Merge branch 'x86-bsp-hotplug-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[linux-3.10.git] / mm / vmscan.c
1 /*
2  *  linux/mm/vmscan.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *
6  *  Swap reorganised 29.12.95, Stephen Tweedie.
7  *  kswapd added: 7.1.96  sct
8  *  Removed kswapd_ctl limits, and swap out as many pages as needed
9  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11  *  Multiqueue VM started 5.8.00, Rik van Riel.
12  */
13
14 #include <linux/mm.h>
15 #include <linux/module.h>
16 #include <linux/gfp.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h>  /* for try_to_release_page(),
27                                         buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/backing-dev.h>
30 #include <linux/rmap.h>
31 #include <linux/topology.h>
32 #include <linux/cpu.h>
33 #include <linux/cpuset.h>
34 #include <linux/compaction.h>
35 #include <linux/notifier.h>
36 #include <linux/rwsem.h>
37 #include <linux/delay.h>
38 #include <linux/kthread.h>
39 #include <linux/freezer.h>
40 #include <linux/memcontrol.h>
41 #include <linux/delayacct.h>
42 #include <linux/sysctl.h>
43 #include <linux/oom.h>
44 #include <linux/prefetch.h>
45
46 #include <asm/tlbflush.h>
47 #include <asm/div64.h>
48
49 #include <linux/swapops.h>
50
51 #include "internal.h"
52
53 #define CREATE_TRACE_POINTS
54 #include <trace/events/vmscan.h>
55
56 struct scan_control {
57         /* Incremented by the number of inactive pages that were scanned */
58         unsigned long nr_scanned;
59
60         /* Number of pages freed so far during a call to shrink_zones() */
61         unsigned long nr_reclaimed;
62
63         /* How many pages shrink_list() should reclaim */
64         unsigned long nr_to_reclaim;
65
66         unsigned long hibernation_mode;
67
68         /* This context's GFP mask */
69         gfp_t gfp_mask;
70
71         int may_writepage;
72
73         /* Can mapped pages be reclaimed? */
74         int may_unmap;
75
76         /* Can pages be swapped as part of reclaim? */
77         int may_swap;
78
79         int order;
80
81         /* Scan (total_size >> priority) pages at once */
82         int priority;
83
84         /*
85          * The memory cgroup that hit its limit and as a result is the
86          * primary target of this reclaim invocation.
87          */
88         struct mem_cgroup *target_mem_cgroup;
89
90         /*
91          * Nodemask of nodes allowed by the caller. If NULL, all nodes
92          * are scanned.
93          */
94         nodemask_t      *nodemask;
95 };
96
97 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
98
99 #ifdef ARCH_HAS_PREFETCH
100 #define prefetch_prev_lru_page(_page, _base, _field)                    \
101         do {                                                            \
102                 if ((_page)->lru.prev != _base) {                       \
103                         struct page *prev;                              \
104                                                                         \
105                         prev = lru_to_page(&(_page->lru));              \
106                         prefetch(&prev->_field);                        \
107                 }                                                       \
108         } while (0)
109 #else
110 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
111 #endif
112
113 #ifdef ARCH_HAS_PREFETCHW
114 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
115         do {                                                            \
116                 if ((_page)->lru.prev != _base) {                       \
117                         struct page *prev;                              \
118                                                                         \
119                         prev = lru_to_page(&(_page->lru));              \
120                         prefetchw(&prev->_field);                       \
121                 }                                                       \
122         } while (0)
123 #else
124 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
125 #endif
126
127 /*
128  * From 0 .. 100.  Higher means more swappy.
129  */
130 int vm_swappiness = 60;
131 long vm_total_pages;    /* The total number of pages which the VM controls */
132
133 static LIST_HEAD(shrinker_list);
134 static DECLARE_RWSEM(shrinker_rwsem);
135
136 #ifdef CONFIG_MEMCG
137 static bool global_reclaim(struct scan_control *sc)
138 {
139         return !sc->target_mem_cgroup;
140 }
141 #else
142 static bool global_reclaim(struct scan_control *sc)
143 {
144         return true;
145 }
146 #endif
147
148 static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
149 {
150         if (!mem_cgroup_disabled())
151                 return mem_cgroup_get_lru_size(lruvec, lru);
152
153         return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
154 }
155
156 /*
157  * Add a shrinker callback to be called from the vm
158  */
159 void register_shrinker(struct shrinker *shrinker)
160 {
161         atomic_long_set(&shrinker->nr_in_batch, 0);
162         down_write(&shrinker_rwsem);
163         list_add_tail(&shrinker->list, &shrinker_list);
164         up_write(&shrinker_rwsem);
165 }
166 EXPORT_SYMBOL(register_shrinker);
167
168 /*
169  * Remove one
170  */
171 void unregister_shrinker(struct shrinker *shrinker)
172 {
173         down_write(&shrinker_rwsem);
174         list_del(&shrinker->list);
175         up_write(&shrinker_rwsem);
176 }
177 EXPORT_SYMBOL(unregister_shrinker);
178
179 static inline int do_shrinker_shrink(struct shrinker *shrinker,
180                                      struct shrink_control *sc,
181                                      unsigned long nr_to_scan)
182 {
183         sc->nr_to_scan = nr_to_scan;
184         return (*shrinker->shrink)(shrinker, sc);
185 }
186
187 #define SHRINK_BATCH 128
188 /*
189  * Call the shrink functions to age shrinkable caches
190  *
191  * Here we assume it costs one seek to replace a lru page and that it also
192  * takes a seek to recreate a cache object.  With this in mind we age equal
193  * percentages of the lru and ageable caches.  This should balance the seeks
194  * generated by these structures.
195  *
196  * If the vm encountered mapped pages on the LRU it increase the pressure on
197  * slab to avoid swapping.
198  *
199  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
200  *
201  * `lru_pages' represents the number of on-LRU pages in all the zones which
202  * are eligible for the caller's allocation attempt.  It is used for balancing
203  * slab reclaim versus page reclaim.
204  *
205  * Returns the number of slab objects which we shrunk.
206  */
207 unsigned long shrink_slab(struct shrink_control *shrink,
208                           unsigned long nr_pages_scanned,
209                           unsigned long lru_pages)
210 {
211         struct shrinker *shrinker;
212         unsigned long ret = 0;
213
214         if (nr_pages_scanned == 0)
215                 nr_pages_scanned = SWAP_CLUSTER_MAX;
216
217         if (!down_read_trylock(&shrinker_rwsem)) {
218                 /* Assume we'll be able to shrink next time */
219                 ret = 1;
220                 goto out;
221         }
222
223         list_for_each_entry(shrinker, &shrinker_list, list) {
224                 unsigned long long delta;
225                 long total_scan;
226                 long max_pass;
227                 int shrink_ret = 0;
228                 long nr;
229                 long new_nr;
230                 long batch_size = shrinker->batch ? shrinker->batch
231                                                   : SHRINK_BATCH;
232
233                 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
234                 if (max_pass <= 0)
235                         continue;
236
237                 /*
238                  * copy the current shrinker scan count into a local variable
239                  * and zero it so that other concurrent shrinker invocations
240                  * don't also do this scanning work.
241                  */
242                 nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
243
244                 total_scan = nr;
245                 delta = (4 * nr_pages_scanned) / shrinker->seeks;
246                 delta *= max_pass;
247                 do_div(delta, lru_pages + 1);
248                 total_scan += delta;
249                 if (total_scan < 0) {
250                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
251                                "delete nr=%ld\n",
252                                shrinker->shrink, total_scan);
253                         total_scan = max_pass;
254                 }
255
256                 /*
257                  * We need to avoid excessive windup on filesystem shrinkers
258                  * due to large numbers of GFP_NOFS allocations causing the
259                  * shrinkers to return -1 all the time. This results in a large
260                  * nr being built up so when a shrink that can do some work
261                  * comes along it empties the entire cache due to nr >>>
262                  * max_pass.  This is bad for sustaining a working set in
263                  * memory.
264                  *
265                  * Hence only allow the shrinker to scan the entire cache when
266                  * a large delta change is calculated directly.
267                  */
268                 if (delta < max_pass / 4)
269                         total_scan = min(total_scan, max_pass / 2);
270
271                 /*
272                  * Avoid risking looping forever due to too large nr value:
273                  * never try to free more than twice the estimate number of
274                  * freeable entries.
275                  */
276                 if (total_scan > max_pass * 2)
277                         total_scan = max_pass * 2;
278
279                 trace_mm_shrink_slab_start(shrinker, shrink, nr,
280                                         nr_pages_scanned, lru_pages,
281                                         max_pass, delta, total_scan);
282
283                 while (total_scan >= batch_size) {
284                         int nr_before;
285
286                         nr_before = do_shrinker_shrink(shrinker, shrink, 0);
287                         shrink_ret = do_shrinker_shrink(shrinker, shrink,
288                                                         batch_size);
289                         if (shrink_ret == -1)
290                                 break;
291                         if (shrink_ret < nr_before)
292                                 ret += nr_before - shrink_ret;
293                         count_vm_events(SLABS_SCANNED, batch_size);
294                         total_scan -= batch_size;
295
296                         cond_resched();
297                 }
298
299                 /*
300                  * move the unused scan count back into the shrinker in a
301                  * manner that handles concurrent updates. If we exhausted the
302                  * scan, there is no need to do an update.
303                  */
304                 if (total_scan > 0)
305                         new_nr = atomic_long_add_return(total_scan,
306                                         &shrinker->nr_in_batch);
307                 else
308                         new_nr = atomic_long_read(&shrinker->nr_in_batch);
309
310                 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
311         }
312         up_read(&shrinker_rwsem);
313 out:
314         cond_resched();
315         return ret;
316 }
317
318 static inline int is_page_cache_freeable(struct page *page)
319 {
320         /*
321          * A freeable page cache page is referenced only by the caller
322          * that isolated the page, the page cache radix tree and
323          * optional buffer heads at page->private.
324          */
325         return page_count(page) - page_has_private(page) == 2;
326 }
327
328 static int may_write_to_queue(struct backing_dev_info *bdi,
329                               struct scan_control *sc)
330 {
331         if (current->flags & PF_SWAPWRITE)
332                 return 1;
333         if (!bdi_write_congested(bdi))
334                 return 1;
335         if (bdi == current->backing_dev_info)
336                 return 1;
337         return 0;
338 }
339
340 /*
341  * We detected a synchronous write error writing a page out.  Probably
342  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
343  * fsync(), msync() or close().
344  *
345  * The tricky part is that after writepage we cannot touch the mapping: nothing
346  * prevents it from being freed up.  But we have a ref on the page and once
347  * that page is locked, the mapping is pinned.
348  *
349  * We're allowed to run sleeping lock_page() here because we know the caller has
350  * __GFP_FS.
351  */
352 static void handle_write_error(struct address_space *mapping,
353                                 struct page *page, int error)
354 {
355         lock_page(page);
356         if (page_mapping(page) == mapping)
357                 mapping_set_error(mapping, error);
358         unlock_page(page);
359 }
360
361 /* possible outcome of pageout() */
362 typedef enum {
363         /* failed to write page out, page is locked */
364         PAGE_KEEP,
365         /* move page to the active list, page is locked */
366         PAGE_ACTIVATE,
367         /* page has been sent to the disk successfully, page is unlocked */
368         PAGE_SUCCESS,
369         /* page is clean and locked */
370         PAGE_CLEAN,
371 } pageout_t;
372
373 /*
374  * pageout is called by shrink_page_list() for each dirty page.
375  * Calls ->writepage().
376  */
377 static pageout_t pageout(struct page *page, struct address_space *mapping,
378                          struct scan_control *sc)
379 {
380         /*
381          * If the page is dirty, only perform writeback if that write
382          * will be non-blocking.  To prevent this allocation from being
383          * stalled by pagecache activity.  But note that there may be
384          * stalls if we need to run get_block().  We could test
385          * PagePrivate for that.
386          *
387          * If this process is currently in __generic_file_aio_write() against
388          * this page's queue, we can perform writeback even if that
389          * will block.
390          *
391          * If the page is swapcache, write it back even if that would
392          * block, for some throttling. This happens by accident, because
393          * swap_backing_dev_info is bust: it doesn't reflect the
394          * congestion state of the swapdevs.  Easy to fix, if needed.
395          */
396         if (!is_page_cache_freeable(page))
397                 return PAGE_KEEP;
398         if (!mapping) {
399                 /*
400                  * Some data journaling orphaned pages can have
401                  * page->mapping == NULL while being dirty with clean buffers.
402                  */
403                 if (page_has_private(page)) {
404                         if (try_to_free_buffers(page)) {
405                                 ClearPageDirty(page);
406                                 printk("%s: orphaned page\n", __func__);
407                                 return PAGE_CLEAN;
408                         }
409                 }
410                 return PAGE_KEEP;
411         }
412         if (mapping->a_ops->writepage == NULL)
413                 return PAGE_ACTIVATE;
414         if (!may_write_to_queue(mapping->backing_dev_info, sc))
415                 return PAGE_KEEP;
416
417         if (clear_page_dirty_for_io(page)) {
418                 int res;
419                 struct writeback_control wbc = {
420                         .sync_mode = WB_SYNC_NONE,
421                         .nr_to_write = SWAP_CLUSTER_MAX,
422                         .range_start = 0,
423                         .range_end = LLONG_MAX,
424                         .for_reclaim = 1,
425                 };
426
427                 SetPageReclaim(page);
428                 res = mapping->a_ops->writepage(page, &wbc);
429                 if (res < 0)
430                         handle_write_error(mapping, page, res);
431                 if (res == AOP_WRITEPAGE_ACTIVATE) {
432                         ClearPageReclaim(page);
433                         return PAGE_ACTIVATE;
434                 }
435
436                 if (!PageWriteback(page)) {
437                         /* synchronous write or broken a_ops? */
438                         ClearPageReclaim(page);
439                 }
440                 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
441                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
442                 return PAGE_SUCCESS;
443         }
444
445         return PAGE_CLEAN;
446 }
447
448 /*
449  * Same as remove_mapping, but if the page is removed from the mapping, it
450  * gets returned with a refcount of 0.
451  */
452 static int __remove_mapping(struct address_space *mapping, struct page *page)
453 {
454         BUG_ON(!PageLocked(page));
455         BUG_ON(mapping != page_mapping(page));
456
457         spin_lock_irq(&mapping->tree_lock);
458         /*
459          * The non racy check for a busy page.
