Merge tag 'ceph-for-5.9-rc1' of git://github.com/ceph/ceph-client
[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(!PageLocked(page), page);
141         VM_BUG_ON_PAGE(!PageMovable(page), page);
142         /*
143          * Clear registered address_space val with keeping PAGE_MAPPING_MOVABLE
144          * flag so that VM can catch up released page by driver after isolation.
145          * With it, VM migration doesn't try to put it back.
146          */
147         page->mapping = (void *)((unsigned long)page->mapping &
148                                 PAGE_MAPPING_MOVABLE);
149 }
150 EXPORT_SYMBOL(__ClearPageMovable);
151
152 /* Do not skip compaction more than 64 times */
153 #define COMPACT_MAX_DEFER_SHIFT 6
154
155 /*
156  * Compaction is deferred when compaction fails to result in a page
157  * allocation success. 1 << compact_defer_shift, compactions are skipped up
158  * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
159  */
160 void defer_compaction(struct zone *zone, int order)
161 {
162         zone->compact_considered = 0;
163         zone->compact_defer_shift++;
164
165         if (order < zone->compact_order_failed)
166                 zone->compact_order_failed = order;
167
168         if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
169                 zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
170
171         trace_mm_compaction_defer_compaction(zone, order);
172 }
173
174 /* Returns true if compaction should be skipped this time */
175 bool compaction_deferred(struct zone *zone, int order)
176 {
177         unsigned long defer_limit = 1UL << zone->compact_defer_shift;
178
179         if (order < zone->compact_order_failed)
180                 return false;
181
182         /* Avoid possible overflow */
183         if (++zone->compact_considered > defer_limit)
184                 zone->compact_considered = defer_limit;
185
186         if (zone->compact_considered >= defer_limit)
187                 return false;
188
189         trace_mm_compaction_deferred(zone, order);
190
191         return true;
192 }
193
194 /*
195  * Update defer tracking counters after successful compaction of given order,
196  * which means an allocation either succeeded (alloc_success == true) or is
197  * expected to succeed.
198  */
199 void compaction_defer_reset(struct zone *zone, int order,
200                 bool alloc_success)
201 {
202         if (alloc_success) {
203                 zone->compact_considered = 0;
204                 zone->compact_defer_shift = 0;
205         }
206         if (order >= zone->compact_order_failed)
207                 zone->compact_order_failed = order + 1;
208
209         trace_mm_compaction_defer_reset(zone, order);
210 }
211
212 /* Returns true if restarting compaction after many failures */
213 bool compaction_restarting(struct zone *zone, int order)
214 {
215         if (order < zone->compact_order_failed)
216                 return false;
217
218         return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
219                 zone->compact_considered >= 1UL << zone->compact_defer_shift;
220 }
221
222 /* Returns true if the pageblock should be scanned for pages to isolate. */
223 static inline bool isolation_suitable(struct compact_control *cc,
224                                         struct page *page)
225 {
226         if (cc->ignore_skip_hint)
227                 return true;
228
229         return !get_pageblock_skip(page);
230 }
231
232 static void reset_cached_positions(struct zone *zone)
233 {
234         zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
235         zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
236         zone->compact_cached_free_pfn =
237                                 pageblock_start_pfn(zone_end_pfn(zone) - 1);
238 }
239
240 /*
241  * Compound pages of >= pageblock_order should consistenly be skipped until
242  * released. It is always pointless to compact pages of such order (if they are
243  * migratable), and the pageblocks they occupy cannot contain any free pages.
244  */
245 static bool pageblock_skip_persistent(struct page *page)
246 {
247         if (!PageCompound(page))
248                 return false;
249
250         page = compound_head(page);
251
252         if (compound_order(page) >= pageblock_order)
253                 return true;
254
255         return false;
256 }
257
258 static bool
259 __reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source,
260                                                         bool check_target)
261 {
262         struct page *page = pfn_to_online_page(pfn);
263         struct page *block_page;
264         struct page *end_page;
265         unsigned long block_pfn;
266
267         if (!page)
268                 return false;
269         if (zone != page_zone(page))
270                 return false;
271         if (pageblock_skip_persistent(page))
272                 return false;
273
274         /*
275          * If skip is already cleared do no further checking once the
276          * restart points have been set.
277          */
278         if (check_source && check_target && !get_pageblock_skip(page))
279                 return true;
280
281         /*
282          * If clearing skip for the target scanner, do not select a
283          * non-movable pageblock as the starting point.
284          */
285         if (!check_source && check_target &&
286             get_pageblock_migratetype(page) != MIGRATE_MOVABLE)
287                 return false;
288
289         /* Ensure the start of the pageblock or zone is online and valid */
290         block_pfn = pageblock_start_pfn(pfn);
291         block_pfn = max(block_pfn, zone->zone_start_pfn);
292         block_page = pfn_to_online_page(block_pfn);
293         if (block_page) {
294                 page = block_page;
295                 pfn = block_pfn;
296         }
297
298         /* Ensure the end of the pageblock or zone is online and valid */
299         block_pfn = pageblock_end_pfn(pfn) - 1;
300         block_pfn = min(block_pfn, zone_end_pfn(zone) - 1);
301         end_page = pfn_to_online_page(block_pfn);
302         if (!end_page)
303                 return false;
304
305         /*
306          * Only clear the hint if a sample indicates there is either a
307          * free page or an LRU page in the block. One or other condition
308          * is necessary for the block to be a migration source/target.
309          */
310         do {
311                 if (pfn_valid_within(pfn)) {
312                         if (check_source && PageLRU(page)) {
313                                 clear_pageblock_skip(page);
314                                 return true;
315                         }
316
317                         if (check_target && PageBuddy(page)) {
318                                 clear_pageblock_skip(page);
319                                 return true;
320                         }
321                 }
322
323                 page += (1 << PAGE_ALLOC_COSTLY_ORDER);
324                 pfn += (1 << PAGE_ALLOC_COSTLY_ORDER);
325         } while (page <= end_page);
326
327         return false;
328 }
329
330 /*
331  * This function is called to clear all cached information on pageblocks that
332  * should be skipped for page isolation when the migrate and free page scanner
333  * meet.
334  */
335 static void __reset_isolation_suitable(struct zone *zone)
336 {
337         unsigned long migrate_pfn = zone->zone_start_pfn;
338         unsigned long free_pfn = zone_end_pfn(zone) - 1;
339         unsigned long reset_migrate = free_pfn;
340         unsigned long reset_free = migrate_pfn;
341         bool source_set = false;
342         bool free_set = false;
343
344         if (!zone->compact_blockskip_flush)
345                 return;
346
347         zone->compact_blockskip_flush = false;
348
349         /*
350          * Walk the zone and update pageblock skip information. Source looks
351          * for PageLRU while target looks for PageBuddy. When the scanner
352          * is found, both PageBuddy and PageLRU are checked as the pageblock
353          * is suitable as both source and target.
354          */
355         for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages,
356                                         free_pfn -= pageblock_nr_pages) {
357                 cond_resched();
358
359                 /* Update the migrate PFN */
360                 if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) &&
361                     migrate_pfn < reset_migrate) {
362                         source_set = true;
363                         reset_migrate = migrate_pfn;
364                         zone->compact_init_migrate_pfn = reset_migrate;
365                         zone->compact_cached_migrate_pfn[0] = reset_migrate;
366                         zone->compact_cached_migrate_pfn[1] = reset_migrate;
367                 }
368
369                 /* Update the free PFN */
370                 if (__reset_isolation_pfn(zone, free_pfn, free_set, true) &&
371                     free_pfn > reset_free) {
372                         free_set = true;
373                         reset_free = free_pfn;
374                         zone->compact_init_free_pfn = reset_free;
375                         zone->compact_cached_free_pfn = reset_free;
376                 }
377         }
378
379         /* Leave no distance if no suitable block was reset */
380         if (reset_migrate >= reset_free) {
381                 zone->compact_cached_migrate_pfn[0] = migrate_pfn;
382                 zone->compact_cached_migrate_pfn[1] = migrate_pfn;
383                 zone->compact_cached_free_pfn = free_pfn;
384         }
385 }
386
387 void reset_isolation_suitable(pg_data_t *pgdat)
388 {
389         int zoneid;
390
391         for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
392                 struct zone *zone = &pgdat->node_zones[zoneid];
393                 if (!populated_zone(zone))
394                         continue;
395
396                 /* Only flush if a full compaction finished recently */
397                 if (zone->compact_blockskip_flush)
398                         __reset_isolation_suitable(zone);
399         }
400 }
401
402 /*
403  * Sets the pageblock skip bit if it was clear. Note that this is a hint as
404  * locks are not required for read/writers. Returns true if it was already set.
405  */
406 static bool test_and_set_skip(struct compact_control *cc, struct page *page,
407                                                         unsigned long pfn)
408 {
409         bool skip;
410
411         /* Do no update if skip hint is being ignored */
412         if (cc->ignore_skip_hint)
413                 return false;
414
415         if (!IS_ALIGNED(pfn, pageblock_nr_pages))
416                 return false;
417
418         skip = get_pageblock_skip(page);
419         if (!skip && !cc->no_set_skip_hint)
420                 set_pageblock_skip(page);
421
422         return skip;
423 }
424
425 static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
426 {
427         struct zone *zone = cc->zone;
428
429         pfn = pageblock_end_pfn(pfn);
430
431         /* Set for isolation rather than compaction */
432         if (cc->no_set_skip_hint)
433                 return;
434
435         if (pfn > zone->compact_cached_migrate_pfn[0])
436                 zone->compact_cached_migrate_pfn[0] = pfn;
437         if (cc->mode != MIGRATE_ASYNC &&
438             pfn > zone->compact_cached_migrate_pfn[1])
439                 zone->compact_cached_migrate_pfn[1] = pfn;
440 }
441
442 /*
443  * If no pages were isolated then mark this pageblock to be skipped in the
444  * future. The information is later cleared by __reset_isolation_suitable().
445  */
446 static void update_pageblock_skip(struct compact_control *cc,
447                         struct page *page, unsigned long pfn)
448 {
449         struct zone *zone = cc->zone;
450
451         if (cc->no_set_skip_hint)
452                 return;
453
454         if (!page)
455                 return;
456
457         set_pageblock_skip(page);
458
459         /* Update where async and sync compaction should restart */
460         if (pfn < zone->compact_cached_free_pfn)
461                 zone->compact_cached_free_pfn = pfn;
462 }
463 #else
464 static inline bool isolation_suitable(struct compact_control *cc,
465                                         struct page *page)
466 {
467         return true;
468 }
469
470 static inline bool pageblock_skip_persistent(struct page *page)
471 {
472         return false;
473 }
474
475 static inline void update_pageblock_skip(struct compact_control *cc,
476                         struct page *page, unsigned long pfn)
477 {
478 }
479
480 static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
481 {
482 }
483
484 static bool test_and_set_skip(struct compact_control *cc, struct page *page,
485                                                         unsigned long pfn)
486 {
487         return false;
488 }
489 #endif /* CONFIG_COMPACTION */
490
491 /*
492  * Compaction requires the taking of some coarse locks that are potentially
493  * very heavily contended. For async compaction, trylock and record if the
494  * lock is contended. The lock will still be acquired but compaction will
495  * abort when the current block is finished regardless of success rate.
496  * Sync compaction acquires the lock.
497  *
498  * Always returns true which makes it easier to track lock state in callers.
499  */
500 static bool compact_lock_irqsave(spinlock_t *lock, unsigned long *flags,
501                                                 struct compact_control *cc)
502         __acquires(lock)
503 {
504         /* Track if the lock is contended in async mode */
505         if (cc->mode == MIGRATE_ASYNC && !cc->contended) {
506                 if (spin_trylock_irqsave(lock, *flags))
507                         return true;
508
509                 cc->contended = true;
510         }
511
512         spin_lock_irqsave(lock, *flags);
513         return true;
514 }
515
516 /*
517  * Compaction requires the taking of some coarse locks that are potentially
518  * very heavily contended. The lock should be periodically unlocked to avoid
519  * having disabled IRQs for a long time, even when there is nobody waiting on
520  * the lock. It might also be that allowing the IRQs will result in
521  * need_resched() becoming true. If scheduling is needed, async compaction
522  * aborts. Sync compaction schedules.
