1cc871da3fdae931350eb17c12d8b044bf73d567
[sfrench/cifs-2.6.git] / mm / compaction.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * linux/mm/compaction.c
4  *
5  * Memory compaction for the reduction of external fragmentation. Note that
6  * this heavily depends upon page migration to do all the real heavy
7  * lifting
8  *
9  * Copyright IBM Corp. 2007-2010 Mel Gorman <mel@csn.ul.ie>
10  */
11 #include <linux/cpu.h>
12 #include <linux/swap.h>
13 #include <linux/migrate.h>
14 #include <linux/compaction.h>
15 #include <linux/mm_inline.h>
16 #include <linux/sched/signal.h>
17 #include <linux/backing-dev.h>
18 #include <linux/sysctl.h>
19 #include <linux/sysfs.h>
20 #include <linux/page-isolation.h>
21 #include <linux/kasan.h>
22 #include <linux/kthread.h>
23 #include <linux/freezer.h>
24 #include <linux/page_owner.h>
25 #include <linux/psi.h>
26 #include "internal.h"
27
28 #ifdef CONFIG_COMPACTION
29 static inline void count_compact_event(enum vm_event_item item)
30 {
31         count_vm_event(item);
32 }
33
34 static inline void count_compact_events(enum vm_event_item item, long delta)
35 {
36         count_vm_events(item, delta);
37 }
38 #else
39 #define count_compact_event(item) do { } while (0)
40 #define count_compact_events(item, delta) do { } while (0)
41 #endif
42
43 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
44
45 #define CREATE_TRACE_POINTS
46 #include <trace/events/compaction.h>
47
48 #define block_start_pfn(pfn, order)     round_down(pfn, 1UL << (order))
49 #define block_end_pfn(pfn, order)       ALIGN((pfn) + 1, 1UL << (order))
50 #define pageblock_start_pfn(pfn)        block_start_pfn(pfn, pageblock_order)
51 #define pageblock_end_pfn(pfn)          block_end_pfn(pfn, pageblock_order)
52
53 static unsigned long release_freepages(struct list_head *freelist)
54 {
55         struct page *page, *next;
56         unsigned long high_pfn = 0;
57
58         list_for_each_entry_safe(page, next, freelist, lru) {
59                 unsigned long pfn = page_to_pfn(page);
60                 list_del(&page->lru);
61                 __free_page(page);
62                 if (pfn > high_pfn)
63                         high_pfn = pfn;
64         }
65
66         return high_pfn;
67 }
68
69 static void split_map_pages(struct list_head *list)
70 {
71         unsigned int i, order, nr_pages;
72         struct page *page, *next;
73         LIST_HEAD(tmp_list);
74
75         list_for_each_entry_safe(page, next, list, lru) {
76                 list_del(&page->lru);
77
78                 order = page_private(page);
79                 nr_pages = 1 << order;
80
81                 post_alloc_hook(page, order, __GFP_MOVABLE);
82                 if (order)
83                         split_page(page, order);
84
85                 for (i = 0; i < nr_pages; i++) {
86                         list_add(&page->lru, &tmp_list);
87                         page++;
88                 }
89         }
90
91         list_splice(&tmp_list, list);
92 }
93
94 #ifdef CONFIG_COMPACTION
95
96 int PageMovable(struct page *page)
97 {
98         struct address_space *mapping;
99
100         VM_BUG_ON_PAGE(!PageLocked(page), page);
101         if (!__PageMovable(page))
102                 return 0;
103
104         mapping = page_mapping(page);
105         if (mapping && mapping->a_ops && mapping->a_ops->isolate_page)
106                 return 1;
107
108         return 0;
109 }
110 EXPORT_SYMBOL(PageMovable);
111
112 void __SetPageMovable(struct page *page, struct address_space *mapping)
113 {
114         VM_BUG_ON_PAGE(!PageLocked(page), page);
115         VM_BUG_ON_PAGE((unsigned long)mapping & PAGE_MAPPING_MOVABLE, page);
116         page->mapping = (void *)((unsigned long)mapping | PAGE_MAPPING_MOVABLE);
117 }
118 EXPORT_SYMBOL(__SetPageMovable);
119
120 void __ClearPageMovable(struct page *page)
121 {
122         VM_BUG_ON_PAGE(!PageLocked(page), page);
123         VM_BUG_ON_PAGE(!PageMovable(page), page);
124         /*
125          * Clear registered address_space val with keeping PAGE_MAPPING_MOVABLE
126          * flag so that VM can catch up released page by driver after isolation.
127          * With it, VM migration doesn't try to put it back.
128          */
129         page->mapping = (void *)((unsigned long)page->mapping &
130                                 PAGE_MAPPING_MOVABLE);
131 }
132 EXPORT_SYMBOL(__ClearPageMovable);
133
134 /* Do not skip compaction more than 64 times */
135 #define COMPACT_MAX_DEFER_SHIFT 6
136
137 /*
138  * Compaction is deferred when compaction fails to result in a page
139  * allocation success. 1 << compact_defer_limit compactions are skipped up
140  * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
141  */
142 void defer_compaction(struct zone *zone, int order)
143 {
144         zone->compact_considered = 0;
145         zone->compact_defer_shift++;
146
147         if (order < zone->compact_order_failed)
148                 zone->compact_order_failed = order;
149
150         if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
151                 zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
152
153         trace_mm_compaction_defer_compaction(zone, order);
154 }
155
156 /* Returns true if compaction should be skipped this time */
157 bool compaction_deferred(struct zone *zone, int order)
158 {
159         unsigned long defer_limit = 1UL << zone->compact_defer_shift;
160
161         if (order < zone->compact_order_failed)
162                 return false;
163
164         /* Avoid possible overflow */
165         if (++zone->compact_considered > defer_limit)
166                 zone->compact_considered = defer_limit;
167
168         if (zone->compact_considered >= defer_limit)
169                 return false;
170
171         trace_mm_compaction_deferred(zone, order);
172
173         return true;
174 }
175
176 /*
177  * Update defer tracking counters after successful compaction of given order,
178  * which means an allocation either succeeded (alloc_success == true) or is
179  * expected to succeed.
180  */
181 void compaction_defer_reset(struct zone *zone, int order,
182                 bool alloc_success)
183 {
184         if (alloc_success) {
185                 zone->compact_considered = 0;
186                 zone->compact_defer_shift = 0;
187         }
188         if (order >= zone->compact_order_failed)
189                 zone->compact_order_failed = order + 1;
190
191         trace_mm_compaction_defer_reset(zone, order);
192 }
193
194 /* Returns true if restarting compaction after many failures */
195 bool compaction_restarting(struct zone *zone, int order)
196 {
197         if (order < zone->compact_order_failed)
198                 return false;
199
200         return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
201                 zone->compact_considered >= 1UL << zone->compact_defer_shift;
202 }
203
204 /* Returns true if the pageblock should be scanned for pages to isolate. */
205 static inline bool isolation_suitable(struct compact_control *cc,
206                                         struct page *page)
207 {
208         if (cc->ignore_skip_hint)
209                 return true;
210
211         return !get_pageblock_skip(page);
212 }
213
214 static void reset_cached_positions(struct zone *zone)
215 {
216         zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
217         zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
218         zone->compact_cached_free_pfn =
219                                 pageblock_start_pfn(zone_end_pfn(zone) - 1);
220 }
221
222 /*
223  * Compound pages of >= pageblock_order should consistenly be skipped until
224  * released. It is always pointless to compact pages of such order (if they are
225  * migratable), and the pageblocks they occupy cannot contain any free pages.
226  */
227 static bool pageblock_skip_persistent(struct page *page)
228 {
229         if (!PageCompound(page))
230                 return false;
231
232         page = compound_head(page);
233
234         if (compound_order(page) >= pageblock_order)
235                 return true;
236
237         return false;
238 }
239
240 static bool
241 __reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source,
242                                                         bool check_target)
243 {
244         struct page *page = pfn_to_online_page(pfn);
245         struct page *end_page;
246         unsigned long block_pfn;
247
248         if (!page)
249                 return false;
250         if (zone != page_zone(page))
251                 return false;
252         if (pageblock_skip_persistent(page))
253                 return false;
254
255         /*
256          * If skip is already cleared do no further checking once the
257          * restart points have been set.
258          */
259         if (check_source && check_target && !get_pageblock_skip(page))
260                 return true;
261
262         /*
263          * If clearing skip for the target scanner, do not select a
264          * non-movable pageblock as the starting point.
265          */
266         if (!check_source && check_target &&
267             get_pageblock_migratetype(page) != MIGRATE_MOVABLE)
268                 return false;
269
270         /*
271          * Only clear the hint if a sample indicates there is either a
272          * free page or an LRU page in the block. One or other condition
273          * is necessary for the block to be a migration source/target.
274          */
275         block_pfn = pageblock_start_pfn(pfn);
276         pfn = max(block_pfn, zone->zone_start_pfn);
277         page = pfn_to_page(pfn);
278         if (zone != page_zone(page))
279                 return false;
280         pfn = block_pfn + pageblock_nr_pages;
281         pfn = min(pfn, zone_end_pfn(zone));
282         end_page = pfn_to_page(pfn);
283
284         do {
285                 if (pfn_valid_within(pfn)) {
286                         if (check_source && PageLRU(page)) {
287                                 clear_pageblock_skip(page);
288                                 return true;
289                         }
290
291                         if (check_target && PageBuddy(page)) {
292                                 clear_pageblock_skip(page);
293                                 return true;
294                         }
295                 }
296
297                 page += (1 << PAGE_ALLOC_COSTLY_ORDER);
298                 pfn += (1 << PAGE_ALLOC_COSTLY_ORDER);
299         } while (page < end_page);
300
301         return false;
302 }
303
304 /*
305  * This function is called to clear all cached information on pageblocks that
306  * should be skipped for page isolation when the migrate and free page scanner
307  * meet.
308  */
309 static void __reset_isolation_suitable(struct zone *zone)
310 {
311         unsigned long migrate_pfn = zone->zone_start_pfn;
312         unsigned long free_pfn = zone_end_pfn(zone);
313         unsigned long reset_migrate = free_pfn;
314         unsigned long reset_free = migrate_pfn;
315         bool source_set = false;
316         bool free_set = false;
317
318         if (!zone->compact_blockskip_flush)
319                 return;
320
321         zone->compact_blockskip_flush = false;
322
323         /*
324          * Walk the zone and update pageblock skip information. Source looks
325          * for PageLRU while target looks for PageBuddy. When the scanner
326          * is found, both PageBuddy and PageLRU are checked as the pageblock
327          * is suitable as both source and target.
328          */
329         for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages,
330                                         free_pfn -= pageblock_nr_pages) {
331                 cond_resched();
332
333                 /* Update the migrate PFN */
334                 if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) &&
335                     migrate_pfn < reset_migrate) {
336                         source_set = true;
337                         reset_migrate = migrate_pfn;
338                         zone->compact_init_migrate_pfn = reset_migrate;
339                         zone->compact_cached_migrate_pfn[0] = reset_migrate;
340                         zone->compact_cached_migrate_pfn[1] = reset_migrate;
341                 }
342
343                 /* Update the free PFN */
344                 if (__reset_isolation_pfn(zone, free_pfn, free_set, true) &&
345                     free_pfn > reset_free) {
346                         free_set = true;
347                         reset_free = free_pfn;
348                         zone->compact_init_free_pfn = reset_free;
349                         zone->compact_cached_free_pfn = reset_free;
350                 }
351         }
352
353         /* Leave no distance if no suitable block was reset */
354         if (reset_migrate >= reset_free) {
355                 zone->compact_cached_migrate_pfn[0] = migrate_pfn;
356                 zone->compact_cached_migrate_pfn[1] = migrate_pfn;
357                 zone->compact_cached_free_pfn = free_pfn;
358         }
359 }
360
361 void reset_isolation_suitable(pg_data_t *pgdat)
362 {
363         int zoneid;
364
365         for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
366                 struct zone *zone = &pgdat->node_zones[zoneid];
367                 if (!populated_zone(zone))
368                         continue;
369
370                 /* Only flush if a full compaction finished recently */
371                 if (zone->compact_blockskip_flush)
372                         __reset_isolation_suitable(zone);
373         }
374 }
375
376 /*
377  * Sets the pageblock skip bit if it was clear. Note that this is a hint as
378  * locks are not required for read/writers. Returns true if it was already set.
