ubifs: Set s_uuid in super block to support ima/evm uuid options
[linux-2.6-microblaze.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 /*
54  * Fragmentation score check interval for proactive compaction purposes.
55  */
56 static const unsigned int HPAGE_FRAG_CHECK_INTERVAL_MSEC = 500;
57
58 /*
59  * Page order with-respect-to which proactive compaction
60  * calculates external fragmentation, which is used as
61  * the "fragmentation score" of a node/zone.
62  */
63 #if defined CONFIG_TRANSPARENT_HUGEPAGE
64 #define COMPACTION_HPAGE_ORDER  HPAGE_PMD_ORDER
65 #elif defined CONFIG_HUGETLBFS
66 #define COMPACTION_HPAGE_ORDER  HUGETLB_PAGE_ORDER
67 #else
68 #define COMPACTION_HPAGE_ORDER  (PMD_SHIFT - PAGE_SHIFT)
69 #endif
70
71 static unsigned long release_freepages(struct list_head *freelist)
72 {
73         struct page *page, *next;
74         unsigned long high_pfn = 0;
75
76         list_for_each_entry_safe(page, next, freelist, lru) {
77                 unsigned long pfn = page_to_pfn(page);
78                 list_del(&page->lru);
79                 __free_page(page);
80                 if (pfn > high_pfn)
81                         high_pfn = pfn;
82         }
83
84         return high_pfn;
85 }
86
87 static void split_map_pages(struct list_head *list)
88 {
89         unsigned int i, order, nr_pages;
90         struct page *page, *next;
91         LIST_HEAD(tmp_list);
92
93         list_for_each_entry_safe(page, next, list, lru) {
94                 list_del(&page->lru);
95
96                 order = page_private(page);
97                 nr_pages = 1 << order;
98
99                 post_alloc_hook(page, order, __GFP_MOVABLE);
100                 if (order)
101                         split_page(page, order);
102
103                 for (i = 0; i < nr_pages; i++) {
104                         list_add(&page->lru, &tmp_list);
105                         page++;
106                 }
107         }
108
109         list_splice(&tmp_list, list);
110 }
111
112 #ifdef CONFIG_COMPACTION
113
114 int PageMovable(struct page *page)
115 {
116         struct address_space *mapping;
117
118         VM_BUG_ON_PAGE(!PageLocked(page), page);
119         if (!__PageMovable(page))
120                 return 0;
121
122         mapping = page_mapping(page);
123         if (mapping && mapping->a_ops && mapping->a_ops->isolate_page)
124                 return 1;
125
126         return 0;
127 }
128 EXPORT_SYMBOL(PageMovable);
129
130 void __SetPageMovable(struct page *page, struct address_space *mapping)
131 {
132         VM_BUG_ON_PAGE(!PageLocked(page), page);
133         VM_BUG_ON_PAGE((unsigned long)mapping & PAGE_MAPPING_MOVABLE, page);
134         page->mapping = (void *)((unsigned long)mapping | PAGE_MAPPING_MOVABLE);
135 }
136 EXPORT_SYMBOL(__SetPageMovable);
137
138 void __ClearPageMovable(struct page *page)
139 {
140         VM_BUG_ON_PAGE(!PageMovable(page), page);
141         /*
142          * Clear registered address_space val with keeping PAGE_MAPPING_MOVABLE
143          * flag so that VM can catch up released page by driver after isolation.
144          * With it, VM migration doesn't try to put it back.
145          */
146         page->mapping = (void *)((unsigned long)page->mapping &
147                                 PAGE_MAPPING_MOVABLE);
148 }
149 EXPORT_SYMBOL(__ClearPageMovable);
150
151 /* Do not skip compaction more than 64 times */
152 #define COMPACT_MAX_DEFER_SHIFT 6
153
154 /*
155  * Compaction is deferred when compaction fails to result in a page
156  * allocation success. 1 << compact_defer_shift, compactions are skipped up
157  * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
158  */
159 static void defer_compaction(struct zone *zone, int order)
160 {
161         zone->compact_considered = 0;
162         zone->compact_defer_shift++;
163
164         if (order < zone->compact_order_failed)
165                 zone->compact_order_failed = order;
166
167         if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
168                 zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
169
170         trace_mm_compaction_defer_compaction(zone, order);
171 }
172
173 /* Returns true if compaction should be skipped this time */
174 static bool compaction_deferred(struct zone *zone, int order)
175 {
176         unsigned long defer_limit = 1UL << zone->compact_defer_shift;
177
178         if (order < zone->compact_order_failed)
179                 return false;
180
181         /* Avoid possible overflow */
182         if (++zone->compact_considered >= defer_limit) {
183                 zone->compact_considered = defer_limit;
184                 return false;
185         }
186
187         trace_mm_compaction_deferred(zone, order);
188
189         return true;
190 }
191
192 /*
193  * Update defer tracking counters after successful compaction of given order,
194  * which means an allocation either succeeded (alloc_success == true) or is
195  * expected to succeed.
196  */
197 void compaction_defer_reset(struct zone *zone, int order,
198                 bool alloc_success)
199 {
200         if (alloc_success) {
201                 zone->compact_considered = 0;
202                 zone->compact_defer_shift = 0;
203         }
204         if (order >= zone->compact_order_failed)
205                 zone->compact_order_failed = order + 1;
206
207         trace_mm_compaction_defer_reset(zone, order);
208 }
209
210 /* Returns true if restarting compaction after many failures */
211 static bool compaction_restarting(struct zone *zone, int order)
212 {
213         if (order < zone->compact_order_failed)
214                 return false;
215
216         return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
217                 zone->compact_considered >= 1UL << zone->compact_defer_shift;
218 }
219
220 /* Returns true if the pageblock should be scanned for pages to isolate. */
221 static inline bool isolation_suitable(struct compact_control *cc,
222                                         struct page *page)
223 {
224         if (cc->ignore_skip_hint)
225                 return true;
226
227         return !get_pageblock_skip(page);
228 }
229
230 static void reset_cached_positions(struct zone *zone)
231 {
232         zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
233         zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
234         zone->compact_cached_free_pfn =
235                                 pageblock_start_pfn(zone_end_pfn(zone) - 1);
236 }
237
238 /*
239  * Compound pages of >= pageblock_order should consistently be skipped until
240  * released. It is always pointless to compact pages of such order (if they are
241  * migratable), and the pageblocks they occupy cannot contain any free pages.
242  */
243 static bool pageblock_skip_persistent(struct page *page)
244 {
245         if (!PageCompound(page))
246                 return false;
247
248         page = compound_head(page);
249
250         if (compound_order(page) >= pageblock_order)
251                 return true;
252
253         return false;
254 }
255
256 static bool
257 __reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source,
258                                                         bool check_target)
259 {
260         struct page *page = pfn_to_online_page(pfn);
261         struct page *block_page;
262         struct page *end_page;
263         unsigned long block_pfn;
264
265         if (!page)
266                 return false;
267         if (zone != page_zone(page))
268                 return false;
269         if (pageblock_skip_persistent(page))
270                 return false;
271
272         /*
273          * If skip is already cleared do no further checking once the
274          * restart points have been set.
275          */
276         if (check_source && check_target && !get_pageblock_skip(page))
277                 return true;
278
279         /*
280          * If clearing skip for the target scanner, do not select a
281          * non-movable pageblock as the starting point.
282          */
283         if (!check_source && check_target &&
284             get_pageblock_migratetype(page) != MIGRATE_MOVABLE)
285                 return false;
286
287         /* Ensure the start of the pageblock or zone is online and valid */
288         block_pfn = pageblock_start_pfn(pfn);
289         block_pfn = max(block_pfn, zone->zone_start_pfn);
290         block_page = pfn_to_online_page(block_pfn);
291         if (block_page) {
292                 page = block_page;
293                 pfn = block_pfn;
294         }
295
296         /* Ensure the end of the pageblock or zone is online and valid */
297         block_pfn = pageblock_end_pfn(pfn) - 1;
298         block_pfn = min(block_pfn, zone_end_pfn(zone) - 1);
299         end_page = pfn_to_online_page(block_pfn);
300         if (!end_page)
301                 return false;
302
303         /*
304          * Only clear the hint if a sample indicates there is either a
305          * free page or an LRU page in the block. One or other condition
306          * is necessary for the block to be a migration source/target.
307          */
308         do {
309                 if (pfn_valid_within(pfn)) {
310                         if (check_source && PageLRU(page)) {
311                                 clear_pageblock_skip(page);
312                                 return true;
313                         }
314
315                         if (check_target && PageBuddy(page)) {
316                                 clear_pageblock_skip(page);
317                                 return true;
318                         }
319                 }
320
321                 page += (1 << PAGE_ALLOC_COSTLY_ORDER);
322                 pfn += (1 << PAGE_ALLOC_COSTLY_ORDER);
323         } while (page <= end_page);
324
325         return false;
326 }
327
328 /*
329  * This function is called to clear all cached information on pageblocks that
330  * should be skipped for page isolation when the migrate and free page scanner
331  * meet.
332  */
333 static void __reset_isolation_suitable(struct zone *zone)
334 {
335         unsigned long migrate_pfn = zone->zone_start_pfn;
336         unsigned long free_pfn = zone_end_pfn(zone) - 1;
337         unsigned long reset_migrate = free_pfn;
338         unsigned long reset_free = migrate_pfn;
339         bool source_set = false;
340         bool free_set = false;
341
342         if (!zone->compact_blockskip_flush)
343                 return;
344
345         zone->compact_blockskip_flush = false;
346
347         /*
348          * Walk the zone and update pageblock skip information. Source looks
349          * for PageLRU while target looks for PageBuddy. When the scanner
350          * is found, both PageBuddy and PageLRU are checked as the pageblock
351          * is suitable as both source and target.
352          */
353         for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages,
354                                         free_pfn -= pageblock_nr_pages) {
355                 cond_resched();
356
357                 /* Update the migrate PFN */
358                 if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) &&
359                     migrate_pfn < reset_migrate) {
360                         source_set = true;
361                         reset_migrate = migrate_pfn;
362                         zone->compact_init_migrate_pfn = reset_migrate;
363                         zone->compact_cached_migrate_pfn[0] = reset_migrate;
364                         zone->compact_cached_migrate_pfn[1] = reset_migrate;
365                 }
366
367                 /* Update the free PFN */
368                 if (__reset_isolation_pfn(zone, free_pfn, free_set, true) &&
369                     free_pfn > reset_free) {
370                         free_set = true;
371                         reset_free = free_pfn;
372                         zone->compact_init_free_pfn = reset_free;
373                         zone->compact_cached_free_pfn = reset_free;
374                 }
375         }
376
377         /* Leave no distance if no suitable block was reset */
378         if (reset_migrate >= reset_free) {
379                 zone->compact_cached_migrate_pfn[0] = migrate_pfn;
380                 zone->compact_cached_migrate_pfn[1] = migrate_pfn;
381                 zone->compact_cached_free_pfn = free_pfn;
382         }
383 }
384
385 void reset_isolation_suitable(pg_data_t *pgdat)
386 {
387         int zoneid;
388
389         for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
390                 struct zone *zone = &pgdat->node_zones[zoneid];
391                 if (!populated_zone(zone))
392                         continue;
393
394                 /* Only flush if a full compaction finished recently */
395                 if (zone->compact_blockskip_flush)
396                         __reset_isolation_suitable(zone);
397         }
398 }
399
400 /*
401  * Sets the pageblock skip bit if it was clear. Note that this is a hint as
402  * locks are not required for read/writers. Returns true if it was already set.
403  */
404 static bool test_and_set_skip(struct compact_control *cc, struct page *page,
405                                                         unsigned long pfn)
406 {
407         bool skip;
408
409         /* Do no update if skip hint is being ignored */
410         if (cc->ignore_skip_hint)
411                 return false;
412
413         if (!IS_ALIGNED(pfn, pageblock_nr_pages))
414                 return false;
415
416         skip = get_pageblock_skip(page);
417         if (!skip && !cc->no_set_skip_hint)
418                 set_pageblock_skip(page);
419
420         return skip;
421 }
422
423 static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
424 {
425         struct zone *zone = cc->zone;
426
427         pfn = pageblock_end_pfn(pfn);
428
429         /* Set for isolation rather than compaction */
430         if (cc->no_set_skip_hint)
431                 return;
432
433         if (pfn > zone->compact_cached_migrate_pfn[0])
434                 zone->compact_cached_migrate_pfn[0] = pfn;
435         if (cc->mode != MIGRATE_ASYNC &&
436             pfn > zone->compact_cached_migrate_pfn[1])
437                 zone->compact_cached_migrate_pfn[1] = pfn;
438 }
439
440 /*
441  * If no pages were isolated then mark this pageblock to be skipped in the
442  * future. The information is later cleared by __reset_isolation_suitable().