523  * Either compaction type will also abort if a fatal signal is pending.
524  * In either case if the lock was locked, it is dropped and not regained.
525  *
526  * Returns true if compaction should abort due to fatal signal pending, or
527  *              async compaction due to need_resched()
528  * Returns false when compaction can continue (sync compaction might have
529  *              scheduled)
530  */
531 static bool compact_unlock_should_abort(spinlock_t *lock,
532                 unsigned long flags, bool *locked, struct compact_control *cc)
533 {
534         if (*locked) {
535                 spin_unlock_irqrestore(lock, flags);
536                 *locked = false;
537         }
538
539         if (fatal_signal_pending(current)) {
540                 cc->contended = true;
541                 return true;
542         }
543
544         cond_resched();
545
546         return false;
547 }
548
549 /*
550  * Isolate free pages onto a private freelist. If @strict is true, will abort
551  * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
552  * (even though it may still end up isolating some pages).
553  */
554 static unsigned long isolate_freepages_block(struct compact_control *cc,
555                                 unsigned long *start_pfn,
556                                 unsigned long end_pfn,
557                                 struct list_head *freelist,
558                                 unsigned int stride,
559                                 bool strict)
560 {
561         int nr_scanned = 0, total_isolated = 0;
562         struct page *cursor;
563         unsigned long flags = 0;
564         bool locked = false;
565         unsigned long blockpfn = *start_pfn;
566         unsigned int order;
567
568         /* Strict mode is for isolation, speed is secondary */
569         if (strict)
570                 stride = 1;
571
572         cursor = pfn_to_page(blockpfn);
573
574         /* Isolate free pages. */
575         for (; blockpfn < end_pfn; blockpfn += stride, cursor += stride) {
576                 int isolated;
577                 struct page *page = cursor;
578
579                 /*
580                  * Periodically drop the lock (if held) regardless of its
581                  * contention, to give chance to IRQs. Abort if fatal signal
582                  * pending or async compaction detects need_resched()
583                  */
584                 if (!(blockpfn % SWAP_CLUSTER_MAX)
585                     && compact_unlock_should_abort(&cc->zone->lock, flags,
586                                                                 &locked, cc))
587                         break;
588
589                 nr_scanned++;
590                 if (!pfn_valid_within(blockpfn))
591                         goto isolate_fail;
592
593                 /*
594                  * For compound pages such as THP and hugetlbfs, we can save
595                  * potentially a lot of iterations if we skip them at once.
596                  * The check is racy, but we can consider only valid values
597                  * and the only danger is skipping too much.
598                  */
599                 if (PageCompound(page)) {
600                         const unsigned int order = compound_order(page);
601
602                         if (likely(order < MAX_ORDER)) {
603                                 blockpfn += (1UL << order) - 1;
604                                 cursor += (1UL << order) - 1;
605                         }
606                         goto isolate_fail;
607                 }
608
609                 if (!PageBuddy(page))
610                         goto isolate_fail;
611
612                 /*
613                  * If we already hold the lock, we can skip some rechecking.
614                  * Note that if we hold the lock now, checked_pageblock was
615                  * already set in some previous iteration (or strict is true),
616                  * so it is correct to skip the suitable migration target
617                  * recheck as well.
618                  */
619                 if (!locked) {
620                         locked = compact_lock_irqsave(&cc->zone->lock,
621                                                                 &flags, cc);
622
623                         /* Recheck this is a buddy page under lock */
624                         if (!PageBuddy(page))
625                                 goto isolate_fail;
626                 }
627
628                 /* Found a free page, will break it into order-0 pages */
629                 order = page_order(page);
630                 isolated = __isolate_free_page(page, order);
631                 if (!isolated)
632                         break;
633                 set_page_private(page, order);
634
635                 total_isolated += isolated;
636                 cc->nr_freepages += isolated;
637                 list_add_tail(&page->lru, freelist);
638
639                 if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
640                         blockpfn += isolated;
641                         break;
642                 }
643                 /* Advance to the end of split page */
644                 blockpfn += isolated - 1;
645                 cursor += isolated - 1;
646                 continue;
647
648 isolate_fail:
649                 if (strict)
650                         break;
651                 else
652                         continue;
653
654         }
655
656         if (locked)
657                 spin_unlock_irqrestore(&cc->zone->lock, flags);
658
659         /*
660          * There is a tiny chance that we have read bogus compound_order(),
661          * so be careful to not go outside of the pageblock.
662          */
663         if (unlikely(blockpfn > end_pfn))
664                 blockpfn = end_pfn;
665
666         trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
667                                         nr_scanned, total_isolated);
668
669         /* Record how far we have got within the block */
670         *start_pfn = blockpfn;
671
672         /*
673          * If strict isolation is requested by CMA then check that all the
674          * pages requested were isolated. If there were any failures, 0 is
675          * returned and CMA will fail.
676          */
677         if (strict && blockpfn < end_pfn)
678                 total_isolated = 0;
679
680         cc->total_free_scanned += nr_scanned;
681         if (total_isolated)
682                 count_compact_events(COMPACTISOLATED, total_isolated);
683         return total_isolated;
684 }
685
686 /**
687  * isolate_freepages_range() - isolate free pages.
688  * @cc:        Compaction control structure.
689  * @start_pfn: The first PFN to start isolating.
690  * @end_pfn:   The one-past-last PFN.
691  *
692  * Non-free pages, invalid PFNs, or zone boundaries within the
693  * [start_pfn, end_pfn) range are considered errors, cause function to
694  * undo its actions and return zero.
695  *
696  * Otherwise, function returns one-past-the-last PFN of isolated page
697  * (which may be greater then end_pfn if end fell in a middle of
698  * a free page).
699  */
700 unsigned long
701 isolate_freepages_range(struct compact_control *cc,
702                         unsigned long start_pfn, unsigned long end_pfn)
703 {
704         unsigned long isolated, pfn, block_start_pfn, block_end_pfn;
705         LIST_HEAD(freelist);
706
707         pfn = start_pfn;
708         block_start_pfn = pageblock_start_pfn(pfn);
709         if (block_start_pfn < cc->zone->zone_start_pfn)
710                 block_start_pfn = cc->zone->zone_start_pfn;
711         block_end_pfn = pageblock_end_pfn(pfn);
712
713         for (; pfn < end_pfn; pfn += isolated,
714                                 block_start_pfn = block_end_pfn,
715                                 block_end_pfn += pageblock_nr_pages) {
716                 /* Protect pfn from changing by isolate_freepages_block */
717                 unsigned long isolate_start_pfn = pfn;
718
719                 block_end_pfn = min(block_end_pfn, end_pfn);
720
721                 /*
722                  * pfn could pass the block_end_pfn if isolated freepage
723                  * is more than pageblock order. In this case, we adjust
724                  * scanning range to right one.
725                  */
726                 if (pfn >= block_end_pfn) {
727                         block_start_pfn = pageblock_start_pfn(pfn);
728                         block_end_pfn = pageblock_end_pfn(pfn);
729                         block_end_pfn = min(block_end_pfn, end_pfn);
730                 }
731
732                 if (!pageblock_pfn_to_page(block_start_pfn,
733                                         block_end_pfn, cc->zone))
734                         break;
735
736                 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
737                                         block_end_pfn, &freelist, 0, true);
738
739                 /*
740                  * In strict mode, isolate_freepages_block() returns 0 if
741                  * there are any holes in the block (ie. invalid PFNs or
742                  * non-free pages).
743                  */
744                 if (!isolated)
745                         break;
746
747                 /*
748                  * If we managed to isolate pages, it is always (1 << n) *
749                  * pageblock_nr_pages for some non-negative n.  (Max order
750                  * page may span two pageblocks).
751                  */
752         }
753
754         /* __isolate_free_page() does not map the pages */
755         split_map_pages(&freelist);
756
757         if (pfn < end_pfn) {
758                 /* Loop terminated early, cleanup. */
759                 release_freepages(&freelist);
760                 return 0;
761         }
762
763         /* We don't use freelists for anything. */
764         return pfn;
765 }
766
767 /* Similar to reclaim, but different enough that they don't share logic */
768 static bool too_many_isolated(pg_data_t *pgdat)
769 {
770         unsigned long active, inactive, isolated;
771
772         inactive = node_page_state(pgdat, NR_INACTIVE_FILE) +
773                         node_page_state(pgdat, NR_INACTIVE_ANON);
774         active = node_page_state(pgdat, NR_ACTIVE_FILE) +
775                         node_page_state(pgdat, NR_ACTIVE_ANON);
776         isolated = node_page_state(pgdat, NR_ISOLATED_FILE) +
777                         node_page_state(pgdat, NR_ISOLATED_ANON);
778
779         return isolated > (inactive + active) / 2;
780 }
781
782 /**
783  * isolate_migratepages_block() - isolate all migrate-able pages within
784  *                                a single pageblock
785  * @cc:         Compaction control structure.
786  * @low_pfn:    The first PFN to isolate
787  * @end_pfn:    The one-past-the-last PFN to isolate, within same pageblock
788  * @isolate_mode: Isolation mode to be used.
789  *
790  * Isolate all pages that can be migrated from the range specified by
791  * [low_pfn, end_pfn). The range is expected to be within same pageblock.
792  * Returns zero if there is a fatal signal pending, otherwise PFN of the
793  * first page that was not scanned (which may be both less, equal to or more
794  * than end_pfn).
795  *
796  * The pages are isolated on cc->migratepages list (not required to be empty),
797  * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
798  * is neither read nor updated.
799  */
800 static unsigned long
801 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
802                         unsigned long end_pfn, isolate_mode_t isolate_mode)
803 {
804         pg_data_t *pgdat = cc->zone->zone_pgdat;
805         unsigned long nr_scanned = 0, nr_isolated = 0;
806         struct lruvec *lruvec;
807         unsigned long flags = 0;
808         bool locked = false;
809         struct page *page = NULL, *valid_page = NULL;
810         unsigned long start_pfn = low_pfn;
811         bool skip_on_failure = false;
812         unsigned long next_skip_pfn = 0;
813         bool skip_updated = false;
814
815         /*
816          * Ensure that there are not too many pages isolated from the LRU
817          * list by either parallel reclaimers or compaction. If there are,
818          * delay for some time until fewer pages are isolated
819          */
820         while (unlikely(too_many_isolated(pgdat))) {
821                 /* async migration should just abort */
822                 if (cc->mode == MIGRATE_ASYNC)
823                         return 0;
824
825                 congestion_wait(BLK_RW_ASYNC, HZ/10);
826
827                 if (fatal_signal_pending(current))
828                         return 0;
829         }
830
831         cond_resched();
832
833         if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
834                 skip_on_failure = true;
835                 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
836         }
837
838         /* Time to isolate some pages for migration */
839         for (; low_pfn < end_pfn; low_pfn++) {
840
841                 if (skip_on_failure && low_pfn >= next_skip_pfn) {
842                         /*
843                          * We have isolated all migration candidates in the
844                          * previous order-aligned block, and did not skip it due
845                          * to failure. We should migrate the pages now and
846                          * hopefully succeed compaction.
847                          */
848                         if (nr_isolated)
849                                 break;
850
851                         /*
852                          * We failed to isolate in the previous order-aligned
853                          * block. Set the new boundary to the end of the
854                          * current block. Note we can't simply increase
855                          * next_skip_pfn by 1 << order, as low_pfn might have
856                          * been incremented by a higher number due to skipping
857                          * a compound or a high-order buddy page in the
858                          * previous loop iteration.
859                          */
860                         next_skip_pfn = block_end_pfn(low_pfn, cc->order);
861                 }
862
863                 /*
864                  * Periodically drop the lock (if held) regardless of its
865                  * contention, to give chance to IRQs. Abort completely if
866                  * a fatal signal is pending.