460          *
461          * Must be careful with the order of the tests. When someone has
462          * a ref to the page, it may be possible that they dirty it then
463          * drop the reference. So if PageDirty is tested before page_count
464          * here, then the following race may occur:
465          *
466          * get_user_pages(&page);
467          * [user mapping goes away]
468          * write_to(page);
469          *                              !PageDirty(page)    [good]
470          * SetPageDirty(page);
471          * put_page(page);
472          *                              !page_count(page)   [good, discard it]
473          *
474          * [oops, our write_to data is lost]
475          *
476          * Reversing the order of the tests ensures such a situation cannot
477          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
478          * load is not satisfied before that of page->_count.
479          *
480          * Note that if SetPageDirty is always performed via set_page_dirty,
481          * and thus under tree_lock, then this ordering is not required.
482          */
483         if (!page_freeze_refs(page, 2))
484                 goto cannot_free;
485         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
486         if (unlikely(PageDirty(page))) {
487                 page_unfreeze_refs(page, 2);
488                 goto cannot_free;
489         }
490
491         if (PageSwapCache(page)) {
492                 swp_entry_t swap = { .val = page_private(page) };
493                 __delete_from_swap_cache(page);
494                 spin_unlock_irq(&mapping->tree_lock);
495                 swapcache_free(swap, page);
496         } else {
497                 void (*freepage)(struct page *);
498
499                 freepage = mapping->a_ops->freepage;
500
501                 __delete_from_page_cache(page);
502                 spin_unlock_irq(&mapping->tree_lock);
503                 mem_cgroup_uncharge_cache_page(page);
504
505                 if (freepage != NULL)
506                         freepage(page);
507         }
508
509         return 1;
510
511 cannot_free:
512         spin_unlock_irq(&mapping->tree_lock);
513         return 0;
514 }
515
516 /*
517  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
518  * someone else has a ref on the page, abort and return 0.  If it was
519  * successfully detached, return 1.  Assumes the caller has a single ref on
520  * this page.
521  */
522 int remove_mapping(struct address_space *mapping, struct page *page)
523 {
524         if (__remove_mapping(mapping, page)) {
525                 /*
526                  * Unfreezing the refcount with 1 rather than 2 effectively
527                  * drops the pagecache ref for us without requiring another
528                  * atomic operation.
529                  */
530                 page_unfreeze_refs(page, 1);
531                 return 1;
532         }
533         return 0;
534 }
535
536 /**
537  * putback_lru_page - put previously isolated page onto appropriate LRU list
538  * @page: page to be put back to appropriate lru list
539  *
540  * Add previously isolated @page to appropriate LRU list.
541  * Page may still be unevictable for other reasons.
542  *
543  * lru_lock must not be held, interrupts must be enabled.
544  */
545 void putback_lru_page(struct page *page)
546 {
547         int lru;
548         int active = !!TestClearPageActive(page);
549         int was_unevictable = PageUnevictable(page);
550
551         VM_BUG_ON(PageLRU(page));
552
553 redo:
554         ClearPageUnevictable(page);
555
556         if (page_evictable(page)) {
557                 /*
558                  * For evictable pages, we can use the cache.
559                  * In event of a race, worst case is we end up with an
560                  * unevictable page on [in]active list.
561                  * We know how to handle that.
562                  */
563                 lru = active + page_lru_base_type(page);
564                 lru_cache_add_lru(page, lru);
565         } else {
566                 /*
567                  * Put unevictable pages directly on zone's unevictable
568                  * list.
569                  */
570                 lru = LRU_UNEVICTABLE;
571                 add_page_to_unevictable_list(page);
572                 /*
573                  * When racing with an mlock or AS_UNEVICTABLE clearing
574                  * (page is unlocked) make sure that if the other thread
575                  * does not observe our setting of PG_lru and fails
576                  * isolation/check_move_unevictable_pages,
577                  * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
578                  * the page back to the evictable list.
579                  *
580                  * The other side is TestClearPageMlocked() or shmem_lock().
581                  */
582                 smp_mb();
583         }
584
585         /*
586          * page's status can change while we move it among lru. If an evictable
587          * page is on unevictable list, it never be freed. To avoid that,
588          * check after we added it to the list, again.
589          */
590         if (lru == LRU_UNEVICTABLE && page_evictable(page)) {
591                 if (!isolate_lru_page(page)) {
592                         put_page(page);
593                         goto redo;
594                 }
595                 /* This means someone else dropped this page from LRU
596                  * So, it will be freed or putback to LRU again. There is
597                  * nothing to do here.
598                  */
599         }
600
601         if (was_unevictable && lru != LRU_UNEVICTABLE)
602                 count_vm_event(UNEVICTABLE_PGRESCUED);
603         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
604                 count_vm_event(UNEVICTABLE_PGCULLED);
605
606         put_page(page);         /* drop ref from isolate */
607 }
608
609 enum page_references {
610         PAGEREF_RECLAIM,
611         PAGEREF_RECLAIM_CLEAN,
612         PAGEREF_KEEP,
613         PAGEREF_ACTIVATE,
614 };
615
616 static enum page_references page_check_references(struct page *page,
617                                                   struct scan_control *sc)
618 {
619         int referenced_ptes, referenced_page;
620         unsigned long vm_flags;
621
622         referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
623                                           &vm_flags);
624         referenced_page = TestClearPageReferenced(page);
625
626         /*
627          * Mlock lost the isolation race with us.  Let try_to_unmap()
628          * move the page to the unevictable list.
629          */
630         if (vm_flags & VM_LOCKED)
631                 return PAGEREF_RECLAIM;
632
633         if (referenced_ptes) {
634                 if (PageSwapBacked(page))
635                         return PAGEREF_ACTIVATE;
636                 /*
637                  * All mapped pages start out with page table
638                  * references from the instantiating fault, so we need
639                  * to look twice if a mapped file page is used more
640                  * than once.
641                  *
642                  * Mark it and spare it for another trip around the
643                  * inactive list.  Another page table reference will
644                  * lead to its activation.
645                  *
646                  * Note: the mark is set for activated pages as well
647                  * so that recently deactivated but used pages are
648                  * quickly recovered.
649                  */
650                 SetPageReferenced(page);
651
652                 if (referenced_page || referenced_ptes > 1)
653                         return PAGEREF_ACTIVATE;
654
655                 /*
656                  * Activate file-backed executable pages after first usage.
657                  */
658                 if (vm_flags & VM_EXEC)
659                         return PAGEREF_ACTIVATE;
660
661                 return PAGEREF_KEEP;
662         }
663
664         /* Reclaim if clean, defer dirty pages to writeback */
665         if (referenced_page && !PageSwapBacked(page))
666                 return PAGEREF_RECLAIM_CLEAN;
667
668         return PAGEREF_RECLAIM;
669 }
670
671 /*
672  * shrink_page_list() returns the number of reclaimed pages
673  */
674 static unsigned long shrink_page_list(struct list_head *page_list,
675                                       struct zone *zone,
676                                       struct scan_control *sc,
677                                       enum ttu_flags ttu_flags,
678                                       unsigned long *ret_nr_dirty,
679                                       unsigned long *ret_nr_writeback,
680                                       bool force_reclaim)
681 {
682         LIST_HEAD(ret_pages);
683         LIST_HEAD(free_pages);
684         int pgactivate = 0;
685         unsigned long nr_dirty = 0;
686         unsigned long nr_congested = 0;
687         unsigned long nr_reclaimed = 0;
688         unsigned long nr_writeback = 0;
689
690         cond_resched();
691
692         mem_cgroup_uncharge_start();
693         while (!list_empty(page_list)) {
694                 struct address_space *mapping;
695                 struct page *page;
696                 int may_enter_fs;
697                 enum page_references references = PAGEREF_RECLAIM_CLEAN;
698
699                 cond_resched();
700
701                 page = lru_to_page(page_list);
702                 list_del(&page->lru);
703
704                 if (!trylock_page(page))
705                         goto keep;
706
707                 VM_BUG_ON(PageActive(page));
708                 VM_BUG_ON(page_zone(page) != zone);
709
710                 sc->nr_scanned++;
711
712                 if (unlikely(!page_evictable(page)))
713                         goto cull_mlocked;
714
715                 if (!sc->may_unmap && page_mapped(page))
716                         goto keep_locked;
717
718                 /* Double the slab pressure for mapped and swapcache pages */
719                 if (page_mapped(page) || PageSwapCache(page))
720                         sc->nr_scanned++;
721
722                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
723                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
724
725                 if (PageWriteback(page)) {
726                         /*
727                          * memcg doesn't have any dirty pages throttling so we
728                          * could easily OOM just because too many pages are in
729                          * writeback and there is nothing else to reclaim.
730                          *
731                          * Check __GFP_IO, certainly because a loop driver
732                          * thread might enter reclaim, and deadlock if it waits
733                          * on a page for which it is needed to do the write
734                          * (loop masks off __GFP_IO|__GFP_FS for this reason);
735                          * but more thought would probably show more reasons.
736                          *
737                          * Don't require __GFP_FS, since we're not going into
738                          * the FS, just waiting on its writeback completion.
739                          * Worryingly, ext4 gfs2 and xfs allocate pages with
740                          * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so
741                          * testing may_enter_fs here is liable to OOM on them.
742                          */
743                         if (global_reclaim(sc) ||
744                             !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
745                                 /*
746                                  * This is slightly racy - end_page_writeback()
747                                  * might have just cleared PageReclaim, then
748                                  * setting PageReclaim here end up interpreted
749                                  * as PageReadahead - but that does not matter
750                                  * enough to care.  What we do want is for this
751                                  * page to have PageReclaim set next time memcg
752                                  * reclaim reaches the tests above, so it will
753                                  * then wait_on_page_writeback() to avoid OOM;
754                                  * and it's also appropriate in global reclaim.
755                                  */
756                                 SetPageReclaim(page);
757                                 nr_writeback++;
758                                 goto keep_locked;
759                         }
760                         wait_on_page_writeback(page);
761                 }
762
763                 if (!force_reclaim)
764                         references = page_check_references(page, sc);
765
766                 switch (references) {
767                 case PAGEREF_ACTIVATE:
768                         goto activate_locked;
769                 case PAGEREF_KEEP:
770                         goto keep_locked;
771                 case PAGEREF_RECLAIM:
772                 case PAGEREF_RECLAIM_CLEAN:
773                         ; /* try to reclaim the page below */
774                 }
775
776                 /*
777                  * Anonymous process memory has backing store?
778                  * Try to allocate it some swap space here.
779                  */
780                 if (PageAnon(page) && !PageSwapCache(page)) {
781                         if (!(sc->gfp_mask & __GFP_IO))
782                                 goto keep_locked;
783                         if (!add_to_swap(page))
784                                 goto activate_locked;
785                         may_enter_fs = 1;
786                 }
787
788                 mapping = page_mapping(page);
789
790                 /*
791                  * The page is mapped into the page tables of one or more
792                  * processes. Try to unmap it here.
793                  */
794                 if (page_mapped(page) && mapping) {
795                         switch (try_to_unmap(page, ttu_flags)) {
796                         case SWAP_FAIL:
797                                 goto activate_locked;
798                         case SWAP_AGAIN:
799                                 goto keep_locked;
800                         case SWAP_MLOCK:
801                                 goto cull_mlocked;
802                         case SWAP_SUCCESS:
803                                 ; /* try to free the page below */
804                         }
805                 }
806
807                 if (PageDirty(page)) {
808                         nr_dirty++;
809
810                         /*
811                          * Only kswapd can writeback filesystem pages to
812                          * avoid risk of stack overflow but do not writeback
813                          * unless under significant pressure.
814                          */
815                         if (page_is_file_cache(page) &&
816                                         (!current_is_kswapd() ||
817                                          sc->priority >= DEF_PRIORITY - 2)) {
818                                 /*
819                                  * Immediately reclaim when written back.
820                                  * Similar in principal to deactivate_page()
821                                  * except we already have the page isolated
822                                  * and know it's dirty
823                                  */
824                                 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
825                                 SetPageReclaim(page);
826
827                                 goto keep_locked;
828                         }
829
830                         if (references == PAGEREF_RECLAIM_CLEAN)
831                                 goto keep_locked;
832                         if (!may_enter_fs)
833                                 goto keep_locked;
834                         if (!sc->may_writepage)
835                                 goto keep_locked;
836
837                         /* Page is dirty, try to write it out here */
838                         switch (pageout(page, mapping, sc)) {
839                         case PAGE_KEEP:
840                                 nr_congested++;
841                                 goto keep_locked;
842                         case PAGE_ACTIVATE:
843                                 goto activate_locked;
844                         case PAGE_SUCCESS:
845                                 if (PageWriteback(page))
846                                         goto keep;
847                                 if (PageDirty(page))
848                                         goto keep;
849
850                                 /*
851                                  * A synchronous write - probably a ramdisk.  Go
852                                  * ahead and try to reclaim the page.
853                                  */
854                                 if (!trylock_page(page))
855                                         goto keep;
856                                 if (PageDirty(page) || PageWriteback(page))
857                                         goto keep_locked;
858                                 mapping = page_mapping(page);
859                         case PAGE_CLEAN:
860                                 ; /* try to free the page below */
861                         }
862                 }
863
864                 /*
865                  * If the page has buffers, try to free the buffer mappings
866                  * associated with this page. If we succeed we try to free
867                  * the page as well.
868                  *
869                  * We do this even if the page is PageDirty().
870                  * try_to_release_page() does not perform I/O, but it is
871                  * possible for a page to have PageDirty set, but it is actually
872                  * clean (all its buffers are clean).  This happens if the
873                  * buffers were written out directly, with submit_bh(). ext3
874                  * will do this, as well as the blockdev mapping.
875                  * try_to_release_page() will discover that cleanness and will
876                  * drop the buffers and mark the page clean - it can be freed.
877                  *
878                  * Rarely, pages can have buffers and no ->mapping.  These are
879                  * the pages which were not successfully invalidated in
880                  * truncate_complete_page().  We try to drop those buffers here
881                  * and if that worked, and the page is no longer mapped into
882                  * process address space (page_count == 1) it can be freed.
883                  * Otherwise, leave the page on the LRU so it is swappable.
884                  */
885                 if (page_has_private(page)) {
886                         if (!try_to_release_page(page, sc->gfp_mask))
887                                 goto activate_locked;
888                         if (!mapping && page_count(page) == 1) {
889                                 unlock_page(page);
890                                 if (put_page_testzero(page))
891                                         goto free_it;
892                                 else {
893                                         /*
894                                          * rare race with speculative reference.
895                                          * the speculative reference will free
896                                          * this page shortly, so we may
897                                          * increment nr_reclaimed here (and
898                                          * leave it off the LRU).
