Merge tag 'md/3.14' of git://neil.brown.name/md
[platform/adaptation/renesas_rcar/renesas_kernel.git] / mm / vmscan.c
1 /*
2  *  linux/mm/vmscan.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *
6  *  Swap reorganised 29.12.95, Stephen Tweedie.
7  *  kswapd added: 7.1.96  sct
8  *  Removed kswapd_ctl limits, and swap out as many pages as needed
9  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11  *  Multiqueue VM started 5.8.00, Rik van Riel.
12  */
13
14 #include <linux/mm.h>
15 #include <linux/module.h>
16 #include <linux/gfp.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmpressure.h>
23 #include <linux/vmstat.h>
24 #include <linux/file.h>
25 #include <linux/writeback.h>
26 #include <linux/blkdev.h>
27 #include <linux/buffer_head.h>  /* for try_to_release_page(),
28                                         buffer_heads_over_limit */
29 #include <linux/mm_inline.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/compaction.h>
36 #include <linux/notifier.h>
37 #include <linux/rwsem.h>
38 #include <linux/delay.h>
39 #include <linux/kthread.h>
40 #include <linux/freezer.h>
41 #include <linux/memcontrol.h>
42 #include <linux/delayacct.h>
43 #include <linux/sysctl.h>
44 #include <linux/oom.h>
45 #include <linux/prefetch.h>
46
47 #include <asm/tlbflush.h>
48 #include <asm/div64.h>
49
50 #include <linux/swapops.h>
51 #include <linux/balloon_compaction.h>
52
53 #include "internal.h"
54
55 #define CREATE_TRACE_POINTS
56 #include <trace/events/vmscan.h>
57
58 struct scan_control {
59         /* Incremented by the number of inactive pages that were scanned */
60         unsigned long nr_scanned;
61
62         /* Number of pages freed so far during a call to shrink_zones() */
63         unsigned long nr_reclaimed;
64
65         /* How many pages shrink_list() should reclaim */
66         unsigned long nr_to_reclaim;
67
68         unsigned long hibernation_mode;
69
70         /* This context's GFP mask */
71         gfp_t gfp_mask;
72
73         int may_writepage;
74
75         /* Can mapped pages be reclaimed? */
76         int may_unmap;
77
78         /* Can pages be swapped as part of reclaim? */
79         int may_swap;
80
81         int order;
82
83         /* Scan (total_size >> priority) pages at once */
84         int priority;
85
86         /*
87          * The memory cgroup that hit its limit and as a result is the
88          * primary target of this reclaim invocation.
89          */
90         struct mem_cgroup *target_mem_cgroup;
91
92         /*
93          * Nodemask of nodes allowed by the caller. If NULL, all nodes
94          * are scanned.
95          */
96         nodemask_t      *nodemask;
97 };
98
99 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
100
101 #ifdef ARCH_HAS_PREFETCH
102 #define prefetch_prev_lru_page(_page, _base, _field)                    \
103         do {                                                            \
104                 if ((_page)->lru.prev != _base) {                       \
105                         struct page *prev;                              \
106                                                                         \
107                         prev = lru_to_page(&(_page->lru));              \
108                         prefetch(&prev->_field);                        \
109                 }                                                       \
110         } while (0)
111 #else
112 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
113 #endif
114
115 #ifdef ARCH_HAS_PREFETCHW
116 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
117         do {                                                            \
118                 if ((_page)->lru.prev != _base) {                       \
119                         struct page *prev;                              \
120                                                                         \
121                         prev = lru_to_page(&(_page->lru));              \
122                         prefetchw(&prev->_field);                       \
123                 }                                                       \
124         } while (0)
125 #else
126 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
127 #endif
128
129 /*
130  * From 0 .. 100.  Higher means more swappy.
131  */
132 int vm_swappiness = 60;
133 unsigned long vm_total_pages;   /* The total number of pages which the VM controls */
134
135 static LIST_HEAD(shrinker_list);
136 static DECLARE_RWSEM(shrinker_rwsem);
137
138 #ifdef CONFIG_MEMCG
139 static bool global_reclaim(struct scan_control *sc)
140 {
141         return !sc->target_mem_cgroup;
142 }
143 #else
144 static bool global_reclaim(struct scan_control *sc)
145 {
146         return true;
147 }
148 #endif
149
150 unsigned long zone_reclaimable_pages(struct zone *zone)
151 {
152         int nr;
153
154         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
155              zone_page_state(zone, NR_INACTIVE_FILE);
156
157         if (get_nr_swap_pages() > 0)
158                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
159                       zone_page_state(zone, NR_INACTIVE_ANON);
160
161         return nr;
162 }
163
164 bool zone_reclaimable(struct zone *zone)
165 {
166         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
167 }
168
169 static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
170 {
171         if (!mem_cgroup_disabled())
172                 return mem_cgroup_get_lru_size(lruvec, lru);
173
174         return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
175 }
176
177 /*
178  * Add a shrinker callback to be called from the vm.
179  */
180 int register_shrinker(struct shrinker *shrinker)
181 {
182         size_t size = sizeof(*shrinker->nr_deferred);
183
184         /*
185          * If we only have one possible node in the system anyway, save
186          * ourselves the trouble and disable NUMA aware behavior. This way we
187          * will save memory and some small loop time later.
188          */
189         if (nr_node_ids == 1)
190                 shrinker->flags &= ~SHRINKER_NUMA_AWARE;
191
192         if (shrinker->flags & SHRINKER_NUMA_AWARE)
193                 size *= nr_node_ids;
194
195         shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
196         if (!shrinker->nr_deferred)
197                 return -ENOMEM;
198
199         down_write(&shrinker_rwsem);
200         list_add_tail(&shrinker->list, &shrinker_list);
201         up_write(&shrinker_rwsem);
202         return 0;
203 }
204 EXPORT_SYMBOL(register_shrinker);
205
206 /*
207  * Remove one
208  */
209 void unregister_shrinker(struct shrinker *shrinker)
210 {
211         down_write(&shrinker_rwsem);
212         list_del(&shrinker->list);
213         up_write(&shrinker_rwsem);
214         kfree(shrinker->nr_deferred);
215 }
216 EXPORT_SYMBOL(unregister_shrinker);
217
218 #define SHRINK_BATCH 128
219
220 static unsigned long
221 shrink_slab_node(struct shrink_control *shrinkctl, struct shrinker *shrinker,
222                  unsigned long nr_pages_scanned, unsigned long lru_pages)
223 {
224         unsigned long freed = 0;
225         unsigned long long delta;
226         long total_scan;
227         long max_pass;
228         long nr;
229         long new_nr;
230         int nid = shrinkctl->nid;
231         long batch_size = shrinker->batch ? shrinker->batch
232                                           : SHRINK_BATCH;
233
234         max_pass = shrinker->count_objects(shrinker, shrinkctl);
235         if (max_pass == 0)
236                 return 0;
237
238         /*
239          * copy the current shrinker scan count into a local variable
240          * and zero it so that other concurrent shrinker invocations
241          * don't also do this scanning work.
242          */
243         nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
244
245         total_scan = nr;
246         delta = (4 * nr_pages_scanned) / shrinker->seeks;
247         delta *= max_pass;
248         do_div(delta, lru_pages + 1);
249         total_scan += delta;
250         if (total_scan < 0) {
251                 printk(KERN_ERR
252                 "shrink_slab: %pF negative objects to delete nr=%ld\n",
253                        shrinker->scan_objects, total_scan);
254                 total_scan = max_pass;
255         }
256
257         /*
258          * We need to avoid excessive windup on filesystem shrinkers
259          * due to large numbers of GFP_NOFS allocations causing the
260          * shrinkers to return -1 all the time. This results in a large
261          * nr being built up so when a shrink that can do some work
262          * comes along it empties the entire cache due to nr >>>
263          * max_pass.  This is bad for sustaining a working set in
264          * memory.
265          *
266          * Hence only allow the shrinker to scan the entire cache when
267          * a large delta change is calculated directly.
268          */
269         if (delta < max_pass / 4)
270                 total_scan = min(total_scan, max_pass / 2);
271
272         /*
273          * Avoid risking looping forever due to too large nr value:
274          * never try to free more than twice the estimate number of
275          * freeable entries.
276          */
277         if (total_scan > max_pass * 2)
278                 total_scan = max_pass * 2;
279
280         trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
281                                 nr_pages_scanned, lru_pages,
282                                 max_pass, delta, total_scan);
283
284         /*
285          * Normally, we should not scan less than batch_size objects in one
286          * pass to avoid too frequent shrinker calls, but if the slab has less
287          * than batch_size objects in total and we are really tight on memory,
288          * we will try to reclaim all available objects, otherwise we can end
289          * up failing allocations although there are plenty of reclaimable
290          * objects spread over several slabs with usage less than the
291          * batch_size.
292          *
293          * We detect the "tight on memory" situations by looking at the total
294          * number of objects we want to scan (total_scan). If it is greater
295          * than the total number of objects on slab (max_pass), we must be
296          * scanning at high prio and therefore should try to reclaim as much as
297          * possible.
298          */
299         while (total_scan >= batch_size ||
300                total_scan >= max_pass) {
301                 unsigned long ret;
302                 unsigned long nr_to_scan = min(batch_size, total_scan);
303
304                 shrinkctl->nr_to_scan = nr_to_scan;
305                 ret = shrinker->scan_objects(shrinker, shrinkctl);
306                 if (ret == SHRINK_STOP)
307                         break;
308                 freed += ret;
309
310                 count_vm_events(SLABS_SCANNED, nr_to_scan);
311                 total_scan -= nr_to_scan;
312
313                 cond_resched();
314         }
315
316         /*
317          * move the unused scan count back into the shrinker in a
318          * manner that handles concurrent updates. If we exhausted the
319          * scan, there is no need to do an update.
320          */
321         if (total_scan > 0)
322                 new_nr = atomic_long_add_return(total_scan,
323                                                 &shrinker->nr_deferred[nid]);
324         else
325                 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
326
327         trace_mm_shrink_slab_end(shrinker, freed, nr, new_nr);
328         return freed;
329 }
330
331 /*
332  * Call the shrink functions to age shrinkable caches
333  *
334  * Here we assume it costs one seek to replace a lru page and that it also
335  * takes a seek to recreate a cache object.  With this in mind we age equal
336  * percentages of the lru and ageable caches.  This should balance the seeks
337  * generated by these structures.
338  *
339  * If the vm encountered mapped pages on the LRU it increase the pressure on
340  * slab to avoid swapping.
341  *
342  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
343  *
344  * `lru_pages' represents the number of on-LRU pages in all the zones which
345  * are eligible for the caller's allocation attempt.  It is used for balancing
346  * slab reclaim versus page reclaim.
347  *
348  * Returns the number of slab objects which we shrunk.
349  */
350 unsigned long shrink_slab(struct shrink_control *shrinkctl,
351                           unsigned long nr_pages_scanned,
352                           unsigned long lru_pages)
353 {
354         struct shrinker *shrinker;
355         unsigned long freed = 0;
356
357         if (nr_pages_scanned == 0)
358                 nr_pages_scanned = SWAP_CLUSTER_MAX;
359
360         if (!down_read_trylock(&shrinker_rwsem)) {
361                 /*
362                  * If we would return 0, our callers would understand that we
363                  * have nothing else to shrink and give up trying. By returning
364                  * 1 we keep it going and assume we'll be able to shrink next
365                  * time.
366                  */
367                 freed = 1;
368                 goto out;
369         }
370
371         list_for_each_entry(shrinker, &shrinker_list, list) {
372                 if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) {
373                         shrinkctl->nid = 0;
374                         freed += shrink_slab_node(shrinkctl, shrinker,
375                                         nr_pages_scanned, lru_pages);
376                         continue;
377                 }
378
379                 for_each_node_mask(shrinkctl->nid, shrinkctl->nodes_to_scan) {
380                         if (node_online(shrinkctl->nid))
381                                 freed += shrink_slab_node(shrinkctl, shrinker,
382                                                 nr_pages_scanned, lru_pages);
383
384                 }
385         }
386         up_read(&shrinker_rwsem);
387 out:
388         cond_resched();
389         return freed;
390 }
391
392 static inline int is_page_cache_freeable(struct page *page)
393 {
394         /*
395          * A freeable page cache page is referenced only by the caller
396          * that isolated the page, the page cache radix tree and
397          * optional buffer heads at page->private.
398          */
399         return page_count(page) - page_has_private(page) == 2;
400 }
401
402 static int may_write_to_queue(struct backing_dev_info *bdi,
403                               struct scan_control *sc)
404 {
405         if (current->flags & PF_SWAPWRITE)
406                 return 1;
407         if (!bdi_write_congested(bdi))
408                 return 1;
409         if (bdi == current->backing_dev_info)
410                 return 1;
411         return 0;
412 }
413
414 /*
415  * We detected a synchronous write error writing a page out.  Probably
416  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
417  * fsync(), msync() or close().
418  *
419  * The tricky part is that after writepage we cannot touch the mapping: nothing
420  * prevents it from being freed up.  But we have a ref on the page and once
421  * that page is locked, the mapping is pinned.
422  *
423  * We're allowed to run sleeping lock_page() here because we know the caller has
424  * __GFP_FS.
425  */
426 static void handle_write_error(struct address_space *mapping,
427                                 struct page *page, int error)
428 {
429         lock_page(page);
430         if (page_mapping(page) == mapping)
431                 mapping_set_error(mapping, error);
432         unlock_page(page);
433 }
434
435 /* possible outcome of pageout() */
436 typedef enum {
437         /* failed to write page out, page is locked */
438         PAGE_KEEP,
439         /* move page to the active list, page is locked */
440         PAGE_ACTIVATE,
441         /* page has been sent to the disk successfully, page is unlocked */
442         PAGE_SUCCESS,
443         /* page is clean and locked */
444         PAGE_CLEAN,
445 } pageout_t;
446
447 /*
448  * pageout is called by shrink_page_list() for each dirty page.
449  * Calls ->writepage().
450  */
451 static pageout_t pageout(struct page *page, struct address_space *mapping,
452                          struct scan_control *sc)
453 {
454         /*
455          * If the page is dirty, only perform writeback if that write
456          * will be non-blocking.  To prevent this allocation from being
457          * stalled by pagecache activity.  But note that there may be
458          * stalls if we need to run get_block().  We could test
459          * PagePrivate for that.
460          *
461          * If this process is currently in __generic_file_aio_write() against
462          * this page's queue, we can perform writeback even if that
463          * will block.
464          *
465          * If the page is swapcache, write it back even if that would
466          * block, for some throttling. This happens by accident, because
467          * swap_backing_dev_info is bust: it doesn't reflect the
468          * congestion state of the swapdevs.  Easy to fix, if needed.
469          */
470         if (!is_page_cache_freeable(page))
471                 return PAGE_KEEP;
472         if (!mapping) {
473                 /*
474                  * Some data journaling orphaned pages can have
475                  * page->mapping == NULL while being dirty with clean buffers.
