mm: remove zone_lru_lock() function, access ->lru_lock directly
[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         pg_data_t *pgdat = zone->zone_pgdat;
779         unsigned long nr_scanned = 0, nr_isolated = 0;
780         struct lruvec *lruvec;
781         unsigned long flags = 0;
782         bool locked = false;
783         struct page *page = NULL, *valid_page = NULL;
784         unsigned long start_pfn = low_pfn;
785         bool skip_on_failure = false;
786         unsigned long next_skip_pfn = 0;
787         bool skip_updated = false;
788
789         /*
790          * Ensure that there are not too many pages isolated from the LRU
791          * list by either parallel reclaimers or compaction. If there are,
792          * delay for some time until fewer pages are isolated
793          */
794         while (unlikely(too_many_isolated(zone))) {
795                 /* async migration should just abort */
796                 if (cc->mode == MIGRATE_ASYNC)
797                         return 0;
798
799                 congestion_wait(BLK_RW_ASYNC, HZ/10);
800
801                 if (fatal_signal_pending(current))
802                         return 0;
803         }
804
805         cond_resched();
806
807         if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
808                 skip_on_failure = true;
809                 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
810         }
811
812         /* Time to isolate some pages for migration */
813         for (; low_pfn < end_pfn; low_pfn++) {
814
815                 if (skip_on_failure && low_pfn >= next_skip_pfn) {
816                         /*
817                          * We have isolated all migration candidates in the
818                          * previous order-aligned block, and did not skip it due
819                          * to failure. We should migrate the pages now and
820                          * hopefully succeed compaction.
821                          */
822                         if (nr_isolated)
823                                 break;
824
825                         /*
826                          * We failed to isolate in the previous order-aligned
827                          * block. Set the new boundary to the end of the
828                          * current block. Note we can't simply increase
829                          * next_skip_pfn by 1 << order, as low_pfn might have
830                          * been incremented by a higher number due to skipping
831                          * a compound or a high-order buddy page in the
832                          * previous loop iteration.
833                          */
834                         next_skip_pfn = block_end_pfn(low_pfn, cc->order);
835                 }
836
837                 /*
838                  * Periodically drop the lock (if held) regardless of its
839                  * contention, to give chance to IRQs. Abort async compaction
840                  * if contended.
841                  */
842                 if (!(low_pfn % SWAP_CLUSTER_MAX)
843                     && compact_unlock_should_abort(&pgdat->lru_lock,
844                                             flags, &locked, cc))
845                         break;
846
847                 if (!pfn_valid_within(low_pfn))
848                         goto isolate_fail;
849                 nr_scanned++;
850
851                 page = pfn_to_page(low_pfn);
852
853                 /*
854                  * Check if the pageblock has already been marked skipped.
855                  * Only the aligned PFN is checked as the caller isolates
856                  * COMPACT_CLUSTER_MAX at a time so the second call must
857                  * not falsely conclude that the block should be skipped.
858                  */
859                 if (!valid_page && IS_ALIGNED(low_pfn, pageblock_nr_pages)) {
860                         if (!cc->ignore_skip_hint && get_pageblock_skip(page)) {
861                                 low_pfn = end_pfn;
862                                 goto isolate_abort;
863                         }
864                         valid_page = page;
865                 }
866
867                 /*
868                  * Skip if free. We read page order here without zone lock
869                  * which is generally unsafe, but the race window is small and
870                  * the worst thing that can happen is that we skip some
871                  * potential isolation targets.
872                  */
873                 if (PageBuddy(page)) {
874                         unsigned long freepage_order = page_order_unsafe(page);
875
876                         /*
877                          * Without lock, we cannot be sure that what we got is
878                          * a valid page order. Consider only values in the
879                          * valid order range to prevent low_pfn overflow.
880                          */
881                         if (freepage_order > 0 && freepage_order < MAX_ORDER)
882                                 low_pfn += (1UL << freepage_order) - 1;
883                         continue;
884                 }
885
886                 /*
887                  * Regardless of being on LRU, compound pages such as THP and
888                  * hugetlbfs are not to be compacted. We can potentially save
889                  * a lot of iterations if we skip them at once. The check is
890                  * racy, but we can consider only valid values and the only
891                  * danger is skipping too much.
892                  */
893                 if (PageCompound(page)) {
894                         const unsigned int order = compound_order(page);
895
896                         if (likely(order < MAX_ORDER))
897                                 low_pfn += (1UL << order) - 1;
898                         goto isolate_fail;
899                 }
900
901                 /*
902                  * Check may be lockless but that's ok as we recheck later.
903                  * It's possible to migrate LRU and non-lru movable pages.
904                  * Skip any other type of page
905                  */
906                 if (!PageLRU(page)) {
907                         /*
908                          * __PageMovable can return false positive so we need
909                          * to verify it under page_lock.
910                          */
911                         if (unlikely(__PageMovable(page)) &&
912                                         !PageIsolated(page)) {
913                                 if (locked) {
914                                         spin_unlock_irqrestore(&pgdat->lru_lock,
915                                                                         flags);
916                                         locked = false;
917                                 }
918
919                                 if (!isolate_movable_page(page, isolate_mode))
920                                         goto isolate_success;
921                         }
922
923                         goto isolate_fail;
924                 }
925
926                 /*
927                  * Migration will fail if an anonymous page is pinned in memory,
928                  * so avoid taking lru_lock and isolating it unnecessarily in an
929                  * admittedly racy check.
930                  */
931                 if (!page_mapping(page) &&
932                     page_count(page) > page_mapcount(page))
933                         goto isolate_fail;
934
935                 /*
936                  * Only allow to migrate anonymous pages in GFP_NOFS context
937                  * because those do not depend on fs locks.
938                  */
939                 if (!(cc->gfp_mask & __GFP_FS) && page_mapping(page))
940                         goto isolate_fail;
941
942                 /* If we already hold the lock, we can skip some rechecking */
943                 if (!locked) {
944                         locked = compact_lock_irqsave(&pgdat->lru_lock,
945                                                                 &flags, cc);
946
947                         /* Try get exclusive access under lock */
948                         if (!skip_updated) {
949                                 skip_updated = true;
950                                 if (test_and_set_skip(cc, page, low_pfn))
951                                         goto isolate_abort;
952                         }
953
954                         /* Recheck PageLRU and PageCompound under lock */
955                         if (!PageLRU(page))
956                                 goto isolate_fail;
957
958                         /*
959                          * Page become compound since the non-locked check,
960                          * and it's on LRU. It can only be a THP so the order
961                          * is safe to read and it's 0 for tail pages.
962                          */
963                         if (unlikely(PageCompound(page))) {
964                                 low_pfn += (1UL << compound_order(page)) - 1;
965                                 goto isolate_fail;
966                         }
967                 }
968
969                 lruvec = mem_cgroup_page_lruvec(page, pgdat);
970
971                 /* Try isolate the page */
972                 if (__isolate_lru_page(page, isolate_mode) != 0)
973                         goto isolate_fail;
974
975                 VM_BUG_ON_PAGE(PageCompound(page), page);
976
977                 /* Successfully isolated */
978                 del_page_from_lru_list(page, lruvec, page_lru(page));
979                 inc_node_page_state(page,
980                                 NR_ISOLATED_ANON + page_is_file_cache(page));
981
982 isolate_success:
983                 list_add(&page->lru, &cc->migratepages);
984                 cc->nr_migratepages++;
985                 nr_isolated++;
986
987                 /*
988                  * Avoid isolating too much unless this block is being
989                  * rescanned (e.g. dirty/writeback pages, parallel allocation)
990                  * or a lock is contended. For contention, isolate quickly to
991                  * potentially remove one source of contention.
992                  */
993                 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX &&
994                     !cc->rescan && !cc->contended) {
995                         ++low_pfn;
996                         break;
997                 }
998
999                 continue;
1000 isolate_fail:
1001                 if (!skip_on_failure)
1002                         continue;
1003
1004                 /*
1005                  * We have isolated some pages, but then failed. Release them
1006                  * instead of migrating, as we cannot form the cc->order buddy
1007                  * page anyway.
1008                  */
1009                 if (nr_isolated) {
1010                         if (locked) {
1011                                 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
1012                                 locked = false;
1013                         }
1014                         putback_movable_pages(&cc->migratepages);
1015                         cc->nr_migratepages = 0;
1016                         nr_isolated = 0;
1017                 }
1018
1019                 if (low_pfn < next_skip_pfn) {
1020                         low_pfn = next_skip_pfn - 1;
1021                         /*
1022                          * The check near the loop beginning would have updated
1023                          * next_skip_pfn too, but this is a bit simpler.
