch (try_to_unmap(page, TTU_UNMAP)) {
848                         case SWAP_FAIL:
849                                 goto activate_locked;
850                         case SWAP_AGAIN:
851                                 goto keep_locked;
852                         case SWAP_MLOCK:
853                                 goto cull_mlocked;
854                         case SWAP_SUCCESS:
855                                 ; /* try to free the page below */
856                         }
857                 }
858
859                 if (PageDirty(page)) {
860                         nr_dirty++;
861
862                         /*
863                          * Only kswapd can writeback filesystem pages to
864                          * avoid risk of stack overflow but do not writeback
865                          * unless under significant pressure.
866                          */
867                         if (page_is_file_cache(page) &&
868                                         (!current_is_kswapd() || priority >= DEF_PRIORITY - 2)) {
869                                 inc_zone_page_state(page, NR_VMSCAN_WRITE_SKIP);
870                                 goto keep_locked;
871                         }
872
873                         if (references == PAGEREF_RECLAIM_CLEAN)
874                                 goto keep_locked;
875                         if (!may_enter_fs)
876                                 goto keep_locked;
877                         if (!sc->may_writepage)
878                                 goto keep_locked;
879
880                         /* Page is dirty, try to write it out here */
881                         switch (pageout(page, mapping, sc)) {
882                         case PAGE_KEEP:
883                                 nr_congested++;
884                                 goto keep_locked;
885                         case PAGE_ACTIVATE:
886                                 goto activate_locked;
887                         case PAGE_SUCCESS:
888                                 if (PageWriteback(page))
889                                         goto keep_lumpy;
890                                 if (PageDirty(page))
891                                         goto keep;
892
893                                 /*
894                                  * A synchronous write - probably a ramdisk.  Go
895                                  * ahead and try to reclaim the page.
896                                  */
897                                 if (!trylock_page(page))
898                                         goto keep;
899                                 if (PageDirty(page) || PageWriteback(page))
900                                         goto keep_locked;
901                                 mapping = page_mapping(page);
902                         case PAGE_CLEAN:
903                                 ; /* try to free the page below */
904                         }
905                 }
906
907                 /*
908                  * If the page has buffers, try to free the buffer mappings
909                  * associated with this page. If we succeed we try to free
910                  * the page as well.
911                  *
912                  * We do this even if the page is PageDirty().
913                  * try_to_release_page() does not perform I/O, but it is
914                  * possible for a page to have PageDirty set, but it is actually
915                  * clean (all its buffers are clean).  This happens if the
916                  * buffers were written out directly, with submit_bh(). ext3
917                  * will do this, as well as the blockdev mapping.
918                  * try_to_release_page() will discover that cleanness and will
919                  * drop the buffers and mark the page clean - it can be freed.
920                  *
921                  * Rarely, pages can have buffers and no ->mapping.  These are
922                  * the pages which were not successfully invalidated in
923                  * truncate_complete_page().  We try to drop those buffers here
924                  * and if that worked, and the page is no longer mapped into
925                  * process address space (page_count == 1) it can be freed.
926                  * Otherwise, leave the page on the LRU so it is swappable.
927                  */
928                 if (page_has_private(page)) {
929                         if (!try_to_release_page(page, sc->gfp_mask))
930                                 goto activate_locked;
931                         if (!mapping && page_count(page) == 1) {
932                                 unlock_page(page);
933                                 if (put_page_testzero(page))
934                                         goto free_it;
935                                 else {
936                                         /*
937                                          * rare race with speculative reference.
938                                          * the speculative reference will free
939                                          * this page shortly, so we may
940                                          * increment nr_reclaimed here (and
941                                          * leave it off the LRU).
942                                          */
943                                         nr_reclaimed++;
944                                         continue;
945                                 }
946                         }
947                 }
948
949                 if (!mapping || !__remove_mapping(mapping, page))
950                         goto keep_locked;
951
952                 /*
953                  * At this point, we have no other references and there is
954                  * no way to pick any more up (removed from LRU, removed
955                  * from pagecache). Can use non-atomic bitops now (and
956                  * we obviously don't have to worry about waking up a process
957                  * waiting on the page lock, because there are no references.
958                  */
959                 __clear_page_locked(page);
960 free_it:
961                 nr_reclaimed++;
962
963                 /*
964                  * Is there need to periodically free_page_list? It would
965                  * appear not as the counts should be low
966                  */
967                 list_add(&page->lru, &free_pages);
968                 continue;
969
970 cull_mlocked:
971                 if (PageSwapCache(page))
972                         try_to_free_swap(page);
973                 unlock_page(page);
974                 putback_lru_page(page);
975                 reset_reclaim_mode(sc);
976                 continue;
977
978 activate_locked:
979                 /* Not a candidate for swapping, so reclaim swap space. */
980                 if (PageSwapCache(page) && vm_swap_full())
981                         try_to_free_swap(page);
982                 VM_BUG_ON(PageActive(page));
983                 SetPageActive(page);
984                 pgactivate++;
985 keep_locked:
986                 unlock_page(page);
987 keep:
988                 reset_reclaim_mode(sc);
989 keep_lumpy:
990                 list_add(&page->lru, &ret_pages);
991                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
992         }
993
994         /*
995          * Tag a zone as congested if all the dirty pages encountered were
996          * backed by a congested BDI. In this case, reclaimers should just
997          * back off and wait for congestion to clear because further reclaim
998          * will encounter the same problem
999          */
1000         if (nr_dirty && nr_dirty == nr_congested && scanning_global_lru(sc))
1001                 zone_set_flag(zone, ZONE_CONGESTED);
1002
1003         free_page_list(&free_pages);
1004
1005         list_splice(&ret_pages, page_list);
1006         count_vm_events(PGACTIVATE, pgactivate);
1007         *ret_nr_dirty += nr_dirty;
1008         *ret_nr_writeback += nr_writeback;
1009         return nr_reclaimed;
1010 }
1011
1012 /*
1013  * Attempt to remove the specified page from its LRU.  Only take this page
1014  * if it is of the appropriate PageActive status.  Pages which are being
1015  * freed elsewhere are also ignored.
1016  *
1017  * page:        page to consider
1018  * mode:        one of the LRU isolation modes defined above
1019  *
1020  * returns 0 on success, -ve errno on failure.
1021  */
1022 int __isolate_lru_page(struct page *page, isolate_mode_t mode, int file)
1023 {
1024         bool all_lru_mode;
1025         int ret = -EINVAL;
1026
1027         /* Only take pages on the LRU. */
1028         if (!PageLRU(page))
1029                 return ret;
1030
1031         all_lru_mode = (mode & (ISOLATE_ACTIVE|ISOLATE_INACTIVE)) ==
1032                 (ISOLATE_ACTIVE|ISOLATE_INACTIVE);
1033
1034         /*
1035          * When checking the active state, we need to be sure we are
1036          * dealing with comparible boolean values.  Take the logical not
1037          * of each.
1038          */
1039         if (!all_lru_mode && !PageActive(page) != !(mode & ISOLATE_ACTIVE))
1040                 return ret;
1041
1042         if (!all_lru_mode && !!page_is_file_cache(page) != file)
1043                 return ret;
1044
1045         /*
1046          * When this function is being called for lumpy reclaim, we
1047          * initially look into all LRU pages, active, inactive and
1048          * unevictable; only give shrink_page_list evictable pages.
1049          */
1050         if (PageUnevictable(page))
1051                 return ret;
1052
1053         ret = -EBUSY;
1054
1055         if ((mode & ISOLATE_CLEAN) && (PageDirty(page) || PageWriteback(page)))
1056                 return ret;
1057
1058         if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1059                 return ret;
1060
1061         if (likely(get_page_unless_zero(page))) {
1062                 /*
1063                  * Be careful not to clear PageLRU until after we're
1064                  * sure the page is not being freed elsewhere -- the
1065                  * page release code relies on it.
1066                  */
1067                 ClearPageLRU(page);
1068                 ret = 0;
1069         }
1070
1071         return ret;
1072 }
1073
1074 /*
1075  * zone->lru_lock is heavily contended.  Some of the functions that
1076  * shrink the lists perform better by taking out a batch of pages
1077  * and working on them outside the LRU lock.
1078  *
1079  * For pagecache intensive workloads, this function is the hottest
1080  * spot in the kernel (apart from copy_*_user functions).
1081  *
1082  * Appropriate locks must be held before calling this function.
1083  *
1084  * @nr_to_scan: The number of pages to look through on the list.
1085  * @src:        The LRU list to pull pages off.
1086  * @dst:        The temp list to put pages on to.
1087  * @scanned:    The number of pages that were scanned.
1088  * @order:      The caller's attempted allocation order
1089  * @mode:       One of the LRU isolation modes
1090  * @file:       True [1] if isolating file [!anon] pages
1091  *
1092  * returns how many pages were moved onto *@dst.