379  */
380 static bool test_and_set_skip(struct compact_control *cc, struct page *page,
381                                                         unsigned long pfn)
382 {
383         bool skip;
384
385         /* Do no update if skip hint is being ignored */
386         if (cc->ignore_skip_hint)
387                 return false;
388
389         if (!IS_ALIGNED(pfn, pageblock_nr_pages))
390                 return false;
391
392         skip = get_pageblock_skip(page);
393         if (!skip && !cc->no_set_skip_hint)
394                 set_pageblock_skip(page);
395
396         return skip;
397 }
398
399 static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
400 {
401         struct zone *zone = cc->zone;
402
403         pfn = pageblock_end_pfn(pfn);
404
405         /* Set for isolation rather than compaction */
406         if (cc->no_set_skip_hint)
407                 return;
408
409         if (pfn > zone->compact_cached_migrate_pfn[0])
410                 zone->compact_cached_migrate_pfn[0] = pfn;
411         if (cc->mode != MIGRATE_ASYNC &&
412             pfn > zone->compact_cached_migrate_pfn[1])
413                 zone->compact_cached_migrate_pfn[1] = pfn;
414 }
415
416 /*
417  * If no pages were isolated then mark this pageblock to be skipped in the
418  * future. The information is later cleared by __reset_isolation_suitable().
419  */
420 static void update_pageblock_skip(struct compact_control *cc,
421                         struct page *page, unsigned long pfn)
422 {
423         struct zone *zone = cc->zone;
424
425         if (cc->no_set_skip_hint)
426                 return;
427
428         if (!page)
429                 return;
430
431         set_pageblock_skip(page);
432
433         /* Update where async and sync compaction should restart */
434         if (pfn < zone->compact_cached_free_pfn)
435                 zone->compact_cached_free_pfn = pfn;
436 }
437 #else
438 static inline bool isolation_suitable(struct compact_control *cc,
439                                         struct page *page)
440 {
441         return true;
442 }
443
444 static inline bool pageblock_skip_persistent(struct page *page)
445 {
446         return false;
447 }
448
449 static inline void update_pageblock_skip(struct compact_control *cc,
450                         struct page *page, unsigned long pfn)
451 {
452 }
453
454 static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
455 {
456 }
457
458 static bool test_and_set_skip(struct compact_control *cc, struct page *page,
459                                                         unsigned long pfn)
460 {
461         return false;
462 }
463 #endif /* CONFIG_COMPACTION */
464
465 /*
466  * Compaction requires the taking of some coarse locks that are potentially
467  * very heavily contended. For async compaction, trylock and record if the
468  * lock is contended. The lock will still be acquired but compaction will
469  * abort when the current block is finished regardless of success rate.
470  * Sync compaction acquires the lock.
471  *
472  * Always returns true which makes it easier to track lock state in callers.
473  */
474 static bool compact_lock_irqsave(spinlock_t *lock, unsigned long *flags,
475                                                 struct compact_control *cc)
476 {
477         /* Track if the lock is contended in async mode */
478         if (cc->mode == MIGRATE_ASYNC && !cc->contended) {
479                 if (spin_trylock_irqsave(lock, *flags))
480                         return true;
481
482                 cc->contended = true;
483         }
484
485         spin_lock_irqsave(lock, *flags);
486         return true;
487 }
488
489 /*
490  * Compaction requires the taking of some coarse locks that are potentially
491  * very heavily contended. The lock should be periodically unlocked to avoid
492  * having disabled IRQs for a long time, even when there is nobody waiting on
493  * the lock. It might also be that allowing the IRQs will result in
494  * need_resched() becoming true. If scheduling is needed, async compaction
495  * aborts. Sync compaction schedules.
496  * Either compaction type will also abort if a fatal signal is pending.
497  * In either case if the lock was locked, it is dropped and not regained.
498  *
499  * Returns true if compaction should abort due to fatal signal pending, or
500  *              async compaction due to need_resched()
501  * Returns false when compaction can continue (sync compaction might have
502  *              scheduled)
503  */
504 static bool compact_unlock_should_abort(spinlock_t *lock,
505                 unsigned long flags, bool *locked, struct compact_control *cc)
506 {
507         if (*locked) {
508                 spin_unlock_irqrestore(lock, flags);
509                 *locked = false;
510         }
511
512         if (fatal_signal_pending(current)) {
513                 cc->contended = true;
514                 return true;
515         }
516
517         cond_resched();
518
519         return false;
520 }
521
522 /*
523  * Isolate free pages onto a private freelist. If @strict is true, will abort
524  * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
525  * (even though it may still end up isolating some pages).
526  */
527 static unsigned long isolate_freepages_block(struct compact_control *cc,
528                                 unsigned long *start_pfn,
529                                 unsigned long end_pfn,
530                                 struct list_head *freelist,
531                                 unsigned int stride,
532                                 bool strict)
533 {
534         int nr_scanned = 0, total_isolated = 0;
535         struct page *cursor;
536         unsigned long flags = 0;
537         bool locked = false;
538         unsigned long blockpfn = *start_pfn;
539         unsigned int order;
540
541         /* Strict mode is for isolation, speed is secondary */
542         if (strict)
543                 stride = 1;
544
545         cursor = pfn_to_page(blockpfn);
546
547         /* Isolate free pages. */
548         for (; blockpfn < end_pfn; blockpfn += stride, cursor += stride) {
549                 int isolated;
550                 struct page *page = cursor;
551
552                 /*
553                  * Periodically drop the lock (if held) regardless of its
554                  * contention, to give chance to IRQs. Abort if fatal signal
555                  * pending or async compaction detects need_resched()
556                  */
557                 if (!(blockpfn % SWAP_CLUSTER_MAX)
558                     && compact_unlock_should_abort(&cc->zone->lock, flags,
559                                                                 &locked, cc))
560                         break;
561
562                 nr_scanned++;
563                 if (!pfn_valid_within(blockpfn))
564                         goto isolate_fail;
565
566                 /*
567                  * For compound pages such as THP and hugetlbfs, we can save
568                  * potentially a lot of iterations if we skip them at once.
569                  * The check is racy, but we can consider only valid values
570                  * and the only danger is skipping too much.
571                  */
572                 if (PageCompound(page)) {
573                         const unsigned int order = compound_order(page);
574
575                         if (likely(order < MAX_ORDER)) {
576                                 blockpfn += (1UL << order) - 1;
577                                 cursor += (1UL << order) - 1;
578                         }
579                         goto isolate_fail;
580                 }
581
582                 if (!PageBuddy(page))
583                         goto isolate_fail;
584
585                 /*
586                  * If we already hold the lock, we can skip some rechecking.
587                  * Note that if we hold the lock now, checked_pageblock was
588                  * already set in some previous iteration (or strict is true),
589                  * so it is correct to skip the suitable migration target
590                  * recheck as well.
591                  */
592                 if (!locked) {
593                         locked = compact_lock_irqsave(&cc->zone->lock,
594                                                                 &flags, cc);
595
596                         /* Recheck this is a buddy page under lock */
597                         if (!PageBuddy(page))
598                                 goto isolate_fail;
599                 }
600
601                 /* Found a free page, will break it into order-0 pages */
602                 order = page_order(page);
603                 isolated = __isolate_free_page(page, order);
604                 if (!isolated)
605                         break;
606                 set_page_private(page, order);
607
608                 total_isolated += isolated;
609                 cc->nr_freepages += isolated;
610                 list_add_tail(&page->lru, freelist);
611
612                 if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
613                         blockpfn += isolated;
614                         break;
615                 }
616                 /* Advance to the end of split page */
617                 blockpfn += isolated - 1;
618                 cursor += isolated - 1;
619                 continue;
620
621 isolate_fail:
622                 if (strict)
623                         break;
624                 else
625                         continue;
626
627         }
628
629         if (locked)
630                 spin_unlock_irqrestore(&cc->zone->lock, flags);
631
632         /*
633          * There is a tiny chance that we have read bogus compound_order(),
634          * so be careful to not go outside of the pageblock.
635          */
636         if (unlikely(blockpfn > end_pfn))
637                 blockpfn = end_pfn;
638
639         trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
640                                         nr_scanned, total_isolated);
641
642         /* Record how far we have got within the block */
643         *start_pfn = blockpfn;
644
645         /*
646          * If strict isolation is requested by CMA then check that all the
647          * pages requested were isolated. If there were any failures, 0 is
648          * returned and CMA will fail.
649          */
650         if (strict && blockpfn < end_pfn)
651                 total_isolated = 0;
652
653         cc->total_free_scanned += nr_scanned;
654         if (total_isolated)
655                 count_compact_events(COMPACTISOLATED, total_isolated);
656         return total_isolated;
657 }
658
659 /**
660  * isolate_freepages_range() - isolate free pages.
661  * @cc:        Compaction control structure.
662  * @start_pfn: The first PFN to start isolating.
663  * @end_pfn:   The one-past-last PFN.
664  *
665  * Non-free pages, invalid PFNs, or zone boundaries within the
666  * [start_pfn, end_pfn) range are considered errors, cause function to
667  * undo its actions and return zero.
668  *
669  * Otherwise, function returns one-past-the-last PFN of isolated page
670  * (which may be greater then end_pfn if end fell in a middle of
671  * a free page).
672  */
673 unsigned long
674 isolate_freepages_range(struct compact_control *cc,
675                         unsigned long start_pfn, unsigned long end_pfn)
676 {
677         unsigned long isolated, pfn, block_start_pfn, block_end_pfn;
678         LIST_HEAD(freelist);
679
680         pfn = start_pfn;
681         block_start_pfn = pageblock_start_pfn(pfn);
682         if (block_start_pfn < cc->zone->zone_start_pfn)
683                 block_start_pfn = cc->zone->zone_start_pfn;
684         block_end_pfn = pageblock_end_pfn(pfn);
685
686         for (; pfn < end_pfn; pfn += isolated,
687                                 block_start_pfn = block_end_pfn,
688                                 block_end_pfn += pageblock_nr_pages) {
689                 /* Protect pfn from changing by isolate_freepages_block */
690                 unsigned long isolate_start_pfn = pfn;
691
692                 block_end_pfn = min(block_end_pfn, end_pfn);
693
694                 /*
695                  * pfn could pass the block_end_pfn if isolated freepage
696                  * is more than pageblock order. In this case, we adjust
697                  * scanning range to right one.
698                  */
699                 if (pfn >= block_end_pfn) {
700                         block_start_pfn = pageblock_start_pfn(pfn);
701                         block_end_pfn = pageblock_end_pfn(pfn);
702                         block_end_pfn = min(block_end_pfn, end_pfn);
703                 }
704
705                 if (!pageblock_pfn_to_page(block_start_pfn,
706                                         block_end_pfn, cc->zone))
707                         break;
708
709                 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
710                                         block_end_pfn, &freelist, 0, true);
711
712                 /*
713                  * In strict mode, isolate_freepages_block() returns 0 if
714                  * there are any holes in the block (ie. invalid PFNs or
715                  * non-free pages).
716                  */
717                 if (!isolated)
718                         break;
719
720                 /*
721                  * If we managed to isolate pages, it is always (1 << n) *
722                  * pageblock_nr_pages for some non-negative n.  (Max order
723                  * page may span two pageblocks).
