idle: Remove GENERIC_IDLE_LOOP config switch
[linux-3.10.git] / mm / vmscan.c
1 /*
2  *  linux/mm/vmscan.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *
6  *  Swap reorganised 29.12.95, Stephen Tweedie.
7  *  kswapd added: 7.1.96  sct
8  *  Removed kswapd_ctl limits, and swap out as many pages as needed
9  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11  *  Multiqueue VM started 5.8.00, Rik van Riel.
12  */
13
14 #include <linux/mm.h>
15 #include <linux/module.h>
16 #include <linux/gfp.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h>  /* for try_to_release_page(),
27                                         buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/backing-dev.h>
30 #include <linux/rmap.h>
31 #include <linux/topology.h>
32 #include <linux/cpu.h>
33 #include <linux/cpuset.h>
34 #include <linux/compaction.h>
35 #include <linux/notifier.h>
36 #include <linux/rwsem.h>
37 #include <linux/delay.h>
38 #include <linux/kthread.h>
39 #include <linux/freezer.h>
40 #include <linux/memcontrol.h>
41 #include <linux/delayacct.h>
42 #include <linux/sysctl.h>
43 #include <linux/oom.h>
44 #include <linux/prefetch.h>
45
46 #include <asm/tlbflush.h>
47 #include <asm/div64.h>
48
49 #include <linux/swapops.h>
50
51 #include "internal.h"
52
53 #define CREATE_TRACE_POINTS
54 #include <trace/events/vmscan.h>
55
56 struct scan_control {
57         /* Incremented by the number of inactive pages that were scanned */
58         unsigned long nr_scanned;
59
60         /* Number of pages freed so far during a call to shrink_zones() */
61         unsigned long nr_reclaimed;
62
63         /* How many pages shrink_list() should reclaim */
64         unsigned long nr_to_reclaim;
65
66         unsigned long hibernation_mode;
67
68         /* This context's GFP mask */
69         gfp_t gfp_mask;
70
71         int may_writepage;
72
73         /* Can mapped pages be reclaimed? */
74         int may_unmap;
75
76         /* Can pages be swapped as part of reclaim? */
77         int may_swap;
78
79         int order;
80
81         /* Scan (total_size >> priority) pages at once */
82         int priority;
83
84         /*
85          * The memory cgroup that hit its limit and as a result is the
86          * primary target of this reclaim invocation.
87          */
88         struct mem_cgroup *target_mem_cgroup;
89
90         /*
91          * Nodemask of nodes allowed by the caller. If NULL, all nodes
92          * are scanned.
93          */
94         nodemask_t      *nodemask;
95 };
96
97 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
98
99 #ifdef ARCH_HAS_PREFETCH
100 #define prefetch_prev_lru_page(_page, _base, _field)                    \
101         do {                                                            \
102                 if ((_page)->lru.prev != _base) {                       \
103                         struct page *prev;                              \
104                                                                         \
105                         prev = lru_to_page(&(_page->lru));              \
106                         prefetch(&prev->_field);                        \
107                 }                                                       \
108         } while (0)
109 #else
110 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
111 #endif
112
113 #ifdef ARCH_HAS_PREFETCHW
114 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
115         do {                                                            \
116                 if ((_page)->lru.prev != _base) {                       \
117                         struct page *prev;                              \
118                                                                         \
119                         prev = lru_to_page(&(_page->lru));              \
120                         prefetchw(&prev->_field);                       \
121                 }                                                       \
122         } while (0)
123 #else
124 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
125 #endif
126
127 /*
128  * From 0 .. 100.  Higher means more swappy.
129  */
130 int vm_swappiness = 60;
131 unsigned long vm_total_pages;   /* The total number of pages which the VM controls */
132
133 static LIST_HEAD(shrinker_list);
134 static DECLARE_RWSEM(shrinker_rwsem);
135
136 #ifdef CONFIG_MEMCG
137 static bool global_reclaim(struct scan_control *sc)
138 {
139         return !sc->target_mem_cgroup;
140 }
141 #else
142 static bool global_reclaim(struct scan_control *sc)
143 {
144         return true;
145 }
146 #endif
147
148 static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
149 {
150         if (!mem_cgroup_disabled())
151                 return mem_cgroup_get_lru_size(lruvec, lru);
152
153         return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
154 }
155
156 /*
157  * Add a shrinker callback to be called from the vm
158  */
159 void register_shrinker(struct shrinker *shrinker)
160 {
161         atomic_long_set(&shrinker->nr_in_batch, 0);
162         down_write(&shrinker_rwsem);
163         list_add_tail(&shrinker->list, &shrinker_list);
164         up_write(&shrinker_rwsem);
165 }
166 EXPORT_SYMBOL(register_shrinker);
167
168 /*
169  * Remove one
170  */
171 void unregister_shrinker(struct shrinker *shrinker)
172 {
173         down_write(&shrinker_rwsem);
174         list_del(&shrinker->list);
175         up_write(&shrinker_rwsem);
176 }
177 EXPORT_SYMBOL(unregister_shrinker);
178
179 static inline int do_shrinker_shrink(struct shrinker *shrinker,
180                                      struct shrink_control *sc,
181                                      unsigned long nr_to_scan)
182 {
183         sc->nr_to_scan = nr_to_scan;
184         return (*shrinker->shrink)(shrinker, sc);
185 }
186
187 #define SHRINK_BATCH 128
188 /*
189  * Call the shrink functions to age shrinkable caches
190  *
191  * Here we assume it costs one seek to replace a lru page and that it also
192  * takes a seek to recreate a cache object.  With this in mind we age equal
193  * percentages of the lru and ageable caches.  This should balance the seeks
194  * generated by these structures.
195  *
196  * If the vm encountered mapped pages on the LRU it increase the pressure on
197  * slab to avoid swapping.
198  *
199  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
200  *
201  * `lru_pages' represents the number of on-LRU pages in all the zones which
202  * are eligible for the caller's allocation attempt.  It is used for balancing
203  * slab reclaim versus page reclaim.
204  *
205  * Returns the number of slab objects which we shrunk.
206  */
207 unsigned long shrink_slab(struct shrink_control *shrink,
208                           unsigned long nr_pages_scanned,
209                           unsigned long lru_pages)
210 {
211         struct shrinker *shrinker;
212         unsigned long ret = 0;
213
214         if (nr_pages_scanned == 0)
215                 nr_pages_scanned = SWAP_CLUSTER_MAX;
216
217         if (!down_read_trylock(&shrinker_rwsem)) {
218                 /* Assume we'll be able to shrink next time */
219                 ret = 1;
220                 goto out;
221         }
222
223         list_for_each_entry(shrinker, &shrinker_list, list) {
224                 unsigned long long delta;
225                 long total_scan;
226                 long max_pass;
227                 int shrink_ret = 0;
228                 long nr;
229                 long new_nr;
230                 long batch_size = shrinker->batch ? shrinker->batch
231                                                   : SHRINK_BATCH;
232
233                 max_pass = do_shrinker_shrink(shrinker, shrink, 0);
234                 if (max_pass <= 0)
235                         continue;
236
237                 /*
238                  * copy the current shrinker scan count into a local variable
239                  * and zero it so that other concurrent shrinker invocations
240                  * don't also do this scanning work.
241                  */
242                 nr = atomic_long_xchg(&shrinker->nr_in_batch, 0);
243
244                 total_scan = nr;
245                 delta = (4 * nr_pages_scanned) / shrinker->seeks;
246                 delta *= max_pass;
247                 do_div(delta, lru_pages + 1);
248                 total_scan += delta;
249                 if (total_scan < 0) {
250                         printk(KERN_ERR "shrink_slab: %pF negative objects to "
251                                "delete nr=%ld\n",
252                                shrinker->shrink, total_scan);
253                         total_scan = max_pass;
254                 }
255
256                 /*
257                  * We need to avoid excessive windup on filesystem shrinkers
258                  * due to large numbers of GFP_NOFS allocations causing the
259                  * shrinkers to return -1 all the time. This results in a large
260                  * nr being built up so when a shrink that can do some work
261                  * comes along it empties the entire cache due to nr >>>
262                  * max_pass.  This is bad for sustaining a working set in
263                  * memory.
264                  *
265                  * Hence only allow the shrinker to scan the entire cache when
266                  * a large delta change is calculated directly.
267                  */
268                 if (delta < max_pass / 4)
269                         total_scan = min(total_scan, max_pass / 2);
270
271                 /*
272                  * Avoid risking looping forever due to too large nr value:
273                  * never try to free more than twice the estimate number of
274                  * freeable entries.
275                  */
276                 if (total_scan > max_pass * 2)
277                         total_scan = max_pass * 2;
278
279                 trace_mm_shrink_slab_start(shrinker, shrink, nr,
280                                         nr_pages_scanned, lru_pages,
281                                         max_pass, delta, total_scan);
282
283                 while (total_scan >= batch_size) {
284                         int nr_before;
285
286                         nr_before = do_shrinker_shrink(shrinker, shrink, 0);
287                         shrink_ret = do_shrinker_shrink(shrinker, shrink,
288                                                         batch_size);
289                         if (shrink_ret == -1)
290                                 break;
291                         if (shrink_ret < nr_before)
292                                 ret += nr_before - shrink_ret;
293                         count_vm_events(SLABS_SCANNED, batch_size);
294                         total_scan -= batch_size;
295
296                         cond_resched();
297                 }
298
299                 /*
300                  * move the unused scan count back into the shrinker in a
301                  * manner that handles concurrent updates. If we exhausted the
302                  * scan, there is no need to do an update.
303                  */
304                 if (total_scan > 0)
305                         new_nr = atomic_long_add_return(total_scan,
306                                         &shrinker->nr_in_batch);
307                 else
308                         new_nr = atomic_long_read(&shrinker->nr_in_batch);
309
310                 trace_mm_shrink_slab_end(shrinker, shrink_ret, nr, new_nr);
311         }
312         up_read(&shrinker_rwsem);
313 out:
314         cond_resched();
315         return ret;
316 }
317
318 static inline int is_page_cache_freeable(struct page *page)
319 {
320         /*
321          * A freeable page cache page is referenced only by the caller
322          * that isolated the page, the page cache radix tree and
323          * optional buffer heads at page->private.
324          */
325         return page_count(page) - page_has_private(page) == 2;
326 }
327
328 static int may_write_to_queue(struct backing_dev_info *bdi,
329                               struct scan_control *sc)
330 {
331         if (current->flags & PF_SWAPWRITE)
332                 return 1;
333         if (!bdi_write_congested(bdi))
334                 return 1;
335         if (bdi == current->backing_dev_info)
336                 return 1;
337         return 0;
338 }
339
340 /*
341  * We detected a synchronous write error writing a page out.  Probably
342  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
343  * fsync(), msync() or close().
344  *
345  * The tricky part is that after writepage we cannot touch the mapping: nothing
346  * prevents it from being freed up.  But we have a ref on the page and once
347  * that page is locked, the mapping is pinned.
348  *
349  * We're allowed to run sleeping lock_page() here because we know the caller has
350  * __GFP_FS.
351  */
352 static void handle_write_error(struct address_space *mapping,
353                                 struct page *page, int error)
354 {
355         lock_page(page);
356         if (page_mapping(page) == mapping)
357                 mapping_set_error(mapping, error);
358         unlock_page(page);
359 }
360
361 /* possible outcome of pageout() */
362 typedef enum {
363         /* failed to write page out, page is locked */
364         PAGE_KEEP,
365         /* move page to the active list, page is locked */
366         PAGE_ACTIVATE,
367         /* page has been sent to the disk successfully, page is unlocked */
368         PAGE_SUCCESS,
369         /* page is clean and locked */
370         PAGE_CLEAN,
371 } pageout_t;
372
373 /*
374  * pageout is called by shrink_page_list() for each dirty page.
375  * Calls ->writepage().
376  */
377 static pageout_t pageout(struct page *page, struct address_space *mapping,
378                          struct scan_control *sc)
379 {
380         /*
381          * If the page is dirty, only perform writeback if that write
382          * will be non-blocking.  To prevent this allocation from being
383          * stalled by pagecache activity.  But note that there may be
384          * stalls if we need to run get_block().  We could test
385          * PagePrivate for that.
386          *
387          * If this process is currently in __generic_file_aio_write() against
388          * this page's queue, we can perform writeback even if that
389          * will block.
390          *
391          * If the page is swapcache, write it back even if that would
392          * block, for some throttling. This happens by accident, because
393          * swap_backing_dev_info is bust: it doesn't reflect the
394          * congestion state of the swapdevs.  Easy to fix, if needed.
395          */
396         if (!is_page_cache_freeable(page))
397                 return PAGE_KEEP;
398         if (!mapping) {
399                 /*
400                  * Some data journaling orphaned pages can have
401                  * page->mapping == NULL while being dirty with clean buffers.
402                  */
403                 if (page_has_private(page)) {
404                         if (try_to_free_buffers(page)) {
405                                 ClearPageDirty(page);
406                                 printk("%s: orphaned page\n", __func__);
407                                 return PAGE_CLEAN;
408                         }
409                 }
410                 return PAGE_KEEP;
411         }
412         if (mapping->a_ops->writepage == NULL)
413                 return PAGE_ACTIVATE;
414         if (!may_write_to_queue(mapping->backing_dev_info, sc))
415                 return PAGE_KEEP;
416
417         if (clear_page_dirty_for_io(page)) {
418                 int res;
419                 struct writeback_control wbc = {
420                         .sync_mode = WB_SYNC_NONE,
421                         .nr_to_write = SWAP_CLUSTER_MAX,
422                         .range_start = 0,
423                         .range_end = LLONG_MAX,
424                         .for_reclaim = 1,
425                 };
426
427                 SetPageReclaim(page);
428                 res = mapping->a_ops->writepage(page, &wbc);
429                 if (res < 0)
430                         handle_write_error(mapping, page, res);
431                 if (res == AOP_WRITEPAGE_ACTIVATE) {
432                         ClearPageReclaim(page);
433                         return PAGE_ACTIVATE;
434                 }
435
436                 if (!PageWriteback(page)) {
437                         /* synchronous write or broken a_ops? */
438                         ClearPageReclaim(page);
439                 }
440                 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
441                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
442                 return PAGE_SUCCESS;
443         }
444
445         return PAGE_CLEAN;
446 }
447
448 /*
449  * Same as remove_mapping, but if the page is removed from the mapping, it
450  * gets returned with a refcount of 0.
451  */
452 static int __remove_mapping(struct address_space *mapping, struct page *page)
453 {
454         BUG_ON(!PageLocked(page));
455         BUG_ON(mapping != page_mapping(page));
456
457         spin_lock_irq(&mapping->tree_lock);
458         /*
459          * The non racy check for a busy page.