1093  */
1094 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1095                 struct list_head *src, struct list_head *dst,
1096                 unsigned long *scanned, int order, isolate_mode_t mode,
1097                 int file)
1098 {
1099         unsigned long nr_taken = 0;
1100         unsigned long nr_lumpy_taken = 0;
1101         unsigned long nr_lumpy_dirty = 0;
1102         unsigned long nr_lumpy_failed = 0;
1103         unsigned long scan;
1104
1105         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1106                 struct page *page;
1107                 unsigned long pfn;
1108                 unsigned long end_pfn;
1109                 unsigned long page_pfn;
1110                 int zone_id;
1111
1112                 page = lru_to_page(src);
1113                 prefetchw_prev_lru_page(page, src, flags);
1114
1115                 VM_BUG_ON(!PageLRU(page));
1116
1117                 switch (__isolate_lru_page(page, mode, file)) {
1118                 case 0:
1119                         list_move(&page->lru, dst);
1120                         mem_cgroup_del_lru(page);
1121                         nr_taken += hpage_nr_pages(page);
1122                         break;
1123
1124                 case -EBUSY:
1125                         /* else it is being freed elsewhere */
1126                         list_move(&page->lru, src);
1127                         mem_cgroup_rotate_lru_list(page, page_lru(page));
1128                         continue;
1129
1130                 default:
1131                         BUG();
1132                 }
1133
1134                 if (!order)
1135                         continue;
1136
1137                 /*
1138                  * Attempt to take all pages in the order aligned region
1139                  * surrounding the tag page.  Only take those pages of
1140                  * the same active state as that tag page.  We may safely
1141                  * round the target page pfn down to the requested order
1142                  * as the mem_map is guaranteed valid out to MAX_ORDER,
1143                  * where that page is in a different zone we will detect
1144                  * it from its zone id and abort this block scan.
1145                  */
1146                 zone_id = page_zone_id(page);
1147                 page_pfn = page_to_pfn(page);
1148                 pfn = page_pfn & ~((1 << order) - 1);
1149                 end_pfn = pfn + (1 << order);
1150                 for (; pfn < end_pfn; pfn++) {
1151                         struct page *cursor_page;
1152
1153                         /* The target page is in the block, ignore it. */
1154                         if (unlikely(pfn == page_pfn))
1155                                 continue;
1156
1157                         /* Avoid holes within the zone. */
1158                         if (unlikely(!pfn_valid_within(pfn)))
1159                                 break;
1160
1161                         cursor_page = pfn_to_page(pfn);
1162
1163                         /* Check that we have not crossed a zone boundary. */
1164                         if (unlikely(page_zone_id(cursor_page) != zone_id))
1165                                 break;
1166
1167                         /*
1168                          * If we don't have enough swap space, reclaiming of
1169                          * anon page which don't already have a swap slot is
1170                          * pointless.
1171                          */
1172                         if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
1173                             !PageSwapCache(cursor_page))
1174                                 break;
1175
1176                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
1177                                 list_move(&cursor_page->lru, dst);
1178                                 mem_cgroup_del_lru(cursor_page);
1179                                 nr_taken += hpage_nr_pages(page);
1180                                 nr_lumpy_taken++;
1181                                 if (PageDirty(cursor_page))
1182                                         nr_lumpy_dirty++;
1183                                 scan++;
1184                         } else {
1185                                 /*
1186                                  * Check if the page is freed already.
1187                                  *
1188                                  * We can't use page_count() as that
1189                                  * requires compound_head and we don't
1190                                  * have a pin on the page here. If a
1191                                  * page is tail, we may or may not
1192                                  * have isolated the head, so assume
1193                                  * it's not free, it'd be tricky to
1194                                  * track the head status without a
1195                                  * page pin.
1196                                  */
1197                                 if (!PageTail(cursor_page) &&
1198                                     !atomic_read(&cursor_page->_count))
1199                                         continue;
1200                                 break;
1201                         }
1202                 }
1203
1204                 /* If we break out of the loop above, lumpy reclaim failed */
1205                 if (pfn < end_pfn)
1206                         nr_lumpy_failed++;
1207         }
1208
1209         *scanned = scan;
1210
1211         trace_mm_vmscan_lru_isolate(order,
1212                         nr_to_scan, scan,
1213                         nr_taken,
1214                         nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1215                         mode);
1216         return nr_taken;
1217 }
1218
1219 static unsigned long isolate_pages_global(unsigned long nr,
1220                                         struct list_head *dst,
1221                                         unsigned long *scanned, int order,
1222                                         isolate_mode_t mode,
1223                                         struct zone *z, int active, int file)
1224 {
1225         int lru = LRU_BASE;
1226         if (active)
1227                 lru += LRU_ACTIVE;
1228         if (file)
1229                 lru += LRU_FILE;
1230         return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
1231                                                                 mode, file);
1232 }
1233
1234 /*
1235  * clear_active_flags() is a helper for shrink_active_list(), clearing
1236  * any active bits from the pages in the list.
1237  */
1238 static unsigned long clear_active_flags(struct list_head *page_list,
1239                                         unsigned int *count)
1240 {
1241         int nr_active = 0;
1242         int lru;
1243         struct page *page;
1244
1245         list_for_each_entry(page, page_list, lru) {
1246                 int numpages = hpage_nr_pages(page);
1247                 lru = page_lru_base_type(page);
1248                 if (PageActive(page)) {
1249                         lru += LRU_ACTIVE;
1250                         ClearPageActive(page);
1251                         nr_active += numpages;
1252                 }
1253                 if (count)
1254                         count[lru] += numpages;
1255         }
1256
1257         return nr_active;
1258 }
1259
1260 /**
1261  * isolate_lru_page - tries to isolate a page from its LRU list
1262  * @page: page to isolate from its LRU list
1263  *
1264  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1265  * vmstat statistic corresponding to whatever LRU list the page was on.
1266  *
1267  * Returns 0 if the page was removed from an LRU list.
1268  * Returns -EBUSY if the page was not on an LRU list.
1269  *
1270  * The returned page will have PageLRU() cleared.  If it was found on
1271  * the active list, it will have PageActive set.  If it was found on
1272  * the unevictable list, it will have the PageUnevictable bit set. That flag
1273  * may need to be cleared by the caller before letting the page go.
1274  *
1275  * The vmstat statistic corresponding to the list on which the page was
1276  * found will be decremented.
1277  *
1278  * Restrictions:
1279  * (1) Must be called with an elevated refcount on the page. This is a
1280  *     fundamentnal difference from isolate_lru_pages (which is called
1281  *     without a stable reference).
1282  * (2) the lru_lock must not be held.
1283  * (3) interrupts must be enabled.
1284  */
1285 int isolate_lru_page(struct page *page)
1286 {
1287         int ret = -EBUSY;
1288
1289         VM_BUG_ON(!page_count(page));
1290
1291         if (PageLRU(page)) {
1292                 struct zone *zone = page_zone(page);
1293
1294                 spin_lock_irq(&zone->lru_lock);
1295                 if (PageLRU(page)) {
1296                         int lru = page_lru(page);
1297                         ret = 0;
1298                         get_page(page);
1299                         ClearPageLRU(page);
1300
1301                         del_page_from_lru_list(zone, page, lru);
1302                 }
1303                 spin_unlock_irq(&zone->lru_lock);
1304         }
1305         return ret;
1306 }
1307
1308 /*
1309  * Are there way too many processes in the direct reclaim path already?
1310  */
1311 static int too_many_isolated(struct zone *zone, int file,
1312                 struct scan_control *sc)
1313 {
1314         unsigned long inactive, isolated;
1315
1316         if (current_is_kswapd())
1317                 return 0;
1318
1319         if (!scanning_global_lru(sc))
1320                 return 0;
1321
1322         if (file) {
1323                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1324                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1325         } else {
1326                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1327                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1328         }
1329
1330         return isolated > inactive;
1331 }
1332
1333 /*
1334  * TODO: Try merging with migrations version of putback_lru_pages
1335  */
1336 static noinline_for_stack void
1337 putback_lru_pages(struct zone *zone, struct scan_control *sc,
1338                                 unsigned long nr_anon, unsigned long nr_file,
1339                                 struct list_head *page_list)
1340 {
1341         struct page *page;
1342         struct pagevec pvec;
1343         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1344
1345         pagevec_init(&pvec, 1);
1346
1347         /*
1348          * Put back any unfreeable pages.
1349          */
1350         spin_lock(&zone->lru_lock);
1351         while (!list_empty(page_list)) {
1352                 int lru;
1353                 page = lru_to_page(page_list);
1354                 VM_BUG_ON(PageLRU(page));
1355                 list_del(&page->lru);
1356                 if (unlikely(!page_evictable(page, NULL))) {
1357                         spin_unlock_irq(&zone->lru_lock);
1358                         putback_lru_page(page);
1359                         spin_lock_irq(&zone->lru_lock);
1360                         continue;
1361                 }
1362                 SetPageLRU(page);
1363                 lru = page_lru(page);
1364                 add_page_to_lru_list(zone, page, lru);
1365                 if (is_active_lru(lru)) {
1366                         int file = is_file_lru(lru);
1367                         int numpages = hpage_nr_pages(page);
1368                         reclaim_stat->recent_rotated[file] += numpages;
1369                 }
1370                 if (!pagevec_add(&pvec, page)) {
1371                         spin_unlock_irq(&zone->lru_lock);
1372                         __pagevec_release(&pvec);
1373                         spin_lock_irq(&zone->lru_lock);
1374                 }
1375         }
1376         __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1377         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1378
1379         spin_unlock_irq(&zone->lru_lock);
1380         pagevec_release(&pvec);
1381 }
1382
1383 static noinline_for_stack void update_isolated_counts(struct zone *zone,
1384                                         struct scan_control *sc,
1385                                         unsigned long *nr_anon,
1386                                         unsigned long *nr_file,
1387                                         struct list_head *isolated_list)
1388 {
1389         unsigned long nr_active;
1390         unsigned int count[NR_LRU_LISTS] = { 0, };
1391         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1392
1393         nr_active = clear_active_flags(isolated_list, count);
1394         __count_vm_events(PGDEACTIVATE, nr_active);
1395
1396         __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1397                               -count[LRU_ACTIVE_FILE]);
1398         __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1399                               -count[LRU_INACTIVE_FILE]);
1400         __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1401                               -count[LRU_ACTIVE_ANON]);
1402         __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1403                               -count[LRU_INACTIVE_ANON]);
1404
1405         *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1406         *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1407         __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1408         __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1409
1410         reclaim_stat->recent_scanned[0] += *nr_anon;
1411         reclaim_stat->recent_scanned[1] += *nr_file;
1412 }
1413
1414 /*
1415  * Returns true if a direct reclaim should wait on pages under writeback.