724                  */
725         }
726
727         /* __isolate_free_page() does not map the pages */
728         split_map_pages(&freelist);
729
730         if (pfn < end_pfn) {
731                 /* Loop terminated early, cleanup. */
732                 release_freepages(&freelist);
733                 return 0;
734         }
735
736         /* We don't use freelists for anything. */
737         return pfn;
738 }
739
740 /* Similar to reclaim, but different enough that they don't share logic */
741 static bool too_many_isolated(struct zone *zone)
742 {
743         unsigned long active, inactive, isolated;
744
745         inactive = node_page_state(zone->zone_pgdat, NR_INACTIVE_FILE) +
746                         node_page_state(zone->zone_pgdat, NR_INACTIVE_ANON);
747         active = node_page_state(zone->zone_pgdat, NR_ACTIVE_FILE) +
748                         node_page_state(zone->zone_pgdat, NR_ACTIVE_ANON);
749         isolated = node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE) +
750                         node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON);
751
752         return isolated > (inactive + active) / 2;
753 }
754
755 /**
756  * isolate_migratepages_block() - isolate all migrate-able pages within
757  *                                a single pageblock
758  * @cc:         Compaction control structure.
759  * @low_pfn:    The first PFN to isolate
760  * @end_pfn:    The one-past-the-last PFN to isolate, within same pageblock
761  * @isolate_mode: Isolation mode to be used.
762  *
763  * Isolate all pages that can be migrated from the range specified by
764  * [low_pfn, end_pfn). The range is expected to be within same pageblock.
765  * Returns zero if there is a fatal signal pending, otherwise PFN of the
766  * first page that was not scanned (which may be both less, equal to or more
767  * than end_pfn).
768  *
769  * The pages are isolated on cc->migratepages list (not required to be empty),
770  * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
771  * is neither read nor updated.
772  */
773 static unsigned long
774 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
775                         unsigned long end_pfn, isolate_mode_t isolate_mode)
776 {
777         struct zone *zone = cc->zone;
778         unsigned long nr_scanned = 0, nr_isolated = 0;
779         struct lruvec *lruvec;
780         unsigned long flags = 0;
781         bool locked = false;
782         struct page *page = NULL, *valid_page = NULL;
783         unsigned long start_pfn = low_pfn;
784         bool skip_on_failure = false;
785         unsigned long next_skip_pfn = 0;
786         bool skip_updated = false;
787
788         /*
789          * Ensure that there are not too many pages isolated from the LRU
790          * list by either parallel reclaimers or compaction. If there are,
791          * delay for some time until fewer pages are isolated
792          */
793         while (unlikely(too_many_isolated(zone))) {
794                 /* async migration should just abort */
795                 if (cc->mode == MIGRATE_ASYNC)
796                         return 0;
797
798                 congestion_wait(BLK_RW_ASYNC, HZ/10);
799
800                 if (fatal_signal_pending(current))
801                         return 0;
802         }
803
804         cond_resched();
805
806         if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
807                 skip_on_failure = true;
808                 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
809         }
810
811         /* Time to isolate some pages for migration */
812         for (; low_pfn < end_pfn; low_pfn++) {
813
814                 if (skip_on_failure && low_pfn >= next_skip_pfn) {
815                         /*
816                          * We have isolated all migration candidates in the
817                          * previous order-aligned block, and did not skip it due
818                          * to failure. We should migrate the pages now and
819                          * hopefully succeed compaction.
820                          */
821                         if (nr_isolated)
822                                 break;
823
824                         /*
825                          * We failed to isolate in the previous order-aligned
826                          * block. Set the new boundary to the end of the
827                          * current block. Note we can't simply increase
828                          * next_skip_pfn by 1 << order, as low_pfn might have
829                          * been incremented by a higher number due to skipping
830                          * a compound or a high-order buddy page in the
831                          * previous loop iteration.
832                          */
833                         next_skip_pfn = block_end_pfn(low_pfn, cc->order);
834                 }
835
836                 /*
837                  * Periodically drop the lock (if held) regardless of its
838                  * contention, to give chance to IRQs. Abort async compaction
839                  * if contended.
840                  */
841                 if (!(low_pfn % SWAP_CLUSTER_MAX)
842                     && compact_unlock_should_abort(zone_lru_lock(zone), flags,
843                                                                 &locked, cc))
844                         break;
845
846                 if (!pfn_valid_within(low_pfn))
847                         goto isolate_fail;
848                 nr_scanned++;
849
850                 page = pfn_to_page(low_pfn);
851
852                 /*
853                  * Check if the pageblock has already been marked skipped.
854                  * Only the aligned PFN is checked as the caller isolates
855                  * COMPACT_CLUSTER_MAX at a time so the second call must
856                  * not falsely conclude that the block should be skipped.
857                  */
858                 if (!valid_page && IS_ALIGNED(low_pfn, pageblock_nr_pages)) {
859                         if (!cc->ignore_skip_hint && get_pageblock_skip(page)) {
860                                 low_pfn = end_pfn;
861                                 goto isolate_abort;
862                         }
863                         valid_page = page;
864                 }
865
866                 /*
867                  * Skip if free. We read page order here without zone lock
868                  * which is generally unsafe, but the race window is small and
869                  * the worst thing that can happen is that we skip some
870                  * potential isolation targets.
871                  */
872                 if (PageBuddy(page)) {
873                         unsigned long freepage_order = page_order_unsafe(page);
874
875                         /*
876                          * Without lock, we cannot be sure that what we got is
877                          * a valid page order. Consider only values in the
878                          * valid order range to prevent low_pfn overflow.
879                          */
880                         if (freepage_order > 0 && freepage_order < MAX_ORDER)
881                                 low_pfn += (1UL << freepage_order) - 1;
882                         continue;
883                 }
884
885                 /*
886                  * Regardless of being on LRU, compound pages such as THP and
887                  * hugetlbfs are not to be compacted. We can potentially save
888                  * a lot of iterations if we skip them at once. The check is
889                  * racy, but we can consider only valid values and the only
890                  * danger is skipping too much.
891                  */
892                 if (PageCompound(page)) {
893                         const unsigned int order = compound_order(page);
894
895                         if (likely(order < MAX_ORDER))
896                                 low_pfn += (1UL << order) - 1;
897                         goto isolate_fail;
898                 }
899
900                 /*
901                  * Check may be lockless but that's ok as we recheck later.
902                  * It's possible to migrate LRU and non-lru movable pages.
903                  * Skip any other type of page
904                  */
905                 if (!PageLRU(page)) {
906                         /*
907                          * __PageMovable can return false positive so we need
908                          * to verify it under page_lock.
909                          */
910                         if (unlikely(__PageMovable(page)) &&
911                                         !PageIsolated(page)) {
912                                 if (locked) {
913                                         spin_unlock_irqrestore(zone_lru_lock(zone),
914                                                                         flags);
915                                         locked = false;
916                                 }
917
918                                 if (!isolate_movable_page(page, isolate_mode))
919                                         goto isolate_success;
920                         }
921
922                         goto isolate_fail;
923                 }
924
925                 /*
926                  * Migration will fail if an anonymous page is pinned in memory,
927                  * so avoid taking lru_lock and isolating it unnecessarily in an
928                  * admittedly racy check.
929                  */
930                 if (!page_mapping(page) &&
931                     page_count(page) > page_mapcount(page))
932                         goto isolate_fail;
933
934                 /*
935                  * Only allow to migrate anonymous pages in GFP_NOFS context
936                  * because those do not depend on fs locks.
937                  */
938                 if (!(cc->gfp_mask & __GFP_FS) && page_mapping(page))
939                         goto isolate_fail;
940
941                 /* If we already hold the lock, we can skip some rechecking */
942                 if (!locked) {
943                         locked = compact_lock_irqsave(zone_lru_lock(zone),
944                                                                 &flags, cc);
945
946                         /* Try get exclusive access under lock */
947                         if (!skip_updated) {
948                                 skip_updated = true;
949                                 if (test_and_set_skip(cc, page, low_pfn))
950                                         goto isolate_abort;
951                         }
952
953                         /* Recheck PageLRU and PageCompound under lock */
954                         if (!PageLRU(page))
955                                 goto isolate_fail;
956
957                         /*
958                          * Page become compound since the non-locked check,
959                          * and it's on LRU. It can only be a THP so the order
960                          * is safe to read and it's 0 for tail pages.
961                          */
962                         if (unlikely(PageCompound(page))) {
963                                 low_pfn += (1UL << compound_order(page)) - 1;
964                                 goto isolate_fail;
965                         }
966                 }
967
968                 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
969
970                 /* Try isolate the page */
971                 if (__isolate_lru_page(page, isolate_mode) != 0)
972                         goto isolate_fail;
973
974                 VM_BUG_ON_PAGE(PageCompound(page), page);
975
976                 /* Successfully isolated */
977                 del_page_from_lru_list(page, lruvec, page_lru(page));
978                 inc_node_page_state(page,
979                                 NR_ISOLATED_ANON + page_is_file_cache(page));
980
981 isolate_success:
982                 list_add(&page->lru, &cc->migratepages);
983                 cc->nr_migratepages++;
984                 nr_isolated++;
985
986                 /*
987                  * Avoid isolating too much unless this block is being
988                  * rescanned (e.g. dirty/writeback pages, parallel allocation)
989                  * or a lock is contended. For contention, isolate quickly to
990                  * potentially remove one source of contention.
991                  */
992                 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX &&
993                     !cc->rescan && !cc->contended) {
994                         ++low_pfn;
995                         break;
996                 }
997
998                 continue;
999 isolate_fail:
1000                 if (!skip_on_failure)
1001                         continue;
1002
1003                 /*
1004                  * We have isolated some pages, but then failed. Release them
1005                  * instead of migrating, as we cannot form the cc->order buddy
1006                  * page anyway.
1007                  */
1008                 if (nr_isolated) {
1009                         if (locked) {
1010                                 spin_unlock_irqrestore(zone_lru_lock(zone), flags);
1011                                 locked = false;
1012                         }
1013                         putback_movable_pages(&cc->migratepages);
1014                         cc->nr_migratepages = 0;
1015                         nr_isolated = 0;
1016                 }
1017
1018                 if (low_pfn < next_skip_pfn) {
1019                         low_pfn = next_skip_pfn - 1;
1020                         /*
1021                          * The check near the loop beginning would have updated
1022                          * next_skip_pfn too, but this is a bit simpler.
1023                          */
1024                         next_skip_pfn += 1UL << cc->order;
1025                 }
1026         }
1027
1028         /*
1029          * The PageBuddy() check could have potentially brought us outside
1030          * the range to be scanned.
1031          */
1032         if (unlikely(low_pfn > end_pfn))
1033                 low_pfn = end_pfn;
1034
1035 isolate_abort:
1036         if (locked)
1037                 spin_unlock_irqrestore(zone_lru_lock(zone), flags);
1038
1039         /*
1040          * Updated the cached scanner pfn once the pageblock has been scanned
1041          * Pages will either be migrated in which case there is no point
1042          * scanning in the near future or migration failed in which case the
1043          * failure reason may persist. The block is marked for skipping if
1044          * there were no pages isolated in the block or if the block is
1045          * rescanned twice in a row.
1046          */
1047         if (low_pfn == end_pfn && (!nr_isolated || cc->rescan)) {
1048                 if (valid_page && !skip_updated)
1049                         set_pageblock_skip(valid_page);
1050                 update_cached_migrate(cc, low_pfn);
1051         }
1052
1053         trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
1054                                                 nr_scanned, nr_isolated);
1055
1056         cc->total_migrate_scanned += nr_scanned;
1057         if (nr_isolated)
1058                 count_compact_events(COMPACTISOLATED, nr_isolated);
1059
1060         return low_pfn;
1061 }
1062
1063 /**
1064  * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
1065  * @cc:        Compaction control structure.