476                  */
477                 if (page_has_private(page)) {
478                         if (try_to_free_buffers(page)) {
479                                 ClearPageDirty(page);
480                                 printk("%s: orphaned page\n", __func__);
481                                 return PAGE_CLEAN;
482                         }
483                 }
484                 return PAGE_KEEP;
485         }
486         if (mapping->a_ops->writepage == NULL)
487                 return PAGE_ACTIVATE;
488         if (!may_write_to_queue(mapping->backing_dev_info, sc))
489                 return PAGE_KEEP;
490
491         if (clear_page_dirty_for_io(page)) {
492                 int res;
493                 struct writeback_control wbc = {
494                         .sync_mode = WB_SYNC_NONE,
495                         .nr_to_write = SWAP_CLUSTER_MAX,
496                         .range_start = 0,
497                         .range_end = LLONG_MAX,
498                         .for_reclaim = 1,
499                 };
500
501                 SetPageReclaim(page);
502                 res = mapping->a_ops->writepage(page, &wbc);
503                 if (res < 0)
504                         handle_write_error(mapping, page, res);
505                 if (res == AOP_WRITEPAGE_ACTIVATE) {
506                         ClearPageReclaim(page);
507                         return PAGE_ACTIVATE;
508                 }
509
510                 if (!PageWriteback(page)) {
511                         /* synchronous write or broken a_ops? */
512                         ClearPageReclaim(page);
513                 }
514                 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
515                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
516                 return PAGE_SUCCESS;
517         }
518
519         return PAGE_CLEAN;
520 }
521
522 /*
523  * Same as remove_mapping, but if the page is removed from the mapping, it
524  * gets returned with a refcount of 0.
525  */
526 static int __remove_mapping(struct address_space *mapping, struct page *page)
527 {
528         BUG_ON(!PageLocked(page));
529         BUG_ON(mapping != page_mapping(page));
530
531         spin_lock_irq(&mapping->tree_lock);
532         /*
533          * The non racy check for a busy page.
534          *
535          * Must be careful with the order of the tests. When someone has
536          * a ref to the page, it may be possible that they dirty it then
537          * drop the reference. So if PageDirty is tested before page_count
538          * here, then the following race may occur:
539          *
540          * get_user_pages(&page);
541          * [user mapping goes away]
542          * write_to(page);
543          *                              !PageDirty(page)    [good]
544          * SetPageDirty(page);
545          * put_page(page);
546          *                              !page_count(page)   [good, discard it]
547          *
548          * [oops, our write_to data is lost]
549          *
550          * Reversing the order of the tests ensures such a situation cannot
551          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
552          * load is not satisfied before that of page->_count.
553          *
554          * Note that if SetPageDirty is always performed via set_page_dirty,
555          * and thus under tree_lock, then this ordering is not required.
556          */
557         if (!page_freeze_refs(page, 2))
558                 goto cannot_free;
559         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
560         if (unlikely(PageDirty(page))) {
561                 page_unfreeze_refs(page, 2);
562                 goto cannot_free;
563         }
564
565         if (PageSwapCache(page)) {
566                 swp_entry_t swap = { .val = page_private(page) };
567                 __delete_from_swap_cache(page);
568                 spin_unlock_irq(&mapping->tree_lock);
569                 swapcache_free(swap, page);
570         } else {
571                 void (*freepage)(struct page *);
572
573                 freepage = mapping->a_ops->freepage;
574
575                 __delete_from_page_cache(page);
576                 spin_unlock_irq(&mapping->tree_lock);
577                 mem_cgroup_uncharge_cache_page(page);
578
579                 if (freepage != NULL)
580                         freepage(page);
581         }
582
583         return 1;
584
585 cannot_free:
586         spin_unlock_irq(&mapping->tree_lock);
587         return 0;
588 }
589
590 /*
591  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
592  * someone else has a ref on the page, abort and return 0.  If it was
593  * successfully detached, return 1.  Assumes the caller has a single ref on
594  * this page.
595  */
596 int remove_mapping(struct address_space *mapping, struct page *page)
597 {
598         if (__remove_mapping(mapping, page)) {
599                 /*
600                  * Unfreezing the refcount with 1 rather than 2 effectively
601                  * drops the pagecache ref for us without requiring another
602                  * atomic operation.
603                  */
604                 page_unfreeze_refs(page, 1);
605                 return 1;
606         }
607         return 0;
608 }
609
610 /**
611  * putback_lru_page - put previously isolated page onto appropriate LRU list
612  * @page: page to be put back to appropriate lru list
613  *
614  * Add previously isolated @page to appropriate LRU list.
615  * Page may still be unevictable for other reasons.
616  *
617  * lru_lock must not be held, interrupts must be enabled.
618  */
619 void putback_lru_page(struct page *page)
620 {
621         bool is_unevictable;
622         int was_unevictable = PageUnevictable(page);
623
624         VM_BUG_ON_PAGE(PageLRU(page), page);
625
626 redo:
627         ClearPageUnevictable(page);
628
629         if (page_evictable(page)) {
630                 /*
631                  * For evictable pages, we can use the cache.
632                  * In event of a race, worst case is we end up with an
633                  * unevictable page on [in]active list.
634                  * We know how to handle that.
635                  */
636                 is_unevictable = false;
637                 lru_cache_add(page);
638         } else {
639                 /*
640                  * Put unevictable pages directly on zone's unevictable
641                  * list.
642                  */
643                 is_unevictable = true;
644                 add_page_to_unevictable_list(page);
645                 /*
646                  * When racing with an mlock or AS_UNEVICTABLE clearing
647                  * (page is unlocked) make sure that if the other thread
648                  * does not observe our setting of PG_lru and fails
649                  * isolation/check_move_unevictable_pages,
650                  * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
651                  * the page back to the evictable list.
652                  *
653                  * The other side is TestClearPageMlocked() or shmem_lock().
654                  */
655                 smp_mb();
656         }
657
658         /*
659          * page's status can change while we move it among lru. If an evictable
660          * page is on unevictable list, it never be freed. To avoid that,
661          * check after we added it to the list, again.
662          */
663         if (is_unevictable && page_evictable(page)) {
664                 if (!isolate_lru_page(page)) {
665                         put_page(page);
666                         goto redo;
667                 }
668                 /* This means someone else dropped this page from LRU
669                  * So, it will be freed or putback to LRU again. There is
670                  * nothing to do here.
671                  */
672         }
673
674         if (was_unevictable && !is_unevictable)
675                 count_vm_event(UNEVICTABLE_PGRESCUED);
676         else if (!was_unevictable && is_unevictable)
677                 count_vm_event(UNEVICTABLE_PGCULLED);
678
679         put_page(page);         /* drop ref from isolate */
680 }
681
682 enum page_references {
683         PAGEREF_RECLAIM,
684         PAGEREF_RECLAIM_CLEAN,
685         PAGEREF_KEEP,
686         PAGEREF_ACTIVATE,
687 };
688
689 static enum page_references page_check_references(struct page *page,
690                                                   struct scan_control *sc)
691 {
692         int referenced_ptes, referenced_page;
693         unsigned long vm_flags;
694
695         referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
696                                           &vm_flags);
697         referenced_page = TestClearPageReferenced(page);
698
699         /*
700          * Mlock lost the isolation race with us.  Let try_to_unmap()
701          * move the page to the unevictable list.
702          */
703         if (vm_flags & VM_LOCKED)
704                 return PAGEREF_RECLAIM;
705
706         if (referenced_ptes) {
707                 if (PageSwapBacked(page))
708                         return PAGEREF_ACTIVATE;
709                 /*
710                  * All mapped pages start out with page table
711                  * references from the instantiating fault, so we need
712                  * to look twice if a mapped file page is used more
713                  * than once.
714                  *
715                  * Mark it and spare it for another trip around the
716                  * inactive list.  Another page table reference will
717                  * lead to its activation.
718                  *
719                  * Note: the mark is set for activated pages as well
720                  * so that recently deactivated but used pages are
721                  * quickly recovered.
722                  */
723                 SetPageReferenced(page);
724
725                 if (referenced_page || referenced_ptes > 1)
726                         return PAGEREF_ACTIVATE;
727
728                 /*
729                  * Activate file-backed executable pages after first usage.
730                  */
731                 if (vm_flags & VM_EXEC)
732                         return PAGEREF_ACTIVATE;
733
734                 return PAGEREF_KEEP;
735         }
736
737         /* Reclaim if clean, defer dirty pages to writeback */
738         if (referenced_page && !PageSwapBacked(page))
739                 return PAGEREF_RECLAIM_CLEAN;
740
741         return PAGEREF_RECLAIM;
742 }
743
744 /* Check if a page is dirty or under writeback */
745 static void page_check_dirty_writeback(struct page *page,
746                                        bool *dirty, bool *writeback)
747 {
748         struct address_space *mapping;
749
750         /*
751          * Anonymous pages are not handled by flushers and must be written
752          * from reclaim context. Do not stall reclaim based on them
753          */
754         if (!page_is_file_cache(page)) {
755                 *dirty = false;
756                 *writeback = false;
757                 return;
758         }
759
760         /* By default assume that the page flags are accurate */
761         *dirty = PageDirty(page);
762         *writeback = PageWriteback(page);
763
764         /* Verify dirty/writeback state if the filesystem supports it */
765         if (!page_has_private(page))
766                 return;
767
768         mapping = page_mapping(page);
769         if (mapping && mapping->a_ops->is_dirty_writeback)
770                 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
771 }
772
773 /*
774  * shrink_page_list() returns the number of reclaimed pages
775  */
776 static unsigned long shrink_page_list(struct list_head *page_list,
777                                       struct zone *zone,
778                                       struct scan_control *sc,
779                                       enum ttu_flags ttu_flags,
780                                       unsigned long *ret_nr_dirty,
781                                       unsigned long *ret_nr_unqueued_dirty,
782                                       unsigned long *ret_nr_congested,
783                                       unsigned long *ret_nr_writeback,
784                                       unsigned long *ret_nr_immediate,
785                                       bool force_reclaim)
786 {
787         LIST_HEAD(ret_pages);
788         LIST_HEAD(free_pages);
789         int pgactivate = 0;
790         unsigned long nr_unqueued_dirty = 0;
791         unsigned long nr_dirty = 0;
792         unsigned long nr_congested = 0;
793         unsigned long nr_reclaimed = 0;
794         unsigned long nr_writeback = 0;
795         unsigned long nr_immediate = 0;
796
797         cond_resched();
798
799         mem_cgroup_uncharge_start();
800         while (!list_empty(page_list)) {
801                 struct address_space *mapping;
802                 struct page *page;
803                 int may_enter_fs;
804                 enum page_references references = PAGEREF_RECLAIM_CLEAN;
805                 bool dirty, writeback;
806
807                 cond_resched();
808
809                 page = lru_to_page(page_list);
810                 list_del(&page->lru);
811
812                 if (!trylock_page(page))
813                         goto keep;
814
815                 VM_BUG_ON_PAGE(PageActive(page), page);
816                 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
817
818                 sc->nr_scanned++;
819
820                 if (unlikely(!page_evictable(page)))
821                         goto cull_mlocked;
822
823                 if (!sc->may_unmap && page_mapped(page))
824                         goto keep_locked;
825
826                 /* Double the slab pressure for mapped and swapcache pages */
827                 if (page_mapped(page) || PageSwapCache(page))
828                         sc->nr_scanned++;
829
830                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
831                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
832
833                 /*
834                  * The number of dirty pages determines if a zone is marked
835                  * reclaim_congested which affects wait_iff_congested. kswapd
836                  * will stall and start writing pages if the tail of the LRU
837                  * is all dirty unqueued pages.
838                  */
839                 page_check_dirty_writeback(page, &dirty, &writeback);
840                 if (dirty || writeback)
841                         nr_dirty++;
842
843                 if (dirty && !writeback)
844                         nr_unqueued_dirty++;
845
846                 /*
847                  * Treat this page as congested if the underlying BDI is or if
848                  * pages are cycling through the LRU so quickly that the
849                  * pages marked for immediate reclaim are making it to the
850                  * end of the LRU a second time.
851                  */
852                 mapping = page_mapping(page);
853                 if ((mapping && bdi_write_congested(mapping->backing_dev_info)) ||
854                     (writeback && PageReclaim(page)))
855                         nr_congested++;
856
857                 /*
858                  * If a page at the tail of the LRU is under writeback, there
859                  * are three cases to consider.
860                  *
861                  * 1) If reclaim is encountering an excessive number of pages
862                  *    under writeback and this page is both under writeback and
863                  *    PageReclaim then it indicates that pages are being queued
864                  *    for IO but are being recycled through the LRU before the
865                  *    IO can complete. Waiting on the page itself risks an
866                  *    indefinite stall if it is impossible to writeback the
867                  *    page due to IO error or disconnected storage so instead
868                  *    note that the LRU is being scanned too quickly and the
869                  *    caller can stall after page list has been processed.
870                  *
871                  * 2) Global reclaim encounters a page, memcg encounters a
872                  *    page that is not marked for immediate reclaim or
873                  *    the caller does not have __GFP_IO. In this case mark
874                  *    the page for immediate reclaim and continue scanning.
875                  *
876                  *    __GFP_IO is checked  because a loop driver thread might
877                  *    enter reclaim, and deadlock if it waits on a page for
878                  *    which it is needed to do the write (loop masks off
879                  *    __GFP_IO|__GFP_FS for this reason); but more thought
880                  *    would probably show more reasons.
881                  *
882                  *    Don't require __GFP_FS, since we're not going into the
883                  *    FS, just waiting on its writeback completion. Worryingly,
884                  *    ext4 gfs2 and xfs allocate pages with
885                  *    grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
886                  *    may_enter_fs here is liable to OOM on them.
887                  *
888                  * 3) memcg encounters a page that is not already marked
889                  *    PageReclaim. memcg does not have any dirty pages
890                  *    throttling so we could easily OOM just because too many
891                  *    pages are in writeback and there is nothing else to
892                  *    reclaim. Wait for the writeback to complete.
893                  */
894                 if (PageWriteback(page)) {
895                         /* Case 1 above */
896                         if (current_is_kswapd() &&
897                             PageReclaim(page) &&
898                             zone_is_reclaim_writeback(zone)) {
899                                 nr_immediate++;
900                                 goto keep_locked;
901
902                         /* Case 2 above */
903                         } else if (global_reclaim(sc) ||
904                             !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
905                                 /*
906                                  * This is slightly racy - end_page_writeback()
907                                  * might have just cleared PageReclaim, then
908                                  * setting PageReclaim here end up interpreted
909                                  * as PageReadahead - but that does not matter
910                                  * enough to care.  What we do want is for this
911                                  * page to have PageReclaim set next time memcg
912                                  * reclaim reaches the tests above, so it will
913                                  * then wait_on_page_writeback() to avoid OOM;
914                                  * and it's also appropriate in global reclaim.
915                                  */
916                                 SetPageReclaim(page);
917                                 nr_writeback++;
918
919                                 goto keep_locked;
920
921                         /* Case 3 above */
922                         } else {
923                                 wait_on_page_writeback(page);
924                         }
925                 }
926
927                 if (!force_reclaim)
928                         references = page_check_references(page, sc);
929
930                 switch (references) {
931                 case PAGEREF_ACTIVATE:
932                         goto activate_locked;
933                 case PAGEREF_KEEP:
934                         goto keep_locked;
935                 case PAGEREF_RECLAIM:
936                 case PAGEREF_RECLAIM_CLEAN:
937                         ; /* try to reclaim the page below */
938                 }
939
940                 /*
941                  * Anonymous process memory has backing store?
942                  * Try to allocate it some swap space here.
943                  */
944                 if (PageAnon(page) && !PageSwapCache(page)) {
945                         if (!(sc->gfp_mask & __GFP_IO))
946                                 goto keep_locked;
947                         if (!add_to_swap(page, page_list))
948                                 goto activate_locked;
949                         may_enter_fs = 1;
950
951                         /* Adding to swap updated mapping */
952                         mapping = page_mapping(page);
953                 }
954
955                 /*
956                  * The page is mapped into the page tables of one or more
957                  * processes. Try to unmap it here.