443  */
444 static void update_pageblock_skip(struct compact_control *cc,
445                         struct page *page, unsigned long pfn)
446 {
447         struct zone *zone = cc->zone;
448
449         if (cc->no_set_skip_hint)
450                 return;
451
452         if (!page)
453                 return;
454
455         set_pageblock_skip(page);
456
457         /* Update where async and sync compaction should restart */
458         if (pfn < zone->compact_cached_free_pfn)
459                 zone->compact_cached_free_pfn = pfn;
460 }
461 #else
462 static inline bool isolation_suitable(struct compact_control *cc,
463                                         struct page *page)
464 {
465         return true;
466 }
467
468 static inline bool pageblock_skip_persistent(struct page *page)
469 {
470         return false;
471 }
472
473 static inline void update_pageblock_skip(struct compact_control *cc,
474                         struct page *page, unsigned long pfn)
475 {
476 }
477
478 static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
479 {
480 }
481
482 static bool test_and_set_skip(struct compact_control *cc, struct page *page,
483                                                         unsigned long pfn)
484 {
485         return false;
486 }
487 #endif /* CONFIG_COMPACTION */
488
489 /*
490  * Compaction requires the taking of some coarse locks that are potentially
491  * very heavily contended. For async compaction, trylock and record if the
492  * lock is contended. The lock will still be acquired but compaction will
493  * abort when the current block is finished regardless of success rate.
494  * Sync compaction acquires the lock.
495  *
496  * Always returns true which makes it easier to track lock state in callers.
497  */
498 static bool compact_lock_irqsave(spinlock_t *lock, unsigned long *flags,
499                                                 struct compact_control *cc)
500         __acquires(lock)
501 {
502         /* Track if the lock is contended in async mode */
503         if (cc->mode == MIGRATE_ASYNC && !cc->contended) {
504                 if (spin_trylock_irqsave(lock, *flags))
505                         return true;
506
507                 cc->contended = true;
508         }
509
510         spin_lock_irqsave(lock, *flags);
511         return true;
512 }
513
514 /*
515  * Compaction requires the taking of some coarse locks that are potentially
516  * very heavily contended. The lock should be periodically unlocked to avoid
517  * having disabled IRQs for a long time, even when there is nobody waiting on
518  * the lock. It might also be that allowing the IRQs will result in
519  * need_resched() becoming true. If scheduling is needed, async compaction
520  * aborts. Sync compaction schedules.
521  * Either compaction type will also abort if a fatal signal is pending.
522  * In either case if the lock was locked, it is dropped and not regained.
523  *
524  * Returns true if compaction should abort due to fatal signal pending, or
525  *              async compaction due to need_resched()
526  * Returns false when compaction can continue (sync compaction might have
527  *              scheduled)
528  */
529 static bool compact_unlock_should_abort(spinlock_t *lock,
530                 unsigned long flags, bool *locked, struct compact_control *cc)
531 {
532         if (*locked) {
533                 spin_unlock_irqrestore(lock, flags);
534                 *locked = false;
535         }
536
537         if (fatal_signal_pending(current)) {
538                 cc->contended = true;
539                 return true;
540         }
541
542         cond_resched();
543
544         return false;
545 }
546
547 /*
548  * Isolate free pages onto a private freelist. If @strict is true, will abort
549  * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
550  * (even though it may still end up isolating some pages).
551  */
552 static unsigned long isolate_freepages_block(struct compact_control *cc,
553                                 unsigned long *start_pfn,
554                                 unsigned long end_pfn,
555                                 struct list_head *freelist,
556                                 unsigned int stride,
557                                 bool strict)
558 {
559         int nr_scanned = 0, total_isolated = 0;
560         struct page *cursor;
561         unsigned long flags = 0;
562         bool locked = false;
563         unsigned long blockpfn = *start_pfn;
564         unsigned int order;
565
566         /* Strict mode is for isolation, speed is secondary */
567         if (strict)
568                 stride = 1;
569
570         cursor = pfn_to_page(blockpfn);
571
572         /* Isolate free pages. */
573         for (; blockpfn < end_pfn; blockpfn += stride, cursor += stride) {
574                 int isolated;
575                 struct page *page = cursor;
576
577                 /*
578                  * Periodically drop the lock (if held) regardless of its
579                  * contention, to give chance to IRQs. Abort if fatal signal
580                  * pending or async compaction detects need_resched()
581                  */
582                 if (!(blockpfn % SWAP_CLUSTER_MAX)
583                     && compact_unlock_should_abort(&cc->zone->lock, flags,
584                                                                 &locked, cc))
585                         break;
586
587                 nr_scanned++;
588                 if (!pfn_valid_within(blockpfn))
589                         goto isolate_fail;
590
591                 /*
592                  * For compound pages such as THP and hugetlbfs, we can save
593                  * potentially a lot of iterations if we skip them at once.
594                  * The check is racy, but we can consider only valid values
595                  * and the only danger is skipping too much.
596                  */
597                 if (PageCompound(page)) {
598                         const unsigned int order = compound_order(page);
599
600                         if (likely(order < MAX_ORDER)) {
601                                 blockpfn += (1UL << order) - 1;
602                                 cursor += (1UL << order) - 1;
603                         }
604                         goto isolate_fail;
605                 }
606
607                 if (!PageBuddy(page))
608                         goto isolate_fail;
609
610                 /*
611                  * If we already hold the lock, we can skip some rechecking.
612                  * Note that if we hold the lock now, checked_pageblock was
613                  * already set in some previous iteration (or strict is true),
614                  * so it is correct to skip the suitable migration target
615                  * recheck as well.
616                  */
617                 if (!locked) {
618                         locked = compact_lock_irqsave(&cc->zone->lock,
619                                                                 &flags, cc);
620
621                         /* Recheck this is a buddy page under lock */
622                         if (!PageBuddy(page))
623                                 goto isolate_fail;
624                 }
625
626                 /* Found a free page, will break it into order-0 pages */
627                 order = buddy_order(page);
628                 isolated = __isolate_free_page(page, order);
629                 if (!isolated)
630                         break;
631                 set_page_private(page, order);
632
633                 total_isolated += isolated;
634                 cc->nr_freepages += isolated;
635                 list_add_tail(&page->lru, freelist);
636
637                 if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
638                         blockpfn += isolated;
639                         break;
640                 }
641                 /* Advance to the end of split page */
642                 blockpfn += isolated - 1;
643                 cursor += isolated - 1;
644                 continue;
645
646 isolate_fail:
647                 if (strict)
648                         break;
649                 else
650                         continue;
651
652         }
653
654         if (locked)
655                 spin_unlock_irqrestore(&cc->zone->lock, flags);
656
657         /*
658          * There is a tiny chance that we have read bogus compound_order(),
659          * so be careful to not go outside of the pageblock.
660          */
661         if (unlikely(blockpfn > end_pfn))
662                 blockpfn = end_pfn;
663
664         trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
665                                         nr_scanned, total_isolated);
666
667         /* Record how far we have got within the block */
668         *start_pfn = blockpfn;
669
670         /*
671          * If strict isolation is requested by CMA then check that all the
672          * pages requested were isolated. If there were any failures, 0 is
673          * returned and CMA will fail.
674          */
675         if (strict && blockpfn < end_pfn)
676                 total_isolated = 0;
677
678         cc->total_free_scanned += nr_scanned;
679         if (total_isolated)
680                 count_compact_events(COMPACTISOLATED, total_isolated);
681         return total_isolated;
682 }
683
684 /**
685  * isolate_freepages_range() - isolate free pages.
686  * @cc:        Compaction control structure.
687  * @start_pfn: The first PFN to start isolating.
688  * @end_pfn:   The one-past-last PFN.
689  *
690  * Non-free pages, invalid PFNs, or zone boundaries within the
691  * [start_pfn, end_pfn) range are considered errors, cause function to
692  * undo its actions and return zero.
693  *
694  * Otherwise, function returns one-past-the-last PFN of isolated page
695  * (which may be greater then end_pfn if end fell in a middle of
696  * a free page).
697  */
698 unsigned long
699 isolate_freepages_range(struct compact_control *cc,
700                         unsigned long start_pfn, unsigned long end_pfn)
701 {
702         unsigned long isolated, pfn, block_start_pfn, block_end_pfn;
703         LIST_HEAD(freelist);
704
705         pfn = start_pfn;
706         block_start_pfn = pageblock_start_pfn(pfn);
707         if (block_start_pfn < cc->zone->zone_start_pfn)
708                 block_start_pfn = cc->zone->zone_start_pfn;
709         block_end_pfn = pageblock_end_pfn(pfn);
710
711         for (; pfn < end_pfn; pfn += isolated,
712                                 block_start_pfn = block_end_pfn,
713                                 block_end_pfn += pageblock_nr_pages) {
714                 /* Protect pfn from changing by isolate_freepages_block */
715                 unsigned long isolate_start_pfn = pfn;
716
717                 block_end_pfn = min(block_end_pfn, end_pfn);
718
719                 /*
720                  * pfn could pass the block_end_pfn if isolated freepage
721                  * is more than pageblock order. In this case, we adjust
722                  * scanning range to right one.
723                  */
724                 if (pfn >= block_end_pfn) {
725                         block_start_pfn = pageblock_start_pfn(pfn);
726                         block_end_pfn = pageblock_end_pfn(pfn);
727                         block_end_pfn = min(block_end_pfn, end_pfn);
728                 }
729
730                 if (!pageblock_pfn_to_page(block_start_pfn,
731                                         block_end_pfn, cc->zone))
732                         break;
733
734                 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
735                                         block_end_pfn, &freelist, 0, true);
736
737                 /*
738                  * In strict mode, isolate_freepages_block() returns 0 if
739                  * there are any holes in the block (ie. invalid PFNs or
740                  * non-free pages).
741                  */
742                 if (!isolated)
743                         break;
744
745                 /*
746                  * If we managed to isolate pages, it is always (1 << n) *
747                  * pageblock_nr_pages for some non-negative n.  (Max order
748                  * page may span two pageblocks).
749                  */
750         }
751
752         /* __isolate_free_page() does not map the pages */
753         split_map_pages(&freelist);
754
755         if (pfn < end_pfn) {
756                 /* Loop terminated early, cleanup. */
757                 release_freepages(&freelist);
758                 return 0;
759         }
760
761         /* We don't use freelists for anything. */
762         return pfn;
763 }
764
765 /* Similar to reclaim, but different enough that they don't share logic */
766 static bool too_many_isolated(pg_data_t *pgdat)
767 {
768         unsigned long active, inactive, isolated;
769
770         inactive = node_page_state(pgdat, NR_INACTIVE_FILE) +
771                         node_page_state(pgdat, NR_INACTIVE_ANON);
772         active = node_page_state(pgdat, NR_ACTIVE_FILE) +
773                         node_page_state(pgdat, NR_ACTIVE_ANON);
774         isolated = node_page_state(pgdat, NR_ISOLATED_FILE) +
775                         node_page_state(pgdat, NR_ISOLATED_ANON);
776
777         return isolated > (inactive + active) / 2;
778 }
779
780 /**
781  * isolate_migratepages_block() - isolate all migrate-able pages within
782  *                                a single pageblock
783  * @cc:         Compaction control structure.
784  * @low_pfn:    The first PFN to isolate
785  * @end_pfn:    The one-past-the-last PFN to isolate, within same pageblock
786  * @isolate_mode: Isolation mode to be used.
787  *
788  * Isolate all pages that can be migrated from the range specified by
789  * [low_pfn, end_pfn). The range is expected to be within same pageblock.
790  * Returns zero if there is a fatal signal pending, otherwise PFN of the
791  * first page that was not scanned (which may be both less, equal to or more
792  * than end_pfn).
793  *
794  * The pages are isolated on cc->migratepages list (not required to be empty),
795  * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
796  * is neither read nor updated.