460          *
461          * Must be careful with the order of the tests. When someone has
462          * a ref to the page, it may be possible that they dirty it then
463          * drop the reference. So if PageDirty is tested before page_count
464          * here, then the following race may occur:
465          *
466          * get_user_pages(&page);
467          * [user mapping goes away]
468          * write_to(page);
469          *                              !PageDirty(page)    [good]
470          * SetPageDirty(page);
471          * put_page(page);
472          *                              !page_count(page)   [good, discard it]
473          *
474          * [oops, our write_to data is lost]
475          *
476          * Reversing the order of the tests ensures such a situation cannot
477          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
478          * load is not satisfied before that of page->_count.
479          *
480          * Note that if SetPageDirty is always performed via set_page_dirty,
481          * and thus under tree_lock, then this ordering is not required.
482          */
483         if (!page_freeze_refs(page, 2))
484                 goto cannot_free;
485         /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
486         if (unlikely(PageDirty(page))) {
487                 page_unfreeze_refs(page, 2);
488                 goto cannot_free;
489         }
490
491         if (PageSwapCache(page)) {
492                 swp_entry_t swap = { .val = page_private(page) };
493                 __delete_from_swap_cache(page);
494                 spin_unlock_irq(&mapping->tree_lock);
495                 swapcache_free(swap, page);
496         } else {
497                 void (*freepage)(struct page *);
498
499                 freepage = mapping->a_ops->freepage;
500
501                 __delete_from_page_cache(page);
502                 spin_unlock_irq(&mapping->tree_lock);
503                 mem_cgroup_uncharge_cache_page(page);
504
505                 if (freepage != NULL)
506                         freepage(page);
507         }
508
509         return 1;
510
511 cannot_free:
512         spin_unlock_irq(&mapping->tree_lock);
513         return 0;
514 }
515
516 /*
517  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
518  * someone else has a ref on the page, abort and return 0.  If it was
519  * successfully detached, return 1.  Assumes the caller has a single ref on
520  * this page.
521  */
522 int remove_mapping(struct address_space *mapping, struct page *page)
523 {
524         if (__remove_mapping(mapping, page)) {
525                 /*
526                  * Unfreezing the refcount with 1 rather than 2 effectively
527                  * drops the pagecache ref for us without requiring another
528                  * atomic operation.
529                  */
530                 page_unfreeze_refs(page, 1);
531                 return 1;
532         }
533         return 0;
534 }
535
536 /**
537  * putback_lru_page - put previously isolated page onto appropriate LRU list
538  * @page: page to be put back to appropriate lru list
539  *
540  * Add previously isolated @page to appropriate LRU list.
541  * Page may still be unevictable for other reasons.
542  *
543  * lru_lock must not be held, interrupts must be enabled.
544  */
545 void putback_lru_page(struct page *page)
546 {
547         int lru;
548         int active = !!TestClearPageActive(page);
549         int was_unevictable = PageUnevictable(page);
550
551         VM_BUG_ON(PageLRU(page));
552
553 redo:
554         ClearPageUnevictable(page);
555
556         if (page_evictable(page)) {
557                 /*
558                  * For evictable pages, we can use the cache.
559                  * In event of a race, worst case is we end up with an
560                  * unevictable page on [in]active list.
561                  * We know how to handle that.
562                  */
563                 lru = active + page_lru_base_type(page);
564                 lru_cache_add_lru(page, lru);
565         } else {
566                 /*
567                  * Put unevictable pages directly on zone's unevictable
568                  * list.
569                  */
570                 lru = LRU_UNEVICTABLE;
571                 add_page_to_unevictable_list(page);
572                 /*
573                  * When racing with an mlock or AS_UNEVICTABLE clearing
574                  * (page is unlocked) make sure that if the other thread
575                  * does not observe our setting of PG_lru and fails
576                  * isolation/check_move_unevictable_pages,
577                  * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
578                  * the page back to the evictable list.
579                  *
580                  * The other side is TestClearPageMlocked() or shmem_lock().
581                  */
582                 smp_mb();
583         }
584
585         /*
586          * page's status can change while we move it among lru. If an evictable
587          * page is on unevictable list, it never be freed. To avoid that,
588          * check after we added it to the list, again.
589          */
590         if (lru == LRU_UNEVICTABLE && page_evictable(page)) {
591                 if (!isolate_lru_page(page)) {
592                         put_page(page);
593                         goto redo;
594                 }
595                 /* This means someone else dropped this page from LRU
596                  * So, it will be freed or putback to LRU again. There is
597                  * nothing to do here.
598                  */
599         }
600
601         if (was_unevictable && lru != LRU_UNEVICTABLE)
602                 count_vm_event(UNEVICTABLE_PGRESCUED);
603         else if (!was_unevictable && lru == LRU_UNEVICTABLE)
604                 count_vm_event(UNEVICTABLE_PGCULLED);
605
606         put_page(page);         /* drop ref from isolate */
607 }
608
609 enum page_references {
610         PAGEREF_RECLAIM,
611         PAGEREF_RECLAIM_CLEAN,
612         PAGEREF_KEEP,
613         PAGEREF_ACTIVATE,
614 };
615
616 static enum page_references page_check_references(struct page *page,
617                                                   struct scan_control *sc)
618 {
619         int referenced_ptes, referenced_page;
620         unsigned long vm_flags;
621
622         referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
623                                           &vm_flags);
624         referenced_page = TestClearPageReferenced(page);
625
626         /*
627          * Mlock lost the isolation race with us.  Let try_to_unmap()
628          * move the page to the unevictable list.
629          */
630         if (vm_flags & VM_LOCKED)
631                 return PAGEREF_RECLAIM;
632
633         if (referenced_ptes) {
634                 if (PageSwapBacked(page))
635                         return PAGEREF_ACTIVATE;
636                 /*
637                  * All mapped pages start out with page table
638                  * references from the instantiating fault, so we need
639                  * to look twice if a mapped file page is used more
640                  * than once.
641                  *
642                  * Mark it and spare it for another trip around the
643                  * inactive list.  Another page table reference will
644                  * lead to its activation.
645                  *
646                  * Note: the mark is set for activated pages as well
647                  * so that recently deactivated but used pages are
648                  * quickly recovered.
649                  */
650                 SetPageReferenced(page);
651
652                 if (referenced_page || referenced_ptes > 1)
653                         return PAGEREF_ACTIVATE;
654
655                 /*
656                  * Activate file-backed executable pages after first usage.
657                  */
658                 if (vm_flags & VM_EXEC)
659                         return PAGEREF_ACTIVATE;
660
661                 return PAGEREF_KEEP;
662         }
663
664         /* Reclaim if clean, defer dirty pages to writeback */
665         if (referenced_page && !PageSwapBacked(page))
666                 return PAGEREF_RECLAIM_CLEAN;
667
668         return PAGEREF_RECLAIM;
669 }
670
671 /*
672  * shrink_page_list() returns the number of reclaimed pages
673  */
674 static unsigned long shrink_page_list(struct list_head *page_list,
675                                       struct zone *zone,
676                                       struct scan_control *sc,
677                                       enum ttu_flags ttu_flags,
678                                       unsigned long *ret_nr_dirty,
679                                       unsigned long *ret_nr_writeback,
680                                       bool force_reclaim)
681 {
682         LIST_HEAD(ret_pages);
683         LIST_HEAD(free_pages);
684         int pgactivate = 0;
685         unsigned long nr_dirty = 0;
686         unsigned long nr_congested = 0;
687         unsigned long nr_reclaimed = 0;
688         unsigned long nr_writeback = 0;
689
690         cond_resched();
691
692         mem_cgroup_uncharge_start();
693         while (!list_empty(page_list)) {
694                 struct address_space *mapping;
695                 struct page *page;
696                 int may_enter_fs;
697                 enum page_references references = PAGEREF_RECLAIM_CLEAN;
698
699                 cond_resched();
700
701                 page = lru_to_page(page_list);
702                 list_del(&page->lru);
703
704                 if (!trylock_page(page))
705                         goto keep;
706
707                 VM_BUG_ON(PageActive(page));
708                 VM_BUG_ON(page_zone(page) != zone);
709
710                 sc->nr_scanned++;
711
712                 if (unlikely(!page_evictable(page)))
713                         goto cull_mlocked;
714
715                 if (!sc->may_unmap && page_mapped(page))
716                         goto keep_locked;
717
718                 /* Double the slab pressure for mapped and swapcache pages */
719                 if (page_mapped(page) || PageSwapCache(page))
720                         sc->nr_scanned++;
721
722                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
723                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
724
725                 if (PageWriteback(page)) {
726                         /*
727                          * memcg doesn't have any dirty pages throttling so we
728                          * could easily OOM just because too many pages are in
729                          * writeback and there is nothing else to reclaim.
730                          *
731                          * Check __GFP_IO, certainly because a loop driver
732                          * thread might enter reclaim, and deadlock if it waits
733                          * on a page for which it is needed to do the write
734                          * (loop masks off __GFP_IO|__GFP_FS for this reason);
735                          * but more thought would probably show more reasons.
736                          *
737                          * Don't require __GFP_FS, since we're not going into
738                          * the FS, just waiting on its writeback completion.
739                          * Worryingly, ext4 gfs2 and xfs allocate pages with
740                          * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so
741                          * testing may_enter_fs here is liable to OOM on them.
742                          */
743                         if (global_reclaim(sc) ||
744                             !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
745                                 /*
746                                  * This is slightly racy - end_page_writeback()
747                                  * might have just cleared PageReclaim, then
748                                  * setting PageReclaim here end up interpreted
749                                  * as PageReadahead - but that does not matter
750                                  * enough to care.  What we do want is for this
751                                  * page to have PageReclaim set next time memcg
752                                  * reclaim reaches the tests above, so it will
753                                  * then wait_on_page_writeback() to avoid OOM;
754                                  * and it's also appropriate in global reclaim.
755                                  */
756                                 SetPageReclaim(page);
757                                 nr_writeback++;
758                                 goto keep_locked;
759                         }
760                         wait_on_page_writeback(page);
761                 }
762
763                 if (!force_reclaim)
764                         references = page_check_references(page, sc);
765
766                 switch (references) {
767                 case PAGEREF_ACTIVATE:
768                         goto activate_locked;
769                 case PAGEREF_KEEP:
770                         goto keep_locked;
771                 case PAGEREF_RECLAIM:
772                 case PAGEREF_RECLAIM_CLEAN:
773                         ; /* try to reclaim the page below */
774                 }
775
776                 /*
777                  * Anonymous process memory has backing store?
778                  * Try to allocate it some swap space here.
779                  */
780                 if (PageAnon(page) && !PageSwapCache(page)) {
781                         if (!(sc->gfp_mask & __GFP_IO))
782                                 goto keep_locked;
783                         if (!add_to_swap(page))
784                                 goto activate_locked;
785                         may_enter_fs = 1;
786                 }
787
788                 mapping = page_mapping(page);
789
790                 /*
791                  * The page is mapped into the page tables of one or more
792                  * processes. Try to unmap it here.
793                  */
794                 if (page_mapped(page) && mapping) {
795                         switch (try_to_unmap(page, ttu_flags)) {
796                         case SWAP_FAIL:
797                                 goto activate_locked;
798                         case SWAP_AGAIN:
799                                 goto keep_locked;
800                         case SWAP_MLOCK:
801                                 goto cull_mlocked;
802                         case SWAP_SUCCESS:
803                                 ; /* try to free the page below */
804                         }
805                 }
806
807                 if (PageDirty(page)) {
808                         nr_dirty++;
809
810                         /*
811                          * Only kswapd can writeback filesystem pages to
812                          * avoid risk of stack overflow but do not writeback
813                          * unless under significant pressure.
814                          */
815                         if (page_is_file_cache(page) &&
816                                         (!current_is_kswapd() ||
817                                          sc->priority >= DEF_PRIORITY - 2)) {
818                                 /*
819                                  * Immediately reclaim when written back.
820                                  * Similar in principal to deactivate_page()
821                                  * except we already have the page isolated
822                                  * and know it's dirty
823                                  */
824                                 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
825                                 SetPageReclaim(page);
826
827                                 goto keep_locked;
828                         }
829
830                         if (references == PAGEREF_RECLAIM_CLEAN)
831                                 goto keep_locked;
832                         if (!may_enter_fs)
833                                 goto keep_locked;
834                         if (!sc->may_writepage)
835                                 goto keep_locked;
836
837                         /* Page is dirty, try to write it out here */
838                         switch (pageout(page, mapping, sc)) {
839                         case PAGE_KEEP:
840                                 nr_congested++;
841                                 goto keep_locked;
842                         case PAGE_ACTIVATE:
843                                 goto activate_locked;
844                         case PAGE_SUCCESS:
845                                 if (PageWriteback(page))
846                                         goto keep;
847                                 if (PageDirty(page))
848                                         goto keep;
849
850                                 /*
851                                  * A synchronous write - probably a ramdisk.  Go
852                                  * ahead and try to reclaim the page.
853                                  */
854                                 if (!trylock_page(page))
855                                         goto keep;
856                                 if (PageDirty(page) || PageWriteback(page))
857                                         goto keep_locked;
858                                 mapping = page_mapping(page);
859                         case PAGE_CLEAN:
860                                 ; /* try to free the page below */
861                         }
862                 }
863
864                 /*
865                  * If the page has buffers, try to free the buffer mappings
866                  * associated with this page. If we succeed we try to free
867                  * the page as well.
868                  *
869                  * We do this even if the page is PageDirty().
870                  * try_to_release_page() does not perform I/O, but it is
871                  * possible for a page to have PageDirty set, but it is actually
872                  * clean (all its buffers are clean).  This happens if the
873                  * buffers were written out directly, with submit_bh(). ext3
874                  * will do this, as well as the blockdev mapping.
875                  * try_to_release_page() will discover that cleanness and will
876                  * drop the buffers and mark the page clean - it can be freed.
877                  *
878                  * Rarely, pages can have buffers and no ->mapping.  These are
879                  * the pages which were not successfully invalidated in
880                  * truncate_complete_page().  We try to drop those buffers here
881                  * and if that worked, and the page is no longer mapped into
882                  * process address space (page_count == 1) it can be freed.
883                  * Otherwise, leave the page on the LRU so it is swappable.
884                  */
885                 if (page_has_private(page)) {
886                         if (!try_to_release_page(page, sc->gfp_mask))
887                                 goto activate_locked;
888                         if (!mapping && page_count(page) == 1) {
889                                 unlock_page(page);
890                                 if (put_page_testzero(page))
891                                         goto free_it;
892                                 else {
893                                         /*
894                                          * rare race with speculative reference.
895                                          * the speculative reference will free
896                                          * this page shortly, so we may
897                                          * increment nr_reclaimed here (and
898                                          * leave it off the LRU).
899                                          */
900                                         nr_reclaimed++;
901                                         continue;
902                                 }
903                         }
904                 }
905
906                 if (!mapping || !__remove_mapping(mapping, page))
907                         goto keep_locked;
908
909                 /*
910                  * At this point, we have no other references and there is
911                  * no way to pick any more up (removed from LRU, removed
912                  * from pagecache). Can use non-atomic bitops now (and
913                  * we obviously don't have to worry about waking up a process
914                  * waiting on the page lock, because there are no references.