958                  */
959                 if (page_mapped(page) && mapping) {
960                         switch (try_to_unmap(page, ttu_flags)) {
961                         case SWAP_FAIL:
962                                 goto activate_locked;
963                         case SWAP_AGAIN:
964                                 goto keep_locked;
965                         case SWAP_MLOCK:
966                                 goto cull_mlocked;
967                         case SWAP_SUCCESS:
968                                 ; /* try to free the page below */
969                         }
970                 }
971
972                 if (PageDirty(page)) {
973                         /*
974                          * Only kswapd can writeback filesystem pages to
975                          * avoid risk of stack overflow but only writeback
976                          * if many dirty pages have been encountered.
977                          */
978                         if (page_is_file_cache(page) &&
979                                         (!current_is_kswapd() ||
980                                          !zone_is_reclaim_dirty(zone))) {
981                                 /*
982                                  * Immediately reclaim when written back.
983                                  * Similar in principal to deactivate_page()
984                                  * except we already have the page isolated
985                                  * and know it's dirty
986                                  */
987                                 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
988                                 SetPageReclaim(page);
989
990                                 goto keep_locked;
991                         }
992
993                         if (references == PAGEREF_RECLAIM_CLEAN)
994                                 goto keep_locked;
995                         if (!may_enter_fs)
996                                 goto keep_locked;
997                         if (!sc->may_writepage)
998                                 goto keep_locked;
999
1000                         /* Page is dirty, try to write it out here */
1001                         switch (pageout(page, mapping, sc)) {
1002                         case PAGE_KEEP:
1003                                 goto keep_locked;
1004                         case PAGE_ACTIVATE:
1005                                 goto activate_locked;
1006                         case PAGE_SUCCESS:
1007                                 if (PageWriteback(page))
1008                                         goto keep;
1009                                 if (PageDirty(page))
1010                                         goto keep;
1011
1012                                 /*
1013                                  * A synchronous write - probably a ramdisk.  Go
1014                                  * ahead and try to reclaim the page.
1015                                  */
1016                                 if (!trylock_page(page))
1017                                         goto keep;
1018                                 if (PageDirty(page) || PageWriteback(page))
1019                                         goto keep_locked;
1020                                 mapping = page_mapping(page);
1021                         case PAGE_CLEAN:
1022                                 ; /* try to free the page below */
1023                         }
1024                 }
1025
1026                 /*
1027                  * If the page has buffers, try to free the buffer mappings
1028                  * associated with this page. If we succeed we try to free
1029                  * the page as well.
1030                  *
1031                  * We do this even if the page is PageDirty().
1032                  * try_to_release_page() does not perform I/O, but it is
1033                  * possible for a page to have PageDirty set, but it is actually
1034                  * clean (all its buffers are clean).  This happens if the
1035                  * buffers were written out directly, with submit_bh(). ext3
1036                  * will do this, as well as the blockdev mapping.
1037                  * try_to_release_page() will discover that cleanness and will
1038                  * drop the buffers and mark the page clean - it can be freed.
1039                  *
1040                  * Rarely, pages can have buffers and no ->mapping.  These are
1041                  * the pages which were not successfully invalidated in
1042                  * truncate_complete_page().  We try to drop those buffers here
1043                  * and if that worked, and the page is no longer mapped into
1044                  * process address space (page_count == 1) it can be freed.
1045                  * Otherwise, leave the page on the LRU so it is swappable.
1046                  */
1047                 if (page_has_private(page)) {
1048                         if (!try_to_release_page(page, sc->gfp_mask))
1049                                 goto activate_locked;
1050                         if (!mapping && page_count(page) == 1) {
1051                                 unlock_page(page);
1052                                 if (put_page_testzero(page))
1053                                         goto free_it;
1054                                 else {
1055                                         /*
1056                                          * rare race with speculative reference.
1057                                          * the speculative reference will free
1058                                          * this page shortly, so we may
1059                                          * increment nr_reclaimed here (and
1060                                          * leave it off the LRU).
1061                                          */
1062                                         nr_reclaimed++;
1063                                         continue;
1064                                 }
1065                         }
1066                 }
1067
1068                 if (!mapping || !__remove_mapping(mapping, page))
1069                         goto keep_locked;
1070
1071                 /*
1072                  * At this point, we have no other references and there is
1073                  * no way to pick any more up (removed from LRU, removed
1074                  * from pagecache). Can use non-atomic bitops now (and
1075                  * we obviously don't have to worry about waking up a process
1076                  * waiting on the page lock, because there are no references.
1077                  */
1078                 __clear_page_locked(page);
1079 free_it:
1080                 nr_reclaimed++;
1081
1082                 /*
1083                  * Is there need to periodically free_page_list? It would
1084                  * appear not as the counts should be low
1085                  */
1086                 list_add(&page->lru, &free_pages);
1087                 continue;
1088
1089 cull_mlocked:
1090                 if (PageSwapCache(page))
1091                         try_to_free_swap(page);
1092                 unlock_page(page);
1093                 putback_lru_page(page);
1094                 continue;
1095
1096 activate_locked:
1097                 /* Not a candidate for swapping, so reclaim swap space. */
1098                 if (PageSwapCache(page) && vm_swap_full())
1099                         try_to_free_swap(page);
1100                 VM_BUG_ON_PAGE(PageActive(page), page);
1101                 SetPageActive(page);
1102                 pgactivate++;
1103 keep_locked:
1104                 unlock_page(page);
1105 keep:
1106                 list_add(&page->lru, &ret_pages);
1107                 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1108         }
1109
1110         free_hot_cold_page_list(&free_pages, 1);
1111
1112         list_splice(&ret_pages, page_list);
1113         count_vm_events(PGACTIVATE, pgactivate);
1114         mem_cgroup_uncharge_end();
1115         *ret_nr_dirty += nr_dirty;
1116         *ret_nr_congested += nr_congested;
1117         *ret_nr_unqueued_dirty += nr_unqueued_dirty;
1118         *ret_nr_writeback += nr_writeback;
1119         *ret_nr_immediate += nr_immediate;
1120         return nr_reclaimed;
1121 }
1122
1123 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1124                                             struct list_head *page_list)
1125 {
1126         struct scan_control sc = {
1127                 .gfp_mask = GFP_KERNEL,
1128                 .priority = DEF_PRIORITY,
1129                 .may_unmap = 1,
1130         };
1131         unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
1132         struct page *page, *next;
1133         LIST_HEAD(clean_pages);
1134
1135         list_for_each_entry_safe(page, next, page_list, lru) {
1136                 if (page_is_file_cache(page) && !PageDirty(page) &&
1137                     !isolated_balloon_page(page)) {
1138                         ClearPageActive(page);
1139                         list_move(&page->lru, &clean_pages);
1140                 }
1141         }
1142
1143         ret = shrink_page_list(&clean_pages, zone, &sc,
1144                         TTU_UNMAP|TTU_IGNORE_ACCESS,
1145                         &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
1146         list_splice(&clean_pages, page_list);
1147         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
1148         return ret;
1149 }
1150
1151 /*
1152  * Attempt to remove the specified page from its LRU.  Only take this page
1153  * if it is of the appropriate PageActive status.  Pages which are being
1154  * freed elsewhere are also ignored.
1155  *
1156  * page:        page to consider
1157  * mode:        one of the LRU isolation modes defined above
1158  *
1159  * returns 0 on success, -ve errno on failure.
1160  */
1161 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1162 {
1163         int ret = -EINVAL;
1164
1165         /* Only take pages on the LRU. */
1166         if (!PageLRU(page))
1167                 return ret;
1168
1169         /* Compaction should not handle unevictable pages but CMA can do so */
1170         if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1171                 return ret;
1172
1173         ret = -EBUSY;
1174
1175         /*
1176          * To minimise LRU disruption, the caller can indicate that it only
1177          * wants to isolate pages it will be able to operate on without
1178          * blocking - clean pages for the most part.
1179          *
1180          * ISOLATE_CLEAN means that only clean pages should be isolated. This
1181          * is used by reclaim when it is cannot write to backing storage
1182          *
1183          * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1184          * that it is possible to migrate without blocking
1185          */
1186         if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1187                 /* All the caller can do on PageWriteback is block */
1188                 if (PageWriteback(page))
1189                         return ret;
1190
1191                 if (PageDirty(page)) {
1192                         struct address_space *mapping;
1193
1194                         /* ISOLATE_CLEAN means only clean pages */
1195                         if (mode & ISOLATE_CLEAN)
1196                                 return ret;
1197
1198                         /*
1199                          * Only pages without mappings or that have a
1200                          * ->migratepage callback are possible to migrate
1201                          * without blocking
1202                          */
1203                         mapping = page_mapping(page);
1204                         if (mapping && !mapping->a_ops->migratepage)
1205                                 return ret;
1206                 }
1207         }
1208
1209         if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1210                 return ret;
1211
1212         if (likely(get_page_unless_zero(page))) {
1213                 /*
1214                  * Be careful not to clear PageLRU until after we're
1215                  * sure the page is not being freed elsewhere -- the
1216                  * page release code relies on it.
1217                  */
1218                 ClearPageLRU(page);
1219                 ret = 0;
1220         }
1221
1222         return ret;
1223 }
1224
1225 /*
1226  * zone->lru_lock is heavily contended.  Some of the functions that
1227  * shrink the lists perform better by taking out a batch of pages
1228  * and working on them outside the LRU lock.
1229  *
1230  * For pagecache intensive workloads, this function is the hottest
1231  * spot in the kernel (apart from copy_*_user functions).
1232  *
1233  * Appropriate locks must be held before calling this function.
1234  *
1235  * @nr_to_scan: The number of pages to look through on the list.
1236  * @lruvec:     The LRU vector to pull pages from.
1237  * @dst:        The temp list to put pages on to.
1238  * @nr_scanned: The number of pages that were scanned.
1239  * @sc:         The scan_control struct for this reclaim session
1240  * @mode:       One of the LRU isolation modes
1241  * @lru:        LRU list id for isolating
1242  *
1243  * returns how many pages were moved onto *@dst.
1244  */
1245 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1246                 struct lruvec *lruvec, struct list_head *dst,
1247                 unsigned long *nr_scanned, struct scan_control *sc,
1248                 isolate_mode_t mode, enum lru_list lru)
1249 {
1250         struct list_head *src = &lruvec->lists[lru];
1251         unsigned long nr_taken = 0;
1252         unsigned long scan;
1253
1254         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1255                 struct page *page;
1256                 int nr_pages;
1257
1258                 page = lru_to_page(src);
1259                 prefetchw_prev_lru_page(page, src, flags);
1260
1261                 VM_BUG_ON_PAGE(!PageLRU(page), page);
1262
1263                 switch (__isolate_lru_page(page, mode)) {
1264                 case 0:
1265                         nr_pages = hpage_nr_pages(page);
1266                         mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1267                         list_move(&page->lru, dst);
1268                         nr_taken += nr_pages;
1269                         break;
1270
1271                 case -EBUSY:
1272                         /* else it is being freed elsewhere */
1273                         list_move(&page->lru, src);
1274                         continue;
1275
1276                 default:
1277                         BUG();
1278                 }
1279         }
1280
1281         *nr_scanned = scan;
1282         trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1283                                     nr_taken, mode, is_file_lru(lru));
1284         return nr_taken;
1285 }
1286
1287 /**
1288  * isolate_lru_page - tries to isolate a page from its LRU list
1289  * @page: page to isolate from its LRU list
1290  *
1291  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1292  * vmstat statistic corresponding to whatever LRU list the page was on.
1293  *
1294  * Returns 0 if the page was removed from an LRU list.
1295  * Returns -EBUSY if the page was not on an LRU list.
1296  *
1297  * The returned page will have PageLRU() cleared.  If it was found on
1298  * the active list, it will have PageActive set.  If it was found on
1299  * the unevictable list, it will have the PageUnevictable bit set. That flag
1300  * may need to be cleared by the caller before letting the page go.
1301  *
1302  * The vmstat statistic corresponding to the list on which the page was
1303  * found will be decremented.
1304  *
1305  * Restrictions:
1306  * (1) Must be called with an elevated refcount on the page. This is a
1307  *     fundamentnal difference from isolate_lru_pages (which is called
1308  *     without a stable reference).
1309  * (2) the lru_lock must not be held.
1310  * (3) interrupts must be enabled.
1311  */
1312 int isolate_lru_page(struct page *page)
1313 {
1314         int ret = -EBUSY;
1315
1316         VM_BUG_ON_PAGE(!page_count(page), page);
1317
1318         if (PageLRU(page)) {
1319                 struct zone *zone = page_zone(page);
1320                 struct lruvec *lruvec;
1321
1322                 spin_lock_irq(&zone->lru_lock);
1323                 lruvec = mem_cgroup_page_lruvec(page, zone);
1324                 if (PageLRU(page)) {
1325                         int lru = page_lru(page);
1326                         get_page(page);
1327                         ClearPageLRU(page);
1328                         del_page_from_lru_list(page, lruvec, lru);
1329                         ret = 0;
1330                 }
1331                 spin_unlock_irq(&zone->lru_lock);
1332         }
1333         return ret;
1334 }
1335
1336 /*
1337  * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1338  * then get resheduled. When there are massive number of tasks doing page
1339  * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1340  * the LRU list will go small and be scanned faster than necessary, leading to
1341  * unnecessary swapping, thrashing and OOM.
1342  */
1343 static int too_many_isolated(struct zone *zone, int file,
1344                 struct scan_control *sc)
1345 {
1346         unsigned long inactive, isolated;
1347
1348         if (current_is_kswapd())
1349                 return 0;
1350
1351         if (!global_reclaim(sc))
1352                 return 0;
1353
1354         if (file) {
1355                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1356                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1357         } else {
1358                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1359                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1360         }
1361
1362         /*
1363          * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1364          * won't get blocked by normal direct-reclaimers, forming a circular
1365          * deadlock.
1366          */
1367         if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1368                 inactive >>= 3;
1369
1370         return isolated > inactive;
1371 }
1372
1373 static noinline_for_stack void
1374 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1375 {
1376         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1377         struct zone *zone = lruvec_zone(lruvec);
1378         LIST_HEAD(pages_to_free);
1379
1380         /*
1381          * Put back any unfreeable pages.
1382          */
1383         while (!list_empty(page_list)) {
1384                 struct page *page = lru_to_page(page_list);
1385                 int lru;
1386
1387                 VM_BUG_ON_PAGE(PageLRU(page), page);
1388                 list_del(&page->lru);
1389                 if (unlikely(!page_evictable(page))) {
1390                         spin_unlock_irq(&zone->lru_lock);
1391                         putback_lru_page(page);
1392                         spin_lock_irq(&zone->lru_lock);
1393                         continue;
1394                 }
1395
1396                 lruvec = mem_cgroup_page_lruvec(page, zone);
1397
1398                 SetPageLRU(page);
1399                 lru = page_lru(page);
1400                 add_page_to_lru_list(page, lruvec, lru);
1401
1402                 if (is_active_lru(lru)) {
1403                         int file = is_file_lru(lru);
1404                         int numpages = hpage_nr_pages(page);
1405                         reclaim_stat->recent_rotated[file] += numpages;
1406                 }
1407                 if (put_page_testzero(page)) {
1408                         __ClearPageLRU(page);
1409                         __ClearPageActive(page);
1410                         del_page_from_lru_list(page, lruvec, lru);
1411
1412                         if (unlikely(PageCompound(page))) {
1413                                 spin_unlock_irq(&zone->lru_lock);
1414                                 (*get_compound_page_dtor(page))(page);
1415                                 spin_lock_irq(&zone->lru_lock);
1416                         } else
1417                                 list_add(&page->lru, &pages_to_free);
1418                 }
1419         }
1420
1421         /*
1422          * To save our caller's stack, now use input list for pages to free.
