Loading...
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/*
30 * Fragmentation score check interval for proactive compaction purposes.
31 */
32#define HPAGE_FRAG_CHECK_INTERVAL_MSEC (500)
33
34static inline void count_compact_event(enum vm_event_item item)
35{
36 count_vm_event(item);
37}
38
39static inline void count_compact_events(enum vm_event_item item, long delta)
40{
41 count_vm_events(item, delta);
42}
43#else
44#define count_compact_event(item) do { } while (0)
45#define count_compact_events(item, delta) do { } while (0)
46#endif
47
48#if defined CONFIG_COMPACTION || defined CONFIG_CMA
49
50#define CREATE_TRACE_POINTS
51#include <trace/events/compaction.h>
52
53#define block_start_pfn(pfn, order) round_down(pfn, 1UL << (order))
54#define block_end_pfn(pfn, order) ALIGN((pfn) + 1, 1UL << (order))
55
56/*
57 * Page order with-respect-to which proactive compaction
58 * calculates external fragmentation, which is used as
59 * the "fragmentation score" of a node/zone.
60 */
61#if defined CONFIG_TRANSPARENT_HUGEPAGE
62#define COMPACTION_HPAGE_ORDER HPAGE_PMD_ORDER
63#elif defined CONFIG_HUGETLBFS
64#define COMPACTION_HPAGE_ORDER HUGETLB_PAGE_ORDER
65#else
66#define COMPACTION_HPAGE_ORDER (PMD_SHIFT - PAGE_SHIFT)
67#endif
68
69static unsigned long release_freepages(struct list_head *freelist)
70{
71 struct page *page, *next;
72 unsigned long high_pfn = 0;
73
74 list_for_each_entry_safe(page, next, freelist, lru) {
75 unsigned long pfn = page_to_pfn(page);
76 list_del(&page->lru);
77 __free_page(page);
78 if (pfn > high_pfn)
79 high_pfn = pfn;
80 }
81
82 return high_pfn;
83}
84
85static void split_map_pages(struct list_head *list)
86{
87 unsigned int i, order, nr_pages;
88 struct page *page, *next;
89 LIST_HEAD(tmp_list);
90
91 list_for_each_entry_safe(page, next, list, lru) {
92 list_del(&page->lru);
93
94 order = page_private(page);
95 nr_pages = 1 << order;
96
97 post_alloc_hook(page, order, __GFP_MOVABLE);
98 if (order)
99 split_page(page, order);
100
101 for (i = 0; i < nr_pages; i++) {
102 list_add(&page->lru, &tmp_list);
103 page++;
104 }
105 }
106
107 list_splice(&tmp_list, list);
108}
109
110#ifdef CONFIG_COMPACTION
111bool PageMovable(struct page *page)
112{
113 const struct movable_operations *mops;
114
115 VM_BUG_ON_PAGE(!PageLocked(page), page);
116 if (!__PageMovable(page))
117 return false;
118
119 mops = page_movable_ops(page);
120 if (mops)
121 return true;
122
123 return false;
124}
125
126void __SetPageMovable(struct page *page, const struct movable_operations *mops)
127{
128 VM_BUG_ON_PAGE(!PageLocked(page), page);
129 VM_BUG_ON_PAGE((unsigned long)mops & PAGE_MAPPING_MOVABLE, page);
130 page->mapping = (void *)((unsigned long)mops | PAGE_MAPPING_MOVABLE);
131}
132EXPORT_SYMBOL(__SetPageMovable);
133
134void __ClearPageMovable(struct page *page)
135{
136 VM_BUG_ON_PAGE(!PageMovable(page), page);
137 /*
138 * This page still has the type of a movable page, but it's
139 * actually not movable any more.
140 */
141 page->mapping = (void *)PAGE_MAPPING_MOVABLE;
142}
143EXPORT_SYMBOL(__ClearPageMovable);
144
145/* Do not skip compaction more than 64 times */
146#define COMPACT_MAX_DEFER_SHIFT 6
147
148/*
149 * Compaction is deferred when compaction fails to result in a page
150 * allocation success. 1 << compact_defer_shift, compactions are skipped up
151 * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
152 */
153static void defer_compaction(struct zone *zone, int order)
154{
155 zone->compact_considered = 0;
156 zone->compact_defer_shift++;
157
158 if (order < zone->compact_order_failed)
159 zone->compact_order_failed = order;
160
161 if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
162 zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
163
164 trace_mm_compaction_defer_compaction(zone, order);
165}
166
167/* Returns true if compaction should be skipped this time */
168static bool compaction_deferred(struct zone *zone, int order)
169{
170 unsigned long defer_limit = 1UL << zone->compact_defer_shift;
171
172 if (order < zone->compact_order_failed)
173 return false;
174
175 /* Avoid possible overflow */
176 if (++zone->compact_considered >= defer_limit) {
177 zone->compact_considered = defer_limit;
178 return false;
179 }
180
181 trace_mm_compaction_deferred(zone, order);
182
183 return true;
184}
185
186/*
187 * Update defer tracking counters after successful compaction of given order,
188 * which means an allocation either succeeded (alloc_success == true) or is
189 * expected to succeed.
190 */
191void compaction_defer_reset(struct zone *zone, int order,
192 bool alloc_success)
193{
194 if (alloc_success) {
195 zone->compact_considered = 0;
196 zone->compact_defer_shift = 0;
197 }
198 if (order >= zone->compact_order_failed)
199 zone->compact_order_failed = order + 1;
200
201 trace_mm_compaction_defer_reset(zone, order);
202}
203
204/* Returns true if restarting compaction after many failures */
205static bool compaction_restarting(struct zone *zone, int order)
206{
207 if (order < zone->compact_order_failed)
208 return false;
209
210 return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
211 zone->compact_considered >= 1UL << zone->compact_defer_shift;
212}
213
214/* Returns true if the pageblock should be scanned for pages to isolate. */
215static inline bool isolation_suitable(struct compact_control *cc,
216 struct page *page)
217{
218 if (cc->ignore_skip_hint)
219 return true;
220
221 return !get_pageblock_skip(page);
222}
223
224static void reset_cached_positions(struct zone *zone)
225{
226 zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
227 zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
228 zone->compact_cached_free_pfn =
229 pageblock_start_pfn(zone_end_pfn(zone) - 1);
230}
231
232#ifdef CONFIG_SPARSEMEM
233/*
234 * If the PFN falls into an offline section, return the start PFN of the
235 * next online section. If the PFN falls into an online section or if
236 * there is no next online section, return 0.
237 */
238static unsigned long skip_offline_sections(unsigned long start_pfn)
239{
240 unsigned long start_nr = pfn_to_section_nr(start_pfn);
241
242 if (online_section_nr(start_nr))
243 return 0;
244
245 while (++start_nr <= __highest_present_section_nr) {
246 if (online_section_nr(start_nr))
247 return section_nr_to_pfn(start_nr);
248 }
249
250 return 0;
251}
252
253/*
254 * If the PFN falls into an offline section, return the end PFN of the
255 * next online section in reverse. If the PFN falls into an online section
256 * or if there is no next online section in reverse, return 0.
257 */
258static unsigned long skip_offline_sections_reverse(unsigned long start_pfn)
259{
260 unsigned long start_nr = pfn_to_section_nr(start_pfn);
261
262 if (!start_nr || online_section_nr(start_nr))
263 return 0;
264
265 while (start_nr-- > 0) {
266 if (online_section_nr(start_nr))
267 return section_nr_to_pfn(start_nr) + PAGES_PER_SECTION;
268 }
269
270 return 0;
271}
272#else
273static unsigned long skip_offline_sections(unsigned long start_pfn)
274{
275 return 0;
276}
277
278static unsigned long skip_offline_sections_reverse(unsigned long start_pfn)
279{
280 return 0;
281}
282#endif
283
284/*
285 * Compound pages of >= pageblock_order should consistently be skipped until
286 * released. It is always pointless to compact pages of such order (if they are
287 * migratable), and the pageblocks they occupy cannot contain any free pages.
288 */
289static bool pageblock_skip_persistent(struct page *page)
290{
291 if (!PageCompound(page))
292 return false;
293
294 page = compound_head(page);
295
296 if (compound_order(page) >= pageblock_order)
297 return true;
298
299 return false;
300}
301
302static bool
303__reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source,
304 bool check_target)
305{
306 struct page *page = pfn_to_online_page(pfn);
307 struct page *block_page;
308 struct page *end_page;
309 unsigned long block_pfn;
310
311 if (!page)
312 return false;
313 if (zone != page_zone(page))
314 return false;
315 if (pageblock_skip_persistent(page))
316 return false;
317
318 /*
319 * If skip is already cleared do no further checking once the
320 * restart points have been set.
321 */
322 if (check_source && check_target && !get_pageblock_skip(page))
323 return true;
324
325 /*
326 * If clearing skip for the target scanner, do not select a
327 * non-movable pageblock as the starting point.
328 */
329 if (!check_source && check_target &&
330 get_pageblock_migratetype(page) != MIGRATE_MOVABLE)
331 return false;
332
333 /* Ensure the start of the pageblock or zone is online and valid */
334 block_pfn = pageblock_start_pfn(pfn);
335 block_pfn = max(block_pfn, zone->zone_start_pfn);
336 block_page = pfn_to_online_page(block_pfn);
337 if (block_page) {
338 page = block_page;
339 pfn = block_pfn;
340 }
341
342 /* Ensure the end of the pageblock or zone is online and valid */
343 block_pfn = pageblock_end_pfn(pfn) - 1;
344 block_pfn = min(block_pfn, zone_end_pfn(zone) - 1);
345 end_page = pfn_to_online_page(block_pfn);
346 if (!end_page)
347 return false;
348
349 /*
350 * Only clear the hint if a sample indicates there is either a
351 * free page or an LRU page in the block. One or other condition
352 * is necessary for the block to be a migration source/target.
353 */
354 do {
355 if (check_source && PageLRU(page)) {
356 clear_pageblock_skip(page);
357 return true;
358 }
359
360 if (check_target && PageBuddy(page)) {
361 clear_pageblock_skip(page);
362 return true;
363 }
364
365 page += (1 << PAGE_ALLOC_COSTLY_ORDER);
366 } while (page <= end_page);
367
368 return false;
369}
370
371/*
372 * This function is called to clear all cached information on pageblocks that
373 * should be skipped for page isolation when the migrate and free page scanner
374 * meet.
375 */
376static void __reset_isolation_suitable(struct zone *zone)
377{
378 unsigned long migrate_pfn = zone->zone_start_pfn;
379 unsigned long free_pfn = zone_end_pfn(zone) - 1;
380 unsigned long reset_migrate = free_pfn;
381 unsigned long reset_free = migrate_pfn;
382 bool source_set = false;
383 bool free_set = false;
384
385 /* Only flush if a full compaction finished recently */
386 if (!zone->compact_blockskip_flush)
387 return;
388
389 zone->compact_blockskip_flush = false;
390
391 /*
392 * Walk the zone and update pageblock skip information. Source looks
393 * for PageLRU while target looks for PageBuddy. When the scanner
394 * is found, both PageBuddy and PageLRU are checked as the pageblock
395 * is suitable as both source and target.
396 */
397 for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages,
398 free_pfn -= pageblock_nr_pages) {
399 cond_resched();
400
401 /* Update the migrate PFN */
402 if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) &&
403 migrate_pfn < reset_migrate) {
404 source_set = true;
405 reset_migrate = migrate_pfn;
406 zone->compact_init_migrate_pfn = reset_migrate;
407 zone->compact_cached_migrate_pfn[0] = reset_migrate;
408 zone->compact_cached_migrate_pfn[1] = reset_migrate;
409 }
410
411 /* Update the free PFN */
412 if (__reset_isolation_pfn(zone, free_pfn, free_set, true) &&
413 free_pfn > reset_free) {
414 free_set = true;
415 reset_free = free_pfn;
416 zone->compact_init_free_pfn = reset_free;
417 zone->compact_cached_free_pfn = reset_free;
418 }
419 }
420
421 /* Leave no distance if no suitable block was reset */
422 if (reset_migrate >= reset_free) {
423 zone->compact_cached_migrate_pfn[0] = migrate_pfn;
424 zone->compact_cached_migrate_pfn[1] = migrate_pfn;
425 zone->compact_cached_free_pfn = free_pfn;
426 }
427}
428
429void reset_isolation_suitable(pg_data_t *pgdat)
430{
431 int zoneid;
432
433 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
434 struct zone *zone = &pgdat->node_zones[zoneid];
435 if (!populated_zone(zone))
436 continue;
437
438 __reset_isolation_suitable(zone);
439 }
440}
441
442/*
443 * Sets the pageblock skip bit if it was clear. Note that this is a hint as
444 * locks are not required for read/writers. Returns true if it was already set.
445 */
446static bool test_and_set_skip(struct compact_control *cc, struct page *page)
447{
448 bool skip;
449
450 /* Do not update if skip hint is being ignored */
451 if (cc->ignore_skip_hint)
452 return false;
453
454 skip = get_pageblock_skip(page);
455 if (!skip && !cc->no_set_skip_hint)
456 set_pageblock_skip(page);
457
458 return skip;
459}
460
461static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
462{
463 struct zone *zone = cc->zone;
464
465 /* Set for isolation rather than compaction */
466 if (cc->no_set_skip_hint)
467 return;
468
469 pfn = pageblock_end_pfn(pfn);
470
471 /* Update where async and sync compaction should restart */
472 if (pfn > zone->compact_cached_migrate_pfn[0])
473 zone->compact_cached_migrate_pfn[0] = pfn;
474 if (cc->mode != MIGRATE_ASYNC &&
475 pfn > zone->compact_cached_migrate_pfn[1])
476 zone->compact_cached_migrate_pfn[1] = pfn;
477}
478
479/*
480 * If no pages were isolated then mark this pageblock to be skipped in the
481 * future. The information is later cleared by __reset_isolation_suitable().
482 */
483static void update_pageblock_skip(struct compact_control *cc,
484 struct page *page, unsigned long pfn)
485{
486 struct zone *zone = cc->zone;
487
488 if (cc->no_set_skip_hint)
489 return;
490
491 set_pageblock_skip(page);
492
493 if (pfn < zone->compact_cached_free_pfn)
494 zone->compact_cached_free_pfn = pfn;
495}
496#else
497static inline bool isolation_suitable(struct compact_control *cc,
498 struct page *page)
499{
500 return true;
501}
502
503static inline bool pageblock_skip_persistent(struct page *page)
504{
505 return false;
506}
507
508static inline void update_pageblock_skip(struct compact_control *cc,
509 struct page *page, unsigned long pfn)
510{
511}
512
513static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
514{
515}
516
517static bool test_and_set_skip(struct compact_control *cc, struct page *page)
518{
519 return false;
520}
521#endif /* CONFIG_COMPACTION */
522
523/*
524 * Compaction requires the taking of some coarse locks that are potentially
525 * very heavily contended. For async compaction, trylock and record if the
526 * lock is contended. The lock will still be acquired but compaction will
527 * abort when the current block is finished regardless of success rate.
528 * Sync compaction acquires the lock.