1024                          */
1025                         next_skip_pfn += 1UL << cc->order;
1026                 }
1027         }
1028
1029         /*
1030          * The PageBuddy() check could have potentially brought us outside
1031          * the range to be scanned.
1032          */
1033         if (unlikely(low_pfn > end_pfn))
1034                 low_pfn = end_pfn;
1035
1036 isolate_abort:
1037         if (locked)
1038                 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
1039
1040         /*
1041          * Updated the cached scanner pfn once the pageblock has been scanned
1042          * Pages will either be migrated in which case there is no point
1043          * scanning in the near future or migration failed in which case the
1044          * failure reason may persist. The block is marked for skipping if
1045          * there were no pages isolated in the block or if the block is
1046          * rescanned twice in a row.
1047          */
1048         if (low_pfn == end_pfn && (!nr_isolated || cc->rescan)) {
1049                 if (valid_page && !skip_updated)
1050                         set_pageblock_skip(valid_page);
1051                 update_cached_migrate(cc, low_pfn);
1052         }
1053
1054         trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
1055                                                 nr_scanned, nr_isolated);
1056
1057         cc->total_migrate_scanned += nr_scanned;
1058         if (nr_isolated)
1059                 count_compact_events(COMPACTISOLATED, nr_isolated);
1060
1061         return low_pfn;
1062 }
1063
1064 /**
1065  * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
1066  * @cc:        Compaction control structure.
1067  * @start_pfn: The first PFN to start isolating.
1068  * @end_pfn:   The one-past-last PFN.
1069  *
1070  * Returns zero if isolation fails fatally due to e.g. pending signal.
1071  * Otherwise, function returns one-past-the-last PFN of isolated page
1072  * (which may be greater than end_pfn if end fell in a middle of a THP page).
1073  */
1074 unsigned long
1075 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
1076                                                         unsigned long end_pfn)
1077 {
1078         unsigned long pfn, block_start_pfn, block_end_pfn;
1079
1080         /* Scan block by block. First and last block may be incomplete */
1081         pfn = start_pfn;
1082         block_start_pfn = pageblock_start_pfn(pfn);
1083         if (block_start_pfn < cc->zone->zone_start_pfn)
1084                 block_start_pfn = cc->zone->zone_start_pfn;
1085         block_end_pfn = pageblock_end_pfn(pfn);
1086
1087         for (; pfn < end_pfn; pfn = block_end_pfn,
1088                                 block_start_pfn = block_end_pfn,
1089                                 block_end_pfn += pageblock_nr_pages) {
1090
1091                 block_end_pfn = min(block_end_pfn, end_pfn);
1092
1093                 if (!pageblock_pfn_to_page(block_start_pfn,
1094                                         block_end_pfn, cc->zone))
1095                         continue;
1096
1097                 pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
1098                                                         ISOLATE_UNEVICTABLE);
1099
1100                 if (!pfn)
1101                         break;
1102
1103                 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
1104                         break;
1105         }
1106
1107         return pfn;
1108 }
1109
1110 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
1111 #ifdef CONFIG_COMPACTION
1112
1113 static bool suitable_migration_source(struct compact_control *cc,
1114                                                         struct page *page)
1115 {
1116         int block_mt;
1117
1118         if (pageblock_skip_persistent(page))
1119                 return false;
1120
1121         if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction)
1122                 return true;
1123
1124         block_mt = get_pageblock_migratetype(page);
1125
1126         if (cc->migratetype == MIGRATE_MOVABLE)
1127                 return is_migrate_movable(block_mt);
1128         else
1129                 return block_mt == cc->migratetype;
1130 }
1131
1132 /* Returns true if the page is within a block suitable for migration to */
1133 static bool suitable_migration_target(struct compact_control *cc,
1134                                                         struct page *page)
1135 {
1136         /* If the page is a large free page, then disallow migration */
1137         if (PageBuddy(page)) {
1138                 /*
1139                  * We are checking page_order without zone->lock taken. But
1140                  * the only small danger is that we skip a potentially suitable
1141                  * pageblock, so it's not worth to check order for valid range.
1142                  */
1143                 if (page_order_unsafe(page) >= pageblock_order)
1144                         return false;
1145         }
1146
1147         if (cc->ignore_block_suitable)
1148                 return true;
1149
1150         /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1151         if (is_migrate_movable(get_pageblock_migratetype(page)))
1152                 return true;
1153
1154         /* Otherwise skip the block */
1155         return false;
1156 }
1157
1158 static inline unsigned int
1159 freelist_scan_limit(struct compact_control *cc)
1160 {
1161         return (COMPACT_CLUSTER_MAX >> cc->fast_search_fail) + 1;
1162 }
1163
1164 /*
1165  * Test whether the free scanner has reached the same or lower pageblock than
1166  * the migration scanner, and compaction should thus terminate.
1167  */
1168 static inline bool compact_scanners_met(struct compact_control *cc)
1169 {
1170         return (cc->free_pfn >> pageblock_order)
1171                 <= (cc->migrate_pfn >> pageblock_order);
1172 }
1173
1174 /*
1175  * Used when scanning for a suitable migration target which scans freelists
1176  * in reverse. Reorders the list such as the unscanned pages are scanned
1177  * first on the next iteration of the free scanner
1178  */
1179 static void
1180 move_freelist_head(struct list_head *freelist, struct page *freepage)
1181 {
1182         LIST_HEAD(sublist);
1183
1184         if (!list_is_last(freelist, &freepage->lru)) {
1185                 list_cut_before(&sublist, freelist, &freepage->lru);
1186                 if (!list_empty(&sublist))
1187                         list_splice_tail(&sublist, freelist);
1188         }
1189 }
1190
1191 /*
1192  * Similar to move_freelist_head except used by the migration scanner
1193  * when scanning forward. It's possible for these list operations to
1194  * move against each other if they search the free list exactly in
1195  * lockstep.
1196  */
1197 static void
1198 move_freelist_tail(struct list_head *freelist, struct page *freepage)
1199 {
1200         LIST_HEAD(sublist);
1201
1202         if (!list_is_first(freelist, &freepage->lru)) {
1203                 list_cut_position(&sublist, freelist, &freepage->lru);
1204                 if (!list_empty(&sublist))
1205                         list_splice_tail(&sublist, freelist);
1206         }
1207 }
1208
1209 static void
1210 fast_isolate_around(struct compact_control *cc, unsigned long pfn, unsigned long nr_isolated)
1211 {
1212         unsigned long start_pfn, end_pfn;
1213         struct page *page = pfn_to_page(pfn);
1214
1215         /* Do not search around if there are enough pages already */
1216         if (cc->nr_freepages >= cc->nr_migratepages)
1217                 return;
1218
1219         /* Minimise scanning during async compaction */
1220         if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC)
1221                 return;
1222
1223         /* Pageblock boundaries */
1224         start_pfn = pageblock_start_pfn(pfn);
1225         end_pfn = min(start_pfn + pageblock_nr_pages, zone_end_pfn(cc->zone));
1226
1227         /* Scan before */
1228         if (start_pfn != pfn) {
1229                 isolate_freepages_block(cc, &start_pfn, pfn, &cc->freepages, 1, false);
1230                 if (cc->nr_freepages >= cc->nr_migratepages)
1231                         return;
1232         }
1233
1234         /* Scan after */
1235         start_pfn = pfn + nr_isolated;
1236         if (start_pfn != end_pfn)
1237                 isolate_freepages_block(cc, &start_pfn, end_pfn, &cc->freepages, 1, false);
1238
1239         /* Skip this pageblock in the future as it's full or nearly full */
1240         if (cc->nr_freepages < cc->nr_migratepages)
1241                 set_pageblock_skip(page);
1242 }
1243
1244 /* Search orders in round-robin fashion */
1245 static int next_search_order(struct compact_control *cc, int order)
1246 {
1247         order--;
1248         if (order < 0)
1249                 order = cc->order - 1;
1250
1251         /* Search wrapped around? */
1252         if (order == cc->search_order) {
1253                 cc->search_order--;
1254                 if (cc->search_order < 0)
1255                         cc->search_order = cc->order - 1;
1256                 return -1;
1257         }
1258
1259         return order;
1260 }
1261
1262 static unsigned long
1263 fast_isolate_freepages(struct compact_control *cc)
1264 {
1265         unsigned int limit = min(1U, freelist_scan_limit(cc) >> 1);
1266         unsigned int nr_scanned = 0;
1267         unsigned long low_pfn, min_pfn, high_pfn = 0, highest = 0;
1268         unsigned long nr_isolated = 0;
1269         unsigned long distance;
1270         struct page *page = NULL;
1271         bool scan_start = false;
1272         int order;
1273
1274         /* Full compaction passes in a negative order */
1275         if (cc->order <= 0)
1276                 return cc->free_pfn;
1277
1278         /*
1279          * If starting the scan, use a deeper search and use the highest
1280          * PFN found if a suitable one is not found.