915                  */
916                 __clear_page_locked(page);
917 free_it:
918                 nr_reclaimed++;
919
920                 /*
921                  * Is there need to periodically free_page_list? It would
922                  * appear not as the counts should be low
923                  */
924                 list_add(&page->lru, &free_pages);
925                 continue;
926
927 cull_mlocked:
928                 if (PageSwapCache(page))
929                         try_to_free_swap(page);
930                 unlock_page(page);
931                 putback_lru_page(page);
932                 continue;
933
934 activate_locked:
935                 /* Not a candidate for swapping, so reclaim swap space. */
936                 if (PageSwapCache(page) && vm_swap_full())
937                         try_to_free_swap(page);
938                 VM_BUG_ON(PageActive(page));
939                 SetPageActive(page);
940                 pgactivate++;
941 keep_locked:
942                 unlock_page(page);
943 keep:
944                 list_add(&page->lru, &ret_pages);
945                 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
946         }
947
948         /*
949          * Tag a zone as congested if all the dirty pages encountered were
950          * backed by a congested BDI. In this case, reclaimers should just
951          * back off and wait for congestion to clear because further reclaim
952          * will encounter the same problem
953          */
954         if (nr_dirty && nr_dirty == nr_congested && global_reclaim(sc))
955                 zone_set_flag(zone, ZONE_CONGESTED);
956
957         free_hot_cold_page_list(&free_pages, 1);
958
959         list_splice(&ret_pages, page_list);
960         count_vm_events(PGACTIVATE, pgactivate);
961         mem_cgroup_uncharge_end();
962         *ret_nr_dirty += nr_dirty;
963         *ret_nr_writeback += nr_writeback;
964         return nr_reclaimed;
965 }
966
967 unsigned long reclaim_clean_pages_from_list(struct zone *zone,
968                                             struct list_head *page_list)
969 {
970         struct scan_control sc = {
971                 .gfp_mask = GFP_KERNEL,
972                 .priority = DEF_PRIORITY,
973                 .may_unmap = 1,
974         };
975         unsigned long ret, dummy1, dummy2;
976         struct page *page, *next;
977         LIST_HEAD(clean_pages);
978
979         list_for_each_entry_safe(page, next, page_list, lru) {
980                 if (page_is_file_cache(page) && !PageDirty(page)) {
981                         ClearPageActive(page);
982                         list_move(&page->lru, &clean_pages);
983                 }
984         }
985
986         ret = shrink_page_list(&clean_pages, zone, &sc,
987                                 TTU_UNMAP|TTU_IGNORE_ACCESS,
988                                 &dummy1, &dummy2, true);
989         list_splice(&clean_pages, page_list);
990         __mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
991         return ret;
992 }
993
994 /*
995  * Attempt to remove the specified page from its LRU.  Only take this page
996  * if it is of the appropriate PageActive status.  Pages which are being
997  * freed elsewhere are also ignored.
998  *
999  * page:        page to consider
1000  * mode:        one of the LRU isolation modes defined above
1001  *
1002  * returns 0 on success, -ve errno on failure.
1003  */
1004 int __isolate_lru_page(struct page *page, isolate_mode_t mode)
1005 {
1006         int ret = -EINVAL;
1007
1008         /* Only take pages on the LRU. */
1009         if (!PageLRU(page))
1010                 return ret;
1011
1012         /* Compaction should not handle unevictable pages but CMA can do so */
1013         if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
1014                 return ret;
1015
1016         ret = -EBUSY;
1017
1018         /*
1019          * To minimise LRU disruption, the caller can indicate that it only
1020          * wants to isolate pages it will be able to operate on without
1021          * blocking - clean pages for the most part.
1022          *
1023          * ISOLATE_CLEAN means that only clean pages should be isolated. This
1024          * is used by reclaim when it is cannot write to backing storage
1025          *
1026          * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1027          * that it is possible to migrate without blocking
1028          */
1029         if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1030                 /* All the caller can do on PageWriteback is block */
1031                 if (PageWriteback(page))
1032                         return ret;
1033
1034                 if (PageDirty(page)) {
1035                         struct address_space *mapping;
1036
1037                         /* ISOLATE_CLEAN means only clean pages */
1038                         if (mode & ISOLATE_CLEAN)
1039                                 return ret;
1040
1041                         /*
1042                          * Only pages without mappings or that have a
1043                          * ->migratepage callback are possible to migrate
1044                          * without blocking
1045                          */
1046                         mapping = page_mapping(page);
1047                         if (mapping && !mapping->a_ops->migratepage)
1048                                 return ret;
1049                 }
1050         }
1051
1052         if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1053                 return ret;
1054
1055         if (likely(get_page_unless_zero(page))) {
1056                 /*
1057                  * Be careful not to clear PageLRU until after we're
1058                  * sure the page is not being freed elsewhere -- the
1059                  * page release code relies on it.
1060                  */
1061                 ClearPageLRU(page);
1062                 ret = 0;
1063         }
1064
1065         return ret;
1066 }
1067
1068 /*
1069  * zone->lru_lock is heavily contended.  Some of the functions that
1070  * shrink the lists perform better by taking out a batch of pages
1071  * and working on them outside the LRU lock.
1072  *
1073  * For pagecache intensive workloads, this function is the hottest
1074  * spot in the kernel (apart from copy_*_user functions).
1075  *
1076  * Appropriate locks must be held before calling this function.
1077  *
1078  * @nr_to_scan: The number of pages to look through on the list.
1079  * @lruvec:     The LRU vector to pull pages from.
1080  * @dst:        The temp list to put pages on to.
1081  * @nr_scanned: The number of pages that were scanned.
1082  * @sc:         The scan_control struct for this reclaim session
1083  * @mode:       One of the LRU isolation modes
1084  * @lru:        LRU list id for isolating
1085  *
1086  * returns how many pages were moved onto *@dst.
1087  */
1088 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1089                 struct lruvec *lruvec, struct list_head *dst,
1090                 unsigned long *nr_scanned, struct scan_control *sc,
1091                 isolate_mode_t mode, enum lru_list lru)
1092 {
1093         struct list_head *src = &lruvec->lists[lru];
1094         unsigned long nr_taken = 0;
1095         unsigned long scan;
1096
1097         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
1098                 struct page *page;
1099                 int nr_pages;
1100
1101                 page = lru_to_page(src);
1102                 prefetchw_prev_lru_page(page, src, flags);
1103
1104                 VM_BUG_ON(!PageLRU(page));
1105
1106                 switch (__isolate_lru_page(page, mode)) {
1107                 case 0:
1108                         nr_pages = hpage_nr_pages(page);
1109                         mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
1110                         list_move(&page->lru, dst);
1111                         nr_taken += nr_pages;
1112                         break;
1113
1114                 case -EBUSY:
1115                         /* else it is being freed elsewhere */
1116                         list_move(&page->lru, src);
1117                         continue;
1118
1119                 default:
1120                         BUG();
1121                 }
1122         }
1123
1124         *nr_scanned = scan;
1125         trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1126                                     nr_taken, mode, is_file_lru(lru));
1127         return nr_taken;
1128 }
1129
1130 /**
1131  * isolate_lru_page - tries to isolate a page from its LRU list
1132  * @page: page to isolate from its LRU list
1133  *
1134  * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1135  * vmstat statistic corresponding to whatever LRU list the page was on.
1136  *
1137  * Returns 0 if the page was removed from an LRU list.
1138  * Returns -EBUSY if the page was not on an LRU list.
1139  *
1140  * The returned page will have PageLRU() cleared.  If it was found on
1141  * the active list, it will have PageActive set.  If it was found on
1142  * the unevictable list, it will have the PageUnevictable bit set. That flag
1143  * may need to be cleared by the caller before letting the page go.
1144  *
1145  * The vmstat statistic corresponding to the list on which the page was
1146  * found will be decremented.
1147  *
1148  * Restrictions:
1149  * (1) Must be called with an elevated refcount on the page. This is a
1150  *     fundamentnal difference from isolate_lru_pages (which is called
1151  *     without a stable reference).
1152  * (2) the lru_lock must not be held.
1153  * (3) interrupts must be enabled.
1154  */
1155 int isolate_lru_page(struct page *page)
1156 {
1157         int ret = -EBUSY;
1158
1159         VM_BUG_ON(!page_count(page));
1160
1161         if (PageLRU(page)) {
1162                 struct zone *zone = page_zone(page);
1163                 struct lruvec *lruvec;
1164
1165                 spin_lock_irq(&zone->lru_lock);
1166                 lruvec = mem_cgroup_page_lruvec(page, zone);
1167                 if (PageLRU(page)) {
1168                         int lru = page_lru(page);
1169                         get_page(page);
1170                         ClearPageLRU(page);
1171                         del_page_from_lru_list(page, lruvec, lru);
1172                         ret = 0;
1173                 }
1174                 spin_unlock_irq(&zone->lru_lock);
1175         }
1176         return ret;
1177 }
1178
1179 /*
1180  * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1181  * then get resheduled. When there are massive number of tasks doing page
1182  * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1183  * the LRU list will go small and be scanned faster than necessary, leading to
1184  * unnecessary swapping, thrashing and OOM.
1185  */
1186 static int too_many_isolated(struct zone *zone, int file,
1187                 struct scan_control *sc)
1188 {
1189         unsigned long inactive, isolated;
1190
1191         if (current_is_kswapd())
1192                 return 0;
1193
1194         if (!global_reclaim(sc))
1195                 return 0;
1196
1197         if (file) {
1198                 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1199                 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1200         } else {
1201                 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1202                 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1203         }
1204
1205         /*
1206          * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1207          * won't get blocked by normal direct-reclaimers, forming a circular
1208          * deadlock.
1209          */
1210         if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1211                 inactive >>= 3;
1212
1213         return isolated > inactive;
1214 }
1215
1216 static noinline_for_stack void
1217 putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
1218 {
1219         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1220         struct zone *zone = lruvec_zone(lruvec);
1221         LIST_HEAD(pages_to_free);
1222
1223         /*
1224          * Put back any unfreeable pages.
1225          */
1226         while (!list_empty(page_list)) {
1227                 struct page *page = lru_to_page(page_list);
1228                 int lru;
1229
1230                 VM_BUG_ON(PageLRU(page));
1231                 list_del(&page->lru);
1232                 if (unlikely(!page_evictable(page))) {
1233                         spin_unlock_irq(&zone->lru_lock);
1234                         putback_lru_page(page);
1235                         spin_lock_irq(&zone->lru_lock);
1236                         continue;
1237                 }
1238
1239                 lruvec = mem_cgroup_page_lruvec(page, zone);
1240
1241                 SetPageLRU(page);
1242                 lru = page_lru(page);
1243                 add_page_to_lru_list(page, lruvec, lru);
1244
1245                 if (is_active_lru(lru)) {
1246                         int file = is_file_lru(lru);
1247                         int numpages = hpage_nr_pages(page);
1248                         reclaim_stat->recent_rotated[file] += numpages;
1249                 }
1250                 if (put_page_testzero(page)) {
1251                         __ClearPageLRU(page);
1252                         __ClearPageActive(page);
1253                         del_page_from_lru_list(page, lruvec, lru);
1254
1255                         if (unlikely(PageCompound(page))) {
1256                                 spin_unlock_irq(&zone->lru_lock);
1257                                 (*get_compound_page_dtor(page))(page);
1258                                 spin_lock_irq(&zone->lru_lock);
1259                         } else
1260                                 list_add(&page->lru, &pages_to_free);
1261                 }
1262         }
1263
1264         /*
1265          * To save our caller's stack, now use input list for pages to free.
1266          */
1267         list_splice(&pages_to_free, page_list);
1268 }
1269
1270 /*
1271  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
1272  * of reclaimed pages
1273  */
1274 static noinline_for_stack unsigned long
1275 shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
1276                      struct scan_control *sc, enum lru_list lru)
1277 {
1278         LIST_HEAD(page_list);
1279         unsigned long nr_scanned;
1280         unsigned long nr_reclaimed = 0;
1281         unsigned long nr_taken;
1282         unsigned long nr_dirty = 0;
1283         unsigned long nr_writeback = 0;
1284         isolate_mode_t isolate_mode = 0;
1285         int file = is_file_lru(lru);
1286         struct zone *zone = lruvec_zone(lruvec);
1287         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1288
1289         while (unlikely(too_many_isolated(zone, file, sc))) {
1290                 congestion_wait(BLK_RW_ASYNC, HZ/10);
1291
1292                 /* We are about to die and free our memory. Return now. */
1293                 if (fatal_signal_pending(current))
1294                         return SWAP_CLUSTER_MAX;
1295         }
1296
1297         lru_add_drain();
1298
1299         if (!sc->may_unmap)
1300                 isolate_mode |= ISOLATE_UNMAPPED;
1301         if (!sc->may_writepage)
1302                 isolate_mode |= ISOLATE_CLEAN;
1303
1304         spin_lock_irq(&zone->lru_lock);
1305
1306         nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1307                                      &nr_scanned, sc, isolate_mode, lru);
1308
1309         __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1310         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1311
1312         if (global_reclaim(sc)) {
1313                 zone->pages_scanned += nr_scanned;
1314                 if (current_is_kswapd())
1315                         __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
1316                 else
1317                         __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
1318         }
1319         spin_unlock_irq(&zone->lru_lock);
1320
1321         if (nr_taken == 0)
1322                 return 0;
1323
1324         nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
1325                                         &nr_dirty, &nr_writeback, false);
1326
1327         spin_lock_irq(&zone->lru_lock);
1328
1329         reclaim_stat->recent_scanned[file] += nr_taken;
1330
1331         if (global_reclaim(sc)) {
1332                 if (current_is_kswapd())
1333                         __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1334                                                nr_reclaimed);
1335                 else
1336                         __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1337                                                nr_reclaimed);
1338         }
1339
1340         putback_inactive_pages(lruvec, &page_list);
1341
1342         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1343
1344         spin_unlock_irq(&zone->lru_lock);
1345
1346         free_hot_cold_page_list(&page_list, 1);
1347
1348         /*
1349          * If reclaim is isolating dirty pages under writeback, it implies
1350          * that the long-lived page allocation rate is exceeding the page
1351          * laundering rate. Either the global limits are not being effective
1352          * at throttling processes due to the page distribution throughout
1353          * zones or there is heavy usage of a slow backing device. The
1354          * only option is to throttle from reclaim context which is not ideal
1355          * as there is no guarantee the dirtying process is throttled in the
1356          * same way balance_dirty_pages() manages.
1357          *
1358          * This scales the number of dirty pages that must be under writeback
1359          * before throttling depending on priority. It is a simple backoff
1360          * function that has the most effect in the range DEF_PRIORITY to
1361          * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1362          * in trouble and reclaim is considered to be in trouble.