899                                          */
900                                         nr_reclaimed++;
901                                         continue;
902                                 }
903                         }
904                 }
905
906                 if (!mapping || !__remove_mapping(mapping, page))
907                         goto keep_locked;
908
909                 /*
910                  * At this point, we have no other references and there is
911                  * no way to pick any more up (removed from LRU, removed
912                  * from pagecache). Can use non-atomic bitops now (and
913                  * we obviously don't have to worry about waking up a process
914                  * waiting on the page lock, because there are no references.
915                  */
916                 __clear_page_locked(page);
917 free_it:
918                 nr_reclaimed++;
919
920                 /*
921                  * Is there need to periodically free_page_list? It would
922                  * appear not as the counts should be low
923                  */
924                 list_add(&page->lru, &free_pages);
925                 continue;
926
927 cull_mlocked:
928                 if (PageSwapCache(page))
929                         try_to_free_swap(page);
930                 unlock_page(page);
931                 putback_lru_page(page);
932                 continue;
933
934 activate_locked:
935                 /* Not a candidate for swapping, so reclaim swap space. */
936                 if (PageSwapCache(page) && vm_swap_full())
937                         try_to_free_swap(page);
938                 VM_BUG_ON(PageActive(page));
939                 SetPageActive(page);
940                 pgactivate++;
941 keep_locked:
942                 unlock_page(page);
943 keep:
944                 list_add(&page->lru, &ret_pages);
945                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
946         }
947
948         /*
949          * Tag a zone as congested if all the dirty pages encountered were
950          * backed by a congested BDI. In this case, reclaimers should just
951          * back off and wait for congestion to clear because further reclaim
952          * will encounter the same problem
953          */
954         if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc))
955                 zone_set_flag(zone, ZONE_CONGESTED);
956
957         free_hot_cold_page_list(&free_pages, 1);
958
959         list_splice(&ret_pages, page_list);
960         count_vm_events(PGACTIVATE, pgactivate);
961         mem_cgroup_uncharge_end();
962         *ret_nr_dirty += nr_dirty;
963         *ret_nr_writeback += nr_writeback;
964         return nr_reclaimed;
965 }
966
967 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
968                                             struct list_head *page_list)
969 {
970         struct scan_control sc = {
971                 .gfp_mask = GFP_KERNEL,
972                 .priority = DEF_PRIORITY,
973                 .may_unmap = 1,
974         };
975         unsigned long ret, dummy1, dummy2;
976         struct page *page, *next;
977         LIST_HEAD(clean_pages);
978
979         list_for_each_entry_safe(page, next, page_list, lru) {
980                 if (page_is_file_cache(page) && !PageDirty(page)) {
981                         ClearPageActive(page);
982                         list_move(&page->lru, &clean_pages);
983                 }
984         }
985
986         ret = shrink_page_list(&clean_pages, zone, &sc,
987                                 TTU_UNMAP|TTU_IGNORE_ACCESS,
988                                 &dummy1, &dummy2, true);
989         list_splice(&clean_pages, page_list);
990         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
991         return ret;
992 }
993
994 /*
995  * Attempt to remove the specified page from its LRU.  Only take this page
996  * if it is of the appropriate PageActive status.  Pages which are being
997  * freed elsewhere are also ignored.
998  *
999  * page:        page to consider
1000  * mode:        one of the LRU isolation modes defined above
1001  *
1002  * returns 0 on success, -ve errno on failure.
1003  */
1004 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1005 {
1006         int ret = -EINVAL;
1007
1008         /* Only take pages on the LRU. */
1009         if (!PageLRU(page))
1010                 return ret;
1011
1012         /* Compaction should not handle unevictable pages but CMA can do so */
1013         if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1014                 return ret;
1015
1016         ret = -EBUSY;
1017
1018         /*
1019          * To minimise LRU disruption, the caller can indicate that it only
1020          * wants to isolate pages it will be able to operate on without
1021          * blocking - clean pages for the most part.
1022          *
1023          * ISOLATE_CLEAN means that only clean pages should be isolated. This
1024          * is used by reclaim when it is cannot write to backing storage
1025          *
1026          * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1027          * that it is possible to migrate without blocking
1028          */
1029         if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1030                 /* All the caller can do on PageWriteback is block */
1031                 if (PageWriteback(page))
1032                         return ret;
1033
1034                 if (PageDirty(page)) {
1035                         struct address_space *mapping;
1036
1037                         /* ISOLATE_CLEAN means only clean pages */
1038                         if (mode & ISOLATE_CLEAN)
1039                                 return ret;
1040
1041                         /*
1042                          * Only pages without mappings or that have a
1043                          * ->migratepage callback are possible to migrate
1044                          * without blocking
1045                          */
1046                         mapping = page_mapping(page);
1047                         if (mapping && !mapping->a_ops->migratepage)
1048                                 return ret;
1049                 }
1050         }
1051
1052         if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1053                 return ret;
1054
1055         if (likely(get_page_unless_zero(page))) {
1056                 /*
1057                  * Be careful not to clear PageLRU until after we're
1058                  * sure the page is not being freed elsewhere -- the
1059                  * page release code relies on it.
1060                  */
1061                 ClearPageLRU(page);
1062                 ret = 0;
1063         }
1064
1065         return ret;
1066 }
1067
1068 /*
1069  * zone->lru_lock is heavily contended.  Some of the functions that
1070  * shrink the lists perform better by taking out a batch of pages
1071  * and working on them outside the LRU lock.
1072  *
1073  * For pagecache intensive workloads, this function is the hottest
1074  * spot in the kernel (apart from copy_*_user functions).
1075  *
1076  * Appropriate locks must be held before calling this function.
1077  *
1078  * @nr_to_scan: The number of pages to look through on the list.
1079  * @lruvec:     The LRU vector to pull pages from.
1080  * @dst:        The temp list to put pages on to.
1081  * @nr_scanned: The number of pages that were scanned.
1082  * @sc:         The scan_control struct for this reclaim session
1083  * @mode:       One of the LRU isolation modes
1084  * @lru:        LRU list id for isolating
1085  *
1086  * returns how many pages were moved onto *@dst.
1087  */
1088 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1089                 struct lruvec *lruvec, struct list_head *dst,
1090                 unsigned long *nr_scanned, struct scan_control *sc,
1091                 isolate_mode_t mode, enum lru_list lru)
1092 {
1093         struct list_head *src = &lruvec->lists[lru];
1094         unsigned long nr_taken = 0;
1095         unsigned long scan;
1096
1097         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1098                 struct page *page;
1099                 int nr_pages;
1100
1101                 page = lru_to_page(src);
1102                 prefetchw_prev_lru_page(page, src, flags);
1103
1104                 VM_BUG_ON(!PageLRU(page));
1105
1106                 switch (__isolate_lru_page(page, mode)) {
1107                 case 0:
1108                         nr_pages = hpage_nr_pages(page);
1109                         mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1110                         list_move(&page->lru, dst);
1111                         nr_taken += nr_pages;
1112                         break;
1113
1114                 case -EBUSY:
1115                         /* else it is being freed elsewhere */
1116                         list_move(&page->lru, src);
1117                         continue;
1118
1119                 default:
1120                         BUG();
1121                 }
1122         }
1123
1124         *nr_scanned = scan;
1125         trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1126                                     nr_taken, mode, is_file_lru(lru));
1127         return nr_taken;
1128 }
1129
1130 /**
1131  * isolate_lru_page - tries to isolate a page from its LRU list
1132  * @page: page to isolate from its LRU list
1133  *
1134  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1135  * vmstat statistic corresponding to whatever LRU list the page was on.
1136  *
1137  * Returns 0 if the page was removed from an LRU list.
1138  * Returns -EBUSY if the page was not on an LRU list.
1139  *
1140  * The returned page will have PageLRU() cleared.  If it was found on
1141  * the active list, it will have PageActive set.  If it was found on
1142  * the unevictable list, it will have the PageUnevictable bit set. That flag
1143  * may need to be cleared by the caller before letting the page go.
1144  *
1145  * The vmstat statistic corresponding to the list on which the page was
1146  * found will be decremented.
1147  *
1148  * Restrictions:
1149  * (1) Must be called with an elevated refcount on the page. This is a
1150  *     fundamentnal difference from isolate_lru_pages (which is called
1151  *     without a stable reference).
1152  * (2) the lru_lock must not be held.
1153  * (3) interrupts must be enabled.
1154  */
1155 int isolate_lru_page(struct page *page)
1156 {
1157         int ret = -EBUSY;
1158
1159         VM_BUG_ON(!page_count(page));
1160
1161         if (PageLRU(page)) {
1162                 struct zone *zone = page_zone(page);
1163                 struct lruvec *lruvec;
1164
1165                 spin_lock_irq(&zone->lru_lock);
1166                 lruvec = mem_cgroup_page_lruvec(page, zone);
1167                 if (PageLRU(page)) {
1168                         int lru = page_lru(page);
1169                         get_page(page);
1170                         ClearPageLRU(page);
1171                         del_page_from_lru_list(page, lruvec, lru);
1172                         ret = 0;
1173                 }
1174                 spin_unlock_irq(&zone->lru_lock);
1175         }
1176         return ret;
1177 }
1178
1179 /*
1180  * Are there way too many processes in the direct reclaim path already?
1181  */
1182 static int too_many_isolated(struct zone *zone, int file,
1183                 struct scan_control *sc)
1184 {
1185         unsigned long inactive, isolated;
1186
1187         if (current_is_kswapd())
1188                 return 0;
1189
1190         if (!global_reclaim(sc))
1191                 return 0;
1192
1193         if (file) {
1194                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1195                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1196         } else {
1197                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1198                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1199         }
1200
1201         return isolated > inactive;
1202 }
1203
1204 static noinline_for_stack void
1205 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1206 {
1207         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1208         struct zone *zone = lruvec_zone(lruvec);
1209         LIST_HEAD(pages_to_free);
1210
1211         /*
1212          * Put back any unfreeable pages.
1213          */
1214         while (!list_empty(page_list)) {
1215                 struct page *page = lru_to_page(page_list);
1216                 int lru;
1217
1218                 VM_BUG_ON(PageLRU(page));
1219                 list_del(&page->lru);
1220                 if (unlikely(!page_evictable(page))) {
1221                         spin_unlock_irq(&zone->lru_lock);
1222                         putback_lru_page(page);
1223                         spin_lock_irq(&zone->lru_lock);
1224                         continue;
1225                 }
1226
1227                 lruvec = mem_cgroup_page_lruvec(page, zone);
1228
1229                 SetPageLRU(page);
1230                 lru = page_lru(page);
1231                 add_page_to_lru_list(page, lruvec, lru);
1232
1233                 if (is_active_lru(lru)) {
1234                         int file = is_file_lru(lru);
1235                         int numpages = hpage_nr_pages(page);
1236                         reclaim_stat->recent_rotated[file] += numpages;
1237                 }
1238                 if (put_page_testzero(page)) {
1239                         __ClearPageLRU(page);
1240                         __ClearPageActive(page);
1241                         del_page_from_lru_list(page, lruvec, lru);
1242
1243                         if (unlikely(PageCompound(page))) {
1244                                 spin_unlock_irq(&zone->lru_lock);
1245                                 (*get_compound_page_dtor(page))(page);
1246                                 spin_lock_irq(&zone->lru_lock);
1247                         } else
1248                                 list_add(&page->lru, &pages_to_free);
1249                 }
1250         }
1251
1252         /*
1253          * To save our caller's stack, now use input list for pages to free.
1254          */
1255         list_splice(&pages_to_free, page_list);
1256 }
1257
1258 /*
1259  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1260  * of reclaimed pages
1261  */
1262 static noinline_for_stack unsigned long
1263 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1264                      struct scan_control *sc, enum lru_list lru)
1265 {
1266         LIST_HEAD(page_list);
1267         unsigned long nr_scanned;
1268         unsigned long nr_reclaimed = 0;
1269         unsigned long nr_taken;
1270         unsigned long nr_dirty = 0;
1271         unsigned long nr_writeback = 0;
1272         isolate_mode_t isolate_mode = 0;
1273         int file = is_file_lru(lru);
1274         struct zone *zone = lruvec_zone(lruvec);
1275         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1276
1277         while (unlikely(too_many_isolated(zone, file, sc))) {
1278                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1279
1280                 /* We are about to die and free our memory. Return now. */
1281                 if (fatal_signal_pending(current))
1282                         return SWAP_CLUSTER_MAX;
1283         }
1284
1285         lru_add_drain();
1286
1287         if (!sc->may_unmap)
1288                 isolate_mode |= ISOLATE_UNMAPPED;
1289         if (!sc->may_writepage)
1290                 isolate_mode |= ISOLATE_CLEAN;
1291
1292         spin_lock_irq(&zone->lru_lock);
1293
1294         nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1295                                      &nr_scanned, sc, isolate_mode, lru);
1296
1297         __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1298         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1299
1300         if (global_reclaim(sc)) {
1301                 zone->pages_scanned += nr_scanned;
1302                 if (current_is_kswapd())
1303                         __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1304                 else
1305                         __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1306         }
1307         spin_unlock_irq(&zone->lru_lock);
1308
1309         if (nr_taken == 0)
1310                 return 0;
1311
1312         nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1313                                         &nr_dirty, &nr_writeback, false);
1314
1315         spin_lock_irq(&zone->lru_lock);
1316
1317         reclaim_stat->recent_scanned[file] += nr_taken;
1318
1319         if (global_reclaim(sc)) {
1320                 if (current_is_kswapd())
1321                         __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1322                                                nr_reclaimed);
1323                 else
1324                         __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1325                                                nr_reclaimed);
1326         }
1327
1328         putback_inactive_pages(lruvec, &page_list);
1329
1330         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1331
1332         spin_unlock_irq(&zone->lru_lock);
1333
1334         free_hot_cold_page_list(&page_list, 1);
1335
1336         /*
1337          * If reclaim is isolating dirty pages under writeback, it implies
1338          * that the long-lived page allocation rate is exceeding the page
1339          * laundering rate. Either the global limits are not being effective
1340          * at throttling processes due to the page distribution throughout
1341          * zones or there is heavy usage of a slow backing device. The
1342          * only option is to throttle from reclaim context which is not ideal
1343          * as there is no guarantee the dirtying process is throttled in the
1344          * same way balance_dirty_pages() manages.
1345          *
1346          * This scales the number of dirty pages that must be under writeback
1347          * before throttling depending on priority. It is a simple backoff
1348          * function that has the most effect in the range DEF_PRIORITY to
1349          * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1350          * in trouble and reclaim is considered to be in trouble.