1423          */
1424         list_splice(&pages_to_free, page_list);
1425 }
1426
1427 /*
1428  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1429  * of reclaimed pages
1430  */
1431 static noinline_for_stack unsigned long
1432 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1433                      struct scan_control *sc, enum lru_list lru)
1434 {
1435         LIST_HEAD(page_list);
1436         unsigned long nr_scanned;
1437         unsigned long nr_reclaimed = 0;
1438         unsigned long nr_taken;
1439         unsigned long nr_dirty = 0;
1440         unsigned long nr_congested = 0;
1441         unsigned long nr_unqueued_dirty = 0;
1442         unsigned long nr_writeback = 0;
1443         unsigned long nr_immediate = 0;
1444         isolate_mode_t isolate_mode = 0;
1445         int file = is_file_lru(lru);
1446         struct zone *zone = lruvec_zone(lruvec);
1447         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1448
1449         while (unlikely(too_many_isolated(zone, file, sc))) {
1450                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1451
1452                 /* We are about to die and free our memory. Return now. */
1453                 if (fatal_signal_pending(current))
1454                         return SWAP_CLUSTER_MAX;
1455         }
1456
1457         lru_add_drain();
1458
1459         if (!sc->may_unmap)
1460                 isolate_mode |= ISOLATE_UNMAPPED;
1461         if (!sc->may_writepage)
1462                 isolate_mode |= ISOLATE_CLEAN;
1463
1464         spin_lock_irq(&zone->lru_lock);
1465
1466         nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1467                                      &nr_scanned, sc, isolate_mode, lru);
1468
1469         __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1470         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1471
1472         if (global_reclaim(sc)) {
1473                 zone->pages_scanned += nr_scanned;
1474                 if (current_is_kswapd())
1475                         __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1476                 else
1477                         __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1478         }
1479         spin_unlock_irq(&zone->lru_lock);
1480
1481         if (nr_taken == 0)
1482                 return 0;
1483
1484         nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1485                                 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1486                                 &nr_writeback, &nr_immediate,
1487                                 false);
1488
1489         spin_lock_irq(&zone->lru_lock);
1490
1491         reclaim_stat->recent_scanned[file] += nr_taken;
1492
1493         if (global_reclaim(sc)) {
1494                 if (current_is_kswapd())
1495                         __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1496                                                nr_reclaimed);
1497                 else
1498                         __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1499                                                nr_reclaimed);
1500         }
1501
1502         putback_inactive_pages(lruvec, &page_list);
1503
1504         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1505
1506         spin_unlock_irq(&zone->lru_lock);
1507
1508         free_hot_cold_page_list(&page_list, 1);
1509
1510         /*
1511          * If reclaim is isolating dirty pages under writeback, it implies
1512          * that the long-lived page allocation rate is exceeding the page
1513          * laundering rate. Either the global limits are not being effective
1514          * at throttling processes due to the page distribution throughout
1515          * zones or there is heavy usage of a slow backing device. The
1516          * only option is to throttle from reclaim context which is not ideal
1517          * as there is no guarantee the dirtying process is throttled in the
1518          * same way balance_dirty_pages() manages.
1519          *
1520          * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1521          * of pages under pages flagged for immediate reclaim and stall if any
1522          * are encountered in the nr_immediate check below.
1523          */
1524         if (nr_writeback && nr_writeback == nr_taken)
1525                 zone_set_flag(zone, ZONE_WRITEBACK);
1526
1527         /*
1528          * memcg will stall in page writeback so only consider forcibly
1529          * stalling for global reclaim
1530          */
1531         if (global_reclaim(sc)) {
1532                 /*
1533                  * Tag a zone as congested if all the dirty pages scanned were
1534                  * backed by a congested BDI and wait_iff_congested will stall.
1535                  */
1536                 if (nr_dirty && nr_dirty == nr_congested)
1537                         zone_set_flag(zone, ZONE_CONGESTED);
1538
1539                 /*
1540                  * If dirty pages are scanned that are not queued for IO, it
1541                  * implies that flushers are not keeping up. In this case, flag
1542                  * the zone ZONE_TAIL_LRU_DIRTY and kswapd will start writing
1543                  * pages from reclaim context. It will forcibly stall in the
1544                  * next check.
1545                  */
1546                 if (nr_unqueued_dirty == nr_taken)
1547                         zone_set_flag(zone, ZONE_TAIL_LRU_DIRTY);
1548
1549                 /*
1550                  * In addition, if kswapd scans pages marked marked for
1551                  * immediate reclaim and under writeback (nr_immediate), it
1552                  * implies that pages are cycling through the LRU faster than
1553                  * they are written so also forcibly stall.
1554                  */
1555                 if (nr_unqueued_dirty == nr_taken || nr_immediate)
1556                         congestion_wait(BLK_RW_ASYNC, HZ/10);
1557         }
1558
1559         /*
1560          * Stall direct reclaim for IO completions if underlying BDIs or zone
1561          * is congested. Allow kswapd to continue until it starts encountering
1562          * unqueued dirty pages or cycling through the LRU too quickly.
1563          */
1564         if (!sc->hibernation_mode && !current_is_kswapd())
1565                 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1566
1567         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1568                 zone_idx(zone),
1569                 nr_scanned, nr_reclaimed,
1570                 sc->priority,
1571                 trace_shrink_flags(file));
1572         return nr_reclaimed;
1573 }
1574
1575 /*
1576  * This moves pages from the active list to the inactive list.
1577  *
1578  * We move them the other way if the page is referenced by one or more
1579  * processes, from rmap.
1580  *
1581  * If the pages are mostly unmapped, the processing is fast and it is
1582  * appropriate to hold zone->lru_lock across the whole operation.  But if
1583  * the pages are mapped, the processing is slow (page_referenced()) so we
1584  * should drop zone->lru_lock around each page.  It's impossible to balance
1585  * this, so instead we remove the pages from the LRU while processing them.
1586  * It is safe to rely on PG_active against the non-LRU pages in here because
1587  * nobody will play with that bit on a non-LRU page.
1588  *
1589  * The downside is that we have to touch page->_count against each page.
1590  * But we had to alter page->flags anyway.
1591  */
1592
1593 static void move_active_pages_to_lru(struct lruvec *lruvec,
1594                                      struct list_head *list,
1595                                      struct list_head *pages_to_free,
1596                                      enum lru_list lru)
1597 {
1598         struct zone *zone = lruvec_zone(lruvec);
1599         unsigned long pgmoved = 0;
1600         struct page *page;
1601         int nr_pages;
1602
1603         while (!list_empty(list)) {
1604                 page = lru_to_page(list);
1605                 lruvec = mem_cgroup_page_lruvec(page, zone);
1606
1607                 VM_BUG_ON_PAGE(PageLRU(page), page);
1608                 SetPageLRU(page);
1609
1610                 nr_pages = hpage_nr_pages(page);
1611                 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1612                 list_move(&page->lru, &lruvec->lists[lru]);
1613                 pgmoved += nr_pages;
1614
1615                 if (put_page_testzero(page)) {
1616                         __ClearPageLRU(page);
1617                         __ClearPageActive(page);
1618                         del_page_from_lru_list(page, lruvec, lru);
1619
1620                         if (unlikely(PageCompound(page))) {
1621                                 spin_unlock_irq(&zone->lru_lock);
1622                                 (*get_compound_page_dtor(page))(page);
1623                                 spin_lock_irq(&zone->lru_lock);
1624                         } else
1625                                 list_add(&page->lru, pages_to_free);
1626                 }
1627         }
1628         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1629         if (!is_active_lru(lru))
1630                 __count_vm_events(PGDEACTIVATE, pgmoved);
1631 }
1632
1633 static void shrink_active_list(unsigned long nr_to_scan,
1634                                struct lruvec *lruvec,
1635                                struct scan_control *sc,
1636                                enum lru_list lru)
1637 {
1638         unsigned long nr_taken;
1639         unsigned long nr_scanned;
1640         unsigned long vm_flags;
1641         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1642         LIST_HEAD(l_active);
1643         LIST_HEAD(l_inactive);
1644         struct page *page;
1645         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1646         unsigned long nr_rotated = 0;
1647         isolate_mode_t isolate_mode = 0;
1648         int file = is_file_lru(lru);
1649         struct zone *zone = lruvec_zone(lruvec);
1650
1651         lru_add_drain();
1652
1653         if (!sc->may_unmap)
1654                 isolate_mode |= ISOLATE_UNMAPPED;
1655         if (!sc->may_writepage)
1656                 isolate_mode |= ISOLATE_CLEAN;
1657
1658         spin_lock_irq(&zone->lru_lock);
1659
1660         nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1661                                      &nr_scanned, sc, isolate_mode, lru);
1662         if (global_reclaim(sc))
1663                 zone->pages_scanned += nr_scanned;
1664
1665         reclaim_stat->recent_scanned[file] += nr_taken;
1666
1667         __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1668         __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1669         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1670         spin_unlock_irq(&zone->lru_lock);
1671
1672         while (!list_empty(&l_hold)) {
1673                 cond_resched();
1674                 page = lru_to_page(&l_hold);
1675                 list_del(&page->lru);
1676
1677                 if (unlikely(!page_evictable(page))) {
1678                         putback_lru_page(page);
1679                         continue;
1680                 }
1681
1682                 if (unlikely(buffer_heads_over_limit)) {
1683                         if (page_has_private(page) && trylock_page(page)) {
1684                                 if (page_has_private(page))
1685                                         try_to_release_page(page, 0);
1686                                 unlock_page(page);
1687                         }
1688                 }
1689
1690                 if (page_referenced(page, 0, sc->target_mem_cgroup,
1691                                     &vm_flags)) {
1692                         nr_rotated += hpage_nr_pages(page);
1693                         /*
1694                          * Identify referenced, file-backed active pages and
1695                          * give them one more trip around the active list. So
1696                          * that executable code get better chances to stay in
1697                          * memory under moderate memory pressure.  Anon pages
1698                          * are not likely to be evicted by use-once streaming
1699                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1700                          * so we ignore them here.
1701                          */
1702                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1703                                 list_add(&page->lru, &l_active);
1704                                 continue;
1705                         }
1706                 }
1707
1708                 ClearPageActive(page);  /* we are de-activating */
1709                 list_add(&page->lru, &l_inactive);
1710         }
1711
1712         /*
1713          * Move pages back to the lru list.
1714          */
1715         spin_lock_irq(&zone->lru_lock);
1716         /*
1717          * Count referenced pages from currently used mappings as rotated,
1718          * even though only some of them are actually re-activated.  This
1719          * helps balance scan pressure between file and anonymous pages in
1720          * get_scan_ratio.
1721          */
1722         reclaim_stat->recent_rotated[file] += nr_rotated;
1723
1724         move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1725         move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1726         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1727         spin_unlock_irq(&zone->lru_lock);
1728
1729         free_hot_cold_page_list(&l_hold, 1);
1730 }
1731
1732 #ifdef CONFIG_SWAP
1733 static int inactive_anon_is_low_global(struct zone *zone)
1734 {
1735         unsigned long active, inactive;
1736
1737         active = zone_page_state(zone, NR_ACTIVE_ANON);
1738         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1739
1740         if (inactive * zone->inactive_ratio < active)
1741                 return 1;
1742
1743         return 0;
1744 }
1745
1746 /**
1747  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1748  * @lruvec: LRU vector to check
1749  *
1750  * Returns true if the zone does not have enough inactive anon pages,
1751  * meaning some active anon pages need to be deactivated.
1752  */
1753 static int inactive_anon_is_low(struct lruvec *lruvec)
1754 {
1755         /*
1756          * If we don't have swap space, anonymous page deactivation
1757          * is pointless.
1758          */
1759         if (!total_swap_pages)
1760                 return 0;
1761
1762         if (!mem_cgroup_disabled())
1763                 return mem_cgroup_inactive_anon_is_low(lruvec);
1764
1765         return inactive_anon_is_low_global(lruvec_zone(lruvec));
1766 }
1767 #else
1768 static inline int inactive_anon_is_low(struct lruvec *lruvec)
1769 {
1770         return 0;
1771 }
1772 #endif
1773
1774 /**
1775  * inactive_file_is_low - check if file pages need to be deactivated
1776  * @lruvec: LRU vector to check
1777  *
1778  * When the system is doing streaming IO, memory pressure here
1779  * ensures that active file pages get deactivated, until more
1780  * than half of the file pages are on the inactive list.
1781  *
1782  * Once we get to that situation, protect the system's working
1783  * set from being evicted by disabling active file page aging.
1784  *
1785  * This uses a different ratio than the anonymous pages, because
1786  * the page cache uses a use-once replacement algorithm.
1787  */
1788 static int inactive_file_is_low(struct lruvec *lruvec)
1789 {
1790         unsigned long inactive;
1791         unsigned long active;
1792
1793         inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1794         active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1795
1796         return active > inactive;
1797 }
1798
1799 static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1800 {
1801         if (is_file_lru(lru))
1802                 return inactive_file_is_low(lruvec);
1803         else
1804                 return inactive_anon_is_low(lruvec);
1805 }
1806
1807 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1808                                  struct lruvec *lruvec, struct scan_control *sc)
1809 {
1810         if (is_active_lru(lru)) {
1811                 if (inactive_list_is_low(lruvec, lru))
1812                         shrink_active_list(nr_to_scan, lruvec, sc, lru);
1813                 return 0;
1814         }
1815
1816         return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1817 }
1818
1819 static int vmscan_swappiness(struct scan_control *sc)
1820 {
1821         if (global_reclaim(sc))
1822                 return vm_swappiness;
1823         return mem_cgroup_swappiness(sc->target_mem_cgroup);
1824 }
1825
1826 enum scan_balance {
1827         SCAN_EQUAL,
1828         SCAN_FRACT,
1829         SCAN_ANON,
1830         SCAN_FILE,
1831 };
1832
1833 /*
1834  * Determine how aggressively the anon and file LRU lists should be
1835  * scanned.  The relative value of each set of LRU lists is determined
1836  * by looking at the fraction of the pages scanned we did rotate back
1837  * onto the active list instead of evict.
1838  *
1839  * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1840  * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1841  */
1842 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
1843                            unsigned long *nr)
1844 {
1845         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1846         u64 fraction[2];
1847         u64 denominator = 0;    /* gcc */
1848         struct zone *zone = lruvec_zone(lruvec);
1849         unsigned long anon_prio, file_prio;
1850         enum scan_balance scan_balance;
1851         unsigned long anon, file, free;
1852         bool force_scan = false;
1853         unsigned long ap, fp;
1854         enum lru_list lru;
1855
1856         /*
1857          * If the zone or memcg is small, nr[l] can be 0.  This
1858          * results in no scanning on this priority and a potential
1859          * priority drop.  Global direct reclaim can go to the next
1860          * zone and tends to have no problems. Global kswapd is for
1861          * zone balancing and it needs to scan a minimum amount. When
1862          * reclaiming for a memcg, a priority drop can cause high
1863          * latencies, so it's better to scan a minimum amount there as
1864          * well.
1865          */
1866         if (current_is_kswapd() && !zone_reclaimable(zone))
1867                 force_scan = true;
1868         if (!global_reclaim(sc))
1869                 force_scan = true;
1870
1871         /* If we have no swap space, do not bother scanning anon pages. */
1872         if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
1873                 scan_balance = SCAN_FILE;
1874                 goto out;
1875         }
1876
1877         /*
1878          * Global reclaim will swap to prevent OOM even with no
1879          * swappiness, but memcg users want to use this knob to
1880          * disable swapping for individual groups completely when
1881          * using the memory controller's swap limit feature would be
1882          * too expensive.
1883          */
1884         if (!global_reclaim(sc) && !vmscan_swappiness(sc)) {
1885                 scan_balance = SCAN_FILE;
1886                 goto out;
1887         }
1888
1889         /*
1890          * Do not apply any pressure balancing cleverness when the
1891          * system is close to OOM, scan both anon and file equally
1892          * (unless the swappiness setting disagrees with swapping).