1363          *
1364          * DEF_PRIORITY   100% isolated pages must be PageWriteback to throttle
1365          * DEF_PRIORITY-1  50% must be PageWriteback
1366          * DEF_PRIORITY-2  25% must be PageWriteback, kswapd in trouble
1367          * ...
1368          * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1369          *                     isolated page is PageWriteback
1370          */
1371         if (nr_writeback && nr_writeback >=
1372                         (nr_taken >> (DEF_PRIORITY - sc->priority)))
1373                 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1374
1375         trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1376                 zone_idx(zone),
1377                 nr_scanned, nr_reclaimed,
1378                 sc->priority,
1379                 trace_shrink_flags(file));
1380         return nr_reclaimed;
1381 }
1382
1383 /*
1384  * This moves pages from the active list to the inactive list.
1385  *
1386  * We move them the other way if the page is referenced by one or more
1387  * processes, from rmap.
1388  *
1389  * If the pages are mostly unmapped, the processing is fast and it is
1390  * appropriate to hold zone->lru_lock across the whole operation.  But if
1391  * the pages are mapped, the processing is slow (page_referenced()) so we
1392  * should drop zone->lru_lock around each page.  It's impossible to balance
1393  * this, so instead we remove the pages from the LRU while processing them.
1394  * It is safe to rely on PG_active against the non-LRU pages in here because
1395  * nobody will play with that bit on a non-LRU page.
1396  *
1397  * The downside is that we have to touch page->_count against each page.
1398  * But we had to alter page->flags anyway.
1399  */
1400
1401 static void move_active_pages_to_lru(struct lruvec *lruvec,
1402                                      struct list_head *list,
1403                                      struct list_head *pages_to_free,
1404                                      enum lru_list lru)
1405 {
1406         struct zone *zone = lruvec_zone(lruvec);
1407         unsigned long pgmoved = 0;
1408         struct page *page;
1409         int nr_pages;
1410
1411         while (!list_empty(list)) {
1412                 page = lru_to_page(list);
1413                 lruvec = mem_cgroup_page_lruvec(page, zone);
1414
1415                 VM_BUG_ON(PageLRU(page));
1416                 SetPageLRU(page);
1417
1418                 nr_pages = hpage_nr_pages(page);
1419                 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
1420                 list_move(&page->lru, &lruvec->lists[lru]);
1421                 pgmoved += nr_pages;
1422
1423                 if (put_page_testzero(page)) {
1424                         __ClearPageLRU(page);
1425                         __ClearPageActive(page);
1426                         del_page_from_lru_list(page, lruvec, lru);
1427
1428                         if (unlikely(PageCompound(page))) {
1429                                 spin_unlock_irq(&zone->lru_lock);
1430                                 (*get_compound_page_dtor(page))(page);
1431                                 spin_lock_irq(&zone->lru_lock);
1432                         } else
1433                                 list_add(&page->lru, pages_to_free);
1434                 }
1435         }
1436         __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1437         if (!is_active_lru(lru))
1438                 __count_vm_events(PGDEACTIVATE, pgmoved);
1439 }
1440
1441 static void shrink_active_list(unsigned long nr_to_scan,
1442                                struct lruvec *lruvec,
1443                                struct scan_control *sc,
1444                                enum lru_list lru)
1445 {
1446         unsigned long nr_taken;
1447         unsigned long nr_scanned;
1448         unsigned long vm_flags;
1449         LIST_HEAD(l_hold);      /* The pages which were snipped off */
1450         LIST_HEAD(l_active);
1451         LIST_HEAD(l_inactive);
1452         struct page *page;
1453         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1454         unsigned long nr_rotated = 0;
1455         isolate_mode_t isolate_mode = 0;
1456         int file = is_file_lru(lru);
1457         struct zone *zone = lruvec_zone(lruvec);
1458
1459         lru_add_drain();
1460
1461         if (!sc->may_unmap)
1462                 isolate_mode |= ISOLATE_UNMAPPED;
1463         if (!sc->may_writepage)
1464                 isolate_mode |= ISOLATE_CLEAN;
1465
1466         spin_lock_irq(&zone->lru_lock);
1467
1468         nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1469                                      &nr_scanned, sc, isolate_mode, lru);
1470         if (global_reclaim(sc))
1471                 zone->pages_scanned += nr_scanned;
1472
1473         reclaim_stat->recent_scanned[file] += nr_taken;
1474
1475         __count_zone_vm_events(PGREFILL, zone, nr_scanned);
1476         __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1477         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1478         spin_unlock_irq(&zone->lru_lock);
1479
1480         while (!list_empty(&l_hold)) {
1481                 cond_resched();
1482                 page = lru_to_page(&l_hold);
1483                 list_del(&page->lru);
1484
1485                 if (unlikely(!page_evictable(page))) {
1486                         putback_lru_page(page);
1487                         continue;
1488                 }
1489
1490                 if (unlikely(buffer_heads_over_limit)) {
1491                         if (page_has_private(page) && trylock_page(page)) {
1492                                 if (page_has_private(page))
1493                                         try_to_release_page(page, 0);
1494                                 unlock_page(page);
1495                         }
1496                 }
1497
1498                 if (page_referenced(page, 0, sc->target_mem_cgroup,
1499                                     &vm_flags)) {
1500                         nr_rotated += hpage_nr_pages(page);
1501                         /*
1502                          * Identify referenced, file-backed active pages and
1503                          * give them one more trip around the active list. So
1504                          * that executable code get better chances to stay in
1505                          * memory under moderate memory pressure.  Anon pages
1506                          * are not likely to be evicted by use-once streaming
1507                          * IO, plus JVM can create lots of anon VM_EXEC pages,
1508                          * so we ignore them here.
1509                          */
1510                         if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
1511                                 list_add(&page->lru, &l_active);
1512                                 continue;
1513                         }
1514                 }
1515
1516                 ClearPageActive(page);  /* we are de-activating */
1517                 list_add(&page->lru, &l_inactive);
1518         }
1519
1520         /*
1521          * Move pages back to the lru list.
1522          */
1523         spin_lock_irq(&zone->lru_lock);
1524         /*
1525          * Count referenced pages from currently used mappings as rotated,
1526          * even though only some of them are actually re-activated.  This
1527          * helps balance scan pressure between file and anonymous pages in
1528          * get_scan_ratio.
1529          */
1530         reclaim_stat->recent_rotated[file] += nr_rotated;
1531
1532         move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1533         move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
1534         __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
1535         spin_unlock_irq(&zone->lru_lock);
1536
1537         free_hot_cold_page_list(&l_hold, 1);
1538 }
1539
1540 #ifdef CONFIG_SWAP
1541 static int inactive_anon_is_low_global(struct zone *zone)
1542 {
1543         unsigned long active, inactive;
1544
1545         active = zone_page_state(zone, NR_ACTIVE_ANON);
1546         inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1547
1548         if (inactive * zone->inactive_ratio < active)
1549                 return 1;
1550
1551         return 0;
1552 }
1553
1554 /**
1555  * inactive_anon_is_low - check if anonymous pages need to be deactivated
1556  * @lruvec: LRU vector to check
1557  *
1558  * Returns true if the zone does not have enough inactive anon pages,
1559  * meaning some active anon pages need to be deactivated.
1560  */
1561 static int inactive_anon_is_low(struct lruvec *lruvec)
1562 {
1563         /*
1564          * If we don't have swap space, anonymous page deactivation
1565          * is pointless.
1566          */
1567         if (!total_swap_pages)
1568                 return 0;
1569
1570         if (!mem_cgroup_disabled())
1571                 return mem_cgroup_inactive_anon_is_low(lruvec);
1572
1573         return inactive_anon_is_low_global(lruvec_zone(lruvec));
1574 }
1575 #else
1576 static inline int inactive_anon_is_low(struct lruvec *lruvec)
1577 {
1578         return 0;
1579 }
1580 #endif
1581
1582 /**
1583  * inactive_file_is_low - check if file pages need to be deactivated
1584  * @lruvec: LRU vector to check
1585  *
1586  * When the system is doing streaming IO, memory pressure here
1587  * ensures that active file pages get deactivated, until more
1588  * than half of the file pages are on the inactive list.
1589  *
1590  * Once we get to that situation, protect the system's working
1591  * set from being evicted by disabling active file page aging.
1592  *
1593  * This uses a different ratio than the anonymous pages, because
1594  * the page cache uses a use-once replacement algorithm.
1595  */
1596 static int inactive_file_is_low(struct lruvec *lruvec)
1597 {
1598         unsigned long inactive;
1599         unsigned long active;
1600
1601         inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1602         active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1603
1604         return active > inactive;
1605 }
1606
1607 static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
1608 {
1609         if (is_file_lru(lru))
1610                 return inactive_file_is_low(lruvec);
1611         else
1612                 return inactive_anon_is_low(lruvec);
1613 }
1614
1615 static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1616                                  struct lruvec *lruvec, struct scan_control *sc)
1617 {
1618         if (is_active_lru(lru)) {
1619                 if (inactive_list_is_low(lruvec, lru))
1620                         shrink_active_list(nr_to_scan, lruvec, sc, lru);
1621                 return 0;
1622         }
1623
1624         return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
1625 }
1626
1627 static int vmscan_swappiness(struct scan_control *sc)
1628 {
1629         if (global_reclaim(sc))
1630                 return vm_swappiness;
1631         return mem_cgroup_swappiness(sc->target_mem_cgroup);
1632 }
1633
1634 enum scan_balance {
1635         SCAN_EQUAL,
1636         SCAN_FRACT,
1637         SCAN_ANON,
1638         SCAN_FILE,
1639 };
1640
1641 /*
1642  * Determine how aggressively the anon and file LRU lists should be
1643  * scanned.  The relative value of each set of LRU lists is determined
1644  * by looking at the fraction of the pages scanned we did rotate back
1645  * onto the active list instead of evict.
1646  *
1647  * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1648  * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
1649  */
1650 static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
1651                            unsigned long *nr)
1652 {
1653         struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1654         u64 fraction[2];
1655         u64 denominator = 0;    /* gcc */
1656         struct zone *zone = lruvec_zone(lruvec);
1657         unsigned long anon_prio, file_prio;
1658         enum scan_balance scan_balance;
1659         unsigned long anon, file, free;
1660         bool force_scan = false;
1661         unsigned long ap, fp;
1662         enum lru_list lru;
1663
1664         /*
1665          * If the zone or memcg is small, nr[l] can be 0.  This
1666          * results in no scanning on this priority and a potential
1667          * priority drop.  Global direct reclaim can go to the next
1668          * zone and tends to have no problems. Global kswapd is for
1669          * zone balancing and it needs to scan a minimum amount. When
1670          * reclaiming for a memcg, a priority drop can cause high
1671          * latencies, so it's better to scan a minimum amount there as
1672          * well.
1673          */
1674         if (current_is_kswapd() && zone->all_unreclaimable)
1675                 force_scan = true;
1676         if (!global_reclaim(sc))
1677                 force_scan = true;
1678
1679         /* If we have no swap space, do not bother scanning anon pages. */
1680         if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
1681                 scan_balance = SCAN_FILE;
1682                 goto out;
1683         }
1684
1685         /*
1686          * Global reclaim will swap to prevent OOM even with no
1687          * swappiness, but memcg users want to use this knob to
1688          * disable swapping for individual groups completely when
1689          * using the memory controller's swap limit feature would be
1690          * too expensive.
1691          */
1692         if (!global_reclaim(sc) && !vmscan_swappiness(sc)) {
1693                 scan_balance = SCAN_FILE;
1694                 goto out;
1695         }
1696
1697         /*
1698          * Do not apply any pressure balancing cleverness when the
1699          * system is close to OOM, scan both anon and file equally
1700          * (unless the swappiness setting disagrees with swapping).
1701          */
1702         if (!sc->priority && vmscan_swappiness(sc)) {
1703                 scan_balance = SCAN_EQUAL;
1704                 goto out;
1705         }
1706
1707         anon  = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1708                 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1709         file  = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1710                 get_lru_size(lruvec, LRU_INACTIVE_FILE);
1711
1712         /*
1713          * If it's foreseeable that reclaiming the file cache won't be
1714          * enough to get the zone back into a desirable shape, we have
1715          * to swap.  Better start now and leave the - probably heavily
1716          * thrashing - remaining file pages alone.
1717          */
1718         if (global_reclaim(sc)) {
1719                 free = zone_page_state(zone, NR_FREE_PAGES);
1720                 if (unlikely(file + free <= high_wmark_pages(zone))) {
1721                         scan_balance = SCAN_ANON;
1722                         goto out;
1723                 }
1724         }
1725
1726         /*
1727          * There is enough inactive page cache, do not reclaim
1728          * anything from the anonymous working set right now.
1729          */
1730         if (!inactive_file_is_low(lruvec)) {
1731                 scan_balance = SCAN_FILE;
1732                 goto out;
1733         }
1734
1735         scan_balance = SCAN_FRACT;
1736
1737         /*
1738          * With swappiness at 100, anonymous and file have the same priority.
1739          * This scanning priority is essentially the inverse of IO cost.
1740          */
1741         anon_prio = vmscan_swappiness(sc);
1742         file_prio = 200 - anon_prio;
1743
1744         /*
1745          * OK, so we have swap space and a fair amount of page cache
1746          * pages.  We use the recently rotated / recently scanned
1747          * ratios to determine how valuable each cache is.
1748          *
1749          * Because workloads change over time (and to avoid overflow)
1750          * we keep these statistics as a floating average, which ends
1751          * up weighing recent references more than old ones.
1752          *
1753          * anon in [0], file in [1]
1754          */
1755         spin_lock_irq(&zone->lru_lock);
1756         if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
1757                 reclaim_stat->recent_scanned[0] /= 2;
1758                 reclaim_stat->recent_rotated[0] /= 2;
1759         }
1760
1761         if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
1762                 reclaim_stat->recent_scanned[1] /= 2;
1763                 reclaim_stat->recent_rotated[1] /= 2;
1764         }
1765
1766         /*
1767          * The amount of pressure on anon vs file pages is inversely
1768          * proportional to the fraction of recently scanned pages on
1769          * each list that were recently referenced and in active use.
1770          */
1771         ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
1772         ap /= reclaim_stat->recent_rotated[0] + 1;
1773
1774         fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
1775         fp /= reclaim_stat->recent_rotated[1] + 1;
1776         spin_unlock_irq(&zone->lru_lock);
1777
1778         fraction[0] = ap;
1779         fraction[1] = fp;
1780         denominator = ap + fp + 1;
1781 out:
1782         for_each_evictable_lru(lru) {
1783                 int file = is_file_lru(lru);
1784                 unsigned long size;
1785                 unsigned long scan;
1786
1787                 size = get_lru_size(lruvec, lru);
1788                 scan = size >> sc->priority;
1789
1790                 if (!scan && force_scan)
1791                         scan = min(size, SWAP_CLUSTER_MAX);
1792
1793                 switch (scan_balance) {
1794                 case SCAN_EQUAL:
1795                         /* Scan lists relative to size */
1796                         break;
1797                 case SCAN_FRACT:
1798                         /*
1799                          * Scan types proportional to swappiness and
1800                          * their relative recent reclaim efficiency.