1281          */
1282         if (cc->free_pfn >= cc->zone->compact_init_free_pfn) {
1283                 limit = pageblock_nr_pages >> 1;
1284                 scan_start = true;
1285         }
1286
1287         /*
1288          * Preferred point is in the top quarter of the scan space but take
1289          * a pfn from the top half if the search is problematic.
1290          */
1291         distance = (cc->free_pfn - cc->migrate_pfn);
1292         low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2));
1293         min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1));
1294
1295         if (WARN_ON_ONCE(min_pfn > low_pfn))
1296                 low_pfn = min_pfn;
1297
1298         /*
1299          * Search starts from the last successful isolation order or the next
1300          * order to search after a previous failure
1301          */
1302         cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order);
1303
1304         for (order = cc->search_order;
1305              !page && order >= 0;
1306              order = next_search_order(cc, order)) {
1307                 struct free_area *area = &cc->zone->free_area[order];
1308                 struct list_head *freelist;
1309                 struct page *freepage;
1310                 unsigned long flags;
1311                 unsigned int order_scanned = 0;
1312
1313                 if (!area->nr_free)
1314                         continue;
1315
1316                 spin_lock_irqsave(&cc->zone->lock, flags);
1317                 freelist = &area->free_list[MIGRATE_MOVABLE];
1318                 list_for_each_entry_reverse(freepage, freelist, lru) {
1319                         unsigned long pfn;
1320
1321                         order_scanned++;
1322                         nr_scanned++;
1323                         pfn = page_to_pfn(freepage);
1324
1325                         if (pfn >= highest)
1326                                 highest = pageblock_start_pfn(pfn);
1327
1328                         if (pfn >= low_pfn) {
1329                                 cc->fast_search_fail = 0;
1330                                 cc->search_order = order;
1331                                 page = freepage;
1332                                 break;
1333                         }
1334
1335                         if (pfn >= min_pfn && pfn > high_pfn) {
1336                                 high_pfn = pfn;
1337
1338                                 /* Shorten the scan if a candidate is found */
1339                                 limit >>= 1;
1340                         }
1341
1342                         if (order_scanned >= limit)
1343                                 break;
1344                 }
1345
1346                 /* Use a minimum pfn if a preferred one was not found */
1347                 if (!page && high_pfn) {
1348                         page = pfn_to_page(high_pfn);
1349
1350                         /* Update freepage for the list reorder below */
1351                         freepage = page;
1352                 }
1353
1354                 /* Reorder to so a future search skips recent pages */
1355                 move_freelist_head(freelist, freepage);
1356
1357                 /* Isolate the page if available */
1358                 if (page) {
1359                         if (__isolate_free_page(page, order)) {
1360                                 set_page_private(page, order);
1361                                 nr_isolated = 1 << order;
1362                                 cc->nr_freepages += nr_isolated;
1363                                 list_add_tail(&page->lru, &cc->freepages);
1364                                 count_compact_events(COMPACTISOLATED, nr_isolated);
1365                         } else {
1366                                 /* If isolation fails, abort the search */
1367                                 order = -1;
1368                                 page = NULL;
1369                         }
1370                 }
1371
1372                 spin_unlock_irqrestore(&cc->zone->lock, flags);
1373
1374                 /*
1375                  * Smaller scan on next order so the total scan ig related
1376                  * to freelist_scan_limit.
1377                  */
1378                 if (order_scanned >= limit)
1379                         limit = min(1U, limit >> 1);
1380         }
1381
1382         if (!page) {
1383                 cc->fast_search_fail++;
1384                 if (scan_start) {
1385                         /*
1386                          * Use the highest PFN found above min. If one was
1387                          * not found, be pessemistic for direct compaction
1388                          * and use the min mark.
1389                          */
1390                         if (highest) {
1391                                 page = pfn_to_page(highest);
1392                                 cc->free_pfn = highest;
1393                         } else {
1394                                 if (cc->direct_compaction) {
1395                                         page = pfn_to_page(min_pfn);
1396                                         cc->free_pfn = min_pfn;
1397                                 }
1398                         }
1399                 }
1400         }
1401
1402         if (highest && highest >= cc->zone->compact_cached_free_pfn) {
1403                 highest -= pageblock_nr_pages;
1404                 cc->zone->compact_cached_free_pfn = highest;
1405         }
1406
1407         cc->total_free_scanned += nr_scanned;
1408         if (!page)
1409                 return cc->free_pfn;
1410
1411         low_pfn = page_to_pfn(page);
1412         fast_isolate_around(cc, low_pfn, nr_isolated);
1413         return low_pfn;
1414 }
1415
1416 /*
1417  * Based on information in the current compact_control, find blocks
1418  * suitable for isolating free pages from and then isolate them.
1419  */
1420 static void isolate_freepages(struct compact_control *cc)
1421 {
1422         struct zone *zone = cc->zone;
1423         struct page *page;
1424         unsigned long block_start_pfn;  /* start of current pageblock */
1425         unsigned long isolate_start_pfn; /* exact pfn we start at */
1426         unsigned long block_end_pfn;    /* end of current pageblock */
1427         unsigned long low_pfn;       /* lowest pfn scanner is able to scan */
1428         struct list_head *freelist = &cc->freepages;
1429         unsigned int stride;
1430
1431         /* Try a small search of the free lists for a candidate */
1432         isolate_start_pfn = fast_isolate_freepages(cc);
1433         if (cc->nr_freepages)
1434                 goto splitmap;
1435
1436         /*
1437          * Initialise the free scanner. The starting point is where we last
1438          * successfully isolated from, zone-cached value, or the end of the
1439          * zone when isolating for the first time. For looping we also need
1440          * this pfn aligned down to the pageblock boundary, because we do
1441          * block_start_pfn -= pageblock_nr_pages in the for loop.
1442          * For ending point, take care when isolating in last pageblock of a
1443          * a zone which ends in the middle of a pageblock.
1444          * The low boundary is the end of the pageblock the migration scanner
1445          * is using.
1446          */
1447         isolate_start_pfn = cc->free_pfn;
1448         block_start_pfn = pageblock_start_pfn(isolate_start_pfn);
1449         block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1450                                                 zone_end_pfn(zone));
1451         low_pfn = pageblock_end_pfn(cc->migrate_pfn);
1452         stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1;
1453
1454         /*
1455          * Isolate free pages until enough are available to migrate the
1456          * pages on cc->migratepages. We stop searching if the migrate
1457          * and free page scanners meet or enough free pages are isolated.
1458          */
1459         for (; block_start_pfn >= low_pfn;
1460                                 block_end_pfn = block_start_pfn,
1461                                 block_start_pfn -= pageblock_nr_pages,
1462                                 isolate_start_pfn = block_start_pfn) {
1463                 unsigned long nr_isolated;
1464
1465                 /*
1466                  * This can iterate a massively long zone without finding any
1467                  * suitable migration targets, so periodically check resched.
1468                  */
1469                 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)))
1470                         cond_resched();
1471
1472                 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1473                                                                         zone);
1474                 if (!page)
1475                         continue;
1476
1477                 /* Check the block is suitable for migration */
1478                 if (!suitable_migration_target(cc, page))
1479                         continue;
1480
1481                 /* If isolation recently failed, do not retry */
1482                 if (!isolation_suitable(cc, page))
1483                         continue;
1484
1485                 /* Found a block suitable for isolating free pages from. */
1486                 nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn,
1487                                         block_end_pfn, freelist, stride, false);
1488
1489                 /* Update the skip hint if the full pageblock was scanned */
1490                 if (isolate_start_pfn == block_end_pfn)
1491                         update_pageblock_skip(cc, page, block_start_pfn);
1492
1493                 /* Are enough freepages isolated? */
1494                 if (cc->nr_freepages >= cc->nr_migratepages) {
1495                         if (isolate_start_pfn >= block_end_pfn) {
1496                                 /*
1497                                  * Restart at previous pageblock if more
1498                                  * freepages can be isolated next time.
1499                                  */
1500                                 isolate_start_pfn =
1501                                         block_start_pfn - pageblock_nr_pages;
1502                         }
1503                         break;
1504                 } else if (isolate_start_pfn < block_end_pfn) {
1505                         /*
1506                          * If isolation failed early, do not continue
1507                          * needlessly.