529 *
530 * Always returns true which makes it easier to track lock state in callers.
531 */
532static bool compact_lock_irqsave(spinlock_t *lock, unsigned long *flags,
533 struct compact_control *cc)
534 __acquires(lock)
535{
536 /* Track if the lock is contended in async mode */
537 if (cc->mode == MIGRATE_ASYNC && !cc->contended) {
538 if (spin_trylock_irqsave(lock, *flags))
539 return true;
540
541 cc->contended = true;
542 }
543
544 spin_lock_irqsave(lock, *flags);
545 return true;
546}
547
548/*
549 * Compaction requires the taking of some coarse locks that are potentially
550 * very heavily contended. The lock should be periodically unlocked to avoid
551 * having disabled IRQs for a long time, even when there is nobody waiting on
552 * the lock. It might also be that allowing the IRQs will result in
553 * need_resched() becoming true. If scheduling is needed, compaction schedules.
554 * Either compaction type will also abort if a fatal signal is pending.
555 * In either case if the lock was locked, it is dropped and not regained.
556 *
557 * Returns true if compaction should abort due to fatal signal pending.
558 * Returns false when compaction can continue.
559 */
560static bool compact_unlock_should_abort(spinlock_t *lock,
561 unsigned long flags, bool *locked, struct compact_control *cc)
562{
563 if (*locked) {
564 spin_unlock_irqrestore(lock, flags);
565 *locked = false;
566 }
567
568 if (fatal_signal_pending(current)) {
569 cc->contended = true;
570 return true;
571 }
572
573 cond_resched();
574
575 return false;
576}
577
578/*
579 * Isolate free pages onto a private freelist. If @strict is true, will abort
580 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
581 * (even though it may still end up isolating some pages).
582 */
583static unsigned long isolate_freepages_block(struct compact_control *cc,
584 unsigned long *start_pfn,
585 unsigned long end_pfn,
586 struct list_head *freelist,
587 unsigned int stride,
588 bool strict)
589{
590 int nr_scanned = 0, total_isolated = 0;
591 struct page *page;
592 unsigned long flags = 0;
593 bool locked = false;
594 unsigned long blockpfn = *start_pfn;
595 unsigned int order;
596
597 /* Strict mode is for isolation, speed is secondary */
598 if (strict)
599 stride = 1;
600
601 page = pfn_to_page(blockpfn);
602
603 /* Isolate free pages. */
604 for (; blockpfn < end_pfn; blockpfn += stride, page += stride) {
605 int isolated;
606
607 /*
608 * Periodically drop the lock (if held) regardless of its
609 * contention, to give chance to IRQs. Abort if fatal signal
610 * pending.
611 */
612 if (!(blockpfn % COMPACT_CLUSTER_MAX)
613 && compact_unlock_should_abort(&cc->zone->lock, flags,
614 &locked, cc))
615 break;
616
617 nr_scanned++;
618
619 /*
620 * For compound pages such as THP and hugetlbfs, we can save
621 * potentially a lot of iterations if we skip them at once.
622 * The check is racy, but we can consider only valid values
623 * and the only danger is skipping too much.
624 */
625 if (PageCompound(page)) {
626 const unsigned int order = compound_order(page);
627
628 if (blockpfn + (1UL << order) <= end_pfn) {
629 blockpfn += (1UL << order) - 1;
630 page += (1UL << order) - 1;
631 nr_scanned += (1UL << order) - 1;
632 }
633
634 goto isolate_fail;
635 }
636
637 if (!PageBuddy(page))
638 goto isolate_fail;
639
640 /* If we already hold the lock, we can skip some rechecking. */
641 if (!locked) {
642 locked = compact_lock_irqsave(&cc->zone->lock,
643 &flags, cc);
644
645 /* Recheck this is a buddy page under lock */
646 if (!PageBuddy(page))
647 goto isolate_fail;
648 }
649
650 /* Found a free page, will break it into order-0 pages */
651 order = buddy_order(page);
652 isolated = __isolate_free_page(page, order);
653 if (!isolated)
654 break;
655 set_page_private(page, order);
656
657 nr_scanned += isolated - 1;
658 total_isolated += isolated;
659 cc->nr_freepages += isolated;
660 list_add_tail(&page->lru, freelist);
661
662 if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
663 blockpfn += isolated;
664 break;
665 }
666 /* Advance to the end of split page */
667 blockpfn += isolated - 1;
668 page += isolated - 1;
669 continue;
670
671isolate_fail:
672 if (strict)
673 break;
674
675 }
676
677 if (locked)
678 spin_unlock_irqrestore(&cc->zone->lock, flags);
679
680 /*
681 * Be careful to not go outside of the pageblock.
682 */
683 if (unlikely(blockpfn > end_pfn))
684 blockpfn = end_pfn;
685
686 trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
687 nr_scanned, total_isolated);
688
689 /* Record how far we have got within the block */
690 *start_pfn = blockpfn;
691
692 /*
693 * If strict isolation is requested by CMA then check that all the
694 * pages requested were isolated. If there were any failures, 0 is
695 * returned and CMA will fail.
696 */
697 if (strict && blockpfn < end_pfn)
698 total_isolated = 0;
699
700 cc->total_free_scanned += nr_scanned;
701 if (total_isolated)
702 count_compact_events(COMPACTISOLATED, total_isolated);
703 return total_isolated;
704}
705
706/**
707 * isolate_freepages_range() - isolate free pages.
708 * @cc: Compaction control structure.
709 * @start_pfn: The first PFN to start isolating.
710 * @end_pfn: The one-past-last PFN.
711 *
712 * Non-free pages, invalid PFNs, or zone boundaries within the
713 * [start_pfn, end_pfn) range are considered errors, cause function to
714 * undo its actions and return zero.
715 *
716 * Otherwise, function returns one-past-the-last PFN of isolated page
717 * (which may be greater then end_pfn if end fell in a middle of
718 * a free page).
719 */
720unsigned long
721isolate_freepages_range(struct compact_control *cc,
722 unsigned long start_pfn, unsigned long end_pfn)
723{
724 unsigned long isolated, pfn, block_start_pfn, block_end_pfn;
725 LIST_HEAD(freelist);
726
727 pfn = start_pfn;
728 block_start_pfn = pageblock_start_pfn(pfn);
729 if (block_start_pfn < cc->zone->zone_start_pfn)
730 block_start_pfn = cc->zone->zone_start_pfn;
731 block_end_pfn = pageblock_end_pfn(pfn);
732
733 for (; pfn < end_pfn; pfn += isolated,
734 block_start_pfn = block_end_pfn,
735 block_end_pfn += pageblock_nr_pages) {
736 /* Protect pfn from changing by isolate_freepages_block */
737 unsigned long isolate_start_pfn = pfn;
738
739 /*
740 * pfn could pass the block_end_pfn if isolated freepage
741 * is more than pageblock order. In this case, we adjust
742 * scanning range to right one.
743 */
744 if (pfn >= block_end_pfn) {
745 block_start_pfn = pageblock_start_pfn(pfn);
746 block_end_pfn = pageblock_end_pfn(pfn);
747 }
748
749 block_end_pfn = min(block_end_pfn, end_pfn);
750
751 if (!pageblock_pfn_to_page(block_start_pfn,
752 block_end_pfn, cc->zone))
753 break;
754
755 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
756 block_end_pfn, &freelist, 0, true);
757
758 /*
759 * In strict mode, isolate_freepages_block() returns 0 if
760 * there are any holes in the block (ie. invalid PFNs or
761 * non-free pages).
762 */
763 if (!isolated)
764 break;
765
766 /*
767 * If we managed to isolate pages, it is always (1 << n) *
768 * pageblock_nr_pages for some non-negative n. (Max order
769 * page may span two pageblocks).
770 */
771 }
772
773 /* __isolate_free_page() does not map the pages */
774 split_map_pages(&freelist);
775
776 if (pfn < end_pfn) {
777 /* Loop terminated early, cleanup. */
778 release_freepages(&freelist);
779 return 0;
780 }
781
782 /* We don't use freelists for anything. */
783 return pfn;
784}
785
786/* Similar to reclaim, but different enough that they don't share logic */
787static bool too_many_isolated(struct compact_control *cc)
788{
789 pg_data_t *pgdat = cc->zone->zone_pgdat;
790 bool too_many;
791
792 unsigned long active, inactive, isolated;
793
794 inactive = node_page_state(pgdat, NR_INACTIVE_FILE) +
795 node_page_state(pgdat, NR_INACTIVE_ANON);
796 active = node_page_state(pgdat, NR_ACTIVE_FILE) +
797 node_page_state(pgdat, NR_ACTIVE_ANON);
798 isolated = node_page_state(pgdat, NR_ISOLATED_FILE) +
799 node_page_state(pgdat, NR_ISOLATED_ANON);
800
801 /*
802 * Allow GFP_NOFS to isolate past the limit set for regular
803 * compaction runs. This prevents an ABBA deadlock when other
804 * compactors have already isolated to the limit, but are
805 * blocked on filesystem locks held by the GFP_NOFS thread.
806 */
807 if (cc->gfp_mask & __GFP_FS) {
808 inactive >>= 3;
809 active >>= 3;
810 }
811
812 too_many = isolated > (inactive + active) / 2;
813 if (!too_many)
814 wake_throttle_isolated(pgdat);
815
816 return too_many;
817}
818
819/**
820 * isolate_migratepages_block() - isolate all migrate-able pages within
821 * a single pageblock
822 * @cc: Compaction control structure.
823 * @low_pfn: The first PFN to isolate
824 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
825 * @mode: Isolation mode to be used.
826 *
827 * Isolate all pages that can be migrated from the range specified by
828 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
829 * Returns errno, like -EAGAIN or -EINTR in case e.g signal pending or congestion,
830 * -ENOMEM in case we could not allocate a page, or 0.
831 * cc->migrate_pfn will contain the next pfn to scan.
832 *
833 * The pages are isolated on cc->migratepages list (not required to be empty),
834 * and cc->nr_migratepages is updated accordingly.
835 */
836static int
837isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
838 unsigned long end_pfn, isolate_mode_t mode)
839{
840 pg_data_t *pgdat = cc->zone->zone_pgdat;
841 unsigned long nr_scanned = 0, nr_isolated = 0;
842 struct lruvec *lruvec;
843 unsigned long flags = 0;
844 struct lruvec *locked = NULL;
845 struct folio *folio = NULL;
846 struct page *page = NULL, *valid_page = NULL;
847 struct address_space *mapping;
848 unsigned long start_pfn = low_pfn;
849 bool skip_on_failure = false;
850 unsigned long next_skip_pfn = 0;
851 bool skip_updated = false;
852 int ret = 0;
853
854 cc->migrate_pfn = low_pfn;
855
856 /*
857 * Ensure that there are not too many pages isolated from the LRU
858 * list by either parallel reclaimers or compaction. If there are,
859 * delay for some time until fewer pages are isolated
860 */
861 while (unlikely(too_many_isolated(cc))) {
862 /* stop isolation if there are still pages not migrated */
863 if (cc->nr_migratepages)
864 return -EAGAIN;
865
866 /* async migration should just abort */
867 if (cc->mode == MIGRATE_ASYNC)
868 return -EAGAIN;
869
870 reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED);
871
872 if (fatal_signal_pending(current))
873 return -EINTR;
874 }
875
876 cond_resched();
877
878 if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
879 skip_on_failure = true;
880 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
881 }
882
883 /* Time to isolate some pages for migration */
884 for (; low_pfn < end_pfn; low_pfn++) {
885 bool is_dirty, is_unevictable;
886
887 if (skip_on_failure && low_pfn >= next_skip_pfn) {
888 /*
889 * We have isolated all migration candidates in the
890 * previous order-aligned block, and did not skip it due
891 * to failure. We should migrate the pages now and
892 * hopefully succeed compaction.
893 */
894 if (nr_isolated)
895 break;
896
897 /*
898 * We failed to isolate in the previous order-aligned
899 * block. Set the new boundary to the end of the
900 * current block. Note we can't simply increase
901 * next_skip_pfn by 1 << order, as low_pfn might have
902 * been incremented by a higher number due to skipping
903 * a compound or a high-order buddy page in the
904 * previous loop iteration.
905 */
906 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
907 }
908
909 /*
910 * Periodically drop the lock (if held) regardless of its
911 * contention, to give chance to IRQs. Abort completely if
912 * a fatal signal is pending.
913 */
914 if (!(low_pfn % COMPACT_CLUSTER_MAX)) {
915 if (locked) {
916 unlock_page_lruvec_irqrestore(locked, flags);
917 locked = NULL;
918 }
919
920 if (fatal_signal_pending(current)) {
921 cc->contended = true;
922 ret = -EINTR;
923
924 goto fatal_pending;
925 }
926
927 cond_resched();
928 }
929
930 nr_scanned++;
931
932 page = pfn_to_page(low_pfn);
933
934 /*
935 * Check if the pageblock has already been marked skipped.
936 * Only the first PFN is checked as the caller isolates
937 * COMPACT_CLUSTER_MAX at a time so the second call must
938 * not falsely conclude that the block should be skipped.
939 */
940 if (!valid_page && (pageblock_aligned(low_pfn) ||
941 low_pfn == cc->zone->zone_start_pfn)) {
942 if (!isolation_suitable(cc, page)) {
943 low_pfn = end_pfn;
944 folio = NULL;
945 goto isolate_abort;
946 }
947 valid_page = page;
948 }
949
950 if (PageHuge(page) && cc->alloc_contig) {
951 if (locked) {
952 unlock_page_lruvec_irqrestore(locked, flags);
953 locked = NULL;
954 }
955
956 ret = isolate_or_dissolve_huge_page(page, &cc->migratepages);
957
958 /*
959 * Fail isolation in case isolate_or_dissolve_huge_page()
960 * reports an error. In case of -ENOMEM, abort right away.
961 */
962 if (ret < 0) {
963 /* Do not report -EBUSY down the chain */
964 if (ret == -EBUSY)
965 ret = 0;
966 low_pfn += compound_nr(page) - 1;
967 nr_scanned += compound_nr(page) - 1;
968 goto isolate_fail;
969 }
970
971 if (PageHuge(page)) {
972 /*
973 * Hugepage was successfully isolated and placed
974 * on the cc->migratepages list.
975 */
976 folio = page_folio(page);
977 low_pfn += folio_nr_pages(folio) - 1;
978 goto isolate_success_no_list;
979 }
980
981 /*
982 * Ok, the hugepage was dissolved. Now these pages are
983 * Buddy and cannot be re-allocated because they are
984 * isolated. Fall-through as the check below handles
985 * Buddy pages.
986 */
987 }
988
989 /*
990 * Skip if free. We read page order here without zone lock
991 * which is generally unsafe, but the race window is small and
992 * the worst thing that can happen is that we skip some
993 * potential isolation targets.
994 */
995 if (PageBuddy(page)) {
996 unsigned long freepage_order = buddy_order_unsafe(page);
997
998 /*
999 * Without lock, we cannot be sure that what we got is
1000 * a valid page order. Consider only values in the
1001 * valid order range to prevent low_pfn overflow.