1066  * @start_pfn: The first PFN to start isolating.
1067  * @end_pfn:   The one-past-last PFN.
1068  *
1069  * Returns zero if isolation fails fatally due to e.g. pending signal.
1070  * Otherwise, function returns one-past-the-last PFN of isolated page
1071  * (which may be greater than end_pfn if end fell in a middle of a THP page).
1072  */
1073 unsigned long
1074 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
1075                                                         unsigned long end_pfn)
1076 {
1077         unsigned long pfn, block_start_pfn, block_end_pfn;
1078
1079         /* Scan block by block. First and last block may be incomplete */
1080         pfn = start_pfn;
1081         block_start_pfn = pageblock_start_pfn(pfn);
1082         if (block_start_pfn < cc->zone->zone_start_pfn)
1083                 block_start_pfn = cc->zone->zone_start_pfn;
1084         block_end_pfn = pageblock_end_pfn(pfn);
1085
1086         for (; pfn < end_pfn; pfn = block_end_pfn,
1087                                 block_start_pfn = block_end_pfn,
1088                                 block_end_pfn += pageblock_nr_pages) {
1089
1090                 block_end_pfn = min(block_end_pfn, end_pfn);
1091
1092                 if (!pageblock_pfn_to_page(block_start_pfn,
1093                                         block_end_pfn, cc->zone))
1094                         continue;
1095
1096                 pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
1097                                                         ISOLATE_UNEVICTABLE);
1098
1099                 if (!pfn)
1100                         break;
1101
1102                 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
1103                         break;
1104         }
1105
1106         return pfn;
1107 }
1108
1109 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
1110 #ifdef CONFIG_COMPACTION
1111
1112 static bool suitable_migration_source(struct compact_control *cc,
1113                                                         struct page *page)
1114 {
1115         int block_mt;
1116
1117         if (pageblock_skip_persistent(page))
1118                 return false;
1119
1120         if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction)
1121                 return true;
1122
1123         block_mt = get_pageblock_migratetype(page);
1124
1125         if (cc->migratetype == MIGRATE_MOVABLE)
1126                 return is_migrate_movable(block_mt);
1127         else
1128                 return block_mt == cc->migratetype;
1129 }
1130
1131 /* Returns true if the page is within a block suitable for migration to */
1132 static bool suitable_migration_target(struct compact_control *cc,
1133                                                         struct page *page)
1134 {
1135         /* If the page is a large free page, then disallow migration */
1136         if (PageBuddy(page)) {
1137                 /*
1138                  * We are checking page_order without zone->lock taken. But
1139                  * the only small danger is that we skip a potentially suitable
1140                  * pageblock, so it's not worth to check order for valid range.
1141                  */
1142                 if (page_order_unsafe(page) >= pageblock_order)
1143                         return false;
1144         }
1145
1146         if (cc->ignore_block_suitable)
1147                 return true;
1148
1149         /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1150         if (is_migrate_movable(get_pageblock_migratetype(page)))
1151                 return true;
1152
1153         /* Otherwise skip the block */
1154         return false;
1155 }
1156
1157 static inline unsigned int
1158 freelist_scan_limit(struct compact_control *cc)
1159 {
1160         return (COMPACT_CLUSTER_MAX >> cc->fast_search_fail) + 1;
1161 }
1162
1163 /*
1164  * Test whether the free scanner has reached the same or lower pageblock than
1165  * the migration scanner, and compaction should thus terminate.
1166  */
1167 static inline bool compact_scanners_met(struct compact_control *cc)
1168 {
1169         return (cc->free_pfn >> pageblock_order)
1170                 <= (cc->migrate_pfn >> pageblock_order);
1171 }
1172
1173 /*
1174  * Used when scanning for a suitable migration target which scans freelists
1175  * in reverse. Reorders the list such as the unscanned pages are scanned
1176  * first on the next iteration of the free scanner
1177  */
1178 static void
1179 move_freelist_head(struct list_head *freelist, struct page *freepage)
1180 {
1181         LIST_HEAD(sublist);
1182
1183         if (!list_is_last(freelist, &freepage->lru)) {
1184                 list_cut_before(&sublist, freelist, &freepage->lru);
1185                 if (!list_empty(&sublist))
1186                         list_splice_tail(&sublist, freelist);
1187         }
1188 }
1189
1190 /*
1191  * Similar to move_freelist_head except used by the migration scanner
1192  * when scanning forward. It's possible for these list operations to
1193  * move against each other if they search the free list exactly in
1194  * lockstep.
1195  */
1196 static void
1197 move_freelist_tail(struct list_head *freelist, struct page *freepage)
1198 {
1199         LIST_HEAD(sublist);
1200
1201         if (!list_is_first(freelist, &freepage->lru)) {
1202                 list_cut_position(&sublist, freelist, &freepage->lru);
1203                 if (!list_empty(&sublist))
1204                         list_splice_tail(&sublist, freelist);
1205         }
1206 }
1207
1208 static void
1209 fast_isolate_around(struct compact_control *cc, unsigned long pfn, unsigned long nr_isolated)
1210 {
1211         unsigned long start_pfn, end_pfn;
1212         struct page *page = pfn_to_page(pfn);
1213
1214         /* Do not search around if there are enough pages already */
1215         if (cc->nr_freepages >= cc->nr_migratepages)
1216                 return;
1217
1218         /* Minimise scanning during async compaction */
1219         if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC)
1220                 return;
1221
1222         /* Pageblock boundaries */
1223         start_pfn = pageblock_start_pfn(pfn);
1224         end_pfn = min(start_pfn + pageblock_nr_pages, zone_end_pfn(cc->zone));
1225
1226         /* Scan before */
1227         if (start_pfn != pfn) {
1228                 isolate_freepages_block(cc, &start_pfn, pfn, &cc->freepages, 1, false);
1229                 if (cc->nr_freepages >= cc->nr_migratepages)
1230                         return;
1231         }
1232
1233         /* Scan after */
1234         start_pfn = pfn + nr_isolated;
1235         if (start_pfn != end_pfn)
1236                 isolate_freepages_block(cc, &start_pfn, end_pfn, &cc->freepages, 1, false);
1237
1238         /* Skip this pageblock in the future as it's full or nearly full */
1239         if (cc->nr_freepages < cc->nr_migratepages)
1240                 set_pageblock_skip(page);
1241 }
1242
1243 /* Search orders in round-robin fashion */
1244 static int next_search_order(struct compact_control *cc, int order)
1245 {
1246         order--;
1247         if (order < 0)
1248                 order = cc->order - 1;
1249
1250         /* Search wrapped around? */
1251         if (order == cc->search_order) {
1252                 cc->search_order--;
1253                 if (cc->search_order < 0)
1254                         cc->search_order = cc->order - 1;
1255                 return -1;
1256         }
1257
1258         return order;
1259 }
1260
1261 static unsigned long
1262 fast_isolate_freepages(struct compact_control *cc)
1263 {
1264         unsigned int limit = min(1U, freelist_scan_limit(cc) >> 1);
1265         unsigned int nr_scanned = 0;
1266         unsigned long low_pfn, min_pfn, high_pfn = 0, highest = 0;
1267         unsigned long nr_isolated = 0;
1268         unsigned long distance;
1269         struct page *page = NULL;
1270         bool scan_start = false;
1271         int order;
1272
1273         /* Full compaction passes in a negative order */
1274         if (cc->order <= 0)
1275                 return cc->free_pfn;
1276
1277         /*
1278          * If starting the scan, use a deeper search and use the highest
1279          * PFN found if a suitable one is not found.
1280          */
1281         if (cc->free_pfn >= cc->zone->compact_init_free_pfn) {
1282                 limit = pageblock_nr_pages >> 1;
1283                 scan_start = true;
1284         }
1285
1286         /*
1287          * Preferred point is in the top quarter of the scan space but take
1288          * a pfn from the top half if the search is problematic.
1289          */
1290         distance = (cc->free_pfn - cc->migrate_pfn);
1291         low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2));
1292         min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1));
1293
1294         if (WARN_ON_ONCE(min_pfn > low_pfn))
1295                 low_pfn = min_pfn;
1296
1297         /*
1298          * Search starts from the last successful isolation order or the next
1299          * order to search after a previous failure
1300          */
1301         cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order);
1302
1303         for (order = cc->search_order;
1304              !page && order >= 0;
1305              order = next_search_order(cc, order)) {
1306                 struct free_area *area = &cc->zone->free_area[order];
1307                 struct list_head *freelist;
1308                 struct page *freepage;
1309                 unsigned long flags;
1310                 unsigned int order_scanned = 0;
1311
1312                 if (!area->nr_free)
1313                         continue;
1314
1315                 spin_lock_irqsave(&cc->zone->lock, flags);
1316                 freelist = &area->free_list[MIGRATE_MOVABLE];
1317                 list_for_each_entry_reverse(freepage, freelist, lru) {
1318                         unsigned long pfn;
1319
1320                         order_scanned++;
1321                         nr_scanned++;
1322                         pfn = page_to_pfn(freepage);
1323
1324                         if (pfn >= highest)
1325                                 highest = pageblock_start_pfn(pfn);
1326
1327                         if (pfn >= low_pfn) {
1328                                 cc->fast_search_fail = 0;
1329                                 cc->search_order = order;
1330                                 page = freepage;
1331                                 break;
1332                         }
1333
1334                         if (pfn >= min_pfn && pfn > high_pfn) {
1335                                 high_pfn = pfn;
1336
1337                                 /* Shorten the scan if a candidate is found */
1338                                 limit >>= 1;
1339                         }
1340
1341                         if (order_scanned >= limit)
1342                                 break;
1343                 }
1344
1345                 /* Use a minimum pfn if a preferred one was not found */
1346                 if (!page && high_pfn) {
1347                         page = pfn_to_page(high_pfn);
1348
1349                         /* Update freepage for the list reorder below */
1350                         freepage = page;
1351                 }
1352
1353                 /* Reorder to so a future search skips recent pages */
1354                 move_freelist_head(freelist, freepage);
1355
1356                 /* Isolate the page if available */
1357                 if (page) {
1358                         if (__isolate_free_page(page, order)) {
1359                                 set_page_private(page, order);
1360                                 nr_isolated = 1 << order;
1361                                 cc->nr_freepages += nr_isolated;
1362                                 list_add_tail(&page->lru, &cc->freepages);
1363                                 count_compact_events(COMPACTISOLATED, nr_isolated);
1364                         } else {
1365                                 /* If isolation fails, abort the search */
1366                                 order = -1;
1367                                 page = NULL;
1368                         }
1369                 }
1370
1371                 spin_unlock_irqrestore(&cc->zone->lock, flags);
1372
1373                 /*
1374                  * Smaller scan on next order so the total scan ig related
1375                  * to freelist_scan_limit.
1376                  */
1377                 if (order_scanned >= limit)
1378                         limit = min(1U, limit >> 1);
1379         }
1380
1381         if (!page) {
1382                 cc->fast_search_fail++;
1383                 if (scan_start) {
1384                         /*
1385                          * Use the highest PFN found above min. If one was
1386                          * not found, be pessemistic for direct compaction
1387                          * and use the min mark.
1388                          */
1389                         if (highest) {
1390                                 page = pfn_to_page(highest);
1391                                 cc->free_pfn = highest;
1392                         } else {
1393                                 if (cc->direct_compaction) {
1394                                         page = pfn_to_page(min_pfn);
1395                                         cc->free_pfn = min_pfn;
1396                                 }
1397                         }
1398                 }
1399         }
1400
1401         if (highest && highest >= cc->zone->compact_cached_free_pfn) {
1402                 highest -= pageblock_nr_pages;
1403                 cc->zone->compact_cached_free_pfn = highest;
1404         }
1405
1406         cc->total_free_scanned += nr_scanned;
1407         if (!page)
1408                 return cc->free_pfn;
1409
1410         low_pfn = page_to_pfn(page);
1411         fast_isolate_around(cc, low_pfn, nr_isolated);
1412         return low_pfn;
1413 }
1414
1415 /*
1416  * Based on information in the current compact_control, find blocks
1417  * suitable for isolating free pages from and then isolate them.