1508                          */
1509                         break;
1510                 }
1511
1512                 /* Adjust stride depending on isolation */
1513                 if (nr_isolated) {
1514                         stride = 1;
1515                         continue;
1516                 }
1517                 stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1);
1518         }
1519
1520         /*
1521          * Record where the free scanner will restart next time. Either we
1522          * broke from the loop and set isolate_start_pfn based on the last
1523          * call to isolate_freepages_block(), or we met the migration scanner
1524          * and the loop terminated due to isolate_start_pfn < low_pfn
1525          */
1526         cc->free_pfn = isolate_start_pfn;
1527
1528 splitmap:
1529         /* __isolate_free_page() does not map the pages */
1530         split_map_pages(freelist);
1531 }
1532
1533 /*
1534  * This is a migrate-callback that "allocates" freepages by taking pages
1535  * from the isolated freelists in the block we are migrating to.
1536  */
1537 static struct page *compaction_alloc(struct page *migratepage,
1538                                         unsigned long data)
1539 {
1540         struct compact_control *cc = (struct compact_control *)data;
1541         struct page *freepage;
1542
1543         if (list_empty(&cc->freepages)) {
1544                 isolate_freepages(cc);
1545
1546                 if (list_empty(&cc->freepages))
1547                         return NULL;
1548         }
1549
1550         freepage = list_entry(cc->freepages.next, struct page, lru);
1551         list_del(&freepage->lru);
1552         cc->nr_freepages--;
1553
1554         return freepage;
1555 }
1556
1557 /*
1558  * This is a migrate-callback that "frees" freepages back to the isolated
1559  * freelist.  All pages on the freelist are from the same zone, so there is no
1560  * special handling needed for NUMA.
1561  */
1562 static void compaction_free(struct page *page, unsigned long data)
1563 {
1564         struct compact_control *cc = (struct compact_control *)data;
1565
1566         list_add(&page->lru, &cc->freepages);
1567         cc->nr_freepages++;
1568 }
1569
1570 /* possible outcome of isolate_migratepages */
1571 typedef enum {
1572         ISOLATE_ABORT,          /* Abort compaction now */
1573         ISOLATE_NONE,           /* No pages isolated, continue scanning */
1574         ISOLATE_SUCCESS,        /* Pages isolated, migrate */
1575 } isolate_migrate_t;
1576
1577 /*
1578  * Allow userspace to control policy on scanning the unevictable LRU for
1579  * compactable pages.
1580  */
1581 int sysctl_compact_unevictable_allowed __read_mostly = 1;
1582
1583 static inline void
1584 update_fast_start_pfn(struct compact_control *cc, unsigned long pfn)
1585 {
1586         if (cc->fast_start_pfn == ULONG_MAX)
1587                 return;
1588
1589         if (!cc->fast_start_pfn)
1590                 cc->fast_start_pfn = pfn;
1591
1592         cc->fast_start_pfn = min(cc->fast_start_pfn, pfn);
1593 }
1594
1595 static inline unsigned long
1596 reinit_migrate_pfn(struct compact_control *cc)
1597 {
1598         if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX)
1599                 return cc->migrate_pfn;
1600
1601         cc->migrate_pfn = cc->fast_start_pfn;
1602         cc->fast_start_pfn = ULONG_MAX;
1603
1604         return cc->migrate_pfn;
1605 }
1606
1607 /*
1608  * Briefly search the free lists for a migration source that already has
1609  * some free pages to reduce the number of pages that need migration
1610  * before a pageblock is free.
1611  */
1612 static unsigned long fast_find_migrateblock(struct compact_control *cc)
1613 {
1614         unsigned int limit = freelist_scan_limit(cc);
1615         unsigned int nr_scanned = 0;
1616         unsigned long distance;
1617         unsigned long pfn = cc->migrate_pfn;
1618         unsigned long high_pfn;
1619         int order;
1620
1621         /* Skip hints are relied on to avoid repeats on the fast search */
1622         if (cc->ignore_skip_hint)
1623                 return pfn;
1624
1625         /*
1626          * If the migrate_pfn is not at the start of a zone or the start
1627          * of a pageblock then assume this is a continuation of a previous
1628          * scan restarted due to COMPACT_CLUSTER_MAX.
1629          */
1630         if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn))
1631                 return pfn;
1632
1633         /*
1634          * For smaller orders, just linearly scan as the number of pages
1635          * to migrate should be relatively small and does not necessarily
1636          * justify freeing up a large block for a small allocation.
1637          */
1638         if (cc->order <= PAGE_ALLOC_COSTLY_ORDER)
1639                 return pfn;
1640
1641         /*
1642          * Only allow kcompactd and direct requests for movable pages to
1643          * quickly clear out a MOVABLE pageblock for allocation. This
1644          * reduces the risk that a large movable pageblock is freed for
1645          * an unmovable/reclaimable small allocation.
1646          */
1647         if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE)
1648                 return pfn;
1649
1650         /*
1651          * When starting the migration scanner, pick any pageblock within the
1652          * first half of the search space. Otherwise try and pick a pageblock
1653          * within the first eighth to reduce the chances that a migration
1654          * target later becomes a source.
1655          */
1656         distance = (cc->free_pfn - cc->migrate_pfn) >> 1;
1657         if (cc->migrate_pfn != cc->zone->zone_start_pfn)
1658                 distance >>= 2;
1659         high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance);
1660
1661         for (order = cc->order - 1;
1662              order >= PAGE_ALLOC_COSTLY_ORDER && pfn == cc->migrate_pfn && nr_scanned < limit;
1663              order--) {
1664                 struct free_area *area = &cc->zone->free_area[order];
1665                 struct list_head *freelist;
1666                 unsigned long flags;
1667                 struct page *freepage;
1668
1669                 if (!area->nr_free)
1670                         continue;
1671
1672                 spin_lock_irqsave(&cc->zone->lock, flags);
1673                 freelist = &area->free_list[MIGRATE_MOVABLE];
1674                 list_for_each_entry(freepage, freelist, lru) {
1675                         unsigned long free_pfn;
1676
1677                         nr_scanned++;
1678                         free_pfn = page_to_pfn(freepage);
1679                         if (free_pfn < high_pfn) {
1680                                 /*
1681                                  * Avoid if skipped recently. Ideally it would
1682                                  * move to the tail but even safe iteration of
1683                                  * the list assumes an entry is deleted, not
1684                                  * reordered.
1685                                  */
1686                                 if (get_pageblock_skip(freepage)) {
1687                                         if (list_is_last(freelist, &freepage->lru))
1688                                                 break;
1689
1690                                         continue;
1691                                 }
1692
1693                                 /* Reorder to so a future search skips recent pages */
1694                                 move_freelist_tail(freelist, freepage);
1695
1696                                 update_fast_start_pfn(cc, free_pfn);
1697                                 pfn = pageblock_start_pfn(free_pfn);
1698                                 cc->fast_search_fail = 0;
1699                                 set_pageblock_skip(freepage);
1700                                 break;
1701                         }
1702
1703                         if (nr_scanned >= limit) {
1704                                 cc->fast_search_fail++;
1705                                 move_freelist_tail(freelist, freepage);
1706                                 break;
1707                         }
1708                 }
1709                 spin_unlock_irqrestore(&cc->zone->lock, flags);
1710         }
1711
1712         cc->total_migrate_scanned += nr_scanned;
1713
1714         /*
1715          * If fast scanning failed then use a cached entry for a page block
1716          * that had free pages as the basis for starting a linear scan.
1717          */
1718         if (pfn == cc->migrate_pfn)
1719                 pfn = reinit_migrate_pfn(cc);
1720
1721         return pfn;
1722 }
1723
1724 /*
1725  * Isolate all pages that can be migrated from the first suitable block,
1726  * starting at the block pointed to by the migrate scanner pfn within
1727  * compact_control.
1728  */
1729 static isolate_migrate_t isolate_migratepages(struct zone *zone,
1730                                         struct compact_control *cc)
1731 {
1732         unsigned long block_start_pfn;
1733         unsigned long block_end_pfn;
1734         unsigned long low_pfn;
1735         struct page *page;
1736         const isolate_mode_t isolate_mode =
1737                 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1738                 (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1739         bool fast_find_block;
1740
1741         /*
1742          * Start at where we last stopped, or beginning of the zone as
1743          * initialized by compact_zone(). The first failure will use
1744          * the lowest PFN as the starting point for linear scanning.
1745          */
1746         low_pfn = fast_find_migrateblock(cc);
1747         block_start_pfn = pageblock_start_pfn(low_pfn);
1748         if (block_start_pfn < zone->zone_start_pfn)
1749                 block_start_pfn = zone->zone_start_pfn;
1750
1751         /*
1752          * fast_find_migrateblock marks a pageblock skipped so to avoid
1753          * the isolation_suitable check below, check whether the fast
1754          * search was successful.