797  */
798 static unsigned long
799 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
800                         unsigned long end_pfn, isolate_mode_t isolate_mode)
801 {
802         pg_data_t *pgdat = cc->zone->zone_pgdat;
803         unsigned long nr_scanned = 0, nr_isolated = 0;
804         struct lruvec *lruvec;
805         unsigned long flags = 0;
806         struct lruvec *locked = NULL;
807         struct page *page = NULL, *valid_page = NULL;
808         unsigned long start_pfn = low_pfn;
809         bool skip_on_failure = false;
810         unsigned long next_skip_pfn = 0;
811         bool skip_updated = false;
812
813         /*
814          * Ensure that there are not too many pages isolated from the LRU
815          * list by either parallel reclaimers or compaction. If there are,
816          * delay for some time until fewer pages are isolated
817          */
818         while (unlikely(too_many_isolated(pgdat))) {
819                 /* stop isolation if there are still pages not migrated */
820                 if (cc->nr_migratepages)
821                         return 0;
822
823                 /* async migration should just abort */
824                 if (cc->mode == MIGRATE_ASYNC)
825                         return 0;
826
827                 congestion_wait(BLK_RW_ASYNC, HZ/10);
828
829                 if (fatal_signal_pending(current))
830                         return 0;
831         }
832
833         cond_resched();
834
835         if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
836                 skip_on_failure = true;
837                 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
838         }
839
840         /* Time to isolate some pages for migration */
841         for (; low_pfn < end_pfn; low_pfn++) {
842
843                 if (skip_on_failure && low_pfn >= next_skip_pfn) {
844                         /*
845                          * We have isolated all migration candidates in the
846                          * previous order-aligned block, and did not skip it due
847                          * to failure. We should migrate the pages now and
848                          * hopefully succeed compaction.
849                          */
850                         if (nr_isolated)
851                                 break;
852
853                         /*
854                          * We failed to isolate in the previous order-aligned
855                          * block. Set the new boundary to the end of the
856                          * current block. Note we can't simply increase
857                          * next_skip_pfn by 1 << order, as low_pfn might have
858                          * been incremented by a higher number due to skipping
859                          * a compound or a high-order buddy page in the
860                          * previous loop iteration.
861                          */
862                         next_skip_pfn = block_end_pfn(low_pfn, cc->order);
863                 }
864
865                 /*
866                  * Periodically drop the lock (if held) regardless of its
867                  * contention, to give chance to IRQs. Abort completely if
868                  * a fatal signal is pending.
869                  */
870                 if (!(low_pfn % SWAP_CLUSTER_MAX)) {
871                         if (locked) {
872                                 unlock_page_lruvec_irqrestore(locked, flags);
873                                 locked = NULL;
874                         }
875
876                         if (fatal_signal_pending(current)) {
877                                 cc->contended = true;
878
879                                 low_pfn = 0;
880                                 goto fatal_pending;
881                         }
882
883                         cond_resched();
884                 }
885
886                 if (!pfn_valid_within(low_pfn))
887                         goto isolate_fail;
888                 nr_scanned++;
889
890                 page = pfn_to_page(low_pfn);
891
892                 /*
893                  * Check if the pageblock has already been marked skipped.
894                  * Only the aligned PFN is checked as the caller isolates
895                  * COMPACT_CLUSTER_MAX at a time so the second call must
896                  * not falsely conclude that the block should be skipped.
897                  */
898                 if (!valid_page && IS_ALIGNED(low_pfn, pageblock_nr_pages)) {
899                         if (!cc->ignore_skip_hint && get_pageblock_skip(page)) {
900                                 low_pfn = end_pfn;
901                                 page = NULL;
902                                 goto isolate_abort;
903                         }
904                         valid_page = page;
905                 }
906
907                 /*
908                  * Skip if free. We read page order here without zone lock
909                  * which is generally unsafe, but the race window is small and
910                  * the worst thing that can happen is that we skip some
911                  * potential isolation targets.
912                  */
913                 if (PageBuddy(page)) {
914                         unsigned long freepage_order = buddy_order_unsafe(page);
915
916                         /*
917                          * Without lock, we cannot be sure that what we got is
918                          * a valid page order. Consider only values in the
919                          * valid order range to prevent low_pfn overflow.
920                          */
921                         if (freepage_order > 0 && freepage_order < MAX_ORDER)
922                                 low_pfn += (1UL << freepage_order) - 1;
923                         continue;
924                 }
925
926                 /*
927                  * Regardless of being on LRU, compound pages such as THP and
928                  * hugetlbfs are not to be compacted unless we are attempting
929                  * an allocation much larger than the huge page size (eg CMA).
930                  * We can potentially save a lot of iterations if we skip them
931                  * at once. The check is racy, but we can consider only valid
932                  * values and the only danger is skipping too much.
933                  */
934                 if (PageCompound(page) && !cc->alloc_contig) {
935                         const unsigned int order = compound_order(page);
936
937                         if (likely(order < MAX_ORDER))
938                                 low_pfn += (1UL << order) - 1;
939                         goto isolate_fail;
940                 }
941
942                 /*
943                  * Check may be lockless but that's ok as we recheck later.
944                  * It's possible to migrate LRU and non-lru movable pages.
945                  * Skip any other type of page
946                  */
947                 if (!PageLRU(page)) {
948                         /*
949                          * __PageMovable can return false positive so we need
950                          * to verify it under page_lock.
951                          */
952                         if (unlikely(__PageMovable(page)) &&
953                                         !PageIsolated(page)) {
954                                 if (locked) {
955                                         unlock_page_lruvec_irqrestore(locked, flags);
956                                         locked = NULL;
957                                 }
958
959                                 if (!isolate_movable_page(page, isolate_mode))
960                                         goto isolate_success;
961                         }
962
963                         goto isolate_fail;
964                 }
965
966                 /*
967                  * Migration will fail if an anonymous page is pinned in memory,
968                  * so avoid taking lru_lock and isolating it unnecessarily in an
969                  * admittedly racy check.
970                  */
971                 if (!page_mapping(page) &&
972                     page_count(page) > page_mapcount(page))
973                         goto isolate_fail;
974
975                 /*
976                  * Only allow to migrate anonymous pages in GFP_NOFS context
977                  * because those do not depend on fs locks.
978                  */
979                 if (!(cc->gfp_mask & __GFP_FS) && page_mapping(page))
980                         goto isolate_fail;
981
982                 /*
983                  * Be careful not to clear PageLRU until after we're
984                  * sure the page is not being freed elsewhere -- the
985                  * page release code relies on it.
986                  */
987                 if (unlikely(!get_page_unless_zero(page)))
988                         goto isolate_fail;
989
990                 if (!__isolate_lru_page_prepare(page, isolate_mode))
991                         goto isolate_fail_put;
992
993                 /* Try isolate the page */
994                 if (!TestClearPageLRU(page))
995                         goto isolate_fail_put;
996
997                 lruvec = mem_cgroup_page_lruvec(page, pgdat);
998
999                 /* If we already hold the lock, we can skip some rechecking */
1000                 if (lruvec != locked) {
1001                         if (locked)
1002                                 unlock_page_lruvec_irqrestore(locked, flags);
1003
1004                         compact_lock_irqsave(&lruvec->lru_lock, &flags, cc);
1005                         locked = lruvec;
1006
1007                         lruvec_memcg_debug(lruvec, page);
1008
1009                         /* Try get exclusive access under lock */
1010                         if (!skip_updated) {
1011                                 skip_updated = true;
1012                                 if (test_and_set_skip(cc, page, low_pfn))
1013                                         goto isolate_abort;
1014                         }
1015
1016                         /*
1017                          * Page become compound since the non-locked check,
1018                          * and it's on LRU. It can only be a THP so the order
1019                          * is safe to read and it's 0 for tail pages.
1020                          */
1021                         if (unlikely(PageCompound(page) && !cc->alloc_contig)) {
1022                                 low_pfn += compound_nr(page) - 1;
1023                                 SetPageLRU(page);
1024                                 goto isolate_fail_put;
1025                         }
1026                 }
1027
1028                 /* The whole page is taken off the LRU; skip the tail pages. */
1029                 if (PageCompound(page))
1030                         low_pfn += compound_nr(page) - 1;
1031
1032                 /* Successfully isolated */
1033                 del_page_from_lru_list(page, lruvec);
1034                 mod_node_page_state(page_pgdat(page),
1035                                 NR_ISOLATED_ANON + page_is_file_lru(page),
1036                                 thp_nr_pages(page));
1037
1038 isolate_success:
1039                 list_add(&page->lru, &cc->migratepages);
1040                 cc->nr_migratepages += compound_nr(page);
1041                 nr_isolated += compound_nr(page);
1042
1043                 /*
1044                  * Avoid isolating too much unless this block is being
1045                  * rescanned (e.g. dirty/writeback pages, parallel allocation)
1046                  * or a lock is contended. For contention, isolate quickly to
1047                  * potentially remove one source of contention.
1048                  */
1049                 if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX &&
1050                     !cc->rescan && !cc->contended) {
1051                         ++low_pfn;
1052                         break;
1053                 }
1054
1055                 continue;
1056
1057 isolate_fail_put:
1058                 /* Avoid potential deadlock in freeing page under lru_lock */
1059                 if (locked) {
1060                         unlock_page_lruvec_irqrestore(locked, flags);
1061                         locked = NULL;
1062                 }
1063                 put_page(page);
1064
1065 isolate_fail:
1066                 if (!skip_on_failure)
1067                         continue;
1068
1069                 /*
1070                  * We have isolated some pages, but then failed. Release them
1071                  * instead of migrating, as we cannot form the cc->order buddy
1072                  * page anyway.
1073                  */
1074                 if (nr_isolated) {
1075                         if (locked) {
1076                                 unlock_page_lruvec_irqrestore(locked, flags);
1077                                 locked = NULL;
1078                         }
1079                         putback_movable_pages(&cc->migratepages);
1080                         cc->nr_migratepages = 0;
1081                         nr_isolated = 0;
1082                 }
1083
1084                 if (low_pfn < next_skip_pfn) {
1085                         low_pfn = next_skip_pfn - 1;
1086                         /*
1087                          * The check near the loop beginning would have updated
1088                          * next_skip_pfn too, but this is a bit simpler.
1089                          */
1090                         next_skip_pfn += 1UL << cc->order;
1091                 }
1092         }
1093
1094         /*
1095          * The PageBuddy() check could have potentially brought us outside
1096          * the range to be scanned.
1097          */
1098         if (unlikely(low_pfn > end_pfn))
1099                 low_pfn = end_pfn;
1100
1101         page = NULL;
1102
1103 isolate_abort:
1104         if (locked)
1105                 unlock_page_lruvec_irqrestore(locked, flags);
1106         if (page) {
1107                 SetPageLRU(page);
1108                 put_page(page);
1109         }
1110
1111         /*
1112          * Updated the cached scanner pfn once the pageblock has been scanned
1113          * Pages will either be migrated in which case there is no point
1114          * scanning in the near future or migration failed in which case the
1115          * failure reason may persist. The block is marked for skipping if
1116          * there were no pages isolated in the block or if the block is
1117          * rescanned twice in a row.
1118          */
1119         if (low_pfn == end_pfn && (!nr_isolated || cc->rescan)) {
1120                 if (valid_page && !skip_updated)
1121                         set_pageblock_skip(valid_page);
1122                 update_cached_migrate(cc, low_pfn);
1123         }
1124
1125         trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
1126                                                 nr_scanned, nr_isolated);
1127
1128 fatal_pending:
1129         cc->total_migrate_scanned += nr_scanned;
1130         if (nr_isolated)
1131                 count_compact_events(COMPACTISOLATED, nr_isolated);
1132
1133         return low_pfn;
1134 }
1135
1136 /**
1137  * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
1138  * @cc:        Compaction control structure.
1139  * @start_pfn: The first PFN to start isolating.
1140  * @end_pfn:   The one-past-last PFN.
1141  *
1142  * Returns zero if isolation fails fatally due to e.g. pending signal.
1143  * Otherwise, function returns one-past-the-last PFN of isolated page
1144  * (which may be greater than end_pfn if end fell in a middle of a THP page).
1145  */
1146 unsigned long
1147 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
1148                                                         unsigned long end_pfn)
1149 {
1150         unsigned long pfn, block_start_pfn, block_end_pfn;
1151
1152         /* Scan block by block. First and last block may be incomplete */
1153         pfn = start_pfn;
1154         block_start_pfn = pageblock_start_pfn(pfn);
1155         if (block_start_pfn < cc->zone->zone_start_pfn)
1156                 block_start_pfn = cc->zone->zone_start_pfn;
1157         block_end_pfn = pageblock_end_pfn(pfn);
1158
1159         for (; pfn < end_pfn; pfn = block_end_pfn,
1160                                 block_start_pfn = block_end_pfn,
1161                                 block_end_pfn += pageblock_nr_pages) {
1162
1163                 block_end_pfn = min(block_end_pfn, end_pfn);
1164
1165                 if (!pageblock_pfn_to_page(block_start_pfn,
1166                                         block_end_pfn, cc->zone))
1167                         continue;
1168
1169                 pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
1170                                                         ISOLATE_UNEVICTABLE);
1171
1172                 if (!pfn)
1173                         break;
1174
1175                 if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX)
1176                         break;
1177         }
1178
1179         return pfn;
1180 }
1181
1182 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
1183 #ifdef CONFIG_COMPACTION
1184
1185 static bool suitable_migration_source(struct compact_control *cc,
1186                                                         struct page *page)
1187 {
1188         int block_mt;
1189
1190         if (pageblock_skip_persistent(page))
1191                 return false;
1192
1193         if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction)
1194                 return true;
1195
1196         block_mt = get_pageblock_migratetype(page);
1197
1198         if (cc->migratetype == MIGRATE_MOVABLE)
1199                 return is_migrate_movable(block_mt);
1200         else
1201                 return block_mt == cc->migratetype;
1202 }
1203
1204 /* Returns true if the page is within a block suitable for migration to */
1205 static bool suitable_migration_target(struct compact_control *cc,
1206                                                         struct page *page)
1207 {
1208         /* If the page is a large free page, then disallow migration */
1209         if (PageBuddy(page)) {
1210                 /*
1211                  * We are checking page_order without zone->lock taken. But
1212                  * the only small danger is that we skip a potentially suitable
1213                  * pageblock, so it's not worth to check order for valid range.