1418  */
1419 static void isolate_freepages(struct compact_control *cc)
1420 {
1421         struct zone *zone = cc->zone;
1422         struct page *page;
1423         unsigned long block_start_pfn;  /* start of current pageblock */
1424         unsigned long isolate_start_pfn; /* exact pfn we start at */
1425         unsigned long block_end_pfn;    /* end of current pageblock */
1426         unsigned long low_pfn;       /* lowest pfn scanner is able to scan */
1427         struct list_head *freelist = &cc->freepages;
1428         unsigned int stride;
1429
1430         /* Try a small search of the free lists for a candidate */
1431         isolate_start_pfn = fast_isolate_freepages(cc);
1432         if (cc->nr_freepages)
1433                 goto splitmap;
1434
1435         /*
1436          * Initialise the free scanner. The starting point is where we last
1437          * successfully isolated from, zone-cached value, or the end of the
1438          * zone when isolating for the first time. For looping we also need
1439          * this pfn aligned down to the pageblock boundary, because we do
1440          * block_start_pfn -= pageblock_nr_pages in the for loop.
1441          * For ending point, take care when isolating in last pageblock of a
1442          * a zone which ends in the middle of a pageblock.
1443          * The low boundary is the end of the pageblock the migration scanner
1444          * is using.
1445          */
1446         isolate_start_pfn = cc->free_pfn;
1447         block_start_pfn = pageblock_start_pfn(isolate_start_pfn);
1448         block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1449                                                 zone_end_pfn(zone));
1450         low_pfn = pageblock_end_pfn(cc->migrate_pfn);
1451         stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1;
1452
1453         /*
1454          * Isolate free pages until enough are available to migrate the
1455          * pages on cc->migratepages. We stop searching if the migrate
1456          * and free page scanners meet or enough free pages are isolated.
1457          */
1458         for (; block_start_pfn >= low_pfn;
1459                                 block_end_pfn = block_start_pfn,
1460                                 block_start_pfn -= pageblock_nr_pages,
1461                                 isolate_start_pfn = block_start_pfn) {
1462                 unsigned long nr_isolated;
1463
1464                 /*
1465                  * This can iterate a massively long zone without finding any
1466                  * suitable migration targets, so periodically check resched.
1467                  */
1468                 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)))
1469                         cond_resched();
1470
1471                 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1472                                                                         zone);
1473                 if (!page)
1474                         continue;
1475
1476                 /* Check the block is suitable for migration */
1477                 if (!suitable_migration_target(cc, page))
1478                         continue;
1479
1480                 /* If isolation recently failed, do not retry */
1481                 if (!isolation_suitable(cc, page))
1482                         continue;
1483
1484                 /* Found a block suitable for isolating free pages from. */
1485                 nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn,
1486                                         block_end_pfn, freelist, stride, false);
1487
1488                 /* Update the skip hint if the full pageblock was scanned */
1489                 if (isolate_start_pfn == block_end_pfn)
1490                         update_pageblock_skip(cc, page, block_start_pfn);
1491
1492                 /* Are enough freepages isolated? */
1493                 if (cc->nr_freepages >= cc->nr_migratepages) {
1494                         if (isolate_start_pfn >= block_end_pfn) {
1495                                 /*
1496                                  * Restart at previous pageblock if more
1497                                  * freepages can be isolated next time.
1498                                  */
1499                                 isolate_start_pfn =
1500                                         block_start_pfn - pageblock_nr_pages;
1501                         }
1502                         break;
1503                 } else if (isolate_start_pfn < block_end_pfn) {
1504                         /*
1505                          * If isolation failed early, do not continue
1506                          * needlessly.
1507                          */
1508                         break;
1509                 }
1510
1511                 /* Adjust stride depending on isolation */
1512                 if (nr_isolated) {
1513                         stride = 1;
1514                         continue;
1515                 }
1516                 stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1);
1517         }
1518
1519         /*
1520          * Record where the free scanner will restart next time. Either we
1521          * broke from the loop and set isolate_start_pfn based on the last
1522          * call to isolate_freepages_block(), or we met the migration scanner
1523          * and the loop terminated due to isolate_start_pfn < low_pfn
1524          */
1525         cc->free_pfn = isolate_start_pfn;
1526
1527 splitmap:
1528         /* __isolate_free_page() does not map the pages */
1529         split_map_pages(freelist);
1530 }
1531
1532 /*
1533  * This is a migrate-callback that "allocates" freepages by taking pages
1534  * from the isolated freelists in the block we are migrating to.
1535  */
1536 static struct page *compaction_alloc(struct page *migratepage,
1537                                         unsigned long data)
1538 {
1539         struct compact_control *cc = (struct compact_control *)data;
1540         struct page *freepage;
1541
1542         if (list_empty(&cc->freepages)) {
1543                 isolate_freepages(cc);
1544
1545                 if (list_empty(&cc->freepages))
1546                         return NULL;
1547         }
1548
1549         freepage = list_entry(cc->freepages.next, struct page, lru);
1550         list_del(&freepage->lru);
1551         cc->nr_freepages--;
1552
1553         return freepage;
1554 }
1555
1556 /*
1557  * This is a migrate-callback that "frees" freepages back to the isolated
1558  * freelist.  All pages on the freelist are from the same zone, so there is no
1559  * special handling needed for NUMA.
1560  */
1561 static void compaction_free(struct page *page, unsigned long data)
1562 {
1563         struct compact_control *cc = (struct compact_control *)data;
1564
1565         list_add(&page->lru, &cc->freepages);
1566         cc->nr_freepages++;
1567 }
1568
1569 /* possible outcome of isolate_migratepages */
1570 typedef enum {
1571         ISOLATE_ABORT,          /* Abort compaction now */
1572         ISOLATE_NONE,           /* No pages isolated, continue scanning */
1573         ISOLATE_SUCCESS,        /* Pages isolated, migrate */
1574 } isolate_migrate_t;
1575
1576 /*
1577  * Allow userspace to control policy on scanning the unevictable LRU for
1578  * compactable pages.
1579  */
1580 int sysctl_compact_unevictable_allowed __read_mostly = 1;
1581
1582 static inline void
1583 update_fast_start_pfn(struct compact_control *cc, unsigned long pfn)
1584 {
1585         if (cc->fast_start_pfn == ULONG_MAX)
1586                 return;
1587
1588         if (!cc->fast_start_pfn)
1589                 cc->fast_start_pfn = pfn;
1590
1591         cc->fast_start_pfn = min(cc->fast_start_pfn, pfn);
1592 }
1593
1594 static inline unsigned long
1595 reinit_migrate_pfn(struct compact_control *cc)
1596 {
1597         if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX)
1598                 return cc->migrate_pfn;
1599
1600         cc->migrate_pfn = cc->fast_start_pfn;
1601         cc->fast_start_pfn = ULONG_MAX;
1602
1603         return cc->migrate_pfn;
1604 }
1605
1606 /*
1607  * Briefly search the free lists for a migration source that already has
1608  * some free pages to reduce the number of pages that need migration
1609  * before a pageblock is free.
1610  */
1611 static unsigned long fast_find_migrateblock(struct compact_control *cc)
1612 {
1613         unsigned int limit = freelist_scan_limit(cc);
1614         unsigned int nr_scanned = 0;
1615         unsigned long distance;
1616         unsigned long pfn = cc->migrate_pfn;
1617         unsigned long high_pfn;
1618         int order;
1619
1620         /* Skip hints are relied on to avoid repeats on the fast search */
1621         if (cc->ignore_skip_hint)
1622                 return pfn;
1623
1624         /*
1625          * If the migrate_pfn is not at the start of a zone or the start
1626          * of a pageblock then assume this is a continuation of a previous
1627          * scan restarted due to COMPACT_CLUSTER_MAX.
1628          */
1629         if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn))
1630                 return pfn;
1631
1632         /*
1633          * For smaller orders, just linearly scan as the number of pages
1634          * to migrate should be relatively small and does not necessarily
1635          * justify freeing up a large block for a small allocation.
1636          */
1637         if (cc->order <= PAGE_ALLOC_COSTLY_ORDER)
1638                 return pfn;
1639
1640         /*
1641          * Only allow kcompactd and direct requests for movable pages to
1642          * quickly clear out a MOVABLE pageblock for allocation. This
1643          * reduces the risk that a large movable pageblock is freed for
1644          * an unmovable/reclaimable small allocation.
1645          */
1646         if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE)
1647                 return pfn;
1648
1649         /*
1650          * When starting the migration scanner, pick any pageblock within the
1651          * first half of the search space. Otherwise try and pick a pageblock
1652          * within the first eighth to reduce the chances that a migration
1653          * target later becomes a source.
1654          */
1655         distance = (cc->free_pfn - cc->migrate_pfn) >> 1;
1656         if (cc->migrate_pfn != cc->zone->zone_start_pfn)
1657                 distance >>= 2;
1658         high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance);
1659
1660         for (order = cc->order - 1;
1661              order >= PAGE_ALLOC_COSTLY_ORDER && pfn == cc->migrate_pfn && nr_scanned < limit;
1662              order--) {
1663                 struct free_area *area = &cc->zone->free_area[order];
1664                 struct list_head *freelist;
1665                 unsigned long flags;
1666                 struct page *freepage;
1667
1668                 if (!area->nr_free)
1669                         continue;
1670
1671                 spin_lock_irqsave(&cc->zone->lock, flags);
1672                 freelist = &area->free_list[MIGRATE_MOVABLE];
1673                 list_for_each_entry(freepage, freelist, lru) {
1674                         unsigned long free_pfn;
1675
1676                         nr_scanned++;
1677                         free_pfn = page_to_pfn(freepage);
1678                         if (free_pfn < high_pfn) {
1679                                 /*
1680                                  * Avoid if skipped recently. Ideally it would
1681                                  * move to the tail but even safe iteration of
1682                                  * the list assumes an entry is deleted, not
1683                                  * reordered.
1684                                  */
1685                                 if (get_pageblock_skip(freepage)) {
1686                                         if (list_is_last(freelist, &freepage->lru))
1687                                                 break;
1688
1689                                         continue;
1690                                 }
1691
1692                                 /* Reorder to so a future search skips recent pages */
1693                                 move_freelist_tail(freelist, freepage);
1694
1695                                 update_fast_start_pfn(cc, free_pfn);
1696                                 pfn = pageblock_start_pfn(free_pfn);
1697                                 cc->fast_search_fail = 0;
1698                                 set_pageblock_skip(freepage);
1699                                 break;
1700                         }
1701
1702                         if (nr_scanned >= limit) {
1703                                 cc->fast_search_fail++;
1704                                 move_freelist_tail(freelist, freepage);
1705                                 break;
1706                         }
1707                 }
1708                 spin_unlock_irqrestore(&cc->zone->lock, flags);
1709         }
1710
1711         cc->total_migrate_scanned += nr_scanned;
1712
1713         /*
1714          * If fast scanning failed then use a cached entry for a page block
1715          * that had free pages as the basis for starting a linear scan.
1716          */
1717         if (pfn == cc->migrate_pfn)
1718                 pfn = reinit_migrate_pfn(cc);
1719
1720         return pfn;
1721 }
1722
1723 /*
1724  * Isolate all pages that can be migrated from the first suitable block,
1725  * starting at the block pointed to by the migrate scanner pfn within
1726  * compact_control.