1801                          */
1802                         scan = div64_u64(scan * fraction[file], denominator);
1803                         break;
1804                 case SCAN_FILE:
1805                 case SCAN_ANON:
1806                         /* Scan one type exclusively */
1807                         if ((scan_balance == SCAN_FILE) != file)
1808                                 scan = 0;
1809                         break;
1810                 default:
1811                         /* Look ma, no brain */
1812                         BUG();
1813                 }
1814                 nr[lru] = scan;
1815         }
1816 }
1817
1818 /*
1819  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1820  */
1821 static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
1822 {
1823         unsigned long nr[NR_LRU_LISTS];
1824         unsigned long nr_to_scan;
1825         enum lru_list lru;
1826         unsigned long nr_reclaimed = 0;
1827         unsigned long nr_to_reclaim = sc->nr_to_reclaim;
1828         struct blk_plug plug;
1829
1830         get_scan_count(lruvec, sc, nr);
1831
1832         blk_start_plug(&plug);
1833         while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1834                                         nr[LRU_INACTIVE_FILE]) {
1835                 for_each_evictable_lru(lru) {
1836                         if (nr[lru]) {
1837                                 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
1838                                 nr[lru] -= nr_to_scan;
1839
1840                                 nr_reclaimed += shrink_list(lru, nr_to_scan,
1841                                                             lruvec, sc);
1842                         }
1843                 }
1844                 /*
1845                  * On large memory systems, scan >> priority can become
1846                  * really large. This is fine for the starting priority;
1847                  * we want to put equal scanning pressure on each zone.
1848                  * However, if the VM has a harder time of freeing pages,
1849                  * with multiple processes reclaiming pages, the total
1850                  * freeing target can get unreasonably large.
1851                  */
1852                 if (nr_reclaimed >= nr_to_reclaim &&
1853                     sc->priority < DEF_PRIORITY)
1854                         break;
1855         }
1856         blk_finish_plug(&plug);
1857         sc->nr_reclaimed += nr_reclaimed;
1858
1859         /*
1860          * Even if we did not try to evict anon pages at all, we want to
1861          * rebalance the anon lru active/inactive ratio.
1862          */
1863         if (inactive_anon_is_low(lruvec))
1864                 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
1865                                    sc, LRU_ACTIVE_ANON);
1866
1867         throttle_vm_writeout(sc->gfp_mask);
1868 }
1869
1870 /* Use reclaim/compaction for costly allocs or under memory pressure */
1871 static bool in_reclaim_compaction(struct scan_control *sc)
1872 {
1873         if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
1874                         (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
1875                          sc->priority < DEF_PRIORITY - 2))
1876                 return true;
1877
1878         return false;
1879 }
1880
1881 /*
1882  * Reclaim/compaction is used for high-order allocation requests. It reclaims
1883  * order-0 pages before compacting the zone. should_continue_reclaim() returns
1884  * true if more pages should be reclaimed such that when the page allocator
1885  * calls try_to_compact_zone() that it will have enough free pages to succeed.
1886  * It will give up earlier than that if there is difficulty reclaiming pages.
1887  */
1888 static inline bool should_continue_reclaim(struct zone *zone,
1889                                         unsigned long nr_reclaimed,
1890                                         unsigned long nr_scanned,
1891                                         struct scan_control *sc)
1892 {
1893         unsigned long pages_for_compaction;
1894         unsigned long inactive_lru_pages;
1895
1896         /* If not in reclaim/compaction mode, stop */
1897         if (!in_reclaim_compaction(sc))
1898                 return false;
1899
1900         /* Consider stopping depending on scan and reclaim activity */
1901         if (sc->gfp_mask & __GFP_REPEAT) {
1902                 /*
1903                  * For __GFP_REPEAT allocations, stop reclaiming if the
1904                  * full LRU list has been scanned and we are still failing
1905                  * to reclaim pages. This full LRU scan is potentially
1906                  * expensive but a __GFP_REPEAT caller really wants to succeed
1907                  */
1908                 if (!nr_reclaimed && !nr_scanned)
1909                         return false;
1910         } else {
1911                 /*
1912                  * For non-__GFP_REPEAT allocations which can presumably
1913                  * fail without consequence, stop if we failed to reclaim
1914                  * any pages from the last SWAP_CLUSTER_MAX number of
1915                  * pages that were scanned. This will return to the
1916                  * caller faster at the risk reclaim/compaction and
1917                  * the resulting allocation attempt fails
1918                  */
1919                 if (!nr_reclaimed)
1920                         return false;
1921         }
1922
1923         /*
1924          * If we have not reclaimed enough pages for compaction and the
1925          * inactive lists are large enough, continue reclaiming
1926          */
1927         pages_for_compaction = (2UL << sc->order);
1928         inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
1929         if (get_nr_swap_pages() > 0)
1930                 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
1931         if (sc->nr_reclaimed < pages_for_compaction &&
1932                         inactive_lru_pages > pages_for_compaction)
1933                 return true;
1934
1935         /* If compaction would go ahead or the allocation would succeed, stop */
1936         switch (compaction_suitable(zone, sc->order)) {
1937         case COMPACT_PARTIAL:
1938         case COMPACT_CONTINUE:
1939                 return false;
1940         default:
1941                 return true;
1942         }
1943 }
1944
1945 static void shrink_zone(struct zone *zone, struct scan_control *sc)
1946 {
1947         unsigned long nr_reclaimed, nr_scanned;
1948
1949         do {
1950                 struct mem_cgroup *root = sc->target_mem_cgroup;
1951                 struct mem_cgroup_reclaim_cookie reclaim = {
1952                         .zone = zone,
1953                         .priority = sc->priority,
1954                 };
1955                 struct mem_cgroup *memcg;
1956
1957                 nr_reclaimed = sc->nr_reclaimed;
1958                 nr_scanned = sc->nr_scanned;
1959
1960                 memcg = mem_cgroup_iter(root, NULL, &reclaim);
1961                 do {
1962                         struct lruvec *lruvec;
1963
1964                         lruvec = mem_cgroup_zone_lruvec(zone, memcg);
1965
1966                         shrink_lruvec(lruvec, sc);
1967
1968                         /*
1969                          * Direct reclaim and kswapd have to scan all memory
1970                          * cgroups to fulfill the overall scan target for the
1971                          * zone.
1972                          *
1973                          * Limit reclaim, on the other hand, only cares about
1974                          * nr_to_reclaim pages to be reclaimed and it will
1975                          * retry with decreasing priority if one round over the
1976                          * whole hierarchy is not sufficient.
1977                          */
1978                         if (!global_reclaim(sc) &&
1979                                         sc->nr_reclaimed >= sc->nr_to_reclaim) {
1980                                 mem_cgroup_iter_break(root, memcg);
1981                                 break;
1982                         }
1983                         memcg = mem_cgroup_iter(root, memcg, &reclaim);
1984                 } while (memcg);
1985         } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
1986                                          sc->nr_scanned - nr_scanned, sc));
1987 }
1988
1989 /* Returns true if compaction should go ahead for a high-order request */
1990 static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
1991 {
1992         unsigned long balance_gap, watermark;
1993         bool watermark_ok;
1994
1995         /* Do not consider compaction for orders reclaim is meant to satisfy */
1996         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
1997                 return false;
1998
1999         /*
2000          * Compaction takes time to run and there are potentially other
2001          * callers using the pages just freed. Continue reclaiming until
2002          * there is a buffer of free pages available to give compaction
2003          * a reasonable chance of completing and allocating the page
2004          */
2005         balance_gap = min(low_wmark_pages(zone),
2006                 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2007                         KSWAPD_ZONE_BALANCE_GAP_RATIO);
2008         watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2009         watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2010
2011         /*
2012          * If compaction is deferred, reclaim up to a point where
2013          * compaction will have a chance of success when re-enabled
2014          */
2015         if (compaction_deferred(zone, sc->order))
2016                 return watermark_ok;
2017
2018         /* If compaction is not ready to start, keep reclaiming */
2019         if (!compaction_suitable(zone, sc->order))
2020                 return false;
2021
2022         return watermark_ok;
2023 }
2024
2025 /*
2026  * This is the direct reclaim path, for page-allocating processes.  We only
2027  * try to reclaim pages from zones which will satisfy the caller's allocation
2028  * request.
2029  *
2030  * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2031  * Because:
2032  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2033  *    allocation or
2034  * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2035  *    must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2036  *    zone defense algorithm.
2037  *
2038  * If a zone is deemed to be full of pinned pages then just give it a light
2039  * scan then give up on it.
2040  *
2041  * This function returns true if a zone is being reclaimed for a costly
2042  * high-order allocation and compaction is ready to begin. This indicates to
2043  * the caller that it should consider retrying the allocation instead of
2044  * further reclaim.
2045  */
2046 static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
2047 {
2048         struct zoneref *z;
2049         struct zone *zone;
2050         unsigned long nr_soft_reclaimed;
2051         unsigned long nr_soft_scanned;
2052         bool aborted_reclaim = false;
2053
2054         /*
2055          * If the number of buffer_heads in the machine exceeds the maximum
2056          * allowed level, force direct reclaim to scan the highmem zone as
2057          * highmem pages could be pinning lowmem pages storing buffer_heads
2058          */
2059         if (buffer_heads_over_limit)
2060                 sc->gfp_mask |= __GFP_HIGHMEM;
2061
2062         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2063                                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2064                 if (!populated_zone(zone))
2065                         continue;
2066                 /*
2067                  * Take care memory controller reclaiming has small influence
2068                  * to global LRU.
2069                  */
2070                 if (global_reclaim(sc)) {
2071                         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2072                                 continue;
2073                         if (zone->all_unreclaimable &&
2074                                         sc->priority != DEF_PRIORITY)
2075                                 continue;       /* Let kswapd poll it */
2076                         if (IS_ENABLED(CONFIG_COMPACTION)) {
2077                                 /*
2078                                  * If we already have plenty of memory free for
2079                                  * compaction in this zone, don't free any more.
2080                                  * Even though compaction is invoked for any
2081                                  * non-zero order, only frequent costly order
2082                                  * reclamation is disruptive enough to become a
2083                                  * noticeable problem, like transparent huge
2084                                  * page allocations.
2085                                  */
2086                                 if (compaction_ready(zone, sc)) {
2087                                         aborted_reclaim = true;
2088                                         continue;
2089                                 }
2090                         }
2091                         /*
2092                          * This steals pages from memory cgroups over softlimit
2093                          * and returns the number of reclaimed pages and
2094                          * scanned pages. This works for global memory pressure
2095                          * and balancing, not for a memcg's limit.
2096                          */
2097                         nr_soft_scanned = 0;
2098                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2099                                                 sc->order, sc->gfp_mask,
2100                                                 &nr_soft_scanned);
2101                         sc->nr_reclaimed += nr_soft_reclaimed;
2102                         sc->nr_scanned += nr_soft_scanned;
2103                         /* need some check for avoid more shrink_zone() */
2104                 }
2105
2106                 shrink_zone(zone, sc);
2107         }
2108
2109         return aborted_reclaim;
2110 }
2111
2112 static bool zone_reclaimable(struct zone *zone)
2113 {
2114         return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
2115 }
2116
2117 /* All zones in zonelist are unreclaimable? */
2118 static bool all_unreclaimable(struct zonelist *zonelist,
2119                 struct scan_control *sc)
2120 {
2121         struct zoneref *z;
2122         struct zone *zone;
2123
2124         for_each_zone_zonelist_nodemask(zone, z, zonelist,
2125                         gfp_zone(sc->gfp_mask), sc->nodemask) {
2126                 if (!populated_zone(zone))
2127                         continue;
2128                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2129                         continue;
2130                 if (!zone->all_unreclaimable)
2131                         return false;
2132         }
2133
2134         return true;
2135 }
2136
2137 /*
2138  * This is the main entry point to direct page reclaim.
2139  *
2140  * If a full scan of the inactive list fails to free enough memory then we
2141  * are "out of memory" and something needs to be killed.
2142  *
2143  * If the caller is !__GFP_FS then the probability of a failure is reasonably
2144  * high - the zone may be full of dirty or under-writeback pages, which this
2145  * caller can't do much about.  We kick the writeback threads and take explicit
2146  * naps in the hope that some of these pages can be written.  But if the
2147  * allocating task holds filesystem locks which prevent writeout this might not
2148  * work, and the allocation attempt will fail.
2149  *
2150  * returns:     0, if no pages reclaimed
2151  *              else, the number of pages reclaimed
2152  */
2153 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2154                                         struct scan_control *sc,
2155                                         struct shrink_control *shrink)
2156 {
2157         unsigned long total_scanned = 0;
2158         struct reclaim_state *reclaim_state = current->reclaim_state;
2159         struct zoneref *z;
2160         struct zone *zone;
2161         unsigned long writeback_threshold;
2162         bool aborted_reclaim;
2163
2164         delayacct_freepages_start();
2165
2166         if (global_reclaim(sc))
2167                 count_vm_event(ALLOCSTALL);
2168
2169         do {
2170                 sc->nr_scanned = 0;
2171                 aborted_reclaim = shrink_zones(zonelist, sc);
2172
2173                 /*
2174                  * Don't shrink slabs when reclaiming memory from
2175                  * over limit cgroups
2176                  */
2177                 if (global_reclaim(sc)) {
2178                         unsigned long lru_pages = 0;
2179                         for_each_zone_zonelist(zone, z, zonelist,
2180                                         gfp_zone(sc->gfp_mask)) {
2181                                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2182                                         continue;
2183
2184                                 lru_pages += zone_reclaimable_pages(zone);
2185                         }
2186
2187                         shrink_slab(shrink, sc->nr_scanned, lru_pages);
2188                         if (reclaim_state) {
2189                                 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2190                                 reclaim_state->reclaimed_slab = 0;
2191                         }
2192                 }
2193                 total_scanned += sc->nr_scanned;
2194                 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
2195                         goto out;
2196
2197                 /*
2198                  * If we're getting trouble reclaiming, start doing
2199                  * writepage even in laptop mode.
2200                  */
2201                 if (sc->priority < DEF_PRIORITY - 2)
2202                         sc->may_writepage = 1;
2203
2204                 /*
2205                  * Try to write back as many pages as we just scanned.  This
2206                  * tends to cause slow streaming writers to write data to the
2207                  * disk smoothly, at the dirtying rate, which is nice.   But
2208                  * that's undesirable in laptop mode, where we *want* lumpy
2209                  * writeout.  So in laptop mode, write out the whole world.