1002 */
1003 if (freepage_order > 0 && freepage_order <= MAX_PAGE_ORDER) {
1004 low_pfn += (1UL << freepage_order) - 1;
1005 nr_scanned += (1UL << freepage_order) - 1;
1006 }
1007 continue;
1008 }
1009
1010 /*
1011 * Regardless of being on LRU, compound pages such as THP and
1012 * hugetlbfs are not to be compacted unless we are attempting
1013 * an allocation much larger than the huge page size (eg CMA).
1014 * We can potentially save a lot of iterations if we skip them
1015 * at once. The check is racy, but we can consider only valid
1016 * values and the only danger is skipping too much.
1017 */
1018 if (PageCompound(page) && !cc->alloc_contig) {
1019 const unsigned int order = compound_order(page);
1020
1021 if (likely(order <= MAX_PAGE_ORDER)) {
1022 low_pfn += (1UL << order) - 1;
1023 nr_scanned += (1UL << order) - 1;
1024 }
1025 goto isolate_fail;
1026 }
1027
1028 /*
1029 * Check may be lockless but that's ok as we recheck later.
1030 * It's possible to migrate LRU and non-lru movable pages.
1031 * Skip any other type of page
1032 */
1033 if (!PageLRU(page)) {
1034 /*
1035 * __PageMovable can return false positive so we need
1036 * to verify it under page_lock.
1037 */
1038 if (unlikely(__PageMovable(page)) &&
1039 !PageIsolated(page)) {
1040 if (locked) {
1041 unlock_page_lruvec_irqrestore(locked, flags);
1042 locked = NULL;
1043 }
1044
1045 if (isolate_movable_page(page, mode)) {
1046 folio = page_folio(page);
1047 goto isolate_success;
1048 }
1049 }
1050
1051 goto isolate_fail;
1052 }
1053
1054 /*
1055 * Be careful not to clear PageLRU until after we're
1056 * sure the page is not being freed elsewhere -- the
1057 * page release code relies on it.
1058 */
1059 folio = folio_get_nontail_page(page);
1060 if (unlikely(!folio))
1061 goto isolate_fail;
1062
1063 /*
1064 * Migration will fail if an anonymous page is pinned in memory,
1065 * so avoid taking lru_lock and isolating it unnecessarily in an
1066 * admittedly racy check.
1067 */
1068 mapping = folio_mapping(folio);
1069 if (!mapping && (folio_ref_count(folio) - 1) > folio_mapcount(folio))
1070 goto isolate_fail_put;
1071
1072 /*
1073 * Only allow to migrate anonymous pages in GFP_NOFS context
1074 * because those do not depend on fs locks.
1075 */
1076 if (!(cc->gfp_mask & __GFP_FS) && mapping)
1077 goto isolate_fail_put;
1078
1079 /* Only take pages on LRU: a check now makes later tests safe */
1080 if (!folio_test_lru(folio))
1081 goto isolate_fail_put;
1082
1083 is_unevictable = folio_test_unevictable(folio);
1084
1085 /* Compaction might skip unevictable pages but CMA takes them */
1086 if (!(mode & ISOLATE_UNEVICTABLE) && is_unevictable)
1087 goto isolate_fail_put;
1088
1089 /*
1090 * To minimise LRU disruption, the caller can indicate with
1091 * ISOLATE_ASYNC_MIGRATE that it only wants to isolate pages
1092 * it will be able to migrate without blocking - clean pages
1093 * for the most part. PageWriteback would require blocking.
1094 */
1095 if ((mode & ISOLATE_ASYNC_MIGRATE) && folio_test_writeback(folio))
1096 goto isolate_fail_put;
1097
1098 is_dirty = folio_test_dirty(folio);
1099
1100 if (((mode & ISOLATE_ASYNC_MIGRATE) && is_dirty) ||
1101 (mapping && is_unevictable)) {
1102 bool migrate_dirty = true;
1103 bool is_unmovable;
1104
1105 /*
1106 * Only folios without mappings or that have
1107 * a ->migrate_folio callback are possible to migrate
1108 * without blocking.
1109 *
1110 * Folios from unmovable mappings are not migratable.
1111 *
1112 * However, we can be racing with truncation, which can
1113 * free the mapping that we need to check. Truncation
1114 * holds the folio lock until after the folio is removed
1115 * from the page so holding it ourselves is sufficient.
1116 *
1117 * To avoid locking the folio just to check unmovable,
1118 * assume every unmovable folio is also unevictable,
1119 * which is a cheaper test. If our assumption goes
1120 * wrong, it's not a correctness bug, just potentially
1121 * wasted cycles.
1122 */
1123 if (!folio_trylock(folio))
1124 goto isolate_fail_put;
1125
1126 mapping = folio_mapping(folio);
1127 if ((mode & ISOLATE_ASYNC_MIGRATE) && is_dirty) {
1128 migrate_dirty = !mapping ||
1129 mapping->a_ops->migrate_folio;
1130 }
1131 is_unmovable = mapping && mapping_unmovable(mapping);
1132 folio_unlock(folio);
1133 if (!migrate_dirty || is_unmovable)
1134 goto isolate_fail_put;
1135 }
1136
1137 /* Try isolate the folio */
1138 if (!folio_test_clear_lru(folio))
1139 goto isolate_fail_put;
1140
1141 lruvec = folio_lruvec(folio);
1142
1143 /* If we already hold the lock, we can skip some rechecking */
1144 if (lruvec != locked) {
1145 if (locked)
1146 unlock_page_lruvec_irqrestore(locked, flags);
1147
1148 compact_lock_irqsave(&lruvec->lru_lock, &flags, cc);
1149 locked = lruvec;
1150
1151 lruvec_memcg_debug(lruvec, folio);
1152
1153 /*
1154 * Try get exclusive access under lock. If marked for
1155 * skip, the scan is aborted unless the current context
1156 * is a rescan to reach the end of the pageblock.
1157 */
1158 if (!skip_updated && valid_page) {
1159 skip_updated = true;
1160 if (test_and_set_skip(cc, valid_page) &&
1161 !cc->finish_pageblock) {
1162 low_pfn = end_pfn;
1163 goto isolate_abort;
1164 }
1165 }
1166
1167 /*
1168 * folio become large since the non-locked check,
1169 * and it's on LRU.
1170 */
1171 if (unlikely(folio_test_large(folio) && !cc->alloc_contig)) {
1172 low_pfn += folio_nr_pages(folio) - 1;
1173 nr_scanned += folio_nr_pages(folio) - 1;
1174 folio_set_lru(folio);
1175 goto isolate_fail_put;
1176 }
1177 }
1178
1179 /* The folio is taken off the LRU */
1180 if (folio_test_large(folio))
1181 low_pfn += folio_nr_pages(folio) - 1;
1182
1183 /* Successfully isolated */
1184 lruvec_del_folio(lruvec, folio);
1185 node_stat_mod_folio(folio,
1186 NR_ISOLATED_ANON + folio_is_file_lru(folio),
1187 folio_nr_pages(folio));
1188
1189isolate_success:
1190 list_add(&folio->lru, &cc->migratepages);
1191isolate_success_no_list:
1192 cc->nr_migratepages += folio_nr_pages(folio);
1193 nr_isolated += folio_nr_pages(folio);
1194 nr_scanned += folio_nr_pages(folio) - 1;
1195
1196 /*
1197 * Avoid isolating too much unless this block is being
1198 * fully scanned (e.g. dirty/writeback pages, parallel allocation)
1199 * or a lock is contended. For contention, isolate quickly to
1200 * potentially remove one source of contention.
1201 */
1202 if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX &&
1203 !cc->finish_pageblock && !cc->contended) {
1204 ++low_pfn;
1205 break;
1206 }
1207
1208 continue;
1209
1210isolate_fail_put:
1211 /* Avoid potential deadlock in freeing page under lru_lock */
1212 if (locked) {
1213 unlock_page_lruvec_irqrestore(locked, flags);
1214 locked = NULL;
1215 }
1216 folio_put(folio);
1217
1218isolate_fail:
1219 if (!skip_on_failure && ret != -ENOMEM)
1220 continue;
1221
1222 /*
1223 * We have isolated some pages, but then failed. Release them
1224 * instead of migrating, as we cannot form the cc->order buddy
1225 * page anyway.
1226 */
1227 if (nr_isolated) {
1228 if (locked) {
1229 unlock_page_lruvec_irqrestore(locked, flags);
1230 locked = NULL;
1231 }
1232 putback_movable_pages(&cc->migratepages);
1233 cc->nr_migratepages = 0;
1234 nr_isolated = 0;
1235 }
1236
1237 if (low_pfn < next_skip_pfn) {
1238 low_pfn = next_skip_pfn - 1;
1239 /*
1240 * The check near the loop beginning would have updated
1241 * next_skip_pfn too, but this is a bit simpler.
1242 */
1243 next_skip_pfn += 1UL << cc->order;
1244 }
1245
1246 if (ret == -ENOMEM)
1247 break;
1248 }
1249
1250 /*
1251 * The PageBuddy() check could have potentially brought us outside
1252 * the range to be scanned.
1253 */
1254 if (unlikely(low_pfn > end_pfn))
1255 low_pfn = end_pfn;
1256
1257 folio = NULL;
1258
1259isolate_abort:
1260 if (locked)
1261 unlock_page_lruvec_irqrestore(locked, flags);
1262 if (folio) {
1263 folio_set_lru(folio);
1264 folio_put(folio);
1265 }
1266
1267 /*
1268 * Update the cached scanner pfn once the pageblock has been scanned.
1269 * Pages will either be migrated in which case there is no point
1270 * scanning in the near future or migration failed in which case the
1271 * failure reason may persist. The block is marked for skipping if
1272 * there were no pages isolated in the block or if the block is
1273 * rescanned twice in a row.
1274 */
1275 if (low_pfn == end_pfn && (!nr_isolated || cc->finish_pageblock)) {
1276 if (!cc->no_set_skip_hint && valid_page && !skip_updated)
1277 set_pageblock_skip(valid_page);
1278 update_cached_migrate(cc, low_pfn);
1279 }
1280
1281 trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
1282 nr_scanned, nr_isolated);
1283
1284fatal_pending:
1285 cc->total_migrate_scanned += nr_scanned;
1286 if (nr_isolated)
1287 count_compact_events(COMPACTISOLATED, nr_isolated);
1288
1289 cc->migrate_pfn = low_pfn;
1290
1291 return ret;
1292}
1293
1294/**
1295 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
1296 * @cc: Compaction control structure.
1297 * @start_pfn: The first PFN to start isolating.
1298 * @end_pfn: The one-past-last PFN.
1299 *
1300 * Returns -EAGAIN when contented, -EINTR in case of a signal pending, -ENOMEM
1301 * in case we could not allocate a page, or 0.
1302 */
1303int
1304isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
1305 unsigned long end_pfn)
1306{
1307 unsigned long pfn, block_start_pfn, block_end_pfn;
1308 int ret = 0;
1309
1310 /* Scan block by block. First and last block may be incomplete */
1311 pfn = start_pfn;
1312 block_start_pfn = pageblock_start_pfn(pfn);
1313 if (block_start_pfn < cc->zone->zone_start_pfn)
1314 block_start_pfn = cc->zone->zone_start_pfn;
1315 block_end_pfn = pageblock_end_pfn(pfn);
1316
1317 for (; pfn < end_pfn; pfn = block_end_pfn,
1318 block_start_pfn = block_end_pfn,
1319 block_end_pfn += pageblock_nr_pages) {
1320
1321 block_end_pfn = min(block_end_pfn, end_pfn);
1322
1323 if (!pageblock_pfn_to_page(block_start_pfn,
1324 block_end_pfn, cc->zone))
1325 continue;
1326
1327 ret = isolate_migratepages_block(cc, pfn, block_end_pfn,
1328 ISOLATE_UNEVICTABLE);
1329
1330 if (ret)
1331 break;
1332
1333 if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX)
1334 break;
1335 }
1336
1337 return ret;
1338}
1339
1340#endif /* CONFIG_COMPACTION || CONFIG_CMA */
1341#ifdef CONFIG_COMPACTION
1342
1343static bool suitable_migration_source(struct compact_control *cc,
1344 struct page *page)
1345{
1346 int block_mt;
1347
1348 if (pageblock_skip_persistent(page))
1349 return false;
1350
1351 if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction)
1352 return true;
1353
1354 block_mt = get_pageblock_migratetype(page);
1355
1356 if (cc->migratetype == MIGRATE_MOVABLE)
1357 return is_migrate_movable(block_mt);
1358 else
1359 return block_mt == cc->migratetype;
1360}
1361
1362/* Returns true if the page is within a block suitable for migration to */
1363static bool suitable_migration_target(struct compact_control *cc,
1364 struct page *page)
1365{
1366 /* If the page is a large free page, then disallow migration */
1367 if (PageBuddy(page)) {
1368 /*
1369 * We are checking page_order without zone->lock taken. But
1370 * the only small danger is that we skip a potentially suitable
1371 * pageblock, so it's not worth to check order for valid range.
1372 */
1373 if (buddy_order_unsafe(page) >= pageblock_order)
1374 return false;
1375 }
1376
1377 if (cc->ignore_block_suitable)
1378 return true;
1379
1380 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1381 if (is_migrate_movable(get_pageblock_migratetype(page)))
1382 return true;
1383
1384 /* Otherwise skip the block */
1385 return false;
1386}
1387
1388static inline unsigned int
1389freelist_scan_limit(struct compact_control *cc)
1390{
1391 unsigned short shift = BITS_PER_LONG - 1;
1392
1393 return (COMPACT_CLUSTER_MAX >> min(shift, cc->fast_search_fail)) + 1;
1394}
1395
1396/*
1397 * Test whether the free scanner has reached the same or lower pageblock than
1398 * the migration scanner, and compaction should thus terminate.
1399 */
1400static inline bool compact_scanners_met(struct compact_control *cc)
1401{
1402 return (cc->free_pfn >> pageblock_order)
1403 <= (cc->migrate_pfn >> pageblock_order);
1404}
1405
1406/*
1407 * Used when scanning for a suitable migration target which scans freelists
1408 * in reverse. Reorders the list such as the unscanned pages are scanned
1409 * first on the next iteration of the free scanner
1410 */
1411static void
1412move_freelist_head(struct list_head *freelist, struct page *freepage)
1413{
1414 LIST_HEAD(sublist);
1415
1416 if (!list_is_first(&freepage->buddy_list, freelist)) {
1417 list_cut_before(&sublist, freelist, &freepage->buddy_list);
1418 list_splice_tail(&sublist, freelist);
1419 }
1420}
1421
1422/*
1423 * Similar to move_freelist_head except used by the migration scanner
1424 * when scanning forward. It's possible for these list operations to
1425 * move against each other if they search the free list exactly in
1426 * lockstep.