1351          *
1352          * DEF_PRIORITY   100% isolated pages must be PageWriteback to throttle
1353          * DEF_PRIORITY-1  50% must be PageWriteback
1354          * DEF_PRIORITY-2  25% must be PageWriteback, kswapd in trouble
1355          * ...
1356          * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1357          *                     isolated page is PageWriteback
1358          */
1359         if (nr_writeback && nr_writeback >=
1360                         (nr_taken >> (DEF_PRIORITY - sc->priority)))
1361                 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1362
1363         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1364                 zone_idx(zone),
1365                 nr_scanned, nr_reclaimed,
1366                 sc->priority,
1367                 trace_shrink_flags(file));
1368         return nr_reclaimed;
1369 }
1370
1371 /*
1372  * This moves pages from the active list to the inactive list.
1373  *
1374  * We move them the other way if the page is referenced by one or more
1375  * processes, from rmap.
1376  *
1377  * If the pages are mostly unmapped, the processing is fast and it is
1378  * appropriate to hold zone->lru_lock across the whole operation.  But if
1379  * the pages are mapped, the processing is slow (page_referenced()) so we
1380  * should drop zone->lru_lock around each page.  It's impossible to balance
1381  * this, so instead we remove the pages from the LRU while processing them.
1382  * It is safe to rely on PG_active against the non-LRU pages in here because
1383  * nobody will play with that bit on a non-LRU page.
1384  *
1385  * The downside is that we have to touch page->_count against each page.
1386  * But we had to alter page->flags anyway.
1387  */
1388
1389 static void move_active_pages_to_lru(struct lruvec *lruvec,
1390                                      struct list_head *list,
1391                                      struct list_head *pages_to_free,
1392                                      enum lru_list lru)
1393 {
1394         struct zone *zone = lruvec_zone(lruvec);
1395         unsigned long pgmoved = 0;
1396         struct page *page;
1397         int nr_pages;
1398
1399         while (!list_empty(list)) {
1400                 page = lru_to_page(list);
1401                 lruvec = mem_cgroup_page_lruvec(page, zone);
1402
1403                 VM_BUG_ON(PageLRU(page));
1404                 SetPageLRU(page);
1405
1406                 nr_pages = hpage_nr_pages(page);
1407                 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1408                 list_move(&page->lru, &lruvec->lists[lru]);
1409                 pgmoved += nr_pages;
1410
1411                 if (put_page_testzero(page)) {
1412                         __ClearPageLRU(page);
1413                         __ClearPageActive(page);
1414                         del_page_from_lru_list(page, lruvec, lru);
1415
1416                         if (unlikely(PageCompound(page))) {
1417                                 spin_unlock_irq(&zone->lru_lock);
1418                                 (*get_compound_page_dtor(page))(page);
1419                                 spin_lock_irq(&zone->lru_lock);
1420                         } else
1421                                 list_add(&page->lru, pages_to_free);
1422                 }
1423         }
1424         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1425         if (!is_active_lru(lru))
1426                 __count_vm_events(PGDEACTIVATE, pgmoved);
1427 }
1428
1429 static void shrink_active_list(unsigned long nr_to_scan,
1430                                struct lruvec *lruvec,
1431                                struct scan_control *sc,
1432                                enum lru_list lru)
1433 {
1434         unsigned long nr_taken;
1435         unsigned long nr_scanned;
1436         unsigned long vm_flags;
1437         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1438         LIST_HEAD(l_active);
1439         LIST_HEAD(l_inactive);
1440         struct page *page;
1441         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1442         unsigned long nr_rotated = 0;
1443         isolate_mode_t isolate_mode = 0;
1444         int file = is_file_lru(lru);
1445         struct zone *zone = lruvec_zone(lruvec);
1446
1447         lru_add_drain();
1448
1449         if (!sc->may_unmap)
1450                 isolate_mode |= ISOLATE_UNMAPPED;
1451         if (!sc->may_writepage)
1452                 isolate_mode |= ISOLATE_CLEAN;
1453
1454         spin_lock_irq(&zone->lru_lock);
1455
1456         nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1457                                      &nr_scanned, sc, isolate_mode, lru);
1458         if (global_reclaim(sc))
1459                 zone->pages_scanned += nr_scanned;
1460
1461         reclaim_stat->recent_scanned[file] += nr_taken;
1462
1463         __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1464         __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1465         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1466         spin_unlock_irq(&zone->lru_lock);
1467
1468         while (!list_empty(&l_hold)) {
1469                 cond_resched();
1470                 page = lru_to_page(&l_hold);
1471                 list_del(&page->lru);
1472
1473                 if (unlikely(!page_evictable(page))) {
1474                         putback_lru_page(page);
1475                         continue;
1476                 }
1477
1478                 if (unlikely(buffer_heads_over_limit)) {
1479                         if (page_has_private(page) && trylock_page(page)) {
1480                                 if (page_has_private(page))
1481                                         try_to_release_page(page, 0);
1482                                 unlock_page(page);
1483                         }
1484                 }
1485
1486                 if (page_referenced(page, 0, sc->target_mem_cgroup,
1487                                     &vm_flags)) {
1488                         nr_rotated += hpage_nr_pages(page);
1489                         /*
1490                          * Identify referenced, file-backed active pages and
1491                          * give them one more trip around the active list. So
1492                          * that executable code get better chances to stay in
1493                          * memory under moderate memory pressure.  Anon pages
1494                          * are not likely to be evicted by use-once streaming
1495                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1496                          * so we ignore them here.
1497                          */
1498                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1499                                 list_add(&page->lru, &l_active);
1500                                 continue;
1501                         }
1502                 }
1503
1504                 ClearPageActive(page);  /* we are de-activating */
1505                 list_add(&page->lru, &l_inactive);
1506         }
1507
1508         /*
1509          * Move pages back to the lru list.
1510          */
1511         spin_lock_irq(&zone->lru_lock);
1512         /*
1513          * Count referenced pages from currently used mappings as rotated,
1514          * even though only some of them are actually re-activated.  This
1515          * helps balance scan pressure between file and anonymous pages in
1516          * get_scan_ratio.
1517          */
1518         reclaim_stat->recent_rotated[file] += nr_rotated;
1519
1520         move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1521         move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1522         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1523         spin_unlock_irq(&zone->lru_lock);
1524
1525         free_hot_cold_page_list(&l_hold, 1);
1526 }
1527
1528 #ifdef CONFIG_SWAP
1529 static int inactive_anon_is_low_global(struct zone *zone)
1530 {
1531         unsigned long active, inactive;
1532
1533         active = zone_page_state(zone, NR_ACTIVE_ANON);
1534         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1535
1536         if (inactive * zone->inactive_ratio < active)
1537                 return 1;
1538
1539         return 0;
1540 }
1541
1542 /**
1543  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1544  * @lruvec: LRU vector to check
1545  *
1546  * Returns true if the zone does not have enough inactive anon pages,
1547  * meaning some active anon pages need to be deactivated.
1548  */
1549 static int inactive_anon_is_low(struct lruvec *lruvec)
1550 {
1551         /*
1552          * If we don't have swap space, anonymous page deactivation
1553          * is pointless.
1554          */
1555         if (!total_swap_pages)
1556                 return 0;
1557
1558         if (!mem_cgroup_disabled())
1559                 return mem_cgroup_inactive_anon_is_low(lruvec);
1560
1561         return inactive_anon_is_low_global(lruvec_zone(lruvec));
1562 }
1563 #else
1564 static inline int inactive_anon_is_low(struct lruvec *lruvec)
1565 {
1566         return 0;
1567 }
1568 #endif
1569
1570 static int inactive_file_is_low_global(struct zone *zone)
1571 {
1572         unsigned long active, inactive;
1573
1574         active = zone_page_state(zone, NR_ACTIVE_FILE);
1575         inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1576
1577         return (active > inactive);
1578 }
1579
1580 /**
1581  * inactive_file_is_low - check if file pages need to be deactivated
1582  * @lruvec: LRU vector to check
1583  *
1584  * When the system is doing streaming IO, memory pressure here
1585  * ensures that active file pages get deactivated, until more
1586  * than half of the file pages are on the inactive list.
1587  *
1588  * Once we get to that situation, protect the system's working
1589  * set from being evicted by disabling active file page aging.
1590  *
1591  * This uses a different ratio than the anonymous pages, because
1592  * the page cache uses a use-once replacement algorithm.
1593  */
1594 static int inactive_file_is_low(struct lruvec *lruvec)
1595 {
1596         if (!mem_cgroup_disabled())
1597                 return mem_cgroup_inactive_file_is_low(lruvec);
1598
1599         return inactive_file_is_low_global(lruvec_zone(lruvec));
1600 }
1601
1602 static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1603 {
1604         if (is_file_lru(lru))
1605                 return inactive_file_is_low(lruvec);
1606         else
1607                 return inactive_anon_is_low(lruvec);
1608 }
1609
1610 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1611                                  struct lruvec *lruvec, struct scan_control *sc)
1612 {
1613         if (is_active_lru(lru)) {
1614                 if (inactive_list_is_low(lruvec, lru))
1615                         shrink_active_list(nr_to_scan, lruvec, sc, lru);
1616                 return 0;
1617         }
1618
1619         return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1620 }
1621
1622 static int vmscan_swappiness(struct scan_control *sc)
1623 {
1624         if (global_reclaim(sc))
1625                 return vm_swappiness;
1626         return mem_cgroup_swappiness(sc->target_mem_cgroup);
1627 }
1628
1629 /*
1630  * Determine how aggressively the anon and file LRU lists should be
1631  * scanned.  The relative value of each set of LRU lists is determined
1632  * by looking at the fraction of the pages scanned we did rotate back
1633  * onto the active list instead of evict.
1634  *
1635  * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1636  * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1637  */
1638 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
1639                            unsigned long *nr)
1640 {
1641         unsigned long anon, file, free;
1642         unsigned long anon_prio, file_prio;
1643         unsigned long ap, fp;
1644         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1645         u64 fraction[2], denominator;
1646         enum lru_list lru;
1647         int noswap = 0;
1648         bool force_scan = false;
1649         struct zone *zone = lruvec_zone(lruvec);
1650
1651         /*
1652          * If the zone or memcg is small, nr[l] can be 0.  This
1653          * results in no scanning on this priority and a potential
1654          * priority drop.  Global direct reclaim can go to the next
1655          * zone and tends to have no problems. Global kswapd is for
1656          * zone balancing and it needs to scan a minimum amount. When
1657          * reclaiming for a memcg, a priority drop can cause high
1658          * latencies, so it's better to scan a minimum amount there as
1659          * well.
1660          */
1661         if (current_is_kswapd() && zone->all_unreclaimable)
1662                 force_scan = true;
1663         if (!global_reclaim(sc))
1664                 force_scan = true;
1665
1666         /* If we have no swap space, do not bother scanning anon pages. */
1667         if (!sc->may_swap || (nr_swap_pages <= 0)) {
1668                 noswap = 1;
1669                 fraction[0] = 0;
1670                 fraction[1] = 1;
1671                 denominator = 1;
1672                 goto out;
1673         }
1674
1675         anon  = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1676                 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1677         file  = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1678                 get_lru_size(lruvec, LRU_INACTIVE_FILE);
1679
1680         if (global_reclaim(sc)) {
1681                 free  = zone_page_state(zone, NR_FREE_PAGES);
1682                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1683                         /*
1684                          * If we have very few page cache pages, force-scan
1685                          * anon pages.
1686                          */
1687                         fraction[0] = 1;
1688                         fraction[1] = 0;
1689                         denominator = 1;
1690                         goto out;
1691                 } else if (!inactive_file_is_low_global(zone)) {
1692                         /*
1693                          * There is enough inactive page cache, do not
1694                          * reclaim anything from the working set right now.
1695                          */
1696                         fraction[0] = 0;
1697                         fraction[1] = 1;
1698                         denominator = 1;
1699                         goto out;
1700                 }
1701         }
1702
1703         /*
1704          * With swappiness at 100, anonymous and file have the same priority.
1705          * This scanning priority is essentially the inverse of IO cost.
1706          */
1707         anon_prio = vmscan_swappiness(sc);
1708         file_prio = 200 - anon_prio;
1709
1710         /*
1711          * OK, so we have swap space and a fair amount of page cache
1712          * pages.  We use the recently rotated / recently scanned
1713          * ratios to determine how valuable each cache is.
1714          *
1715          * Because workloads change over time (and to avoid overflow)
1716          * we keep these statistics as a floating average, which ends
1717          * up weighing recent references more than old ones.
1718          *
1719          * anon in [0], file in [1]
1720          */
1721         spin_lock_irq(&zone->lru_lock);
1722         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1723                 reclaim_stat->recent_scanned[0] /= 2;
1724                 reclaim_stat->recent_rotated[0] /= 2;
1725         }
1726
1727         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1728                 reclaim_stat->recent_scanned[1] /= 2;
1729                 reclaim_stat->recent_rotated[1] /= 2;
1730         }
1731
1732         /*
1733          * The amount of pressure on anon vs file pages is inversely
1734          * proportional to the fraction of recently scanned pages on
1735          * each list that were recently referenced and in active use.
1736          */
1737         ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
1738         ap /= reclaim_stat->recent_rotated[0] + 1;
1739
1740         fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
1741         fp /= reclaim_stat->recent_rotated[1] + 1;
1742         spin_unlock_irq(&zone->lru_lock);
1743
1744         fraction[0] = ap;
1745         fraction[1] = fp;
1746         denominator = ap + fp + 1;
1747 out:
1748         for_each_evictable_lru(lru) {
1749                 int file = is_file_lru(lru);
1750                 unsigned long scan;
1751
1752                 scan = get_lru_size(lruvec, lru);
1753                 if (sc->priority || noswap || !vmscan_swappiness(sc)) {
1754                         scan >>= sc->priority;
1755                         if (!scan && force_scan)
1756                                 scan = SWAP_CLUSTER_MAX;
1757                         scan = div64_u64(scan * fraction[file], denominator);
1758                 }
1759                 nr[lru] = scan;
1760         }
1761 }
1762
1763 /* Use reclaim/compaction for costly allocs or under memory pressure */
1764 static bool in_reclaim_compaction(struct scan_control *sc)
1765 {
1766         if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
1767                         (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
1768                          sc->priority < DEF_PRIORITY - 2))
1769                 return true;
1770
1771         return false;
1772 }
1773
1774 /*
1775  * Reclaim/compaction is used for high-order allocation requests. It reclaims
1776  * order-0 pages before compacting the zone. should_continue_reclaim() returns
1777  * true if more pages should be reclaimed such that when the page allocator
1778  * calls try_to_compact_zone() that it will have enough free pages to succeed.