1727  */
1728 static isolate_migrate_t isolate_migratepages(struct zone *zone,
1729                                         struct compact_control *cc)
1730 {
1731         unsigned long block_start_pfn;
1732         unsigned long block_end_pfn;
1733         unsigned long low_pfn;
1734         struct page *page;
1735         const isolate_mode_t isolate_mode =
1736                 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1737                 (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1738         bool fast_find_block;
1739
1740         /*
1741          * Start at where we last stopped, or beginning of the zone as
1742          * initialized by compact_zone(). The first failure will use
1743          * the lowest PFN as the starting point for linear scanning.
1744          */
1745         low_pfn = fast_find_migrateblock(cc);
1746         block_start_pfn = pageblock_start_pfn(low_pfn);
1747         if (block_start_pfn < zone->zone_start_pfn)
1748                 block_start_pfn = zone->zone_start_pfn;
1749
1750         /*
1751          * fast_find_migrateblock marks a pageblock skipped so to avoid
1752          * the isolation_suitable check below, check whether the fast
1753          * search was successful.
1754          */
1755         fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail;
1756
1757         /* Only scan within a pageblock boundary */
1758         block_end_pfn = pageblock_end_pfn(low_pfn);
1759
1760         /*
1761          * Iterate over whole pageblocks until we find the first suitable.
1762          * Do not cross the free scanner.
1763          */
1764         for (; block_end_pfn <= cc->free_pfn;
1765                         fast_find_block = false,
1766                         low_pfn = block_end_pfn,
1767                         block_start_pfn = block_end_pfn,
1768                         block_end_pfn += pageblock_nr_pages) {
1769
1770                 /*
1771                  * This can potentially iterate a massively long zone with
1772                  * many pageblocks unsuitable, so periodically check if we
1773                  * need to schedule.
1774                  */
1775                 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)))
1776                         cond_resched();
1777
1778                 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1779                                                                         zone);
1780                 if (!page)
1781                         continue;
1782
1783                 /*
1784                  * If isolation recently failed, do not retry. Only check the
1785                  * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock
1786                  * to be visited multiple times. Assume skip was checked
1787                  * before making it "skip" so other compaction instances do
1788                  * not scan the same block.
1789                  */
1790                 if (IS_ALIGNED(low_pfn, pageblock_nr_pages) &&
1791                     !fast_find_block && !isolation_suitable(cc, page))
1792                         continue;
1793
1794                 /*
1795                  * For async compaction, also only scan in MOVABLE blocks
1796                  * without huge pages. Async compaction is optimistic to see
1797                  * if the minimum amount of work satisfies the allocation.
1798                  * The cached PFN is updated as it's possible that all
1799                  * remaining blocks between source and target are unsuitable
1800                  * and the compaction scanners fail to meet.
1801                  */
1802                 if (!suitable_migration_source(cc, page)) {
1803                         update_cached_migrate(cc, block_end_pfn);
1804                         continue;
1805                 }
1806
1807                 /* Perform the isolation */
1808                 low_pfn = isolate_migratepages_block(cc, low_pfn,
1809                                                 block_end_pfn, isolate_mode);
1810
1811                 if (!low_pfn)
1812                         return ISOLATE_ABORT;
1813
1814                 /*
1815                  * Either we isolated something and proceed with migration. Or
1816                  * we failed and compact_zone should decide if we should
1817                  * continue or not.
1818                  */
1819                 break;
1820         }
1821
1822         /* Record where migration scanner will be restarted. */
1823         cc->migrate_pfn = low_pfn;
1824
1825         return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
1826 }
1827
1828 /*
1829  * order == -1 is expected when compacting via
1830  * /proc/sys/vm/compact_memory
1831  */
1832 static inline bool is_via_compact_memory(int order)
1833 {
1834         return order == -1;
1835 }
1836
1837 static enum compact_result __compact_finished(struct compact_control *cc)
1838 {
1839         unsigned int order;
1840         const int migratetype = cc->migratetype;
1841         int ret;
1842
1843         /* Compaction run completes if the migrate and free scanner meet */
1844         if (compact_scanners_met(cc)) {
1845                 /* Let the next compaction start anew. */
1846                 reset_cached_positions(cc->zone);
1847
1848                 /*
1849                  * Mark that the PG_migrate_skip information should be cleared
1850                  * by kswapd when it goes to sleep. kcompactd does not set the
1851                  * flag itself as the decision to be clear should be directly
1852                  * based on an allocation request.
1853                  */
1854                 if (cc->direct_compaction)
1855                         cc->zone->compact_blockskip_flush = true;
1856
1857                 if (cc->whole_zone)
1858                         return COMPACT_COMPLETE;
1859                 else
1860                         return COMPACT_PARTIAL_SKIPPED;
1861         }
1862
1863         if (is_via_compact_memory(cc->order))
1864                 return COMPACT_CONTINUE;
1865
1866         /*
1867          * Always finish scanning a pageblock to reduce the possibility of
1868          * fallbacks in the future. This is particularly important when
1869          * migration source is unmovable/reclaimable but it's not worth
1870          * special casing.
1871          */
1872         if (!IS_ALIGNED(cc->migrate_pfn, pageblock_nr_pages))
1873                 return COMPACT_CONTINUE;
1874
1875         /* Direct compactor: Is a suitable page free? */
1876         ret = COMPACT_NO_SUITABLE_PAGE;
1877         for (order = cc->order; order < MAX_ORDER; order++) {
1878                 struct free_area *area = &cc->zone->free_area[order];
1879                 bool can_steal;
1880
1881                 /* Job done if page is free of the right migratetype */
1882                 if (!list_empty(&area->free_list[migratetype]))
1883                         return COMPACT_SUCCESS;
1884
1885 #ifdef CONFIG_CMA
1886                 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
1887                 if (migratetype == MIGRATE_MOVABLE &&
1888                         !list_empty(&area->free_list[MIGRATE_CMA]))
1889                         return COMPACT_SUCCESS;
1890 #endif
1891                 /*
1892                  * Job done if allocation would steal freepages from
1893                  * other migratetype buddy lists.
1894                  */
1895                 if (find_suitable_fallback(area, order, migratetype,
1896                                                 true, &can_steal) != -1) {
1897
1898                         /* movable pages are OK in any pageblock */
1899                         if (migratetype == MIGRATE_MOVABLE)
1900                                 return COMPACT_SUCCESS;
1901
1902                         /*
1903                          * We are stealing for a non-movable allocation. Make
1904                          * sure we finish compacting the current pageblock
1905                          * first so it is as free as possible and we won't
1906                          * have to steal another one soon. This only applies
1907                          * to sync compaction, as async compaction operates
1908                          * on pageblocks of the same migratetype.
1909                          */
1910                         if (cc->mode == MIGRATE_ASYNC ||
1911                                         IS_ALIGNED(cc->migrate_pfn,
1912                                                         pageblock_nr_pages)) {
1913                                 return COMPACT_SUCCESS;
1914                         }
1915
1916                         ret = COMPACT_CONTINUE;
1917                         break;
1918                 }
1919         }
1920
1921         if (cc->contended || fatal_signal_pending(current))
1922                 ret = COMPACT_CONTENDED;
1923
1924         return ret;
1925 }
1926
1927 static enum compact_result compact_finished(struct compact_control *cc)
1928 {
1929         int ret;
1930
1931         ret = __compact_finished(cc);
1932         trace_mm_compaction_finished(cc->zone, cc->order, ret);
1933         if (ret == COMPACT_NO_SUITABLE_PAGE)
1934                 ret = COMPACT_CONTINUE;
1935
1936         return ret;
1937 }
1938
1939 /*
1940  * compaction_suitable: Is this suitable to run compaction on this zone now?
1941  * Returns
1942  *   COMPACT_SKIPPED  - If there are too few free pages for compaction
1943  *   COMPACT_SUCCESS  - If the allocation would succeed without compaction
1944  *   COMPACT_CONTINUE - If compaction should run now
1945  */
1946 static enum compact_result __compaction_suitable(struct zone *zone, int order,
1947                                         unsigned int alloc_flags,
1948                                         int classzone_idx,
1949                                         unsigned long wmark_target)
1950 {
1951         unsigned long watermark;
1952
1953         if (is_via_compact_memory(order))
1954                 return COMPACT_CONTINUE;
1955
1956         watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
1957         /*
1958          * If watermarks for high-order allocation are already met, there
1959          * should be no need for compaction at all.
1960          */
1961         if (zone_watermark_ok(zone, order, watermark, classzone_idx,
1962                                                                 alloc_flags))
1963                 return COMPACT_SUCCESS;
1964
1965         /*
1966          * Watermarks for order-0 must be met for compaction to be able to
1967          * isolate free pages for migration targets. This means that the
1968          * watermark and alloc_flags have to match, or be more pessimistic than
1969          * the check in __isolate_free_page(). We don't use the direct
1970          * compactor's alloc_flags, as they are not relevant for freepage
1971          * isolation. We however do use the direct compactor's classzone_idx to
1972          * skip over zones where lowmem reserves would prevent allocation even
1973          * if compaction succeeds.
1974          * For costly orders, we require low watermark instead of min for
1975          * compaction to proceed to increase its chances.
1976          * ALLOC_CMA is used, as pages in CMA pageblocks are considered
1977          * suitable migration targets
1978          */
1979         watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
1980                                 low_wmark_pages(zone) : min_wmark_pages(zone);
1981         watermark += compact_gap(order);
1982         if (!__zone_watermark_ok(zone, 0, watermark, classzone_idx,
1983                                                 ALLOC_CMA, wmark_target))
1984                 return COMPACT_SKIPPED;
1985
1986         return COMPACT_CONTINUE;
1987 }
1988
1989 enum compact_result compaction_suitable(struct zone *zone, int order,
1990                                         unsigned int alloc_flags,
1991                                         int classzone_idx)
1992 {
1993         enum compact_result ret;
1994         int fragindex;
1995
1996         ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx,
1997                                     zone_page_state(zone, NR_FREE_PAGES));
1998         /*
1999          * fragmentation index determines if allocation failures are due to
2000          * low memory or external fragmentation
2001          *
2002          * index of -1000 would imply allocations might succeed depending on
2003          * watermarks, but we already failed the high-order watermark check
2004          * index towards 0 implies failure is due to lack of memory
2005          * index towards 1000 implies failure is due to fragmentation
2006          *
2007          * Only compact if a failure would be due to fragmentation. Also
2008          * ignore fragindex for non-costly orders where the alternative to
2009          * a successful reclaim/compaction is OOM. Fragindex and the
2010          * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
2011          * excessive compaction for costly orders, but it should not be at the
2012          * expense of system stability.
2013          */
2014         if (ret == COMPACT_CONTINUE && (order > PAGE_ALLOC_COSTLY_ORDER)) {
2015                 fragindex = fragmentation_index(zone, order);
2016                 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
2017                         ret = COMPACT_NOT_SUITABLE_ZONE;
2018         }
2019
2020         trace_mm_compaction_suitable(zone, order, ret);
2021         if (ret == COMPACT_NOT_SUITABLE_ZONE)
2022                 ret = COMPACT_SKIPPED;
2023
2024         return ret;
2025 }
2026
2027 bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
2028                 int alloc_flags)
2029 {
2030         struct zone *zone;
2031         struct zoneref *z;
2032
2033         /*
2034          * Make sure at least one zone would pass __compaction_suitable if we continue
2035          * retrying the reclaim.
2036          */
2037         for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
2038                                         ac->nodemask) {
2039                 unsigned long available;
2040                 enum compact_result compact_result;
2041
2042                 /*
2043                  * Do not consider all the reclaimable memory because we do not
2044                  * want to trash just for a single high order allocation which
2045                  * is even not guaranteed to appear even if __compaction_suitable
2046                  * is happy about the watermark check.