1416  *
1417  * If we are direct reclaiming for contiguous pages and we do not reclaim
1418  * everything in the list, try again and wait for writeback IO to complete.
1419  * This will stall high-order allocations noticeably. Only do that when really
1420  * need to free the pages under high memory pressure.
1421  */
1422 static inline bool should_reclaim_stall(unsigned long nr_taken,
1423                                         unsigned long nr_freed,
1424                                         int priority,
1425                                         struct scan_control *sc)
1426 {
1427         int lumpy_stall_priority;
1428
1429         /* kswapd should not stall on sync IO */
1430         if (current_is_kswapd())
1431                 return false;
1432
1433         /* Only stall on lumpy reclaim */
1434         if (sc->reclaim_mode & RECLAIM_MODE_SINGLE)
1435                 return false;
1436
1437         /* If we have reclaimed everything on the isolated list, no stall */
1438         if (nr_freed == nr_taken)
1439                 return false;
1440
1441         /*
1442          * For high-order allocations, there are two stall thresholds.
1443          * High-cost allocations stall immediately where as lower
1444          * order allocations such as stacks require the scanning
1445          * priority to be much higher before stalling.
1446          */
1447         if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1448                 lumpy_stall_priority = DEF_PRIORITY;
1449         else
1450                 lumpy_stall_priority = DEF_PRIORITY / 3;
1451
1452         return priority <= lumpy_stall_priority;
1453 }
1454
1455 /*
1456  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1457  * of reclaimed pages
1458  */
1459 static noinline_for_stack unsigned long
1460 shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1461                         struct scan_control *sc, int priority, int file)
1462 {
1463         LIST_HEAD(page_list);
1464         unsigned long nr_scanned;
1465         unsigned long nr_reclaimed = 0;
1466         unsigned long nr_taken;
1467         unsigned long nr_anon;
1468         unsigned long nr_file;
1469         unsigned long nr_dirty = 0;
1470         unsigned long nr_writeback = 0;
1471         isolate_mode_t reclaim_mode = ISOLATE_INACTIVE;
1472
1473         while (unlikely(too_many_isolated(zone, file, sc))) {
1474                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1475
1476                 /* We are about to die and free our memory. Return now. */
1477                 if (fatal_signal_pending(current))
1478                         return SWAP_CLUSTER_MAX;
1479         }
1480
1481         set_reclaim_mode(priority, sc, false);
1482         if (sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM)
1483                 reclaim_mode |= ISOLATE_ACTIVE;
1484
1485         lru_add_drain();
1486
1487         if (!sc->may_unmap)
1488                 reclaim_mode |= ISOLATE_UNMAPPED;
1489         if (!sc->may_writepage)
1490                 reclaim_mode |= ISOLATE_CLEAN;
1491
1492         spin_lock_irq(&zone->lru_lock);
1493
1494         if (scanning_global_lru(sc)) {
1495                 nr_taken = isolate_pages_global(nr_to_scan, &page_list,
1496                         &nr_scanned, sc->order, reclaim_mode, zone, 0, file);
1497                 zone->pages_scanned += nr_scanned;
1498                 if (current_is_kswapd())
1499                         __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1500                                                nr_scanned);
1501                 else
1502                         __count_zone_vm_events(PGSCAN_DIRECT, zone,
1503                                                nr_scanned);
1504         } else {
1505                 nr_taken = mem_cgroup_isolate_pages(nr_to_scan, &page_list,
1506                         &nr_scanned, sc->order, reclaim_mode, zone,
1507                         sc->mem_cgroup, 0, file);
1508                 /*
1509                  * mem_cgroup_isolate_pages() keeps track of
1510                  * scanned pages on its own.
1511                  */
1512         }
1513
1514         if (nr_taken == 0) {
1515                 spin_unlock_irq(&zone->lru_lock);
1516                 return 0;
1517         }
1518
1519         update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
1520
1521         spin_unlock_irq(&zone->lru_lock);
1522
1523         nr_reclaimed = shrink_page_list(&page_list, zone, sc, priority,
1524                                                 &nr_dirty, &nr_writeback);
1525
1526         /* Check if we should syncronously wait for writeback */
1527         if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
1528                 set_reclaim_mode(priority, sc, true);
1529                 nr_reclaimed += shrink_page_list(&page_list, zone, sc,
1530                                         priority, &nr_dirty, &nr_writeback);
1531         }
1532
1533         local_irq_disable();
1534         if (current_is_kswapd())
1535                 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1536         __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1537
1538         putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
1539
1540         /*
1541          * If reclaim is isolating dirty pages under writeback, it implies
1542          * that the long-lived page allocation rate is exceeding the page
1543          * laundering rate. Either the global limits are not being effective
1544          * at throttling processes due to the page distribution throughout
1545          * zones or there is heavy usage of a slow backing device. The
1546          * only option is to throttle from reclaim context which is not ideal
1547          * as there is no guarantee the dirtying process is throttled in the
1548          * same way balance_dirty_pages() manages.
1549          *
1550          * This scales the number of dirty pages that must be under writeback
1551          * before throttling depending on priority. It is a simple backoff
1552          * function that has the most effect in the range DEF_PRIORITY to
1553          * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1554          * in trouble and reclaim is considered to be in trouble.
1555          *
1556          * DEF_PRIORITY   100% isolated pages must be PageWriteback to throttle
1557          * DEF_PRIORITY-1  50% must be PageWriteback
1558          * DEF_PRIORITY-2  25% must be PageWriteback, kswapd in trouble
1559          * ...
1560          * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1561          *                     isolated page is PageWriteback
1562          */
1563         if (nr_writeback && nr_writeback >= (nr_taken >> (DEF_PRIORITY-priority)))
1564                 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1565
1566         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1567                 zone_idx(zone),
1568                 nr_scanned, nr_reclaimed,
1569                 priority,
1570                 trace_shrink_flags(file, sc->reclaim_mode));
1571         return nr_reclaimed;
1572 }
1573
1574 /*
1575  * This moves pages from the active list to the inactive list.
1576  *
1577  * We move them the other way if the page is referenced by one or more
1578  * processes, from rmap.
1579  *
1580  * If the pages are mostly unmapped, the processing is fast and it is
1581  * appropriate to hold zone->lru_lock across the whole operation.  But if
1582  * the pages are mapped, the processing is slow (page_referenced()) so we
1583  * should drop zone->lru_lock around each page.  It's impossible to balance
1584  * this, so instead we remove the pages from the LRU while processing them.
1585  * It is safe to rely on PG_active against the non-LRU pages in here because
1586  * nobody will play with that bit on a non-LRU page.
1587  *
1588  * The downside is that we have to touch page->_count against each page.
1589  * But we had to alter page->flags anyway.
1590  */
1591
1592 static void move_active_pages_to_lru(struct zone *zone,
1593                                      struct list_head *list,
1594                                      enum lru_list lru)
1595 {
1596         unsigned long pgmoved = 0;
1597         struct pagevec pvec;
1598         struct page *page;
1599
1600         pagevec_init(&pvec, 1);
1601
1602         while (!list_empty(list)) {
1603                 page = lru_to_page(list);
1604
1605                 VM_BUG_ON(PageLRU(page));
1606                 SetPageLRU(page);
1607
1608                 list_move(&page->lru, &zone->lru[lru].list);
1609                 mem_cgroup_add_lru_list(page, lru);
1610                 pgmoved += hpage_nr_pages(page);
1611
1612                 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1613                         spin_unlock_irq(&zone->lru_lock);
1614                         if (buffer_heads_over_limit)
1615                                 pagevec_strip(&pvec);
1616                         __pagevec_release(&pvec);
1617                         spin_lock_irq(&zone->lru_lock);
1618                 }
1619         }
1620         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1621         if (!is_active_lru(lru))
1622                 __count_vm_events(PGDEACTIVATE, pgmoved);
1623 }
1624
1625 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
1626                         struct scan_control *sc, int priority, int file)
1627 {
1628         unsigned long nr_taken;
1629         unsigned long pgscanned;
1630         unsigned long vm_flags;
1631         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1632         LIST_HEAD(l_active);
1633         LIST_HEAD(l_inactive);
1634         struct page *page;
1635         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1636         unsigned long nr_rotated = 0;
1637         isolate_mode_t reclaim_mode = ISOLATE_ACTIVE;
1638
1639         lru_add_drain();
1640
1641         if (!sc->may_unmap)
1642                 reclaim_mode |= ISOLATE_UNMAPPED;
1643         if (!sc->may_writepage)
1644                 reclaim_mode |= ISOLATE_CLEAN;
1645
1646         spin_lock_irq(&zone->lru_lock);
1647         if (scanning_global_lru(sc)) {
1648                 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1649                                                 &pgscanned, sc->order,
1650                                                 reclaim_mode, zone,
1651                                                 1, file);
1652                 zone->pages_scanned += pgscanned;
1653         } else {
1654                 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1655                                                 &pgscanned, sc->order,
1656                                                 reclaim_mode, zone,
1657                                                 sc->mem_cgroup, 1, file);
1658                 /*
1659                  * mem_cgroup_isolate_pages() keeps track of
1660                  * scanned pages on its own.