1779  * It will give up earlier than that if there is difficulty reclaiming pages.
1780  */
1781 static inline bool should_continue_reclaim(struct lruvec *lruvec,
1782                                         unsigned long nr_reclaimed,
1783                                         unsigned long nr_scanned,
1784                                         struct scan_control *sc)
1785 {
1786         unsigned long pages_for_compaction;
1787         unsigned long inactive_lru_pages;
1788
1789         /* If not in reclaim/compaction mode, stop */
1790         if (!in_reclaim_compaction(sc))
1791                 return false;
1792
1793         /* Consider stopping depending on scan and reclaim activity */
1794         if (sc->gfp_mask & __GFP_REPEAT) {
1795                 /*
1796                  * For __GFP_REPEAT allocations, stop reclaiming if the
1797                  * full LRU list has been scanned and we are still failing
1798                  * to reclaim pages. This full LRU scan is potentially
1799                  * expensive but a __GFP_REPEAT caller really wants to succeed
1800                  */
1801                 if (!nr_reclaimed && !nr_scanned)
1802                         return false;
1803         } else {
1804                 /*
1805                  * For non-__GFP_REPEAT allocations which can presumably
1806                  * fail without consequence, stop if we failed to reclaim
1807                  * any pages from the last SWAP_CLUSTER_MAX number of
1808                  * pages that were scanned. This will return to the
1809                  * caller faster at the risk reclaim/compaction and
1810                  * the resulting allocation attempt fails
1811                  */
1812                 if (!nr_reclaimed)
1813                         return false;
1814         }
1815
1816         /*
1817          * If we have not reclaimed enough pages for compaction and the
1818          * inactive lists are large enough, continue reclaiming
1819          */
1820         pages_for_compaction = (2UL << sc->order);
1821         inactive_lru_pages = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1822         if (nr_swap_pages > 0)
1823                 inactive_lru_pages += get_lru_size(lruvec, LRU_INACTIVE_ANON);
1824         if (sc->nr_reclaimed < pages_for_compaction &&
1825                         inactive_lru_pages > pages_for_compaction)
1826                 return true;
1827
1828         /* If compaction would go ahead or the allocation would succeed, stop */
1829         switch (compaction_suitable(lruvec_zone(lruvec), sc->order)) {
1830         case COMPACT_PARTIAL:
1831         case COMPACT_CONTINUE:
1832                 return false;
1833         default:
1834                 return true;
1835         }
1836 }
1837
1838 /*
1839  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1840  */
1841 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
1842 {
1843         unsigned long nr[NR_LRU_LISTS];
1844         unsigned long nr_to_scan;
1845         enum lru_list lru;
1846         unsigned long nr_reclaimed, nr_scanned;
1847         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1848         struct blk_plug plug;
1849
1850 restart:
1851         nr_reclaimed = 0;
1852         nr_scanned = sc->nr_scanned;
1853         get_scan_count(lruvec, sc, nr);
1854
1855         blk_start_plug(&plug);
1856         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1857                                         nr[LRU_INACTIVE_FILE]) {
1858                 for_each_evictable_lru(lru) {
1859                         if (nr[lru]) {
1860                                 nr_to_scan = min_t(unsigned long,
1861                                                    nr[lru], SWAP_CLUSTER_MAX);
1862                                 nr[lru] -= nr_to_scan;
1863
1864                                 nr_reclaimed += shrink_list(lru, nr_to_scan,
1865                                                             lruvec, sc);
1866                         }
1867                 }
1868                 /*
1869                  * On large memory systems, scan >> priority can become
1870                  * really large. This is fine for the starting priority;
1871                  * we want to put equal scanning pressure on each zone.
1872                  * However, if the VM has a harder time of freeing pages,
1873                  * with multiple processes reclaiming pages, the total
1874                  * freeing target can get unreasonably large.
1875                  */
1876                 if (nr_reclaimed >= nr_to_reclaim &&
1877                     sc->priority < DEF_PRIORITY)
1878                         break;
1879         }
1880         blk_finish_plug(&plug);
1881         sc->nr_reclaimed += nr_reclaimed;
1882
1883         /*
1884          * Even if we did not try to evict anon pages at all, we want to
1885          * rebalance the anon lru active/inactive ratio.
1886          */
1887         if (inactive_anon_is_low(lruvec))
1888                 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
1889                                    sc, LRU_ACTIVE_ANON);
1890
1891         /* reclaim/compaction might need reclaim to continue */
1892         if (should_continue_reclaim(lruvec, nr_reclaimed,
1893                                     sc->nr_scanned - nr_scanned, sc))
1894                 goto restart;
1895
1896         throttle_vm_writeout(sc->gfp_mask);
1897 }
1898
1899 static void shrink_zone(struct zone *zone, struct scan_control *sc)
1900 {
1901         struct mem_cgroup *root = sc->target_mem_cgroup;
1902         struct mem_cgroup_reclaim_cookie reclaim = {
1903                 .zone = zone,
1904                 .priority = sc->priority,
1905         };
1906         struct mem_cgroup *memcg;
1907
1908         memcg = mem_cgroup_iter(root, NULL, &reclaim);
1909         do {
1910                 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
1911
1912                 shrink_lruvec(lruvec, sc);
1913
1914                 /*
1915                  * Limit reclaim has historically picked one memcg and
1916                  * scanned it with decreasing priority levels until
1917                  * nr_to_reclaim had been reclaimed.  This priority
1918                  * cycle is thus over after a single memcg.
1919                  *
1920                  * Direct reclaim and kswapd, on the other hand, have
1921                  * to scan all memory cgroups to fulfill the overall
1922                  * scan target for the zone.
1923                  */
1924                 if (!global_reclaim(sc)) {
1925                         mem_cgroup_iter_break(root, memcg);
1926                         break;
1927                 }
1928                 memcg = mem_cgroup_iter(root, memcg, &reclaim);
1929         } while (memcg);
1930 }
1931
1932 /* Returns true if compaction should go ahead for a high-order request */
1933 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
1934 {
1935         unsigned long balance_gap, watermark;
1936         bool watermark_ok;
1937
1938         /* Do not consider compaction for orders reclaim is meant to satisfy */
1939         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
1940                 return false;
1941
1942         /*
1943          * Compaction takes time to run and there are potentially other
1944          * callers using the pages just freed. Continue reclaiming until
1945          * there is a buffer of free pages available to give compaction
1946          * a reasonable chance of completing and allocating the page
1947          */
1948         balance_gap = min(low_wmark_pages(zone),
1949                 (zone->present_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
1950                         KSWAPD_ZONE_BALANCE_GAP_RATIO);
1951         watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
1952         watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
1953
1954         /*
1955          * If compaction is deferred, reclaim up to a point where
1956          * compaction will have a chance of success when re-enabled
1957          */
1958         if (compaction_deferred(zone, sc->order))
1959                 return watermark_ok;
1960
1961         /* If compaction is not ready to start, keep reclaiming */
1962         if (!compaction_suitable(zone, sc->order))
1963                 return false;
1964
1965         return watermark_ok;
1966 }
1967
1968 /*
1969  * This is the direct reclaim path, for page-allocating processes.  We only
1970  * try to reclaim pages from zones which will satisfy the caller's allocation
1971  * request.
1972  *
1973  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1974  * Because:
1975  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1976  *    allocation or
1977  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1978  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1979  *    zone defense algorithm.
1980  *
1981  * If a zone is deemed to be full of pinned pages then just give it a light
1982  * scan then give up on it.
1983  *
1984  * This function returns true if a zone is being reclaimed for a costly
1985  * high-order allocation and compaction is ready to begin. This indicates to
1986  * the caller that it should consider retrying the allocation instead of
1987  * further reclaim.
1988  */
1989 static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
1990 {
1991         struct zoneref *z;
1992         struct zone *zone;
1993         unsigned long nr_soft_reclaimed;
1994         unsigned long nr_soft_scanned;
1995         bool aborted_reclaim = false;
1996
1997         /*
1998          * If the number of buffer_heads in the machine exceeds the maximum
1999          * allowed level, force direct reclaim to scan the highmem zone as
2000          * highmem pages could be pinning lowmem pages storing buffer_heads
2001          */
2002         if (buffer_heads_over_limit)
2003                 sc->gfp_mask |= __GFP_HIGHMEM;
2004
2005         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2006                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2007                 if (!populated_zone(zone))
2008                         continue;
2009                 /*
2010                  * Take care memory controller reclaiming has small influence
2011                  * to global LRU.
2012                  */
2013                 if (global_reclaim(sc)) {
2014                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2015                                 continue;
2016                         if (zone->all_unreclaimable &&
2017                                         sc->priority != DEF_PRIORITY)
2018                                 continue;       /* Let kswapd poll it */
2019                         if (IS_ENABLED(CONFIG_COMPACTION)) {
2020                                 /*
2021                                  * If we already have plenty of memory free for
2022                                  * compaction in this zone, don't free any more.
2023                                  * Even though compaction is invoked for any
2024                                  * non-zero order, only frequent costly order
2025                                  * reclamation is disruptive enough to become a
2026                                  * noticeable problem, like transparent huge
2027                                  * page allocations.
2028                                  */
2029                                 if (compaction_ready(zone, sc)) {
2030                                         aborted_reclaim = true;
2031                                         continue;
2032                                 }
2033                         }
2034                         /*
2035                          * This steals pages from memory cgroups over softlimit
2036                          * and returns the number of reclaimed pages and
2037                          * scanned pages. This works for global memory pressure
2038                          * and balancing, not for a memcg's limit.
2039                          */
2040                         nr_soft_scanned = 0;
2041                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2042                                                 sc->order, sc->gfp_mask,
2043                                                 &nr_soft_scanned);
2044                         sc->nr_reclaimed += nr_soft_reclaimed;
2045                         sc->nr_scanned += nr_soft_scanned;
2046                         /* need some check for avoid more shrink_zone() */
2047                 }
2048
2049                 shrink_zone(zone, sc);
2050         }
2051
2052         return aborted_reclaim;
2053 }
2054
2055 static bool zone_reclaimable(struct zone *zone)
2056 {
2057         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2058 }
2059
2060 /* All zones in zonelist are unreclaimable? */
2061 static bool all_unreclaimable(struct zonelist *zonelist,
2062                 struct scan_control *sc)
2063 {
2064         struct zoneref *z;
2065         struct zone *zone;
2066
2067         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2068                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2069                 if (!populated_zone(zone))
2070                         continue;
2071                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2072                         continue;
2073                 if (!zone->all_unreclaimable)
2074                         return false;
2075         }
2076
2077         return true;
2078 }
2079
2080 /*
2081  * This is the main entry point to direct page reclaim.
2082  *
2083  * If a full scan of the inactive list fails to free enough memory then we
2084  * are "out of memory" and something needs to be killed.
2085  *
2086  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2087  * high - the zone may be full of dirty or under-writeback pages, which this
2088  * caller can't do much about.  We kick the writeback threads and take explicit
2089  * naps in the hope that some of these pages can be written.  But if the
2090  * allocating task holds filesystem locks which prevent writeout this might not
2091  * work, and the allocation attempt will fail.
2092  *
2093  * returns:     0, if no pages reclaimed
2094  *              else, the number of pages reclaimed
2095  */
2096 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2097                                         struct scan_control *sc,
2098                                         struct shrink_control *shrink)
2099 {
2100         unsigned long total_scanned = 0;
2101         struct reclaim_state *reclaim_state = current->reclaim_state;
2102         struct zoneref *z;
2103         struct zone *zone;
2104         unsigned long writeback_threshold;
2105         bool aborted_reclaim;
2106
2107         delayacct_freepages_start();
2108
2109         if (global_reclaim(sc))
2110                 count_vm_event(ALLOCSTALL);
2111
2112         do {
2113                 sc->nr_scanned = 0;
2114                 aborted_reclaim = shrink_zones(zonelist, sc);
2115
2116                 /*
2117                  * Don't shrink slabs when reclaiming memory from
2118                  * over limit cgroups
2119                  */
2120                 if (global_reclaim(sc)) {
2121                         unsigned long lru_pages = 0;
2122                         for_each_zone_zonelist(zone, z, zonelist,
2123                                         gfp_zone(sc->gfp_mask)) {
2124                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2125                                         continue;
2126
2127                                 lru_pages += zone_reclaimable_pages(zone);
2128                         }
2129
2130                         shrink_slab(shrink, sc->nr_scanned, lru_pages);
2131                         if (reclaim_state) {
2132                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2133                                 reclaim_state->reclaimed_slab = 0;
2134                         }
2135                 }
2136                 total_scanned += sc->nr_scanned;
2137                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2138                         goto out;
2139
2140                 /*
2141                  * Try to write back as many pages as we just scanned.  This
2142                  * tends to cause slow streaming writers to write data to the
2143                  * disk smoothly, at the dirtying rate, which is nice.   But
2144                  * that's undesirable in laptop mode, where we *want* lumpy
2145                  * writeout.  So in laptop mode, write out the whole world.
2146                  */
2147                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2148                 if (total_scanned > writeback_threshold) {
2149                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2150                                                 WB_REASON_TRY_TO_FREE_PAGES);
2151                         sc->may_writepage = 1;
2152                 }
2153
2154                 /* Take a nap, wait for some writeback to complete */
2155                 if (!sc->hibernation_mode && sc->nr_scanned &&
2156                     sc->priority < DEF_PRIORITY - 2) {
2157                         struct zone *preferred_zone;
2158
2159                         first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2160                                                 &cpuset_current_mems_allowed,
2161                                                 &preferred_zone);
2162                         wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2163                 }
2164         } while (--sc->priority >= 0);
2165
2166 out:
2167         delayacct_freepages_end();
2168
2169         if (sc->nr_reclaimed)
2170                 return sc->nr_reclaimed;
2171
2172         /*
2173          * As hibernation is going on, kswapd is freezed so that it can't mark
2174          * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2175          * check.