2047                  */
2048                 available = zone_reclaimable_pages(zone) / order;
2049                 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
2050                 compact_result = __compaction_suitable(zone, order, alloc_flags,
2051                                 ac_classzone_idx(ac), available);
2052                 if (compact_result != COMPACT_SKIPPED)
2053                         return true;
2054         }
2055
2056         return false;
2057 }
2058
2059 static enum compact_result
2060 compact_zone(struct compact_control *cc, struct capture_control *capc)
2061 {
2062         enum compact_result ret;
2063         unsigned long start_pfn = cc->zone->zone_start_pfn;
2064         unsigned long end_pfn = zone_end_pfn(cc->zone);
2065         unsigned long last_migrated_pfn;
2066         const bool sync = cc->mode != MIGRATE_ASYNC;
2067         bool update_cached;
2068
2069         cc->migratetype = gfpflags_to_migratetype(cc->gfp_mask);
2070         ret = compaction_suitable(cc->zone, cc->order, cc->alloc_flags,
2071                                                         cc->classzone_idx);
2072         /* Compaction is likely to fail */
2073         if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED)
2074                 return ret;
2075
2076         /* huh, compaction_suitable is returning something unexpected */
2077         VM_BUG_ON(ret != COMPACT_CONTINUE);
2078
2079         /*
2080          * Clear pageblock skip if there were failures recently and compaction
2081          * is about to be retried after being deferred.
2082          */
2083         if (compaction_restarting(cc->zone, cc->order))
2084                 __reset_isolation_suitable(cc->zone);
2085
2086         /*
2087          * Setup to move all movable pages to the end of the zone. Used cached
2088          * information on where the scanners should start (unless we explicitly
2089          * want to compact the whole zone), but check that it is initialised
2090          * by ensuring the values are within zone boundaries.
2091          */
2092         cc->fast_start_pfn = 0;
2093         if (cc->whole_zone) {
2094                 cc->migrate_pfn = start_pfn;
2095                 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2096         } else {
2097                 cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync];
2098                 cc->free_pfn = cc->zone->compact_cached_free_pfn;
2099                 if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
2100                         cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2101                         cc->zone->compact_cached_free_pfn = cc->free_pfn;
2102                 }
2103                 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
2104                         cc->migrate_pfn = start_pfn;
2105                         cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
2106                         cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
2107                 }
2108
2109                 if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn)
2110                         cc->whole_zone = true;
2111         }
2112
2113         last_migrated_pfn = 0;
2114
2115         /*
2116          * Migrate has separate cached PFNs for ASYNC and SYNC* migration on
2117          * the basis that some migrations will fail in ASYNC mode. However,
2118          * if the cached PFNs match and pageblocks are skipped due to having
2119          * no isolation candidates, then the sync state does not matter.
2120          * Until a pageblock with isolation candidates is found, keep the
2121          * cached PFNs in sync to avoid revisiting the same blocks.
2122          */
2123         update_cached = !sync &&
2124                 cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1];
2125
2126         trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
2127                                 cc->free_pfn, end_pfn, sync);
2128
2129         migrate_prep_local();
2130
2131         while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) {
2132                 int err;
2133                 unsigned long start_pfn = cc->migrate_pfn;
2134
2135                 /*
2136                  * Avoid multiple rescans which can happen if a page cannot be
2137                  * isolated (dirty/writeback in async mode) or if the migrated
2138                  * pages are being allocated before the pageblock is cleared.
2139                  * The first rescan will capture the entire pageblock for
2140                  * migration. If it fails, it'll be marked skip and scanning
2141                  * will proceed as normal.
2142                  */
2143                 cc->rescan = false;
2144                 if (pageblock_start_pfn(last_migrated_pfn) ==
2145                     pageblock_start_pfn(start_pfn)) {
2146                         cc->rescan = true;
2147                 }
2148
2149                 switch (isolate_migratepages(cc->zone, cc)) {
2150                 case ISOLATE_ABORT:
2151                         ret = COMPACT_CONTENDED;
2152                         putback_movable_pages(&cc->migratepages);
2153                         cc->nr_migratepages = 0;
2154                         last_migrated_pfn = 0;
2155                         goto out;
2156                 case ISOLATE_NONE:
2157                         if (update_cached) {
2158                                 cc->zone->compact_cached_migrate_pfn[1] =
2159                                         cc->zone->compact_cached_migrate_pfn[0];
2160                         }
2161
2162                         /*
2163                          * We haven't isolated and migrated anything, but
2164                          * there might still be unflushed migrations from
2165                          * previous cc->order aligned block.
2166                          */
2167                         goto check_drain;
2168                 case ISOLATE_SUCCESS:
2169                         update_cached = false;
2170                         last_migrated_pfn = start_pfn;
2171                         ;
2172                 }
2173
2174                 err = migrate_pages(&cc->migratepages, compaction_alloc,
2175                                 compaction_free, (unsigned long)cc, cc->mode,
2176                                 MR_COMPACTION);
2177
2178                 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
2179                                                         &cc->migratepages);
2180
2181                 /* All pages were either migrated or will be released */
2182                 cc->nr_migratepages = 0;
2183                 if (err) {
2184                         putback_movable_pages(&cc->migratepages);
2185                         /*
2186                          * migrate_pages() may return -ENOMEM when scanners meet
2187                          * and we want compact_finished() to detect it
2188                          */
2189                         if (err == -ENOMEM && !compact_scanners_met(cc)) {
2190                                 ret = COMPACT_CONTENDED;
2191                                 goto out;
2192                         }
2193                         /*
2194                          * We failed to migrate at least one page in the current
2195                          * order-aligned block, so skip the rest of it.
2196                          */
2197                         if (cc->direct_compaction &&
2198                                                 (cc->mode == MIGRATE_ASYNC)) {
2199                                 cc->migrate_pfn = block_end_pfn(
2200                                                 cc->migrate_pfn - 1, cc->order);
2201                                 /* Draining pcplists is useless in this case */
2202                                 last_migrated_pfn = 0;
2203                         }
2204                 }
2205
2206 check_drain:
2207                 /*
2208                  * Has the migration scanner moved away from the previous
2209                  * cc->order aligned block where we migrated from? If yes,
2210                  * flush the pages that were freed, so that they can merge and
2211                  * compact_finished() can detect immediately if allocation
2212                  * would succeed.
2213                  */
2214                 if (cc->order > 0 && last_migrated_pfn) {
2215                         int cpu;
2216                         unsigned long current_block_start =
2217                                 block_start_pfn(cc->migrate_pfn, cc->order);
2218
2219                         if (last_migrated_pfn < current_block_start) {
2220                                 cpu = get_cpu();
2221                                 lru_add_drain_cpu(cpu);
2222                                 drain_local_pages(cc->zone);
2223                                 put_cpu();
2224                                 /* No more flushing until we migrate again */
2225                                 last_migrated_pfn = 0;
2226                         }
2227                 }
2228
2229                 /* Stop if a page has been captured */
2230                 if (capc && capc->page) {
2231                         ret = COMPACT_SUCCESS;
2232                         break;
2233                 }
2234         }
2235
2236 out:
2237         /*
2238          * Release free pages and update where the free scanner should restart,
2239          * so we don't leave any returned pages behind in the next attempt.
2240          */
2241         if (cc->nr_freepages > 0) {
2242                 unsigned long free_pfn = release_freepages(&cc->freepages);
2243
2244                 cc->nr_freepages = 0;
2245                 VM_BUG_ON(free_pfn == 0);
2246                 /* The cached pfn is always the first in a pageblock */
2247                 free_pfn = pageblock_start_pfn(free_pfn);
2248                 /*
2249                  * Only go back, not forward. The cached pfn might have been
2250                  * already reset to zone end in compact_finished()
2251                  */
2252                 if (free_pfn > cc->zone->compact_cached_free_pfn)
2253                         cc->zone->compact_cached_free_pfn = free_pfn;
2254         }
2255
2256         count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned);
2257         count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned);
2258
2259         trace_mm_compaction_end(start_pfn, cc->migrate_pfn,
2260                                 cc->free_pfn, end_pfn, sync, ret);
2261
2262         return ret;
2263 }
2264
2265 static enum compact_result compact_zone_order(struct zone *zone, int order,
2266                 gfp_t gfp_mask, enum compact_priority prio,
2267                 unsigned int alloc_flags, int classzone_idx,
2268                 struct page **capture)
2269 {
2270         enum compact_result ret;
2271         struct compact_control cc = {
2272                 .nr_freepages = 0,
2273                 .nr_migratepages = 0,
2274                 .total_migrate_scanned = 0,
2275                 .total_free_scanned = 0,
2276                 .order = order,
2277                 .search_order = order,
2278                 .gfp_mask = gfp_mask,
2279                 .zone = zone,
2280                 .mode = (prio == COMPACT_PRIO_ASYNC) ?
2281                                         MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT,
2282                 .alloc_flags = alloc_flags,
2283                 .classzone_idx = classzone_idx,
2284                 .direct_compaction = true,
2285                 .whole_zone = (prio == MIN_COMPACT_PRIORITY),
2286                 .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
2287                 .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
2288         };
2289         struct capture_control capc = {
2290                 .cc = &cc,
2291                 .page = NULL,
2292         };
2293
2294         if (capture)
2295                 current->capture_control = &capc;
2296         INIT_LIST_HEAD(&cc.freepages);
2297         INIT_LIST_HEAD(&cc.migratepages);
2298
2299         ret = compact_zone(&cc, &capc);
2300
2301         VM_BUG_ON(!list_empty(&cc.freepages));
2302         VM_BUG_ON(!list_empty(&cc.migratepages));
2303
2304         *capture = capc.page;
2305         current->capture_control = NULL;
2306
2307         return ret;
2308 }
2309
2310 int sysctl_extfrag_threshold = 500;
2311
2312 /**
2313  * try_to_compact_pages - Direct compact to satisfy a high-order allocation
2314  * @gfp_mask: The GFP mask of the current allocation
2315  * @order: The order of the current allocation
2316  * @alloc_flags: The allocation flags of the current allocation
2317  * @ac: The context of current allocation
2318  * @prio: Determines how hard direct compaction should try to succeed
2319  *
2320  * This is the main entry point for direct page compaction.
2321  */
2322 enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
2323                 unsigned int alloc_flags, const struct alloc_context *ac,
2324                 enum compact_priority prio, struct page **capture)
2325 {
2326         int may_perform_io = gfp_mask & __GFP_IO;
2327         struct zoneref *z;
2328         struct zone *zone;
2329         enum compact_result rc = COMPACT_SKIPPED;
2330
2331         /*
2332          * Check if the GFP flags allow compaction - GFP_NOIO is really
2333          * tricky context because the migration might require IO
2334          */
2335         if (!may_perform_io)
2336                 return COMPACT_SKIPPED;
2337
2338         trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
2339
2340         /* Compact each zone in the list */
2341         for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
2342                                                                 ac->nodemask) {
2343                 enum compact_result status;
2344
2345                 if (prio > MIN_COMPACT_PRIORITY
2346                                         && compaction_deferred(zone, order)) {
2347                         rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
2348                         continue;
2349                 }
2350
2351                 status = compact_zone_order(zone, order, gfp_mask, prio,
2352                                 alloc_flags, ac_classzone_idx(ac), capture);
2353                 rc = max(status, rc);
2354
2355                 /* The allocation should succeed, stop compacting */
2356                 if (status == COMPACT_SUCCESS) {
2357                         /*
2358                          * We think the allocation will succeed in this zone,
2359                          * but it is not certain, hence the false. The caller
2360                          * will repeat this with true if allocation indeed
2361                          * succeeds in this zone.
2362                          */
2363                         compaction_defer_reset(zone, order, false);
2364
2365                         break;
2366                 }
2367
2368                 if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
2369                                         status == COMPACT_PARTIAL_SKIPPED))
2370                         /*
2371                          * We think that allocation won't succeed in this zone
2372                          * so we defer compaction there. If it ends up
2373                          * succeeding after all, it will be reset.