1214                  */
1215                 if (buddy_order_unsafe(page) >= pageblock_order)
1216                         return false;
1217         }
1218
1219         if (cc->ignore_block_suitable)
1220                 return true;
1221
1222         /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1223         if (is_migrate_movable(get_pageblock_migratetype(page)))
1224                 return true;
1225
1226         /* Otherwise skip the block */
1227         return false;
1228 }
1229
1230 static inline unsigned int
1231 freelist_scan_limit(struct compact_control *cc)
1232 {
1233         unsigned short shift = BITS_PER_LONG - 1;
1234
1235         return (COMPACT_CLUSTER_MAX >> min(shift, cc->fast_search_fail)) + 1;
1236 }
1237
1238 /*
1239  * Test whether the free scanner has reached the same or lower pageblock than
1240  * the migration scanner, and compaction should thus terminate.
1241  */
1242 static inline bool compact_scanners_met(struct compact_control *cc)
1243 {
1244         return (cc->free_pfn >> pageblock_order)
1245                 <= (cc->migrate_pfn >> pageblock_order);
1246 }
1247
1248 /*
1249  * Used when scanning for a suitable migration target which scans freelists
1250  * in reverse. Reorders the list such as the unscanned pages are scanned
1251  * first on the next iteration of the free scanner
1252  */
1253 static void
1254 move_freelist_head(struct list_head *freelist, struct page *freepage)
1255 {
1256         LIST_HEAD(sublist);
1257
1258         if (!list_is_last(freelist, &freepage->lru)) {
1259                 list_cut_before(&sublist, freelist, &freepage->lru);
1260                 if (!list_empty(&sublist))
1261                         list_splice_tail(&sublist, freelist);
1262         }
1263 }
1264
1265 /*
1266  * Similar to move_freelist_head except used by the migration scanner
1267  * when scanning forward. It's possible for these list operations to
1268  * move against each other if they search the free list exactly in
1269  * lockstep.
1270  */
1271 static void
1272 move_freelist_tail(struct list_head *freelist, struct page *freepage)
1273 {
1274         LIST_HEAD(sublist);
1275
1276         if (!list_is_first(freelist, &freepage->lru)) {
1277                 list_cut_position(&sublist, freelist, &freepage->lru);
1278                 if (!list_empty(&sublist))
1279                         list_splice_tail(&sublist, freelist);
1280         }
1281 }
1282
1283 static void
1284 fast_isolate_around(struct compact_control *cc, unsigned long pfn, unsigned long nr_isolated)
1285 {
1286         unsigned long start_pfn, end_pfn;
1287         struct page *page;
1288
1289         /* Do not search around if there are enough pages already */
1290         if (cc->nr_freepages >= cc->nr_migratepages)
1291                 return;
1292
1293         /* Minimise scanning during async compaction */
1294         if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC)
1295                 return;
1296
1297         /* Pageblock boundaries */
1298         start_pfn = max(pageblock_start_pfn(pfn), cc->zone->zone_start_pfn);
1299         end_pfn = min(pageblock_end_pfn(pfn), zone_end_pfn(cc->zone));
1300
1301         page = pageblock_pfn_to_page(start_pfn, end_pfn, cc->zone);
1302         if (!page)
1303                 return;
1304
1305         /* Scan before */
1306         if (start_pfn != pfn) {
1307                 isolate_freepages_block(cc, &start_pfn, pfn, &cc->freepages, 1, false);
1308                 if (cc->nr_freepages >= cc->nr_migratepages)
1309                         return;
1310         }
1311
1312         /* Scan after */
1313         start_pfn = pfn + nr_isolated;
1314         if (start_pfn < end_pfn)
1315                 isolate_freepages_block(cc, &start_pfn, end_pfn, &cc->freepages, 1, false);
1316
1317         /* Skip this pageblock in the future as it's full or nearly full */
1318         if (cc->nr_freepages < cc->nr_migratepages)
1319                 set_pageblock_skip(page);
1320 }
1321
1322 /* Search orders in round-robin fashion */
1323 static int next_search_order(struct compact_control *cc, int order)
1324 {
1325         order--;
1326         if (order < 0)
1327                 order = cc->order - 1;
1328
1329         /* Search wrapped around? */
1330         if (order == cc->search_order) {
1331                 cc->search_order--;
1332                 if (cc->search_order < 0)
1333                         cc->search_order = cc->order - 1;
1334                 return -1;
1335         }
1336
1337         return order;
1338 }
1339
1340 static unsigned long
1341 fast_isolate_freepages(struct compact_control *cc)
1342 {
1343         unsigned int limit = min(1U, freelist_scan_limit(cc) >> 1);
1344         unsigned int nr_scanned = 0;
1345         unsigned long low_pfn, min_pfn, highest = 0;
1346         unsigned long nr_isolated = 0;
1347         unsigned long distance;
1348         struct page *page = NULL;
1349         bool scan_start = false;
1350         int order;
1351
1352         /* Full compaction passes in a negative order */
1353         if (cc->order <= 0)
1354                 return cc->free_pfn;
1355
1356         /*
1357          * If starting the scan, use a deeper search and use the highest
1358          * PFN found if a suitable one is not found.
1359          */
1360         if (cc->free_pfn >= cc->zone->compact_init_free_pfn) {
1361                 limit = pageblock_nr_pages >> 1;
1362                 scan_start = true;
1363         }
1364
1365         /*
1366          * Preferred point is in the top quarter of the scan space but take
1367          * a pfn from the top half if the search is problematic.
1368          */
1369         distance = (cc->free_pfn - cc->migrate_pfn);
1370         low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2));
1371         min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1));
1372
1373         if (WARN_ON_ONCE(min_pfn > low_pfn))
1374                 low_pfn = min_pfn;
1375
1376         /*
1377          * Search starts from the last successful isolation order or the next
1378          * order to search after a previous failure
1379          */
1380         cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order);
1381
1382         for (order = cc->search_order;
1383              !page && order >= 0;
1384              order = next_search_order(cc, order)) {
1385                 struct free_area *area = &cc->zone->free_area[order];
1386                 struct list_head *freelist;
1387                 struct page *freepage;
1388                 unsigned long flags;
1389                 unsigned int order_scanned = 0;
1390                 unsigned long high_pfn = 0;
1391
1392                 if (!area->nr_free)
1393                         continue;
1394
1395                 spin_lock_irqsave(&cc->zone->lock, flags);
1396                 freelist = &area->free_list[MIGRATE_MOVABLE];
1397                 list_for_each_entry_reverse(freepage, freelist, lru) {
1398                         unsigned long pfn;
1399
1400                         order_scanned++;
1401                         nr_scanned++;
1402                         pfn = page_to_pfn(freepage);
1403
1404                         if (pfn >= highest)
1405                                 highest = max(pageblock_start_pfn(pfn),
1406                                               cc->zone->zone_start_pfn);
1407
1408                         if (pfn >= low_pfn) {
1409                                 cc->fast_search_fail = 0;
1410                                 cc->search_order = order;
1411                                 page = freepage;
1412                                 break;
1413                         }
1414
1415                         if (pfn >= min_pfn && pfn > high_pfn) {
1416                                 high_pfn = pfn;
1417
1418                                 /* Shorten the scan if a candidate is found */
1419                                 limit >>= 1;
1420                         }
1421
1422                         if (order_scanned >= limit)
1423                                 break;
1424                 }
1425
1426                 /* Use a minimum pfn if a preferred one was not found */
1427                 if (!page && high_pfn) {
1428                         page = pfn_to_page(high_pfn);
1429
1430                         /* Update freepage for the list reorder below */
1431                         freepage = page;
1432                 }
1433
1434                 /* Reorder to so a future search skips recent pages */
1435                 move_freelist_head(freelist, freepage);
1436
1437                 /* Isolate the page if available */
1438                 if (page) {
1439                         if (__isolate_free_page(page, order)) {
1440                                 set_page_private(page, order);
1441                                 nr_isolated = 1 << order;
1442                                 cc->nr_freepages += nr_isolated;
1443                                 list_add_tail(&page->lru, &cc->freepages);
1444                                 count_compact_events(COMPACTISOLATED, nr_isolated);
1445                         } else {
1446                                 /* If isolation fails, abort the search */
1447                                 order = cc->search_order + 1;
1448                                 page = NULL;
1449                         }
1450                 }
1451
1452                 spin_unlock_irqrestore(&cc->zone->lock, flags);
1453
1454                 /*
1455                  * Smaller scan on next order so the total scan ig related
1456                  * to freelist_scan_limit.
1457                  */
1458                 if (order_scanned >= limit)
1459                         limit = min(1U, limit >> 1);
1460         }
1461
1462         if (!page) {
1463                 cc->fast_search_fail++;
1464                 if (scan_start) {
1465                         /*
1466                          * Use the highest PFN found above min. If one was
1467                          * not found, be pessimistic for direct compaction
1468                          * and use the min mark.
1469                          */
1470                         if (highest) {
1471                                 page = pfn_to_page(highest);
1472                                 cc->free_pfn = highest;
1473                         } else {
1474                                 if (cc->direct_compaction && pfn_valid(min_pfn)) {
1475                                         page = pageblock_pfn_to_page(min_pfn,
1476                                                 min(pageblock_end_pfn(min_pfn),
1477                                                     zone_end_pfn(cc->zone)),
1478                                                 cc->zone);
1479                                         cc->free_pfn = min_pfn;
1480                                 }
1481                         }
1482                 }
1483         }
1484
1485         if (highest && highest >= cc->zone->compact_cached_free_pfn) {
1486                 highest -= pageblock_nr_pages;
1487                 cc->zone->compact_cached_free_pfn = highest;
1488         }
1489
1490         cc->total_free_scanned += nr_scanned;
1491         if (!page)
1492                 return cc->free_pfn;
1493
1494         low_pfn = page_to_pfn(page);
1495         fast_isolate_around(cc, low_pfn, nr_isolated);
1496         return low_pfn;
1497 }
1498
1499 /*
1500  * Based on information in the current compact_control, find blocks
1501  * suitable for isolating free pages from and then isolate them.
1502  */
1503 static void isolate_freepages(struct compact_control *cc)
1504 {
1505         struct zone *zone = cc->zone;
1506         struct page *page;
1507         unsigned long block_start_pfn;  /* start of current pageblock */
1508         unsigned long isolate_start_pfn; /* exact pfn we start at */
1509         unsigned long block_end_pfn;    /* end of current pageblock */
1510         unsigned long low_pfn;       /* lowest pfn scanner is able to scan */
1511         struct list_head *freelist = &cc->freepages;
1512         unsigned int stride;
1513
1514         /* Try a small search of the free lists for a candidate */
1515         isolate_start_pfn = fast_isolate_freepages(cc);
1516         if (cc->nr_freepages)
1517                 goto splitmap;
1518
1519         /*
1520          * Initialise the free scanner. The starting point is where we last
1521          * successfully isolated from, zone-cached value, or the end of the
1522          * zone when isolating for the first time. For looping we also need
1523          * this pfn aligned down to the pageblock boundary, because we do
1524          * block_start_pfn -= pageblock_nr_pages in the for loop.
1525          * For ending point, take care when isolating in last pageblock of a
1526          * zone which ends in the middle of a pageblock.
1527          * The low boundary is the end of the pageblock the migration scanner
1528          * is using.
1529          */
1530         isolate_start_pfn = cc->free_pfn;
1531         block_start_pfn = pageblock_start_pfn(isolate_start_pfn);
1532         block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1533                                                 zone_end_pfn(zone));
1534         low_pfn = pageblock_end_pfn(cc->migrate_pfn);
1535         stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1;
1536
1537         /*
1538          * Isolate free pages until enough are available to migrate the
1539          * pages on cc->migratepages. We stop searching if the migrate
1540          * and free page scanners meet or enough free pages are isolated.