2210                  */
2211                 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2212                 if (total_scanned > writeback_threshold) {
2213                         wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2214                                                 WB_REASON_TRY_TO_FREE_PAGES);
2215                         sc->may_writepage = 1;
2216                 }
2217
2218                 /* Take a nap, wait for some writeback to complete */
2219                 if (!sc->hibernation_mode && sc->nr_scanned &&
2220                     sc->priority < DEF_PRIORITY - 2) {
2221                         struct zone *preferred_zone;
2222
2223                         first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2224                                                 &cpuset_current_mems_allowed,
2225                                                 &preferred_zone);
2226                         wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2227                 }
2228         } while (--sc->priority >= 0);
2229
2230 out:
2231         delayacct_freepages_end();
2232
2233         if (sc->nr_reclaimed)
2234                 return sc->nr_reclaimed;
2235
2236         /*
2237          * As hibernation is going on, kswapd is freezed so that it can't mark
2238          * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2239          * check.
2240          */
2241         if (oom_killer_disabled)
2242                 return 0;
2243
2244         /* Aborted reclaim to try compaction? don't OOM, then */
2245         if (aborted_reclaim)
2246                 return 1;
2247
2248         /* top priority shrink_zones still had more to do? don't OOM, then */
2249         if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
2250                 return 1;
2251
2252         return 0;
2253 }
2254
2255 static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2256 {
2257         struct zone *zone;
2258         unsigned long pfmemalloc_reserve = 0;
2259         unsigned long free_pages = 0;
2260         int i;
2261         bool wmark_ok;
2262
2263         for (i = 0; i <= ZONE_NORMAL; i++) {
2264                 zone = &pgdat->node_zones[i];
2265                 pfmemalloc_reserve += min_wmark_pages(zone);
2266                 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2267         }
2268
2269         wmark_ok = free_pages > pfmemalloc_reserve / 2;
2270
2271         /* kswapd must be awake if processes are being throttled */
2272         if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2273                 pgdat->classzone_idx = min(pgdat->classzone_idx,
2274                                                 (enum zone_type)ZONE_NORMAL);
2275                 wake_up_interruptible(&pgdat->kswapd_wait);
2276         }
2277
2278         return wmark_ok;
2279 }
2280
2281 /*
2282  * Throttle direct reclaimers if backing storage is backed by the network
2283  * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2284  * depleted. kswapd will continue to make progress and wake the processes
2285  * when the low watermark is reached.
2286  *
2287  * Returns true if a fatal signal was delivered during throttling. If this
2288  * happens, the page allocator should not consider triggering the OOM killer.
2289  */
2290 static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
2291                                         nodemask_t *nodemask)
2292 {
2293         struct zone *zone;
2294         int high_zoneidx = gfp_zone(gfp_mask);
2295         pg_data_t *pgdat;
2296
2297         /*
2298          * Kernel threads should not be throttled as they may be indirectly
2299          * responsible for cleaning pages necessary for reclaim to make forward
2300          * progress. kjournald for example may enter direct reclaim while
2301          * committing a transaction where throttling it could forcing other
2302          * processes to block on log_wait_commit().
2303          */
2304         if (current->flags & PF_KTHREAD)
2305                 goto out;
2306
2307         /*
2308          * If a fatal signal is pending, this process should not throttle.
2309          * It should return quickly so it can exit and free its memory
2310          */
2311         if (fatal_signal_pending(current))
2312                 goto out;
2313
2314         /* Check if the pfmemalloc reserves are ok */
2315         first_zones_zonelist(zonelist, high_zoneidx, NULL, &zone);
2316         pgdat = zone->zone_pgdat;
2317         if (pfmemalloc_watermark_ok(pgdat))
2318                 goto out;
2319
2320         /* Account for the throttling */
2321         count_vm_event(PGSCAN_DIRECT_THROTTLE);
2322
2323         /*
2324          * If the caller cannot enter the filesystem, it's possible that it
2325          * is due to the caller holding an FS lock or performing a journal
2326          * transaction in the case of a filesystem like ext[3|4]. In this case,
2327          * it is not safe to block on pfmemalloc_wait as kswapd could be
2328          * blocked waiting on the same lock. Instead, throttle for up to a
2329          * second before continuing.
2330          */
2331         if (!(gfp_mask & __GFP_FS)) {
2332                 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2333                         pfmemalloc_watermark_ok(pgdat), HZ);
2334
2335                 goto check_pending;
2336         }
2337
2338         /* Throttle until kswapd wakes the process */
2339         wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2340                 pfmemalloc_watermark_ok(pgdat));
2341
2342 check_pending:
2343         if (fatal_signal_pending(current))
2344                 return true;
2345
2346 out:
2347         return false;
2348 }
2349
2350 unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
2351                                 gfp_t gfp_mask, nodemask_t *nodemask)
2352 {
2353         unsigned long nr_reclaimed;
2354         struct scan_control sc = {
2355                 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2356                 .may_writepage = !laptop_mode,
2357                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2358                 .may_unmap = 1,
2359                 .may_swap = 1,
2360                 .order = order,
2361                 .priority = DEF_PRIORITY,
2362                 .target_mem_cgroup = NULL,
2363                 .nodemask = nodemask,
2364         };
2365         struct shrink_control shrink = {
2366                 .gfp_mask = sc.gfp_mask,
2367         };
2368
2369         /*
2370          * Do not enter reclaim if fatal signal was delivered while throttled.
2371          * 1 is returned so that the page allocator does not OOM kill at this
2372          * point.
2373          */
2374         if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
2375                 return 1;
2376
2377         trace_mm_vmscan_direct_reclaim_begin(order,
2378                                 sc.may_writepage,
2379                                 gfp_mask);
2380
2381         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2382
2383         trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2384
2385         return nr_reclaimed;
2386 }
2387
2388 #ifdef CONFIG_MEMCG
2389
2390 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
2391                                                 gfp_t gfp_mask, bool noswap,
2392                                                 struct zone *zone,
2393                                                 unsigned long *nr_scanned)
2394 {
2395         struct scan_control sc = {
2396                 .nr_scanned = 0,
2397                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2398                 .may_writepage = !laptop_mode,
2399                 .may_unmap = 1,
2400                 .may_swap = !noswap,
2401                 .order = 0,
2402                 .priority = 0,
2403                 .target_mem_cgroup = memcg,
2404         };
2405         struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2406
2407         sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2408                         (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2409
2410         trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
2411                                                       sc.may_writepage,
2412                                                       sc.gfp_mask);
2413
2414         /*
2415          * NOTE: Although we can get the priority field, using it
2416          * here is not a good idea, since it limits the pages we can scan.
2417          * if we don't reclaim here, the shrink_zone from balance_pgdat
2418          * will pick up pages from other mem cgroup's as well. We hack
2419          * the priority and make it zero.
2420          */
2421         shrink_lruvec(lruvec, &sc);
2422
2423         trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2424
2425         *nr_scanned = sc.nr_scanned;
2426         return sc.nr_reclaimed;
2427 }
2428
2429 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
2430                                            gfp_t gfp_mask,
2431                                            bool noswap)
2432 {
2433         struct zonelist *zonelist;
2434         unsigned long nr_reclaimed;
2435         int nid;
2436         struct scan_control sc = {
2437                 .may_writepage = !laptop_mode,
2438                 .may_unmap = 1,
2439                 .may_swap = !noswap,
2440                 .nr_to_reclaim = SWAP_CLUSTER_MAX,
2441                 .order = 0,
2442                 .priority = DEF_PRIORITY,
2443                 .target_mem_cgroup = memcg,
2444                 .nodemask = NULL, /* we don't care the placement */
2445                 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2446                                 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2447         };
2448         struct shrink_control shrink = {
2449                 .gfp_mask = sc.gfp_mask,
2450         };
2451
2452         /*
2453          * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2454          * take care of from where we get pages. So the node where we start the
2455          * scan does not need to be the current node.
2456          */
2457         nid = mem_cgroup_select_victim_node(memcg);
2458
2459         zonelist = NODE_DATA(nid)->node_zonelists;
2460
2461         trace_mm_vmscan_memcg_reclaim_begin(0,
2462                                             sc.may_writepage,
2463                                             sc.gfp_mask);
2464
2465         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
2466
2467         trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2468
2469         return nr_reclaimed;
2470 }
2471 #endif
2472
2473 static void age_active_anon(struct zone *zone, struct scan_control *sc)
2474 {
2475         struct mem_cgroup *memcg;
2476
2477         if (!total_swap_pages)
2478                 return;
2479
2480         memcg = mem_cgroup_iter(NULL, NULL, NULL);
2481         do {
2482                 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2483
2484                 if (inactive_anon_is_low(lruvec))
2485                         shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2486                                            sc, LRU_ACTIVE_ANON);
2487
2488                 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2489         } while (memcg);
2490 }
2491
2492 static bool zone_balanced(struct zone *zone, int order,
2493                           unsigned long balance_gap, int classzone_idx)
2494 {
2495         if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2496                                     balance_gap, classzone_idx, 0))
2497                 return false;
2498
2499         if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2500             !compaction_suitable(zone, order))
2501                 return false;
2502
2503         return true;
2504 }
2505
2506 /*
2507  * pgdat_balanced() is used when checking if a node is balanced.
2508  *
2509  * For order-0, all zones must be balanced!
2510  *
2511  * For high-order allocations only zones that meet watermarks and are in a
2512  * zone allowed by the callers classzone_idx are added to balanced_pages. The
2513  * total of balanced pages must be at least 25% of the zones allowed by
2514  * classzone_idx for the node to be considered balanced. Forcing all zones to
2515  * be balanced for high orders can cause excessive reclaim when there are
2516  * imbalanced zones.
2517  * The choice of 25% is due to
2518  *   o a 16M DMA zone that is balanced will not balance a zone on any
2519  *     reasonable sized machine
2520  *   o On all other machines, the top zone must be at least a reasonable
2521  *     percentage of the middle zones. For example, on 32-bit x86, highmem
2522  *     would need to be at least 256M for it to be balance a whole node.
2523  *     Similarly, on x86-64 the Normal zone would need to be at least 1G
2524  *     to balance a node on its own. These seemed like reasonable ratios.
2525  */
2526 static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
2527 {
2528         unsigned long managed_pages = 0;
2529         unsigned long balanced_pages = 0;
2530         int i;
2531
2532         /* Check the watermark levels */
2533         for (i = 0; i <= classzone_idx; i++) {
2534                 struct zone *zone = pgdat->node_zones + i;
2535
2536                 if (!populated_zone(zone))
2537                         continue;
2538
2539                 managed_pages += zone->managed_pages;
2540
2541                 /*
2542                  * A special case here:
2543                  *
2544                  * balance_pgdat() skips over all_unreclaimable after
2545                  * DEF_PRIORITY. Effectively, it considers them balanced so
2546                  * they must be considered balanced here as well!
2547                  */
2548                 if (zone->all_unreclaimable) {
2549                         balanced_pages += zone->managed_pages;
2550                         continue;
2551                 }
2552
2553                 if (zone_balanced(zone, order, 0, i))
2554                         balanced_pages += zone->managed_pages;
2555                 else if (!order)
2556                         return false;
2557         }
2558
2559         if (order)
2560                 return balanced_pages >= (managed_pages >> 2);
2561         else
2562                 return true;
2563 }
2564
2565 /*
2566  * Prepare kswapd for sleeping. This verifies that there are no processes
2567  * waiting in throttle_direct_reclaim() and that watermarks have been met.
2568  *
2569  * Returns true if kswapd is ready to sleep
2570  */
2571 static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
2572                                         int classzone_idx)
2573 {
2574         /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2575         if (remaining)
2576                 return false;
2577
2578         /*
2579          * There is a potential race between when kswapd checks its watermarks
2580          * and a process gets throttled. There is also a potential race if
2581          * processes get throttled, kswapd wakes, a large process exits therby
2582          * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2583          * is going to sleep, no process should be sleeping on pfmemalloc_wait
2584          * so wake them now if necessary. If necessary, processes will wake
2585          * kswapd and get throttled again
2586          */
2587         if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2588                 wake_up(&pgdat->pfmemalloc_wait);
2589                 return false;
2590         }
2591
2592         return pgdat_balanced(pgdat, order, classzone_idx);
2593 }
2594
2595 /*
2596  * For kswapd, balance_pgdat() will work across all this node's zones until
2597  * they are all at high_wmark_pages(zone).
2598  *
2599  * Returns the final order kswapd was reclaiming at
2600  *
2601  * There is special handling here for zones which are full of pinned pages.
2602  * This can happen if the pages are all mlocked, or if they are all used by
2603  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
2604  * What we do is to detect the case where all pages in the zone have been
2605  * scanned twice and there has been zero successful reclaim.  Mark the zone as
2606  * dead and from now on, only perform a short scan.  Basically we're polling
2607  * the zone for when the problem goes away.
2608  *
2609  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
2610  * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2611  * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2612  * lower zones regardless of the number of free pages in the lower zones. This
2613  * interoperates with the page allocator fallback scheme to ensure that aging
2614  * of pages is balanced across the zones.
2615  */
2616 static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
2617                                                         int *classzone_idx)
2618 {
2619         bool pgdat_is_balanced = false;
2620         int i;
2621         int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
2622         unsigned long total_scanned;
2623         struct reclaim_state *reclaim_state = current->reclaim_state;
2624         unsigned long nr_soft_reclaimed;
2625         unsigned long nr_soft_scanned;
2626         struct scan_control sc = {
2627                 .gfp_mask = GFP_KERNEL,
2628                 .may_unmap = 1,
2629                 .may_swap = 1,
2630                 /*
2631                  * kswapd doesn't want to be bailed out while reclaim. because
2632                  * we want to put equal scanning pressure on each zone.
2633                  */
2634                 .nr_to_reclaim = ULONG_MAX,
2635                 .order = order,
2636                 .target_mem_cgroup = NULL,
2637         };
2638         struct shrink_control shrink = {
2639                 .gfp_mask = sc.gfp_mask,
2640         };
2641 loop_again:
2642         total_scanned = 0;
2643         sc.priority = DEF_PRIORITY;
2644         sc.nr_reclaimed = 0;
2645         sc.may_writepage = !laptop_mode;
2646         count_vm_event(PAGEOUTRUN);
2647
2648         do {
2649                 unsigned long lru_pages = 0;
2650
2651                 /*
2652                  * Scan in the highmem->dma direction for the highest
2653                  * zone which needs scanning
2654                  */
2655                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2656                         struct zone *zone = pgdat->node_zones + i;
2657
2658                         if (!populated_zone(zone))
2659                                 continue;
2660
2661                         if (zone->all_unreclaimable &&
2662                             sc.priority != DEF_PRIORITY)
2663                                 continue;
2664
2665                         /*
2666                          * Do some background aging of the anon list, to give
2667                          * pages a chance to be referenced before reclaiming.