1893          */
1894         if (!sc->priority && vmscan_swappiness(sc)) {
1895                 scan_balance = SCAN_EQUAL;
1896                 goto out;
1897         }
1898
1899         anon  = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1900                 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1901         file  = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1902                 get_lru_size(lruvec, LRU_INACTIVE_FILE);
1903
1904         /*
1905          * If it's foreseeable that reclaiming the file cache won't be
1906          * enough to get the zone back into a desirable shape, we have
1907          * to swap.  Better start now and leave the - probably heavily
1908          * thrashing - remaining file pages alone.
1909          */
1910         if (global_reclaim(sc)) {
1911                 free = zone_page_state(zone, NR_FREE_PAGES);
1912                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1913                         scan_balance = SCAN_ANON;
1914                         goto out;
1915                 }
1916         }
1917
1918         /*
1919          * There is enough inactive page cache, do not reclaim
1920          * anything from the anonymous working set right now.
1921          */
1922         if (!inactive_file_is_low(lruvec)) {
1923                 scan_balance = SCAN_FILE;
1924                 goto out;
1925         }
1926
1927         scan_balance = SCAN_FRACT;
1928
1929         /*
1930          * With swappiness at 100, anonymous and file have the same priority.
1931          * This scanning priority is essentially the inverse of IO cost.
1932          */
1933         anon_prio = vmscan_swappiness(sc);
1934         file_prio = 200 - anon_prio;
1935
1936         /*
1937          * OK, so we have swap space and a fair amount of page cache
1938          * pages.  We use the recently rotated / recently scanned
1939          * ratios to determine how valuable each cache is.
1940          *
1941          * Because workloads change over time (and to avoid overflow)
1942          * we keep these statistics as a floating average, which ends
1943          * up weighing recent references more than old ones.
1944          *
1945          * anon in [0], file in [1]
1946          */
1947         spin_lock_irq(&zone->lru_lock);
1948         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1949                 reclaim_stat->recent_scanned[0] /= 2;
1950                 reclaim_stat->recent_rotated[0] /= 2;
1951         }
1952
1953         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1954                 reclaim_stat->recent_scanned[1] /= 2;
1955                 reclaim_stat->recent_rotated[1] /= 2;
1956         }
1957
1958         /*
1959          * The amount of pressure on anon vs file pages is inversely
1960          * proportional to the fraction of recently scanned pages on
1961          * each list that were recently referenced and in active use.
1962          */
1963         ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
1964         ap /= reclaim_stat->recent_rotated[0] + 1;
1965
1966         fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
1967         fp /= reclaim_stat->recent_rotated[1] + 1;
1968         spin_unlock_irq(&zone->lru_lock);
1969
1970         fraction[0] = ap;
1971         fraction[1] = fp;
1972         denominator = ap + fp + 1;
1973 out:
1974         for_each_evictable_lru(lru) {
1975                 int file = is_file_lru(lru);
1976                 unsigned long size;
1977                 unsigned long scan;
1978
1979                 size = get_lru_size(lruvec, lru);
1980                 scan = size >> sc->priority;
1981
1982                 if (!scan && force_scan)
1983                         scan = min(size, SWAP_CLUSTER_MAX);
1984
1985                 switch (scan_balance) {
1986                 case SCAN_EQUAL:
1987                         /* Scan lists relative to size */
1988                         break;
1989                 case SCAN_FRACT:
1990                         /*
1991                          * Scan types proportional to swappiness and
1992                          * their relative recent reclaim efficiency.
1993                          */
1994                         scan = div64_u64(scan * fraction[file], denominator);
1995                         break;
1996                 case SCAN_FILE:
1997                 case SCAN_ANON:
1998                         /* Scan one type exclusively */
1999                         if ((scan_balance == SCAN_FILE) != file)
2000                                 scan = 0;
2001                         break;
2002                 default:
2003                         /* Look ma, no brain */
2004                         BUG();
2005                 }
2006                 nr[lru] = scan;
2007         }
2008 }
2009
2010 /*
2011  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
2012  */
2013 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
2014 {
2015         unsigned long nr[NR_LRU_LISTS];
2016         unsigned long targets[NR_LRU_LISTS];
2017         unsigned long nr_to_scan;
2018         enum lru_list lru;
2019         unsigned long nr_reclaimed = 0;
2020         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2021         struct blk_plug plug;
2022         bool scan_adjusted = false;
2023
2024         get_scan_count(lruvec, sc, nr);
2025
2026         /* Record the original scan target for proportional adjustments later */
2027         memcpy(targets, nr, sizeof(nr));
2028
2029         blk_start_plug(&plug);
2030         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2031                                         nr[LRU_INACTIVE_FILE]) {
2032                 unsigned long nr_anon, nr_file, percentage;
2033                 unsigned long nr_scanned;
2034
2035                 for_each_evictable_lru(lru) {
2036                         if (nr[lru]) {
2037                                 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2038                                 nr[lru] -= nr_to_scan;
2039
2040                                 nr_reclaimed += shrink_list(lru, nr_to_scan,
2041                                                             lruvec, sc);
2042                         }
2043                 }
2044
2045                 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2046                         continue;
2047
2048                 /*
2049                  * For global direct reclaim, reclaim only the number of pages
2050                  * requested. Less care is taken to scan proportionally as it
2051                  * is more important to minimise direct reclaim stall latency
2052                  * than it is to properly age the LRU lists.
2053                  */
2054                 if (global_reclaim(sc) && !current_is_kswapd())
2055                         break;
2056
2057                 /*
2058                  * For kswapd and memcg, reclaim at least the number of pages
2059                  * requested. Ensure that the anon and file LRUs shrink
2060                  * proportionally what was requested by get_scan_count(). We
2061                  * stop reclaiming one LRU and reduce the amount scanning
2062                  * proportional to the original scan target.
2063                  */
2064                 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2065                 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2066
2067                 if (nr_file > nr_anon) {
2068                         unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2069                                                 targets[LRU_ACTIVE_ANON] + 1;
2070                         lru = LRU_BASE;
2071                         percentage = nr_anon * 100 / scan_target;
2072                 } else {
2073                         unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2074                                                 targets[LRU_ACTIVE_FILE] + 1;
2075                         lru = LRU_FILE;
2076                         percentage = nr_file * 100 / scan_target;
2077                 }
2078
2079                 /* Stop scanning the smaller of the LRU */
2080                 nr[lru] = 0;
2081                 nr[lru + LRU_ACTIVE] = 0;
2082
2083                 /*
2084                  * Recalculate the other LRU scan count based on its original
2085                  * scan target and the percentage scanning already complete
2086                  */
2087                 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2088                 nr_scanned = targets[lru] - nr[lru];
2089                 nr[lru] = targets[lru] * (100 - percentage) / 100;
2090                 nr[lru] -= min(nr[lru], nr_scanned);
2091
2092                 lru += LRU_ACTIVE;
2093                 nr_scanned = targets[lru] - nr[lru];
2094                 nr[lru] = targets[lru] * (100 - percentage) / 100;
2095                 nr[lru] -= min(nr[lru], nr_scanned);
2096
2097                 scan_adjusted = true;
2098         }
2099         blk_finish_plug(&plug);
2100         sc->nr_reclaimed += nr_reclaimed;
2101
2102         /*
2103          * Even if we did not try to evict anon pages at all, we want to
2104          * rebalance the anon lru active/inactive ratio.
2105          */
2106         if (inactive_anon_is_low(lruvec))
2107                 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2108                                    sc, LRU_ACTIVE_ANON);
2109
2110         throttle_vm_writeout(sc->gfp_mask);
2111 }
2112
2113 /* Use reclaim/compaction for costly allocs or under memory pressure */
2114 static bool in_reclaim_compaction(struct scan_control *sc)
2115 {
2116         if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2117                         (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
2118                          sc->priority < DEF_PRIORITY - 2))
2119                 return true;
2120
2121         return false;
2122 }
2123
2124 /*
2125  * Reclaim/compaction is used for high-order allocation requests. It reclaims
2126  * order-0 pages before compacting the zone. should_continue_reclaim() returns
2127  * true if more pages should be reclaimed such that when the page allocator
2128  * calls try_to_compact_zone() that it will have enough free pages to succeed.
2129  * It will give up earlier than that if there is difficulty reclaiming pages.
2130  */
2131 static inline bool should_continue_reclaim(struct zone *zone,
2132                                         unsigned long nr_reclaimed,
2133                                         unsigned long nr_scanned,
2134                                         struct scan_control *sc)
2135 {
2136         unsigned long pages_for_compaction;
2137         unsigned long inactive_lru_pages;
2138
2139         /* If not in reclaim/compaction mode, stop */
2140         if (!in_reclaim_compaction(sc))
2141                 return false;
2142
2143         /* Consider stopping depending on scan and reclaim activity */
2144         if (sc->gfp_mask & __GFP_REPEAT) {
2145                 /*
2146                  * For __GFP_REPEAT allocations, stop reclaiming if the
2147                  * full LRU list has been scanned and we are still failing
2148                  * to reclaim pages. This full LRU scan is potentially
2149                  * expensive but a __GFP_REPEAT caller really wants to succeed
2150                  */
2151                 if (!nr_reclaimed && !nr_scanned)
2152                         return false;
2153         } else {
2154                 /*
2155                  * For non-__GFP_REPEAT allocations which can presumably
2156                  * fail without consequence, stop if we failed to reclaim
2157                  * any pages from the last SWAP_CLUSTER_MAX number of
2158                  * pages that were scanned. This will return to the
2159                  * caller faster at the risk reclaim/compaction and
2160                  * the resulting allocation attempt fails
2161                  */
2162                 if (!nr_reclaimed)
2163                         return false;
2164         }
2165
2166         /*
2167          * If we have not reclaimed enough pages for compaction and the
2168          * inactive lists are large enough, continue reclaiming
2169          */
2170         pages_for_compaction = (2UL << sc->order);
2171         inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
2172         if (get_nr_swap_pages() > 0)
2173                 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
2174         if (sc->nr_reclaimed < pages_for_compaction &&
2175                         inactive_lru_pages > pages_for_compaction)
2176                 return true;
2177
2178         /* If compaction would go ahead or the allocation would succeed, stop */
2179         switch (compaction_suitable(zone, sc->order)) {
2180         case COMPACT_PARTIAL:
2181         case COMPACT_CONTINUE:
2182                 return false;
2183         default:
2184                 return true;
2185         }
2186 }
2187
2188 static void shrink_zone(struct zone *zone, struct scan_control *sc)
2189 {
2190         unsigned long nr_reclaimed, nr_scanned;
2191
2192         do {
2193                 struct mem_cgroup *root = sc->target_mem_cgroup;
2194                 struct mem_cgroup_reclaim_cookie reclaim = {
2195                         .zone = zone,
2196                         .priority = sc->priority,
2197                 };
2198                 struct mem_cgroup *memcg;
2199
2200                 nr_reclaimed = sc->nr_reclaimed;
2201                 nr_scanned = sc->nr_scanned;
2202
2203                 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2204                 do {
2205                         struct lruvec *lruvec;
2206
2207                         lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2208
2209                         shrink_lruvec(lruvec, sc);
2210
2211                         /*
2212                          * Direct reclaim and kswapd have to scan all memory
2213                          * cgroups to fulfill the overall scan target for the
2214                          * zone.
2215                          *
2216                          * Limit reclaim, on the other hand, only cares about
2217                          * nr_to_reclaim pages to be reclaimed and it will
2218                          * retry with decreasing priority if one round over the
2219                          * whole hierarchy is not sufficient.
2220                          */
2221                         if (!global_reclaim(sc) &&
2222                                         sc->nr_reclaimed >= sc->nr_to_reclaim) {
2223                                 mem_cgroup_iter_break(root, memcg);
2224                                 break;
2225                         }
2226                         memcg = mem_cgroup_iter(root, memcg, &reclaim);
2227                 } while (memcg);
2228
2229                 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2230                            sc->nr_scanned - nr_scanned,
2231                            sc->nr_reclaimed - nr_reclaimed);
2232
2233         } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2234                                          sc->nr_scanned - nr_scanned, sc));
2235 }
2236
2237 /* Returns true if compaction should go ahead for a high-order request */
2238 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2239 {
2240         unsigned long balance_gap, watermark;
2241         bool watermark_ok;
2242
2243         /* Do not consider compaction for orders reclaim is meant to satisfy */
2244         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2245                 return false;
2246
2247         /*
2248          * Compaction takes time to run and there are potentially other
2249          * callers using the pages just freed. Continue reclaiming until
2250          * there is a buffer of free pages available to give compaction
2251          * a reasonable chance of completing and allocating the page
2252          */
2253         balance_gap = min(low_wmark_pages(zone),
2254                 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2255                         KSWAPD_ZONE_BALANCE_GAP_RATIO);
2256         watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2257         watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2258
2259         /*
2260          * If compaction is deferred, reclaim up to a point where
2261          * compaction will have a chance of success when re-enabled
2262          */
2263         if (compaction_deferred(zone, sc->order))
2264                 return watermark_ok;
2265
2266         /* If compaction is not ready to start, keep reclaiming */
2267         if (!compaction_suitable(zone, sc->order))
2268                 return false;
2269
2270         return watermark_ok;
2271 }
2272
2273 /*
2274  * This is the direct reclaim path, for page-allocating processes.  We only
2275  * try to reclaim pages from zones which will satisfy the caller's allocation
2276  * request.
2277  *
2278  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2279  * Because:
2280  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2281  *    allocation or
2282  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2283  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2284  *    zone defense algorithm.
2285  *
2286  * If a zone is deemed to be full of pinned pages then just give it a light
2287  * scan then give up on it.
2288  *
2289  * This function returns true if a zone is being reclaimed for a costly
2290  * high-order allocation and compaction is ready to begin. This indicates to
2291  * the caller that it should consider retrying the allocation instead of
2292  * further reclaim.
2293  */
2294 static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2295 {
2296         struct zoneref *z;
2297         struct zone *zone;
2298         unsigned long nr_soft_reclaimed;
2299         unsigned long nr_soft_scanned;
2300         bool aborted_reclaim = false;
2301
2302         /*
2303          * If the number of buffer_heads in the machine exceeds the maximum
2304          * allowed level, force direct reclaim to scan the highmem zone as
2305          * highmem pages could be pinning lowmem pages storing buffer_heads
2306          */
2307         if (buffer_heads_over_limit)
2308                 sc->gfp_mask |= __GFP_HIGHMEM;
2309
2310         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2311                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2312                 if (!populated_zone(zone))
2313                         continue;
2314                 /*
2315                  * Take care memory controller reclaiming has small influence
2316                  * to global LRU.
2317                  */
2318                 if (global_reclaim(sc)) {
2319                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2320                                 continue;
2321                         if (sc->priority != DEF_PRIORITY &&
2322                             !zone_reclaimable(zone))
2323                                 continue;       /* Let kswapd poll it */
2324                         if (IS_ENABLED(CONFIG_COMPACTION)) {
2325                                 /*
2326                                  * If we already have plenty of memory free for
2327                                  * compaction in this zone, don't free any more.
2328                                  * Even though compaction is invoked for any
2329                                  * non-zero order, only frequent costly order
2330                                  * reclamation is disruptive enough to become a
2331                                  * noticeable problem, like transparent huge
2332                                  * page allocations.
2333                                  */
2334                                 if (compaction_ready(zone, sc)) {
2335                                         aborted_reclaim = true;
2336                                         continue;
2337                                 }
2338                         }
2339                         /*
2340                          * This steals pages from memory cgroups over softlimit
2341                          * and returns the number of reclaimed pages and
2342                          * scanned pages. This works for global memory pressure
2343                          * and balancing, not for a memcg's limit.