1541          */
1542         for (; block_start_pfn >= low_pfn;
1543                                 block_end_pfn = block_start_pfn,
1544                                 block_start_pfn -= pageblock_nr_pages,
1545                                 isolate_start_pfn = block_start_pfn) {
1546                 unsigned long nr_isolated;
1547
1548                 /*
1549                  * This can iterate a massively long zone without finding any
1550                  * suitable migration targets, so periodically check resched.
1551                  */
1552                 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)))
1553                         cond_resched();
1554
1555                 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1556                                                                         zone);
1557                 if (!page)
1558                         continue;
1559
1560                 /* Check the block is suitable for migration */
1561                 if (!suitable_migration_target(cc, page))
1562                         continue;
1563
1564                 /* If isolation recently failed, do not retry */
1565                 if (!isolation_suitable(cc, page))
1566                         continue;
1567
1568                 /* Found a block suitable for isolating free pages from. */
1569                 nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn,
1570                                         block_end_pfn, freelist, stride, false);
1571
1572                 /* Update the skip hint if the full pageblock was scanned */
1573                 if (isolate_start_pfn == block_end_pfn)
1574                         update_pageblock_skip(cc, page, block_start_pfn);
1575
1576                 /* Are enough freepages isolated? */
1577                 if (cc->nr_freepages >= cc->nr_migratepages) {
1578                         if (isolate_start_pfn >= block_end_pfn) {
1579                                 /*
1580                                  * Restart at previous pageblock if more
1581                                  * freepages can be isolated next time.
1582                                  */
1583                                 isolate_start_pfn =
1584                                         block_start_pfn - pageblock_nr_pages;
1585                         }
1586                         break;
1587                 } else if (isolate_start_pfn < block_end_pfn) {
1588                         /*
1589                          * If isolation failed early, do not continue
1590                          * needlessly.
1591                          */
1592                         break;
1593                 }
1594
1595                 /* Adjust stride depending on isolation */
1596                 if (nr_isolated) {
1597                         stride = 1;
1598                         continue;
1599                 }
1600                 stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1);
1601         }
1602
1603         /*
1604          * Record where the free scanner will restart next time. Either we
1605          * broke from the loop and set isolate_start_pfn based on the last
1606          * call to isolate_freepages_block(), or we met the migration scanner
1607          * and the loop terminated due to isolate_start_pfn < low_pfn
1608          */
1609         cc->free_pfn = isolate_start_pfn;
1610
1611 splitmap:
1612         /* __isolate_free_page() does not map the pages */
1613         split_map_pages(freelist);
1614 }
1615
1616 /*
1617  * This is a migrate-callback that "allocates" freepages by taking pages
1618  * from the isolated freelists in the block we are migrating to.
1619  */
1620 static struct page *compaction_alloc(struct page *migratepage,
1621                                         unsigned long data)
1622 {
1623         struct compact_control *cc = (struct compact_control *)data;
1624         struct page *freepage;
1625
1626         if (list_empty(&cc->freepages)) {
1627                 isolate_freepages(cc);
1628
1629                 if (list_empty(&cc->freepages))
1630                         return NULL;
1631         }
1632
1633         freepage = list_entry(cc->freepages.next, struct page, lru);
1634         list_del(&freepage->lru);
1635         cc->nr_freepages--;
1636
1637         return freepage;
1638 }
1639
1640 /*
1641  * This is a migrate-callback that "frees" freepages back to the isolated
1642  * freelist.  All pages on the freelist are from the same zone, so there is no
1643  * special handling needed for NUMA.
1644  */
1645 static void compaction_free(struct page *page, unsigned long data)
1646 {
1647         struct compact_control *cc = (struct compact_control *)data;
1648
1649         list_add(&page->lru, &cc->freepages);
1650         cc->nr_freepages++;
1651 }
1652
1653 /* possible outcome of isolate_migratepages */
1654 typedef enum {
1655         ISOLATE_ABORT,          /* Abort compaction now */
1656         ISOLATE_NONE,           /* No pages isolated, continue scanning */
1657         ISOLATE_SUCCESS,        /* Pages isolated, migrate */
1658 } isolate_migrate_t;
1659
1660 /*
1661  * Allow userspace to control policy on scanning the unevictable LRU for
1662  * compactable pages.
1663  */
1664 #ifdef CONFIG_PREEMPT_RT
1665 int sysctl_compact_unevictable_allowed __read_mostly = 0;
1666 #else
1667 int sysctl_compact_unevictable_allowed __read_mostly = 1;
1668 #endif
1669
1670 static inline void
1671 update_fast_start_pfn(struct compact_control *cc, unsigned long pfn)
1672 {
1673         if (cc->fast_start_pfn == ULONG_MAX)
1674                 return;
1675
1676         if (!cc->fast_start_pfn)
1677                 cc->fast_start_pfn = pfn;
1678
1679         cc->fast_start_pfn = min(cc->fast_start_pfn, pfn);
1680 }
1681
1682 static inline unsigned long
1683 reinit_migrate_pfn(struct compact_control *cc)
1684 {
1685         if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX)
1686                 return cc->migrate_pfn;
1687
1688         cc->migrate_pfn = cc->fast_start_pfn;
1689         cc->fast_start_pfn = ULONG_MAX;
1690
1691         return cc->migrate_pfn;
1692 }
1693
1694 /*
1695  * Briefly search the free lists for a migration source that already has
1696  * some free pages to reduce the number of pages that need migration
1697  * before a pageblock is free.
1698  */
1699 static unsigned long fast_find_migrateblock(struct compact_control *cc)
1700 {
1701         unsigned int limit = freelist_scan_limit(cc);
1702         unsigned int nr_scanned = 0;
1703         unsigned long distance;
1704         unsigned long pfn = cc->migrate_pfn;
1705         unsigned long high_pfn;
1706         int order;
1707         bool found_block = false;
1708
1709         /* Skip hints are relied on to avoid repeats on the fast search */
1710         if (cc->ignore_skip_hint)
1711                 return pfn;
1712
1713         /*
1714          * If the migrate_pfn is not at the start of a zone or the start
1715          * of a pageblock then assume this is a continuation of a previous
1716          * scan restarted due to COMPACT_CLUSTER_MAX.
1717          */
1718         if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn))
1719                 return pfn;
1720
1721         /*
1722          * For smaller orders, just linearly scan as the number of pages
1723          * to migrate should be relatively small and does not necessarily
1724          * justify freeing up a large block for a small allocation.
1725          */
1726         if (cc->order <= PAGE_ALLOC_COSTLY_ORDER)
1727                 return pfn;
1728
1729         /*
1730          * Only allow kcompactd and direct requests for movable pages to
1731          * quickly clear out a MOVABLE pageblock for allocation. This
1732          * reduces the risk that a large movable pageblock is freed for
1733          * an unmovable/reclaimable small allocation.
1734          */
1735         if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE)
1736                 return pfn;
1737
1738         /*
1739          * When starting the migration scanner, pick any pageblock within the
1740          * first half of the search space. Otherwise try and pick a pageblock
1741          * within the first eighth to reduce the chances that a migration
1742          * target later becomes a source.
1743          */
1744         distance = (cc->free_pfn - cc->migrate_pfn) >> 1;
1745         if (cc->migrate_pfn != cc->zone->zone_start_pfn)
1746                 distance >>= 2;
1747         high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance);
1748
1749         for (order = cc->order - 1;
1750              order >= PAGE_ALLOC_COSTLY_ORDER && !found_block && nr_scanned < limit;
1751              order--) {
1752                 struct free_area *area = &cc->zone->free_area[order];
1753                 struct list_head *freelist;
1754                 unsigned long flags;
1755                 struct page *freepage;
1756
1757                 if (!area->nr_free)
1758                         continue;
1759
1760                 spin_lock_irqsave(&cc->zone->lock, flags);
1761                 freelist = &area->free_list[MIGRATE_MOVABLE];
1762                 list_for_each_entry(freepage, freelist, lru) {
1763                         unsigned long free_pfn;
1764
1765                         if (nr_scanned++ >= limit) {
1766                                 move_freelist_tail(freelist, freepage);
1767                                 break;
1768                         }
1769
1770                         free_pfn = page_to_pfn(freepage);
1771                         if (free_pfn < high_pfn) {
1772                                 /*
1773                                  * Avoid if skipped recently. Ideally it would
1774                                  * move to the tail but even safe iteration of
1775                                  * the list assumes an entry is deleted, not
1776                                  * reordered.
1777                                  */
1778                                 if (get_pageblock_skip(freepage))
1779                                         continue;
1780
1781                                 /* Reorder to so a future search skips recent pages */
1782                                 move_freelist_tail(freelist, freepage);
1783
1784                                 update_fast_start_pfn(cc, free_pfn);
1785                                 pfn = pageblock_start_pfn(free_pfn);
1786                                 cc->fast_search_fail = 0;
1787                                 found_block = true;
1788                                 set_pageblock_skip(freepage);
1789                                 break;
1790                         }
1791                 }
1792                 spin_unlock_irqrestore(&cc->zone->lock, flags);
1793         }
1794
1795         cc->total_migrate_scanned += nr_scanned;
1796
1797         /*
1798          * If fast scanning failed then use a cached entry for a page block
1799          * that had free pages as the basis for starting a linear scan.
1800          */
1801         if (!found_block) {
1802                 cc->fast_search_fail++;
1803                 pfn = reinit_migrate_pfn(cc);
1804         }
1805         return pfn;
1806 }
1807
1808 /*
1809  * Isolate all pages that can be migrated from the first suitable block,
1810  * starting at the block pointed to by the migrate scanner pfn within
1811  * compact_control.
1812  */
1813 static isolate_migrate_t isolate_migratepages(struct compact_control *cc)
1814 {
1815         unsigned long block_start_pfn;
1816         unsigned long block_end_pfn;
1817         unsigned long low_pfn;
1818         struct page *page;
1819         const isolate_mode_t isolate_mode =
1820                 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1821                 (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1822         bool fast_find_block;
1823
1824         /*
1825          * Start at where we last stopped, or beginning of the zone as
1826          * initialized by compact_zone(). The first failure will use
1827          * the lowest PFN as the starting point for linear scanning.
1828          */
1829         low_pfn = fast_find_migrateblock(cc);
1830         block_start_pfn = pageblock_start_pfn(low_pfn);
1831         if (block_start_pfn < cc->zone->zone_start_pfn)
1832                 block_start_pfn = cc->zone->zone_start_pfn;
1833
1834         /*
1835          * fast_find_migrateblock marks a pageblock skipped so to avoid
1836          * the isolation_suitable check below, check whether the fast
1837          * search was successful.
1838          */
1839         fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail;
1840
1841         /* Only scan within a pageblock boundary */
1842         block_end_pfn = pageblock_end_pfn(low_pfn);
1843
1844         /*
1845          * Iterate over whole pageblocks until we find the first suitable.
1846          * Do not cross the free scanner.
1847          */
1848         for (; block_end_pfn <= cc->free_pfn;
1849                         fast_find_block = false,
1850                         low_pfn = block_end_pfn,
1851                         block_start_pfn = block_end_pfn,
1852                         block_end_pfn += pageblock_nr_pages) {
1853
1854                 /*
1855                  * This can potentially iterate a massively long zone with
1856                  * many pageblocks unsuitable, so periodically check if we
1857                  * need to schedule.
1858                  */
1859                 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)))
1860                         cond_resched();
1861
1862                 page = pageblock_pfn_to_page(block_start_pfn,
1863                                                 block_end_pfn, cc->zone);
1864                 if (!page)
1865                         continue;
1866
1867                 /*
1868                  * If isolation recently failed, do not retry. Only check the
1869                  * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock
1870                  * to be visited multiple times. Assume skip was checked
1871                  * before making it "skip" so other compaction instances do
1872                  * not scan the same block.
1873                  */
1874                 if (IS_ALIGNED(low_pfn, pageblock_nr_pages) &&
1875                     !fast_find_block && !isolation_suitable(cc, page))
1876                         continue;
1877
1878                 /*
1879                  * For async compaction, also only scan in MOVABLE blocks
1880                  * without huge pages. Async compaction is optimistic to see
1881                  * if the minimum amount of work satisfies the allocation.
1882                  * The cached PFN is updated as it's possible that all
1883                  * remaining blocks between source and target are unsuitable
1884                  * and the compaction scanners fail to meet.
1885                  */
1886                 if (!suitable_migration_source(cc, page)) {
1887                         update_cached_migrate(cc, block_end_pfn);
1888                         continue;
1889                 }
1890
1891                 /* Perform the isolation */
1892                 low_pfn = isolate_migratepages_block(cc, low_pfn,
1893                                                 block_end_pfn, isolate_mode);
1894
1895                 if (!low_pfn)
1896                         return ISOLATE_ABORT;
1897
1898                 /*
1899                  * Either we isolated something and proceed with migration. Or
1900                  * we failed and compact_zone should decide if we should
1901                  * continue or not.