2374                          */
2375                         defer_compaction(zone, order);
2376
2377                 /*
2378                  * We might have stopped compacting due to need_resched() in
2379                  * async compaction, or due to a fatal signal detected. In that
2380                  * case do not try further zones
2381                  */
2382                 if ((prio == COMPACT_PRIO_ASYNC && need_resched())
2383                                         || fatal_signal_pending(current))
2384                         break;
2385         }
2386
2387         return rc;
2388 }
2389
2390
2391 /* Compact all zones within a node */
2392 static void compact_node(int nid)
2393 {
2394         pg_data_t *pgdat = NODE_DATA(nid);
2395         int zoneid;
2396         struct zone *zone;
2397         struct compact_control cc = {
2398                 .order = -1,
2399                 .total_migrate_scanned = 0,
2400                 .total_free_scanned = 0,
2401                 .mode = MIGRATE_SYNC,
2402                 .ignore_skip_hint = true,
2403                 .whole_zone = true,
2404                 .gfp_mask = GFP_KERNEL,
2405         };
2406
2407
2408         for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2409
2410                 zone = &pgdat->node_zones[zoneid];
2411                 if (!populated_zone(zone))
2412                         continue;
2413
2414                 cc.nr_freepages = 0;
2415                 cc.nr_migratepages = 0;
2416                 cc.zone = zone;
2417                 INIT_LIST_HEAD(&cc.freepages);
2418                 INIT_LIST_HEAD(&cc.migratepages);
2419
2420                 compact_zone(&cc, NULL);
2421
2422                 VM_BUG_ON(!list_empty(&cc.freepages));
2423                 VM_BUG_ON(!list_empty(&cc.migratepages));
2424         }
2425 }
2426
2427 /* Compact all nodes in the system */
2428 static void compact_nodes(void)
2429 {
2430         int nid;
2431
2432         /* Flush pending updates to the LRU lists */
2433         lru_add_drain_all();
2434
2435         for_each_online_node(nid)
2436                 compact_node(nid);
2437 }
2438
2439 /* The written value is actually unused, all memory is compacted */
2440 int sysctl_compact_memory;
2441
2442 /*
2443  * This is the entry point for compacting all nodes via
2444  * /proc/sys/vm/compact_memory
2445  */
2446 int sysctl_compaction_handler(struct ctl_table *table, int write,
2447                         void __user *buffer, size_t *length, loff_t *ppos)
2448 {
2449         if (write)
2450                 compact_nodes();
2451
2452         return 0;
2453 }
2454
2455 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
2456 static ssize_t sysfs_compact_node(struct device *dev,
2457                         struct device_attribute *attr,
2458                         const char *buf, size_t count)
2459 {
2460         int nid = dev->id;
2461
2462         if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
2463                 /* Flush pending updates to the LRU lists */
2464                 lru_add_drain_all();
2465
2466                 compact_node(nid);
2467         }
2468
2469         return count;
2470 }
2471 static DEVICE_ATTR(compact, 0200, NULL, sysfs_compact_node);
2472
2473 int compaction_register_node(struct node *node)
2474 {
2475         return device_create_file(&node->dev, &dev_attr_compact);
2476 }
2477
2478 void compaction_unregister_node(struct node *node)
2479 {
2480         return device_remove_file(&node->dev, &dev_attr_compact);
2481 }
2482 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
2483
2484 static inline bool kcompactd_work_requested(pg_data_t *pgdat)
2485 {
2486         return pgdat->kcompactd_max_order > 0 || kthread_should_stop();
2487 }
2488
2489 static bool kcompactd_node_suitable(pg_data_t *pgdat)
2490 {
2491         int zoneid;
2492         struct zone *zone;
2493         enum zone_type classzone_idx = pgdat->kcompactd_classzone_idx;
2494
2495         for (zoneid = 0; zoneid <= classzone_idx; zoneid++) {
2496                 zone = &pgdat->node_zones[zoneid];
2497
2498                 if (!populated_zone(zone))
2499                         continue;
2500
2501                 if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0,
2502                                         classzone_idx) == COMPACT_CONTINUE)
2503                         return true;
2504         }
2505
2506         return false;
2507 }
2508
2509 static void kcompactd_do_work(pg_data_t *pgdat)
2510 {
2511         /*
2512          * With no special task, compact all zones so that a page of requested
2513          * order is allocatable.
2514          */
2515         int zoneid;
2516         struct zone *zone;
2517         struct compact_control cc = {
2518                 .order = pgdat->kcompactd_max_order,
2519                 .search_order = pgdat->kcompactd_max_order,
2520                 .total_migrate_scanned = 0,
2521                 .total_free_scanned = 0,
2522                 .classzone_idx = pgdat->kcompactd_classzone_idx,
2523                 .mode = MIGRATE_SYNC_LIGHT,
2524                 .ignore_skip_hint = false,
2525                 .gfp_mask = GFP_KERNEL,
2526         };
2527         trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
2528                                                         cc.classzone_idx);
2529         count_compact_event(KCOMPACTD_WAKE);
2530
2531         for (zoneid = 0; zoneid <= cc.classzone_idx; zoneid++) {
2532                 int status;
2533
2534                 zone = &pgdat->node_zones[zoneid];
2535                 if (!populated_zone(zone))
2536                         continue;
2537
2538                 if (compaction_deferred(zone, cc.order))
2539                         continue;
2540
2541                 if (compaction_suitable(zone, cc.order, 0, zoneid) !=
2542                                                         COMPACT_CONTINUE)
2543                         continue;
2544
2545                 cc.nr_freepages = 0;
2546                 cc.nr_migratepages = 0;
2547                 cc.total_migrate_scanned = 0;
2548                 cc.total_free_scanned = 0;
2549                 cc.zone = zone;
2550                 INIT_LIST_HEAD(&cc.freepages);
2551                 INIT_LIST_HEAD(&cc.migratepages);
2552
2553                 if (kthread_should_stop())
2554                         return;
2555                 status = compact_zone(&cc, NULL);
2556
2557                 if (status == COMPACT_SUCCESS) {
2558                         compaction_defer_reset(zone, cc.order, false);
2559                 } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
2560                         /*
2561                          * Buddy pages may become stranded on pcps that could
2562                          * otherwise coalesce on the zone's free area for
2563                          * order >= cc.order.  This is ratelimited by the
2564                          * upcoming deferral.
2565                          */
2566                         drain_all_pages(zone);
2567
2568                         /*
2569                          * We use sync migration mode here, so we defer like
2570                          * sync direct compaction does.
2571                          */
2572                         defer_compaction(zone, cc.order);
2573                 }
2574
2575                 count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
2576                                      cc.total_migrate_scanned);
2577                 count_compact_events(KCOMPACTD_FREE_SCANNED,
2578                                      cc.total_free_scanned);
2579
2580                 VM_BUG_ON(!list_empty(&cc.freepages));
2581                 VM_BUG_ON(!list_empty(&cc.migratepages));
2582         }
2583
2584         /*
2585          * Regardless of success, we are done until woken up next. But remember
2586          * the requested order/classzone_idx in case it was higher/tighter than
2587          * our current ones
2588          */
2589         if (pgdat->kcompactd_max_order <= cc.order)
2590                 pgdat->kcompactd_max_order = 0;
2591         if (pgdat->kcompactd_classzone_idx >= cc.classzone_idx)
2592                 pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
2593 }
2594
2595 void wakeup_kcompactd(pg_data_t *pgdat, int order, int classzone_idx)
2596 {
2597         if (!order)
2598                 return;
2599
2600         if (pgdat->kcompactd_max_order < order)
2601                 pgdat->kcompactd_max_order = order;
2602
2603         if (pgdat->kcompactd_classzone_idx > classzone_idx)
2604                 pgdat->kcompactd_classzone_idx = classzone_idx;
2605
2606         /*
2607          * Pairs with implicit barrier in wait_event_freezable()
2608          * such that wakeups are not missed.
2609          */
2610         if (!wq_has_sleeper(&pgdat->kcompactd_wait))
2611                 return;
2612
2613         if (!kcompactd_node_suitable(pgdat))
2614                 return;
2615
2616         trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
2617                                                         classzone_idx);
2618         wake_up_interruptible(&pgdat->kcompactd_wait);
2619 }
2620
2621 /*
2622  * The background compaction daemon, started as a kernel thread
2623  * from the init process.
2624  */
2625 static int kcompactd(void *p)
2626 {
2627         pg_data_t *pgdat = (pg_data_t*)p;
2628         struct task_struct *tsk = current;
2629
2630         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2631
2632         if (!cpumask_empty(cpumask))
2633                 set_cpus_allowed_ptr(tsk, cpumask);
2634
2635         set_freezable();
2636
2637         pgdat->kcompactd_max_order = 0;
2638         pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
2639
2640         while (!kthread_should_stop()) {
2641                 unsigned long pflags;
2642
2643                 trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
2644                 wait_event_freezable(pgdat->kcompactd_wait,
2645                                 kcompactd_work_requested(pgdat));
2646
2647                 psi_memstall_enter(&pflags);
2648                 kcompactd_do_work(pgdat);
2649                 psi_memstall_leave(&pflags);
2650         }
2651
2652         return 0;
2653 }
2654
2655 /*
2656  * This kcompactd start function will be called by init and node-hot-add.
2657  * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
2658  */
2659 int kcompactd_run(int nid)
2660 {
2661         pg_data_t *pgdat = NODE_DATA(nid);
2662         int ret = 0;
2663
2664         if (pgdat->kcompactd)
2665                 return 0;
2666
2667         pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
2668         if (IS_ERR(pgdat->kcompactd)) {
2669                 pr_err("Failed to start kcompactd on node %d\n", nid);
2670                 ret = PTR_ERR(pgdat->kcompactd);
2671                 pgdat->kcompactd = NULL;
2672         }
2673         return ret;
2674 }
2675
2676 /*
2677  * Called by memory hotplug when all memory in a node is offlined. Caller must
2678  * hold mem_hotplug_begin/end().
2679  */
2680 void kcompactd_stop(int nid)
2681 {
2682         struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
2683
2684         if (kcompactd) {
2685                 kthread_stop(kcompactd);
2686                 NODE_DATA(nid)->kcompactd = NULL;
2687         }
2688 }
2689
2690 /*
2691  * It's optimal to keep kcompactd on the same CPUs as their memory, but
2692  * not required for correctness. So if the last cpu in a node goes
2693  * away, we get changed to run anywhere: as the first one comes back,
2694  * restore their cpu bindings.
2695  */
2696 static int kcompactd_cpu_online(unsigned int cpu)
2697 {
2698         int nid;
2699
2700         for_each_node_state(nid, N_MEMORY) {
2701                 pg_data_t *pgdat = NODE_DATA(nid);
2702                 const struct cpumask *mask;
2703
2704                 mask = cpumask_of_node(pgdat->node_id);
2705
2706                 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2707                         /* One of our CPUs online: restore mask */
2708                         set_cpus_allowed_ptr(pgdat->kcompactd, mask);
2709         }
2710         return 0;
2711 }
2712
2713 static int __init kcompactd_init(void)
2714 {
2715         int nid;
2716         int ret;
2717
2718         ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
2719                                         "mm/compaction:online",
2720                                         kcompactd_cpu_online, NULL);
2721         if (ret < 0) {
2722                 pr_err("kcompactd: failed to register hotplug callbacks.\n");
2723                 return ret;
2724         }
2725
2726         for_each_node_state(nid, N_MEMORY)
2727                 kcompactd_run(nid);
2728         return 0;
2729 }
2730 subsys_initcall(kcompactd_init)
2731
2732 #endif /* CONFIG_COMPACTION */