1661                  */
1662         }
1663
1664         reclaim_stat->recent_scanned[file] += nr_taken;
1665
1666         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1667         if (file)
1668                 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
1669         else
1670                 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
1671         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1672         spin_unlock_irq(&zone->lru_lock);
1673
1674         while (!list_empty(&l_hold)) {
1675                 cond_resched();
1676                 page = lru_to_page(&l_hold);
1677                 list_del(&page->lru);
1678
1679                 if (unlikely(!page_evictable(page, NULL))) {
1680                         putback_lru_page(page);
1681                         continue;
1682                 }
1683
1684                 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
1685                         nr_rotated += hpage_nr_pages(page);
1686                         /*
1687                          * Identify referenced, file-backed active pages and
1688                          * give them one more trip around the active list. So
1689                          * that executable code get better chances to stay in
1690                          * memory under moderate memory pressure.  Anon pages
1691                          * are not likely to be evicted by use-once streaming
1692                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1693                          * so we ignore them here.
1694                          */
1695                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1696                                 list_add(&page->lru, &l_active);
1697                                 continue;
1698                         }
1699                 }
1700
1701                 ClearPageActive(page);  /* we are de-activating */
1702                 list_add(&page->lru, &l_inactive);
1703         }
1704
1705         /*
1706          * Move pages back to the lru list.
1707          */
1708         spin_lock_irq(&zone->lru_lock);
1709         /*
1710          * Count referenced pages from currently used mappings as rotated,
1711          * even though only some of them are actually re-activated.  This
1712          * helps balance scan pressure between file and anonymous pages in
1713          * get_scan_ratio.
1714          */
1715         reclaim_stat->recent_rotated[file] += nr_rotated;
1716
1717         move_active_pages_to_lru(zone, &l_active,
1718                                                 LRU_ACTIVE + file * LRU_FILE);
1719         move_active_pages_to_lru(zone, &l_inactive,
1720                                                 LRU_BASE   + file * LRU_FILE);
1721         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1722         spin_unlock_irq(&zone->lru_lock);
1723 }
1724
1725 #ifdef CONFIG_SWAP
1726 static int inactive_anon_is_low_global(struct zone *zone)
1727 {
1728         unsigned long active, inactive;
1729
1730         active = zone_page_state(zone, NR_ACTIVE_ANON);
1731         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1732
1733         if (inactive * zone->inactive_ratio < active)
1734                 return 1;
1735
1736         return 0;
1737 }
1738
1739 /**
1740  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1741  * @zone: zone to check
1742  * @sc:   scan control of this context
1743  *
1744  * Returns true if the zone does not have enough inactive anon pages,
1745  * meaning some active anon pages need to be deactivated.
1746  */
1747 static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1748 {
1749         int low;
1750
1751         /*
1752          * If we don't have swap space, anonymous page deactivation
1753          * is pointless.
1754          */
1755         if (!total_swap_pages)
1756                 return 0;
1757
1758         if (scanning_global_lru(sc))
1759                 low = inactive_anon_is_low_global(zone);
1760         else
1761                 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
1762         return low;
1763 }
1764 #else
1765 static inline int inactive_anon_is_low(struct zone *zone,
1766                                         struct scan_control *sc)
1767 {
1768         return 0;
1769 }
1770 #endif
1771
1772 static int inactive_file_is_low_global(struct zone *zone)
1773 {
1774         unsigned long active, inactive;
1775
1776         active = zone_page_state(zone, NR_ACTIVE_FILE);
1777         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1778
1779         return (active > inactive);
1780 }
1781
1782 /**
1783  * inactive_file_is_low - check if file pages need to be deactivated
1784  * @zone: zone to check
1785  * @sc:   scan control of this context
1786  *
1787  * When the system is doing streaming IO, memory pressure here
1788  * ensures that active file pages get deactivated, until more
1789  * than half of the file pages are on the inactive list.
1790  *
1791  * Once we get to that situation, protect the system's working
1792  * set from being evicted by disabling active file page aging.
1793  *
1794  * This uses a different ratio than the anonymous pages, because
1795  * the page cache uses a use-once replacement algorithm.
1796  */
1797 static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1798 {
1799         int low;
1800
1801         if (scanning_global_lru(sc))
1802                 low = inactive_file_is_low_global(zone);
1803         else
1804                 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1805         return low;
1806 }
1807
1808 static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1809                                 int file)
1810 {
1811         if (file)
1812                 return inactive_file_is_low(zone, sc);
1813         else
1814                 return inactive_anon_is_low(zone, sc);
1815 }
1816
1817 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1818         struct zone *zone, struct scan_control *sc, int priority)
1819 {
1820         int file = is_file_lru(lru);
1821
1822         if (is_active_lru(lru)) {
1823                 if (inactive_list_is_low(zone, sc, file))
1824                     shrink_active_list(nr_to_scan, zone, sc, priority, file);
1825                 return 0;
1826         }
1827
1828         return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
1829 }
1830
1831 static int vmscan_swappiness(struct scan_control *sc)
1832 {
1833         if (scanning_global_lru(sc))
1834                 return vm_swappiness;
1835         return mem_cgroup_swappiness(sc->mem_cgroup);
1836 }
1837
1838 /*
1839  * Determine how aggressively the anon and file LRU lists should be
1840  * scanned.  The relative value of each set of LRU lists is determined
1841  * by looking at the fraction of the pages scanned we did rotate back
1842  * onto the active list instead of evict.
1843  *
1844  * nr[0] = anon pages to scan; nr[1] = file pages to scan
1845  */
1846 static void get_scan_count(struct zone *zone, struct scan_control *sc,
1847                                         unsigned long *nr, int priority)
1848 {
1849         unsigned long anon, file, free;
1850         unsigned long anon_prio, file_prio;
1851         unsigned long ap, fp;
1852         struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1853         u64 fraction[2], denominator;
1854         enum lru_list l;
1855         int noswap = 0;
1856         bool force_scan = false;
1857
1858         /*
1859          * If the zone or memcg is small, nr[l] can be 0.  This
1860          * results in no scanning on this priority and a potential
1861          * priority drop.  Global direct reclaim can go to the next
1862          * zone and tends to have no problems. Global kswapd is for
1863          * zone balancing and it needs to scan a minimum amount. When
1864          * reclaiming for a memcg, a priority drop can cause high
1865          * latencies, so it's better to scan a minimum amount there as
1866          * well.
1867          */
1868         if (scanning_global_lru(sc) && current_is_kswapd())
1869                 force_scan = true;
1870         if (!scanning_global_lru(sc))
1871                 force_scan = true;
1872
1873         /* If we have no swap space, do not bother scanning anon pages. */
1874         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1875                 noswap = 1;
1876                 fraction[0] = 0;
1877                 fraction[1] = 1;
1878                 denominator = 1;
1879                 goto out;
1880         }
1881
1882         anon  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1883                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1884         file  = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1885                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
1886
1887         if (scanning_global_lru(sc)) {
1888                 free  = zone_page_state(zone, NR_FREE_PAGES);
1889                 /* If we have very few page cache pages,
1890                    force-scan anon pages. */
1891                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1892                         fraction[0] = 1;
1893                         fraction[1] = 0;
1894                         denominator = 1;
1895                         goto out;
1896                 }
1897         }
1898
1899         /*
1900          * With swappiness at 100, anonymous and file have the same priority.
1901          * This scanning priority is essentially the inverse of IO cost.
1902          */
1903         anon_prio = vmscan_swappiness(sc);
1904         file_prio = 200 - vmscan_swappiness(sc);
1905
1906         /*
1907          * OK, so we have swap space and a fair amount of page cache
1908          * pages.  We use the recently rotated / recently scanned
1909          * ratios to determine how valuable each cache is.
1910          *
1911          * Because workloads change over time (and to avoid overflow)
1912          * we keep these statistics as a floating average, which ends
1913          * up weighing recent references more than old ones.
1914          *
1915          * anon in [0], file in [1]
1916          */
1917         spin_lock_irq(&zone->lru_lock);
1918         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1919                 reclaim_stat->recent_scanned[0] /= 2;
1920                 reclaim_stat->recent_rotated[0] /= 2;
1921         }
1922
1923         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1924                 reclaim_stat->recent_scanned[1] /= 2;
1925                 reclaim_stat->recent_rotated[1] /= 2;
1926         }
1927
1928         /*
1929          * The amount of pressure on anon vs file pages is inversely
1930          * proportional to the fraction of recently scanned pages on
1931          * each list that were recently referenced and in active use.