1902                  */
1903                 break;
1904         }
1905
1906         /* Record where migration scanner will be restarted. */
1907         cc->migrate_pfn = low_pfn;
1908
1909         return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
1910 }
1911
1912 /*
1913  * order == -1 is expected when compacting via
1914  * /proc/sys/vm/compact_memory
1915  */
1916 static inline bool is_via_compact_memory(int order)
1917 {
1918         return order == -1;
1919 }
1920
1921 static bool kswapd_is_running(pg_data_t *pgdat)
1922 {
1923         return pgdat->kswapd && (pgdat->kswapd->state == TASK_RUNNING);
1924 }
1925
1926 /*
1927  * A zone's fragmentation score is the external fragmentation wrt to the
1928  * COMPACTION_HPAGE_ORDER. It returns a value in the range [0, 100].
1929  */
1930 static unsigned int fragmentation_score_zone(struct zone *zone)
1931 {
1932         return extfrag_for_order(zone, COMPACTION_HPAGE_ORDER);
1933 }
1934
1935 /*
1936  * A weighted zone's fragmentation score is the external fragmentation
1937  * wrt to the COMPACTION_HPAGE_ORDER scaled by the zone's size. It
1938  * returns a value in the range [0, 100].
1939  *
1940  * The scaling factor ensures that proactive compaction focuses on larger
1941  * zones like ZONE_NORMAL, rather than smaller, specialized zones like
1942  * ZONE_DMA32. For smaller zones, the score value remains close to zero,
1943  * and thus never exceeds the high threshold for proactive compaction.
1944  */
1945 static unsigned int fragmentation_score_zone_weighted(struct zone *zone)
1946 {
1947         unsigned long score;
1948
1949         score = zone->present_pages * fragmentation_score_zone(zone);
1950         return div64_ul(score, zone->zone_pgdat->node_present_pages + 1);
1951 }
1952
1953 /*
1954  * The per-node proactive (background) compaction process is started by its
1955  * corresponding kcompactd thread when the node's fragmentation score
1956  * exceeds the high threshold. The compaction process remains active till
1957  * the node's score falls below the low threshold, or one of the back-off
1958  * conditions is met.
1959  */
1960 static unsigned int fragmentation_score_node(pg_data_t *pgdat)
1961 {
1962         unsigned int score = 0;
1963         int zoneid;
1964
1965         for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1966                 struct zone *zone;
1967
1968                 zone = &pgdat->node_zones[zoneid];
1969                 score += fragmentation_score_zone_weighted(zone);
1970         }
1971
1972         return score;
1973 }
1974
1975 static unsigned int fragmentation_score_wmark(pg_data_t *pgdat, bool low)
1976 {
1977         unsigned int wmark_low;
1978
1979         /*
1980          * Cap the low watermak to avoid excessive compaction
1981          * activity in case a user sets the proactivess tunable
1982          * close to 100 (maximum).
1983          */
1984         wmark_low = max(100U - sysctl_compaction_proactiveness, 5U);
1985         return low ? wmark_low : min(wmark_low + 10, 100U);
1986 }
1987
1988 static bool should_proactive_compact_node(pg_data_t *pgdat)
1989 {
1990         int wmark_high;
1991
1992         if (!sysctl_compaction_proactiveness || kswapd_is_running(pgdat))
1993                 return false;
1994
1995         wmark_high = fragmentation_score_wmark(pgdat, false);
1996         return fragmentation_score_node(pgdat) > wmark_high;
1997 }
1998
1999 static enum compact_result __compact_finished(struct compact_control *cc)
2000 {
2001         unsigned int order;
2002         const int migratetype = cc->migratetype;
2003         int ret;
2004
2005         /* Compaction run completes if the migrate and free scanner meet */
2006         if (compact_scanners_met(cc)) {
2007                 /* Let the next compaction start anew. */
2008                 reset_cached_positions(cc->zone);
2009
2010                 /*
2011                  * Mark that the PG_migrate_skip information should be cleared
2012                  * by kswapd when it goes to sleep. kcompactd does not set the
2013                  * flag itself as the decision to be clear should be directly
2014                  * based on an allocation request.
2015                  */
2016                 if (cc->direct_compaction)
2017                         cc->zone->compact_blockskip_flush = true;
2018
2019                 if (cc->whole_zone)
2020                         return COMPACT_COMPLETE;
2021                 else
2022                         return COMPACT_PARTIAL_SKIPPED;
2023         }
2024
2025         if (cc->proactive_compaction) {
2026                 int score, wmark_low;
2027                 pg_data_t *pgdat;
2028
2029                 pgdat = cc->zone->zone_pgdat;
2030                 if (kswapd_is_running(pgdat))
2031                         return COMPACT_PARTIAL_SKIPPED;
2032
2033                 score = fragmentation_score_zone(cc->zone);
2034                 wmark_low = fragmentation_score_wmark(pgdat, true);
2035
2036                 if (score > wmark_low)
2037                         ret = COMPACT_CONTINUE;
2038                 else
2039                         ret = COMPACT_SUCCESS;
2040
2041                 goto out;
2042         }
2043
2044         if (is_via_compact_memory(cc->order))
2045                 return COMPACT_CONTINUE;
2046
2047         /*
2048          * Always finish scanning a pageblock to reduce the possibility of
2049          * fallbacks in the future. This is particularly important when
2050          * migration source is unmovable/reclaimable but it's not worth
2051          * special casing.
2052          */
2053         if (!IS_ALIGNED(cc->migrate_pfn, pageblock_nr_pages))
2054                 return COMPACT_CONTINUE;
2055
2056         /* Direct compactor: Is a suitable page free? */
2057         ret = COMPACT_NO_SUITABLE_PAGE;
2058         for (order = cc->order; order < MAX_ORDER; order++) {
2059                 struct free_area *area = &cc->zone->free_area[order];
2060                 bool can_steal;
2061
2062                 /* Job done if page is free of the right migratetype */
2063                 if (!free_area_empty(area, migratetype))
2064                         return COMPACT_SUCCESS;
2065
2066 #ifdef CONFIG_CMA
2067                 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
2068                 if (migratetype == MIGRATE_MOVABLE &&
2069                         !free_area_empty(area, MIGRATE_CMA))
2070                         return COMPACT_SUCCESS;
2071 #endif
2072                 /*
2073                  * Job done if allocation would steal freepages from
2074                  * other migratetype buddy lists.
2075                  */
2076                 if (find_suitable_fallback(area, order, migratetype,
2077                                                 true, &can_steal) != -1) {
2078
2079                         /* movable pages are OK in any pageblock */
2080                         if (migratetype == MIGRATE_MOVABLE)
2081                                 return COMPACT_SUCCESS;
2082
2083                         /*
2084                          * We are stealing for a non-movable allocation. Make
2085                          * sure we finish compacting the current pageblock
2086                          * first so it is as free as possible and we won't
2087                          * have to steal another one soon. This only applies
2088                          * to sync compaction, as async compaction operates
2089                          * on pageblocks of the same migratetype.
2090                          */
2091                         if (cc->mode == MIGRATE_ASYNC ||
2092                                         IS_ALIGNED(cc->migrate_pfn,
2093                                                         pageblock_nr_pages)) {
2094                                 return COMPACT_SUCCESS;
2095                         }
2096
2097                         ret = COMPACT_CONTINUE;
2098                         break;
2099                 }
2100         }
2101
2102 out:
2103         if (cc->contended || fatal_signal_pending(current))
2104                 ret = COMPACT_CONTENDED;
2105
2106         return ret;
2107 }
2108
2109 static enum compact_result compact_finished(struct compact_control *cc)
2110 {
2111         int ret;
2112
2113         ret = __compact_finished(cc);
2114         trace_mm_compaction_finished(cc->zone, cc->order, ret);
2115         if (ret == COMPACT_NO_SUITABLE_PAGE)
2116                 ret = COMPACT_CONTINUE;
2117
2118         return ret;
2119 }
2120
2121 static enum compact_result __compaction_suitable(struct zone *zone, int order,
2122                                         unsigned int alloc_flags,
2123                                         int highest_zoneidx,
2124                                         unsigned long wmark_target)
2125 {
2126         unsigned long watermark;
2127
2128         if (is_via_compact_memory(order))
2129                 return COMPACT_CONTINUE;
2130
2131         watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
2132         /*
2133          * If watermarks for high-order allocation are already met, there
2134          * should be no need for compaction at all.
2135          */
2136         if (zone_watermark_ok(zone, order, watermark, highest_zoneidx,
2137                                                                 alloc_flags))
2138                 return COMPACT_SUCCESS;
2139
2140         /*
2141          * Watermarks for order-0 must be met for compaction to be able to
2142          * isolate free pages for migration targets. This means that the
2143          * watermark and alloc_flags have to match, or be more pessimistic than
2144          * the check in __isolate_free_page(). We don't use the direct
2145          * compactor's alloc_flags, as they are not relevant for freepage
2146          * isolation. We however do use the direct compactor's highest_zoneidx
2147          * to skip over zones where lowmem reserves would prevent allocation
2148          * even if compaction succeeds.
2149          * For costly orders, we require low watermark instead of min for
2150          * compaction to proceed to increase its chances.
2151          * ALLOC_CMA is used, as pages in CMA pageblocks are considered
2152          * suitable migration targets
2153          */
2154         watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
2155                                 low_wmark_pages(zone) : min_wmark_pages(zone);
2156         watermark += compact_gap(order);
2157         if (!__zone_watermark_ok(zone, 0, watermark, highest_zoneidx,
2158                                                 ALLOC_CMA, wmark_target))
2159                 return COMPACT_SKIPPED;
2160
2161         return COMPACT_CONTINUE;
2162 }
2163
2164 /*
2165  * compaction_suitable: Is this suitable to run compaction on this zone now?
2166  * Returns
2167  *   COMPACT_SKIPPED  - If there are too few free pages for compaction
2168  *   COMPACT_SUCCESS  - If the allocation would succeed without compaction
2169  *   COMPACT_CONTINUE - If compaction should run now
2170  */
2171 enum compact_result compaction_suitable(struct zone *zone, int order,
2172                                         unsigned int alloc_flags,
2173                                         int highest_zoneidx)
2174 {
2175         enum compact_result ret;
2176         int fragindex;
2177
2178         ret = __compaction_suitable(zone, order, alloc_flags, highest_zoneidx,
2179                                     zone_page_state(zone, NR_FREE_PAGES));
2180         /*
2181          * fragmentation index determines if allocation failures are due to
2182          * low memory or external fragmentation
2183          *
2184          * index of -1000 would imply allocations might succeed depending on
2185          * watermarks, but we already failed the high-order watermark check
2186          * index towards 0 implies failure is due to lack of memory
2187          * index towards 1000 implies failure is due to fragmentation
2188          *
2189          * Only compact if a failure would be due to fragmentation. Also
2190          * ignore fragindex for non-costly orders where the alternative to
2191          * a successful reclaim/compaction is OOM. Fragindex and the
2192          * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
2193          * excessive compaction for costly orders, but it should not be at the
2194          * expense of system stability.
2195          */
2196         if (ret == COMPACT_CONTINUE && (order > PAGE_ALLOC_COSTLY_ORDER)) {
2197                 fragindex = fragmentation_index(zone, order);
2198                 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
2199                         ret = COMPACT_NOT_SUITABLE_ZONE;
2200         }
2201
2202         trace_mm_compaction_suitable(zone, order, ret);
2203         if (ret == COMPACT_NOT_SUITABLE_ZONE)
2204                 ret = COMPACT_SKIPPED;
2205
2206         return ret;
2207 }
2208
2209 bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
2210                 int alloc_flags)
2211 {
2212         struct zone *zone;
2213         struct zoneref *z;
2214
2215         /*
2216          * Make sure at least one zone would pass __compaction_suitable if we continue
2217          * retrying the reclaim.
2218          */
2219         for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2220                                 ac->highest_zoneidx, ac->nodemask) {
2221                 unsigned long available;
2222                 enum compact_result compact_result;
2223
2224                 /*
2225                  * Do not consider all the reclaimable memory because we do not
2226                  * want to trash just for a single high order allocation which
2227                  * is even not guaranteed to appear even if __compaction_suitable
2228                  * is happy about the watermark check.