2344                          */
2345                         nr_soft_scanned = 0;
2346                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2347                                                 sc->order, sc->gfp_mask,
2348                                                 &nr_soft_scanned);
2349                         sc->nr_reclaimed += nr_soft_reclaimed;
2350                         sc->nr_scanned += nr_soft_scanned;
2351                         /* need some check for avoid more shrink_zone() */
2352                 }
2353
2354                 shrink_zone(zone, sc);
2355         }
2356
2357         return aborted_reclaim;
2358 }
2359
2360 /* All zones in zonelist are unreclaimable? */
2361 static bool all_unreclaimable(struct zonelist *zonelist,
2362                 struct scan_control *sc)
2363 {
2364         struct zoneref *z;
2365         struct zone *zone;
2366
2367         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2368                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2369                 if (!populated_zone(zone))
2370                         continue;
2371                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2372                         continue;
2373                 if (zone_reclaimable(zone))
2374                         return false;
2375         }
2376
2377         return true;
2378 }
2379
2380 /*
2381  * This is the main entry point to direct page reclaim.
2382  *
2383  * If a full scan of the inactive list fails to free enough memory then we
2384  * are "out of memory" and something needs to be killed.
2385  *
2386  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2387  * high - the zone may be full of dirty or under-writeback pages, which this
2388  * caller can't do much about.  We kick the writeback threads and take explicit
2389  * naps in the hope that some of these pages can be written.  But if the
2390  * allocating task holds filesystem locks which prevent writeout this might not
2391  * work, and the allocation attempt will fail.
2392  *
2393  * returns:     0, if no pages reclaimed
2394  *              else, the number of pages reclaimed
2395  */
2396 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2397                                         struct scan_control *sc,
2398                                         struct shrink_control *shrink)
2399 {
2400         unsigned long total_scanned = 0;
2401         struct reclaim_state *reclaim_state = current->reclaim_state;
2402         struct zoneref *z;
2403         struct zone *zone;
2404         unsigned long writeback_threshold;
2405         bool aborted_reclaim;
2406
2407         delayacct_freepages_start();
2408
2409         if (global_reclaim(sc))
2410                 count_vm_event(ALLOCSTALL);
2411
2412         do {
2413                 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2414                                 sc->priority);
2415                 sc->nr_scanned = 0;
2416                 aborted_reclaim = shrink_zones(zonelist, sc);
2417
2418                 /*
2419                  * Don't shrink slabs when reclaiming memory from over limit
2420                  * cgroups but do shrink slab at least once when aborting
2421                  * reclaim for compaction to avoid unevenly scanning file/anon
2422                  * LRU pages over slab pages.
2423                  */
2424                 if (global_reclaim(sc)) {
2425                         unsigned long lru_pages = 0;
2426
2427                         nodes_clear(shrink->nodes_to_scan);
2428                         for_each_zone_zonelist(zone, z, zonelist,
2429                                         gfp_zone(sc->gfp_mask)) {
2430                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2431                                         continue;
2432
2433                                 lru_pages += zone_reclaimable_pages(zone);
2434                                 node_set(zone_to_nid(zone),
2435                                          shrink->nodes_to_scan);
2436                         }
2437
2438                         shrink_slab(shrink, sc->nr_scanned, lru_pages);
2439                         if (reclaim_state) {
2440                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2441                                 reclaim_state->reclaimed_slab = 0;
2442                         }
2443                 }
2444                 total_scanned += sc->nr_scanned;
2445                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2446                         goto out;
2447
2448                 /*
2449                  * If we're getting trouble reclaiming, start doing
2450                  * writepage even in laptop mode.
2451                  */
2452                 if (sc->priority < DEF_PRIORITY - 2)
2453                         sc->may_writepage = 1;
2454
2455                 /*
2456                  * Try to write back as many pages as we just scanned.  This
2457                  * tends to cause slow streaming writers to write data to the
2458                  * disk smoothly, at the dirtying rate, which is nice.   But
2459                  * that's undesirable in laptop mode, where we *want* lumpy
2460                  * writeout.  So in laptop mode, write out the whole world.
2461                  */
2462                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2463                 if (total_scanned > writeback_threshold) {
2464                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2465                                                 WB_REASON_TRY_TO_FREE_PAGES);
2466                         sc->may_writepage = 1;
2467                 }
2468         } while (--sc->priority >= 0 && !aborted_reclaim);
2469
2470 out:
2471         delayacct_freepages_end();
2472
2473         if (sc->nr_reclaimed)
2474                 return sc->nr_reclaimed;
2475
2476         /*
2477          * As hibernation is going on, kswapd is freezed so that it can't mark
2478          * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2479          * check.
2480          */
2481         if (oom_killer_disabled)
2482                 return 0;
2483
2484         /* Aborted reclaim to try compaction? don't OOM, then */
2485         if (aborted_reclaim)
2486                 return 1;
2487
2488         /* top priority shrink_zones still had more to do? don't OOM, then */
2489         if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2490                 return 1;
2491
2492         return 0;
2493 }
2494
2495 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2496 {
2497         struct zone *zone;
2498         unsigned long pfmemalloc_reserve = 0;
2499         unsigned long free_pages = 0;
2500         int i;
2501         bool wmark_ok;
2502
2503         for (i = 0; i <= ZONE_NORMAL; i++) {
2504                 zone = &pgdat->node_zones[i];
2505                 pfmemalloc_reserve += min_wmark_pages(zone);
2506                 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2507         }
2508
2509         wmark_ok = free_pages > pfmemalloc_reserve / 2;
2510
2511         /* kswapd must be awake if processes are being throttled */
2512         if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2513                 pgdat->classzone_idx = min(pgdat->classzone_idx,
2514                                                 (enum zone_type)ZONE_NORMAL);
2515                 wake_up_interruptible(&pgdat->kswapd_wait);
2516         }
2517
2518         return wmark_ok;
2519 }
2520
2521 /*
2522  * Throttle direct reclaimers if backing storage is backed by the network
2523  * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2524  * depleted. kswapd will continue to make progress and wake the processes
2525  * when the low watermark is reached.
2526  *
2527  * Returns true if a fatal signal was delivered during throttling. If this
2528  * happens, the page allocator should not consider triggering the OOM killer.
2529  */
2530 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2531                                         nodemask_t *nodemask)
2532 {
2533         struct zone *zone;
2534         int high_zoneidx = gfp_zone(gfp_mask);
2535         pg_data_t *pgdat;
2536
2537         /*
2538          * Kernel threads should not be throttled as they may be indirectly
2539          * responsible for cleaning pages necessary for reclaim to make forward
2540          * progress. kjournald for example may enter direct reclaim while
2541          * committing a transaction where throttling it could forcing other
2542          * processes to block on log_wait_commit().
2543          */
2544         if (current->flags & PF_KTHREAD)
2545                 goto out;
2546
2547         /*
2548          * If a fatal signal is pending, this process should not throttle.
2549          * It should return quickly so it can exit and free its memory
2550          */
2551         if (fatal_signal_pending(current))
2552                 goto out;
2553
2554         /* Check if the pfmemalloc reserves are ok */
2555         first_zones_zonelist(zonelist, high_zoneidx, NULL, &zone);
2556         pgdat = zone->zone_pgdat;
2557         if (pfmemalloc_watermark_ok(pgdat))
2558                 goto out;
2559
2560         /* Account for the throttling */
2561         count_vm_event(PGSCAN_DIRECT_THROTTLE);
2562
2563         /*
2564          * If the caller cannot enter the filesystem, it's possible that it
2565          * is due to the caller holding an FS lock or performing a journal
2566          * transaction in the case of a filesystem like ext[3|4]. In this case,
2567          * it is not safe to block on pfmemalloc_wait as kswapd could be
2568          * blocked waiting on the same lock. Instead, throttle for up to a
2569          * second before continuing.
2570          */
2571         if (!(gfp_mask & __GFP_FS)) {
2572                 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2573                         pfmemalloc_watermark_ok(pgdat), HZ);
2574
2575                 goto check_pending;
2576         }
2577
2578         /* Throttle until kswapd wakes the process */
2579         wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2580                 pfmemalloc_watermark_ok(pgdat));
2581
2582 check_pending:
2583         if (fatal_signal_pending(current))
2584                 return true;
2585
2586 out:
2587         return false;
2588 }
2589
2590 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2591                                 gfp_t gfp_mask, nodemask_t *nodemask)
2592 {
2593         unsigned long nr_reclaimed;
2594         struct scan_control sc = {
2595                 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2596                 .may_writepage = !laptop_mode,
2597                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2598                 .may_unmap = 1,
2599                 .may_swap = 1,
2600                 .order = order,
2601                 .priority = DEF_PRIORITY,
2602                 .target_mem_cgroup = NULL,
2603                 .nodemask = nodemask,
2604         };
2605         struct shrink_control shrink = {
2606                 .gfp_mask = sc.gfp_mask,
2607         };
2608
2609         /*
2610          * Do not enter reclaim if fatal signal was delivered while throttled.
2611          * 1 is returned so that the page allocator does not OOM kill at this
2612          * point.
2613          */
2614         if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2615                 return 1;
2616
2617         trace_mm_vmscan_direct_reclaim_begin(order,
2618                                 sc.may_writepage,
2619                                 gfp_mask);
2620
2621         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2622
2623         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2624
2625         return nr_reclaimed;
2626 }
2627
2628 #ifdef CONFIG_MEMCG
2629
2630 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2631                                                 gfp_t gfp_mask, bool noswap,
2632                                                 struct zone *zone,
2633                                                 unsigned long *nr_scanned)
2634 {
2635         struct scan_control sc = {
2636                 .nr_scanned = 0,
2637                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2638                 .may_writepage = !laptop_mode,
2639                 .may_unmap = 1,
2640                 .may_swap = !noswap,
2641                 .order = 0,
2642                 .priority = 0,
2643                 .target_mem_cgroup = memcg,
2644         };
2645         struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2646
2647         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2648                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2649
2650         trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2651                                                       sc.may_writepage,
2652                                                       sc.gfp_mask);
2653
2654         /*
2655          * NOTE: Although we can get the priority field, using it
2656          * here is not a good idea, since it limits the pages we can scan.
2657          * if we don't reclaim here, the shrink_zone from balance_pgdat
2658          * will pick up pages from other mem cgroup's as well. We hack
2659          * the priority and make it zero.
2660          */
2661         shrink_lruvec(lruvec, &sc);
2662
2663         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2664
2665         *nr_scanned = sc.nr_scanned;
2666         return sc.nr_reclaimed;
2667 }
2668
2669 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2670                                            gfp_t gfp_mask,
2671                                            bool noswap)
2672 {
2673         struct zonelist *zonelist;
2674         unsigned long nr_reclaimed;
2675         int nid;
2676         struct scan_control sc = {
2677                 .may_writepage = !laptop_mode,
2678                 .may_unmap = 1,
2679                 .may_swap = !noswap,
2680                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2681                 .order = 0,
2682                 .priority = DEF_PRIORITY,
2683                 .target_mem_cgroup = memcg,
2684                 .nodemask = NULL, /* we don't care the placement */
2685                 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2686                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2687         };
2688         struct shrink_control shrink = {
2689                 .gfp_mask = sc.gfp_mask,
2690         };
2691
2692         /*
2693          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2694          * take care of from where we get pages. So the node where we start the
2695          * scan does not need to be the current node.
2696          */
2697         nid = mem_cgroup_select_victim_node(memcg);
2698
2699         zonelist = NODE_DATA(nid)->node_zonelists;
2700
2701         trace_mm_vmscan_memcg_reclaim_begin(0,
2702                                             sc.may_writepage,
2703                                             sc.gfp_mask);
2704
2705         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2706
2707         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2708
2709         return nr_reclaimed;
2710 }
2711 #endif
2712
2713 static void age_active_anon(struct zone *zone, struct scan_control *sc)
2714 {
2715         struct mem_cgroup *memcg;
2716
2717         if (!total_swap_pages)
2718                 return;
2719
2720         memcg = mem_cgroup_iter(NULL, NULL, NULL);
2721         do {
2722                 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2723
2724                 if (inactive_anon_is_low(lruvec))
2725                         shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2726                                            sc, LRU_ACTIVE_ANON);
2727
2728                 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2729         } while (memcg);
2730 }
2731
2732 static bool zone_balanced(struct zone *zone, int order,
2733                           unsigned long balance_gap, int classzone_idx)
2734 {
2735         if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2736                                     balance_gap, classzone_idx, 0))
2737                 return false;
2738
2739         if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2740             !compaction_suitable(zone, order))
2741                 return false;
2742
2743         return true;
2744 }
2745
2746 /*
2747  * pgdat_balanced() is used when checking if a node is balanced.
2748  *
2749  * For order-0, all zones must be balanced!
2750  *
2751  * For high-order allocations only zones that meet watermarks and are in a
2752  * zone allowed by the callers classzone_idx are added to balanced_pages. The
2753  * total of balanced pages must be at least 25% of the zones allowed by
2754  * classzone_idx for the node to be considered balanced. Forcing all zones to
2755  * be balanced for high orders can cause excessive reclaim when there are
2756  * imbalanced zones.
2757  * The choice of 25% is due to
2758  *   o a 16M DMA zone that is balanced will not balance a zone on any
2759  *     reasonable sized machine
2760  *   o On all other machines, the top zone must be at least a reasonable
2761  *     percentage of the middle zones. For example, on 32-bit x86, highmem
2762  *     would need to be at least 256M for it to be balance a whole node.
2763  *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2764  *     to balance a node on its own. These seemed like reasonable ratios.
2765  */
2766 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
2767 {
2768         unsigned long managed_pages = 0;
2769         unsigned long balanced_pages = 0;
2770         int i;
2771
2772         /* Check the watermark levels */
2773         for (i = 0; i <= classzone_idx; i++) {
2774                 struct zone *zone = pgdat->node_zones + i;
2775
2776                 if (!populated_zone(zone))
2777                         continue;
2778
2779                 managed_pages += zone->managed_pages;
2780
2781                 /*
2782                  * A special case here:
2783                  *
2784                  * balance_pgdat() skips over all_unreclaimable after
2785                  * DEF_PRIORITY. Effectively, it considers them balanced so
2786                  * they must be considered balanced here as well!
2787                  */
2788                 if (!zone_reclaimable(zone)) {
2789                         balanced_pages += zone->managed_pages;
2790                         continue;
2791                 }
2792
2793                 if (zone_balanced(zone, order, 0, i))
2794                         balanced_pages += zone->managed_pages;
2795                 else if (!order)
2796                         return false;
2797         }
2798
2799         if (order)
2800                 return balanced_pages >= (managed_pages >> 2);
2801         else
2802                 return true;
2803 }
2804
2805 /*
2806  * Prepare kswapd for sleeping. This verifies that there are no processes
2807  * waiting in throttle_direct_reclaim() and that watermarks have been met.
2808  *
2809  * Returns true if kswapd is ready to sleep
2810  */
2811 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
2812                                         int classzone_idx)
2813 {
2814         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2815         if (remaining)
2816                 return false;
2817
2818         /*
2819          * There is a potential race between when kswapd checks its watermarks
2820          * and a process gets throttled. There is also a potential race if
2821          * processes get throttled, kswapd wakes, a large process exits therby
2822          * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2823          * is going to sleep, no process should be sleeping on pfmemalloc_wait
2824          * so wake them now if necessary. If necessary, processes will wake
2825          * kswapd and get throttled again
2826          */
2827         if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2828                 wake_up(&pgdat->pfmemalloc_wait);
2829                 return false;
2830         }
2831
2832         return pgdat_balanced(pgdat, order, classzone_idx);
2833 }
2834
2835 /*
2836  * kswapd shrinks the zone by the number of pages required to reach
2837  * the high watermark.