2668                          */
2669                         age_active_anon(zone, &sc);
2670
2671                         /*
2672                          * If the number of buffer_heads in the machine
2673                          * exceeds the maximum allowed level and this node
2674                          * has a highmem zone, force kswapd to reclaim from
2675                          * it to relieve lowmem pressure.
2676                          */
2677                         if (buffer_heads_over_limit && is_highmem_idx(i)) {
2678                                 end_zone = i;
2679                                 break;
2680                         }
2681
2682                         if (!zone_balanced(zone, order, 0, 0)) {
2683                                 end_zone = i;
2684                                 break;
2685                         } else {
2686                                 /* If balanced, clear the congested flag */
2687                                 zone_clear_flag(zone, ZONE_CONGESTED);
2688                         }
2689                 }
2690
2691                 if (i < 0) {
2692                         pgdat_is_balanced = true;
2693                         goto out;
2694                 }
2695
2696                 for (i = 0; i <= end_zone; i++) {
2697                         struct zone *zone = pgdat->node_zones + i;
2698
2699                         lru_pages += zone_reclaimable_pages(zone);
2700                 }
2701
2702                 /*
2703                  * Now scan the zone in the dma->highmem direction, stopping
2704                  * at the last zone which needs scanning.
2705                  *
2706                  * We do this because the page allocator works in the opposite
2707                  * direction.  This prevents the page allocator from allocating
2708                  * pages behind kswapd's direction of progress, which would
2709                  * cause too much scanning of the lower zones.
2710                  */
2711                 for (i = 0; i <= end_zone; i++) {
2712                         struct zone *zone = pgdat->node_zones + i;
2713                         int nr_slab, testorder;
2714                         unsigned long balance_gap;
2715
2716                         if (!populated_zone(zone))
2717                                 continue;
2718
2719                         if (zone->all_unreclaimable &&
2720                             sc.priority != DEF_PRIORITY)
2721                                 continue;
2722
2723                         sc.nr_scanned = 0;
2724
2725                         nr_soft_scanned = 0;
2726                         /*
2727                          * Call soft limit reclaim before calling shrink_zone.
2728                          */
2729                         nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2730                                                         order, sc.gfp_mask,
2731                                                         &nr_soft_scanned);
2732                         sc.nr_reclaimed += nr_soft_reclaimed;
2733                         total_scanned += nr_soft_scanned;
2734
2735                         /*
2736                          * We put equal pressure on every zone, unless
2737                          * one zone has way too many pages free
2738                          * already. The "too many pages" is defined
2739                          * as the high wmark plus a "gap" where the
2740                          * gap is either the low watermark or 1%
2741                          * of the zone, whichever is smaller.
2742                          */
2743                         balance_gap = min(low_wmark_pages(zone),
2744                                 (zone->managed_pages +
2745                                         KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2746                                 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2747                         /*
2748                          * Kswapd reclaims only single pages with compaction
2749                          * enabled. Trying too hard to reclaim until contiguous
2750                          * free pages have become available can hurt performance
2751                          * by evicting too much useful data from memory.
2752                          * Do not reclaim more than needed for compaction.
2753                          */
2754                         testorder = order;
2755                         if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2756                                         compaction_suitable(zone, order) !=
2757                                                 COMPACT_SKIPPED)
2758                                 testorder = 0;
2759
2760                         if ((buffer_heads_over_limit && is_highmem_idx(i)) ||
2761                             !zone_balanced(zone, testorder,
2762                                            balance_gap, end_zone)) {
2763                                 shrink_zone(zone, &sc);
2764
2765                                 reclaim_state->reclaimed_slab = 0;
2766                                 nr_slab = shrink_slab(&shrink, sc.nr_scanned, lru_pages);
2767                                 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
2768                                 total_scanned += sc.nr_scanned;
2769
2770                                 if (nr_slab == 0 && !zone_reclaimable(zone))
2771                                         zone->all_unreclaimable = 1;
2772                         }
2773
2774                         /*
2775                          * If we're getting trouble reclaiming, start doing
2776                          * writepage even in laptop mode.
2777                          */
2778                         if (sc.priority < DEF_PRIORITY - 2)
2779                                 sc.may_writepage = 1;
2780
2781                         if (zone->all_unreclaimable) {
2782                                 if (end_zone && end_zone == i)
2783                                         end_zone--;
2784                                 continue;
2785                         }
2786
2787                         if (zone_balanced(zone, testorder, 0, end_zone))
2788                                 /*
2789                                  * If a zone reaches its high watermark,
2790                                  * consider it to be no longer congested. It's
2791                                  * possible there are dirty pages backed by
2792                                  * congested BDIs but as pressure is relieved,
2793                                  * speculatively avoid congestion waits
2794                                  */
2795                                 zone_clear_flag(zone, ZONE_CONGESTED);
2796                 }
2797
2798                 /*
2799                  * If the low watermark is met there is no need for processes
2800                  * to be throttled on pfmemalloc_wait as they should not be
2801                  * able to safely make forward progress. Wake them
2802                  */
2803                 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
2804                                 pfmemalloc_watermark_ok(pgdat))
2805                         wake_up(&pgdat->pfmemalloc_wait);
2806
2807                 if (pgdat_balanced(pgdat, order, *classzone_idx)) {
2808                         pgdat_is_balanced = true;
2809                         break;          /* kswapd: all done */
2810                 }
2811
2812                 /*
2813                  * We do this so kswapd doesn't build up large priorities for
2814                  * example when it is freeing in parallel with allocators. It
2815                  * matches the direct reclaim path behaviour in terms of impact
2816                  * on zone->*_priority.
2817                  */
2818                 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
2819                         break;
2820         } while (--sc.priority >= 0);
2821
2822 out:
2823         if (!pgdat_is_balanced) {
2824                 cond_resched();
2825
2826                 try_to_freeze();
2827
2828                 /*
2829                  * Fragmentation may mean that the system cannot be
2830                  * rebalanced for high-order allocations in all zones.
2831                  * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2832                  * it means the zones have been fully scanned and are still
2833                  * not balanced. For high-order allocations, there is
2834                  * little point trying all over again as kswapd may
2835                  * infinite loop.
2836                  *
2837                  * Instead, recheck all watermarks at order-0 as they
2838                  * are the most important. If watermarks are ok, kswapd will go
2839                  * back to sleep. High-order users can still perform direct
2840                  * reclaim if they wish.
2841                  */
2842                 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2843                         order = sc.order = 0;
2844
2845                 goto loop_again;
2846         }
2847
2848         /*
2849          * If kswapd was reclaiming at a higher order, it has the option of
2850          * sleeping without all zones being balanced. Before it does, it must
2851          * ensure that the watermarks for order-0 on *all* zones are met and
2852          * that the congestion flags are cleared. The congestion flag must
2853          * be cleared as kswapd is the only mechanism that clears the flag
2854          * and it is potentially going to sleep here.
2855          */
2856         if (order) {
2857                 int zones_need_compaction = 1;
2858
2859                 for (i = 0; i <= end_zone; i++) {
2860                         struct zone *zone = pgdat->node_zones + i;
2861
2862                         if (!populated_zone(zone))
2863                                 continue;
2864
2865                         /* Check if the memory needs to be defragmented. */
2866                         if (zone_watermark_ok(zone, order,
2867                                     low_wmark_pages(zone), *classzone_idx, 0))
2868                                 zones_need_compaction = 0;
2869                 }
2870
2871                 if (zones_need_compaction)
2872                         compact_pgdat(pgdat, order);
2873         }
2874
2875         /*
2876          * Return the order we were reclaiming at so prepare_kswapd_sleep()
2877          * makes a decision on the order we were last reclaiming at. However,
2878          * if another caller entered the allocator slow path while kswapd
2879          * was awake, order will remain at the higher level
2880          */
2881         *classzone_idx = end_zone;
2882         return order;
2883 }
2884
2885 static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
2886 {
2887         long remaining = 0;
2888         DEFINE_WAIT(wait);
2889
2890         if (freezing(current) || kthread_should_stop())
2891                 return;
2892
2893         prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2894
2895         /* Try to sleep for a short interval */
2896         if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
2897                 remaining = schedule_timeout(HZ/10);
2898                 finish_wait(&pgdat->kswapd_wait, &wait);
2899                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2900         }
2901
2902         /*
2903          * After a short sleep, check if it was a premature sleep. If not, then
2904          * go fully to sleep until explicitly woken up.
2905          */
2906         if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
2907                 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
2908
2909                 /*
2910                  * vmstat counters are not perfectly accurate and the estimated
2911                  * value for counters such as NR_FREE_PAGES can deviate from the
2912                  * true value by nr_online_cpus * threshold. To avoid the zone
2913                  * watermarks being breached while under pressure, we reduce the
2914                  * per-cpu vmstat threshold while kswapd is awake and restore
2915                  * them before going back to sleep.
2916                  */
2917                 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
2918
2919                 /*
2920                  * Compaction records what page blocks it recently failed to
2921                  * isolate pages from and skips them in the future scanning.
2922                  * When kswapd is going to sleep, it is reasonable to assume
2923                  * that pages and compaction may succeed so reset the cache.
2924                  */
2925                 reset_isolation_suitable(pgdat);
2926
2927                 if (!kthread_should_stop())
2928                         schedule();
2929
2930                 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
2931         } else {
2932                 if (remaining)
2933                         count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
2934                 else
2935                         count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
2936         }
2937         finish_wait(&pgdat->kswapd_wait, &wait);
2938 }
2939
2940 /*
2941  * The background pageout daemon, started as a kernel thread
2942  * from the init process.
2943  *
2944  * This basically trickles out pages so that we have _some_
2945  * free memory available even if there is no other activity
2946  * that frees anything up. This is needed for things like routing
2947  * etc, where we otherwise might have all activity going on in
2948  * asynchronous contexts that cannot page things out.
2949  *
2950  * If there are applications that are active memory-allocators
2951  * (most normal use), this basically shouldn't matter.
2952  */
2953 static int kswapd(void *p)
2954 {
2955         unsigned long order, new_order;
2956         unsigned balanced_order;
2957         int classzone_idx, new_classzone_idx;
2958         int balanced_classzone_idx;
2959         pg_data_t *pgdat = (pg_data_t*)p;
2960         struct task_struct *tsk = current;
2961
2962         struct reclaim_state reclaim_state = {
2963                 .reclaimed_slab = 0,
2964         };
2965         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2966
2967         lockdep_set_current_reclaim_state(GFP_KERNEL);
2968
2969         if (!cpumask_empty(cpumask))
2970                 set_cpus_allowed_ptr(tsk, cpumask);
2971         current->reclaim_state = &reclaim_state;
2972
2973         /*
2974          * Tell the memory management that we're a "memory allocator",
2975          * and that if we need more memory we should get access to it
2976          * regardless (see "__alloc_pages()"). "kswapd" should
2977          * never get caught in the normal page freeing logic.
2978          *
2979          * (Kswapd normally doesn't need memory anyway, but sometimes
2980          * you need a small amount of memory in order to be able to
2981          * page out something else, and this flag essentially protects
2982          * us from recursively trying to free more memory as we're
2983          * trying to free the first piece of memory in the first place).
2984          */
2985         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
2986         set_freezable();
2987
2988         order = new_order = 0;
2989         balanced_order = 0;
2990         classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
2991         balanced_classzone_idx = classzone_idx;
2992         for ( ; ; ) {
2993                 bool ret;
2994
2995                 /*
2996                  * If the last balance_pgdat was unsuccessful it's unlikely a
2997                  * new request of a similar or harder type will succeed soon
2998                  * so consider going to sleep on the basis we reclaimed at
2999                  */
3000                 if (balanced_classzone_idx >= new_classzone_idx &&
3001                                         balanced_order == new_order) {
3002                         new_order = pgdat->kswapd_max_order;
3003                         new_classzone_idx = pgdat->classzone_idx;
3004                         pgdat->kswapd_max_order =  0;
3005                         pgdat->classzone_idx = pgdat->nr_zones - 1;
3006                 }
3007
3008                 if (order < new_order || classzone_idx > new_classzone_idx) {
3009                         /*
3010                          * Don't sleep if someone wants a larger 'order'
3011                          * allocation or has tigher zone constraints
3012                          */
3013                         order = new_order;
3014                         classzone_idx = new_classzone_idx;
3015                 } else {
3016                         kswapd_try_to_sleep(pgdat, balanced_order,
3017                                                 balanced_classzone_idx);
3018                         order = pgdat->kswapd_max_order;
3019                         classzone_idx = pgdat->classzone_idx;
3020                         new_order = order;
3021                         new_classzone_idx = classzone_idx;
3022                         pgdat->kswapd_max_order = 0;
3023                         pgdat->classzone_idx = pgdat->nr_zones - 1;
3024                 }
3025
3026                 ret = try_to_freeze();
3027                 if (kthread_should_stop())
3028                         break;
3029
3030                 /*
3031                  * We can speed up thawing tasks if we don't call balance_pgdat
3032                  * after returning from the refrigerator
3033                  */
3034                 if (!ret) {
3035                         trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
3036                         balanced_classzone_idx = classzone_idx;
3037                         balanced_order = balance_pgdat(pgdat, order,
3038                                                 &balanced_classzone_idx);
3039                 }
3040         }
3041
3042         current->reclaim_state = NULL;
3043         return 0;
3044 }
3045
3046 /*
3047  * A zone is low on free memory, so wake its kswapd task to service it.
3048  */
3049 void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
3050 {
3051         pg_data_t *pgdat;
3052
3053         if (!populated_zone(zone))
3054                 return;
3055
3056         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
3057                 return;
3058         pgdat = zone->zone_pgdat;
3059         if (pgdat->kswapd_max_order < order) {
3060                 pgdat->kswapd_max_order = order;
3061                 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3062         }
3063         if (!waitqueue_active(&pgdat->kswapd_wait))
3064                 return;
3065         if (zone_watermark_ok_safe(zone, order, low_wmark_pages(zone), 0, 0))
3066                 return;
3067
3068         trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
3069         wake_up_interruptible(&pgdat->kswapd_wait);
3070 }
3071
3072 /*
3073  * The reclaimable count would be mostly accurate.
3074  * The less reclaimable pages may be
3075  * - mlocked pages, which will be moved to unevictable list when encountered
3076  * - mapped pages, which may require several travels to be reclaimed
3077  * - dirty pages, which is not "instantly" reclaimable
3078  */
3079 unsigned long global_reclaimable_pages(void)
3080 {
3081         int nr;
3082
3083         nr = global_page_state(NR_ACTIVE_FILE) +
3084              global_page_state(NR_INACTIVE_FILE);
3085
3086         if (get_nr_swap_pages() > 0)
3087                 nr += global_page_state(NR_ACTIVE_ANON) +
3088                       global_page_state(NR_INACTIVE_ANON);
3089
3090         return nr;
3091 }
3092
3093 unsigned long zone_reclaimable_pages(struct zone *zone)
3094 {
3095         int nr;
3096
3097         nr = zone_page_state(zone, NR_ACTIVE_FILE) +
3098              zone_page_state(zone, NR_INACTIVE_FILE);
3099
3100         if (get_nr_swap_pages() > 0)
3101                 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
3102                       zone_page_state(zone, NR_INACTIVE_ANON);
3103
3104         return nr;
3105 }
3106
3107 #ifdef CONFIG_HIBERNATION
3108 /*
3109  * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
3110  * freed pages.