2176          */
2177         if (oom_killer_disabled)
2178                 return 0;
2179
2180         /* Aborted reclaim to try compaction? don't OOM, then */
2181         if (aborted_reclaim)
2182                 return 1;
2183
2184         /* top priority shrink_zones still had more to do? don't OOM, then */
2185         if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2186                 return 1;
2187
2188         return 0;
2189 }
2190
2191 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2192 {
2193         struct zone *zone;
2194         unsigned long pfmemalloc_reserve = 0;
2195         unsigned long free_pages = 0;
2196         int i;
2197         bool wmark_ok;
2198
2199         for (i = 0; i <= ZONE_NORMAL; i++) {
2200                 zone = &pgdat->node_zones[i];
2201                 pfmemalloc_reserve += min_wmark_pages(zone);
2202                 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2203         }
2204
2205         wmark_ok = free_pages > pfmemalloc_reserve / 2;
2206
2207         /* kswapd must be awake if processes are being throttled */
2208         if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2209                 pgdat->classzone_idx = min(pgdat->classzone_idx,
2210                                                 (enum zone_type)ZONE_NORMAL);
2211                 wake_up_interruptible(&pgdat->kswapd_wait);
2212         }
2213
2214         return wmark_ok;
2215 }
2216
2217 /*
2218  * Throttle direct reclaimers if backing storage is backed by the network
2219  * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2220  * depleted. kswapd will continue to make progress and wake the processes
2221  * when the low watermark is reached.
2222  *
2223  * Returns true if a fatal signal was delivered during throttling. If this
2224  * happens, the page allocator should not consider triggering the OOM killer.
2225  */
2226 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2227                                         nodemask_t *nodemask)
2228 {
2229         struct zone *zone;
2230         int high_zoneidx = gfp_zone(gfp_mask);
2231         pg_data_t *pgdat;
2232
2233         /*
2234          * Kernel threads should not be throttled as they may be indirectly
2235          * responsible for cleaning pages necessary for reclaim to make forward
2236          * progress. kjournald for example may enter direct reclaim while
2237          * committing a transaction where throttling it could forcing other
2238          * processes to block on log_wait_commit().
2239          */
2240         if (current->flags & PF_KTHREAD)
2241                 goto out;
2242
2243         /*
2244          * If a fatal signal is pending, this process should not throttle.
2245          * It should return quickly so it can exit and free its memory
2246          */
2247         if (fatal_signal_pending(current))
2248                 goto out;
2249
2250         /* Check if the pfmemalloc reserves are ok */
2251         first_zones_zonelist(zonelist, high_zoneidx, NULL, &zone);
2252         pgdat = zone->zone_pgdat;
2253         if (pfmemalloc_watermark_ok(pgdat))
2254                 goto out;
2255
2256         /* Account for the throttling */
2257         count_vm_event(PGSCAN_DIRECT_THROTTLE);
2258
2259         /*
2260          * If the caller cannot enter the filesystem, it's possible that it
2261          * is due to the caller holding an FS lock or performing a journal
2262          * transaction in the case of a filesystem like ext[3|4]. In this case,
2263          * it is not safe to block on pfmemalloc_wait as kswapd could be
2264          * blocked waiting on the same lock. Instead, throttle for up to a
2265          * second before continuing.
2266          */
2267         if (!(gfp_mask & __GFP_FS)) {
2268                 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2269                         pfmemalloc_watermark_ok(pgdat), HZ);
2270
2271                 goto check_pending;
2272         }
2273
2274         /* Throttle until kswapd wakes the process */
2275         wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2276                 pfmemalloc_watermark_ok(pgdat));
2277
2278 check_pending:
2279         if (fatal_signal_pending(current))
2280                 return true;
2281
2282 out:
2283         return false;
2284 }
2285
2286 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2287                                 gfp_t gfp_mask, nodemask_t *nodemask)
2288 {
2289         unsigned long nr_reclaimed;
2290         struct scan_control sc = {
2291                 .gfp_mask = gfp_mask,
2292                 .may_writepage = !laptop_mode,
2293                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2294                 .may_unmap = 1,
2295                 .may_swap = 1,
2296                 .order = order,
2297                 .priority = DEF_PRIORITY,
2298                 .target_mem_cgroup = NULL,
2299                 .nodemask = nodemask,
2300         };
2301         struct shrink_control shrink = {
2302                 .gfp_mask = sc.gfp_mask,
2303         };
2304
2305         /*
2306          * Do not enter reclaim if fatal signal was delivered while throttled.
2307          * 1 is returned so that the page allocator does not OOM kill at this
2308          * point.
2309          */
2310         if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2311                 return 1;
2312
2313         trace_mm_vmscan_direct_reclaim_begin(order,
2314                                 sc.may_writepage,
2315                                 gfp_mask);
2316
2317         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2318
2319         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2320
2321         return nr_reclaimed;
2322 }
2323
2324 #ifdef CONFIG_MEMCG
2325
2326 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2327                                                 gfp_t gfp_mask, bool noswap,
2328                                                 struct zone *zone,
2329                                                 unsigned long *nr_scanned)
2330 {
2331         struct scan_control sc = {
2332                 .nr_scanned = 0,
2333                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2334                 .may_writepage = !laptop_mode,
2335                 .may_unmap = 1,
2336                 .may_swap = !noswap,
2337                 .order = 0,
2338                 .priority = 0,
2339                 .target_mem_cgroup = memcg,
2340         };
2341         struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2342
2343         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2344                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2345
2346         trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2347                                                       sc.may_writepage,
2348                                                       sc.gfp_mask);
2349
2350         /*
2351          * NOTE: Although we can get the priority field, using it
2352          * here is not a good idea, since it limits the pages we can scan.
2353          * if we don't reclaim here, the shrink_zone from balance_pgdat
2354          * will pick up pages from other mem cgroup's as well. We hack
2355          * the priority and make it zero.
2356          */
2357         shrink_lruvec(lruvec, &sc);
2358
2359         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2360
2361         *nr_scanned = sc.nr_scanned;
2362         return sc.nr_reclaimed;
2363 }
2364
2365 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2366                                            gfp_t gfp_mask,
2367                                            bool noswap)
2368 {
2369         struct zonelist *zonelist;
2370         unsigned long nr_reclaimed;
2371         int nid;
2372         struct scan_control sc = {
2373                 .may_writepage = !laptop_mode,
2374                 .may_unmap = 1,
2375                 .may_swap = !noswap,
2376                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2377                 .order = 0,
2378                 .priority = DEF_PRIORITY,
2379                 .target_mem_cgroup = memcg,
2380                 .nodemask = NULL, /* we don't care the placement */
2381                 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2382                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2383         };
2384         struct shrink_control shrink = {
2385                 .gfp_mask = sc.gfp_mask,
2386         };
2387
2388         /*
2389          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2390          * take care of from where we get pages. So the node where we start the
2391          * scan does not need to be the current node.
2392          */
2393         nid = mem_cgroup_select_victim_node(memcg);
2394
2395         zonelist = NODE_DATA(nid)->node_zonelists;
2396
2397         trace_mm_vmscan_memcg_reclaim_begin(0,
2398                                             sc.may_writepage,
2399                                             sc.gfp_mask);
2400
2401         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2402
2403         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2404
2405         return nr_reclaimed;
2406 }
2407 #endif
2408
2409 static void age_active_anon(struct zone *zone, struct scan_control *sc)
2410 {
2411         struct mem_cgroup *memcg;
2412
2413         if (!total_swap_pages)
2414                 return;
2415
2416         memcg = mem_cgroup_iter(NULL, NULL, NULL);
2417         do {
2418                 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2419
2420                 if (inactive_anon_is_low(lruvec))
2421                         shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2422                                            sc, LRU_ACTIVE_ANON);
2423
2424                 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2425         } while (memcg);
2426 }
2427
2428 static bool zone_balanced(struct zone *zone, int order,
2429                           unsigned long balance_gap, int classzone_idx)
2430 {
2431         if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2432                                     balance_gap, classzone_idx, 0))
2433                 return false;
2434
2435         if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2436             !compaction_suitable(zone, order))
2437                 return false;
2438
2439         return true;
2440 }
2441
2442 /*
2443  * pgdat_balanced is used when checking if a node is balanced for high-order
2444  * allocations. Only zones that meet watermarks and are in a zone allowed
2445  * by the callers classzone_idx are added to balanced_pages. The total of
2446  * balanced pages must be at least 25% of the zones allowed by classzone_idx
2447  * for the node to be considered balanced. Forcing all zones to be balanced
2448  * for high orders can cause excessive reclaim when there are imbalanced zones.
2449  * The choice of 25% is due to
2450  *   o a 16M DMA zone that is balanced will not balance a zone on any
2451  *     reasonable sized machine
2452  *   o On all other machines, the top zone must be at least a reasonable
2453  *     percentage of the middle zones. For example, on 32-bit x86, highmem
2454  *     would need to be at least 256M for it to be balance a whole node.
2455  *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2456  *     to balance a node on its own. These seemed like reasonable ratios.
2457  */
2458 static bool pgdat_balanced(pg_data_t *pgdat, unsigned long balanced_pages,
2459                                                 int classzone_idx)
2460 {
2461         unsigned long present_pages = 0;
2462         int i;
2463
2464         for (i = 0; i <= classzone_idx; i++)
2465                 present_pages += pgdat->node_zones[i].present_pages;
2466
2467         /* A special case here: if zone has no page, we think it's balanced */
2468         return balanced_pages >= (present_pages >> 2);
2469 }
2470
2471 /*
2472  * Prepare kswapd for sleeping. This verifies that there are no processes
2473  * waiting in throttle_direct_reclaim() and that watermarks have been met.
2474  *
2475  * Returns true if kswapd is ready to sleep
2476  */
2477 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
2478                                         int classzone_idx)
2479 {
2480         int i;
2481         unsigned long balanced = 0;
2482         bool all_zones_ok = true;
2483
2484         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2485         if (remaining)
2486                 return false;
2487
2488         /*
2489          * There is a potential race between when kswapd checks its watermarks
2490          * and a process gets throttled. There is also a potential race if
2491          * processes get throttled, kswapd wakes, a large process exits therby
2492          * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2493          * is going to sleep, no process should be sleeping on pfmemalloc_wait
2494          * so wake them now if necessary. If necessary, processes will wake
2495          * kswapd and get throttled again
2496          */
2497         if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2498                 wake_up(&pgdat->pfmemalloc_wait);
2499                 return false;
2500         }
2501
2502         /* Check the watermark levels */
2503         for (i = 0; i <= classzone_idx; i++) {
2504                 struct zone *zone = pgdat->node_zones + i;
2505
2506                 if (!populated_zone(zone))
2507                         continue;
2508
2509                 /*
2510                  * balance_pgdat() skips over all_unreclaimable after
2511                  * DEF_PRIORITY. Effectively, it considers them balanced so
2512                  * they must be considered balanced here as well if kswapd
2513                  * is to sleep
2514                  */
2515                 if (zone->all_unreclaimable) {
2516                         balanced += zone->present_pages;
2517                         continue;
2518                 }
2519
2520                 if (!zone_balanced(zone, order, 0, i))
2521                         all_zones_ok = false;
2522                 else
2523                         balanced += zone->present_pages;
2524         }
2525
2526         /*
2527          * For high-order requests, the balanced zones must contain at least
2528          * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2529          * must be balanced
2530          */
2531         if (order)
2532                 return pgdat_balanced(pgdat, balanced, classzone_idx);
2533         else
2534                 return all_zones_ok;
2535 }
2536
2537 /*
2538  * For kswapd, balance_pgdat() will work across all this node's zones until
2539  * they are all at high_wmark_pages(zone).
2540  *
2541  * Returns the final order kswapd was reclaiming at
2542  *
2543  * There is special handling here for zones which are full of pinned pages.
2544  * This can happen if the pages are all mlocked, or if they are all used by
2545  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2546  * What we do is to detect the case where all pages in the zone have been
2547  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2548  * dead and from now on, only perform a short scan.  Basically we're polling
2549  * the zone for when the problem goes away.
2550  *
2551  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2552  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2553  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2554  * lower zones regardless of the number of free pages in the lower zones. This
2555  * interoperates with the page allocator fallback scheme to ensure that aging
2556  * of pages is balanced across the zones.
2557  */
2558 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2559                                                         int *classzone_idx)
2560 {
2561         int all_zones_ok;
2562         unsigned long balanced;
2563         int i;
2564         int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2565         unsigned long total_scanned;
2566         struct reclaim_state *reclaim_state = current->reclaim_state;
2567         unsigned long nr_soft_reclaimed;
2568         unsigned long nr_soft_scanned;
2569         struct scan_control sc = {
2570                 .gfp_mask = GFP_KERNEL,
2571                 .may_unmap = 1,
2572                 .may_swap = 1,
2573                 /*
2574                  * kswapd doesn't want to be bailed out while reclaim. because
2575                  * we want to put equal scanning pressure on each zone.
2576                  */
2577                 .nr_to_reclaim = ULONG_MAX,
2578                 .order = order,
2579                 .target_mem_cgroup = NULL,
2580         };
2581         struct shrink_control shrink = {
2582                 .gfp_mask = sc.gfp_mask,
2583         };
2584 loop_again:
2585         total_scanned = 0;
2586         sc.priority = DEF_PRIORITY;
2587         sc.nr_reclaimed = 0;
2588         sc.may_writepage = !laptop_mode;
2589         count_vm_event(PAGEOUTRUN);
2590
2591         do {
2592                 unsigned long lru_pages = 0;
2593                 int has_under_min_watermark_zone = 0;
2594
2595                 all_zones_ok = 1;
2596                 balanced = 0;
2597
2598                 /*
2599                  * Scan in the highmem->dma direction for the highest
2600                  * zone which needs scanning
2601                  */
2602                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2603                         struct zone *zone = pgdat->node_zones + i;
2604
2605                         if (!populated_zone(zone))
2606                                 continue;
2607
2608                         if (zone->all_unreclaimable &&
2609                             sc.priority != DEF_PRIORITY)
2610                                 continue;
2611
2612                         /*
2613                          * Do some background aging of the anon list, to give
2614                          * pages a chance to be referenced before reclaiming.
2615                          */
2616                         age_active_anon(zone, &sc);
2617
2618                         /*
2619                          * If the number of buffer_heads in the machine
2620                          * exceeds the maximum allowed level and this node
2621                          * has a highmem zone, force kswapd to reclaim from
2622                          * it to relieve lowmem pressure.
2623                          */
2624                         if (buffer_heads_over_limit && is_highmem_idx(i)) {
2625                                 end_zone = i;
2626                                 break;
2627                         }
2628
2629                         if (!zone_balanced(zone, order, 0, 0)) {
2630                                 end_zone = i;
2631                                 break;
2632                         } else {
2633                                 /* If balanced, clear the congested flag */
2634                                 zone_clear_flag(zone, ZONE_CONGESTED);
2635                         }
2636                 }
2637                 if (i < 0)
2638                         goto out;
2639
2640                 for (i = 0; i <= end_zone; i++) {
2641                         struct zone *zone = pgdat->node_zones + i;
2642
2643                         lru_pages += zone_reclaimable_pages(zone);
2644                 }
2645
2646                 /*
2647                  * Now scan the zone in the dma->highmem direction, stopping
2648                  * at the last zone which needs scanning.