1427 */
1428static void
1429move_freelist_tail(struct list_head *freelist, struct page *freepage)
1430{
1431 LIST_HEAD(sublist);
1432
1433 if (!list_is_last(&freepage->buddy_list, freelist)) {
1434 list_cut_position(&sublist, freelist, &freepage->buddy_list);
1435 list_splice_tail(&sublist, freelist);
1436 }
1437}
1438
1439static void
1440fast_isolate_around(struct compact_control *cc, unsigned long pfn)
1441{
1442 unsigned long start_pfn, end_pfn;
1443 struct page *page;
1444
1445 /* Do not search around if there are enough pages already */
1446 if (cc->nr_freepages >= cc->nr_migratepages)
1447 return;
1448
1449 /* Minimise scanning during async compaction */
1450 if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC)
1451 return;
1452
1453 /* Pageblock boundaries */
1454 start_pfn = max(pageblock_start_pfn(pfn), cc->zone->zone_start_pfn);
1455 end_pfn = min(pageblock_end_pfn(pfn), zone_end_pfn(cc->zone));
1456
1457 page = pageblock_pfn_to_page(start_pfn, end_pfn, cc->zone);
1458 if (!page)
1459 return;
1460
1461 isolate_freepages_block(cc, &start_pfn, end_pfn, &cc->freepages, 1, false);
1462
1463 /* Skip this pageblock in the future as it's full or nearly full */
1464 if (start_pfn == end_pfn && !cc->no_set_skip_hint)
1465 set_pageblock_skip(page);
1466}
1467
1468/* Search orders in round-robin fashion */
1469static int next_search_order(struct compact_control *cc, int order)
1470{
1471 order--;
1472 if (order < 0)
1473 order = cc->order - 1;
1474
1475 /* Search wrapped around? */
1476 if (order == cc->search_order) {
1477 cc->search_order--;
1478 if (cc->search_order < 0)
1479 cc->search_order = cc->order - 1;
1480 return -1;
1481 }
1482
1483 return order;
1484}
1485
1486static void fast_isolate_freepages(struct compact_control *cc)
1487{
1488 unsigned int limit = max(1U, freelist_scan_limit(cc) >> 1);
1489 unsigned int nr_scanned = 0, total_isolated = 0;
1490 unsigned long low_pfn, min_pfn, highest = 0;
1491 unsigned long nr_isolated = 0;
1492 unsigned long distance;
1493 struct page *page = NULL;
1494 bool scan_start = false;
1495 int order;
1496
1497 /* Full compaction passes in a negative order */
1498 if (cc->order <= 0)
1499 return;
1500
1501 /*
1502 * If starting the scan, use a deeper search and use the highest
1503 * PFN found if a suitable one is not found.
1504 */
1505 if (cc->free_pfn >= cc->zone->compact_init_free_pfn) {
1506 limit = pageblock_nr_pages >> 1;
1507 scan_start = true;
1508 }
1509
1510 /*
1511 * Preferred point is in the top quarter of the scan space but take
1512 * a pfn from the top half if the search is problematic.
1513 */
1514 distance = (cc->free_pfn - cc->migrate_pfn);
1515 low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2));
1516 min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1));
1517
1518 if (WARN_ON_ONCE(min_pfn > low_pfn))
1519 low_pfn = min_pfn;
1520
1521 /*
1522 * Search starts from the last successful isolation order or the next
1523 * order to search after a previous failure
1524 */
1525 cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order);
1526
1527 for (order = cc->search_order;
1528 !page && order >= 0;
1529 order = next_search_order(cc, order)) {
1530 struct free_area *area = &cc->zone->free_area[order];
1531 struct list_head *freelist;
1532 struct page *freepage;
1533 unsigned long flags;
1534 unsigned int order_scanned = 0;
1535 unsigned long high_pfn = 0;
1536
1537 if (!area->nr_free)
1538 continue;
1539
1540 spin_lock_irqsave(&cc->zone->lock, flags);
1541 freelist = &area->free_list[MIGRATE_MOVABLE];
1542 list_for_each_entry_reverse(freepage, freelist, buddy_list) {
1543 unsigned long pfn;
1544
1545 order_scanned++;
1546 nr_scanned++;
1547 pfn = page_to_pfn(freepage);
1548
1549 if (pfn >= highest)
1550 highest = max(pageblock_start_pfn(pfn),
1551 cc->zone->zone_start_pfn);
1552
1553 if (pfn >= low_pfn) {
1554 cc->fast_search_fail = 0;
1555 cc->search_order = order;
1556 page = freepage;
1557 break;
1558 }
1559
1560 if (pfn >= min_pfn && pfn > high_pfn) {
1561 high_pfn = pfn;
1562
1563 /* Shorten the scan if a candidate is found */
1564 limit >>= 1;
1565 }
1566
1567 if (order_scanned >= limit)
1568 break;
1569 }
1570
1571 /* Use a maximum candidate pfn if a preferred one was not found */
1572 if (!page && high_pfn) {
1573 page = pfn_to_page(high_pfn);
1574
1575 /* Update freepage for the list reorder below */
1576 freepage = page;
1577 }
1578
1579 /* Reorder to so a future search skips recent pages */
1580 move_freelist_head(freelist, freepage);
1581
1582 /* Isolate the page if available */
1583 if (page) {
1584 if (__isolate_free_page(page, order)) {
1585 set_page_private(page, order);
1586 nr_isolated = 1 << order;
1587 nr_scanned += nr_isolated - 1;
1588 total_isolated += nr_isolated;
1589 cc->nr_freepages += nr_isolated;
1590 list_add_tail(&page->lru, &cc->freepages);
1591 count_compact_events(COMPACTISOLATED, nr_isolated);
1592 } else {
1593 /* If isolation fails, abort the search */
1594 order = cc->search_order + 1;
1595 page = NULL;
1596 }
1597 }
1598
1599 spin_unlock_irqrestore(&cc->zone->lock, flags);
1600
1601 /* Skip fast search if enough freepages isolated */
1602 if (cc->nr_freepages >= cc->nr_migratepages)
1603 break;
1604
1605 /*
1606 * Smaller scan on next order so the total scan is related
1607 * to freelist_scan_limit.
1608 */
1609 if (order_scanned >= limit)
1610 limit = max(1U, limit >> 1);
1611 }
1612
1613 trace_mm_compaction_fast_isolate_freepages(min_pfn, cc->free_pfn,
1614 nr_scanned, total_isolated);
1615
1616 if (!page) {
1617 cc->fast_search_fail++;
1618 if (scan_start) {
1619 /*
1620 * Use the highest PFN found above min. If one was
1621 * not found, be pessimistic for direct compaction
1622 * and use the min mark.
1623 */
1624 if (highest >= min_pfn) {
1625 page = pfn_to_page(highest);
1626 cc->free_pfn = highest;
1627 } else {
1628 if (cc->direct_compaction && pfn_valid(min_pfn)) {
1629 page = pageblock_pfn_to_page(min_pfn,
1630 min(pageblock_end_pfn(min_pfn),
1631 zone_end_pfn(cc->zone)),
1632 cc->zone);
1633 if (page && !suitable_migration_target(cc, page))
1634 page = NULL;
1635
1636 cc->free_pfn = min_pfn;
1637 }
1638 }
1639 }
1640 }
1641
1642 if (highest && highest >= cc->zone->compact_cached_free_pfn) {
1643 highest -= pageblock_nr_pages;
1644 cc->zone->compact_cached_free_pfn = highest;
1645 }
1646
1647 cc->total_free_scanned += nr_scanned;
1648 if (!page)
1649 return;
1650
1651 low_pfn = page_to_pfn(page);
1652 fast_isolate_around(cc, low_pfn);
1653}
1654
1655/*
1656 * Based on information in the current compact_control, find blocks
1657 * suitable for isolating free pages from and then isolate them.
1658 */
1659static void isolate_freepages(struct compact_control *cc)
1660{
1661 struct zone *zone = cc->zone;
1662 struct page *page;
1663 unsigned long block_start_pfn; /* start of current pageblock */
1664 unsigned long isolate_start_pfn; /* exact pfn we start at */
1665 unsigned long block_end_pfn; /* end of current pageblock */
1666 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
1667 struct list_head *freelist = &cc->freepages;
1668 unsigned int stride;
1669
1670 /* Try a small search of the free lists for a candidate */
1671 fast_isolate_freepages(cc);
1672 if (cc->nr_freepages)
1673 goto splitmap;
1674
1675 /*
1676 * Initialise the free scanner. The starting point is where we last
1677 * successfully isolated from, zone-cached value, or the end of the
1678 * zone when isolating for the first time. For looping we also need
1679 * this pfn aligned down to the pageblock boundary, because we do
1680 * block_start_pfn -= pageblock_nr_pages in the for loop.
1681 * For ending point, take care when isolating in last pageblock of a
1682 * zone which ends in the middle of a pageblock.
1683 * The low boundary is the end of the pageblock the migration scanner
1684 * is using.
1685 */
1686 isolate_start_pfn = cc->free_pfn;
1687 block_start_pfn = pageblock_start_pfn(isolate_start_pfn);
1688 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1689 zone_end_pfn(zone));
1690 low_pfn = pageblock_end_pfn(cc->migrate_pfn);
1691 stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1;
1692
1693 /*
1694 * Isolate free pages until enough are available to migrate the
1695 * pages on cc->migratepages. We stop searching if the migrate
1696 * and free page scanners meet or enough free pages are isolated.
1697 */
1698 for (; block_start_pfn >= low_pfn;
1699 block_end_pfn = block_start_pfn,
1700 block_start_pfn -= pageblock_nr_pages,
1701 isolate_start_pfn = block_start_pfn) {
1702 unsigned long nr_isolated;
1703
1704 /*
1705 * This can iterate a massively long zone without finding any
1706 * suitable migration targets, so periodically check resched.
1707 */
1708 if (!(block_start_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages)))
1709 cond_resched();
1710
1711 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1712 zone);
1713 if (!page) {
1714 unsigned long next_pfn;
1715
1716 next_pfn = skip_offline_sections_reverse(block_start_pfn);
1717 if (next_pfn)
1718 block_start_pfn = max(next_pfn, low_pfn);
1719
1720 continue;
1721 }
1722
1723 /* Check the block is suitable for migration */
1724 if (!suitable_migration_target(cc, page))
1725 continue;
1726
1727 /* If isolation recently failed, do not retry */
1728 if (!isolation_suitable(cc, page))
1729 continue;
1730
1731 /* Found a block suitable for isolating free pages from. */
1732 nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn,
1733 block_end_pfn, freelist, stride, false);
1734
1735 /* Update the skip hint if the full pageblock was scanned */
1736 if (isolate_start_pfn == block_end_pfn)
1737 update_pageblock_skip(cc, page, block_start_pfn -
1738 pageblock_nr_pages);
1739
1740 /* Are enough freepages isolated? */
1741 if (cc->nr_freepages >= cc->nr_migratepages) {
1742 if (isolate_start_pfn >= block_end_pfn) {
1743 /*
1744 * Restart at previous pageblock if more
1745 * freepages can be isolated next time.
1746 */
1747 isolate_start_pfn =
1748 block_start_pfn - pageblock_nr_pages;
1749 }
1750 break;
1751 } else if (isolate_start_pfn < block_end_pfn) {
1752 /*
1753 * If isolation failed early, do not continue
1754 * needlessly.
1755 */
1756 break;
1757 }
1758
1759 /* Adjust stride depending on isolation */
1760 if (nr_isolated) {
1761 stride = 1;
1762 continue;
1763 }
1764 stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1);
1765 }
1766
1767 /*
1768 * Record where the free scanner will restart next time. Either we
1769 * broke from the loop and set isolate_start_pfn based on the last
1770 * call to isolate_freepages_block(), or we met the migration scanner
1771 * and the loop terminated due to isolate_start_pfn < low_pfn
1772 */
1773 cc->free_pfn = isolate_start_pfn;
1774
1775splitmap:
1776 /* __isolate_free_page() does not map the pages */
1777 split_map_pages(freelist);
1778}
1779
1780/*
1781 * This is a migrate-callback that "allocates" freepages by taking pages
1782 * from the isolated freelists in the block we are migrating to.
1783 */
1784static struct folio *compaction_alloc(struct folio *src, unsigned long data)
1785{
1786 struct compact_control *cc = (struct compact_control *)data;
1787 struct folio *dst;
1788
1789 if (list_empty(&cc->freepages)) {
1790 isolate_freepages(cc);
1791
1792 if (list_empty(&cc->freepages))
1793 return NULL;
1794 }
1795
1796 dst = list_entry(cc->freepages.next, struct folio, lru);
1797 list_del(&dst->lru);
1798 cc->nr_freepages--;
1799
1800 return dst;
1801}
1802
1803/*
1804 * This is a migrate-callback that "frees" freepages back to the isolated
1805 * freelist. All pages on the freelist are from the same zone, so there is no
1806 * special handling needed for NUMA.
1807 */
1808static void compaction_free(struct folio *dst, unsigned long data)
1809{
1810 struct compact_control *cc = (struct compact_control *)data;
1811
1812 list_add(&dst->lru, &cc->freepages);
1813 cc->nr_freepages++;
1814}
1815
1816/* possible outcome of isolate_migratepages */
1817typedef enum {
1818 ISOLATE_ABORT, /* Abort compaction now */
1819 ISOLATE_NONE, /* No pages isolated, continue scanning */
1820 ISOLATE_SUCCESS, /* Pages isolated, migrate */
1821} isolate_migrate_t;
1822
1823/*
1824 * Allow userspace to control policy on scanning the unevictable LRU for
1825 * compactable pages.
1826 */
1827static int sysctl_compact_unevictable_allowed __read_mostly = CONFIG_COMPACT_UNEVICTABLE_DEFAULT;
1828/*
1829 * Tunable for proactive compaction. It determines how
1830 * aggressively the kernel should compact memory in the
1831 * background. It takes values in the range [0, 100].
1832 */
1833static unsigned int __read_mostly sysctl_compaction_proactiveness = 20;
1834static int sysctl_extfrag_threshold = 500;
1835static int __read_mostly sysctl_compact_memory;
1836
1837static inline void
1838update_fast_start_pfn(struct compact_control *cc, unsigned long pfn)
1839{
1840 if (cc->fast_start_pfn == ULONG_MAX)
1841 return;
1842
1843 if (!cc->fast_start_pfn)
1844 cc->fast_start_pfn = pfn;
1845
1846 cc->fast_start_pfn = min(cc->fast_start_pfn, pfn);
1847}
1848
1849static inline unsigned long
1850reinit_migrate_pfn(struct compact_control *cc)
1851{
1852 if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX)
1853 return cc->migrate_pfn;
1854
1855 cc->migrate_pfn = cc->fast_start_pfn;
1856 cc->fast_start_pfn = ULONG_MAX;
1857
1858 return cc->migrate_pfn;
1859}
1860
1861/*
1862 * Briefly search the free lists for a migration source that already has
1863 * some free pages to reduce the number of pages that need migration
1864 * before a pageblock is free.
1865 */
1866static unsigned long fast_find_migrateblock(struct compact_control *cc)
1867{
1868 unsigned int limit = freelist_scan_limit(cc);
1869 unsigned int nr_scanned = 0;
1870 unsigned long distance;
1871 unsigned long pfn = cc->migrate_pfn;
1872 unsigned long high_pfn;
1873 int order;
1874 bool found_block = false;
1875
1876 /* Skip hints are relied on to avoid repeats on the fast search */
1877 if (cc->ignore_skip_hint)
1878 return pfn;
1879
1880 /*
1881 * If the pageblock should be finished then do not select a different
1882 * pageblock.