3111  *
3112  * Rather than trying to age LRUs the aim is to preserve the overall
3113  * LRU order by reclaiming preferentially
3114  * inactive > active > active referenced > active mapped
3115  */
3116 unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
3117 {
3118         struct reclaim_state reclaim_state;
3119         struct scan_control sc = {
3120                 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3121                 .may_swap = 1,
3122                 .may_unmap = 1,
3123                 .may_writepage = 1,
3124                 .nr_to_reclaim = nr_to_reclaim,
3125                 .hibernation_mode = 1,
3126                 .order = 0,
3127                 .priority = DEF_PRIORITY,
3128         };
3129         struct shrink_control shrink = {
3130                 .gfp_mask = sc.gfp_mask,
3131         };
3132         struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
3133         struct task_struct *p = current;
3134         unsigned long nr_reclaimed;
3135
3136         p->flags |= PF_MEMALLOC;
3137         lockdep_set_current_reclaim_state(sc.gfp_mask);
3138         reclaim_state.reclaimed_slab = 0;
3139         p->reclaim_state = &reclaim_state;
3140
3141         nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
3142
3143         p->reclaim_state = NULL;
3144         lockdep_clear_current_reclaim_state();
3145         p->flags &= ~PF_MEMALLOC;
3146
3147         return nr_reclaimed;
3148 }
3149 #endif /* CONFIG_HIBERNATION */
3150
3151 /* It's optimal to keep kswapds on the same CPUs as their memory, but
3152    not required for correctness.  So if the last cpu in a node goes
3153    away, we get changed to run anywhere: as the first one comes back,
3154    restore their cpu bindings. */
3155 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3156                         void *hcpu)
3157 {
3158         int nid;
3159
3160         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
3161                 for_each_node_state(nid, N_MEMORY) {
3162                         pg_data_t *pgdat = NODE_DATA(nid);
3163                         const struct cpumask *mask;
3164
3165                         mask = cpumask_of_node(pgdat->node_id);
3166
3167                         if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3168                                 /* One of our CPUs online: restore mask */
3169                                 set_cpus_allowed_ptr(pgdat->kswapd, mask);
3170                 }
3171         }
3172         return NOTIFY_OK;
3173 }
3174
3175 /*
3176  * This kswapd start function will be called by init and node-hot-add.
3177  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3178  */
3179 int kswapd_run(int nid)
3180 {
3181         pg_data_t *pgdat = NODE_DATA(nid);
3182         int ret = 0;
3183
3184         if (pgdat->kswapd)
3185                 return 0;
3186
3187         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3188         if (IS_ERR(pgdat->kswapd)) {
3189                 /* failure at boot is fatal */
3190                 BUG_ON(system_state == SYSTEM_BOOTING);
3191                 pgdat->kswapd = NULL;
3192                 pr_err("Failed to start kswapd on node %d\n", nid);
3193                 ret = PTR_ERR(pgdat->kswapd);
3194         }
3195         return ret;
3196 }
3197
3198 /*
3199  * Called by memory hotplug when all memory in a node is offlined.  Caller must
3200  * hold lock_memory_hotplug().
3201  */
3202 void kswapd_stop(int nid)
3203 {
3204         struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3205
3206         if (kswapd) {
3207                 kthread_stop(kswapd);
3208                 NODE_DATA(nid)->kswapd = NULL;
3209         }
3210 }
3211
3212 static int __init kswapd_init(void)
3213 {
3214         int nid;
3215
3216         swap_setup();
3217         for_each_node_state(nid, N_MEMORY)
3218                 kswapd_run(nid);
3219         hotcpu_notifier(cpu_callback, 0);
3220         return 0;
3221 }
3222
3223 module_init(kswapd_init)
3224
3225 #ifdef CONFIG_NUMA
3226 /*
3227  * Zone reclaim mode
3228  *
3229  * If non-zero call zone_reclaim when the number of free pages falls below
3230  * the watermarks.
3231  */
3232 int zone_reclaim_mode __read_mostly;
3233
3234 #define RECLAIM_OFF 0
3235 #define RECLAIM_ZONE (1<<0)     /* Run shrink_inactive_list on the zone */
3236 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
3237 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
3238
3239 /*
3240  * Priority for ZONE_RECLAIM. This determines the fraction of pages
3241  * of a node considered for each zone_reclaim. 4 scans 1/16th of
3242  * a zone.
3243  */
3244 #define ZONE_RECLAIM_PRIORITY 4
3245
3246 /*
3247  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3248  * occur.
3249  */
3250 int sysctl_min_unmapped_ratio = 1;
3251
3252 /*
3253  * If the number of slab pages in a zone grows beyond this percentage then
3254  * slab reclaim needs to occur.
3255  */
3256 int sysctl_min_slab_ratio = 5;
3257
3258 static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3259 {
3260         unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3261         unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3262                 zone_page_state(zone, NR_ACTIVE_FILE);
3263
3264         /*
3265          * It's possible for there to be more file mapped pages than
3266          * accounted for by the pages on the file LRU lists because
3267          * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3268          */
3269         return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3270 }
3271
3272 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3273 static long zone_pagecache_reclaimable(struct zone *zone)
3274 {
3275         long nr_pagecache_reclaimable;
3276         long delta = 0;
3277
3278         /*
3279          * If RECLAIM_SWAP is set, then all file pages are considered
3280          * potentially reclaimable. Otherwise, we have to worry about
3281          * pages like swapcache and zone_unmapped_file_pages() provides
3282          * a better estimate
3283          */
3284         if (zone_reclaim_mode & RECLAIM_SWAP)
3285                 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3286         else
3287                 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3288
3289         /* If we can't clean pages, remove dirty pages from consideration */
3290         if (!(zone_reclaim_mode & RECLAIM_WRITE))
3291                 delta += zone_page_state(zone, NR_FILE_DIRTY);
3292
3293         /* Watch for any possible underflows due to delta */
3294         if (unlikely(delta > nr_pagecache_reclaimable))
3295                 delta = nr_pagecache_reclaimable;
3296
3297         return nr_pagecache_reclaimable - delta;
3298 }
3299
3300 /*
3301  * Try to free up some pages from this zone through reclaim.
3302  */
3303 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3304 {
3305         /* Minimum pages needed in order to stay on node */
3306         const unsigned long nr_pages = 1 << order;
3307         struct task_struct *p = current;
3308         struct reclaim_state reclaim_state;
3309         struct scan_control sc = {
3310                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3311                 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3312                 .may_swap = 1,
3313                 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
3314                 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
3315                 .order = order,
3316                 .priority = ZONE_RECLAIM_PRIORITY,
3317         };
3318         struct shrink_control shrink = {
3319                 .gfp_mask = sc.gfp_mask,
3320         };
3321         unsigned long nr_slab_pages0, nr_slab_pages1;
3322
3323         cond_resched();
3324         /*
3325          * We need to be able to allocate from the reserves for RECLAIM_SWAP
3326          * and we also need to be able to write out pages for RECLAIM_WRITE
3327          * and RECLAIM_SWAP.
3328          */
3329         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
3330         lockdep_set_current_reclaim_state(gfp_mask);
3331         reclaim_state.reclaimed_slab = 0;
3332         p->reclaim_state = &reclaim_state;
3333
3334         if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
3335                 /*
3336                  * Free memory by calling shrink zone with increasing
3337                  * priorities until we have enough memory freed.
3338                  */
3339                 do {
3340                         shrink_zone(zone, &sc);
3341                 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
3342         }
3343
3344         nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3345         if (nr_slab_pages0 > zone->min_slab_pages) {
3346                 /*
3347                  * shrink_slab() does not currently allow us to determine how
3348                  * many pages were freed in this zone. So we take the current
3349                  * number of slab pages and shake the slab until it is reduced
3350                  * by the same nr_pages that we used for reclaiming unmapped
3351                  * pages.
3352                  *
3353                  * Note that shrink_slab will free memory on all zones and may
3354                  * take a long time.
3355                  */
3356                 for (;;) {
3357                         unsigned long lru_pages = zone_reclaimable_pages(zone);
3358
3359                         /* No reclaimable slab or very low memory pressure */
3360                         if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
3361                                 break;
3362
3363                         /* Freed enough memory */
3364                         nr_slab_pages1 = zone_page_state(zone,
3365                                                         NR_SLAB_RECLAIMABLE);
3366                         if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3367                                 break;
3368                 }
3369
3370                 /*
3371                  * Update nr_reclaimed by the number of slab pages we
3372                  * reclaimed from this zone.
3373                  */
3374                 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3375                 if (nr_slab_pages1 < nr_slab_pages0)
3376                         sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
3377         }
3378
3379         p->reclaim_state = NULL;
3380         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
3381         lockdep_clear_current_reclaim_state();
3382         return sc.nr_reclaimed >= nr_pages;
3383 }
3384
3385 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3386 {
3387         int node_id;
3388         int ret;
3389
3390         /*
3391          * Zone reclaim reclaims unmapped file backed pages and
3392          * slab pages if we are over the defined limits.
3393          *
3394          * A small portion of unmapped file backed pages is needed for
3395          * file I/O otherwise pages read by file I/O will be immediately
3396          * thrown out if the zone is overallocated. So we do not reclaim
3397          * if less than a specified percentage of the zone is used by
3398          * unmapped file backed pages.
3399          */
3400         if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3401             zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
3402                 return ZONE_RECLAIM_FULL;
3403
3404         if (zone->all_unreclaimable)
3405                 return ZONE_RECLAIM_FULL;
3406
3407         /*
3408          * Do not scan if the allocation should not be delayed.
3409          */
3410         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
3411                 return ZONE_RECLAIM_NOSCAN;
3412
3413         /*
3414          * Only run zone reclaim on the local zone or on zones that do not
3415          * have associated processors. This will favor the local processor
3416          * over remote processors and spread off node memory allocations
3417          * as wide as possible.
3418          */
3419         node_id = zone_to_nid(zone);
3420         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
3421                 return ZONE_RECLAIM_NOSCAN;
3422
3423         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
3424                 return ZONE_RECLAIM_NOSCAN;
3425
3426         ret = __zone_reclaim(zone, gfp_mask, order);
3427         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3428
3429         if (!ret)
3430                 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3431
3432         return ret;
3433 }
3434 #endif
3435
3436 /*
3437  * page_evictable - test whether a page is evictable
3438  * @page: the page to test
3439  *
3440  * Test whether page is evictable--i.e., should be placed on active/inactive
3441  * lists vs unevictable list.
3442  *
3443  * Reasons page might not be evictable:
3444  * (1) page's mapping marked unevictable
3445  * (2) page is part of an mlocked VMA
3446  *
3447  */
3448 int page_evictable(struct page *page)
3449 {
3450         return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
3451 }
3452
3453 #ifdef CONFIG_SHMEM
3454 /**
3455  * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3456  * @pages:      array of pages to check
3457  * @nr_pages:   number of pages to check
3458  *
3459  * Checks pages for evictability and moves them to the appropriate lru list.
3460  *
3461  * This function is only used for SysV IPC SHM_UNLOCK.
3462  */
3463 void check_move_unevictable_pages(struct page **pages, int nr_pages)
3464 {
3465         struct lruvec *lruvec;
3466         struct zone *zone = NULL;
3467         int pgscanned = 0;
3468         int pgrescued = 0;
3469         int i;
3470
3471         for (i = 0; i < nr_pages; i++) {
3472                 struct page *page = pages[i];
3473                 struct zone *pagezone;
3474
3475                 pgscanned++;
3476                 pagezone = page_zone(page);
3477                 if (pagezone != zone) {
3478                         if (zone)
3479                                 spin_unlock_irq(&zone->lru_lock);
3480                         zone = pagezone;
3481                         spin_lock_irq(&zone->lru_lock);
3482                 }
3483                 lruvec = mem_cgroup_page_lruvec(page, zone);
3484
3485                 if (!PageLRU(page) || !PageUnevictable(page))
3486                         continue;
3487
3488                 if (page_evictable(page)) {
3489                         enum lru_list lru = page_lru_base_type(page);
3490
3491                         VM_BUG_ON(PageActive(page));
3492                         ClearPageUnevictable(page);
3493                         del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3494                         add_page_to_lru_list(page, lruvec, lru);
3495                         pgrescued++;
3496                 }
3497         }
3498
3499         if (zone) {
3500                 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3501                 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3502                 spin_unlock_irq(&zone->lru_lock);
3503         }
3504 }
3505 #endif /* CONFIG_SHMEM */
3506
3507 static void warn_scan_unevictable_pages(void)
3508 {
3509         printk_once(KERN_WARNING
3510                     "%s: The scan_unevictable_pages sysctl/node-interface has been "
3511                     "disabled for lack of a legitimate use case.  If you have "
3512                     "one, please send an email to linux-mm@kvack.org.\n",
3513                     current->comm);
3514 }
3515
3516 /*
3517  * scan_unevictable_pages [vm] sysctl handler.  On demand re-scan of
3518  * all nodes' unevictable lists for evictable pages
3519  */
3520 unsigned long scan_unevictable_pages;
3521
3522 int scan_unevictable_handler(struct ctl_table *table, int write,
3523                            void __user *buffer,
3524                            size_t *length, loff_t *ppos)
3525 {
3526         warn_scan_unevictable_pages();
3527         proc_doulongvec_minmax(table, write, buffer, length, ppos);
3528         scan_unevictable_pages = 0;
3529         return 0;
3530 }
3531
3532 #ifdef CONFIG_NUMA
3533 /*
3534  * per node 'scan_unevictable_pages' attribute.  On demand re-scan of
3535  * a specified node's per zone unevictable lists for evictable pages.
3536  */
3537
3538 static ssize_t read_scan_unevictable_node(struct device *dev,
3539                                           struct device_attribute *attr,
3540                                           char *buf)
3541 {
3542         warn_scan_unevictable_pages();
3543         return sprintf(buf, "0\n");     /* always zero; should fit... */
3544 }
3545
3546 static ssize_t write_scan_unevictable_node(struct device *dev,
3547                                            struct device_attribute *attr,
3548                                         const char *buf, size_t count)
3549 {
3550         warn_scan_unevictable_pages();
3551         return 1;
3552 }
3553
3554
3555 static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3556                         read_scan_unevictable_node,
3557                         write_scan_unevictable_node);
3558
3559 int scan_unevictable_register_node(struct node *node)
3560 {
3561         return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
3562 }
3563
3564 void scan_unevictable_unregister_node(struct node *node)
3565 {
3566         device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
3567 }
3568 #endif