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