867                  */
868                 if (!(low_pfn % SWAP_CLUSTER_MAX)
869                     && compact_unlock_should_abort(&pgdat->lru_lock,
870                                             flags, &locked, cc)) {
871                         low_pfn = 0;
872                         goto fatal_pending;
873                 }
874
875                 if (!pfn_valid_within(low_pfn))
876                         goto isolate_fail;
877                 nr_scanned++;
878
879                 page = pfn_to_page(low_pfn);
880
881                 /*
882                  * Check if the pageblock has already been marked skipped.
883                  * Only the aligned PFN is checked as the caller isolates
884                  * COMPACT_CLUSTER_MAX at a time so the second call must
885                  * not falsely conclude that the block should be skipped.
886                  */
887                 if (!valid_page && IS_ALIGNED(low_pfn, pageblock_nr_pages)) {
888                         if (!cc->ignore_skip_hint && get_pageblock_skip(page)) {
889                                 low_pfn = end_pfn;
890                                 goto isolate_abort;
891                         }
892                         valid_page = page;
893                 }
894
895                 /*
896                  * Skip if free. We read page order here without zone lock
897                  * which is generally unsafe, but the race window is small and
898                  * the worst thing that can happen is that we skip some
899                  * potential isolation targets.
900                  */
901                 if (PageBuddy(page)) {
902                         unsigned long freepage_order = page_order_unsafe(page);
903
904                         /*
905                          * Without lock, we cannot be sure that what we got is
906                          * a valid page order. Consider only values in the
907                          * valid order range to prevent low_pfn overflow.
908                          */
909                         if (freepage_order > 0 && freepage_order < MAX_ORDER)
910                                 low_pfn += (1UL << freepage_order) - 1;
911                         continue;
912                 }
913
914                 /*
915                  * Regardless of being on LRU, compound pages such as THP and
916                  * hugetlbfs are not to be compacted unless we are attempting
917                  * an allocation much larger than the huge page size (eg CMA).
918                  * We can potentially save a lot of iterations if we skip them
919                  * at once. The check is racy, but we can consider only valid
920                  * values and the only danger is skipping too much.
921                  */
922                 if (PageCompound(page) && !cc->alloc_contig) {
923                         const unsigned int order = compound_order(page);
924
925                         if (likely(order < MAX_ORDER))
926                                 low_pfn += (1UL << order) - 1;
927                         goto isolate_fail;
928                 }
929
930                 /*
931                  * Check may be lockless but that's ok as we recheck later.
932                  * It's possible to migrate LRU and non-lru movable pages.
933                  * Skip any other type of page
934                  */
935                 if (!PageLRU(page)) {
936                         /*
937                          * __PageMovable can return false positive so we need
938                          * to verify it under page_lock.
939                          */
940                         if (unlikely(__PageMovable(page)) &&
941                                         !PageIsolated(page)) {
942                                 if (locked) {
943                                         spin_unlock_irqrestore(&pgdat->lru_lock,
944                                                                         flags);
945                                         locked = false;
946                                 }
947
948                                 if (!isolate_movable_page(page, isolate_mode))
949                                         goto isolate_success;
950                         }
951
952                         goto isolate_fail;
953                 }
954
955                 /*
956                  * Migration will fail if an anonymous page is pinned in memory,
957                  * so avoid taking lru_lock and isolating it unnecessarily in an
958                  * admittedly racy check.
959                  */
960                 if (!page_mapping(page) &&
961                     page_count(page) > page_mapcount(page))
962                         goto isolate_fail;
963
964                 /*
965                  * Only allow to migrate anonymous pages in GFP_NOFS context
966                  * because those do not depend on fs locks.
967                  */
968                 if (!(cc->gfp_mask & __GFP_FS) && page_mapping(page))
969                         goto isolate_fail;
970
971                 /* If we already hold the lock, we can skip some rechecking */
972                 if (!locked) {
973                         locked = compact_lock_irqsave(&pgdat->lru_lock,
974                                                                 &flags, cc);
975
976                         /* Try get exclusive access under lock */
977                         if (!skip_updated) {
978                                 skip_updated = true;
979                                 if (test_and_set_skip(cc, page, low_pfn))
980                                         goto isolate_abort;
981                         }
982
983                         /* Recheck PageLRU and PageCompound under lock */
984                         if (!PageLRU(page))
985                                 goto isolate_fail;
986
987                         /*
988                          * Page become compound since the non-locked check,
989                          * and it's on LRU. It can only be a THP so the order
990                          * is safe to read and it's 0 for tail pages.
991                          */
992                         if (unlikely(PageCompound(page) && !cc->alloc_contig)) {
993                                 low_pfn += compound_nr(page) - 1;
994                                 goto isolate_fail;
995                         }
996                 }
997
998                 lruvec = mem_cgroup_page_lruvec(page, pgdat);
999
1000                 /* Try isolate the page */
1001                 if (__isolate_lru_page(page, isolate_mode) != 0)
1002                         goto isolate_fail;
1003
1004                 /* The whole page is taken off the LRU; skip the tail pages. */
1005                 if (PageCompound(page))
1006                         low_pfn += compound_nr(page) - 1;
1007
1008                 /* Successfully isolated */
1009                 del_page_from_lru_list(page, lruvec, page_lru(page));
1010                 mod_node_page_state(page_pgdat(page),
1011                                 NR_ISOLATED_ANON + page_is_file_lru(page),
1012                                 hpage_nr_pages(page));
1013
1014 isolate_success:
1015                 list_add(&page->lru, &cc->migratepages);
1016                 cc->nr_migratepages++;
1017                 nr_isolated++;
1018
1019                 /*
1020                  * Avoid isolating too much unless this block is being
1021                  * rescanned (e.g. dirty/writeback pages, parallel allocation)
1022                  * or a lock is contended. For contention, isolate quickly to
1023                  * potentially remove one source of contention.
1024                  */
1025                 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX &&
1026                     !cc->rescan && !cc->contended) {
1027                         ++low_pfn;
1028                         break;
1029                 }
1030
1031                 continue;
1032 isolate_fail:
1033                 if (!skip_on_failure)
1034                         continue;
1035
1036                 /*
1037                  * We have isolated some pages, but then failed. Release them
1038                  * instead of migrating, as we cannot form the cc->order buddy
1039                  * page anyway.
1040                  */
1041                 if (nr_isolated) {
1042                         if (locked) {
1043                                 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
1044                                 locked = false;
1045                         }
1046                         putback_movable_pages(&cc->migratepages);
1047                         cc->nr_migratepages = 0;
1048                         nr_isolated = 0;
1049                 }
1050
1051                 if (low_pfn < next_skip_pfn) {
1052                         low_pfn = next_skip_pfn - 1;
1053                         /*
1054                          * The check near the loop beginning would have updated
1055                          * next_skip_pfn too, but this is a bit simpler.
1056                          */
1057                         next_skip_pfn += 1UL << cc->order;
1058                 }
1059         }
1060
1061         /*
1062          * The PageBuddy() check could have potentially brought us outside
1063          * the range to be scanned.
1064          */
1065         if (unlikely(low_pfn > end_pfn))
1066                 low_pfn = end_pfn;
1067
1068 isolate_abort:
1069         if (locked)
1070                 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
1071
1072         /*
1073          * Updated the cached scanner pfn once the pageblock has been scanned
1074          * Pages will either be migrated in which case there is no point
1075          * scanning in the near future or migration failed in which case the
1076          * failure reason may persist. The block is marked for skipping if
1077          * there were no pages isolated in the block or if the block is
1078          * rescanned twice in a row.
1079          */
1080         if (low_pfn == end_pfn && (!nr_isolated || cc->rescan)) {
1081                 if (valid_page && !skip_updated)
1082                         set_pageblock_skip(valid_page);
1083                 update_cached_migrate(cc, low_pfn);
1084         }
1085
1086         trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
1087                                                 nr_scanned, nr_isolated);
1088
1089 fatal_pending:
1090         cc->total_migrate_scanned += nr_scanned;
1091         if (nr_isolated)
1092                 count_compact_events(COMPACTISOLATED, nr_isolated);
1093
1094         return low_pfn;
1095 }
1096
1097 /**
1098  * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
1099  * @cc:        Compaction control structure.
1100  * @start_pfn: The first PFN to start isolating.
1101  * @end_pfn:   The one-past-last PFN.
1102  *
1103  * Returns zero if isolation fails fatally due to e.g. pending signal.
1104  * Otherwise, function returns one-past-the-last PFN of isolated page
1105  * (which may be greater than end_pfn if end fell in a middle of a THP page).
1106  */
1107 unsigned long
1108 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
1109                                                         unsigned long end_pfn)
1110 {
1111         unsigned long pfn, block_start_pfn, block_end_pfn;
1112
1113         /* Scan block by block. First and last block may be incomplete */
1114         pfn = start_pfn;
1115         block_start_pfn = pageblock_start_pfn(pfn);
1116         if (block_start_pfn < cc->zone->zone_start_pfn)
1117                 block_start_pfn = cc->zone->zone_start_pfn;
1118         block_end_pfn = pageblock_end_pfn(pfn);
1119
1120         for (; pfn < end_pfn; pfn = block_end_pfn,
1121                                 block_start_pfn = block_end_pfn,
1122                                 block_end_pfn += pageblock_nr_pages) {
1123
1124                 block_end_pfn = min(block_end_pfn, end_pfn);
1125
1126                 if (!pageblock_pfn_to_page(block_start_pfn,
1127                                         block_end_pfn, cc->zone))
1128                         continue;
1129
1130                 pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
1131                                                         ISOLATE_UNEVICTABLE);
1132
1133                 if (!pfn)
1134                         break;
1135
1136                 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
1137                         break;
1138         }
1139
1140         return pfn;
1141 }
1142
1143 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
1144 #ifdef CONFIG_COMPACTION
1145
1146 static bool suitable_migration_source(struct compact_control *cc,
1147                                                         struct page *page)
1148 {
1149         int block_mt;
1150
1151         if (pageblock_skip_persistent(page))
1152                 return false;
1153
1154         if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction)
1155                 return true;
1156
1157         block_mt = get_pageblock_migratetype(page);
1158
1159         if (cc->migratetype == MIGRATE_MOVABLE)
1160                 return is_migrate_movable(block_mt);
1161         else
1162                 return block_mt == cc->migratetype;
1163 }
1164
1165 /* Returns true if the page is within a block suitable for migration to */
1166 static bool suitable_migration_target(struct compact_control *cc,
1167                                                         struct page *page)
1168 {
1169         /* If the page is a large free page, then disallow migration */
1170         if (PageBuddy(page)) {
1171                 /*
1172                  * We are checking page_order without zone->lock taken. But
1173                  * the only small danger is that we skip a potentially suitable
1174                  * pageblock, so it's not worth to check order for valid range.
1175                  */
1176                 if (page_order_unsafe(page) >= pageblock_order)
1177                         return false;
1178         }
1179
1180         if (cc->ignore_block_suitable)
1181                 return true;
1182
1183         /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1184         if (is_migrate_movable(get_pageblock_migratetype(page)))
1185                 return true;
1186
1187         /* Otherwise skip the block */
1188         return false;
1189 }
1190
1191 static inline unsigned int
1192 freelist_scan_limit(struct compact_control *cc)
1193 {
1194         unsigned short shift = BITS_PER_LONG - 1;
1195
1196         return (COMPACT_CLUSTER_MAX >> min(shift, cc->fast_search_fail)) + 1;
1197 }
1198
1199 /*
1200  * Test whether the free scanner has reached the same or lower pageblock than
1201  * the migration scanner, and compaction should thus terminate.
1202  */
1203 static inline bool compact_scanners_met(struct compact_control *cc)
1204 {
1205         return (cc->free_pfn >> pageblock_order)
1206                 <= (cc->migrate_pfn >> pageblock_order);
1207 }
1208
1209 /*
1210  * Used when scanning for a suitable migration target which scans freelists
1211  * in reverse. Reorders the list such as the unscanned pages are scanned
1212  * first on the next iteration of the free scanner
1213  */
1214 static void
1215 move_freelist_head(struct list_head *freelist, struct page *freepage)
1216 {
1217         LIST_HEAD(sublist);
1218
1219         if (!list_is_last(freelist, &freepage->lru)) {
1220                 list_cut_before(&sublist, freelist, &freepage->lru);
1221                 if (!list_empty(&sublist))
1222                         list_splice_tail(&sublist, freelist);
1223         }
1224 }
1225
1226 /*
1227  * Similar to move_freelist_head except used by the migration scanner
1228  * when scanning forward. It's possible for these list operations to
1229  * move against each other if they search the free list exactly in
1230  * lockstep.