1932          */
1933         ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1934         ap /= reclaim_stat->recent_rotated[0] + 1;
1935
1936         fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1937         fp /= reclaim_stat->recent_rotated[1] + 1;
1938         spin_unlock_irq(&zone->lru_lock);
1939
1940         fraction[0] = ap;
1941         fraction[1] = fp;
1942         denominator = ap + fp + 1;
1943 out:
1944         for_each_evictable_lru(l) {
1945                 int file = is_file_lru(l);
1946                 unsigned long scan;
1947
1948                 scan = zone_nr_lru_pages(zone, sc, l);
1949                 if (priority || noswap) {
1950                         scan >>= priority;
1951                         if (!scan && force_scan)
1952                                 scan = SWAP_CLUSTER_MAX;
1953                         scan = div64_u64(scan * fraction[file], denominator);
1954                 }
1955                 nr[l] = scan;
1956         }
1957 }
1958
1959 /*
1960  * Reclaim/compaction depends on a number of pages being freed. To avoid
1961  * disruption to the system, a small number of order-0 pages continue to be
1962  * rotated and reclaimed in the normal fashion. However, by the time we get
1963  * back to the allocator and call try_to_compact_zone(), we ensure that
1964  * there are enough free pages for it to be likely successful
1965  */
1966 static inline bool should_continue_reclaim(struct zone *zone,
1967                                         unsigned long nr_reclaimed,
1968                                         unsigned long nr_scanned,
1969                                         struct scan_control *sc)
1970 {
1971         unsigned long pages_for_compaction;
1972         unsigned long inactive_lru_pages;
1973
1974         /* If not in reclaim/compaction mode, stop */
1975         if (!(sc->reclaim_mode & RECLAIM_MODE_COMPACTION))
1976                 return false;
1977
1978         /* Consider stopping depending on scan and reclaim activity */
1979         if (sc->gfp_mask & __GFP_REPEAT) {
1980                 /*
1981                  * For __GFP_REPEAT allocations, stop reclaiming if the
1982                  * full LRU list has been scanned and we are still failing
1983                  * to reclaim pages. This full LRU scan is potentially
1984                  * expensive but a __GFP_REPEAT caller really wants to succeed
1985                  */
1986                 if (!nr_reclaimed && !nr_scanned)
1987                         return false;
1988         } else {
1989                 /*
1990                  * For non-__GFP_REPEAT allocations which can presumably
1991                  * fail without consequence, stop if we failed to reclaim
1992                  * any pages from the last SWAP_CLUSTER_MAX number of
1993                  * pages that were scanned. This will return to the
1994                  * caller faster at the risk reclaim/compaction and
1995                  * the resulting allocation attempt fails
1996                  */
1997                 if (!nr_reclaimed)
1998                         return false;
1999         }
2000
2001         /*
2002          * If we have not reclaimed enough pages for compaction and the
2003          * inactive lists are large enough, continue reclaiming
2004          */
2005         pages_for_compaction = (2UL << sc->order);
2006         inactive_lru_pages = zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON) +
2007                                 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
2008         if (sc->nr_reclaimed < pages_for_compaction &&
2009                         inactive_lru_pages > pages_for_compaction)
2010                 return true;
2011
2012         /* If compaction would go ahead or the allocation would succeed, stop */
2013         switch (compaction_suitable(zone, sc->order)) {
2014         case COMPACT_PARTIAL:
2015         case COMPACT_CONTINUE:
2016                 return false;
2017         default:
2018                 return true;
2019         }
2020 }
2021
2022 /*
2023  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
2024  */
2025 static void shrink_zone(int priority, struct zone *zone,
2026                                 struct scan_control *sc)
2027 {
2028         unsigned long nr[NR_LRU_LISTS];
2029         unsigned long nr_to_scan;
2030         enum lru_list l;
2031         unsigned long nr_reclaimed, nr_scanned;
2032         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2033         struct blk_plug plug;
2034
2035 restart:
2036         nr_reclaimed = 0;
2037         nr_scanned = sc->nr_scanned;
2038         get_scan_count(zone, sc, nr, priority);
2039
2040         blk_start_plug(&plug);
2041         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2042                                         nr[LRU_INACTIVE_FILE]) {
2043                 for_each_evictable_lru(l) {
2044                         if (nr[l]) {
2045                                 nr_to_scan = min_t(unsigned long,
2046                                                    nr[l], SWAP_CLUSTER_MAX);
2047                                 nr[l] -= nr_to_scan;
2048
2049                                 nr_reclaimed += shrink_list(l, nr_to_scan,
2050                                                             zone, sc, priority);
2051                         }
2052                 }
2053                 /*
2054                  * On large memory systems, scan >> priority can become
2055                  * really large. This is fine for the starting priority;
2056                  * we want to put equal scanning pressure on each zone.
2057                  * However, if the VM has a harder time of freeing pages,
2058                  * with multiple processes reclaiming pages, the total
2059                  * freeing target can get unreasonably large.
2060                  */
2061                 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
2062                         break;
2063         }
2064         blk_finish_plug(&plug);
2065         sc->nr_reclaimed += nr_reclaimed;
2066
2067         /*
2068          * Even if we did not try to evict anon pages at all, we want to
2069          * rebalance the anon lru active/inactive ratio.
2070          */
2071         if (inactive_anon_is_low(zone, sc))
2072                 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
2073
2074         /* reclaim/compaction might need reclaim to continue */
2075         if (should_continue_reclaim(zone, nr_reclaimed,
2076                                         sc->nr_scanned - nr_scanned, sc))
2077                 goto restart;
2078
2079         throttle_vm_writeout(sc->gfp_mask);
2080 }
2081
2082 /*
2083  * This is the direct reclaim path, for page-allocating processes.  We only
2084  * try to reclaim pages from zones which will satisfy the caller's allocation
2085  * request.
2086  *
2087  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2088  * Because:
2089  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2090  *    allocation or
2091  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2092  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2093  *    zone defense algorithm.
2094  *
2095  * If a zone is deemed to be full of pinned pages then just give it a light
2096  * scan then give up on it.
2097  */
2098 static void shrink_zones(int priority, struct zonelist *zonelist,
2099                                         struct scan_control *sc)
2100 {
2101         struct zoneref *z;
2102         struct zone *zone;
2103         unsigned long nr_soft_reclaimed;
2104         unsigned long nr_soft_scanned;
2105
2106         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2107                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2108                 if (!populated_zone(zone))
2109                         continue;
2110                 /*
2111                  * Take care memory controller reclaiming has small influence
2112                  * to global LRU.
2113                  */
2114                 if (scanning_global_lru(sc)) {
2115                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2116                                 continue;
2117                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2118                                 continue;       /* Let kswapd poll it */
2119                         /*
2120                          * This steals pages from memory cgroups over softlimit
2121                          * and returns the number of reclaimed pages and
2122                          * scanned pages. This works for global memory pressure
2123                          * and balancing, not for a memcg's limit.
2124                          */
2125                         nr_soft_scanned = 0;
2126                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2127                                                 sc->order, sc->gfp_mask,
2128                                                 &nr_soft_scanned);
2129                         sc->nr_reclaimed += nr_soft_reclaimed;
2130                         sc->nr_scanned += nr_soft_scanned;
2131                         /* need some check for avoid more shrink_zone() */
2132                 }
2133
2134                 shrink_zone(priority, zone, sc);
2135         }
2136 }
2137
2138 static bool zone_reclaimable(struct zone *zone)
2139 {
2140         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2141 }
2142
2143 /* All zones in zonelist are unreclaimable? */
2144 static bool all_unreclaimable(struct zonelist *zonelist,
2145                 struct scan_control *sc)
2146 {
2147         struct zoneref *z;
2148         struct zone *zone;
2149
2150         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2151                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2152                 if (!populated_zone(zone))
2153                         continue;
2154                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2155                         continue;
2156                 if (!zone->all_unreclaimable)
2157                         return false;
2158         }
2159
2160         return true;
2161 }
2162
2163 /*
2164  * This is the main entry point to direct page reclaim.
2165  *
2166  * If a full scan of the inactive list fails to free enough memory then we
2167  * are "out of memory" and something needs to be killed.
2168  *
2169  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2170  * high - the zone may be full of dirty or under-writeback pages, which this
2171  * caller can't do much about.  We kick the writeback threads and take explicit
2172  * naps in the hope that some of these pages can be written.  But if the
2173  * allocating task holds filesystem locks which prevent writeout this might not
2174  * work, and the allocation attempt will fail.
2175  *
2176  * returns:     0, if no pages reclaimed
2177  *              else, the number of pages reclaimed
2178  */
2179 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2180                                         struct scan_control *sc,
2181                                         struct shrink_control *shrink)
2182 {
2183         int priority;
2184         unsigned long total_scanned = 0;
2185         struct reclaim_state *reclaim_state = current->reclaim_state;
2186         struct zoneref *z;
2187         struct zone *zone;
2188         unsigned long writeback_threshold;
2189
2190         get_mems_allowed();
2191         delayacct_freepages_start();
2192
2193         if (scanning_global_lru(sc))
2194                 count_vm_event(ALLOCSTALL);
2195
2196         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2197                 sc->nr_scanned = 0;
2198                 if (!priority)
2199                         disable_swap_token(sc->mem_cgroup);
2200                 shrink_zones(priority, zonelist, sc);
2201                 /*
2202                  * Don't shrink slabs when reclaiming memory from
2203                  * over limit cgroups
2204                  */
2205                 if (scanning_global_lru(sc)) {
2206                         unsigned long lru_pages = 0;
2207                         for_each_zone_zonelist(zone, z, zonelist,
2208                                         gfp_zone(sc->gfp_mask)) {
2209                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2210                                         continue;
2211
2212                                 lru_pages += zone_reclaimable_pages(zone);
2213                         }
2214
2215                         shrink_slab(shrink, sc->nr_scanned, lru_pages);
2216                         if (reclaim_state) {
2217                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2218                                 reclaim_state->reclaimed_slab = 0;
2219                         }
2220                 }
2221                 total_scanned += sc->nr_scanned;
2222                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2223                         goto out;
2224
2225                 /*
2226                  * Try to write back as many pages as we just scanned.  This
2227                  * tends to cause slow streaming writers to write data to the
2228                  * disk smoothly, at the dirtying rate, which is nice.   But
2229                  * that's undesirable in laptop mode, where we *want* lumpy
2230                  * writeout.  So in laptop mode, write out the whole world.