2649                  *
2650                  * We do this because the page allocator works in the opposite
2651                  * direction.  This prevents the page allocator from allocating
2652                  * pages behind kswapd's direction of progress, which would
2653                  * cause too much scanning of the lower zones.
2654                  */
2655                 for (i = 0; i <= end_zone; i++) {
2656                         struct zone *zone = pgdat->node_zones + i;
2657                         int nr_slab, testorder;
2658                         unsigned long balance_gap;
2659
2660                         if (!populated_zone(zone))
2661                                 continue;
2662
2663                         if (zone->all_unreclaimable &&
2664                             sc.priority != DEF_PRIORITY)
2665                                 continue;
2666
2667                         sc.nr_scanned = 0;
2668
2669                         nr_soft_scanned = 0;
2670                         /*
2671                          * Call soft limit reclaim before calling shrink_zone.
2672                          */
2673                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2674                                                         order, sc.gfp_mask,
2675                                                         &nr_soft_scanned);
2676                         sc.nr_reclaimed += nr_soft_reclaimed;
2677                         total_scanned += nr_soft_scanned;
2678
2679                         /*
2680                          * We put equal pressure on every zone, unless
2681                          * one zone has way too many pages free
2682                          * already. The "too many pages" is defined
2683                          * as the high wmark plus a "gap" where the
2684                          * gap is either the low watermark or 1%
2685                          * of the zone, whichever is smaller.
2686                          */
2687                         balance_gap = min(low_wmark_pages(zone),
2688                                 (zone->present_pages +
2689                                         KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2690                                 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2691                         /*
2692                          * Kswapd reclaims only single pages with compaction
2693                          * enabled. Trying too hard to reclaim until contiguous
2694                          * free pages have become available can hurt performance
2695                          * by evicting too much useful data from memory.
2696                          * Do not reclaim more than needed for compaction.
2697                          */
2698                         testorder = order;
2699                         if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2700                                         compaction_suitable(zone, order) !=
2701                                                 COMPACT_SKIPPED)
2702                                 testorder = 0;
2703
2704                         if ((buffer_heads_over_limit && is_highmem_idx(i)) ||
2705                             !zone_balanced(zone, testorder,
2706                                            balance_gap, end_zone)) {
2707                                 shrink_zone(zone, &sc);
2708
2709                                 reclaim_state->reclaimed_slab = 0;
2710                                 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2711                                 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2712                                 total_scanned += sc.nr_scanned;
2713
2714                                 if (nr_slab == 0 && !zone_reclaimable(zone))
2715                                         zone->all_unreclaimable = 1;
2716                         }
2717
2718                         /*
2719                          * If we've done a decent amount of scanning and
2720                          * the reclaim ratio is low, start doing writepage
2721                          * even in laptop mode
2722                          */
2723                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
2724                             total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
2725                                 sc.may_writepage = 1;
2726
2727                         if (zone->all_unreclaimable) {
2728                                 if (end_zone && end_zone == i)
2729                                         end_zone--;
2730                                 continue;
2731                         }
2732
2733                         if (!zone_balanced(zone, testorder, 0, end_zone)) {
2734                                 all_zones_ok = 0;
2735                                 /*
2736                                  * We are still under min water mark.  This
2737                                  * means that we have a GFP_ATOMIC allocation
2738                                  * failure risk. Hurry up!
2739                                  */
2740                                 if (!zone_watermark_ok_safe(zone, order,
2741                                             min_wmark_pages(zone), end_zone, 0))
2742                                         has_under_min_watermark_zone = 1;
2743                         } else {
2744                                 /*
2745                                  * If a zone reaches its high watermark,
2746                                  * consider it to be no longer congested. It's
2747                                  * possible there are dirty pages backed by
2748                                  * congested BDIs but as pressure is relieved,
2749                                  * speculatively avoid congestion waits
2750                                  */
2751                                 zone_clear_flag(zone, ZONE_CONGESTED);
2752                                 if (i <= *classzone_idx)
2753                                         balanced += zone->present_pages;
2754                         }
2755
2756                 }
2757
2758                 /*
2759                  * If the low watermark is met there is no need for processes
2760                  * to be throttled on pfmemalloc_wait as they should not be
2761                  * able to safely make forward progress. Wake them
2762                  */
2763                 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
2764                                 pfmemalloc_watermark_ok(pgdat))
2765                         wake_up(&pgdat->pfmemalloc_wait);
2766
2767                 if (all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))
2768                         break;          /* kswapd: all done */
2769                 /*
2770                  * OK, kswapd is getting into trouble.  Take a nap, then take
2771                  * another pass across the zones.
2772                  */
2773                 if (total_scanned && (sc.priority < DEF_PRIORITY - 2)) {
2774                         if (has_under_min_watermark_zone)
2775                                 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2776                         else
2777                                 congestion_wait(BLK_RW_ASYNC, HZ/10);
2778                 }
2779
2780                 /*
2781                  * We do this so kswapd doesn't build up large priorities for
2782                  * example when it is freeing in parallel with allocators. It
2783                  * matches the direct reclaim path behaviour in terms of impact
2784                  * on zone->*_priority.
2785                  */
2786                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2787                         break;
2788         } while (--sc.priority >= 0);
2789 out:
2790
2791         /*
2792          * order-0: All zones must meet high watermark for a balanced node
2793          * high-order: Balanced zones must make up at least 25% of the node
2794          *             for the node to be balanced
2795          */
2796         if (!(all_zones_ok || (order && pgdat_balanced(pgdat, balanced, *classzone_idx)))) {
2797                 cond_resched();
2798
2799                 try_to_freeze();
2800
2801                 /*
2802                  * Fragmentation may mean that the system cannot be
2803                  * rebalanced for high-order allocations in all zones.
2804                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2805                  * it means the zones have been fully scanned and are still
2806                  * not balanced. For high-order allocations, there is
2807                  * little point trying all over again as kswapd may
2808                  * infinite loop.
2809                  *
2810                  * Instead, recheck all watermarks at order-0 as they
2811                  * are the most important. If watermarks are ok, kswapd will go
2812                  * back to sleep. High-order users can still perform direct
2813                  * reclaim if they wish.
2814                  */
2815                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2816                         order = sc.order = 0;
2817
2818                 goto loop_again;
2819         }
2820
2821         /*
2822          * If kswapd was reclaiming at a higher order, it has the option of
2823          * sleeping without all zones being balanced. Before it does, it must
2824          * ensure that the watermarks for order-0 on *all* zones are met and
2825          * that the congestion flags are cleared. The congestion flag must
2826          * be cleared as kswapd is the only mechanism that clears the flag
2827          * and it is potentially going to sleep here.
2828          */
2829         if (order) {
2830                 int zones_need_compaction = 1;
2831
2832                 for (i = 0; i <= end_zone; i++) {
2833                         struct zone *zone = pgdat->node_zones + i;
2834
2835                         if (!populated_zone(zone))
2836                                 continue;
2837
2838                         /* Check if the memory needs to be defragmented. */
2839                         if (zone_watermark_ok(zone, order,
2840                                     low_wmark_pages(zone), *classzone_idx, 0))
2841                                 zones_need_compaction = 0;
2842                 }
2843
2844                 if (zones_need_compaction)
2845                         compact_pgdat(pgdat, order);
2846         }
2847
2848         /*
2849          * Return the order we were reclaiming at so prepare_kswapd_sleep()
2850          * makes a decision on the order we were last reclaiming at. However,
2851          * if another caller entered the allocator slow path while kswapd
2852          * was awake, order will remain at the higher level
2853          */
2854         *classzone_idx = end_zone;
2855         return order;
2856 }
2857
2858 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2859 {
2860         long remaining = 0;
2861         DEFINE_WAIT(wait);
2862
2863         if (freezing(current) || kthread_should_stop())
2864                 return;
2865
2866         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2867
2868         /* Try to sleep for a short interval */
2869         if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
2870                 remaining = schedule_timeout(HZ/10);
2871                 finish_wait(&pgdat->kswapd_wait, &wait);
2872                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2873         }
2874
2875         /*
2876          * After a short sleep, check if it was a premature sleep. If not, then
2877          * go fully to sleep until explicitly woken up.
2878          */
2879         if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
2880                 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2881
2882                 /*
2883                  * vmstat counters are not perfectly accurate and the estimated
2884                  * value for counters such as NR_FREE_PAGES can deviate from the
2885                  * true value by nr_online_cpus * threshold. To avoid the zone
2886                  * watermarks being breached while under pressure, we reduce the
2887                  * per-cpu vmstat threshold while kswapd is awake and restore
2888                  * them before going back to sleep.
2889                  */
2890                 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2891
2892                 /*
2893                  * Compaction records what page blocks it recently failed to
2894                  * isolate pages from and skips them in the future scanning.
2895                  * When kswapd is going to sleep, it is reasonable to assume
2896                  * that pages and compaction may succeed so reset the cache.
2897                  */
2898                 reset_isolation_suitable(pgdat);
2899
2900                 if (!kthread_should_stop())
2901                         schedule();
2902
2903                 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2904         } else {
2905                 if (remaining)
2906                         count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2907                 else
2908                         count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2909         }
2910         finish_wait(&pgdat->kswapd_wait, &wait);
2911 }
2912
2913 /*
2914  * The background pageout daemon, started as a kernel thread
2915  * from the init process.
2916  *
2917  * This basically trickles out pages so that we have _some_
2918  * free memory available even if there is no other activity
2919  * that frees anything up. This is needed for things like routing
2920  * etc, where we otherwise might have all activity going on in
2921  * asynchronous contexts that cannot page things out.
2922  *
2923  * If there are applications that are active memory-allocators
2924  * (most normal use), this basically shouldn't matter.
2925  */
2926 static int kswapd(void *p)
2927 {
2928         unsigned long order, new_order;
2929         unsigned balanced_order;
2930         int classzone_idx, new_classzone_idx;
2931         int balanced_classzone_idx;
2932         pg_data_t *pgdat = (pg_data_t*)p;
2933         struct task_struct *tsk = current;
2934
2935         struct reclaim_state reclaim_state = {
2936                 .reclaimed_slab = 0,
2937         };
2938         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2939
2940         lockdep_set_current_reclaim_state(GFP_KERNEL);
2941
2942         if (!cpumask_empty(cpumask))
2943                 set_cpus_allowed_ptr(tsk, cpumask);
2944         current->reclaim_state = &reclaim_state;
2945
2946         /*
2947          * Tell the memory management that we're a "memory allocator",
2948          * and that if we need more memory we should get access to it
2949          * regardless (see "__alloc_pages()"). "kswapd" should
2950          * never get caught in the normal page freeing logic.
2951          *
2952          * (Kswapd normally doesn't need memory anyway, but sometimes
2953          * you need a small amount of memory in order to be able to
2954          * page out something else, and this flag essentially protects
2955          * us from recursively trying to free more memory as we're
2956          * trying to free the first piece of memory in the first place).
2957          */
2958         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2959         set_freezable();
2960
2961         order = new_order = 0;
2962         balanced_order = 0;
2963         classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2964         balanced_classzone_idx = classzone_idx;
2965         for ( ; ; ) {
2966                 bool ret;
2967
2968                 /*
2969                  * If the last balance_pgdat was unsuccessful it's unlikely a
2970                  * new request of a similar or harder type will succeed soon
2971                  * so consider going to sleep on the basis we reclaimed at
2972                  */
2973                 if (balanced_classzone_idx >= new_classzone_idx &&
2974                                         balanced_order == new_order) {
2975                         new_order = pgdat->kswapd_max_order;
2976                         new_classzone_idx = pgdat->classzone_idx;
2977                         pgdat->kswapd_max_order =  0;
2978                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2979                 }
2980
2981                 if (order < new_order || classzone_idx > new_classzone_idx) {
2982                         /*
2983                          * Don't sleep if someone wants a larger 'order'
2984                          * allocation or has tigher zone constraints
2985                          */
2986                         order = new_order;
2987                         classzone_idx = new_classzone_idx;
2988                 } else {
2989                         kswapd_try_to_sleep(pgdat, balanced_order,
2990                                                 balanced_classzone_idx);
2991                         order = pgdat->kswapd_max_order;
2992                         classzone_idx = pgdat->classzone_idx;
2993                         new_order = order;
2994                         new_classzone_idx = classzone_idx;
2995                         pgdat->kswapd_max_order = 0;
2996                         pgdat->classzone_idx = pgdat->nr_zones - 1;
2997                 }
2998
2999                 ret = try_to_freeze();
3000                 if (kthread_should_stop())
3001                         break;
3002
3003                 /*
3004                  * We can speed up thawing tasks if we don't call balance_pgdat
3005                  * after returning from the refrigerator
3006                  */
3007                 if (!ret) {
3008                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3009                         balanced_classzone_idx = classzone_idx;
3010                         balanced_order = balance_pgdat(pgdat, order,
3011                                                 &balanced_classzone_idx);
3012                 }
3013         }
3014
3015         current->reclaim_state = NULL;
3016         return 0;
3017 }
3018
3019 /*
3020  * A zone is low on free memory, so wake its kswapd task to service it.
3021  */
3022 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3023 {
3024         pg_data_t *pgdat;
3025
3026         if (!populated_zone(zone))
3027                 return;
3028
3029         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3030                 return;
3031         pgdat = zone->zone_pgdat;
3032         if (pgdat->kswapd_max_order < order) {
3033                 pgdat->kswapd_max_order = order;
3034                 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3035         }
3036         if (!waitqueue_active(&pgdat->kswapd_wait))
3037                 return;
3038         if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
3039                 return;
3040
3041         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3042         wake_up_interruptible(&pgdat->kswapd_wait);
3043 }
3044
3045 /*
3046  * The reclaimable count would be mostly accurate.
3047  * The less reclaimable pages may be
3048  * - mlocked pages, which will be moved to unevictable list when encountered
3049  * - mapped pages, which may require several travels to be reclaimed
3050  * - dirty pages, which is not "instantly" reclaimable
3051  */
3052 unsigned long global_reclaimable_pages(void)
3053 {
3054         int nr;
3055
3056         nr = global_page_state(NR_ACTIVE_FILE) +
3057              global_page_state(NR_INACTIVE_FILE);
3058
3059         if (nr_swap_pages > 0)
3060                 nr += global_page_state(NR_ACTIVE_ANON) +
3061                       global_page_state(NR_INACTIVE_ANON);
3062
3063         return nr;
3064 }
3065
3066 unsigned long zone_reclaimable_pages(struct zone *zone)
3067 {
3068         int nr;
3069
3070         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3071              zone_page_state(zone, NR_INACTIVE_FILE);
3072
3073         if (nr_swap_pages > 0)
3074                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3075                       zone_page_state(zone, NR_INACTIVE_ANON);
3076
3077         return nr;
3078 }
3079
3080 #ifdef CONFIG_HIBERNATION
3081 /*
3082  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3083  * freed pages.