1755          */
1756         fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail;
1757
1758         /* Only scan within a pageblock boundary */
1759         block_end_pfn = pageblock_end_pfn(low_pfn);
1760
1761         /*
1762          * Iterate over whole pageblocks until we find the first suitable.
1763          * Do not cross the free scanner.
1764          */
1765         for (; block_end_pfn <= cc->free_pfn;
1766                         fast_find_block = false,
1767                         low_pfn = block_end_pfn,
1768                         block_start_pfn = block_end_pfn,
1769                         block_end_pfn += pageblock_nr_pages) {
1770
1771                 /*
1772                  * This can potentially iterate a massively long zone with
1773                  * many pageblocks unsuitable, so periodically check if we
1774                  * need to schedule.
1775                  */
1776                 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)))
1777                         cond_resched();
1778
1779                 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1780                                                                         zone);
1781                 if (!page)
1782                         continue;
1783
1784                 /*
1785                  * If isolation recently failed, do not retry. Only check the
1786                  * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock
1787                  * to be visited multiple times. Assume skip was checked
1788                  * before making it "skip" so other compaction instances do
1789                  * not scan the same block.
1790                  */
1791                 if (IS_ALIGNED(low_pfn, pageblock_nr_pages) &&
1792                     !fast_find_block && !isolation_suitable(cc, page))
1793                         continue;
1794
1795                 /*
1796                  * For async compaction, also only scan in MOVABLE blocks
1797                  * without huge pages. Async compaction is optimistic to see
1798                  * if the minimum amount of work satisfies the allocation.
1799                  * The cached PFN is updated as it's possible that all
1800                  * remaining blocks between source and target are unsuitable
1801                  * and the compaction scanners fail to meet.
1802                  */
1803                 if (!suitable_migration_source(cc, page)) {
1804                         update_cached_migrate(cc, block_end_pfn);
1805                         continue;
1806                 }
1807
1808                 /* Perform the isolation */
1809                 low_pfn = isolate_migratepages_block(cc, low_pfn,
1810                                                 block_end_pfn, isolate_mode);
1811
1812                 if (!low_pfn)
1813                         return ISOLATE_ABORT;
1814
1815                 /*
1816                  * Either we isolated something and proceed with migration. Or
1817                  * we failed and compact_zone should decide if we should
1818                  * continue or not.
1819                  */
1820                 break;
1821         }
1822
1823         /* Record where migration scanner will be restarted. */
1824         cc->migrate_pfn = low_pfn;
1825
1826         return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
1827 }
1828
1829 /*
1830  * order == -1 is expected when compacting via
1831  * /proc/sys/vm/compact_memory
1832  */
1833 static inline bool is_via_compact_memory(int order)
1834 {
1835         return order == -1;
1836 }
1837
1838 static enum compact_result __compact_finished(struct compact_control *cc)
1839 {
1840         unsigned int order;
1841         const int migratetype = cc->migratetype;
1842         int ret;
1843
1844         /* Compaction run completes if the migrate and free scanner meet */
1845         if (compact_scanners_met(cc)) {
1846                 /* Let the next compaction start anew. */
1847                 reset_cached_positions(cc->zone);
1848
1849                 /*
1850                  * Mark that the PG_migrate_skip information should be cleared
1851                  * by kswapd when it goes to sleep. kcompactd does not set the
1852                  * flag itself as the decision to be clear should be directly
1853                  * based on an allocation request.
1854                  */
1855                 if (cc->direct_compaction)
1856                         cc->zone->compact_blockskip_flush = true;
1857
1858                 if (cc->whole_zone)
1859                         return COMPACT_COMPLETE;
1860                 else
1861                         return COMPACT_PARTIAL_SKIPPED;
1862         }
1863
1864         if (is_via_compact_memory(cc->order))
1865                 return COMPACT_CONTINUE;
1866
1867         /*
1868          * Always finish scanning a pageblock to reduce the possibility of
1869          * fallbacks in the future. This is particularly important when
1870          * migration source is unmovable/reclaimable but it's not worth
1871          * special casing.
1872          */
1873         if (!IS_ALIGNED(cc->migrate_pfn, pageblock_nr_pages))
1874                 return COMPACT_CONTINUE;
1875
1876         /* Direct compactor: Is a suitable page free? */
1877         ret = COMPACT_NO_SUITABLE_PAGE;
1878         for (order = cc->order; order < MAX_ORDER; order++) {
1879                 struct free_area *area = &cc->zone->free_area[order];
1880                 bool can_steal;
1881
1882                 /* Job done if page is free of the right migratetype */
1883                 if (!list_empty(&area->free_list[migratetype]))
1884                         return COMPACT_SUCCESS;
1885
1886 #ifdef CONFIG_CMA
1887                 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
1888                 if (migratetype == MIGRATE_MOVABLE &&
1889                         !list_empty(&area->free_list[MIGRATE_CMA]))
1890                         return COMPACT_SUCCESS;
1891 #endif
1892                 /*
1893                  * Job done if allocation would steal freepages from
1894                  * other migratetype buddy lists.
1895                  */
1896                 if (find_suitable_fallback(area, order, migratetype,
1897                                                 true, &can_steal) != -1) {
1898
1899                         /* movable pages are OK in any pageblock */
1900                         if (migratetype == MIGRATE_MOVABLE)
1901                                 return COMPACT_SUCCESS;
1902
1903                         /*
1904                          * We are stealing for a non-movable allocation. Make
1905                          * sure we finish compacting the current pageblock
1906                          * first so it is as free as possible and we won't
1907                          * have to steal another one soon. This only applies
1908                          * to sync compaction, as async compaction operates
1909                          * on pageblocks of the same migratetype.
1910                          */
1911                         if (cc->mode == MIGRATE_ASYNC ||
1912                                         IS_ALIGNED(cc->migrate_pfn,
1913                                                         pageblock_nr_pages)) {
1914                                 return COMPACT_SUCCESS;
1915                         }
1916
1917                         ret = COMPACT_CONTINUE;
1918                         break;
1919                 }
1920         }
1921
1922         if (cc->contended || fatal_signal_pending(current))
1923                 ret = COMPACT_CONTENDED;
1924
1925         return ret;
1926 }
1927
1928 static enum compact_result compact_finished(struct compact_control *cc)
1929 {
1930         int ret;
1931
1932         ret = __compact_finished(cc);
1933         trace_mm_compaction_finished(cc->zone, cc->order, ret);
1934         if (ret == COMPACT_NO_SUITABLE_PAGE)
1935                 ret = COMPACT_CONTINUE;
1936
1937         return ret;
1938 }
1939
1940 /*
1941  * compaction_suitable: Is this suitable to run compaction on this zone now?
1942  * Returns
1943  *   COMPACT_SKIPPED  - If there are too few free pages for compaction
1944  *   COMPACT_SUCCESS  - If the allocation would succeed without compaction
1945  *   COMPACT_CONTINUE - If compaction should run now
1946  */
1947 static enum compact_result __compaction_suitable(struct zone *zone, int order,
1948                                         unsigned int alloc_flags,
1949                                         int classzone_idx,
1950                                         unsigned long wmark_target)
1951 {
1952         unsigned long watermark;
1953
1954         if (is_via_compact_memory(order))
1955                 return COMPACT_CONTINUE;
1956
1957         watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
1958         /*
1959          * If watermarks for high-order allocation are already met, there
1960          * should be no need for compaction at all.
1961          */
1962         if (zone_watermark_ok(zone, order, watermark, classzone_idx,
1963                                                                 alloc_flags))
1964                 return COMPACT_SUCCESS;
1965
1966         /*
1967          * Watermarks for order-0 must be met for compaction to be able to
1968          * isolate free pages for migration targets. This means that the
1969          * watermark and alloc_flags have to match, or be more pessimistic than
1970          * the check in __isolate_free_page(). We don't use the direct
1971          * compactor's alloc_flags, as they are not relevant for freepage
1972          * isolation. We however do use the direct compactor's classzone_idx to
1973          * skip over zones where lowmem reserves would prevent allocation even
1974          * if compaction succeeds.
1975          * For costly orders, we require low watermark instead of min for
1976          * compaction to proceed to increase its chances.