2838  *
2839  * Returns true if kswapd scanned at least the requested number of pages to
2840  * reclaim or if the lack of progress was due to pages under writeback.
2841  * This is used to determine if the scanning priority needs to be raised.
2842  */
2843 static bool kswapd_shrink_zone(struct zone *zone,
2844                                int classzone_idx,
2845                                struct scan_control *sc,
2846                                unsigned long lru_pages,
2847                                unsigned long *nr_attempted)
2848 {
2849         int testorder = sc->order;
2850         unsigned long balance_gap;
2851         struct reclaim_state *reclaim_state = current->reclaim_state;
2852         struct shrink_control shrink = {
2853                 .gfp_mask = sc->gfp_mask,
2854         };
2855         bool lowmem_pressure;
2856
2857         /* Reclaim above the high watermark. */
2858         sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
2859
2860         /*
2861          * Kswapd reclaims only single pages with compaction enabled. Trying
2862          * too hard to reclaim until contiguous free pages have become
2863          * available can hurt performance by evicting too much useful data
2864          * from memory. Do not reclaim more than needed for compaction.
2865          */
2866         if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2867                         compaction_suitable(zone, sc->order) !=
2868                                 COMPACT_SKIPPED)
2869                 testorder = 0;
2870
2871         /*
2872          * We put equal pressure on every zone, unless one zone has way too
2873          * many pages free already. The "too many pages" is defined as the
2874          * high wmark plus a "gap" where the gap is either the low
2875          * watermark or 1% of the zone, whichever is smaller.
2876          */
2877         balance_gap = min(low_wmark_pages(zone),
2878                 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2879                 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2880
2881         /*
2882          * If there is no low memory pressure or the zone is balanced then no
2883          * reclaim is necessary
2884          */
2885         lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
2886         if (!lowmem_pressure && zone_balanced(zone, testorder,
2887                                                 balance_gap, classzone_idx))
2888                 return true;
2889
2890         shrink_zone(zone, sc);
2891         nodes_clear(shrink.nodes_to_scan);
2892         node_set(zone_to_nid(zone), shrink.nodes_to_scan);
2893
2894         reclaim_state->reclaimed_slab = 0;
2895         shrink_slab(&shrink, sc->nr_scanned, lru_pages);
2896         sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2897
2898         /* Account for the number of pages attempted to reclaim */
2899         *nr_attempted += sc->nr_to_reclaim;
2900
2901         zone_clear_flag(zone, ZONE_WRITEBACK);
2902
2903         /*
2904          * If a zone reaches its high watermark, consider it to be no longer
2905          * congested. It's possible there are dirty pages backed by congested
2906          * BDIs but as pressure is relieved, speculatively avoid congestion
2907          * waits.
2908          */
2909         if (zone_reclaimable(zone) &&
2910             zone_balanced(zone, testorder, 0, classzone_idx)) {
2911                 zone_clear_flag(zone, ZONE_CONGESTED);
2912                 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
2913         }
2914
2915         return sc->nr_scanned >= sc->nr_to_reclaim;
2916 }
2917
2918 /*
2919  * For kswapd, balance_pgdat() will work across all this node's zones until
2920  * they are all at high_wmark_pages(zone).
2921  *
2922  * Returns the final order kswapd was reclaiming at
2923  *
2924  * There is special handling here for zones which are full of pinned pages.
2925  * This can happen if the pages are all mlocked, or if they are all used by
2926  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2927  * What we do is to detect the case where all pages in the zone have been
2928  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2929  * dead and from now on, only perform a short scan.  Basically we're polling
2930  * the zone for when the problem goes away.
2931  *
2932  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2933  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2934  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2935  * lower zones regardless of the number of free pages in the lower zones. This
2936  * interoperates with the page allocator fallback scheme to ensure that aging
2937  * of pages is balanced across the zones.
2938  */
2939 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2940                                                         int *classzone_idx)
2941 {
2942         int i;
2943         int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2944         unsigned long nr_soft_reclaimed;
2945         unsigned long nr_soft_scanned;
2946         struct scan_control sc = {
2947                 .gfp_mask = GFP_KERNEL,
2948                 .priority = DEF_PRIORITY,
2949                 .may_unmap = 1,
2950                 .may_swap = 1,
2951                 .may_writepage = !laptop_mode,
2952                 .order = order,
2953                 .target_mem_cgroup = NULL,
2954         };
2955         count_vm_event(PAGEOUTRUN);
2956
2957         do {
2958                 unsigned long lru_pages = 0;
2959                 unsigned long nr_attempted = 0;
2960                 bool raise_priority = true;
2961                 bool pgdat_needs_compaction = (order > 0);
2962
2963                 sc.nr_reclaimed = 0;
2964
2965                 /*
2966                  * Scan in the highmem->dma direction for the highest
2967                  * zone which needs scanning
2968                  */
2969                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2970                         struct zone *zone = pgdat->node_zones + i;
2971
2972                         if (!populated_zone(zone))
2973                                 continue;
2974
2975                         if (sc.priority != DEF_PRIORITY &&
2976                             !zone_reclaimable(zone))
2977                                 continue;
2978
2979                         /*
2980                          * Do some background aging of the anon list, to give
2981                          * pages a chance to be referenced before reclaiming.
2982                          */
2983                         age_active_anon(zone, &sc);
2984
2985                         /*
2986                          * If the number of buffer_heads in the machine
2987                          * exceeds the maximum allowed level and this node
2988                          * has a highmem zone, force kswapd to reclaim from
2989                          * it to relieve lowmem pressure.
2990                          */
2991                         if (buffer_heads_over_limit && is_highmem_idx(i)) {
2992                                 end_zone = i;
2993                                 break;
2994                         }
2995
2996                         if (!zone_balanced(zone, order, 0, 0)) {
2997                                 end_zone = i;
2998                                 break;
2999                         } else {
3000                                 /*
3001                                  * If balanced, clear the dirty and congested
3002                                  * flags
3003                                  */
3004                                 zone_clear_flag(zone, ZONE_CONGESTED);
3005                                 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
3006                         }
3007                 }
3008
3009                 if (i < 0)
3010                         goto out;
3011
3012                 for (i = 0; i <= end_zone; i++) {
3013                         struct zone *zone = pgdat->node_zones + i;
3014
3015                         if (!populated_zone(zone))
3016                                 continue;
3017
3018                         lru_pages += zone_reclaimable_pages(zone);
3019
3020                         /*
3021                          * If any zone is currently balanced then kswapd will
3022                          * not call compaction as it is expected that the
3023                          * necessary pages are already available.
3024                          */
3025                         if (pgdat_needs_compaction &&
3026                                         zone_watermark_ok(zone, order,
3027                                                 low_wmark_pages(zone),
3028                                                 *classzone_idx, 0))
3029                                 pgdat_needs_compaction = false;
3030                 }
3031
3032                 /*
3033                  * If we're getting trouble reclaiming, start doing writepage
3034                  * even in laptop mode.
3035                  */
3036                 if (sc.priority < DEF_PRIORITY - 2)
3037                         sc.may_writepage = 1;
3038
3039                 /*
3040                  * Now scan the zone in the dma->highmem direction, stopping
3041                  * at the last zone which needs scanning.
3042                  *
3043                  * We do this because the page allocator works in the opposite
3044                  * direction.  This prevents the page allocator from allocating
3045                  * pages behind kswapd's direction of progress, which would
3046                  * cause too much scanning of the lower zones.
3047                  */
3048                 for (i = 0; i <= end_zone; i++) {
3049                         struct zone *zone = pgdat->node_zones + i;
3050
3051                         if (!populated_zone(zone))
3052                                 continue;
3053
3054                         if (sc.priority != DEF_PRIORITY &&
3055                             !zone_reclaimable(zone))
3056                                 continue;
3057
3058                         sc.nr_scanned = 0;
3059
3060                         nr_soft_scanned = 0;
3061                         /*
3062                          * Call soft limit reclaim before calling shrink_zone.
3063                          */
3064                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3065                                                         order, sc.gfp_mask,
3066                                                         &nr_soft_scanned);
3067                         sc.nr_reclaimed += nr_soft_reclaimed;
3068
3069                         /*
3070                          * There should be no need to raise the scanning
3071                          * priority if enough pages are already being scanned
3072                          * that that high watermark would be met at 100%
3073                          * efficiency.
3074                          */
3075                         if (kswapd_shrink_zone(zone, end_zone, &sc,
3076                                         lru_pages, &nr_attempted))
3077                                 raise_priority = false;
3078                 }
3079
3080                 /*
3081                  * If the low watermark is met there is no need for processes
3082                  * to be throttled on pfmemalloc_wait as they should not be
3083                  * able to safely make forward progress. Wake them
3084                  */
3085                 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3086                                 pfmemalloc_watermark_ok(pgdat))
3087                         wake_up(&pgdat->pfmemalloc_wait);
3088
3089                 /*
3090                  * Fragmentation may mean that the system cannot be rebalanced
3091                  * for high-order allocations in all zones. If twice the
3092                  * allocation size has been reclaimed and the zones are still
3093                  * not balanced then recheck the watermarks at order-0 to
3094                  * prevent kswapd reclaiming excessively. Assume that a
3095                  * process requested a high-order can direct reclaim/compact.
3096                  */
3097                 if (order && sc.nr_reclaimed >= 2UL << order)
3098                         order = sc.order = 0;
3099
3100                 /* Check if kswapd should be suspending */
3101                 if (try_to_freeze() || kthread_should_stop())
3102                         break;
3103
3104                 /*
3105                  * Compact if necessary and kswapd is reclaiming at least the
3106                  * high watermark number of pages as requsted
3107                  */
3108                 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3109                         compact_pgdat(pgdat, order);
3110
3111                 /*
3112                  * Raise priority if scanning rate is too low or there was no
3113                  * progress in reclaiming pages
3114                  */
3115                 if (raise_priority || !sc.nr_reclaimed)
3116                         sc.priority--;
3117         } while (sc.priority >= 1 &&
3118                  !pgdat_balanced(pgdat, order, *classzone_idx));
3119
3120 out:
3121         /*
3122          * Return the order we were reclaiming at so prepare_kswapd_sleep()
3123          * makes a decision on the order we were last reclaiming at. However,
3124          * if another caller entered the allocator slow path while kswapd
3125          * was awake, order will remain at the higher level
3126          */
3127         *classzone_idx = end_zone;
3128         return order;
3129 }
3130
3131 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
3132 {
3133         long remaining = 0;
3134         DEFINE_WAIT(wait);
3135
3136         if (freezing(current) || kthread_should_stop())
3137                 return;
3138
3139         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3140
3141         /* Try to sleep for a short interval */
3142         if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3143                 remaining = schedule_timeout(HZ/10);
3144                 finish_wait(&pgdat->kswapd_wait, &wait);
3145                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3146         }
3147
3148         /*
3149          * After a short sleep, check if it was a premature sleep. If not, then
3150          * go fully to sleep until explicitly woken up.
3151          */
3152         if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
3153                 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3154
3155                 /*
3156                  * vmstat counters are not perfectly accurate and the estimated
3157                  * value for counters such as NR_FREE_PAGES can deviate from the
3158                  * true value by nr_online_cpus * threshold. To avoid the zone
3159                  * watermarks being breached while under pressure, we reduce the
3160                  * per-cpu vmstat threshold while kswapd is awake and restore
3161                  * them before going back to sleep.
3162                  */
3163                 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
3164
3165                 /*
3166                  * Compaction records what page blocks it recently failed to
3167                  * isolate pages from and skips them in the future scanning.
3168                  * When kswapd is going to sleep, it is reasonable to assume
3169                  * that pages and compaction may succeed so reset the cache.
3170                  */
3171                 reset_isolation_suitable(pgdat);
3172
3173                 if (!kthread_should_stop())
3174                         schedule();
3175
3176                 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3177         } else {
3178                 if (remaining)
3179                         count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3180                 else
3181                         count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3182         }
3183         finish_wait(&pgdat->kswapd_wait, &wait);
3184 }
3185
3186 /*
3187  * The background pageout daemon, started as a kernel thread
3188  * from the init process.
3189  *
3190  * This basically trickles out pages so that we have _some_
3191  * free memory available even if there is no other activity
3192  * that frees anything up. This is needed for things like routing
3193  * etc, where we otherwise might have all activity going on in
3194  * asynchronous contexts that cannot page things out.
3195  *
3196  * If there are applications that are active memory-allocators
3197  * (most normal use), this basically shouldn't matter.
3198  */
3199 static int kswapd(void *p)
3200 {
3201         unsigned long order, new_order;
3202         unsigned balanced_order;
3203         int classzone_idx, new_classzone_idx;
3204         int balanced_classzone_idx;
3205         pg_data_t *pgdat = (pg_data_t*)p;
3206         struct task_struct *tsk = current;
3207
3208         struct reclaim_state reclaim_state = {
3209                 .reclaimed_slab = 0,
3210         };
3211         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3212
3213         lockdep_set_current_reclaim_state(GFP_KERNEL);
3214
3215         if (!cpumask_empty(cpumask))
3216                 set_cpus_allowed_ptr(tsk, cpumask);
3217         current->reclaim_state = &reclaim_state;
3218
3219         /*
3220          * Tell the memory management that we're a "memory allocator",
3221          * and that if we need more memory we should get access to it
3222          * regardless (see "__alloc_pages()"). "kswapd" should
3223          * never get caught in the normal page freeing logic.
3224          *
3225          * (Kswapd normally doesn't need memory anyway, but sometimes
3226          * you need a small amount of memory in order to be able to
3227          * page out something else, and this flag essentially protects
3228          * us from recursively trying to free more memory as we're
3229          * trying to free the first piece of memory in the first place).
3230          */
3231         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
3232         set_freezable();
3233
3234         order = new_order = 0;
3235         balanced_order = 0;
3236         classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
3237         balanced_classzone_idx = classzone_idx;
3238         for ( ; ; ) {
3239                 bool ret;
3240
3241                 /*
3242                  * If the last balance_pgdat was unsuccessful it's unlikely a
3243                  * new request of a similar or harder type will succeed soon
3244                  * so consider going to sleep on the basis we reclaimed at
3245                  */
3246                 if (balanced_classzone_idx >= new_classzone_idx &&
3247                                         balanced_order == new_order) {
3248                         new_order = pgdat->kswapd_max_order;
3249                         new_classzone_idx = pgdat->classzone_idx;
3250                         pgdat->kswapd_max_order =  0;
3251                         pgdat->classzone_idx = pgdat->nr_zones - 1;
3252                 }
3253
3254                 if (order < new_order || classzone_idx > new_classzone_idx) {
3255                         /*
3256                          * Don't sleep if someone wants a larger 'order'
3257                          * allocation or has tigher zone constraints
3258                          */
3259                         order = new_order;
3260                         classzone_idx = new_classzone_idx;
3261                 } else {
3262                         kswapd_try_to_sleep(pgdat, balanced_order,
3263                                                 balanced_classzone_idx);
3264                         order = pgdat->kswapd_max_order;
3265                         classzone_idx = pgdat->classzone_idx;
3266                         new_order = order;
3267                         new_classzone_idx = classzone_idx;
3268                         pgdat->kswapd_max_order = 0;
3269                         pgdat->classzone_idx = pgdat->nr_zones - 1;
3270                 }
3271
3272                 ret = try_to_freeze();
3273                 if (kthread_should_stop())
3274                         break;
3275
3276                 /*
3277                  * We can speed up thawing tasks if we don't call balance_pgdat
3278                  * after returning from the refrigerator
3279                  */
3280                 if (!ret) {
3281                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3282                         balanced_classzone_idx = classzone_idx;
3283                         balanced_order = balance_pgdat(pgdat, order,
3284                                                 &balanced_classzone_idx);
3285                 }
3286         }
3287
3288         current->reclaim_state = NULL;
3289         return 0;
3290 }
3291
3292 /*
3293  * A zone is low on free memory, so wake its kswapd task to service it.