1883 */
1884 if (cc->finish_pageblock)
1885 return pfn;
1886
1887 /*
1888 * If the migrate_pfn is not at the start of a zone or the start
1889 * of a pageblock then assume this is a continuation of a previous
1890 * scan restarted due to COMPACT_CLUSTER_MAX.
1891 */
1892 if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn))
1893 return pfn;
1894
1895 /*
1896 * For smaller orders, just linearly scan as the number of pages
1897 * to migrate should be relatively small and does not necessarily
1898 * justify freeing up a large block for a small allocation.
1899 */
1900 if (cc->order <= PAGE_ALLOC_COSTLY_ORDER)
1901 return pfn;
1902
1903 /*
1904 * Only allow kcompactd and direct requests for movable pages to
1905 * quickly clear out a MOVABLE pageblock for allocation. This
1906 * reduces the risk that a large movable pageblock is freed for
1907 * an unmovable/reclaimable small allocation.
1908 */
1909 if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE)
1910 return pfn;
1911
1912 /*
1913 * When starting the migration scanner, pick any pageblock within the
1914 * first half of the search space. Otherwise try and pick a pageblock
1915 * within the first eighth to reduce the chances that a migration
1916 * target later becomes a source.
1917 */
1918 distance = (cc->free_pfn - cc->migrate_pfn) >> 1;
1919 if (cc->migrate_pfn != cc->zone->zone_start_pfn)
1920 distance >>= 2;
1921 high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance);
1922
1923 for (order = cc->order - 1;
1924 order >= PAGE_ALLOC_COSTLY_ORDER && !found_block && nr_scanned < limit;
1925 order--) {
1926 struct free_area *area = &cc->zone->free_area[order];
1927 struct list_head *freelist;
1928 unsigned long flags;
1929 struct page *freepage;
1930
1931 if (!area->nr_free)
1932 continue;
1933
1934 spin_lock_irqsave(&cc->zone->lock, flags);
1935 freelist = &area->free_list[MIGRATE_MOVABLE];
1936 list_for_each_entry(freepage, freelist, buddy_list) {
1937 unsigned long free_pfn;
1938
1939 if (nr_scanned++ >= limit) {
1940 move_freelist_tail(freelist, freepage);
1941 break;
1942 }
1943
1944 free_pfn = page_to_pfn(freepage);
1945 if (free_pfn < high_pfn) {
1946 /*
1947 * Avoid if skipped recently. Ideally it would
1948 * move to the tail but even safe iteration of
1949 * the list assumes an entry is deleted, not
1950 * reordered.
1951 */
1952 if (get_pageblock_skip(freepage))
1953 continue;
1954
1955 /* Reorder to so a future search skips recent pages */
1956 move_freelist_tail(freelist, freepage);
1957
1958 update_fast_start_pfn(cc, free_pfn);
1959 pfn = pageblock_start_pfn(free_pfn);
1960 if (pfn < cc->zone->zone_start_pfn)
1961 pfn = cc->zone->zone_start_pfn;
1962 cc->fast_search_fail = 0;
1963 found_block = true;
1964 break;
1965 }
1966 }
1967 spin_unlock_irqrestore(&cc->zone->lock, flags);
1968 }
1969
1970 cc->total_migrate_scanned += nr_scanned;
1971
1972 /*
1973 * If fast scanning failed then use a cached entry for a page block
1974 * that had free pages as the basis for starting a linear scan.
1975 */
1976 if (!found_block) {
1977 cc->fast_search_fail++;
1978 pfn = reinit_migrate_pfn(cc);
1979 }
1980 return pfn;
1981}
1982
1983/*
1984 * Isolate all pages that can be migrated from the first suitable block,
1985 * starting at the block pointed to by the migrate scanner pfn within
1986 * compact_control.
1987 */
1988static isolate_migrate_t isolate_migratepages(struct compact_control *cc)
1989{
1990 unsigned long block_start_pfn;
1991 unsigned long block_end_pfn;
1992 unsigned long low_pfn;
1993 struct page *page;
1994 const isolate_mode_t isolate_mode =
1995 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1996 (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1997 bool fast_find_block;
1998
1999 /*
2000 * Start at where we last stopped, or beginning of the zone as
2001 * initialized by compact_zone(). The first failure will use
2002 * the lowest PFN as the starting point for linear scanning.
2003 */
2004 low_pfn = fast_find_migrateblock(cc);
2005 block_start_pfn = pageblock_start_pfn(low_pfn);
2006 if (block_start_pfn < cc->zone->zone_start_pfn)
2007 block_start_pfn = cc->zone->zone_start_pfn;
2008
2009 /*
2010 * fast_find_migrateblock() has already ensured the pageblock is not
2011 * set with a skipped flag, so to avoid the isolation_suitable check
2012 * below again, check whether the fast search was successful.
2013 */
2014 fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail;
2015
2016 /* Only scan within a pageblock boundary */
2017 block_end_pfn = pageblock_end_pfn(low_pfn);
2018
2019 /*
2020 * Iterate over whole pageblocks until we find the first suitable.
2021 * Do not cross the free scanner.
2022 */
2023 for (; block_end_pfn <= cc->free_pfn;
2024 fast_find_block = false,
2025 cc->migrate_pfn = low_pfn = block_end_pfn,
2026 block_start_pfn = block_end_pfn,
2027 block_end_pfn += pageblock_nr_pages) {
2028
2029 /*
2030 * This can potentially iterate a massively long zone with
2031 * many pageblocks unsuitable, so periodically check if we
2032 * need to schedule.
2033 */
2034 if (!(low_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages)))
2035 cond_resched();
2036
2037 page = pageblock_pfn_to_page(block_start_pfn,
2038 block_end_pfn, cc->zone);
2039 if (!page) {
2040 unsigned long next_pfn;
2041
2042 next_pfn = skip_offline_sections(block_start_pfn);
2043 if (next_pfn)
2044 block_end_pfn = min(next_pfn, cc->free_pfn);
2045 continue;
2046 }
2047
2048 /*
2049 * If isolation recently failed, do not retry. Only check the
2050 * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock
2051 * to be visited multiple times. Assume skip was checked
2052 * before making it "skip" so other compaction instances do
2053 * not scan the same block.
2054 */
2055 if ((pageblock_aligned(low_pfn) ||
2056 low_pfn == cc->zone->zone_start_pfn) &&
2057 !fast_find_block && !isolation_suitable(cc, page))
2058 continue;
2059
2060 /*
2061 * For async direct compaction, only scan the pageblocks of the
2062 * same migratetype without huge pages. Async direct compaction
2063 * is optimistic to see if the minimum amount of work satisfies
2064 * the allocation. The cached PFN is updated as it's possible
2065 * that all remaining blocks between source and target are
2066 * unsuitable and the compaction scanners fail to meet.
2067 */
2068 if (!suitable_migration_source(cc, page)) {
2069 update_cached_migrate(cc, block_end_pfn);
2070 continue;
2071 }
2072
2073 /* Perform the isolation */
2074 if (isolate_migratepages_block(cc, low_pfn, block_end_pfn,
2075 isolate_mode))
2076 return ISOLATE_ABORT;
2077
2078 /*
2079 * Either we isolated something and proceed with migration. Or
2080 * we failed and compact_zone should decide if we should
2081 * continue or not.
2082 */
2083 break;
2084 }
2085
2086 return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
2087}
2088
2089/*
2090 * order == -1 is expected when compacting proactively via
2091 * 1. /proc/sys/vm/compact_memory
2092 * 2. /sys/devices/system/node/nodex/compact
2093 * 3. /proc/sys/vm/compaction_proactiveness
2094 */
2095static inline bool is_via_compact_memory(int order)
2096{
2097 return order == -1;
2098}
2099
2100/*
2101 * Determine whether kswapd is (or recently was!) running on this node.
2102 *
2103 * pgdat_kswapd_lock() pins pgdat->kswapd, so a concurrent kswapd_stop() can't
2104 * zero it.
2105 */
2106static bool kswapd_is_running(pg_data_t *pgdat)
2107{
2108 bool running;
2109
2110 pgdat_kswapd_lock(pgdat);
2111 running = pgdat->kswapd && task_is_running(pgdat->kswapd);
2112 pgdat_kswapd_unlock(pgdat);
2113
2114 return running;
2115}
2116
2117/*
2118 * A zone's fragmentation score is the external fragmentation wrt to the
2119 * COMPACTION_HPAGE_ORDER. It returns a value in the range [0, 100].
2120 */
2121static unsigned int fragmentation_score_zone(struct zone *zone)
2122{
2123 return extfrag_for_order(zone, COMPACTION_HPAGE_ORDER);
2124}
2125
2126/*
2127 * A weighted zone's fragmentation score is the external fragmentation
2128 * wrt to the COMPACTION_HPAGE_ORDER scaled by the zone's size. It
2129 * returns a value in the range [0, 100].
2130 *
2131 * The scaling factor ensures that proactive compaction focuses on larger
2132 * zones like ZONE_NORMAL, rather than smaller, specialized zones like
2133 * ZONE_DMA32. For smaller zones, the score value remains close to zero,
2134 * and thus never exceeds the high threshold for proactive compaction.
2135 */
2136static unsigned int fragmentation_score_zone_weighted(struct zone *zone)
2137{
2138 unsigned long score;
2139
2140 score = zone->present_pages * fragmentation_score_zone(zone);
2141 return div64_ul(score, zone->zone_pgdat->node_present_pages + 1);
2142}
2143
2144/*
2145 * The per-node proactive (background) compaction process is started by its
2146 * corresponding kcompactd thread when the node's fragmentation score
2147 * exceeds the high threshold. The compaction process remains active till
2148 * the node's score falls below the low threshold, or one of the back-off
2149 * conditions is met.
2150 */
2151static unsigned int fragmentation_score_node(pg_data_t *pgdat)
2152{
2153 unsigned int score = 0;
2154 int zoneid;
2155
2156 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2157 struct zone *zone;
2158
2159 zone = &pgdat->node_zones[zoneid];
2160 if (!populated_zone(zone))
2161 continue;
2162 score += fragmentation_score_zone_weighted(zone);
2163 }
2164
2165 return score;
2166}
2167
2168static unsigned int fragmentation_score_wmark(bool low)
2169{
2170 unsigned int wmark_low;
2171
2172 /*
2173 * Cap the low watermark to avoid excessive compaction
2174 * activity in case a user sets the proactiveness tunable
2175 * close to 100 (maximum).
2176 */
2177 wmark_low = max(100U - sysctl_compaction_proactiveness, 5U);
2178 return low ? wmark_low : min(wmark_low + 10, 100U);
2179}
2180
2181static bool should_proactive_compact_node(pg_data_t *pgdat)
2182{
2183 int wmark_high;
2184
2185 if (!sysctl_compaction_proactiveness || kswapd_is_running(pgdat))
2186 return false;
2187
2188 wmark_high = fragmentation_score_wmark(false);
2189 return fragmentation_score_node(pgdat) > wmark_high;
2190}
2191
2192static enum compact_result __compact_finished(struct compact_control *cc)
2193{
2194 unsigned int order;
2195 const int migratetype = cc->migratetype;
2196 int ret;
2197
2198 /* Compaction run completes if the migrate and free scanner meet */
2199 if (compact_scanners_met(cc)) {
2200 /* Let the next compaction start anew. */
2201 reset_cached_positions(cc->zone);
2202
2203 /*
2204 * Mark that the PG_migrate_skip information should be cleared
2205 * by kswapd when it goes to sleep. kcompactd does not set the
2206 * flag itself as the decision to be clear should be directly
2207 * based on an allocation request.
2208 */
2209 if (cc->direct_compaction)
2210 cc->zone->compact_blockskip_flush = true;
2211
2212 if (cc->whole_zone)
2213 return COMPACT_COMPLETE;
2214 else
2215 return COMPACT_PARTIAL_SKIPPED;
2216 }
2217
2218 if (cc->proactive_compaction) {
2219 int score, wmark_low;
2220 pg_data_t *pgdat;
2221
2222 pgdat = cc->zone->zone_pgdat;
2223 if (kswapd_is_running(pgdat))
2224 return COMPACT_PARTIAL_SKIPPED;
2225
2226 score = fragmentation_score_zone(cc->zone);
2227 wmark_low = fragmentation_score_wmark(true);
2228
2229 if (score > wmark_low)
2230 ret = COMPACT_CONTINUE;
2231 else
2232 ret = COMPACT_SUCCESS;
2233
2234 goto out;
2235 }
2236
2237 if (is_via_compact_memory(cc->order))
2238 return COMPACT_CONTINUE;
2239
2240 /*
2241 * Always finish scanning a pageblock to reduce the possibility of
2242 * fallbacks in the future. This is particularly important when
2243 * migration source is unmovable/reclaimable but it's not worth
2244 * special casing.
2245 */
2246 if (!pageblock_aligned(cc->migrate_pfn))
2247 return COMPACT_CONTINUE;
2248
2249 /* Direct compactor: Is a suitable page free? */
2250 ret = COMPACT_NO_SUITABLE_PAGE;
2251 for (order = cc->order; order < NR_PAGE_ORDERS; order++) {
2252 struct free_area *area = &cc->zone->free_area[order];
2253 bool can_steal;
2254
2255 /* Job done if page is free of the right migratetype */
2256 if (!free_area_empty(area, migratetype))
2257 return COMPACT_SUCCESS;
2258
2259#ifdef CONFIG_CMA
2260 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
2261 if (migratetype == MIGRATE_MOVABLE &&
2262 !free_area_empty(area, MIGRATE_CMA))
2263 return COMPACT_SUCCESS;
2264#endif
2265 /*
2266 * Job done if allocation would steal freepages from
2267 * other migratetype buddy lists.
2268 */
2269 if (find_suitable_fallback(area, order, migratetype,
2270 true, &can_steal) != -1)
2271 /*
2272 * Movable pages are OK in any pageblock. If we are
2273 * stealing for a non-movable allocation, make sure
2274 * we finish compacting the current pageblock first
2275 * (which is assured by the above migrate_pfn align
2276 * check) so it is as free as possible and we won't
2277 * have to steal another one soon.
2278 */
2279 return COMPACT_SUCCESS;
2280 }
2281
2282out:
2283 if (cc->contended || fatal_signal_pending(current))
2284 ret = COMPACT_CONTENDED;
2285
2286 return ret;
2287}
2288
2289static enum compact_result compact_finished(struct compact_control *cc)
2290{
2291 int ret;
2292
2293 ret = __compact_finished(cc);
2294 trace_mm_compaction_finished(cc->zone, cc->order, ret);
2295 if (ret == COMPACT_NO_SUITABLE_PAGE)
2296 ret = COMPACT_CONTINUE;
2297
2298 return ret;
2299}
2300
2301static bool __compaction_suitable(struct zone *zone, int order,
2302 int highest_zoneidx,
2303 unsigned long wmark_target)
2304{
2305 unsigned long watermark;
2306 /*
2307 * Watermarks for order-0 must be met for compaction to be able to
2308 * isolate free pages for migration targets. This means that the
2309 * watermark and alloc_flags have to match, or be more pessimistic than
2310 * the check in __isolate_free_page(). We don't use the direct
2311 * compactor's alloc_flags, as they are not relevant for freepage
2312 * isolation. We however do use the direct compactor's highest_zoneidx
2313 * to skip over zones where lowmem reserves would prevent allocation
2314 * even if compaction succeeds.