1231  */
1232 static void
1233 move_freelist_tail(struct list_head *freelist, struct page *freepage)
1234 {
1235         LIST_HEAD(sublist);
1236
1237         if (!list_is_first(freelist, &freepage->lru)) {
1238                 list_cut_position(&sublist, freelist, &freepage->lru);
1239                 if (!list_empty(&sublist))
1240                         list_splice_tail(&sublist, freelist);
1241         }
1242 }
1243
1244 static void
1245 fast_isolate_around(struct compact_control *cc, unsigned long pfn, unsigned long nr_isolated)
1246 {
1247         unsigned long start_pfn, end_pfn;
1248         struct page *page = pfn_to_page(pfn);
1249
1250         /* Do not search around if there are enough pages already */
1251         if (cc->nr_freepages >= cc->nr_migratepages)
1252                 return;
1253
1254         /* Minimise scanning during async compaction */
1255         if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC)
1256                 return;
1257
1258         /* Pageblock boundaries */
1259         start_pfn = pageblock_start_pfn(pfn);
1260         end_pfn = min(pageblock_end_pfn(pfn), zone_end_pfn(cc->zone)) - 1;
1261
1262         /* Scan before */
1263         if (start_pfn != pfn) {
1264                 isolate_freepages_block(cc, &start_pfn, pfn, &cc->freepages, 1, false);
1265                 if (cc->nr_freepages >= cc->nr_migratepages)
1266                         return;
1267         }
1268
1269         /* Scan after */
1270         start_pfn = pfn + nr_isolated;
1271         if (start_pfn < end_pfn)
1272                 isolate_freepages_block(cc, &start_pfn, end_pfn, &cc->freepages, 1, false);
1273
1274         /* Skip this pageblock in the future as it's full or nearly full */
1275         if (cc->nr_freepages < cc->nr_migratepages)
1276                 set_pageblock_skip(page);
1277 }
1278
1279 /* Search orders in round-robin fashion */
1280 static int next_search_order(struct compact_control *cc, int order)
1281 {
1282         order--;
1283         if (order < 0)
1284                 order = cc->order - 1;
1285
1286         /* Search wrapped around? */
1287         if (order == cc->search_order) {
1288                 cc->search_order--;
1289                 if (cc->search_order < 0)
1290                         cc->search_order = cc->order - 1;
1291                 return -1;
1292         }
1293
1294         return order;
1295 }
1296
1297 static unsigned long
1298 fast_isolate_freepages(struct compact_control *cc)
1299 {
1300         unsigned int limit = min(1U, freelist_scan_limit(cc) >> 1);
1301         unsigned int nr_scanned = 0;
1302         unsigned long low_pfn, min_pfn, high_pfn = 0, highest = 0;
1303         unsigned long nr_isolated = 0;
1304         unsigned long distance;
1305         struct page *page = NULL;
1306         bool scan_start = false;
1307         int order;
1308
1309         /* Full compaction passes in a negative order */
1310         if (cc->order <= 0)
1311                 return cc->free_pfn;
1312
1313         /*
1314          * If starting the scan, use a deeper search and use the highest
1315          * PFN found if a suitable one is not found.
1316          */
1317         if (cc->free_pfn >= cc->zone->compact_init_free_pfn) {
1318                 limit = pageblock_nr_pages >> 1;
1319                 scan_start = true;
1320         }
1321
1322         /*
1323          * Preferred point is in the top quarter of the scan space but take
1324          * a pfn from the top half if the search is problematic.
1325          */
1326         distance = (cc->free_pfn - cc->migrate_pfn);
1327         low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2));
1328         min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1));
1329
1330         if (WARN_ON_ONCE(min_pfn > low_pfn))
1331                 low_pfn = min_pfn;
1332
1333         /*
1334          * Search starts from the last successful isolation order or the next
1335          * order to search after a previous failure
1336          */
1337         cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order);
1338
1339         for (order = cc->search_order;
1340              !page && order >= 0;
1341              order = next_search_order(cc, order)) {
1342                 struct free_area *area = &cc->zone->free_area[order];
1343                 struct list_head *freelist;
1344                 struct page *freepage;
1345                 unsigned long flags;
1346                 unsigned int order_scanned = 0;
1347
1348                 if (!area->nr_free)
1349                         continue;
1350
1351                 spin_lock_irqsave(&cc->zone->lock, flags);
1352                 freelist = &area->free_list[MIGRATE_MOVABLE];
1353                 list_for_each_entry_reverse(freepage, freelist, lru) {
1354                         unsigned long pfn;
1355
1356                         order_scanned++;
1357                         nr_scanned++;
1358                         pfn = page_to_pfn(freepage);
1359
1360                         if (pfn >= highest)
1361                                 highest = pageblock_start_pfn(pfn);
1362
1363                         if (pfn >= low_pfn) {
1364                                 cc->fast_search_fail = 0;
1365                                 cc->search_order = order;
1366                                 page = freepage;
1367                                 break;
1368                         }
1369
1370                         if (pfn >= min_pfn && pfn > high_pfn) {
1371                                 high_pfn = pfn;
1372
1373                                 /* Shorten the scan if a candidate is found */
1374                                 limit >>= 1;
1375                         }
1376
1377                         if (order_scanned >= limit)
1378                                 break;
1379                 }
1380
1381                 /* Use a minimum pfn if a preferred one was not found */
1382                 if (!page && high_pfn) {
1383                         page = pfn_to_page(high_pfn);
1384
1385                         /* Update freepage for the list reorder below */
1386                         freepage = page;
1387                 }
1388
1389                 /* Reorder to so a future search skips recent pages */
1390                 move_freelist_head(freelist, freepage);
1391
1392                 /* Isolate the page if available */
1393                 if (page) {
1394                         if (__isolate_free_page(page, order)) {
1395                                 set_page_private(page, order);
1396                                 nr_isolated = 1 << order;
1397                                 cc->nr_freepages += nr_isolated;
1398                                 list_add_tail(&page->lru, &cc->freepages);
1399                                 count_compact_events(COMPACTISOLATED, nr_isolated);
1400                         } else {
1401                                 /* If isolation fails, abort the search */
1402                                 order = cc->search_order + 1;
1403                                 page = NULL;
1404                         }
1405                 }
1406
1407                 spin_unlock_irqrestore(&cc->zone->lock, flags);
1408
1409                 /*
1410                  * Smaller scan on next order so the total scan ig related
1411                  * to freelist_scan_limit.
1412                  */
1413                 if (order_scanned >= limit)
1414                         limit = min(1U, limit >> 1);
1415         }
1416
1417         if (!page) {
1418                 cc->fast_search_fail++;
1419                 if (scan_start) {
1420                         /*
1421                          * Use the highest PFN found above min. If one was
1422                          * not found, be pessimistic for direct compaction
1423                          * and use the min mark.
1424                          */
1425                         if (highest) {
1426                                 page = pfn_to_page(highest);
1427                                 cc->free_pfn = highest;
1428                         } else {
1429                                 if (cc->direct_compaction && pfn_valid(min_pfn)) {
1430                                         page = pageblock_pfn_to_page(min_pfn,
1431                                                 pageblock_end_pfn(min_pfn),
1432                                                 cc->zone);
1433                                         cc->free_pfn = min_pfn;
1434                                 }
1435                         }
1436                 }
1437         }
1438
1439         if (highest && highest >= cc->zone->compact_cached_free_pfn) {
1440                 highest -= pageblock_nr_pages;
1441                 cc->zone->compact_cached_free_pfn = highest;
1442         }
1443
1444         cc->total_free_scanned += nr_scanned;
1445         if (!page)
1446                 return cc->free_pfn;
1447
1448         low_pfn = page_to_pfn(page);
1449         fast_isolate_around(cc, low_pfn, nr_isolated);
1450         return low_pfn;
1451 }
1452
1453 /*
1454  * Based on information in the current compact_control, find blocks
1455  * suitable for isolating free pages from and then isolate them.
1456  */
1457 static void isolate_freepages(struct compact_control *cc)
1458 {
1459         struct zone *zone = cc->zone;
1460         struct page *page;
1461         unsigned long block_start_pfn;  /* start of current pageblock */
1462         unsigned long isolate_start_pfn; /* exact pfn we start at */
1463         unsigned long block_end_pfn;    /* end of current pageblock */
1464         unsigned long low_pfn;       /* lowest pfn scanner is able to scan */
1465         struct list_head *freelist = &cc->freepages;
1466         unsigned int stride;
1467
1468         /* Try a small search of the free lists for a candidate */
1469         isolate_start_pfn = fast_isolate_freepages(cc);
1470         if (cc->nr_freepages)
1471                 goto splitmap;
1472
1473         /*
1474          * Initialise the free scanner. The starting point is where we last
1475          * successfully isolated from, zone-cached value, or the end of the
1476          * zone when isolating for the first time. For looping we also need
1477          * this pfn aligned down to the pageblock boundary, because we do
1478          * block_start_pfn -= pageblock_nr_pages in the for loop.
1479          * For ending point, take care when isolating in last pageblock of a
1480          * zone which ends in the middle of a pageblock.
1481          * The low boundary is the end of the pageblock the migration scanner
1482          * is using.
1483          */
1484         isolate_start_pfn = cc->free_pfn;
1485         block_start_pfn = pageblock_start_pfn(isolate_start_pfn);
1486         block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1487                                                 zone_end_pfn(zone));
1488         low_pfn = pageblock_end_pfn(cc->migrate_pfn);
1489         stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1;
1490
1491         /*
1492          * Isolate free pages until enough are available to migrate the
1493          * pages on cc->migratepages. We stop searching if the migrate
1494          * and free page scanners meet or enough free pages are isolated.
1495          */
1496         for (; block_start_pfn >= low_pfn;
1497                                 block_end_pfn = block_start_pfn,
1498                                 block_start_pfn -= pageblock_nr_pages,
1499                                 isolate_start_pfn = block_start_pfn) {
1500                 unsigned long nr_isolated;
1501
1502                 /*
1503                  * This can iterate a massively long zone without finding any
1504                  * suitable migration targets, so periodically check resched.
1505                  */
1506                 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)))
1507                         cond_resched();
1508
1509                 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1510                                                                         zone);
1511                 if (!page)
1512                         continue;
1513
1514                 /* Check the block is suitable for migration */
1515                 if (!suitable_migration_target(cc, page))
1516                         continue;
1517
1518                 /* If isolation recently failed, do not retry */
1519                 if (!isolation_suitable(cc, page))
1520                         continue;
1521
1522                 /* Found a block suitable for isolating free pages from. */
1523                 nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn,
1524                                         block_end_pfn, freelist, stride, false);
1525
1526                 /* Update the skip hint if the full pageblock was scanned */
1527                 if (isolate_start_pfn == block_end_pfn)
1528                         update_pageblock_skip(cc, page, block_start_pfn);
1529
1530                 /* Are enough freepages isolated? */
1531                 if (cc->nr_freepages >= cc->nr_migratepages) {
1532                         if (isolate_start_pfn >= block_end_pfn) {
1533                                 /*
1534                                  * Restart at previous pageblock if more
1535                                  * freepages can be isolated next time.
1536                                  */
1537                                 isolate_start_pfn =
1538                                         block_start_pfn - pageblock_nr_pages;
1539                         }
1540                         break;
1541                 } else if (isolate_start_pfn < block_end_pfn) {
1542                         /*
1543                          * If isolation failed early, do not continue
1544                          * needlessly.
1545                          */
1546                         break;
1547                 }
1548
1549                 /* Adjust stride depending on isolation */
1550                 if (nr_isolated) {
1551                         stride = 1;
1552                         continue;
1553                 }
1554                 stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1);
1555         }
1556
1557         /*
1558          * Record where the free scanner will restart next time. Either we
1559          * broke from the loop and set isolate_start_pfn based on the last
1560          * call to isolate_freepages_block(), or we met the migration scanner
1561          * and the loop terminated due to isolate_start_pfn < low_pfn
1562          */
1563         cc->free_pfn = isolate_start_pfn;
1564
1565 splitmap:
1566         /* __isolate_free_page() does not map the pages */
1567         split_map_pages(freelist);
1568 }
1569
1570 /*
1571  * This is a migrate-callback that "allocates" freepages by taking pages
1572  * from the isolated freelists in the block we are migrating to.