2229                  */
2230                 available = zone_reclaimable_pages(zone) / order;
2231                 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
2232                 compact_result = __compaction_suitable(zone, order, alloc_flags,
2233                                 ac->highest_zoneidx, available);
2234                 if (compact_result != COMPACT_SKIPPED)
2235                         return true;
2236         }
2237
2238         return false;
2239 }
2240
2241 static enum compact_result
2242 compact_zone(struct compact_control *cc, struct capture_control *capc)
2243 {
2244         enum compact_result ret;
2245         unsigned long start_pfn = cc->zone->zone_start_pfn;
2246         unsigned long end_pfn = zone_end_pfn(cc->zone);
2247         unsigned long last_migrated_pfn;
2248         const bool sync = cc->mode != MIGRATE_ASYNC;
2249         bool update_cached;
2250
2251         /*
2252          * These counters track activities during zone compaction.  Initialize
2253          * them before compacting a new zone.
2254          */
2255         cc->total_migrate_scanned = 0;
2256         cc->total_free_scanned = 0;
2257         cc->nr_migratepages = 0;
2258         cc->nr_freepages = 0;
2259         INIT_LIST_HEAD(&cc->freepages);
2260         INIT_LIST_HEAD(&cc->migratepages);
2261
2262         cc->migratetype = gfp_migratetype(cc->gfp_mask);
2263         ret = compaction_suitable(cc->zone, cc->order, cc->alloc_flags,
2264                                                         cc->highest_zoneidx);
2265         /* Compaction is likely to fail */
2266         if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED)
2267                 return ret;
2268
2269         /* huh, compaction_suitable is returning something unexpected */
2270         VM_BUG_ON(ret != COMPACT_CONTINUE);
2271
2272         /*
2273          * Clear pageblock skip if there were failures recently and compaction
2274          * is about to be retried after being deferred.
2275          */
2276         if (compaction_restarting(cc->zone, cc->order))
2277                 __reset_isolation_suitable(cc->zone);
2278
2279         /*
2280          * Setup to move all movable pages to the end of the zone. Used cached
2281          * information on where the scanners should start (unless we explicitly
2282          * want to compact the whole zone), but check that it is initialised
2283          * by ensuring the values are within zone boundaries.
2284          */
2285         cc->fast_start_pfn = 0;
2286         if (cc->whole_zone) {
2287                 cc->migrate_pfn = start_pfn;
2288                 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2289         } else {
2290                 cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync];
2291                 cc->free_pfn = cc->zone->compact_cached_free_pfn;
2292                 if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
2293                         cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2294                         cc->zone->compact_cached_free_pfn = cc->free_pfn;
2295                 }
2296                 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
2297                         cc->migrate_pfn = start_pfn;
2298                         cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
2299                         cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
2300                 }
2301
2302                 if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn)
2303                         cc->whole_zone = true;
2304         }
2305
2306         last_migrated_pfn = 0;
2307
2308         /*
2309          * Migrate has separate cached PFNs for ASYNC and SYNC* migration on
2310          * the basis that some migrations will fail in ASYNC mode. However,
2311          * if the cached PFNs match and pageblocks are skipped due to having
2312          * no isolation candidates, then the sync state does not matter.
2313          * Until a pageblock with isolation candidates is found, keep the
2314          * cached PFNs in sync to avoid revisiting the same blocks.
2315          */
2316         update_cached = !sync &&
2317                 cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1];
2318
2319         trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
2320                                 cc->free_pfn, end_pfn, sync);
2321
2322         migrate_prep_local();
2323
2324         while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) {
2325                 int err;
2326                 unsigned long iteration_start_pfn = cc->migrate_pfn;
2327
2328                 /*
2329                  * Avoid multiple rescans which can happen if a page cannot be
2330                  * isolated (dirty/writeback in async mode) or if the migrated
2331                  * pages are being allocated before the pageblock is cleared.
2332                  * The first rescan will capture the entire pageblock for
2333                  * migration. If it fails, it'll be marked skip and scanning
2334                  * will proceed as normal.
2335                  */
2336                 cc->rescan = false;
2337                 if (pageblock_start_pfn(last_migrated_pfn) ==
2338                     pageblock_start_pfn(iteration_start_pfn)) {
2339                         cc->rescan = true;
2340                 }
2341
2342                 switch (isolate_migratepages(cc)) {
2343                 case ISOLATE_ABORT:
2344                         ret = COMPACT_CONTENDED;
2345                         putback_movable_pages(&cc->migratepages);
2346                         cc->nr_migratepages = 0;
2347                         goto out;
2348                 case ISOLATE_NONE:
2349                         if (update_cached) {
2350                                 cc->zone->compact_cached_migrate_pfn[1] =
2351                                         cc->zone->compact_cached_migrate_pfn[0];
2352                         }
2353
2354                         /*
2355                          * We haven't isolated and migrated anything, but
2356                          * there might still be unflushed migrations from
2357                          * previous cc->order aligned block.
2358                          */
2359                         goto check_drain;
2360                 case ISOLATE_SUCCESS:
2361                         update_cached = false;
2362                         last_migrated_pfn = iteration_start_pfn;
2363                 }
2364
2365                 err = migrate_pages(&cc->migratepages, compaction_alloc,
2366                                 compaction_free, (unsigned long)cc, cc->mode,
2367                                 MR_COMPACTION);
2368
2369                 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
2370                                                         &cc->migratepages);
2371
2372                 /* All pages were either migrated or will be released */
2373                 cc->nr_migratepages = 0;
2374                 if (err) {
2375                         putback_movable_pages(&cc->migratepages);
2376                         /*
2377                          * migrate_pages() may return -ENOMEM when scanners meet
2378                          * and we want compact_finished() to detect it
2379                          */
2380                         if (err == -ENOMEM && !compact_scanners_met(cc)) {
2381                                 ret = COMPACT_CONTENDED;
2382                                 goto out;
2383                         }
2384                         /*
2385                          * We failed to migrate at least one page in the current
2386                          * order-aligned block, so skip the rest of it.
2387                          */
2388                         if (cc->direct_compaction &&
2389                                                 (cc->mode == MIGRATE_ASYNC)) {
2390                                 cc->migrate_pfn = block_end_pfn(
2391                                                 cc->migrate_pfn - 1, cc->order);
2392                                 /* Draining pcplists is useless in this case */
2393                                 last_migrated_pfn = 0;
2394                         }
2395                 }
2396
2397 check_drain:
2398                 /*
2399                  * Has the migration scanner moved away from the previous
2400                  * cc->order aligned block where we migrated from? If yes,
2401                  * flush the pages that were freed, so that they can merge and
2402                  * compact_finished() can detect immediately if allocation
2403                  * would succeed.
2404                  */
2405                 if (cc->order > 0 && last_migrated_pfn) {
2406                         unsigned long current_block_start =
2407                                 block_start_pfn(cc->migrate_pfn, cc->order);
2408
2409                         if (last_migrated_pfn < current_block_start) {
2410                                 lru_add_drain_cpu_zone(cc->zone);
2411                                 /* No more flushing until we migrate again */
2412                                 last_migrated_pfn = 0;
2413                         }
2414                 }
2415
2416                 /* Stop if a page has been captured */
2417                 if (capc && capc->page) {
2418                         ret = COMPACT_SUCCESS;
2419                         break;
2420                 }
2421         }
2422
2423 out:
2424         /*
2425          * Release free pages and update where the free scanner should restart,
2426          * so we don't leave any returned pages behind in the next attempt.
2427          */
2428         if (cc->nr_freepages > 0) {
2429                 unsigned long free_pfn = release_freepages(&cc->freepages);
2430
2431                 cc->nr_freepages = 0;
2432                 VM_BUG_ON(free_pfn == 0);
2433                 /* The cached pfn is always the first in a pageblock */
2434                 free_pfn = pageblock_start_pfn(free_pfn);
2435                 /*
2436                  * Only go back, not forward. The cached pfn might have been
2437                  * already reset to zone end in compact_finished()
2438                  */
2439                 if (free_pfn > cc->zone->compact_cached_free_pfn)
2440                         cc->zone->compact_cached_free_pfn = free_pfn;
2441         }
2442
2443         count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned);
2444         count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned);
2445
2446         trace_mm_compaction_end(start_pfn, cc->migrate_pfn,
2447                                 cc->free_pfn, end_pfn, sync, ret);
2448
2449         return ret;
2450 }
2451
2452 static enum compact_result compact_zone_order(struct zone *zone, int order,
2453                 gfp_t gfp_mask, enum compact_priority prio,
2454                 unsigned int alloc_flags, int highest_zoneidx,
2455                 struct page **capture)
2456 {
2457         enum compact_result ret;
2458         struct compact_control cc = {
2459                 .order = order,
2460                 .search_order = order,
2461                 .gfp_mask = gfp_mask,
2462                 .zone = zone,
2463                 .mode = (prio == COMPACT_PRIO_ASYNC) ?
2464                                         MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT,
2465                 .alloc_flags = alloc_flags,
2466                 .highest_zoneidx = highest_zoneidx,
2467                 .direct_compaction = true,
2468                 .whole_zone = (prio == MIN_COMPACT_PRIORITY),
2469                 .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
2470                 .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
2471         };
2472         struct capture_control capc = {
2473                 .cc = &cc,
2474                 .page = NULL,
2475         };
2476
2477         /*
2478          * Make sure the structs are really initialized before we expose the
2479          * capture control, in case we are interrupted and the interrupt handler
2480          * frees a page.
2481          */
2482         barrier();
2483         WRITE_ONCE(current->capture_control, &capc);
2484
2485         ret = compact_zone(&cc, &capc);
2486
2487         VM_BUG_ON(!list_empty(&cc.freepages));
2488         VM_BUG_ON(!list_empty(&cc.migratepages));
2489
2490         /*
2491          * Make sure we hide capture control first before we read the captured
2492          * page pointer, otherwise an interrupt could free and capture a page
2493          * and we would leak it.
2494          */
2495         WRITE_ONCE(current->capture_control, NULL);
2496         *capture = READ_ONCE(capc.page);
2497
2498         return ret;
2499 }
2500
2501 int sysctl_extfrag_threshold = 500;
2502
2503 /**
2504  * try_to_compact_pages - Direct compact to satisfy a high-order allocation
2505  * @gfp_mask: The GFP mask of the current allocation
2506  * @order: The order of the current allocation
2507  * @alloc_flags: The allocation flags of the current allocation
2508  * @ac: The context of current allocation
2509  * @prio: Determines how hard direct compaction should try to succeed
2510  * @capture: Pointer to free page created by compaction will be stored here
2511  *
2512  * This is the main entry point for direct page compaction.
2513  */
2514 enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
2515                 unsigned int alloc_flags, const struct alloc_context *ac,
2516                 enum compact_priority prio, struct page **capture)
2517 {
2518         int may_perform_io = gfp_mask & __GFP_IO;
2519         struct zoneref *z;
2520         struct zone *zone;
2521         enum compact_result rc = COMPACT_SKIPPED;
2522
2523         /*
2524          * Check if the GFP flags allow compaction - GFP_NOIO is really
2525          * tricky context because the migration might require IO
2526          */
2527         if (!may_perform_io)
2528                 return COMPACT_SKIPPED;
2529
2530         trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
2531
2532         /* Compact each zone in the list */
2533         for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2534                                         ac->highest_zoneidx, ac->nodemask) {
2535                 enum compact_result status;
2536
2537                 if (prio > MIN_COMPACT_PRIORITY
2538                                         && compaction_deferred(zone, order)) {
2539                         rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
2540                         continue;
2541                 }
2542
2543                 status = compact_zone_order(zone, order, gfp_mask, prio,
2544                                 alloc_flags, ac->highest_zoneidx, capture);
2545                 rc = max(status, rc);
2546
2547                 /* The allocation should succeed, stop compacting */
2548                 if (status == COMPACT_SUCCESS) {
2549                         /*
2550                          * We think the allocation will succeed in this zone,
2551                          * but it is not certain, hence the false. The caller
2552                          * will repeat this with true if allocation indeed
2553                          * succeeds in this zone.
2554                          */
2555                         compaction_defer_reset(zone, order, false);
2556
2557                         break;
2558                 }
2559
2560                 if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
2561                                         status == COMPACT_PARTIAL_SKIPPED))
2562                         /*
2563                          * We think that allocation won't succeed in this zone
2564                          * so we defer compaction there. If it ends up
2565                          * succeeding after all, it will be reset.
2566                          */
2567                         defer_compaction(zone, order);
2568
2569                 /*
2570                  * We might have stopped compacting due to need_resched() in
2571                  * async compaction, or due to a fatal signal detected. In that
2572                  * case do not try further zones
2573                  */
2574                 if ((prio == COMPACT_PRIO_ASYNC && need_resched())
2575                                         || fatal_signal_pending(current))
2576                         break;
2577         }
2578
2579         return rc;
2580 }
2581
2582 /*
2583  * Compact all zones within a node till each zone's fragmentation score
2584  * reaches within proactive compaction thresholds (as determined by the
2585  * proactiveness tunable).