2315 * For costly orders, we require low watermark instead of min for
2316 * compaction to proceed to increase its chances.
2317 * ALLOC_CMA is used, as pages in CMA pageblocks are considered
2318 * suitable migration targets
2319 */
2320 watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
2321 low_wmark_pages(zone) : min_wmark_pages(zone);
2322 watermark += compact_gap(order);
2323 return __zone_watermark_ok(zone, 0, watermark, highest_zoneidx,
2324 ALLOC_CMA, wmark_target);
2325}
2326
2327/*
2328 * compaction_suitable: Is this suitable to run compaction on this zone now?
2329 */
2330bool compaction_suitable(struct zone *zone, int order, int highest_zoneidx)
2331{
2332 enum compact_result compact_result;
2333 bool suitable;
2334
2335 suitable = __compaction_suitable(zone, order, highest_zoneidx,
2336 zone_page_state(zone, NR_FREE_PAGES));
2337 /*
2338 * fragmentation index determines if allocation failures are due to
2339 * low memory or external fragmentation
2340 *
2341 * index of -1000 would imply allocations might succeed depending on
2342 * watermarks, but we already failed the high-order watermark check
2343 * index towards 0 implies failure is due to lack of memory
2344 * index towards 1000 implies failure is due to fragmentation
2345 *
2346 * Only compact if a failure would be due to fragmentation. Also
2347 * ignore fragindex for non-costly orders where the alternative to
2348 * a successful reclaim/compaction is OOM. Fragindex and the
2349 * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
2350 * excessive compaction for costly orders, but it should not be at the
2351 * expense of system stability.
2352 */
2353 if (suitable) {
2354 compact_result = COMPACT_CONTINUE;
2355 if (order > PAGE_ALLOC_COSTLY_ORDER) {
2356 int fragindex = fragmentation_index(zone, order);
2357
2358 if (fragindex >= 0 &&
2359 fragindex <= sysctl_extfrag_threshold) {
2360 suitable = false;
2361 compact_result = COMPACT_NOT_SUITABLE_ZONE;
2362 }
2363 }
2364 } else {
2365 compact_result = COMPACT_SKIPPED;
2366 }
2367
2368 trace_mm_compaction_suitable(zone, order, compact_result);
2369
2370 return suitable;
2371}
2372
2373bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
2374 int alloc_flags)
2375{
2376 struct zone *zone;
2377 struct zoneref *z;
2378
2379 /*
2380 * Make sure at least one zone would pass __compaction_suitable if we continue
2381 * retrying the reclaim.
2382 */
2383 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2384 ac->highest_zoneidx, ac->nodemask) {
2385 unsigned long available;
2386
2387 /*
2388 * Do not consider all the reclaimable memory because we do not
2389 * want to trash just for a single high order allocation which
2390 * is even not guaranteed to appear even if __compaction_suitable
2391 * is happy about the watermark check.
2392 */
2393 available = zone_reclaimable_pages(zone) / order;
2394 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
2395 if (__compaction_suitable(zone, order, ac->highest_zoneidx,
2396 available))
2397 return true;
2398 }
2399
2400 return false;
2401}
2402
2403/*
2404 * Should we do compaction for target allocation order.
2405 * Return COMPACT_SUCCESS if allocation for target order can be already
2406 * satisfied
2407 * Return COMPACT_SKIPPED if compaction for target order is likely to fail
2408 * Return COMPACT_CONTINUE if compaction for target order should be ran
2409 */
2410static enum compact_result
2411compaction_suit_allocation_order(struct zone *zone, unsigned int order,
2412 int highest_zoneidx, unsigned int alloc_flags)
2413{
2414 unsigned long watermark;
2415
2416 watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
2417 if (zone_watermark_ok(zone, order, watermark, highest_zoneidx,
2418 alloc_flags))
2419 return COMPACT_SUCCESS;
2420
2421 if (!compaction_suitable(zone, order, highest_zoneidx))
2422 return COMPACT_SKIPPED;
2423
2424 return COMPACT_CONTINUE;
2425}
2426
2427static enum compact_result
2428compact_zone(struct compact_control *cc, struct capture_control *capc)
2429{
2430 enum compact_result ret;
2431 unsigned long start_pfn = cc->zone->zone_start_pfn;
2432 unsigned long end_pfn = zone_end_pfn(cc->zone);
2433 unsigned long last_migrated_pfn;
2434 const bool sync = cc->mode != MIGRATE_ASYNC;
2435 bool update_cached;
2436 unsigned int nr_succeeded = 0;
2437
2438 /*
2439 * These counters track activities during zone compaction. Initialize
2440 * them before compacting a new zone.
2441 */
2442 cc->total_migrate_scanned = 0;
2443 cc->total_free_scanned = 0;
2444 cc->nr_migratepages = 0;
2445 cc->nr_freepages = 0;
2446 INIT_LIST_HEAD(&cc->freepages);
2447 INIT_LIST_HEAD(&cc->migratepages);
2448
2449 cc->migratetype = gfp_migratetype(cc->gfp_mask);
2450
2451 if (!is_via_compact_memory(cc->order)) {
2452 ret = compaction_suit_allocation_order(cc->zone, cc->order,
2453 cc->highest_zoneidx,
2454 cc->alloc_flags);
2455 if (ret != COMPACT_CONTINUE)
2456 return ret;
2457 }
2458
2459 /*
2460 * Clear pageblock skip if there were failures recently and compaction
2461 * is about to be retried after being deferred.
2462 */
2463 if (compaction_restarting(cc->zone, cc->order))
2464 __reset_isolation_suitable(cc->zone);
2465
2466 /*
2467 * Setup to move all movable pages to the end of the zone. Used cached
2468 * information on where the scanners should start (unless we explicitly
2469 * want to compact the whole zone), but check that it is initialised
2470 * by ensuring the values are within zone boundaries.
2471 */
2472 cc->fast_start_pfn = 0;
2473 if (cc->whole_zone) {
2474 cc->migrate_pfn = start_pfn;
2475 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2476 } else {
2477 cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync];
2478 cc->free_pfn = cc->zone->compact_cached_free_pfn;
2479 if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
2480 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2481 cc->zone->compact_cached_free_pfn = cc->free_pfn;
2482 }
2483 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
2484 cc->migrate_pfn = start_pfn;
2485 cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
2486 cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
2487 }
2488
2489 if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn)
2490 cc->whole_zone = true;
2491 }
2492
2493 last_migrated_pfn = 0;
2494
2495 /*
2496 * Migrate has separate cached PFNs for ASYNC and SYNC* migration on
2497 * the basis that some migrations will fail in ASYNC mode. However,
2498 * if the cached PFNs match and pageblocks are skipped due to having
2499 * no isolation candidates, then the sync state does not matter.
2500 * Until a pageblock with isolation candidates is found, keep the
2501 * cached PFNs in sync to avoid revisiting the same blocks.
2502 */
2503 update_cached = !sync &&
2504 cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1];
2505
2506 trace_mm_compaction_begin(cc, start_pfn, end_pfn, sync);
2507
2508 /* lru_add_drain_all could be expensive with involving other CPUs */
2509 lru_add_drain();
2510
2511 while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) {
2512 int err;
2513 unsigned long iteration_start_pfn = cc->migrate_pfn;
2514
2515 /*
2516 * Avoid multiple rescans of the same pageblock which can
2517 * happen if a page cannot be isolated (dirty/writeback in
2518 * async mode) or if the migrated pages are being allocated
2519 * before the pageblock is cleared. The first rescan will
2520 * capture the entire pageblock for migration. If it fails,
2521 * it'll be marked skip and scanning will proceed as normal.
2522 */
2523 cc->finish_pageblock = false;
2524 if (pageblock_start_pfn(last_migrated_pfn) ==
2525 pageblock_start_pfn(iteration_start_pfn)) {
2526 cc->finish_pageblock = true;
2527 }
2528
2529rescan:
2530 switch (isolate_migratepages(cc)) {
2531 case ISOLATE_ABORT:
2532 ret = COMPACT_CONTENDED;
2533 putback_movable_pages(&cc->migratepages);
2534 cc->nr_migratepages = 0;
2535 goto out;
2536 case ISOLATE_NONE:
2537 if (update_cached) {
2538 cc->zone->compact_cached_migrate_pfn[1] =
2539 cc->zone->compact_cached_migrate_pfn[0];
2540 }
2541
2542 /*
2543 * We haven't isolated and migrated anything, but
2544 * there might still be unflushed migrations from
2545 * previous cc->order aligned block.
2546 */
2547 goto check_drain;
2548 case ISOLATE_SUCCESS:
2549 update_cached = false;
2550 last_migrated_pfn = max(cc->zone->zone_start_pfn,
2551 pageblock_start_pfn(cc->migrate_pfn - 1));
2552 }
2553
2554 err = migrate_pages(&cc->migratepages, compaction_alloc,
2555 compaction_free, (unsigned long)cc, cc->mode,
2556 MR_COMPACTION, &nr_succeeded);
2557
2558 trace_mm_compaction_migratepages(cc, nr_succeeded);
2559
2560 /* All pages were either migrated or will be released */
2561 cc->nr_migratepages = 0;
2562 if (err) {
2563 putback_movable_pages(&cc->migratepages);
2564 /*
2565 * migrate_pages() may return -ENOMEM when scanners meet
2566 * and we want compact_finished() to detect it
2567 */
2568 if (err == -ENOMEM && !compact_scanners_met(cc)) {
2569 ret = COMPACT_CONTENDED;
2570 goto out;
2571 }
2572 /*
2573 * If an ASYNC or SYNC_LIGHT fails to migrate a page
2574 * within the pageblock_order-aligned block and
2575 * fast_find_migrateblock may be used then scan the
2576 * remainder of the pageblock. This will mark the
2577 * pageblock "skip" to avoid rescanning in the near
2578 * future. This will isolate more pages than necessary
2579 * for the request but avoid loops due to
2580 * fast_find_migrateblock revisiting blocks that were
2581 * recently partially scanned.
2582 */
2583 if (!pageblock_aligned(cc->migrate_pfn) &&
2584 !cc->ignore_skip_hint && !cc->finish_pageblock &&
2585 (cc->mode < MIGRATE_SYNC)) {
2586 cc->finish_pageblock = true;
2587
2588 /*
2589 * Draining pcplists does not help THP if
2590 * any page failed to migrate. Even after
2591 * drain, the pageblock will not be free.
2592 */
2593 if (cc->order == COMPACTION_HPAGE_ORDER)
2594 last_migrated_pfn = 0;
2595
2596 goto rescan;
2597 }
2598 }
2599
2600 /* Stop if a page has been captured */
2601 if (capc && capc->page) {
2602 ret = COMPACT_SUCCESS;
2603 break;
2604 }
2605
2606check_drain:
2607 /*
2608 * Has the migration scanner moved away from the previous
2609 * cc->order aligned block where we migrated from? If yes,
2610 * flush the pages that were freed, so that they can merge and
2611 * compact_finished() can detect immediately if allocation
2612 * would succeed.
2613 */
2614 if (cc->order > 0 && last_migrated_pfn) {
2615 unsigned long current_block_start =
2616 block_start_pfn(cc->migrate_pfn, cc->order);
2617
2618 if (last_migrated_pfn < current_block_start) {
2619 lru_add_drain_cpu_zone(cc->zone);
2620 /* No more flushing until we migrate again */
2621 last_migrated_pfn = 0;
2622 }
2623 }
2624 }
2625
2626out:
2627 /*
2628 * Release free pages and update where the free scanner should restart,
2629 * so we don't leave any returned pages behind in the next attempt.
2630 */
2631 if (cc->nr_freepages > 0) {
2632 unsigned long free_pfn = release_freepages(&cc->freepages);
2633
2634 cc->nr_freepages = 0;
2635 VM_BUG_ON(free_pfn == 0);
2636 /* The cached pfn is always the first in a pageblock */
2637 free_pfn = pageblock_start_pfn(free_pfn);
2638 /*
2639 * Only go back, not forward. The cached pfn might have been
2640 * already reset to zone end in compact_finished()
2641 */
2642 if (free_pfn > cc->zone->compact_cached_free_pfn)
2643 cc->zone->compact_cached_free_pfn = free_pfn;
2644 }
2645
2646 count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned);
2647 count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned);
2648
2649 trace_mm_compaction_end(cc, start_pfn, end_pfn, sync, ret);
2650
2651 VM_BUG_ON(!list_empty(&cc->freepages));
2652 VM_BUG_ON(!list_empty(&cc->migratepages));
2653
2654 return ret;
2655}
2656
2657static enum compact_result compact_zone_order(struct zone *zone, int order,
2658 gfp_t gfp_mask, enum compact_priority prio,
2659 unsigned int alloc_flags, int highest_zoneidx,
2660 struct page **capture)
2661{
2662 enum compact_result ret;
2663 struct compact_control cc = {
2664 .order = order,
2665 .search_order = order,
2666 .gfp_mask = gfp_mask,
2667 .zone = zone,
2668 .mode = (prio == COMPACT_PRIO_ASYNC) ?
2669 MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT,
2670 .alloc_flags = alloc_flags,
2671 .highest_zoneidx = highest_zoneidx,
2672 .direct_compaction = true,
2673 .whole_zone = (prio == MIN_COMPACT_PRIORITY),
2674 .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
2675 .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
2676 };
2677 struct capture_control capc = {
2678 .cc = &cc,
2679 .page = NULL,
2680 };
2681
2682 /*
2683 * Make sure the structs are really initialized before we expose the
2684 * capture control, in case we are interrupted and the interrupt handler
2685 * frees a page.
2686 */
2687 barrier();
2688 WRITE_ONCE(current->capture_control, &capc);
2689
2690 ret = compact_zone(&cc, &capc);
2691
2692 /*
2693 * Make sure we hide capture control first before we read the captured
2694 * page pointer, otherwise an interrupt could free and capture a page
2695 * and we would leak it.
2696 */
2697 WRITE_ONCE(current->capture_control, NULL);
2698 *capture = READ_ONCE(capc.page);
2699 /*
2700 * Technically, it is also possible that compaction is skipped but
2701 * the page is still captured out of luck(IRQ came and freed the page).
2702 * Returning COMPACT_SUCCESS in such cases helps in properly accounting
2703 * the COMPACT[STALL|FAIL] when compaction is skipped.
2704 */
2705 if (*capture)
2706 ret = COMPACT_SUCCESS;
2707
2708 return ret;
2709}
2710
2711/**
2712 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
2713 * @gfp_mask: The GFP mask of the current allocation
2714 * @order: The order of the current allocation
2715 * @alloc_flags: The allocation flags of the current allocation
2716 * @ac: The context of current allocation
2717 * @prio: Determines how hard direct compaction should try to succeed
2718 * @capture: Pointer to free page created by compaction will be stored here
2719 *
2720 * This is the main entry point for direct page compaction.