1977          * ALLOC_CMA is used, as pages in CMA pageblocks are considered
1978          * suitable migration targets
1979          */
1980         watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
1981                                 low_wmark_pages(zone) : min_wmark_pages(zone);
1982         watermark += compact_gap(order);
1983         if (!__zone_watermark_ok(zone, 0, watermark, classzone_idx,
1984                                                 ALLOC_CMA, wmark_target))
1985                 return COMPACT_SKIPPED;
1986
1987         return COMPACT_CONTINUE;
1988 }
1989
1990 enum compact_result compaction_suitable(struct zone *zone, int order,
1991                                         unsigned int alloc_flags,
1992                                         int classzone_idx)
1993 {
1994         enum compact_result ret;
1995         int fragindex;
1996
1997         ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx,
1998                                     zone_page_state(zone, NR_FREE_PAGES));
1999         /*
2000          * fragmentation index determines if allocation failures are due to
2001          * low memory or external fragmentation
2002          *
2003          * index of -1000 would imply allocations might succeed depending on
2004          * watermarks, but we already failed the high-order watermark check
2005          * index towards 0 implies failure is due to lack of memory
2006          * index towards 1000 implies failure is due to fragmentation
2007          *
2008          * Only compact if a failure would be due to fragmentation. Also
2009          * ignore fragindex for non-costly orders where the alternative to
2010          * a successful reclaim/compaction is OOM. Fragindex and the
2011          * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
2012          * excessive compaction for costly orders, but it should not be at the
2013          * expense of system stability.
2014          */
2015         if (ret == COMPACT_CONTINUE && (order > PAGE_ALLOC_COSTLY_ORDER)) {
2016                 fragindex = fragmentation_index(zone, order);
2017                 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
2018                         ret = COMPACT_NOT_SUITABLE_ZONE;
2019         }
2020
2021         trace_mm_compaction_suitable(zone, order, ret);
2022         if (ret == COMPACT_NOT_SUITABLE_ZONE)
2023                 ret = COMPACT_SKIPPED;
2024
2025         return ret;
2026 }
2027
2028 bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
2029                 int alloc_flags)
2030 {
2031         struct zone *zone;
2032         struct zoneref *z;
2033
2034         /*
2035          * Make sure at least one zone would pass __compaction_suitable if we continue
2036          * retrying the reclaim.
2037          */
2038         for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
2039                                         ac->nodemask) {
2040                 unsigned long available;
2041                 enum compact_result compact_result;
2042
2043                 /*
2044                  * Do not consider all the reclaimable memory because we do not
2045                  * want to trash just for a single high order allocation which
2046                  * is even not guaranteed to appear even if __compaction_suitable
2047                  * is happy about the watermark check.
2048                  */
2049                 available = zone_reclaimable_pages(zone) / order;
2050                 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
2051                 compact_result = __compaction_suitable(zone, order, alloc_flags,
2052                                 ac_classzone_idx(ac), available);
2053                 if (compact_result != COMPACT_SKIPPED)
2054                         return true;
2055         }
2056
2057         return false;
2058 }
2059
2060 static enum compact_result
2061 compact_zone(struct compact_control *cc, struct capture_control *capc)
2062 {
2063         enum compact_result ret;
2064         unsigned long start_pfn = cc->zone->zone_start_pfn;
2065         unsigned long end_pfn = zone_end_pfn(cc->zone);
2066         unsigned long last_migrated_pfn;
2067         const bool sync = cc->mode != MIGRATE_ASYNC;
2068         bool update_cached;
2069
2070         cc->migratetype = gfpflags_to_migratetype(cc->gfp_mask);
2071         ret = compaction_suitable(cc->zone, cc->order, cc->alloc_flags,
2072                                                         cc->classzone_idx);
2073         /* Compaction is likely to fail */
2074         if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED)
2075                 return ret;
2076
2077         /* huh, compaction_suitable is returning something unexpected */
2078         VM_BUG_ON(ret != COMPACT_CONTINUE);
2079
2080         /*
2081          * Clear pageblock skip if there were failures recently and compaction
2082          * is about to be retried after being deferred.
2083          */
2084         if (compaction_restarting(cc->zone, cc->order))
2085                 __reset_isolation_suitable(cc->zone);
2086
2087         /*
2088          * Setup to move all movable pages to the end of the zone. Used cached
2089          * information on where the scanners should start (unless we explicitly
2090          * want to compact the whole zone), but check that it is initialised
2091          * by ensuring the values are within zone boundaries.
2092          */
2093         cc->fast_start_pfn = 0;
2094         if (cc->whole_zone) {
2095                 cc->migrate_pfn = start_pfn;
2096                 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2097         } else {
2098                 cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync];
2099                 cc->free_pfn = cc->zone->compact_cached_free_pfn;
2100                 if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
2101                         cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2102                         cc->zone->compact_cached_free_pfn = cc->free_pfn;
2103                 }
2104                 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
2105                         cc->migrate_pfn = start_pfn;
2106                         cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
2107                         cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
2108                 }
2109
2110                 if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn)
2111                         cc->whole_zone = true;
2112         }
2113
2114         last_migrated_pfn = 0;
2115
2116         /*
2117          * Migrate has separate cached PFNs for ASYNC and SYNC* migration on
2118          * the basis that some migrations will fail in ASYNC mode. However,
2119          * if the cached PFNs match and pageblocks are skipped due to having
2120          * no isolation candidates, then the sync state does not matter.
2121          * Until a pageblock with isolation candidates is found, keep the
2122          * cached PFNs in sync to avoid revisiting the same blocks.
2123          */
2124         update_cached = !sync &&
2125                 cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1];
2126
2127         trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
2128                                 cc->free_pfn, end_pfn, sync);
2129
2130         migrate_prep_local();
2131
2132         while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) {
2133                 int err;
2134                 unsigned long start_pfn = cc->migrate_pfn;
2135
2136                 /*
2137                  * Avoid multiple rescans which can happen if a page cannot be
2138                  * isolated (dirty/writeback in async mode) or if the migrated
2139                  * pages are being allocated before the pageblock is cleared.
2140                  * The first rescan will capture the entire pageblock for
2141                  * migration. If it fails, it'll be marked skip and scanning
2142                  * will proceed as normal.
2143                  */
2144                 cc->rescan = false;
2145                 if (pageblock_start_pfn(last_migrated_pfn) ==
2146                     pageblock_start_pfn(start_pfn)) {
2147                         cc->rescan = true;
2148                 }
2149
2150                 switch (isolate_migratepages(cc->zone, cc)) {
2151                 case ISOLATE_ABORT:
2152                         ret = COMPACT_CONTENDED;
2153                         putback_movable_pages(&cc->migratepages);
2154                         cc->nr_migratepages = 0;
2155                         last_migrated_pfn = 0;
2156                         goto out;
2157                 case ISOLATE_NONE:
2158                         if (update_cached) {
2159                                 cc->zone->compact_cached_migrate_pfn[1] =
2160                                         cc->zone->compact_cached_migrate_pfn[0];
2161                         }
2162
2163                         /*
2164                          * We haven't isolated and migrated anything, but
2165                          * there might still be unflushed migrations from
2166                          * previous cc->order aligned block.
2167                          */
2168                         goto check_drain;
2169                 case ISOLATE_SUCCESS:
2170                         update_cached = false;
2171                         last_migrated_pfn = start_pfn;
2172                         ;
2173                 }
2174
2175                 err = migrate_pages(&cc->migratepages, compaction_alloc,
2176                                 compaction_free, (unsigned long)cc, cc->mode,
2177                                 MR_COMPACTION);
2178
2179                 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
2180                                                         &cc->migratepages);
2181
2182                 /* All pages were either migrated or will be released */
2183                 cc->nr_migratepages = 0;
2184                 if (err) {
2185                         putback_movable_pages(&cc->migratepages);
2186                         /*
2187                          * migrate_pages() may return -ENOMEM when scanners meet
2188                          * and we want compact_finished() to detect it
2189                          */
2190                         if (err == -ENOMEM && !compact_scanners_met(cc)) {
2191                                 ret = COMPACT_CONTENDED;
2192                                 goto out;
2193                         }
2194                         /*
2195                          * We failed to migrate at least one page in the current
2196                          * order-aligned block, so skip the rest of it.
2197                          */
2198                         if (cc->direct_compaction &&
2199                                                 (cc->mode == MIGRATE_ASYNC)) {
2200                                 cc->migrate_pfn = block_end_pfn(
2201                                                 cc->migrate_pfn - 1, cc->order);
2202                                 /* Draining pcplists is useless in this case */
2203                                 last_migrated_pfn = 0;
2204                         }
2205                 }
2206
2207 check_drain:
2208                 /*
2209                  * Has the migration scanner moved away from the previous
2210                  * cc->order aligned block where we migrated from? If yes,
2211                  * flush the pages that were freed, so that they can merge and
2212                  * compact_finished() can detect immediately if allocation
2213                  * would succeed.