1573  */
1574 static struct page *compaction_alloc(struct page *migratepage,
1575                                         unsigned long data)
1576 {
1577         struct compact_control *cc = (struct compact_control *)data;
1578         struct page *freepage;
1579
1580         if (list_empty(&cc->freepages)) {
1581                 isolate_freepages(cc);
1582
1583                 if (list_empty(&cc->freepages))
1584                         return NULL;
1585         }
1586
1587         freepage = list_entry(cc->freepages.next, struct page, lru);
1588         list_del(&freepage->lru);
1589         cc->nr_freepages--;
1590
1591         return freepage;
1592 }
1593
1594 /*
1595  * This is a migrate-callback that "frees" freepages back to the isolated
1596  * freelist.  All pages on the freelist are from the same zone, so there is no
1597  * special handling needed for NUMA.
1598  */
1599 static void compaction_free(struct page *page, unsigned long data)
1600 {
1601         struct compact_control *cc = (struct compact_control *)data;
1602
1603         list_add(&page->lru, &cc->freepages);
1604         cc->nr_freepages++;
1605 }
1606
1607 /* possible outcome of isolate_migratepages */
1608 typedef enum {
1609         ISOLATE_ABORT,          /* Abort compaction now */
1610         ISOLATE_NONE,           /* No pages isolated, continue scanning */
1611         ISOLATE_SUCCESS,        /* Pages isolated, migrate */
1612 } isolate_migrate_t;
1613
1614 /*
1615  * Allow userspace to control policy on scanning the unevictable LRU for
1616  * compactable pages.
1617  */
1618 #ifdef CONFIG_PREEMPT_RT
1619 int sysctl_compact_unevictable_allowed __read_mostly = 0;
1620 #else
1621 int sysctl_compact_unevictable_allowed __read_mostly = 1;
1622 #endif
1623
1624 static inline void
1625 update_fast_start_pfn(struct compact_control *cc, unsigned long pfn)
1626 {
1627         if (cc->fast_start_pfn == ULONG_MAX)
1628                 return;
1629
1630         if (!cc->fast_start_pfn)
1631                 cc->fast_start_pfn = pfn;
1632
1633         cc->fast_start_pfn = min(cc->fast_start_pfn, pfn);
1634 }
1635
1636 static inline unsigned long
1637 reinit_migrate_pfn(struct compact_control *cc)
1638 {
1639         if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX)
1640                 return cc->migrate_pfn;
1641
1642         cc->migrate_pfn = cc->fast_start_pfn;
1643         cc->fast_start_pfn = ULONG_MAX;
1644
1645         return cc->migrate_pfn;
1646 }
1647
1648 /*
1649  * Briefly search the free lists for a migration source that already has
1650  * some free pages to reduce the number of pages that need migration
1651  * before a pageblock is free.
1652  */
1653 static unsigned long fast_find_migrateblock(struct compact_control *cc)
1654 {
1655         unsigned int limit = freelist_scan_limit(cc);
1656         unsigned int nr_scanned = 0;
1657         unsigned long distance;
1658         unsigned long pfn = cc->migrate_pfn;
1659         unsigned long high_pfn;
1660         int order;
1661
1662         /* Skip hints are relied on to avoid repeats on the fast search */
1663         if (cc->ignore_skip_hint)
1664                 return pfn;
1665
1666         /*
1667          * If the migrate_pfn is not at the start of a zone or the start
1668          * of a pageblock then assume this is a continuation of a previous
1669          * scan restarted due to COMPACT_CLUSTER_MAX.
1670          */
1671         if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn))
1672                 return pfn;
1673
1674         /*
1675          * For smaller orders, just linearly scan as the number of pages
1676          * to migrate should be relatively small and does not necessarily
1677          * justify freeing up a large block for a small allocation.
1678          */
1679         if (cc->order <= PAGE_ALLOC_COSTLY_ORDER)
1680                 return pfn;
1681
1682         /*
1683          * Only allow kcompactd and direct requests for movable pages to
1684          * quickly clear out a MOVABLE pageblock for allocation. This
1685          * reduces the risk that a large movable pageblock is freed for
1686          * an unmovable/reclaimable small allocation.
1687          */
1688         if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE)
1689                 return pfn;
1690
1691         /*
1692          * When starting the migration scanner, pick any pageblock within the
1693          * first half of the search space. Otherwise try and pick a pageblock
1694          * within the first eighth to reduce the chances that a migration
1695          * target later becomes a source.
1696          */
1697         distance = (cc->free_pfn - cc->migrate_pfn) >> 1;
1698         if (cc->migrate_pfn != cc->zone->zone_start_pfn)
1699                 distance >>= 2;
1700         high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance);
1701
1702         for (order = cc->order - 1;
1703              order >= PAGE_ALLOC_COSTLY_ORDER && pfn == cc->migrate_pfn && nr_scanned < limit;
1704              order--) {
1705                 struct free_area *area = &cc->zone->free_area[order];
1706                 struct list_head *freelist;
1707                 unsigned long flags;
1708                 struct page *freepage;
1709
1710                 if (!area->nr_free)
1711                         continue;
1712
1713                 spin_lock_irqsave(&cc->zone->lock, flags);
1714                 freelist = &area->free_list[MIGRATE_MOVABLE];
1715                 list_for_each_entry(freepage, freelist, lru) {
1716                         unsigned long free_pfn;
1717
1718                         nr_scanned++;
1719                         free_pfn = page_to_pfn(freepage);
1720                         if (free_pfn < high_pfn) {
1721                                 /*
1722                                  * Avoid if skipped recently. Ideally it would
1723                                  * move to the tail but even safe iteration of
1724                                  * the list assumes an entry is deleted, not
1725                                  * reordered.
1726                                  */
1727                                 if (get_pageblock_skip(freepage)) {
1728                                         if (list_is_last(freelist, &freepage->lru))
1729                                                 break;
1730
1731                                         continue;
1732                                 }
1733
1734                                 /* Reorder to so a future search skips recent pages */
1735                                 move_freelist_tail(freelist, freepage);
1736
1737                                 update_fast_start_pfn(cc, free_pfn);
1738                                 pfn = pageblock_start_pfn(free_pfn);
1739                                 cc->fast_search_fail = 0;
1740                                 set_pageblock_skip(freepage);
1741                                 break;
1742                         }
1743
1744                         if (nr_scanned >= limit) {
1745                                 cc->fast_search_fail++;
1746                                 move_freelist_tail(freelist, freepage);
1747                                 break;
1748                         }
1749                 }
1750                 spin_unlock_irqrestore(&cc->zone->lock, flags);
1751         }
1752
1753         cc->total_migrate_scanned += nr_scanned;
1754
1755         /*
1756          * If fast scanning failed then use a cached entry for a page block
1757          * that had free pages as the basis for starting a linear scan.
1758          */
1759         if (pfn == cc->migrate_pfn)
1760                 pfn = reinit_migrate_pfn(cc);
1761
1762         return pfn;
1763 }
1764
1765 /*
1766  * Isolate all pages that can be migrated from the first suitable block,
1767  * starting at the block pointed to by the migrate scanner pfn within
1768  * compact_control.
1769  */
1770 static isolate_migrate_t isolate_migratepages(struct compact_control *cc)
1771 {
1772         unsigned long block_start_pfn;
1773         unsigned long block_end_pfn;
1774         unsigned long low_pfn;
1775         struct page *page;
1776         const isolate_mode_t isolate_mode =
1777                 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1778                 (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1779         bool fast_find_block;
1780
1781         /*
1782          * Start at where we last stopped, or beginning of the zone as
1783          * initialized by compact_zone(). The first failure will use
1784          * the lowest PFN as the starting point for linear scanning.
1785          */
1786         low_pfn = fast_find_migrateblock(cc);
1787         block_start_pfn = pageblock_start_pfn(low_pfn);
1788         if (block_start_pfn < cc->zone->zone_start_pfn)
1789                 block_start_pfn = cc->zone->zone_start_pfn;
1790
1791         /*
1792          * fast_find_migrateblock marks a pageblock skipped so to avoid
1793          * the isolation_suitable check below, check whether the fast
1794          * search was successful.
1795          */
1796         fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail;
1797
1798         /* Only scan within a pageblock boundary */
1799         block_end_pfn = pageblock_end_pfn(low_pfn);
1800
1801         /*
1802          * Iterate over whole pageblocks until we find the first suitable.
1803          * Do not cross the free scanner.
1804          */
1805         for (; block_end_pfn <= cc->free_pfn;
1806                         fast_find_block = false,
1807                         low_pfn = block_end_pfn,
1808                         block_start_pfn = block_end_pfn,
1809                         block_end_pfn += pageblock_nr_pages) {
1810
1811                 /*
1812                  * This can potentially iterate a massively long zone with
1813                  * many pageblocks unsuitable, so periodically check if we
1814                  * need to schedule.
1815                  */
1816                 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)))
1817                         cond_resched();
1818
1819                 page = pageblock_pfn_to_page(block_start_pfn,
1820                                                 block_end_pfn, cc->zone);
1821                 if (!page)
1822                         continue;
1823
1824                 /*
1825                  * If isolation recently failed, do not retry. Only check the
1826                  * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock
1827                  * to be visited multiple times. Assume skip was checked
1828                  * before making it "skip" so other compaction instances do
1829                  * not scan the same block.
1830                  */
1831                 if (IS_ALIGNED(low_pfn, pageblock_nr_pages) &&
1832                     !fast_find_block && !isolation_suitable(cc, page))
1833                         continue;
1834
1835                 /*
1836                  * For async compaction, also only scan in MOVABLE blocks
1837                  * without huge pages. Async compaction is optimistic to see
1838                  * if the minimum amount of work satisfies the allocation.
1839                  * The cached PFN is updated as it's possible that all
1840                  * remaining blocks between source and target are unsuitable
1841                  * and the compaction scanners fail to meet.
1842                  */
1843                 if (!suitable_migration_source(cc, page)) {
1844                         update_cached_migrate(cc, block_end_pfn);
1845                         continue;
1846                 }
1847
1848                 /* Perform the isolation */
1849                 low_pfn = isolate_migratepages_block(cc, low_pfn,
1850                                                 block_end_pfn, isolate_mode);
1851
1852                 if (!low_pfn)
1853                         return ISOLATE_ABORT;
1854
1855                 /*
1856                  * Either we isolated something and proceed with migration. Or
1857                  * we failed and compact_zone should decide if we should
1858                  * continue or not.
1859                  */
1860                 break;
1861         }
1862
1863         /* Record where migration scanner will be restarted. */
1864         cc->migrate_pfn = low_pfn;
1865
1866         return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
1867 }
1868
1869 /*
1870  * order == -1 is expected when compacting via
1871  * /proc/sys/vm/compact_memory
1872  */
1873 static inline bool is_via_compact_memory(int order)
1874 {
1875         return order == -1;
1876 }
1877
1878 static bool kswapd_is_running(pg_data_t *pgdat)
1879 {
1880         return pgdat->kswapd && (pgdat->kswapd->state == TASK_RUNNING);
1881 }
1882
1883 /*
1884  * A zone's fragmentation score is the external fragmentation wrt to the
1885  * COMPACTION_HPAGE_ORDER scaled by the zone's size. It returns a value
1886  * in the range [0, 100].
1887  *
1888  * The scaling factor ensures that proactive compaction focuses on larger
1889  * zones like ZONE_NORMAL, rather than smaller, specialized zones like
1890  * ZONE_DMA32. For smaller zones, the score value remains close to zero,
1891  * and thus never exceeds the high threshold for proactive compaction.