2721 */
2722enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
2723 unsigned int alloc_flags, const struct alloc_context *ac,
2724 enum compact_priority prio, struct page **capture)
2725{
2726 struct zoneref *z;
2727 struct zone *zone;
2728 enum compact_result rc = COMPACT_SKIPPED;
2729
2730 if (!gfp_compaction_allowed(gfp_mask))
2731 return COMPACT_SKIPPED;
2732
2733 trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
2734
2735 /* Compact each zone in the list */
2736 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2737 ac->highest_zoneidx, ac->nodemask) {
2738 enum compact_result status;
2739
2740 if (prio > MIN_COMPACT_PRIORITY
2741 && compaction_deferred(zone, order)) {
2742 rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
2743 continue;
2744 }
2745
2746 status = compact_zone_order(zone, order, gfp_mask, prio,
2747 alloc_flags, ac->highest_zoneidx, capture);
2748 rc = max(status, rc);
2749
2750 /* The allocation should succeed, stop compacting */
2751 if (status == COMPACT_SUCCESS) {
2752 /*
2753 * We think the allocation will succeed in this zone,
2754 * but it is not certain, hence the false. The caller
2755 * will repeat this with true if allocation indeed
2756 * succeeds in this zone.
2757 */
2758 compaction_defer_reset(zone, order, false);
2759
2760 break;
2761 }
2762
2763 if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
2764 status == COMPACT_PARTIAL_SKIPPED))
2765 /*
2766 * We think that allocation won't succeed in this zone
2767 * so we defer compaction there. If it ends up
2768 * succeeding after all, it will be reset.
2769 */
2770 defer_compaction(zone, order);
2771
2772 /*
2773 * We might have stopped compacting due to need_resched() in
2774 * async compaction, or due to a fatal signal detected. In that
2775 * case do not try further zones
2776 */
2777 if ((prio == COMPACT_PRIO_ASYNC && need_resched())
2778 || fatal_signal_pending(current))
2779 break;
2780 }
2781
2782 return rc;
2783}
2784
2785/*
2786 * Compact all zones within a node till each zone's fragmentation score
2787 * reaches within proactive compaction thresholds (as determined by the
2788 * proactiveness tunable).
2789 *
2790 * It is possible that the function returns before reaching score targets
2791 * due to various back-off conditions, such as, contention on per-node or
2792 * per-zone locks.
2793 */
2794static void proactive_compact_node(pg_data_t *pgdat)
2795{
2796 int zoneid;
2797 struct zone *zone;
2798 struct compact_control cc = {
2799 .order = -1,
2800 .mode = MIGRATE_SYNC_LIGHT,
2801 .ignore_skip_hint = true,
2802 .whole_zone = true,
2803 .gfp_mask = GFP_KERNEL,
2804 .proactive_compaction = true,
2805 };
2806
2807 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2808 zone = &pgdat->node_zones[zoneid];
2809 if (!populated_zone(zone))
2810 continue;
2811
2812 cc.zone = zone;
2813
2814 compact_zone(&cc, NULL);
2815
2816 count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
2817 cc.total_migrate_scanned);
2818 count_compact_events(KCOMPACTD_FREE_SCANNED,
2819 cc.total_free_scanned);
2820 }
2821}
2822
2823/* Compact all zones within a node */
2824static void compact_node(int nid)
2825{
2826 pg_data_t *pgdat = NODE_DATA(nid);
2827 int zoneid;
2828 struct zone *zone;
2829 struct compact_control cc = {
2830 .order = -1,
2831 .mode = MIGRATE_SYNC,
2832 .ignore_skip_hint = true,
2833 .whole_zone = true,
2834 .gfp_mask = GFP_KERNEL,
2835 };
2836
2837
2838 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2839
2840 zone = &pgdat->node_zones[zoneid];
2841 if (!populated_zone(zone))
2842 continue;
2843
2844 cc.zone = zone;
2845
2846 compact_zone(&cc, NULL);
2847 }
2848}
2849
2850/* Compact all nodes in the system */
2851static void compact_nodes(void)
2852{
2853 int nid;
2854
2855 /* Flush pending updates to the LRU lists */
2856 lru_add_drain_all();
2857
2858 for_each_online_node(nid)
2859 compact_node(nid);
2860}
2861
2862static int compaction_proactiveness_sysctl_handler(struct ctl_table *table, int write,
2863 void *buffer, size_t *length, loff_t *ppos)
2864{
2865 int rc, nid;
2866
2867 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
2868 if (rc)
2869 return rc;
2870
2871 if (write && sysctl_compaction_proactiveness) {
2872 for_each_online_node(nid) {
2873 pg_data_t *pgdat = NODE_DATA(nid);
2874
2875 if (pgdat->proactive_compact_trigger)
2876 continue;
2877
2878 pgdat->proactive_compact_trigger = true;
2879 trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, -1,
2880 pgdat->nr_zones - 1);
2881 wake_up_interruptible(&pgdat->kcompactd_wait);
2882 }
2883 }
2884
2885 return 0;
2886}
2887
2888/*
2889 * This is the entry point for compacting all nodes via
2890 * /proc/sys/vm/compact_memory
2891 */
2892static int sysctl_compaction_handler(struct ctl_table *table, int write,
2893 void *buffer, size_t *length, loff_t *ppos)
2894{
2895 int ret;
2896
2897 ret = proc_dointvec(table, write, buffer, length, ppos);
2898 if (ret)
2899 return ret;
2900
2901 if (sysctl_compact_memory != 1)
2902 return -EINVAL;
2903
2904 if (write)
2905 compact_nodes();
2906
2907 return 0;
2908}
2909
2910#if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
2911static ssize_t compact_store(struct device *dev,
2912 struct device_attribute *attr,
2913 const char *buf, size_t count)
2914{
2915 int nid = dev->id;
2916
2917 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
2918 /* Flush pending updates to the LRU lists */
2919 lru_add_drain_all();
2920
2921 compact_node(nid);
2922 }
2923
2924 return count;
2925}
2926static DEVICE_ATTR_WO(compact);
2927
2928int compaction_register_node(struct node *node)
2929{
2930 return device_create_file(&node->dev, &dev_attr_compact);
2931}
2932
2933void compaction_unregister_node(struct node *node)
2934{
2935 device_remove_file(&node->dev, &dev_attr_compact);
2936}
2937#endif /* CONFIG_SYSFS && CONFIG_NUMA */
2938
2939static inline bool kcompactd_work_requested(pg_data_t *pgdat)
2940{
2941 return pgdat->kcompactd_max_order > 0 || kthread_should_stop() ||
2942 pgdat->proactive_compact_trigger;
2943}
2944
2945static bool kcompactd_node_suitable(pg_data_t *pgdat)
2946{
2947 int zoneid;
2948 struct zone *zone;
2949 enum zone_type highest_zoneidx = pgdat->kcompactd_highest_zoneidx;
2950 enum compact_result ret;
2951
2952 for (zoneid = 0; zoneid <= highest_zoneidx; zoneid++) {
2953 zone = &pgdat->node_zones[zoneid];
2954
2955 if (!populated_zone(zone))
2956 continue;
2957
2958 ret = compaction_suit_allocation_order(zone,
2959 pgdat->kcompactd_max_order,
2960 highest_zoneidx, ALLOC_WMARK_MIN);
2961 if (ret == COMPACT_CONTINUE)
2962 return true;
2963 }
2964
2965 return false;
2966}
2967
2968static void kcompactd_do_work(pg_data_t *pgdat)
2969{
2970 /*
2971 * With no special task, compact all zones so that a page of requested
2972 * order is allocatable.
2973 */
2974 int zoneid;
2975 struct zone *zone;
2976 struct compact_control cc = {
2977 .order = pgdat->kcompactd_max_order,
2978 .search_order = pgdat->kcompactd_max_order,
2979 .highest_zoneidx = pgdat->kcompactd_highest_zoneidx,
2980 .mode = MIGRATE_SYNC_LIGHT,
2981 .ignore_skip_hint = false,
2982 .gfp_mask = GFP_KERNEL,
2983 };
2984 enum compact_result ret;
2985
2986 trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
2987 cc.highest_zoneidx);
2988 count_compact_event(KCOMPACTD_WAKE);
2989
2990 for (zoneid = 0; zoneid <= cc.highest_zoneidx; zoneid++) {
2991 int status;
2992
2993 zone = &pgdat->node_zones[zoneid];
2994 if (!populated_zone(zone))
2995 continue;
2996
2997 if (compaction_deferred(zone, cc.order))
2998 continue;
2999
3000 ret = compaction_suit_allocation_order(zone,
3001 cc.order, zoneid, ALLOC_WMARK_MIN);
3002 if (ret != COMPACT_CONTINUE)
3003 continue;
3004
3005 if (kthread_should_stop())
3006 return;
3007
3008 cc.zone = zone;
3009 status = compact_zone(&cc, NULL);
3010
3011 if (status == COMPACT_SUCCESS) {
3012 compaction_defer_reset(zone, cc.order, false);
3013 } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
3014 /*
3015 * Buddy pages may become stranded on pcps that could
3016 * otherwise coalesce on the zone's free area for
3017 * order >= cc.order. This is ratelimited by the
3018 * upcoming deferral.
3019 */
3020 drain_all_pages(zone);
3021
3022 /*
3023 * We use sync migration mode here, so we defer like
3024 * sync direct compaction does.
3025 */
3026 defer_compaction(zone, cc.order);
3027 }
3028
3029 count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
3030 cc.total_migrate_scanned);
3031 count_compact_events(KCOMPACTD_FREE_SCANNED,
3032 cc.total_free_scanned);
3033 }
3034
3035 /*
3036 * Regardless of success, we are done until woken up next. But remember
3037 * the requested order/highest_zoneidx in case it was higher/tighter
3038 * than our current ones
3039 */
3040 if (pgdat->kcompactd_max_order <= cc.order)
3041 pgdat->kcompactd_max_order = 0;
3042 if (pgdat->kcompactd_highest_zoneidx >= cc.highest_zoneidx)
3043 pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
3044}
3045
3046void wakeup_kcompactd(pg_data_t *pgdat, int order, int highest_zoneidx)
3047{
3048 if (!order)
3049 return;
3050
3051 if (pgdat->kcompactd_max_order < order)
3052 pgdat->kcompactd_max_order = order;
3053
3054 if (pgdat->kcompactd_highest_zoneidx > highest_zoneidx)
3055 pgdat->kcompactd_highest_zoneidx = highest_zoneidx;
3056
3057 /*
3058 * Pairs with implicit barrier in wait_event_freezable()
3059 * such that wakeups are not missed.
3060 */
3061 if (!wq_has_sleeper(&pgdat->kcompactd_wait))
3062 return;
3063
3064 if (!kcompactd_node_suitable(pgdat))
3065 return;
3066
3067 trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
3068 highest_zoneidx);
3069 wake_up_interruptible(&pgdat->kcompactd_wait);
3070}
3071
3072/*
3073 * The background compaction daemon, started as a kernel thread
3074 * from the init process.
3075 */
3076static int kcompactd(void *p)
3077{
3078 pg_data_t *pgdat = (pg_data_t *)p;
3079 struct task_struct *tsk = current;
3080 long default_timeout = msecs_to_jiffies(HPAGE_FRAG_CHECK_INTERVAL_MSEC);
3081 long timeout = default_timeout;
3082
3083 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3084
3085 if (!cpumask_empty(cpumask))
3086 set_cpus_allowed_ptr(tsk, cpumask);
3087
3088 set_freezable();
3089
3090 pgdat->kcompactd_max_order = 0;
3091 pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
3092
3093 while (!kthread_should_stop()) {
3094 unsigned long pflags;
3095
3096 /*
3097 * Avoid the unnecessary wakeup for proactive compaction
3098 * when it is disabled.
3099 */
3100 if (!sysctl_compaction_proactiveness)
3101 timeout = MAX_SCHEDULE_TIMEOUT;
3102 trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
3103 if (wait_event_freezable_timeout(pgdat->kcompactd_wait,
3104 kcompactd_work_requested(pgdat), timeout) &&
3105 !pgdat->proactive_compact_trigger) {
3106
3107 psi_memstall_enter(&pflags);
3108 kcompactd_do_work(pgdat);
3109 psi_memstall_leave(&pflags);
3110 /*
3111 * Reset the timeout value. The defer timeout from
3112 * proactive compaction is lost here but that is fine
3113 * as the condition of the zone changing substantionally
3114 * then carrying on with the previous defer interval is
3115 * not useful.
3116 */
3117 timeout = default_timeout;
3118 continue;
3119 }
3120
3121 /*
3122 * Start the proactive work with default timeout. Based
3123 * on the fragmentation score, this timeout is updated.
3124 */
3125 timeout = default_timeout;
3126 if (should_proactive_compact_node(pgdat)) {
3127 unsigned int prev_score, score;
3128
3129 prev_score = fragmentation_score_node(pgdat);
3130 proactive_compact_node(pgdat);
3131 score = fragmentation_score_node(pgdat);
3132 /*
3133 * Defer proactive compaction if the fragmentation
3134 * score did not go down i.e. no progress made.
3135 */
3136 if (unlikely(score >= prev_score))
3137 timeout =
3138 default_timeout << COMPACT_MAX_DEFER_SHIFT;
3139 }
3140 if (unlikely(pgdat->proactive_compact_trigger))
3141 pgdat->proactive_compact_trigger = false;
3142 }
3143
3144 return 0;
3145}
3146
3147/*
3148 * This kcompactd start function will be called by init and node-hot-add.
3149 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
3150 */
3151void __meminit kcompactd_run(int nid)
3152{
3153 pg_data_t *pgdat = NODE_DATA(nid);
3154
3155 if (pgdat->kcompactd)
3156 return;
3157
3158 pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
3159 if (IS_ERR(pgdat->kcompactd)) {
3160 pr_err("Failed to start kcompactd on node %d\n", nid);
3161 pgdat->kcompactd = NULL;
3162 }
3163}
3164
3165/*
3166 * Called by memory hotplug when all memory in a node is offlined. Caller must
3167 * be holding mem_hotplug_begin/done().
3168 */
3169void __meminit kcompactd_stop(int nid)
3170{
3171 struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
3172
3173 if (kcompactd) {
3174 kthread_stop(kcompactd);
3175 NODE_DATA(nid)->kcompactd = NULL;
3176 }
3177}
3178
3179/*
3180 * It's optimal to keep kcompactd on the same CPUs as their memory, but
3181 * not required for correctness. So if the last cpu in a node goes
3182 * away, we get changed to run anywhere: as the first one comes back,
3183 * restore their cpu bindings.