2231                  */
2232                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2233                 if (total_scanned > writeback_threshold) {
2234                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
2235                         sc->may_writepage = 1;
2236                 }
2237
2238                 /* Take a nap, wait for some writeback to complete */
2239                 if (!sc->hibernation_mode && sc->nr_scanned &&
2240                     priority < DEF_PRIORITY - 2) {
2241                         struct zone *preferred_zone;
2242
2243                         first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2244                                                 &cpuset_current_mems_allowed,
2245                                                 &preferred_zone);
2246                         wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2247                 }
2248         }
2249
2250 out:
2251         delayacct_freepages_end();
2252         put_mems_allowed();
2253
2254         if (sc->nr_reclaimed)
2255                 return sc->nr_reclaimed;
2256
2257         /*
2258          * As hibernation is going on, kswapd is freezed so that it can't mark
2259          * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2260          * check.
2261          */
2262         if (oom_killer_disabled)
2263                 return 0;
2264
2265         /* top priority shrink_zones still had more to do? don't OOM, then */
2266         if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
2267                 return 1;
2268
2269         return 0;
2270 }
2271
2272 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2273                                 gfp_t gfp_mask, nodemask_t *nodemask)
2274 {
2275         unsigned long nr_reclaimed;
2276         struct scan_control sc = {
2277                 .gfp_mask = gfp_mask,
2278                 .may_writepage = !laptop_mode,
2279                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2280                 .may_unmap = 1,
2281                 .may_swap = 1,
2282                 .order = order,
2283                 .mem_cgroup = NULL,
2284                 .nodemask = nodemask,
2285         };
2286         struct shrink_control shrink = {
2287                 .gfp_mask = sc.gfp_mask,
2288         };
2289
2290         trace_mm_vmscan_direct_reclaim_begin(order,
2291                                 sc.may_writepage,
2292                                 gfp_mask);
2293
2294         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2295
2296         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2297
2298         return nr_reclaimed;
2299 }
2300
2301 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2302
2303 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2304                                                 gfp_t gfp_mask, bool noswap,
2305                                                 struct zone *zone,
2306                                                 unsigned long *nr_scanned)
2307 {
2308         struct scan_control sc = {
2309                 .nr_scanned = 0,
2310                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2311                 .may_writepage = !laptop_mode,
2312                 .may_unmap = 1,
2313                 .may_swap = !noswap,
2314                 .order = 0,
2315                 .mem_cgroup = mem,
2316         };
2317
2318         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2319                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2320
2321         trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2322                                                       sc.may_writepage,
2323                                                       sc.gfp_mask);
2324
2325         /*
2326          * NOTE: Although we can get the priority field, using it
2327          * here is not a good idea, since it limits the pages we can scan.
2328          * if we don't reclaim here, the shrink_zone from balance_pgdat
2329          * will pick up pages from other mem cgroup's as well. We hack
2330          * the priority and make it zero.
2331          */
2332         shrink_zone(0, zone, &sc);
2333
2334         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2335
2336         *nr_scanned = sc.nr_scanned;
2337         return sc.nr_reclaimed;
2338 }
2339
2340 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
2341                                            gfp_t gfp_mask,
2342                                            bool noswap)
2343 {
2344         struct zonelist *zonelist;
2345         unsigned long nr_reclaimed;
2346         int nid;
2347         struct scan_control sc = {
2348                 .may_writepage = !laptop_mode,
2349                 .may_unmap = 1,
2350                 .may_swap = !noswap,
2351                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2352                 .order = 0,
2353                 .mem_cgroup = mem_cont,
2354                 .nodemask = NULL, /* we don't care the placement */
2355                 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2356                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2357         };
2358         struct shrink_control shrink = {
2359                 .gfp_mask = sc.gfp_mask,
2360         };
2361
2362         /*
2363          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2364          * take care of from where we get pages. So the node where we start the
2365          * scan does not need to be the current node.
2366          */
2367         nid = mem_cgroup_select_victim_node(mem_cont);
2368
2369         zonelist = NODE_DATA(nid)->node_zonelists;
2370
2371         trace_mm_vmscan_memcg_reclaim_begin(0,
2372                                             sc.may_writepage,
2373                                             sc.gfp_mask);
2374
2375         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2376
2377         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2378
2379         return nr_reclaimed;
2380 }
2381 #endif
2382
2383 /*
2384  * pgdat_balanced is used when checking if a node is balanced for high-order
2385  * allocations. Only zones that meet watermarks and are in a zone allowed
2386  * by the callers classzone_idx are added to balanced_pages. The total of
2387  * balanced pages must be at least 25% of the zones allowed by classzone_idx
2388  * for the node to be considered balanced. Forcing all zones to be balanced
2389  * for high orders can cause excessive reclaim when there are imbalanced zones.
2390  * The choice of 25% is due to
2391  *   o a 16M DMA zone that is balanced will not balance a zone on any
2392  *     reasonable sized machine
2393  *   o On all other machines, the top zone must be at least a reasonable
2394  *     percentage of the middle zones. For example, on 32-bit x86, highmem
2395  *     would need to be at least 256M for it to be balance a whole node.
2396  *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2397  *     to balance a node on its own. These seemed like reasonable ratios.
2398  */
2399 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2400                                                 int classzone_idx)
2401 {
2402         unsigned long present_pages = 0;
2403         int i;
2404
2405         for (i = 0; i <= classzone_idx; i++)
2406                 present_pages += pgdat->node_zones[i].present_pages;
2407
2408         /* A special case here: if zone has no page, we think it's balanced */
2409         return balanced_pages >= (present_pages >> 2);
2410 }
2411
2412 /* is kswapd sleeping prematurely? */
2413 static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
2414                                         int classzone_idx)
2415 {
2416         int i;
2417         unsigned long balanced = 0;
2418         bool all_zones_ok = true;
2419
2420         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2421         if (remaining)
2422                 return true;
2423
2424         /* Check the watermark levels */
2425         for (i = 0; i <= classzone_idx; i++) {
2426                 struct zone *zone = pgdat->node_zones + i;
2427
2428                 if (!populated_zone(zone))
2429                         continue;
2430
2431                 /*
2432                  * balance_pgdat() skips over all_unreclaimable after
2433                  * DEF_PRIORITY. Effectively, it considers them balanced so
2434                  * they must be considered balanced here as well if kswapd
2435                  * is to sleep
2436                  */
2437                 if (zone->all_unreclaimable) {
2438                         balanced += zone->present_pages;
2439                         continue;
2440                 }
2441
2442                 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone),
2443                                                         i, 0))
2444                         all_zones_ok = false;
2445                 else
2446                         balanced += zone->present_pages;
2447         }
2448
2449         /*
2450          * For high-order requests, the balanced zones must contain at least
2451          * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2452          * must be balanced
2453          */
2454         if (order)
2455                 return !pgdat_balanced(pgdat, balanced, classzone_idx);
2456         else
2457                 return !all_zones_ok;
2458 }
2459
2460 /*
2461  * For kswapd, balance_pgdat() will work across all this node's zones until
2462  * they are all at high_wmark_pages(zone).
2463  *
2464  * Returns the final order kswapd was reclaiming at
2465  *
2466  * There is special handling here for zones which are full of pinned pages.
2467  * This can happen if the pages are all mlocked, or if they are all used by
2468  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2469  * What we do is to detect the case where all pages in the zone have been
2470  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2471  * dead and from now on, only perform a short scan.  Basically we're polling
2472  * the zone for when the problem goes away.
2473  *
2474  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2475  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2476  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2477  * lower zones regardless of the number of free pages in the lower zones. This
2478  * interoperates with the page allocator fallback scheme to ensure that aging
2479  * of pages is balanced across the zones.
2480  */
2481 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2482                                                         int *classzone_idx)
2483 {
2484         int all_zones_ok;
2485         unsigned long balanced;
2486         int priority;
2487         int i;
2488         int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2489         unsigned long total_scanned;
2490         struct reclaim_state *reclaim_state = current->reclaim_state;
2491         unsigned long nr_soft_reclaimed;
2492         unsigned long nr_soft_scanned;
2493         struct scan_control sc = {
2494                 .gfp_mask = GFP_KERNEL,
2495                 .may_unmap = 1,
2496                 .may_swap = 1,
2497                 /*
2498                  * kswapd doesn't want to be bailed out while reclaim. because
2499                  * we want to put equal scanning pressure on each zone.
2500                  */
2501                 .nr_to_reclaim = ULONG_MAX,
2502                 .order = order,
2503                 .mem_cgroup = NULL,
2504         };
2505         struct shrink_control shrink = {
2506                 .gfp_mask = sc.gfp_mask,
2507         };
2508 loop_again:
2509         total_scanned = 0;
2510         sc.nr_reclaimed = 0;
2511         sc.may_writepage = !laptop_mode;
2512         count_vm_event(PAGEOUTRUN);
2513
2514         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2515                 unsigned long lru_pages = 0;
2516                 int has_under_min_watermark_zone = 0;
2517
2518                 /* The swap token gets in the way of swapout... */
2519                 if (!priority)
2520                         disable_swap_token(NULL);
2521
2522                 all_zones_ok = 1;
2523                 balanced = 0;
2524
2525                 /*
2526                  * Scan in the highmem->dma direction for the highest
2527                  * zone which needs scanning
2528                  */
2529                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2530                         struct zone *zone = pgdat->node_zones + i;
2531
2532                         if (!populated_zone(zone))
2533                                 continue;
2534
2535                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2536                                 continue;
2537
2538                         /*
2539                          * Do some background aging of the anon list, to give
2540                          * pages a chance to be referenced before reclaiming.