2214                  */
2215                 if (cc->order > 0 && last_migrated_pfn) {
2216                         int cpu;
2217                         unsigned long current_block_start =
2218                                 block_start_pfn(cc->migrate_pfn, cc->order);
2219
2220                         if (last_migrated_pfn < current_block_start) {
2221                                 cpu = get_cpu();
2222                                 lru_add_drain_cpu(cpu);
2223                                 drain_local_pages(cc->zone);
2224                                 put_cpu();
2225                                 /* No more flushing until we migrate again */
2226                                 last_migrated_pfn = 0;
2227                         }
2228                 }
2229
2230                 /* Stop if a page has been captured */
2231                 if (capc && capc->page) {
2232                         ret = COMPACT_SUCCESS;
2233                         break;
2234                 }
2235         }
2236
2237 out:
2238         /*
2239          * Release free pages and update where the free scanner should restart,
2240          * so we don't leave any returned pages behind in the next attempt.
2241          */
2242         if (cc->nr_freepages > 0) {
2243                 unsigned long free_pfn = release_freepages(&cc->freepages);
2244
2245                 cc->nr_freepages = 0;
2246                 VM_BUG_ON(free_pfn == 0);
2247                 /* The cached pfn is always the first in a pageblock */
2248                 free_pfn = pageblock_start_pfn(free_pfn);
2249                 /*
2250                  * Only go back, not forward. The cached pfn might have been
2251                  * already reset to zone end in compact_finished()
2252                  */
2253                 if (free_pfn > cc->zone->compact_cached_free_pfn)
2254                         cc->zone->compact_cached_free_pfn = free_pfn;
2255         }
2256
2257         count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned);
2258         count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned);
2259
2260         trace_mm_compaction_end(start_pfn, cc->migrate_pfn,
2261                                 cc->free_pfn, end_pfn, sync, ret);
2262
2263         return ret;
2264 }
2265
2266 static enum compact_result compact_zone_order(struct zone *zone, int order,
2267                 gfp_t gfp_mask, enum compact_priority prio,
2268                 unsigned int alloc_flags, int classzone_idx,
2269                 struct page **capture)
2270 {
2271         enum compact_result ret;
2272         struct compact_control cc = {
2273                 .nr_freepages = 0,
2274                 .nr_migratepages = 0,
2275                 .total_migrate_scanned = 0,
2276                 .total_free_scanned = 0,
2277                 .order = order,
2278                 .search_order = order,
2279                 .gfp_mask = gfp_mask,
2280                 .zone = zone,
2281                 .mode = (prio == COMPACT_PRIO_ASYNC) ?
2282                                         MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT,
2283                 .alloc_flags = alloc_flags,
2284                 .classzone_idx = classzone_idx,
2285                 .direct_compaction = true,
2286                 .whole_zone = (prio == MIN_COMPACT_PRIORITY),
2287                 .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
2288                 .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
2289         };
2290         struct capture_control capc = {
2291                 .cc = &cc,
2292                 .page = NULL,
2293         };
2294
2295         if (capture)
2296                 current->capture_control = &capc;
2297         INIT_LIST_HEAD(&cc.freepages);
2298         INIT_LIST_HEAD(&cc.migratepages);
2299
2300         ret = compact_zone(&cc, &capc);
2301
2302         VM_BUG_ON(!list_empty(&cc.freepages));
2303         VM_BUG_ON(!list_empty(&cc.migratepages));
2304
2305         *capture = capc.page;
2306         current->capture_control = NULL;
2307
2308         return ret;
2309 }
2310
2311 int sysctl_extfrag_threshold = 500;
2312
2313 /**
2314  * try_to_compact_pages - Direct compact to satisfy a high-order allocation
2315  * @gfp_mask: The GFP mask of the current allocation
2316  * @order: The order of the current allocation
2317  * @alloc_flags: The allocation flags of the current allocation
2318  * @ac: The context of current allocation
2319  * @prio: Determines how hard direct compaction should try to succeed
2320  *
2321  * This is the main entry point for direct page compaction.
2322  */
2323 enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
2324                 unsigned int alloc_flags, const struct alloc_context *ac,
2325                 enum compact_priority prio, struct page **capture)
2326 {
2327         int may_perform_io = gfp_mask & __GFP_IO;
2328         struct zoneref *z;
2329         struct zone *zone;
2330         enum compact_result rc = COMPACT_SKIPPED;
2331
2332         /*
2333          * Check if the GFP flags allow compaction - GFP_NOIO is really
2334          * tricky context because the migration might require IO
2335          */
2336         if (!may_perform_io)
2337                 return COMPACT_SKIPPED;
2338
2339         trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
2340
2341         /* Compact each zone in the list */
2342         for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
2343                                                                 ac->nodemask) {
2344                 enum compact_result status;
2345
2346                 if (prio > MIN_COMPACT_PRIORITY
2347                                         && compaction_deferred(zone, order)) {
2348                         rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
2349                         continue;
2350                 }
2351
2352                 status = compact_zone_order(zone, order, gfp_mask, prio,
2353                                 alloc_flags, ac_classzone_idx(ac), capture);
2354                 rc = max(status, rc);
2355
2356                 /* The allocation should succeed, stop compacting */
2357                 if (status == COMPACT_SUCCESS) {
2358                         /*
2359                          * We think the allocation will succeed in this zone,
2360                          * but it is not certain, hence the false. The caller
2361                          * will repeat this with true if allocation indeed
2362                          * succeeds in this zone.
2363                          */
2364                         compaction_defer_reset(zone, order, false);
2365
2366                         break;
2367                 }
2368
2369                 if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
2370                                         status == COMPACT_PARTIAL_SKIPPED))
2371                         /*
2372                          * We think that allocation won't succeed in this zone
2373                          * so we defer compaction there. If it ends up
2374                          * succeeding after all, it will be reset.
2375                          */
2376                         defer_compaction(zone, order);
2377
2378                 /*
2379                  * We might have stopped compacting due to need_resched() in
2380                  * async compaction, or due to a fatal signal detected. In that
2381                  * case do not try further zones
2382                  */
2383                 if ((prio == COMPACT_PRIO_ASYNC && need_resched())
2384                                         || fatal_signal_pending(current))
2385                         break;
2386         }
2387
2388         return rc;
2389 }
2390
2391
2392 /* Compact all zones within a node */
2393 static void compact_node(int nid)
2394 {
2395         pg_data_t *pgdat = NODE_DATA(nid);
2396         int zoneid;
2397         struct zone *zone;
2398         struct compact_control cc = {
2399                 .order = -1,
2400                 .total_migrate_scanned = 0,
2401                 .total_free_scanned = 0,
2402                 .mode = MIGRATE_SYNC,
2403                 .ignore_skip_hint = true,
2404                 .whole_zone = true,
2405                 .gfp_mask = GFP_KERNEL,
2406         };
2407
2408
2409         for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2410
2411                 zone = &pgdat->node_zones[zoneid];
2412                 if (!populated_zone(zone))
2413                         continue;
2414
2415                 cc.nr_freepages = 0;
2416                 cc.nr_migratepages = 0;
2417                 cc.zone = zone;
2418                 INIT_LIST_HEAD(&cc.freepages);
2419                 INIT_LIST_HEAD(&cc.migratepages);
2420
2421                 compact_zone(&cc, NULL);
2422
2423                 VM_BUG_ON(!list_empty(&cc.freepages));
2424                 VM_BUG_ON(!list_empty(&cc.migratepages));
2425         }
2426 }
2427
2428 /* Compact all nodes in the system */
2429 static void compact_nodes(void)
2430 {
2431         int nid;
2432
2433         /* Flush pending updates to the LRU lists */
2434         lru_add_drain_all();
2435
2436         for_each_online_node(nid)
2437                 compact_node(nid);
2438 }
2439
2440 /* The written value is actually unused, all memory is compacted */
2441 int sysctl_compact_memory;
2442
2443 /*
2444  * This is the entry point for compacting all nodes via
2445  * /proc/sys/vm/compact_memory
2446  */
2447 int sysctl_compaction_handler(struct ctl_table *table, int write,
2448                         void __user *buffer, size_t *length, loff_t *ppos)
2449 {
2450         if (write)
2451                 compact_nodes();
2452
2453         return 0;
2454 }
2455
2456 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
2457 static ssize_t sysfs_compact_node(struct device *dev,
2458                         struct device_attribute *attr,
2459                         const char *buf, size_t count)
2460 {
2461         int nid = dev->id;
2462
2463         if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
2464                 /* Flush pending updates to the LRU lists */
2465                 lru_add_drain_all();
2466
2467                 compact_node(nid);
2468         }
2469
2470         return count;
2471 }
2472 static DEVICE_ATTR(compact, 0200, NULL, sysfs_compact_node);
2473
2474 int compaction_register_node(struct node *node)
2475 {
2476         return device_create_file(&node->dev, &dev_attr_compact);
2477 }
2478
2479 void compaction_unregister_node(struct node *node)
2480 {
2481         return device_remove_file(&node->dev, &dev_attr_compact);
2482 }
2483 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
2484
2485 static inline bool kcompactd_work_requested(pg_data_t *pgdat)
2486 {
2487         return pgdat->kcompactd_max_order > 0 || kthread_should_stop();
2488 }
2489
2490 static bool kcompactd_node_suitable(pg_data_t *pgdat)
2491 {
2492         int zoneid;
2493         struct zone *zone;
2494         enum zone_type classzone_idx = pgdat->kcompactd_classzone_idx;
2495
2496         for (zoneid = 0; zoneid <= classzone_idx; zoneid++) {
2497                 zone = &pgdat->node_zones[zoneid];
2498
2499                 if (!populated_zone(zone))
2500                         continue;
2501
2502                 if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0,
2503                                         classzone_idx) == COMPACT_CONTINUE)
2504                         return true;
2505         }
2506
2507         return false;
2508 }
2509
2510 static void kcompactd_do_work(pg_data_t *pgdat)
2511 {
2512         /*
2513          * With no special task, compact all zones so that a page of requested
2514          * order is allocatable.