1892  */
1893 static unsigned int fragmentation_score_zone(struct zone *zone)
1894 {
1895         unsigned long score;
1896
1897         score = zone->present_pages *
1898                         extfrag_for_order(zone, COMPACTION_HPAGE_ORDER);
1899         return div64_ul(score, zone->zone_pgdat->node_present_pages + 1);
1900 }
1901
1902 /*
1903  * The per-node proactive (background) compaction process is started by its
1904  * corresponding kcompactd thread when the node's fragmentation score
1905  * exceeds the high threshold. The compaction process remains active till
1906  * the node's score falls below the low threshold, or one of the back-off
1907  * conditions is met.
1908  */
1909 static unsigned int fragmentation_score_node(pg_data_t *pgdat)
1910 {
1911         unsigned int score = 0;
1912         int zoneid;
1913
1914         for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1915                 struct zone *zone;
1916
1917                 zone = &pgdat->node_zones[zoneid];
1918                 score += fragmentation_score_zone(zone);
1919         }
1920
1921         return score;
1922 }
1923
1924 static unsigned int fragmentation_score_wmark(pg_data_t *pgdat, bool low)
1925 {
1926         unsigned int wmark_low;
1927
1928         /*
1929          * Cap the low watermak to avoid excessive compaction
1930          * activity in case a user sets the proactivess tunable
1931          * close to 100 (maximum).
1932          */
1933         wmark_low = max(100U - sysctl_compaction_proactiveness, 5U);
1934         return low ? wmark_low : min(wmark_low + 10, 100U);
1935 }
1936
1937 static bool should_proactive_compact_node(pg_data_t *pgdat)
1938 {
1939         int wmark_high;
1940
1941         if (!sysctl_compaction_proactiveness || kswapd_is_running(pgdat))
1942                 return false;
1943
1944         wmark_high = fragmentation_score_wmark(pgdat, false);
1945         return fragmentation_score_node(pgdat) > wmark_high;
1946 }
1947
1948 static enum compact_result __compact_finished(struct compact_control *cc)
1949 {
1950         unsigned int order;
1951         const int migratetype = cc->migratetype;
1952         int ret;
1953
1954         /* Compaction run completes if the migrate and free scanner meet */
1955         if (compact_scanners_met(cc)) {
1956                 /* Let the next compaction start anew. */
1957                 reset_cached_positions(cc->zone);
1958
1959                 /*
1960                  * Mark that the PG_migrate_skip information should be cleared
1961                  * by kswapd when it goes to sleep. kcompactd does not set the
1962                  * flag itself as the decision to be clear should be directly
1963                  * based on an allocation request.
1964                  */
1965                 if (cc->direct_compaction)
1966                         cc->zone->compact_blockskip_flush = true;
1967
1968                 if (cc->whole_zone)
1969                         return COMPACT_COMPLETE;
1970                 else
1971                         return COMPACT_PARTIAL_SKIPPED;
1972         }
1973
1974         if (cc->proactive_compaction) {
1975                 int score, wmark_low;
1976                 pg_data_t *pgdat;
1977
1978                 pgdat = cc->zone->zone_pgdat;
1979                 if (kswapd_is_running(pgdat))
1980                         return COMPACT_PARTIAL_SKIPPED;
1981
1982                 score = fragmentation_score_zone(cc->zone);
1983                 wmark_low = fragmentation_score_wmark(pgdat, true);
1984
1985                 if (score > wmark_low)
1986                         ret = COMPACT_CONTINUE;
1987                 else
1988                         ret = COMPACT_SUCCESS;
1989
1990                 goto out;
1991         }
1992
1993         if (is_via_compact_memory(cc->order))
1994                 return COMPACT_CONTINUE;
1995
1996         /*
1997          * Always finish scanning a pageblock to reduce the possibility of
1998          * fallbacks in the future. This is particularly important when
1999          * migration source is unmovable/reclaimable but it's not worth
2000          * special casing.
2001          */
2002         if (!IS_ALIGNED(cc->migrate_pfn, pageblock_nr_pages))
2003                 return COMPACT_CONTINUE;
2004
2005         /* Direct compactor: Is a suitable page free? */
2006         ret = COMPACT_NO_SUITABLE_PAGE;
2007         for (order = cc->order; order < MAX_ORDER; order++) {
2008                 struct free_area *area = &cc->zone->free_area[order];
2009                 bool can_steal;
2010
2011                 /* Job done if page is free of the right migratetype */
2012                 if (!free_area_empty(area, migratetype))
2013                         return COMPACT_SUCCESS;
2014
2015 #ifdef CONFIG_CMA
2016                 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
2017                 if (migratetype == MIGRATE_MOVABLE &&
2018                         !free_area_empty(area, MIGRATE_CMA))
2019                         return COMPACT_SUCCESS;
2020 #endif
2021                 /*
2022                  * Job done if allocation would steal freepages from
2023                  * other migratetype buddy lists.
2024                  */
2025                 if (find_suitable_fallback(area, order, migratetype,
2026                                                 true, &can_steal) != -1) {
2027
2028                         /* movable pages are OK in any pageblock */
2029                         if (migratetype == MIGRATE_MOVABLE)
2030                                 return COMPACT_SUCCESS;
2031
2032                         /*
2033                          * We are stealing for a non-movable allocation. Make
2034                          * sure we finish compacting the current pageblock
2035                          * first so it is as free as possible and we won't
2036                          * have to steal another one soon. This only applies
2037                          * to sync compaction, as async compaction operates
2038                          * on pageblocks of the same migratetype.
2039                          */
2040                         if (cc->mode == MIGRATE_ASYNC ||
2041                                         IS_ALIGNED(cc->migrate_pfn,
2042                                                         pageblock_nr_pages)) {
2043                                 return COMPACT_SUCCESS;
2044                         }
2045
2046                         ret = COMPACT_CONTINUE;
2047                         break;
2048                 }
2049         }
2050
2051 out:
2052         if (cc->contended || fatal_signal_pending(current))
2053                 ret = COMPACT_CONTENDED;
2054
2055         return ret;
2056 }
2057
2058 static enum compact_result compact_finished(struct compact_control *cc)
2059 {
2060         int ret;
2061
2062         ret = __compact_finished(cc);
2063         trace_mm_compaction_finished(cc->zone, cc->order, ret);
2064         if (ret == COMPACT_NO_SUITABLE_PAGE)
2065                 ret = COMPACT_CONTINUE;
2066
2067         return ret;
2068 }
2069
2070 /*
2071  * compaction_suitable: Is this suitable to run compaction on this zone now?
2072  * Returns
2073  *   COMPACT_SKIPPED  - If there are too few free pages for compaction
2074  *   COMPACT_SUCCESS  - If the allocation would succeed without compaction
2075  *   COMPACT_CONTINUE - If compaction should run now
2076  */
2077 static enum compact_result __compaction_suitable(struct zone *zone, int order,
2078                                         unsigned int alloc_flags,
2079                                         int highest_zoneidx,
2080                                         unsigned long wmark_target)
2081 {
2082         unsigned long watermark;
2083
2084         if (is_via_compact_memory(order))
2085                 return COMPACT_CONTINUE;
2086
2087         watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
2088         /*
2089          * If watermarks for high-order allocation are already met, there
2090          * should be no need for compaction at all.
2091          */
2092         if (zone_watermark_ok(zone, order, watermark, highest_zoneidx,
2093                                                                 alloc_flags))
2094                 return COMPACT_SUCCESS;
2095
2096         /*
2097          * Watermarks for order-0 must be met for compaction to be able to
2098          * isolate free pages for migration targets. This means that the
2099          * watermark and alloc_flags have to match, or be more pessimistic than
2100          * the check in __isolate_free_page(). We don't use the direct
2101          * compactor's alloc_flags, as they are not relevant for freepage
2102          * isolation. We however do use the direct compactor's highest_zoneidx
2103          * to skip over zones where lowmem reserves would prevent allocation
2104          * even if compaction succeeds.
2105          * For costly orders, we require low watermark instead of min for
2106          * compaction to proceed to increase its chances.
2107          * ALLOC_CMA is used, as pages in CMA pageblocks are considered
2108          * suitable migration targets
2109          */
2110         watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
2111                                 low_wmark_pages(zone) : min_wmark_pages(zone);
2112         watermark += compact_gap(order);
2113         if (!__zone_watermark_ok(zone, 0, watermark, highest_zoneidx,
2114                                                 ALLOC_CMA, wmark_target))
2115                 return COMPACT_SKIPPED;
2116
2117         return COMPACT_CONTINUE;
2118 }
2119
2120 enum compact_result compaction_suitable(struct zone *zone, int order,
2121                                         unsigned int alloc_flags,
2122                                         int highest_zoneidx)
2123 {
2124         enum compact_result ret;
2125         int fragindex;
2126
2127         ret = __compaction_suitable(zone, order, alloc_flags, highest_zoneidx,
2128                                     zone_page_state(zone, NR_FREE_PAGES));
2129         /*
2130          * fragmentation index determines if allocation failures are due to
2131          * low memory or external fragmentation
2132          *
2133          * index of -1000 would imply allocations might succeed depending on
2134          * watermarks, but we already failed the high-order watermark check
2135          * index towards 0 implies failure is due to lack of memory
2136          * index towards 1000 implies failure is due to fragmentation
2137          *
2138          * Only compact if a failure would be due to fragmentation. Also
2139          * ignore fragindex for non-costly orders where the alternative to
2140          * a successful reclaim/compaction is OOM. Fragindex and the
2141          * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
2142          * excessive compaction for costly orders, but it should not be at the
2143          * expense of system stability.
2144          */
2145         if (ret == COMPACT_CONTINUE && (order > PAGE_ALLOC_COSTLY_ORDER)) {
2146                 fragindex = fragmentation_index(zone, order);
2147                 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
2148                         ret = COMPACT_NOT_SUITABLE_ZONE;
2149         }
2150
2151         trace_mm_compaction_suitable(zone, order, ret);
2152         if (ret == COMPACT_NOT_SUITABLE_ZONE)
2153                 ret = COMPACT_SKIPPED;
2154
2155         return ret;
2156 }
2157
2158 bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
2159                 int alloc_flags)
2160 {
2161         struct zone *zone;
2162         struct zoneref *z;
2163
2164         /*
2165          * Make sure at least one zone would pass __compaction_suitable if we continue
2166          * retrying the reclaim.
2167          */
2168         for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2169                                 ac->highest_zoneidx, ac->nodemask) {
2170                 unsigned long available;
2171                 enum compact_result compact_result;
2172
2173                 /*
2174                  * Do not consider all the reclaimable memory because we do not
2175                  * want to trash just for a single high order allocation which
2176                  * is even not guaranteed to appear even if __compaction_suitable
2177                  * is happy about the watermark check.
2178                  */
2179                 available = zone_reclaimable_pages(zone) / order;
2180                 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
2181                 compact_result = __compaction_suitable(zone, order, alloc_flags,
2182                                 ac->highest_zoneidx, available);
2183                 if (compact_result != COMPACT_SKIPPED)
2184                         return true;
2185         }
2186
2187         return false;
2188 }
2189
2190 static enum compact_result
2191 compact_zone(struct compact_control *cc, struct capture_control *capc)
2192 {
2193         enum compact_result ret;
2194         unsigned long start_pfn = cc->zone->zone_start_pfn;
2195         unsigned long end_pfn = zone_end_pfn(cc->zone);
2196         unsigned long last_migrated_pfn;
2197         const bool sync = cc->mode != MIGRATE_ASYNC;
2198         bool update_cached;
2199
2200         /*
2201          * These counters track activities during zone compaction.  Initialize
2202          * them before compacting a new zone.
2203          */
2204         cc->total_migrate_scanned = 0;
2205         cc->total_free_scanned = 0;
2206         cc->nr_migratepages = 0;
2207         cc->nr_freepages = 0;
2208         INIT_LIST_HEAD(&cc->freepages);
2209         INIT_LIST_HEAD(&cc->migratepages);
2210
2211         cc->migratetype = gfp_migratetype(cc->gfp_mask);
2212         ret = compaction_suitable(cc->zone, cc->order, cc->alloc_flags,
2213                                                         cc->highest_zoneidx);
2214         /* Compaction is likely to fail */
2215         if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED)
2216                 return ret;
2217
2218         /* huh, compaction_suitable is returning something unexpected */
2219         VM_BUG_ON(ret != COMPACT_CONTINUE);
2220
2221         /*
2222          * Clear pageblock skip if there were failures recently and compaction
2223          * is about to be retried after being deferred.