2541                          */
2542                         if (inactive_anon_is_low(zone, &sc))
2543                                 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2544                                                         &sc, priority, 0);
2545
2546                         if (!zone_watermark_ok_safe(zone, order,
2547                                         high_wmark_pages(zone), 0, 0)) {
2548                                 end_zone = i;
2549                                 break;
2550                         } else {
2551                                 /* If balanced, clear the congested flag */
2552                                 zone_clear_flag(zone, ZONE_CONGESTED);
2553                         }
2554                 }
2555                 if (i < 0)
2556                         goto out;
2557
2558                 for (i = 0; i <= end_zone; i++) {
2559                         struct zone *zone = pgdat->node_zones + i;
2560
2561                         lru_pages += zone_reclaimable_pages(zone);
2562                 }
2563
2564                 /*
2565                  * Now scan the zone in the dma->highmem direction, stopping
2566                  * at the last zone which needs scanning.
2567                  *
2568                  * We do this because the page allocator works in the opposite
2569                  * direction.  This prevents the page allocator from allocating
2570                  * pages behind kswapd's direction of progress, which would
2571                  * cause too much scanning of the lower zones.
2572                  */
2573                 for (i = 0; i <= end_zone; i++) {
2574                         struct zone *zone = pgdat->node_zones + i;
2575                         int nr_slab;
2576                         unsigned long balance_gap;
2577
2578                         if (!populated_zone(zone))
2579                                 continue;
2580
2581                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2582                                 continue;
2583
2584                         sc.nr_scanned = 0;
2585
2586                         nr_soft_scanned = 0;
2587                         /*
2588                          * Call soft limit reclaim before calling shrink_zone.
2589                          */
2590                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2591                                                         order, sc.gfp_mask,
2592                                                         &nr_soft_scanned);
2593                         sc.nr_reclaimed += nr_soft_reclaimed;
2594                         total_scanned += nr_soft_scanned;
2595
2596                         /*
2597                          * We put equal pressure on every zone, unless
2598                          * one zone has way too many pages free
2599                          * already. The "too many pages" is defined
2600                          * as the high wmark plus a "gap" where the
2601                          * gap is either the low watermark or 1%
2602                          * of the zone, whichever is smaller.
2603                          */
2604                         balance_gap = min(low_wmark_pages(zone),
2605                                 (zone->present_pages +
2606                                         KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2607                                 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2608                         if (!zone_watermark_ok_safe(zone, order,
2609                                         high_wmark_pages(zone) + balance_gap,
2610                                         end_zone, 0)) {
2611                                 shrink_zone(priority, zone, &sc);
2612
2613                                 reclaim_state->reclaimed_slab = 0;
2614                                 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2615                                 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2616                                 total_scanned += sc.nr_scanned;
2617
2618                                 if (nr_slab == 0 && !zone_reclaimable(zone))
2619                                         zone->all_unreclaimable = 1;
2620                         }
2621
2622                         /*
2623                          * If we've done a decent amount of scanning and
2624                          * the reclaim ratio is low, start doing writepage
2625                          * even in laptop mode
2626                          */
2627                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2628                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2629                                 sc.may_writepage = 1;
2630
2631                         if (zone->all_unreclaimable) {
2632                                 if (end_zone && end_zone == i)
2633                                         end_zone--;
2634                                 continue;
2635                         }
2636
2637                         if (!zone_watermark_ok_safe(zone, order,
2638                                         high_wmark_pages(zone), end_zone, 0)) {
2639                                 all_zones_ok = 0;
2640                                 /*
2641                                  * We are still under min water mark.  This
2642                                  * means that we have a GFP_ATOMIC allocation
2643                                  * failure risk. Hurry up!
2644                                  */
2645                                 if (!zone_watermark_ok_safe(zone, order,
2646                                             min_wmark_pages(zone), end_zone, 0))
2647                                         has_under_min_watermark_zone = 1;
2648                         } else {
2649                                 /*
2650                                  * If a zone reaches its high watermark,
2651                                  * consider it to be no longer congested. It's
2652                                  * possible there are dirty pages backed by
2653                                  * congested BDIs but as pressure is relieved,
2654                                  * spectulatively avoid congestion waits
2655                                  */
2656                                 zone_clear_flag(zone, ZONE_CONGESTED);
2657                                 if (i <= *classzone_idx)
2658                                         balanced += zone->present_pages;
2659                         }
2660
2661                 }
2662                 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2663                         break;          /* kswapd: all done */
2664                 /*
2665                  * OK, kswapd is getting into trouble.  Take a nap, then take
2666                  * another pass across the zones.
2667                  */
2668                 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2669                         if (has_under_min_watermark_zone)
2670                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2671                         else
2672                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2673                 }
2674
2675                 /*
2676                  * We do this so kswapd doesn't build up large priorities for
2677                  * example when it is freeing in parallel with allocators. It
2678                  * matches the direct reclaim path behaviour in terms of impact
2679                  * on zone->*_priority.
2680                  */
2681                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2682                         break;
2683         }
2684 out:
2685
2686         /*
2687          * order-0: All zones must meet high watermark for a balanced node
2688          * high-order: Balanced zones must make up at least 25% of the node
2689          *             for the node to be balanced
2690          */
2691         if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2692                 cond_resched();
2693
2694                 try_to_freeze();
2695
2696                 /*
2697                  * Fragmentation may mean that the system cannot be
2698                  * rebalanced for high-order allocations in all zones.
2699                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2700                  * it means the zones have been fully scanned and are still
2701                  * not balanced. For high-order allocations, there is
2702                  * little point trying all over again as kswapd may
2703                  * infinite loop.
2704                  *
2705                  * Instead, recheck all watermarks at order-0 as they
2706                  * are the most important. If watermarks are ok, kswapd will go
2707                  * back to sleep. High-order users can still perform direct
2708                  * reclaim if they wish.
2709                  */
2710                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2711                         order = sc.order = 0;
2712
2713                 goto loop_again;
2714         }
2715
2716         /*
2717          * If kswapd was reclaiming at a higher order, it has the option of
2718          * sleeping without all zones being balanced. Before it does, it must
2719          * ensure that the watermarks for order-0 on *all* zones are met and
2720          * that the congestion flags are cleared. The congestion flag must
2721          * be cleared as kswapd is the only mechanism that clears the flag
2722          * and it is potentially going to sleep here.
2723          */
2724         if (order) {
2725                 for (i = 0; i <= end_zone; i++) {
2726                         struct zone *zone = pgdat->node_zones + i;
2727
2728                         if (!populated_zone(zone))
2729                                 continue;
2730
2731                         if (zone->all_unreclaimable && priority != DEF_PRIORITY)
2732                                 continue;
2733
2734                         /* Confirm the zone is balanced for order-0 */
2735                         if (!zone_watermark_ok(zone, 0,
2736                                         high_wmark_pages(zone), 0, 0)) {
2737                                 order = sc.order = 0;
2738                                 goto loop_again;
2739                         }
2740
2741                         /* If balanced, clear the congested flag */
2742                         zone_clear_flag(zone, ZONE_CONGESTED);
2743                 }
2744         }
2745
2746         /*
2747          * Return the order we were reclaiming at so sleeping_prematurely()
2748          * makes a decision on the order we were last reclaiming at. However,
2749          * if another caller entered the allocator slow path while kswapd
2750          * was awake, order will remain at the higher level
2751          */
2752         *classzone_idx = end_zone;
2753         return order;
2754 }
2755
2756 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2757 {
2758         long remaining = 0;
2759         DEFINE_WAIT(wait);
2760
2761         if (freezing(current) || kthread_should_stop())
2762                 return;
2763
2764         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2765
2766         /* Try to sleep for a short interval */
2767         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2768                 remaining = schedule_timeout(HZ/10);
2769                 finish_wait(&pgdat->kswapd_wait, &wait);
2770                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2771         }
2772
2773         /*
2774          * After a short sleep, check if it was a premature sleep. If not, then
2775          * go fully to sleep until explicitly woken up.
2776          */
2777         if (!sleeping_prematurely(pgdat, order, remaining, classzone_idx)) {
2778                 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2779
2780                 /*
2781                  * vmstat counters are not perfectly accurate and the estimated
2782                  * value for counters such as NR_FREE_PAGES can deviate from the
2783                  * true value by nr_online_cpus * threshold. To avoid the zone
2784                  * watermarks being breached while under pressure, we reduce the
2785                  * per-cpu vmstat threshold while kswapd is awake and restore
2786                  * them before going back to sleep.
2787                  */
2788                 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2789                 schedule();
2790                 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2791         } else {
2792                 if (remaining)
2793                         count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2794                 else
2795                         count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2796         }
2797         finish_wait(&pgdat->kswapd_wait, &wait);
2798 }
2799
2800 /*
2801  * The background pageout daemon, started as a kernel thread
2802  * from the init process.
2803  *
2804  * This basically trickles out pages so that we have _some_
2805  * free memory available even if there is no other activity
2806  * that frees anything up. This is needed for things like routing
2807  * etc, where we otherwise might have all activity going on in
2808  * asynchronous contexts that cannot page things out.
2809  *
2810  * If there are applications that are active memory-allocators
2811  * (most normal use), this basically shouldn't matter.