3084  *
3085  * Rather than trying to age LRUs the aim is to preserve the overall
3086  * LRU order by reclaiming preferentially
3087  * inactive > active > active referenced > active mapped
3088  */
3089 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3090 {
3091         struct reclaim_state reclaim_state;
3092         struct scan_control sc = {
3093                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3094                 .may_swap = 1,
3095                 .may_unmap = 1,
3096                 .may_writepage = 1,
3097                 .nr_to_reclaim = nr_to_reclaim,
3098                 .hibernation_mode = 1,
3099                 .order = 0,
3100                 .priority = DEF_PRIORITY,
3101         };
3102         struct shrink_control shrink = {
3103                 .gfp_mask = sc.gfp_mask,
3104         };
3105         struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3106         struct task_struct *p = current;
3107         unsigned long nr_reclaimed;
3108
3109         p->flags |= PF_MEMALLOC;
3110         lockdep_set_current_reclaim_state(sc.gfp_mask);
3111         reclaim_state.reclaimed_slab = 0;
3112         p->reclaim_state = &reclaim_state;
3113
3114         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3115
3116         p->reclaim_state = NULL;
3117         lockdep_clear_current_reclaim_state();
3118         p->flags &= ~PF_MEMALLOC;
3119
3120         return nr_reclaimed;
3121 }
3122 #endif /* CONFIG_HIBERNATION */
3123
3124 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3125    not required for correctness.  So if the last cpu in a node goes
3126    away, we get changed to run anywhere: as the first one comes back,
3127    restore their cpu bindings. */
3128 static int __devinit cpu_callback(struct notifier_block *nfb,
3129                                   unsigned long action, void *hcpu)
3130 {
3131         int nid;
3132
3133         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3134                 for_each_node_state(nid, N_HIGH_MEMORY) {
3135                         pg_data_t *pgdat = NODE_DATA(nid);
3136                         const struct cpumask *mask;
3137
3138                         mask = cpumask_of_node(pgdat->node_id);
3139
3140                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3141                                 /* One of our CPUs online: restore mask */
3142                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3143                 }
3144         }
3145         return NOTIFY_OK;
3146 }
3147
3148 /*
3149  * This kswapd start function will be called by init and node-hot-add.
3150  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3151  */
3152 int kswapd_run(int nid)
3153 {
3154         pg_data_t *pgdat = NODE_DATA(nid);
3155         int ret = 0;
3156
3157         if (pgdat->kswapd)
3158                 return 0;
3159
3160         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3161         if (IS_ERR(pgdat->kswapd)) {
3162                 /* failure at boot is fatal */
3163                 BUG_ON(system_state == SYSTEM_BOOTING);
3164                 pgdat->kswapd = NULL;
3165                 pr_err("Failed to start kswapd on node %d\n", nid);
3166                 ret = PTR_ERR(pgdat->kswapd);
3167         }
3168         return ret;
3169 }
3170
3171 /*
3172  * Called by memory hotplug when all memory in a node is offlined.  Caller must
3173  * hold lock_memory_hotplug().
3174  */
3175 void kswapd_stop(int nid)
3176 {
3177         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3178
3179         if (kswapd) {
3180                 kthread_stop(kswapd);
3181                 NODE_DATA(nid)->kswapd = NULL;
3182         }
3183 }
3184
3185 static int __init kswapd_init(void)
3186 {
3187         int nid;
3188
3189         swap_setup();
3190         for_each_node_state(nid, N_HIGH_MEMORY)
3191                 kswapd_run(nid);
3192         hotcpu_notifier(cpu_callback, 0);
3193         return 0;
3194 }
3195
3196 module_init(kswapd_init)
3197
3198 #ifdef CONFIG_NUMA
3199 /*
3200  * Zone reclaim mode
3201  *
3202  * If non-zero call zone_reclaim when the number of free pages falls below
3203  * the watermarks.
3204  */
3205 int zone_reclaim_mode __read_mostly;
3206
3207 #define RECLAIM_OFF 0
3208 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
3209 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
3210 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
3211
3212 /*
3213  * Priority for ZONE_RECLAIM. This determines the fraction of pages
3214  * of a node considered for each zone_reclaim. 4 scans 1/16th of
3215  * a zone.
3216  */
3217 #define ZONE_RECLAIM_PRIORITY 4
3218
3219 /*
3220  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3221  * occur.
3222  */
3223 int sysctl_min_unmapped_ratio = 1;
3224
3225 /*
3226  * If the number of slab pages in a zone grows beyond this percentage then
3227  * slab reclaim needs to occur.
3228  */
3229 int sysctl_min_slab_ratio = 5;
3230
3231 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3232 {
3233         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3234         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3235                 zone_page_state(zone, NR_ACTIVE_FILE);
3236
3237         /*
3238          * It's possible for there to be more file mapped pages than
3239          * accounted for by the pages on the file LRU lists because
3240          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3241          */
3242         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3243 }
3244
3245 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3246 static long zone_pagecache_reclaimable(struct zone *zone)
3247 {
3248         long nr_pagecache_reclaimable;
3249         long delta = 0;
3250
3251         /*
3252          * If RECLAIM_SWAP is set, then all file pages are considered
3253          * potentially reclaimable. Otherwise, we have to worry about
3254          * pages like swapcache and zone_unmapped_file_pages() provides
3255          * a better estimate
3256          */
3257         if (zone_reclaim_mode & RECLAIM_SWAP)
3258                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3259         else
3260                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3261
3262         /* If we can't clean pages, remove dirty pages from consideration */
3263         if (!(zone_reclaim_mode & RECLAIM_WRITE))
3264                 delta += zone_page_state(zone, NR_FILE_DIRTY);
3265
3266         /* Watch for any possible underflows due to delta */
3267         if (unlikely(delta > nr_pagecache_reclaimable))
3268                 delta = nr_pagecache_reclaimable;
3269
3270         return nr_pagecache_reclaimable - delta;
3271 }
3272
3273 /*
3274  * Try to free up some pages from this zone through reclaim.
3275  */
3276 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3277 {
3278         /* Minimum pages needed in order to stay on node */
3279         const unsigned long nr_pages = 1 << order;
3280         struct task_struct *p = current;
3281         struct reclaim_state reclaim_state;
3282         struct scan_control sc = {
3283                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3284                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3285                 .may_swap = 1,
3286                 .nr_to_reclaim = max_t(unsigned long, nr_pages,
3287                                        SWAP_CLUSTER_MAX),
3288                 .gfp_mask = gfp_mask,
3289                 .order = order,
3290                 .priority = ZONE_RECLAIM_PRIORITY,
3291         };
3292         struct shrink_control shrink = {
3293                 .gfp_mask = sc.gfp_mask,
3294         };
3295         unsigned long nr_slab_pages0, nr_slab_pages1;
3296
3297         cond_resched();
3298         /*
3299          * We need to be able to allocate from the reserves for RECLAIM_SWAP
3300          * and we also need to be able to write out pages for RECLAIM_WRITE
3301          * and RECLAIM_SWAP.
3302          */
3303         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3304         lockdep_set_current_reclaim_state(gfp_mask);
3305         reclaim_state.reclaimed_slab = 0;
3306         p->reclaim_state = &reclaim_state;
3307
3308         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3309                 /*
3310                  * Free memory by calling shrink zone with increasing
3311                  * priorities until we have enough memory freed.
3312                  */
3313                 do {
3314                         shrink_zone(zone, &sc);
3315                 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3316         }
3317
3318         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3319         if (nr_slab_pages0 > zone->min_slab_pages) {
3320                 /*
3321                  * shrink_slab() does not currently allow us to determine how
3322                  * many pages were freed in this zone. So we take the current
3323                  * number of slab pages and shake the slab until it is reduced
3324                  * by the same nr_pages that we used for reclaiming unmapped
3325                  * pages.
3326                  *
3327                  * Note that shrink_slab will free memory on all zones and may
3328                  * take a long time.
3329                  */
3330                 for (;;) {
3331                         unsigned long lru_pages = zone_reclaimable_pages(zone);
3332
3333                         /* No reclaimable slab or very low memory pressure */
3334                         if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3335                                 break;
3336
3337                         /* Freed enough memory */
3338                         nr_slab_pages1 = zone_page_state(zone,
3339                                                         NR_SLAB_RECLAIMABLE);
3340                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3341                                 break;
3342                 }
3343
3344                 /*
3345                  * Update nr_reclaimed by the number of slab pages we
3346                  * reclaimed from this zone.
3347                  */
3348                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3349                 if (nr_slab_pages1 < nr_slab_pages0)
3350                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3351         }
3352
3353         p->reclaim_state = NULL;
3354         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3355         lockdep_clear_current_reclaim_state();
3356         return sc.nr_reclaimed >= nr_pages;
3357 }
3358
3359 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3360 {
3361         int node_id;
3362         int ret;
3363
3364         /*
3365          * Zone reclaim reclaims unmapped file backed pages and
3366          * slab pages if we are over the defined limits.
3367          *
3368          * A small portion of unmapped file backed pages is needed for
3369          * file I/O otherwise pages read by file I/O will be immediately
3370          * thrown out if the zone is overallocated. So we do not reclaim
3371          * if less than a specified percentage of the zone is used by
3372          * unmapped file backed pages.
3373          */
3374         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3375             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3376                 return ZONE_RECLAIM_FULL;
3377
3378         if (zone->all_unreclaimable)
3379                 return ZONE_RECLAIM_FULL;
3380
3381         /*
3382          * Do not scan if the allocation should not be delayed.
3383          */
3384         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3385                 return ZONE_RECLAIM_NOSCAN;
3386
3387         /*
3388          * Only run zone reclaim on the local zone or on zones that do not
3389          * have associated processors. This will favor the local processor
3390          * over remote processors and spread off node memory allocations
3391          * as wide as possible.
3392          */
3393         node_id = zone_to_nid(zone);
3394         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3395                 return ZONE_RECLAIM_NOSCAN;
3396
3397         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3398                 return ZONE_RECLAIM_NOSCAN;
3399
3400         ret = __zone_reclaim(zone, gfp_mask, order);
3401         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3402
3403         if (!ret)
3404                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3405
3406         return ret;
3407 }
3408 #endif
3409
3410 /*
3411  * page_evictable - test whether a page is evictable
3412  * @page: the page to test
3413  *
3414  * Test whether page is evictable--i.e., should be placed on active/inactive
3415  * lists vs unevictable list.
3416  *
3417  * Reasons page might not be evictable:
3418  * (1) page's mapping marked unevictable
3419  * (2) page is part of an mlocked VMA
3420  *
3421  */
3422 int page_evictable(struct page *page)
3423 {
3424         return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3425 }
3426
3427 #ifdef CONFIG_SHMEM
3428 /**
3429  * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3430  * @pages:      array of pages to check
3431  * @nr_pages:   number of pages to check
3432  *
3433  * Checks pages for evictability and moves them to the appropriate lru list.
3434  *
3435  * This function is only used for SysV IPC SHM_UNLOCK.
3436  */
3437 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3438 {
3439         struct lruvec *lruvec;
3440         struct zone *zone = NULL;
3441         int pgscanned = 0;
3442         int pgrescued = 0;
3443         int i;
3444
3445         for (i = 0; i < nr_pages; i++) {
3446                 struct page *page = pages[i];
3447                 struct zone *pagezone;
3448
3449                 pgscanned++;
3450                 pagezone = page_zone(page);
3451                 if (pagezone != zone) {
3452                         if (zone)
3453                                 spin_unlock_irq(&zone->lru_lock);
3454                         zone = pagezone;
3455                         spin_lock_irq(&zone->lru_lock);
3456                 }
3457                 lruvec = mem_cgroup_page_lruvec(page, zone);
3458
3459                 if (!PageLRU(page) || !PageUnevictable(page))
3460                         continue;
3461
3462                 if (page_evictable(page)) {
3463                         enum lru_list lru = page_lru_base_type(page);
3464
3465                         VM_BUG_ON(PageActive(page));
3466                         ClearPageUnevictable(page);
3467                         del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3468                         add_page_to_lru_list(page, lruvec, lru);
3469                         pgrescued++;
3470                 }
3471         }
3472
3473         if (zone) {
3474                 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3475                 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3476                 spin_unlock_irq(&zone->lru_lock);
3477         }
3478 }
3479 #endif /* CONFIG_SHMEM */
3480
3481 static void warn_scan_unevictable_pages(void)
3482 {
3483         printk_once(KERN_WARNING
3484                     "%s: The scan_unevictable_pages sysctl/node-interface has been "
3485                     "disabled for lack of a legitimate use case.  If you have "
3486                     "one, please send an email to linux-mm@kvack.org.\n",
3487                     current->comm);
3488 }
3489
3490 /*
3491  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3492  * all nodes' unevictable lists for evictable pages
3493  */
3494 unsigned long scan_unevictable_pages;
3495
3496 int scan_unevictable_handler(struct ctl_table *table, int write,
3497                            void __user *buffer,
3498                            size_t *length, loff_t *ppos)
3499 {
3500         warn_scan_unevictable_pages();
3501         proc_doulongvec_minmax(table, write, buffer, length, ppos);
3502         scan_unevictable_pages = 0;
3503         return 0;
3504 }
3505
3506 #ifdef CONFIG_NUMA
3507 /*
3508  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3509  * a specified node's per zone unevictable lists for evictable pages.
3510  */
3511
3512 static ssize_t read_scan_unevictable_node(struct device *dev,
3513                                           struct device_attribute *attr,
3514                                           char *buf)
3515 {
3516         warn_scan_unevictable_pages();
3517         return sprintf(buf, "0\n");     /* always zero; should fit... */
3518 }
3519
3520 static ssize_t write_scan_unevictable_node(struct device *dev,
3521                                            struct device_attribute *attr,
3522                                         const char *buf, size_t count)
3523 {
3524         warn_scan_unevictable_pages();
3525         return 1;
3526 }
3527
3528
3529 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3530                         read_scan_unevictable_node,
3531                         write_scan_unevictable_node);
3532
3533 int scan_unevictable_register_node(struct node *node)
3534 {
3535         return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3536 }
3537
3538 void scan_unevictable_unregister_node(struct node *node)
3539 {
3540         device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
3541 }
3542 #endif