2224          */
2225         if (compaction_restarting(cc->zone, cc->order))
2226                 __reset_isolation_suitable(cc->zone);
2227
2228         /*
2229          * Setup to move all movable pages to the end of the zone. Used cached
2230          * information on where the scanners should start (unless we explicitly
2231          * want to compact the whole zone), but check that it is initialised
2232          * by ensuring the values are within zone boundaries.
2233          */
2234         cc->fast_start_pfn = 0;
2235         if (cc->whole_zone) {
2236                 cc->migrate_pfn = start_pfn;
2237                 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2238         } else {
2239                 cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync];
2240                 cc->free_pfn = cc->zone->compact_cached_free_pfn;
2241                 if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
2242                         cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2243                         cc->zone->compact_cached_free_pfn = cc->free_pfn;
2244                 }
2245                 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
2246                         cc->migrate_pfn = start_pfn;
2247                         cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
2248                         cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
2249                 }
2250
2251                 if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn)
2252                         cc->whole_zone = true;
2253         }
2254
2255         last_migrated_pfn = 0;
2256
2257         /*
2258          * Migrate has separate cached PFNs for ASYNC and SYNC* migration on
2259          * the basis that some migrations will fail in ASYNC mode. However,
2260          * if the cached PFNs match and pageblocks are skipped due to having
2261          * no isolation candidates, then the sync state does not matter.
2262          * Until a pageblock with isolation candidates is found, keep the
2263          * cached PFNs in sync to avoid revisiting the same blocks.
2264          */
2265         update_cached = !sync &&
2266                 cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1];
2267
2268         trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
2269                                 cc->free_pfn, end_pfn, sync);
2270
2271         migrate_prep_local();
2272
2273         while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) {
2274                 int err;
2275                 unsigned long start_pfn = cc->migrate_pfn;
2276
2277                 /*
2278                  * Avoid multiple rescans which can happen if a page cannot be
2279                  * isolated (dirty/writeback in async mode) or if the migrated
2280                  * pages are being allocated before the pageblock is cleared.
2281                  * The first rescan will capture the entire pageblock for
2282                  * migration. If it fails, it'll be marked skip and scanning
2283                  * will proceed as normal.
2284                  */
2285                 cc->rescan = false;
2286                 if (pageblock_start_pfn(last_migrated_pfn) ==
2287                     pageblock_start_pfn(start_pfn)) {
2288                         cc->rescan = true;
2289                 }
2290
2291                 switch (isolate_migratepages(cc)) {
2292                 case ISOLATE_ABORT:
2293                         ret = COMPACT_CONTENDED;
2294                         putback_movable_pages(&cc->migratepages);
2295                         cc->nr_migratepages = 0;
2296                         goto out;
2297                 case ISOLATE_NONE:
2298                         if (update_cached) {
2299                                 cc->zone->compact_cached_migrate_pfn[1] =
2300                                         cc->zone->compact_cached_migrate_pfn[0];
2301                         }
2302
2303                         /*
2304                          * We haven't isolated and migrated anything, but
2305                          * there might still be unflushed migrations from
2306                          * previous cc->order aligned block.
2307                          */
2308                         goto check_drain;
2309                 case ISOLATE_SUCCESS:
2310                         update_cached = false;
2311                         last_migrated_pfn = start_pfn;
2312                         ;
2313                 }
2314
2315                 err = migrate_pages(&cc->migratepages, compaction_alloc,
2316                                 compaction_free, (unsigned long)cc, cc->mode,
2317                                 MR_COMPACTION);
2318
2319                 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
2320                                                         &cc->migratepages);
2321
2322                 /* All pages were either migrated or will be released */
2323                 cc->nr_migratepages = 0;
2324                 if (err) {
2325                         putback_movable_pages(&cc->migratepages);
2326                         /*
2327                          * migrate_pages() may return -ENOMEM when scanners meet
2328                          * and we want compact_finished() to detect it
2329                          */
2330                         if (err == -ENOMEM && !compact_scanners_met(cc)) {
2331                                 ret = COMPACT_CONTENDED;
2332                                 goto out;
2333                         }
2334                         /*
2335                          * We failed to migrate at least one page in the current
2336                          * order-aligned block, so skip the rest of it.
2337                          */
2338                         if (cc->direct_compaction &&
2339                                                 (cc->mode == MIGRATE_ASYNC)) {
2340                                 cc->migrate_pfn = block_end_pfn(
2341                                                 cc->migrate_pfn - 1, cc->order);
2342                                 /* Draining pcplists is useless in this case */
2343                                 last_migrated_pfn = 0;
2344                         }
2345                 }
2346
2347 check_drain:
2348                 /*
2349                  * Has the migration scanner moved away from the previous
2350                  * cc->order aligned block where we migrated from? If yes,
2351                  * flush the pages that were freed, so that they can merge and
2352                  * compact_finished() can detect immediately if allocation
2353                  * would succeed.
2354                  */
2355                 if (cc->order > 0 && last_migrated_pfn) {
2356                         unsigned long current_block_start =
2357                                 block_start_pfn(cc->migrate_pfn, cc->order);
2358
2359                         if (last_migrated_pfn < current_block_start) {
2360                                 lru_add_drain_cpu_zone(cc->zone);
2361                                 /* No more flushing until we migrate again */
2362                                 last_migrated_pfn = 0;
2363                         }
2364                 }
2365
2366                 /* Stop if a page has been captured */
2367                 if (capc && capc->page) {
2368                         ret = COMPACT_SUCCESS;
2369                         break;
2370                 }
2371         }
2372
2373 out:
2374         /*
2375          * Release free pages and update where the free scanner should restart,
2376          * so we don't leave any returned pages behind in the next attempt.
2377          */
2378         if (cc->nr_freepages > 0) {
2379                 unsigned long free_pfn = release_freepages(&cc->freepages);
2380
2381                 cc->nr_freepages = 0;
2382                 VM_BUG_ON(free_pfn == 0);
2383                 /* The cached pfn is always the first in a pageblock */
2384                 free_pfn = pageblock_start_pfn(free_pfn);
2385                 /*
2386                  * Only go back, not forward. The cached pfn might have been
2387                  * already reset to zone end in compact_finished()
2388                  */
2389                 if (free_pfn > cc->zone->compact_cached_free_pfn)
2390                         cc->zone->compact_cached_free_pfn = free_pfn;
2391         }
2392
2393         count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned);
2394         count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned);
2395
2396         trace_mm_compaction_end(start_pfn, cc->migrate_pfn,
2397                                 cc->free_pfn, end_pfn, sync, ret);
2398
2399         return ret;
2400 }
2401
2402 static enum compact_result compact_zone_order(struct zone *zone, int order,
2403                 gfp_t gfp_mask, enum compact_priority prio,
2404                 unsigned int alloc_flags, int highest_zoneidx,
2405                 struct page **capture)
2406 {
2407         enum compact_result ret;
2408         struct compact_control cc = {
2409                 .order = order,
2410                 .search_order = order,
2411                 .gfp_mask = gfp_mask,
2412                 .zone = zone,
2413                 .mode = (prio == COMPACT_PRIO_ASYNC) ?
2414                                         MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT,
2415                 .alloc_flags = alloc_flags,
2416                 .highest_zoneidx = highest_zoneidx,
2417                 .direct_compaction = true,
2418                 .whole_zone = (prio == MIN_COMPACT_PRIORITY),
2419                 .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
2420                 .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
2421         };
2422         struct capture_control capc = {
2423                 .cc = &cc,
2424                 .page = NULL,
2425         };
2426
2427         /*
2428          * Make sure the structs are really initialized before we expose the
2429          * capture control, in case we are interrupted and the interrupt handler
2430          * frees a page.
2431          */
2432         barrier();
2433         WRITE_ONCE(current->capture_control, &capc);
2434
2435         ret = compact_zone(&cc, &capc);
2436
2437         VM_BUG_ON(!list_empty(&cc.freepages));
2438         VM_BUG_ON(!list_empty(&cc.migratepages));
2439
2440         /*
2441          * Make sure we hide capture control first before we read the captured
2442          * page pointer, otherwise an interrupt could free and capture a page
2443          * and we would leak it.
2444          */
2445         WRITE_ONCE(current->capture_control, NULL);
2446         *capture = READ_ONCE(capc.page);
2447
2448         return ret;
2449 }
2450
2451 int sysctl_extfrag_threshold = 500;
2452
2453 /**
2454  * try_to_compact_pages - Direct compact to satisfy a high-order allocation
2455  * @gfp_mask: The GFP mask of the current allocation
2456  * @order: The order of the current allocation
2457  * @alloc_flags: The allocation flags of the current allocation
2458  * @ac: The context of current allocation
2459  * @prio: Determines how hard direct compaction should try to succeed
2460  * @capture: Pointer to free page created by compaction will be stored here
2461  *
2462  * This is the main entry point for direct page compaction.
2463  */
2464 enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
2465                 unsigned int alloc_flags, const struct alloc_context *ac,
2466                 enum compact_priority prio, struct page **capture)
2467 {
2468         int may_perform_io = gfp_mask & __GFP_IO;
2469         struct zoneref *z;
2470         struct zone *zone;
2471         enum compact_result rc = COMPACT_SKIPPED;
2472
2473         /*
2474          * Check if the GFP flags allow compaction - GFP_NOIO is really
2475          * tricky context because the migration might require IO
2476          */
2477         if (!may_perform_io)
2478                 return COMPACT_SKIPPED;
2479
2480         trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
2481
2482         /* Compact each zone in the list */
2483         for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2484                                         ac->highest_zoneidx, ac->nodemask) {
2485                 enum compact_result status;
2486
2487                 if (prio > MIN_COMPACT_PRIORITY
2488                                         && compaction_deferred(zone, order)) {
2489                         rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
2490                         continue;
2491                 }
2492
2493                 status = compact_zone_order(zone, order, gfp_mask, prio,
2494                                 alloc_flags, ac->highest_zoneidx, capture);
2495                 rc = max(status, rc);
2496
2497                 /* The allocation should succeed, stop compacting */
2498                 if (status == COMPACT_SUCCESS) {
2499                         /*
2500                          * We think the allocation will succeed in this zone,
2501                          * but it is not certain, hence the false. The caller
2502                          * will repeat this with true if allocation indeed
2503                          * succeeds in this zone.
2504                          */
2505                         compaction_defer_reset(zone, order, false);
2506
2507                         break;
2508                 }
2509
2510                 if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
2511                                         status == COMPACT_PARTIAL_SKIPPED))
2512                         /*
2513                          * We think that allocation won't succeed in this zone
2514                          * so we defer compaction there. If it ends up
2515                          * succeeding after all, it will be reset.
2516                          */
2517                         defer_compaction(zone, order);
2518
2519                 /*
2520                  * We might have stopped compacting due to need_resched() in
2521                  * async compaction, or due to a fatal signal detected. In that
2522                  * case do not try further zones
2523                  */
2524                 if ((prio == COMPACT_PRIO_ASYNC && need_resched())
2525                                         || fatal_signal_pending(current))
2526                         break;
2527         }
2528
2529         return rc;
2530 }
2531
2532 /*
2533  * Compact all zones within a node till each zone's fragmentation score
2534  * reaches within proactive compaction thresholds (as determined by the
2535  * proactiveness tunable).
2536  *
2537  * It is possible that the function returns before reaching score targets
2538  * due to various back-off conditions, such as, contention on per-node or
2539  * per-zone locks.