2586  *
2587  * It is possible that the function returns before reaching score targets
2588  * due to various back-off conditions, such as, contention on per-node or
2589  * per-zone locks.
2590  */
2591 static void proactive_compact_node(pg_data_t *pgdat)
2592 {
2593         int zoneid;
2594         struct zone *zone;
2595         struct compact_control cc = {
2596                 .order = -1,
2597                 .mode = MIGRATE_SYNC_LIGHT,
2598                 .ignore_skip_hint = true,
2599                 .whole_zone = true,
2600                 .gfp_mask = GFP_KERNEL,
2601                 .proactive_compaction = true,
2602         };
2603
2604         for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2605                 zone = &pgdat->node_zones[zoneid];
2606                 if (!populated_zone(zone))
2607                         continue;
2608
2609                 cc.zone = zone;
2610
2611                 compact_zone(&cc, NULL);
2612
2613                 VM_BUG_ON(!list_empty(&cc.freepages));
2614                 VM_BUG_ON(!list_empty(&cc.migratepages));
2615         }
2616 }
2617
2618 /* Compact all zones within a node */
2619 static void compact_node(int nid)
2620 {
2621         pg_data_t *pgdat = NODE_DATA(nid);
2622         int zoneid;
2623         struct zone *zone;
2624         struct compact_control cc = {
2625                 .order = -1,
2626                 .mode = MIGRATE_SYNC,
2627                 .ignore_skip_hint = true,
2628                 .whole_zone = true,
2629                 .gfp_mask = GFP_KERNEL,
2630         };
2631
2632
2633         for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2634
2635                 zone = &pgdat->node_zones[zoneid];
2636                 if (!populated_zone(zone))
2637                         continue;
2638
2639                 cc.zone = zone;
2640
2641                 compact_zone(&cc, NULL);
2642
2643                 VM_BUG_ON(!list_empty(&cc.freepages));
2644                 VM_BUG_ON(!list_empty(&cc.migratepages));
2645         }
2646 }
2647
2648 /* Compact all nodes in the system */
2649 static void compact_nodes(void)
2650 {
2651         int nid;
2652
2653         /* Flush pending updates to the LRU lists */
2654         lru_add_drain_all();
2655
2656         for_each_online_node(nid)
2657                 compact_node(nid);
2658 }
2659
2660 /* The written value is actually unused, all memory is compacted */
2661 int sysctl_compact_memory;
2662
2663 /*
2664  * Tunable for proactive compaction. It determines how
2665  * aggressively the kernel should compact memory in the
2666  * background. It takes values in the range [0, 100].
2667  */
2668 unsigned int __read_mostly sysctl_compaction_proactiveness = 20;
2669
2670 /*
2671  * This is the entry point for compacting all nodes via
2672  * /proc/sys/vm/compact_memory
2673  */
2674 int sysctl_compaction_handler(struct ctl_table *table, int write,
2675                         void *buffer, size_t *length, loff_t *ppos)
2676 {
2677         if (write)
2678                 compact_nodes();
2679
2680         return 0;
2681 }
2682
2683 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
2684 static ssize_t sysfs_compact_node(struct device *dev,
2685                         struct device_attribute *attr,
2686                         const char *buf, size_t count)
2687 {
2688         int nid = dev->id;
2689
2690         if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
2691                 /* Flush pending updates to the LRU lists */
2692                 lru_add_drain_all();
2693
2694                 compact_node(nid);
2695         }
2696
2697         return count;
2698 }
2699 static DEVICE_ATTR(compact, 0200, NULL, sysfs_compact_node);
2700
2701 int compaction_register_node(struct node *node)
2702 {
2703         return device_create_file(&node->dev, &dev_attr_compact);
2704 }
2705
2706 void compaction_unregister_node(struct node *node)
2707 {
2708         return device_remove_file(&node->dev, &dev_attr_compact);
2709 }
2710 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
2711
2712 static inline bool kcompactd_work_requested(pg_data_t *pgdat)
2713 {
2714         return pgdat->kcompactd_max_order > 0 || kthread_should_stop();
2715 }
2716
2717 static bool kcompactd_node_suitable(pg_data_t *pgdat)
2718 {
2719         int zoneid;
2720         struct zone *zone;
2721         enum zone_type highest_zoneidx = pgdat->kcompactd_highest_zoneidx;
2722
2723         for (zoneid = 0; zoneid <= highest_zoneidx; zoneid++) {
2724                 zone = &pgdat->node_zones[zoneid];
2725
2726                 if (!populated_zone(zone))
2727                         continue;
2728
2729                 if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0,
2730                                         highest_zoneidx) == COMPACT_CONTINUE)
2731                         return true;
2732         }
2733
2734         return false;
2735 }
2736
2737 static void kcompactd_do_work(pg_data_t *pgdat)
2738 {
2739         /*
2740          * With no special task, compact all zones so that a page of requested
2741          * order is allocatable.
2742          */
2743         int zoneid;
2744         struct zone *zone;
2745         struct compact_control cc = {
2746                 .order = pgdat->kcompactd_max_order,
2747                 .search_order = pgdat->kcompactd_max_order,
2748                 .highest_zoneidx = pgdat->kcompactd_highest_zoneidx,
2749                 .mode = MIGRATE_SYNC_LIGHT,
2750                 .ignore_skip_hint = false,
2751                 .gfp_mask = GFP_KERNEL,
2752         };
2753         trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
2754                                                         cc.highest_zoneidx);
2755         count_compact_event(KCOMPACTD_WAKE);
2756
2757         for (zoneid = 0; zoneid <= cc.highest_zoneidx; zoneid++) {
2758                 int status;
2759
2760                 zone = &pgdat->node_zones[zoneid];
2761                 if (!populated_zone(zone))
2762                         continue;
2763
2764                 if (compaction_deferred(zone, cc.order))
2765                         continue;
2766
2767                 if (compaction_suitable(zone, cc.order, 0, zoneid) !=
2768                                                         COMPACT_CONTINUE)
2769                         continue;
2770
2771                 if (kthread_should_stop())
2772                         return;
2773
2774                 cc.zone = zone;
2775                 status = compact_zone(&cc, NULL);
2776
2777                 if (status == COMPACT_SUCCESS) {
2778                         compaction_defer_reset(zone, cc.order, false);
2779                 } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
2780                         /*
2781                          * Buddy pages may become stranded on pcps that could
2782                          * otherwise coalesce on the zone's free area for
2783                          * order >= cc.order.  This is ratelimited by the
2784                          * upcoming deferral.
2785                          */
2786                         drain_all_pages(zone);
2787
2788                         /*
2789                          * We use sync migration mode here, so we defer like
2790                          * sync direct compaction does.
2791                          */
2792                         defer_compaction(zone, cc.order);
2793                 }
2794
2795                 count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
2796                                      cc.total_migrate_scanned);
2797                 count_compact_events(KCOMPACTD_FREE_SCANNED,
2798                                      cc.total_free_scanned);
2799
2800                 VM_BUG_ON(!list_empty(&cc.freepages));
2801                 VM_BUG_ON(!list_empty(&cc.migratepages));
2802         }
2803
2804         /*
2805          * Regardless of success, we are done until woken up next. But remember
2806          * the requested order/highest_zoneidx in case it was higher/tighter
2807          * than our current ones
2808          */
2809         if (pgdat->kcompactd_max_order <= cc.order)
2810                 pgdat->kcompactd_max_order = 0;
2811         if (pgdat->kcompactd_highest_zoneidx >= cc.highest_zoneidx)
2812                 pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
2813 }
2814
2815 void wakeup_kcompactd(pg_data_t *pgdat, int order, int highest_zoneidx)
2816 {
2817         if (!order)
2818                 return;
2819
2820         if (pgdat->kcompactd_max_order < order)
2821                 pgdat->kcompactd_max_order = order;
2822
2823         if (pgdat->kcompactd_highest_zoneidx > highest_zoneidx)
2824                 pgdat->kcompactd_highest_zoneidx = highest_zoneidx;
2825
2826         /*
2827          * Pairs with implicit barrier in wait_event_freezable()
2828          * such that wakeups are not missed.
2829          */
2830         if (!wq_has_sleeper(&pgdat->kcompactd_wait))
2831                 return;
2832
2833         if (!kcompactd_node_suitable(pgdat))
2834                 return;
2835
2836         trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
2837                                                         highest_zoneidx);
2838         wake_up_interruptible(&pgdat->kcompactd_wait);
2839 }
2840
2841 /*
2842  * The background compaction daemon, started as a kernel thread
2843  * from the init process.
2844  */
2845 static int kcompactd(void *p)
2846 {
2847         pg_data_t *pgdat = (pg_data_t*)p;
2848         struct task_struct *tsk = current;
2849         unsigned int proactive_defer = 0;
2850
2851         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2852
2853         if (!cpumask_empty(cpumask))
2854                 set_cpus_allowed_ptr(tsk, cpumask);
2855
2856         set_freezable();
2857
2858         pgdat->kcompactd_max_order = 0;
2859         pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
2860
2861         while (!kthread_should_stop()) {
2862                 unsigned long pflags;
2863
2864                 trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
2865                 if (wait_event_freezable_timeout(pgdat->kcompactd_wait,
2866                         kcompactd_work_requested(pgdat),
2867                         msecs_to_jiffies(HPAGE_FRAG_CHECK_INTERVAL_MSEC))) {
2868
2869                         psi_memstall_enter(&pflags);
2870                         kcompactd_do_work(pgdat);
2871                         psi_memstall_leave(&pflags);
2872                         continue;
2873                 }
2874
2875                 /* kcompactd wait timeout */
2876                 if (should_proactive_compact_node(pgdat)) {
2877                         unsigned int prev_score, score;
2878
2879                         if (proactive_defer) {
2880                                 proactive_defer--;
2881                                 continue;
2882                         }
2883                         prev_score = fragmentation_score_node(pgdat);
2884                         proactive_compact_node(pgdat);
2885                         score = fragmentation_score_node(pgdat);
2886                         /*
2887                          * Defer proactive compaction if the fragmentation
2888                          * score did not go down i.e. no progress made.
2889                          */
2890                         proactive_defer = score < prev_score ?
2891                                         0 : 1 << COMPACT_MAX_DEFER_SHIFT;
2892                 }
2893         }
2894
2895         return 0;
2896 }
2897
2898 /*
2899  * This kcompactd start function will be called by init and node-hot-add.
2900  * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
2901  */
2902 int kcompactd_run(int nid)
2903 {
2904         pg_data_t *pgdat = NODE_DATA(nid);
2905         int ret = 0;
2906
2907         if (pgdat->kcompactd)
2908                 return 0;
2909
2910         pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
2911         if (IS_ERR(pgdat->kcompactd)) {
2912                 pr_err("Failed to start kcompactd on node %d\n", nid);
2913                 ret = PTR_ERR(pgdat->kcompactd);
2914                 pgdat->kcompactd = NULL;
2915         }
2916         return ret;
2917 }
2918
2919 /*
2920  * Called by memory hotplug when all memory in a node is offlined. Caller must
2921  * hold mem_hotplug_begin/end().
2922  */
2923 void kcompactd_stop(int nid)
2924 {
2925         struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
2926
2927         if (kcompactd) {
2928                 kthread_stop(kcompactd);
2929                 NODE_DATA(nid)->kcompactd = NULL;
2930         }
2931 }
2932
2933 /*
2934  * It's optimal to keep kcompactd on the same CPUs as their memory, but
2935  * not required for correctness. So if the last cpu in a node goes
2936  * away, we get changed to run anywhere: as the first one comes back,
2937  * restore their cpu bindings.
2938  */
2939 static int kcompactd_cpu_online(unsigned int cpu)
2940 {
2941         int nid;
2942
2943         for_each_node_state(nid, N_MEMORY) {
2944                 pg_data_t *pgdat = NODE_DATA(nid);
2945                 const struct cpumask *mask;
2946
2947                 mask = cpumask_of_node(pgdat->node_id);
2948
2949                 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2950                         /* One of our CPUs online: restore mask */
2951                         set_cpus_allowed_ptr(pgdat->kcompactd, mask);
2952         }
2953         return 0;
2954 }
2955
2956 static int __init kcompactd_init(void)
2957 {
2958         int nid;
2959         int ret;
2960
2961         ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
2962                                         "mm/compaction:online",
2963                                         kcompactd_cpu_online, NULL);
2964         if (ret < 0) {
2965                 pr_err("kcompactd: failed to register hotplug callbacks.\n");
2966                 return ret;
2967         }
2968
2969         for_each_node_state(nid, N_MEMORY)
2970                 kcompactd_run(nid);
2971         return 0;
2972 }
2973 subsys_initcall(kcompactd_init)
2974
2975 #endif /* CONFIG_COMPACTION */