3294  */
3295 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3296 {
3297         pg_data_t *pgdat;
3298
3299         if (!populated_zone(zone))
3300                 return;
3301
3302         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3303                 return;
3304         pgdat = zone->zone_pgdat;
3305         if (pgdat->kswapd_max_order < order) {
3306                 pgdat->kswapd_max_order = order;
3307                 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3308         }
3309         if (!waitqueue_active(&pgdat->kswapd_wait))
3310                 return;
3311         if (zone_balanced(zone, order, 0, 0))
3312                 return;
3313
3314         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3315         wake_up_interruptible(&pgdat->kswapd_wait);
3316 }
3317
3318 /*
3319  * The reclaimable count would be mostly accurate.
3320  * The less reclaimable pages may be
3321  * - mlocked pages, which will be moved to unevictable list when encountered
3322  * - mapped pages, which may require several travels to be reclaimed
3323  * - dirty pages, which is not "instantly" reclaimable
3324  */
3325 unsigned long global_reclaimable_pages(void)
3326 {
3327         int nr;
3328
3329         nr = global_page_state(NR_ACTIVE_FILE) +
3330              global_page_state(NR_INACTIVE_FILE);
3331
3332         if (get_nr_swap_pages() > 0)
3333                 nr += global_page_state(NR_ACTIVE_ANON) +
3334                       global_page_state(NR_INACTIVE_ANON);
3335
3336         return nr;
3337 }
3338
3339 #ifdef CONFIG_HIBERNATION
3340 /*
3341  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3342  * freed pages.
3343  *
3344  * Rather than trying to age LRUs the aim is to preserve the overall
3345  * LRU order by reclaiming preferentially
3346  * inactive > active > active referenced > active mapped
3347  */
3348 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3349 {
3350         struct reclaim_state reclaim_state;
3351         struct scan_control sc = {
3352                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3353                 .may_swap = 1,
3354                 .may_unmap = 1,
3355                 .may_writepage = 1,
3356                 .nr_to_reclaim = nr_to_reclaim,
3357                 .hibernation_mode = 1,
3358                 .order = 0,
3359                 .priority = DEF_PRIORITY,
3360         };
3361         struct shrink_control shrink = {
3362                 .gfp_mask = sc.gfp_mask,
3363         };
3364         struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3365         struct task_struct *p = current;
3366         unsigned long nr_reclaimed;
3367
3368         p->flags |= PF_MEMALLOC;
3369         lockdep_set_current_reclaim_state(sc.gfp_mask);
3370         reclaim_state.reclaimed_slab = 0;
3371         p->reclaim_state = &reclaim_state;
3372
3373         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3374
3375         p->reclaim_state = NULL;
3376         lockdep_clear_current_reclaim_state();
3377         p->flags &= ~PF_MEMALLOC;
3378
3379         return nr_reclaimed;
3380 }
3381 #endif /* CONFIG_HIBERNATION */
3382
3383 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3384    not required for correctness.  So if the last cpu in a node goes
3385    away, we get changed to run anywhere: as the first one comes back,
3386    restore their cpu bindings. */
3387 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3388                         void *hcpu)
3389 {
3390         int nid;
3391
3392         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3393                 for_each_node_state(nid, N_MEMORY) {
3394                         pg_data_t *pgdat = NODE_DATA(nid);
3395                         const struct cpumask *mask;
3396
3397                         mask = cpumask_of_node(pgdat->node_id);
3398
3399                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3400                                 /* One of our CPUs online: restore mask */
3401                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3402                 }
3403         }
3404         return NOTIFY_OK;
3405 }
3406
3407 /*
3408  * This kswapd start function will be called by init and node-hot-add.
3409  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3410  */
3411 int kswapd_run(int nid)
3412 {
3413         pg_data_t *pgdat = NODE_DATA(nid);
3414         int ret = 0;
3415
3416         if (pgdat->kswapd)
3417                 return 0;
3418
3419         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3420         if (IS_ERR(pgdat->kswapd)) {
3421                 /* failure at boot is fatal */
3422                 BUG_ON(system_state == SYSTEM_BOOTING);
3423                 pr_err("Failed to start kswapd on node %d\n", nid);
3424                 ret = PTR_ERR(pgdat->kswapd);
3425                 pgdat->kswapd = NULL;
3426         }
3427         return ret;
3428 }
3429
3430 /*
3431  * Called by memory hotplug when all memory in a node is offlined.  Caller must
3432  * hold lock_memory_hotplug().
3433  */
3434 void kswapd_stop(int nid)
3435 {
3436         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3437
3438         if (kswapd) {
3439                 kthread_stop(kswapd);
3440                 NODE_DATA(nid)->kswapd = NULL;
3441         }
3442 }
3443
3444 static int __init kswapd_init(void)
3445 {
3446         int nid;
3447
3448         swap_setup();
3449         for_each_node_state(nid, N_MEMORY)
3450                 kswapd_run(nid);
3451         hotcpu_notifier(cpu_callback, 0);
3452         return 0;
3453 }
3454
3455 module_init(kswapd_init)
3456
3457 #ifdef CONFIG_NUMA
3458 /*
3459  * Zone reclaim mode
3460  *
3461  * If non-zero call zone_reclaim when the number of free pages falls below
3462  * the watermarks.
3463  */
3464 int zone_reclaim_mode __read_mostly;
3465
3466 #define RECLAIM_OFF 0
3467 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
3468 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
3469 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
3470
3471 /*
3472  * Priority for ZONE_RECLAIM. This determines the fraction of pages
3473  * of a node considered for each zone_reclaim. 4 scans 1/16th of
3474  * a zone.
3475  */
3476 #define ZONE_RECLAIM_PRIORITY 4
3477
3478 /*
3479  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3480  * occur.
3481  */
3482 int sysctl_min_unmapped_ratio = 1;
3483
3484 /*
3485  * If the number of slab pages in a zone grows beyond this percentage then
3486  * slab reclaim needs to occur.
3487  */
3488 int sysctl_min_slab_ratio = 5;
3489
3490 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3491 {
3492         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3493         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3494                 zone_page_state(zone, NR_ACTIVE_FILE);
3495
3496         /*
3497          * It's possible for there to be more file mapped pages than
3498          * accounted for by the pages on the file LRU lists because
3499          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3500          */
3501         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3502 }
3503
3504 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3505 static long zone_pagecache_reclaimable(struct zone *zone)
3506 {
3507         long nr_pagecache_reclaimable;
3508         long delta = 0;
3509
3510         /*
3511          * If RECLAIM_SWAP is set, then all file pages are considered
3512          * potentially reclaimable. Otherwise, we have to worry about
3513          * pages like swapcache and zone_unmapped_file_pages() provides
3514          * a better estimate
3515          */
3516         if (zone_reclaim_mode & RECLAIM_SWAP)
3517                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3518         else
3519                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3520
3521         /* If we can't clean pages, remove dirty pages from consideration */
3522         if (!(zone_reclaim_mode & RECLAIM_WRITE))
3523                 delta += zone_page_state(zone, NR_FILE_DIRTY);
3524
3525         /* Watch for any possible underflows due to delta */
3526         if (unlikely(delta > nr_pagecache_reclaimable))
3527                 delta = nr_pagecache_reclaimable;
3528
3529         return nr_pagecache_reclaimable - delta;
3530 }
3531
3532 /*
3533  * Try to free up some pages from this zone through reclaim.
3534  */
3535 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3536 {
3537         /* Minimum pages needed in order to stay on node */
3538         const unsigned long nr_pages = 1 << order;
3539         struct task_struct *p = current;
3540         struct reclaim_state reclaim_state;
3541         struct scan_control sc = {
3542                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3543                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3544                 .may_swap = 1,
3545                 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3546                 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3547                 .order = order,
3548                 .priority = ZONE_RECLAIM_PRIORITY,
3549         };
3550         struct shrink_control shrink = {
3551                 .gfp_mask = sc.gfp_mask,
3552         };
3553         unsigned long nr_slab_pages0, nr_slab_pages1;
3554
3555         cond_resched();
3556         /*
3557          * We need to be able to allocate from the reserves for RECLAIM_SWAP
3558          * and we also need to be able to write out pages for RECLAIM_WRITE
3559          * and RECLAIM_SWAP.
3560          */
3561         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3562         lockdep_set_current_reclaim_state(gfp_mask);
3563         reclaim_state.reclaimed_slab = 0;
3564         p->reclaim_state = &reclaim_state;
3565
3566         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3567                 /*
3568                  * Free memory by calling shrink zone with increasing
3569                  * priorities until we have enough memory freed.
3570                  */
3571                 do {
3572                         shrink_zone(zone, &sc);
3573                 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3574         }
3575
3576         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3577         if (nr_slab_pages0 > zone->min_slab_pages) {
3578                 /*
3579                  * shrink_slab() does not currently allow us to determine how
3580                  * many pages were freed in this zone. So we take the current
3581                  * number of slab pages and shake the slab until it is reduced
3582                  * by the same nr_pages that we used for reclaiming unmapped
3583                  * pages.
3584                  */
3585                 nodes_clear(shrink.nodes_to_scan);
3586                 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
3587                 for (;;) {
3588                         unsigned long lru_pages = zone_reclaimable_pages(zone);
3589
3590                         /* No reclaimable slab or very low memory pressure */
3591                         if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3592                                 break;
3593
3594                         /* Freed enough memory */
3595                         nr_slab_pages1 = zone_page_state(zone,
3596                                                         NR_SLAB_RECLAIMABLE);
3597                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3598                                 break;
3599                 }
3600
3601                 /*
3602                  * Update nr_reclaimed by the number of slab pages we
3603                  * reclaimed from this zone.
3604                  */
3605                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3606                 if (nr_slab_pages1 < nr_slab_pages0)
3607                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3608         }
3609
3610         p->reclaim_state = NULL;
3611         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3612         lockdep_clear_current_reclaim_state();
3613         return sc.nr_reclaimed >= nr_pages;
3614 }
3615
3616 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3617 {
3618         int node_id;
3619         int ret;
3620
3621         /*
3622          * Zone reclaim reclaims unmapped file backed pages and
3623          * slab pages if we are over the defined limits.
3624          *
3625          * A small portion of unmapped file backed pages is needed for
3626          * file I/O otherwise pages read by file I/O will be immediately
3627          * thrown out if the zone is overallocated. So we do not reclaim
3628          * if less than a specified percentage of the zone is used by
3629          * unmapped file backed pages.
3630          */
3631         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3632             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3633                 return ZONE_RECLAIM_FULL;
3634
3635         if (!zone_reclaimable(zone))
3636                 return ZONE_RECLAIM_FULL;
3637
3638         /*
3639          * Do not scan if the allocation should not be delayed.
3640          */
3641         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3642                 return ZONE_RECLAIM_NOSCAN;
3643
3644         /*
3645          * Only run zone reclaim on the local zone or on zones that do not
3646          * have associated processors. This will favor the local processor
3647          * over remote processors and spread off node memory allocations
3648          * as wide as possible.
3649          */
3650         node_id = zone_to_nid(zone);
3651         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3652                 return ZONE_RECLAIM_NOSCAN;
3653
3654         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3655                 return ZONE_RECLAIM_NOSCAN;
3656
3657         ret = __zone_reclaim(zone, gfp_mask, order);
3658         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3659
3660         if (!ret)
3661                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3662
3663         return ret;
3664 }
3665 #endif
3666
3667 /*
3668  * page_evictable - test whether a page is evictable
3669  * @page: the page to test
3670  *
3671  * Test whether page is evictable--i.e., should be placed on active/inactive
3672  * lists vs unevictable list.
3673  *
3674  * Reasons page might not be evictable:
3675  * (1) page's mapping marked unevictable
3676  * (2) page is part of an mlocked VMA
3677  *
3678  */
3679 int page_evictable(struct page *page)
3680 {
3681         return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3682 }
3683
3684 #ifdef CONFIG_SHMEM
3685 /**
3686  * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3687  * @pages:      array of pages to check
3688  * @nr_pages:   number of pages to check
3689  *
3690  * Checks pages for evictability and moves them to the appropriate lru list.
3691  *
3692  * This function is only used for SysV IPC SHM_UNLOCK.
3693  */
3694 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3695 {
3696         struct lruvec *lruvec;
3697         struct zone *zone = NULL;
3698         int pgscanned = 0;
3699         int pgrescued = 0;
3700         int i;
3701
3702         for (i = 0; i < nr_pages; i++) {
3703                 struct page *page = pages[i];
3704                 struct zone *pagezone;
3705
3706                 pgscanned++;
3707                 pagezone = page_zone(page);
3708                 if (pagezone != zone) {
3709                         if (zone)
3710                                 spin_unlock_irq(&zone->lru_lock);
3711                         zone = pagezone;
3712                         spin_lock_irq(&zone->lru_lock);
3713                 }
3714                 lruvec = mem_cgroup_page_lruvec(page, zone);
3715
3716                 if (!PageLRU(page) || !PageUnevictable(page))
3717                         continue;
3718
3719                 if (page_evictable(page)) {
3720                         enum lru_list lru = page_lru_base_type(page);
3721
3722                         VM_BUG_ON_PAGE(PageActive(page), page);
3723                         ClearPageUnevictable(page);
3724                         del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3725                         add_page_to_lru_list(page, lruvec, lru);
3726                         pgrescued++;
3727                 }
3728         }
3729
3730         if (zone) {
3731                 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3732                 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3733                 spin_unlock_irq(&zone->lru_lock);
3734         }
3735 }
3736 #endif /* CONFIG_SHMEM */
3737
3738 static void warn_scan_unevictable_pages(void)
3739 {
3740         printk_once(KERN_WARNING
3741                     "%s: The scan_unevictable_pages sysctl/node-interface has been "
3742                     "disabled for lack of a legitimate use case.  If you have "
3743                     "one, please send an email to linux-mm@kvack.org.\n",
3744                     current->comm);
3745 }
3746
3747 /*
3748  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3749  * all nodes' unevictable lists for evictable pages
3750  */
3751 unsigned long scan_unevictable_pages;
3752
3753 int scan_unevictable_handler(struct ctl_table *table, int write,
3754                            void __user *buffer,
3755                            size_t *length, loff_t *ppos)
3756 {
3757         warn_scan_unevictable_pages();
3758         proc_doulongvec_minmax(table, write, buffer, length, ppos);
3759         scan_unevictable_pages = 0;
3760         return 0;
3761 }
3762
3763 #ifdef CONFIG_NUMA
3764 /*
3765  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3766  * a specified node's per zone unevictable lists for evictable pages.
3767  */
3768
3769 static ssize_t read_scan_unevictable_node(struct device *dev,
3770                                           struct device_attribute *attr,
3771                                           char *buf)
3772 {
3773         warn_scan_unevictable_pages();
3774         return sprintf(buf, "0\n");     /* always zero; should fit... */
3775 }
3776
3777 static ssize_t write_scan_unevictable_node(struct device *dev,
3778                                            struct device_attribute *attr,
3779                                         const char *buf, size_t count)
3780 {
3781         warn_scan_unevictable_pages();
3782         return 1;
3783 }
3784
3785
3786 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3787                         read_scan_unevictable_node,
3788                         write_scan_unevictable_node);
3789
3790 int scan_unevictable_register_node(struct node *node)
3791 {
3792         return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3793 }
3794
3795 void scan_unevictable_unregister_node(struct node *node)
3796 {
3797         device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
3798 }
3799 #endif