2540  */
2541 static void proactive_compact_node(pg_data_t *pgdat)
2542 {
2543         int zoneid;
2544         struct zone *zone;
2545         struct compact_control cc = {
2546                 .order = -1,
2547                 .mode = MIGRATE_SYNC_LIGHT,
2548                 .ignore_skip_hint = true,
2549                 .whole_zone = true,
2550                 .gfp_mask = GFP_KERNEL,
2551                 .proactive_compaction = true,
2552         };
2553
2554         for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2555                 zone = &pgdat->node_zones[zoneid];
2556                 if (!populated_zone(zone))
2557                         continue;
2558
2559                 cc.zone = zone;
2560
2561                 compact_zone(&cc, NULL);
2562
2563                 VM_BUG_ON(!list_empty(&cc.freepages));
2564                 VM_BUG_ON(!list_empty(&cc.migratepages));
2565         }
2566 }
2567
2568 /* Compact all zones within a node */
2569 static void compact_node(int nid)
2570 {
2571         pg_data_t *pgdat = NODE_DATA(nid);
2572         int zoneid;
2573         struct zone *zone;
2574         struct compact_control cc = {
2575                 .order = -1,
2576                 .mode = MIGRATE_SYNC,
2577                 .ignore_skip_hint = true,
2578                 .whole_zone = true,
2579                 .gfp_mask = GFP_KERNEL,
2580         };
2581
2582
2583         for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2584
2585                 zone = &pgdat->node_zones[zoneid];
2586                 if (!populated_zone(zone))
2587                         continue;
2588
2589                 cc.zone = zone;
2590
2591                 compact_zone(&cc, NULL);
2592
2593                 VM_BUG_ON(!list_empty(&cc.freepages));
2594                 VM_BUG_ON(!list_empty(&cc.migratepages));
2595         }
2596 }
2597
2598 /* Compact all nodes in the system */
2599 static void compact_nodes(void)
2600 {
2601         int nid;
2602
2603         /* Flush pending updates to the LRU lists */
2604         lru_add_drain_all();
2605
2606         for_each_online_node(nid)
2607                 compact_node(nid);
2608 }
2609
2610 /* The written value is actually unused, all memory is compacted */
2611 int sysctl_compact_memory;
2612
2613 /*
2614  * Tunable for proactive compaction. It determines how
2615  * aggressively the kernel should compact memory in the
2616  * background. It takes values in the range [0, 100].
2617  */
2618 unsigned int __read_mostly sysctl_compaction_proactiveness = 20;
2619
2620 /*
2621  * This is the entry point for compacting all nodes via
2622  * /proc/sys/vm/compact_memory
2623  */
2624 int sysctl_compaction_handler(struct ctl_table *table, int write,
2625                         void *buffer, size_t *length, loff_t *ppos)
2626 {
2627         if (write)
2628                 compact_nodes();
2629
2630         return 0;
2631 }
2632
2633 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
2634 static ssize_t sysfs_compact_node(struct device *dev,
2635                         struct device_attribute *attr,
2636                         const char *buf, size_t count)
2637 {
2638         int nid = dev->id;
2639
2640         if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
2641                 /* Flush pending updates to the LRU lists */
2642                 lru_add_drain_all();
2643
2644                 compact_node(nid);
2645         }
2646
2647         return count;
2648 }
2649 static DEVICE_ATTR(compact, 0200, NULL, sysfs_compact_node);
2650
2651 int compaction_register_node(struct node *node)
2652 {
2653         return device_create_file(&node->dev, &dev_attr_compact);
2654 }
2655
2656 void compaction_unregister_node(struct node *node)
2657 {
2658         return device_remove_file(&node->dev, &dev_attr_compact);
2659 }
2660 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
2661
2662 static inline bool kcompactd_work_requested(pg_data_t *pgdat)
2663 {
2664         return pgdat->kcompactd_max_order > 0 || kthread_should_stop();
2665 }
2666
2667 static bool kcompactd_node_suitable(pg_data_t *pgdat)
2668 {
2669         int zoneid;
2670         struct zone *zone;
2671         enum zone_type highest_zoneidx = pgdat->kcompactd_highest_zoneidx;
2672
2673         for (zoneid = 0; zoneid <= highest_zoneidx; zoneid++) {
2674                 zone = &pgdat->node_zones[zoneid];
2675
2676                 if (!populated_zone(zone))
2677                         continue;
2678
2679                 if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0,
2680                                         highest_zoneidx) == COMPACT_CONTINUE)
2681                         return true;
2682         }
2683
2684         return false;
2685 }
2686
2687 static void kcompactd_do_work(pg_data_t *pgdat)
2688 {
2689         /*
2690          * With no special task, compact all zones so that a page of requested
2691          * order is allocatable.
2692          */
2693         int zoneid;
2694         struct zone *zone;
2695         struct compact_control cc = {
2696                 .order = pgdat->kcompactd_max_order,
2697                 .search_order = pgdat->kcompactd_max_order,
2698                 .highest_zoneidx = pgdat->kcompactd_highest_zoneidx,
2699                 .mode = MIGRATE_SYNC_LIGHT,
2700                 .ignore_skip_hint = false,
2701                 .gfp_mask = GFP_KERNEL,
2702         };
2703         trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
2704                                                         cc.highest_zoneidx);
2705         count_compact_event(KCOMPACTD_WAKE);
2706
2707         for (zoneid = 0; zoneid <= cc.highest_zoneidx; zoneid++) {
2708                 int status;
2709
2710                 zone = &pgdat->node_zones[zoneid];
2711                 if (!populated_zone(zone))
2712                         continue;
2713
2714                 if (compaction_deferred(zone, cc.order))
2715                         continue;
2716
2717                 if (compaction_suitable(zone, cc.order, 0, zoneid) !=
2718                                                         COMPACT_CONTINUE)
2719                         continue;
2720
2721                 if (kthread_should_stop())
2722                         return;
2723
2724                 cc.zone = zone;
2725                 status = compact_zone(&cc, NULL);
2726
2727                 if (status == COMPACT_SUCCESS) {
2728                         compaction_defer_reset(zone, cc.order, false);
2729                 } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
2730                         /*
2731                          * Buddy pages may become stranded on pcps that could
2732                          * otherwise coalesce on the zone's free area for
2733                          * order >= cc.order.  This is ratelimited by the
2734                          * upcoming deferral.
2735                          */
2736                         drain_all_pages(zone);
2737
2738                         /*
2739                          * We use sync migration mode here, so we defer like
2740                          * sync direct compaction does.
2741                          */
2742                         defer_compaction(zone, cc.order);
2743                 }
2744
2745                 count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
2746                                      cc.total_migrate_scanned);
2747                 count_compact_events(KCOMPACTD_FREE_SCANNED,
2748                                      cc.total_free_scanned);
2749
2750                 VM_BUG_ON(!list_empty(&cc.freepages));
2751                 VM_BUG_ON(!list_empty(&cc.migratepages));
2752         }
2753
2754         /*
2755          * Regardless of success, we are done until woken up next. But remember
2756          * the requested order/highest_zoneidx in case it was higher/tighter
2757          * than our current ones
2758          */
2759         if (pgdat->kcompactd_max_order <= cc.order)
2760                 pgdat->kcompactd_max_order = 0;
2761         if (pgdat->kcompactd_highest_zoneidx >= cc.highest_zoneidx)
2762                 pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
2763 }
2764
2765 void wakeup_kcompactd(pg_data_t *pgdat, int order, int highest_zoneidx)
2766 {
2767         if (!order)
2768                 return;
2769
2770         if (pgdat->kcompactd_max_order < order)
2771                 pgdat->kcompactd_max_order = order;
2772
2773         if (pgdat->kcompactd_highest_zoneidx > highest_zoneidx)
2774                 pgdat->kcompactd_highest_zoneidx = highest_zoneidx;
2775
2776         /*
2777          * Pairs with implicit barrier in wait_event_freezable()
2778          * such that wakeups are not missed.
2779          */
2780         if (!wq_has_sleeper(&pgdat->kcompactd_wait))
2781                 return;
2782
2783         if (!kcompactd_node_suitable(pgdat))
2784                 return;
2785
2786         trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
2787                                                         highest_zoneidx);
2788         wake_up_interruptible(&pgdat->kcompactd_wait);
2789 }
2790
2791 /*
2792  * The background compaction daemon, started as a kernel thread
2793  * from the init process.
2794  */
2795 static int kcompactd(void *p)
2796 {
2797         pg_data_t *pgdat = (pg_data_t*)p;
2798         struct task_struct *tsk = current;
2799         unsigned int proactive_defer = 0;
2800
2801         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2802
2803         if (!cpumask_empty(cpumask))
2804                 set_cpus_allowed_ptr(tsk, cpumask);
2805
2806         set_freezable();
2807
2808         pgdat->kcompactd_max_order = 0;
2809         pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
2810
2811         while (!kthread_should_stop()) {
2812                 unsigned long pflags;
2813
2814                 trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
2815                 if (wait_event_freezable_timeout(pgdat->kcompactd_wait,
2816                         kcompactd_work_requested(pgdat),
2817                         msecs_to_jiffies(HPAGE_FRAG_CHECK_INTERVAL_MSEC))) {
2818
2819                         psi_memstall_enter(&pflags);
2820                         kcompactd_do_work(pgdat);
2821                         psi_memstall_leave(&pflags);
2822                         continue;
2823                 }
2824
2825                 /* kcompactd wait timeout */
2826                 if (should_proactive_compact_node(pgdat)) {
2827                         unsigned int prev_score, score;
2828
2829                         if (proactive_defer) {
2830                                 proactive_defer--;
2831                                 continue;
2832                         }
2833                         prev_score = fragmentation_score_node(pgdat);
2834                         proactive_compact_node(pgdat);
2835                         score = fragmentation_score_node(pgdat);
2836                         /*
2837                          * Defer proactive compaction if the fragmentation
2838                          * score did not go down i.e. no progress made.
2839                          */
2840                         proactive_defer = score < prev_score ?
2841                                         0 : 1 << COMPACT_MAX_DEFER_SHIFT;
2842                 }
2843         }
2844
2845         return 0;
2846 }
2847
2848 /*
2849  * This kcompactd start function will be called by init and node-hot-add.
2850  * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
2851  */
2852 int kcompactd_run(int nid)
2853 {
2854         pg_data_t *pgdat = NODE_DATA(nid);
2855         int ret = 0;
2856
2857         if (pgdat->kcompactd)
2858                 return 0;
2859
2860         pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
2861         if (IS_ERR(pgdat->kcompactd)) {
2862                 pr_err("Failed to start kcompactd on node %d\n", nid);
2863                 ret = PTR_ERR(pgdat->kcompactd);
2864                 pgdat->kcompactd = NULL;
2865         }
2866         return ret;
2867 }
2868
2869 /*
2870  * Called by memory hotplug when all memory in a node is offlined. Caller must
2871  * hold mem_hotplug_begin/end().
2872  */
2873 void kcompactd_stop(int nid)
2874 {
2875         struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
2876
2877         if (kcompactd) {
2878                 kthread_stop(kcompactd);
2879                 NODE_DATA(nid)->kcompactd = NULL;
2880         }
2881 }
2882
2883 /*
2884  * It's optimal to keep kcompactd on the same CPUs as their memory, but
2885  * not required for correctness. So if the last cpu in a node goes
2886  * away, we get changed to run anywhere: as the first one comes back,
2887  * restore their cpu bindings.
2888  */
2889 static int kcompactd_cpu_online(unsigned int cpu)
2890 {
2891         int nid;
2892
2893         for_each_node_state(nid, N_MEMORY) {
2894                 pg_data_t *pgdat = NODE_DATA(nid);
2895                 const struct cpumask *mask;
2896
2897                 mask = cpumask_of_node(pgdat->node_id);
2898
2899                 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2900                         /* One of our CPUs online: restore mask */
2901                         set_cpus_allowed_ptr(pgdat->kcompactd, mask);
2902         }
2903         return 0;
2904 }
2905
2906 static int __init kcompactd_init(void)
2907 {
2908         int nid;
2909         int ret;
2910
2911         ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
2912                                         "mm/compaction:online",
2913                                         kcompactd_cpu_online, NULL);
2914         if (ret < 0) {
2915                 pr_err("kcompactd: failed to register hotplug callbacks.\n");
2916                 return ret;
2917         }
2918
2919         for_each_node_state(nid, N_MEMORY)
2920                 kcompactd_run(nid);
2921         return 0;
2922 }
2923 subsys_initcall(kcompactd_init)
2924
2925 #endif /* CONFIG_COMPACTION */