2812  */
2813 static int kswapd(void *p)
2814 {
2815         unsigned long order, new_order;
2816         int classzone_idx, new_classzone_idx;
2817         pg_data_t *pgdat = (pg_data_t*)p;
2818         struct task_struct *tsk = current;
2819
2820         struct reclaim_state reclaim_state = {
2821                 .reclaimed_slab = 0,
2822         };
2823         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2824
2825         lockdep_set_current_reclaim_state(GFP_KERNEL);
2826
2827         if (!cpumask_empty(cpumask))
2828                 set_cpus_allowed_ptr(tsk, cpumask);
2829         current->reclaim_state = &reclaim_state;
2830
2831         /*
2832          * Tell the memory management that we're a "memory allocator",
2833          * and that if we need more memory we should get access to it
2834          * regardless (see "__alloc_pages()"). "kswapd" should
2835          * never get caught in the normal page freeing logic.
2836          *
2837          * (Kswapd normally doesn't need memory anyway, but sometimes
2838          * you need a small amount of memory in order to be able to
2839          * page out something else, and this flag essentially protects
2840          * us from recursively trying to free more memory as we're
2841          * trying to free the first piece of memory in the first place).
2842          */
2843         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2844         set_freezable();
2845
2846         order = new_order = 0;
2847         classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2848         for ( ; ; ) {
2849                 int ret;
2850
2851                 /*
2852                  * If the last balance_pgdat was unsuccessful it's unlikely a
2853                  * new request of a similar or harder type will succeed soon
2854                  * so consider going to sleep on the basis we reclaimed at
2855                  */
2856                 if (classzone_idx >= new_classzone_idx && order == new_order) {
2857                         new_order = pgdat->kswapd_max_order;
2858                         new_classzone_idx = pgdat->classzone_idx;
2859                         pgdat->kswapd_max_order =  0;
2860                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2861                 }
2862
2863                 if (order < new_order || classzone_idx > new_classzone_idx) {
2864                         /*
2865                          * Don't sleep if someone wants a larger 'order'
2866                          * allocation or has tigher zone constraints
2867                          */
2868                         order = new_order;
2869                         classzone_idx = new_classzone_idx;
2870                 } else {
2871                         kswapd_try_to_sleep(pgdat, order, classzone_idx);
2872                         order = pgdat->kswapd_max_order;
2873                         classzone_idx = pgdat->classzone_idx;
2874                         pgdat->kswapd_max_order = 0;
2875                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2876                 }
2877
2878                 ret = try_to_freeze();
2879                 if (kthread_should_stop())
2880                         break;
2881
2882                 /*
2883                  * We can speed up thawing tasks if we don't call balance_pgdat
2884                  * after returning from the refrigerator
2885                  */
2886                 if (!ret) {
2887                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
2888                         order = balance_pgdat(pgdat, order, &classzone_idx);
2889                 }
2890         }
2891         return 0;
2892 }
2893
2894 /*
2895  * A zone is low on free memory, so wake its kswapd task to service it.
2896  */
2897 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
2898 {
2899         pg_data_t *pgdat;
2900
2901         if (!populated_zone(zone))
2902                 return;
2903
2904         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2905                 return;
2906         pgdat = zone->zone_pgdat;
2907         if (pgdat->kswapd_max_order < order) {
2908                 pgdat->kswapd_max_order = order;
2909                 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
2910         }
2911         if (!waitqueue_active(&pgdat->kswapd_wait))
2912                 return;
2913         if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
2914                 return;
2915
2916         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
2917         wake_up_interruptible(&pgdat->kswapd_wait);
2918 }
2919
2920 /*
2921  * The reclaimable count would be mostly accurate.
2922  * The less reclaimable pages may be
2923  * - mlocked pages, which will be moved to unevictable list when encountered
2924  * - mapped pages, which may require several travels to be reclaimed
2925  * - dirty pages, which is not "instantly" reclaimable
2926  */
2927 unsigned long global_reclaimable_pages(void)
2928 {
2929         int nr;
2930
2931         nr = global_page_state(NR_ACTIVE_FILE) +
2932              global_page_state(NR_INACTIVE_FILE);
2933
2934         if (nr_swap_pages > 0)
2935                 nr += global_page_state(NR_ACTIVE_ANON) +
2936                       global_page_state(NR_INACTIVE_ANON);
2937
2938         return nr;
2939 }
2940
2941 unsigned long zone_reclaimable_pages(struct zone *zone)
2942 {
2943         int nr;
2944
2945         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2946              zone_page_state(zone, NR_INACTIVE_FILE);
2947
2948         if (nr_swap_pages > 0)
2949                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2950                       zone_page_state(zone, NR_INACTIVE_ANON);
2951
2952         return nr;
2953 }
2954
2955 #ifdef CONFIG_HIBERNATION
2956 /*
2957  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2958  * freed pages.
2959  *
2960  * Rather than trying to age LRUs the aim is to preserve the overall
2961  * LRU order by reclaiming preferentially
2962  * inactive > active > active referenced > active mapped
2963  */
2964 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
2965 {
2966         struct reclaim_state reclaim_state;
2967         struct scan_control sc = {
2968                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2969                 .may_swap = 1,
2970                 .may_unmap = 1,
2971                 .may_writepage = 1,
2972                 .nr_to_reclaim = nr_to_reclaim,
2973                 .hibernation_mode = 1,
2974                 .order = 0,
2975         };
2976         struct shrink_control shrink = {
2977                 .gfp_mask = sc.gfp_mask,
2978         };
2979         struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2980         struct task_struct *p = current;
2981         unsigned long nr_reclaimed;
2982
2983         p->flags |= PF_MEMALLOC;
2984         lockdep_set_current_reclaim_state(sc.gfp_mask);
2985         reclaim_state.reclaimed_slab = 0;
2986         p->reclaim_state = &reclaim_state;
2987
2988         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2989
2990         p->reclaim_state = NULL;
2991         lockdep_clear_current_reclaim_state();
2992         p->flags &= ~PF_MEMALLOC;
2993
2994         return nr_reclaimed;
2995 }
2996 #endif /* CONFIG_HIBERNATION */
2997
2998 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2999    not required for correctness.  So if the last cpu in a node goes
3000    away, we get changed to run anywhere: as the first one comes back,
3001    restore their cpu bindings. */
3002 static int __devinit cpu_callback(struct notifier_block *nfb,
3003                                   unsigned long action, void *hcpu)
3004 {
3005         int nid;
3006
3007         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3008                 for_each_node_state(nid, N_HIGH_MEMORY) {
3009                         pg_data_t *pgdat = NODE_DATA(nid);
3010                         const struct cpumask *mask;
3011
3012                         mask = cpumask_of_node(pgdat->node_id);
3013
3014                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3015                                 /* One of our CPUs online: restore mask */
3016                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3017                 }
3018         }
3019         return NOTIFY_OK;
3020 }
3021
3022 /*
3023  * This kswapd start function will be called by init and node-hot-add.
3024  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3025  */
3026 int kswapd_run(int nid)
3027 {
3028         pg_data_t *pgdat = NODE_DATA(nid);
3029         int ret = 0;
3030
3031         if (pgdat->kswapd)
3032                 return 0;
3033
3034         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3035         if (IS_ERR(pgdat->kswapd)) {
3036                 /* failure at boot is fatal */
3037                 BUG_ON(system_state == SYSTEM_BOOTING);
3038                 printk("Failed to start kswapd on node %d\n",nid);
3039                 ret = -1;
3040         }
3041         return ret;
3042 }
3043
3044 /*
3045  * Called by memory hotplug when all memory in a node is offlined.
3046  */
3047 void kswapd_stop(int nid)
3048 {
3049         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3050
3051         if (kswapd)
3052                 kthread_stop(kswapd);
3053 }
3054
3055 static int __init kswapd_init(void)
3056 {
3057         int nid;
3058
3059         swap_setup();
3060         for_each_node_state(nid, N_HIGH_MEMORY)
3061                 kswapd_run(nid);
3062         hotcpu_notifier(cpu_callback, 0);
3063         return 0;
3064 }
3065
3066 module_init(kswapd_init)
3067
3068 #ifdef CONFIG_NUMA
3069 /*
3070  * Zone reclaim mode
3071  *
3072  * If non-zero call zone_reclaim when the number of free pages falls below
3073  * the watermarks.
3074  */
3075 int zone_reclaim_mode __read_mostly;
3076
3077 #define RECLAIM_OFF 0
3078 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
3079 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
3080 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
3081
3082 /*
3083  * Priority for ZONE_RECLAIM. This determines the fraction of pages
3084  * of a node considered for each zone_reclaim. 4 scans 1/16th of
3085  * a zone.
3086  */
3087 #define ZONE_RECLAIM_PRIORITY 4
3088
3089 /*
3090  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3091  * occur.
3092  */
3093 int sysctl_min_unmapped_ratio = 1;
3094
3095 /*
3096  * If the number of slab pages in a zone grows beyond this percentage then
3097  * slab reclaim needs to occur.
3098  */
3099 int sysctl_min_slab_ratio = 5;
3100
3101 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3102 {
3103         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3104         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3105                 zone_page_state(zone, NR_ACTIVE_FILE);
3106
3107         /*
3108          * It's possible for there to be more file mapped pages than
3109          * accounted for by the pages on the file LRU lists because
3110          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3111          */
3112         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3113 }
3114
3115 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3116 static long zone_pagecache_reclaimable(struct zone *zone)
3117 {
3118         long nr_pagecache_reclaimable;
3119         long delta = 0;
3120
3121         /*
3122          * If RECLAIM_SWAP is set, then all file pages are considered
3123          * potentially reclaimable. Otherwise, we have to worry about
3124          * pages like swapcache and zone_unmapped_file_pages() provides
3125          * a better estimate
3126          */
3127         if (zone_reclaim_mode & RECLAIM_SWAP)
3128                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3129         else