2515          */
2516         int zoneid;
2517         struct zone *zone;
2518         struct compact_control cc = {
2519                 .order = pgdat->kcompactd_max_order,
2520                 .search_order = pgdat->kcompactd_max_order,
2521                 .total_migrate_scanned = 0,
2522                 .total_free_scanned = 0,
2523                 .classzone_idx = pgdat->kcompactd_classzone_idx,
2524                 .mode = MIGRATE_SYNC_LIGHT,
2525                 .ignore_skip_hint = false,
2526                 .gfp_mask = GFP_KERNEL,
2527         };
2528         trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
2529                                                         cc.classzone_idx);
2530         count_compact_event(KCOMPACTD_WAKE);
2531
2532         for (zoneid = 0; zoneid <= cc.classzone_idx; zoneid++) {
2533                 int status;
2534
2535                 zone = &pgdat->node_zones[zoneid];
2536                 if (!populated_zone(zone))
2537                         continue;
2538
2539                 if (compaction_deferred(zone, cc.order))
2540                         continue;
2541
2542                 if (compaction_suitable(zone, cc.order, 0, zoneid) !=
2543                                                         COMPACT_CONTINUE)
2544                         continue;
2545
2546                 cc.nr_freepages = 0;
2547                 cc.nr_migratepages = 0;
2548                 cc.total_migrate_scanned = 0;
2549                 cc.total_free_scanned = 0;
2550                 cc.zone = zone;
2551                 INIT_LIST_HEAD(&cc.freepages);
2552                 INIT_LIST_HEAD(&cc.migratepages);
2553
2554                 if (kthread_should_stop())
2555                         return;
2556                 status = compact_zone(&cc, NULL);
2557
2558                 if (status == COMPACT_SUCCESS) {
2559                         compaction_defer_reset(zone, cc.order, false);
2560                 } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
2561                         /*
2562                          * Buddy pages may become stranded on pcps that could
2563                          * otherwise coalesce on the zone's free area for
2564                          * order >= cc.order.  This is ratelimited by the
2565                          * upcoming deferral.
2566                          */
2567                         drain_all_pages(zone);
2568
2569                         /*
2570                          * We use sync migration mode here, so we defer like
2571                          * sync direct compaction does.
2572                          */
2573                         defer_compaction(zone, cc.order);
2574                 }
2575
2576                 count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
2577                                      cc.total_migrate_scanned);
2578                 count_compact_events(KCOMPACTD_FREE_SCANNED,
2579                                      cc.total_free_scanned);
2580
2581                 VM_BUG_ON(!list_empty(&cc.freepages));
2582                 VM_BUG_ON(!list_empty(&cc.migratepages));
2583         }
2584
2585         /*
2586          * Regardless of success, we are done until woken up next. But remember
2587          * the requested order/classzone_idx in case it was higher/tighter than
2588          * our current ones
2589          */
2590         if (pgdat->kcompactd_max_order <= cc.order)
2591                 pgdat->kcompactd_max_order = 0;
2592         if (pgdat->kcompactd_classzone_idx >= cc.classzone_idx)
2593                 pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
2594 }
2595
2596 void wakeup_kcompactd(pg_data_t *pgdat, int order, int classzone_idx)
2597 {
2598         if (!order)
2599                 return;
2600
2601         if (pgdat->kcompactd_max_order < order)
2602                 pgdat->kcompactd_max_order = order;
2603
2604         if (pgdat->kcompactd_classzone_idx > classzone_idx)
2605                 pgdat->kcompactd_classzone_idx = classzone_idx;
2606
2607         /*
2608          * Pairs with implicit barrier in wait_event_freezable()
2609          * such that wakeups are not missed.
2610          */
2611         if (!wq_has_sleeper(&pgdat->kcompactd_wait))
2612                 return;
2613
2614         if (!kcompactd_node_suitable(pgdat))
2615                 return;
2616
2617         trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
2618                                                         classzone_idx);
2619         wake_up_interruptible(&pgdat->kcompactd_wait);
2620 }
2621
2622 /*
2623  * The background compaction daemon, started as a kernel thread
2624  * from the init process.
2625  */
2626 static int kcompactd(void *p)
2627 {
2628         pg_data_t *pgdat = (pg_data_t*)p;
2629         struct task_struct *tsk = current;
2630
2631         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2632
2633         if (!cpumask_empty(cpumask))
2634                 set_cpus_allowed_ptr(tsk, cpumask);
2635
2636         set_freezable();
2637
2638         pgdat->kcompactd_max_order = 0;
2639         pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
2640
2641         while (!kthread_should_stop()) {
2642                 unsigned long pflags;
2643
2644                 trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
2645                 wait_event_freezable(pgdat->kcompactd_wait,
2646                                 kcompactd_work_requested(pgdat));
2647
2648                 psi_memstall_enter(&pflags);
2649                 kcompactd_do_work(pgdat);
2650                 psi_memstall_leave(&pflags);
2651         }
2652
2653         return 0;
2654 }
2655
2656 /*
2657  * This kcompactd start function will be called by init and node-hot-add.
2658  * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
2659  */
2660 int kcompactd_run(int nid)
2661 {
2662         pg_data_t *pgdat = NODE_DATA(nid);
2663         int ret = 0;
2664
2665         if (pgdat->kcompactd)
2666                 return 0;
2667
2668         pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
2669         if (IS_ERR(pgdat->kcompactd)) {
2670                 pr_err("Failed to start kcompactd on node %d\n", nid);
2671                 ret = PTR_ERR(pgdat->kcompactd);
2672                 pgdat->kcompactd = NULL;
2673         }
2674         return ret;
2675 }
2676
2677 /*
2678  * Called by memory hotplug when all memory in a node is offlined. Caller must
2679  * hold mem_hotplug_begin/end().
2680  */
2681 void kcompactd_stop(int nid)
2682 {
2683         struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
2684
2685         if (kcompactd) {
2686                 kthread_stop(kcompactd);
2687                 NODE_DATA(nid)->kcompactd = NULL;
2688         }
2689 }
2690
2691 /*
2692  * It's optimal to keep kcompactd on the same CPUs as their memory, but
2693  * not required for correctness. So if the last cpu in a node goes
2694  * away, we get changed to run anywhere: as the first one comes back,
2695  * restore their cpu bindings.
2696  */
2697 static int kcompactd_cpu_online(unsigned int cpu)
2698 {
2699         int nid;
2700
2701         for_each_node_state(nid, N_MEMORY) {
2702                 pg_data_t *pgdat = NODE_DATA(nid);
2703                 const struct cpumask *mask;
2704
2705                 mask = cpumask_of_node(pgdat->node_id);
2706
2707                 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2708                         /* One of our CPUs online: restore mask */
2709                         set_cpus_allowed_ptr(pgdat->kcompactd, mask);
2710         }
2711         return 0;
2712 }
2713
2714 static int __init kcompactd_init(void)
2715 {
2716         int nid;
2717         int ret;
2718
2719         ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
2720                                         "mm/compaction:online",
2721                                         kcompactd_cpu_online, NULL);
2722         if (ret < 0) {
2723                 pr_err("kcompactd: failed to register hotplug callbacks.\n");
2724                 return ret;
2725         }
2726
2727         for_each_node_state(nid, N_MEMORY)
2728                 kcompactd_run(nid);
2729         return 0;
2730 }
2731 subsys_initcall(kcompactd_init)
2732
2733 #endif /* CONFIG_COMPACTION */