3184 */
3185static int kcompactd_cpu_online(unsigned int cpu)
3186{
3187 int nid;
3188
3189 for_each_node_state(nid, N_MEMORY) {
3190 pg_data_t *pgdat = NODE_DATA(nid);
3191 const struct cpumask *mask;
3192
3193 mask = cpumask_of_node(pgdat->node_id);
3194
3195 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3196 /* One of our CPUs online: restore mask */
3197 if (pgdat->kcompactd)
3198 set_cpus_allowed_ptr(pgdat->kcompactd, mask);
3199 }
3200 return 0;
3201}
3202
3203static int proc_dointvec_minmax_warn_RT_change(struct ctl_table *table,
3204 int write, void *buffer, size_t *lenp, loff_t *ppos)
3205{
3206 int ret, old;
3207
3208 if (!IS_ENABLED(CONFIG_PREEMPT_RT) || !write)
3209 return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
3210
3211 old = *(int *)table->data;
3212 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
3213 if (ret)
3214 return ret;
3215 if (old != *(int *)table->data)
3216 pr_warn_once("sysctl attribute %s changed by %s[%d]\n",
3217 table->procname, current->comm,
3218 task_pid_nr(current));
3219 return ret;
3220}
3221
3222static struct ctl_table vm_compaction[] = {
3223 {
3224 .procname = "compact_memory",
3225 .data = &sysctl_compact_memory,
3226 .maxlen = sizeof(int),
3227 .mode = 0200,
3228 .proc_handler = sysctl_compaction_handler,
3229 },
3230 {
3231 .procname = "compaction_proactiveness",
3232 .data = &sysctl_compaction_proactiveness,
3233 .maxlen = sizeof(sysctl_compaction_proactiveness),
3234 .mode = 0644,
3235 .proc_handler = compaction_proactiveness_sysctl_handler,
3236 .extra1 = SYSCTL_ZERO,
3237 .extra2 = SYSCTL_ONE_HUNDRED,
3238 },
3239 {
3240 .procname = "extfrag_threshold",
3241 .data = &sysctl_extfrag_threshold,
3242 .maxlen = sizeof(int),
3243 .mode = 0644,
3244 .proc_handler = proc_dointvec_minmax,
3245 .extra1 = SYSCTL_ZERO,
3246 .extra2 = SYSCTL_ONE_THOUSAND,
3247 },
3248 {
3249 .procname = "compact_unevictable_allowed",
3250 .data = &sysctl_compact_unevictable_allowed,
3251 .maxlen = sizeof(int),
3252 .mode = 0644,
3253 .proc_handler = proc_dointvec_minmax_warn_RT_change,
3254 .extra1 = SYSCTL_ZERO,
3255 .extra2 = SYSCTL_ONE,
3256 },
3257 { }
3258};
3259
3260static int __init kcompactd_init(void)
3261{
3262 int nid;
3263 int ret;
3264
3265 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3266 "mm/compaction:online",
3267 kcompactd_cpu_online, NULL);
3268 if (ret < 0) {
3269 pr_err("kcompactd: failed to register hotplug callbacks.\n");
3270 return ret;
3271 }
3272
3273 for_each_node_state(nid, N_MEMORY)
3274 kcompactd_run(nid);
3275 register_sysctl_init("vm", vm_compaction);
3276 return 0;
3277}
3278subsys_initcall(kcompactd_init)
3279
3280#endif /* CONFIG_COMPACTION */
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 <linux/cpuset.h>
27#include "internal.h"
28
29#ifdef CONFIG_COMPACTION
30/*
31 * Fragmentation score check interval for proactive compaction purposes.
32 */
33#define HPAGE_FRAG_CHECK_INTERVAL_MSEC (500)
34
35static inline void count_compact_event(enum vm_event_item item)
36{
37 count_vm_event(item);
38}
39
40static inline void count_compact_events(enum vm_event_item item, long delta)
41{
42 count_vm_events(item, delta);
43}
44
45/*
46 * order == -1 is expected when compacting proactively via
47 * 1. /proc/sys/vm/compact_memory
48 * 2. /sys/devices/system/node/nodex/compact
49 * 3. /proc/sys/vm/compaction_proactiveness
50 */
51static inline bool is_via_compact_memory(int order)
52{
53 return order == -1;
54}
55
56#else
57#define count_compact_event(item) do { } while (0)
58#define count_compact_events(item, delta) do { } while (0)
59static inline bool is_via_compact_memory(int order) { return false; }
60#endif
61
62#if defined CONFIG_COMPACTION || defined CONFIG_CMA
63
64#define CREATE_TRACE_POINTS
65#include <trace/events/compaction.h>
66
67#define block_start_pfn(pfn, order) round_down(pfn, 1UL << (order))
68#define block_end_pfn(pfn, order) ALIGN((pfn) + 1, 1UL << (order))
69
70/*
71 * Page order with-respect-to which proactive compaction
72 * calculates external fragmentation, which is used as
73 * the "fragmentation score" of a node/zone.
74 */
75#if defined CONFIG_TRANSPARENT_HUGEPAGE
76#define COMPACTION_HPAGE_ORDER HPAGE_PMD_ORDER
77#elif defined CONFIG_HUGETLBFS
78#define COMPACTION_HPAGE_ORDER HUGETLB_PAGE_ORDER
79#else
80#define COMPACTION_HPAGE_ORDER (PMD_SHIFT - PAGE_SHIFT)
81#endif
82
83static struct page *mark_allocated_noprof(struct page *page, unsigned int order, gfp_t gfp_flags)
84{
85 post_alloc_hook(page, order, __GFP_MOVABLE);
86 return page;
87}
88#define mark_allocated(...) alloc_hooks(mark_allocated_noprof(__VA_ARGS__))
89
90static unsigned long release_free_list(struct list_head *freepages)
91{
92 int order;
93 unsigned long high_pfn = 0;
94
95 for (order = 0; order < NR_PAGE_ORDERS; order++) {
96 struct page *page, *next;
97
98 list_for_each_entry_safe(page, next, &freepages[order], lru) {
99 unsigned long pfn = page_to_pfn(page);
100
101 list_del(&page->lru);
102 /*
103 * Convert free pages into post allocation pages, so
104 * that we can free them via __free_page.
105 */
106 mark_allocated(page, order, __GFP_MOVABLE);
107 __free_pages(page, order);
108 if (pfn > high_pfn)
109 high_pfn = pfn;
110 }
111 }
112 return high_pfn;
113}
114
115#ifdef CONFIG_COMPACTION
116bool PageMovable(struct page *page)
117{
118 const struct movable_operations *mops;
119
120 VM_BUG_ON_PAGE(!PageLocked(page), page);
121 if (!__PageMovable(page))
122 return false;
123
124 mops = page_movable_ops(page);
125 if (mops)
126 return true;
127
128 return false;
129}
130
131void __SetPageMovable(struct page *page, const struct movable_operations *mops)
132{
133 VM_BUG_ON_PAGE(!PageLocked(page), page);
134 VM_BUG_ON_PAGE((unsigned long)mops & PAGE_MAPPING_MOVABLE, page);
135 page->mapping = (void *)((unsigned long)mops | PAGE_MAPPING_MOVABLE);
136}
137EXPORT_SYMBOL(__SetPageMovable);
138
139void __ClearPageMovable(struct page *page)
140{
141 VM_BUG_ON_PAGE(!PageMovable(page), page);
142 /*
143 * This page still has the type of a movable page, but it's
144 * actually not movable any more.
145 */
146 page->mapping = (void *)PAGE_MAPPING_MOVABLE;
147}
148EXPORT_SYMBOL(__ClearPageMovable);
149
150/* Do not skip compaction more than 64 times */
151#define COMPACT_MAX_DEFER_SHIFT 6
152
153/*
154 * Compaction is deferred when compaction fails to result in a page
155 * allocation success. 1 << compact_defer_shift, compactions are skipped up
156 * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
157 */
158static void defer_compaction(struct zone *zone, int order)
159{
160 zone->compact_considered = 0;
161 zone->compact_defer_shift++;
162
163 if (order < zone->compact_order_failed)
164 zone->compact_order_failed = order;
165
166 if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
167 zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
168
169 trace_mm_compaction_defer_compaction(zone, order);
170}
171
172/* Returns true if compaction should be skipped this time */
173static bool compaction_deferred(struct zone *zone, int order)
174{
175 unsigned long defer_limit = 1UL << zone->compact_defer_shift;
176
177 if (order < zone->compact_order_failed)
178 return false;
179
180 /* Avoid possible overflow */
181 if (++zone->compact_considered >= defer_limit) {
182 zone->compact_considered = defer_limit;
183 return false;
184 }
185
186 trace_mm_compaction_deferred(zone, order);
187
188 return true;
189}
190
191/*
192 * Update defer tracking counters after successful compaction of given order,
193 * which means an allocation either succeeded (alloc_success == true) or is
194 * expected to succeed.
195 */
196void compaction_defer_reset(struct zone *zone, int order,
197 bool alloc_success)
198{
199 if (alloc_success) {
200 zone->compact_considered = 0;
201 zone->compact_defer_shift = 0;
202 }
203 if (order >= zone->compact_order_failed)
204 zone->compact_order_failed = order + 1;
205
206 trace_mm_compaction_defer_reset(zone, order);
207}
208
209/* Returns true if restarting compaction after many failures */
210static bool compaction_restarting(struct zone *zone, int order)
211{
212 if (order < zone->compact_order_failed)
213 return false;
214
215 return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
216 zone->compact_considered >= 1UL << zone->compact_defer_shift;
217}
218
219/* Returns true if the pageblock should be scanned for pages to isolate. */
220static inline bool isolation_suitable(struct compact_control *cc,
221 struct page *page)
222{
223 if (cc->ignore_skip_hint)
224 return true;
225
226 return !get_pageblock_skip(page);
227}
228
229static void reset_cached_positions(struct zone *zone)
230{
231 zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
232 zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
233 zone->compact_cached_free_pfn =
234 pageblock_start_pfn(zone_end_pfn(zone) - 1);
235}
236
237#ifdef CONFIG_SPARSEMEM
238/*
239 * If the PFN falls into an offline section, return the start PFN of the
240 * next online section. If the PFN falls into an online section or if
241 * there is no next online section, return 0.
242 */
243static unsigned long skip_offline_sections(unsigned long start_pfn)
244{
245 unsigned long start_nr = pfn_to_section_nr(start_pfn);
246
247 if (online_section_nr(start_nr))
248 return 0;
249
250 while (++start_nr <= __highest_present_section_nr) {
251 if (online_section_nr(start_nr))
252 return section_nr_to_pfn(start_nr);
253 }
254
255 return 0;
256}
257
258/*
259 * If the PFN falls into an offline section, return the end PFN of the
260 * next online section in reverse. If the PFN falls into an online section
261 * or if there is no next online section in reverse, return 0.
262 */
263static unsigned long skip_offline_sections_reverse(unsigned long start_pfn)
264{
265 unsigned long start_nr = pfn_to_section_nr(start_pfn);
266
267 if (!start_nr || online_section_nr(start_nr))
268 return 0;
269
270 while (start_nr-- > 0) {
271 if (online_section_nr(start_nr))
272 return section_nr_to_pfn(start_nr) + PAGES_PER_SECTION;
273 }
274
275 return 0;
276}
277#else
278static unsigned long skip_offline_sections(unsigned long start_pfn)
279{
280 return 0;
281}
282
283static unsigned long skip_offline_sections_reverse(unsigned long start_pfn)
284{
285 return 0;
286}
287#endif
288
289/*
290 * Compound pages of >= pageblock_order should consistently be skipped until
291 * released. It is always pointless to compact pages of such order (if they are
292 * migratable), and the pageblocks they occupy cannot contain any free pages.
293 */
294static bool pageblock_skip_persistent(struct page *page)
295{
296 if (!PageCompound(page))
297 return false;
298
299 page = compound_head(page);
300
301 if (compound_order(page) >= pageblock_order)
302 return true;
303
304 return false;
305}
306
307static bool
308__reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source,
309 bool check_target)
310{
311 struct page *page = pfn_to_online_page(pfn);
312 struct page *block_page;
313 struct page *end_page;
314 unsigned long block_pfn;
315
316 if (!page)
317 return false;
318 if (zone != page_zone(page))
319 return false;
320 if (pageblock_skip_persistent(page))
321 return false;
322
323 /*
324 * If skip is already cleared do no further checking once the
325 * restart points have been set.
326 */
327 if (check_source && check_target && !get_pageblock_skip(page))
328 return true;
329
330 /*
331 * If clearing skip for the target scanner, do not select a
332 * non-movable pageblock as the starting point.
333 */
334 if (!check_source && check_target &&
335 get_pageblock_migratetype(page) != MIGRATE_MOVABLE)
336 return false;
337
338 /* Ensure the start of the pageblock or zone is online and valid */
339 block_pfn = pageblock_start_pfn(pfn);
340 block_pfn = max(block_pfn, zone->zone_start_pfn);
341 block_page = pfn_to_online_page(block_pfn);
342 if (block_page) {
343 page = block_page;
344 pfn = block_pfn;
345 }
346
347 /* Ensure the end of the pageblock or zone is online and valid */
348 block_pfn = pageblock_end_pfn(pfn) - 1;
349 block_pfn = min(block_pfn, zone_end_pfn(zone) - 1);
350 end_page = pfn_to_online_page(block_pfn);
351 if (!end_page)
352 return false;
353
354 /*
355 * Only clear the hint if a sample indicates there is either a
356 * free page or an LRU page in the block. One or other condition
357 * is necessary for the block to be a migration source/target.
358 */
359 do {
360 if (check_source && PageLRU(page)) {
361 clear_pageblock_skip(page);
362 return true;
363 }
364
365 if (check_target && PageBuddy(page)) {
366 clear_pageblock_skip(page);
367 return true;
368 }
369
370 page += (1 << PAGE_ALLOC_COSTLY_ORDER);
371 } while (page <= end_page);
372
373 return false;
374}
375
376/*
377 * This function is called to clear all cached information on pageblocks that
378 * should be skipped for page isolation when the migrate and free page scanner
379 * meet.
380 */
381static void __reset_isolation_suitable(struct zone *zone)
382{
383 unsigned long migrate_pfn = zone->zone_start_pfn;
384 unsigned long free_pfn = zone_end_pfn(zone) - 1;
385 unsigned long reset_migrate = free_pfn;
386 unsigned long reset_free = migrate_pfn;
387 bool source_set = false;
388 bool free_set = false;
389
390 /* Only flush if a full compaction finished recently */
391 if (!zone->compact_blockskip_flush)
392 return;
393
394 zone->compact_blockskip_flush = false;
395
396 /*
397 * Walk the zone and update pageblock skip information. Source looks
398 * for PageLRU while target looks for PageBuddy. When the scanner
399 * is found, both PageBuddy and PageLRU are checked as the pageblock
400 * is suitable as both source and target.
401 */
402 for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages,
403 free_pfn -= pageblock_nr_pages) {
404 cond_resched();
405
406 /* Update the migrate PFN */
407 if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) &&
408 migrate_pfn < reset_migrate) {
409 source_set = true;
410 reset_migrate = migrate_pfn;
411 zone->compact_init_migrate_pfn = reset_migrate;
412 zone->compact_cached_migrate_pfn[0] = reset_migrate;
413 zone->compact_cached_migrate_pfn[1] = reset_migrate;
414 }
415
416 /* Update the free PFN */
417 if (__reset_isolation_pfn(zone, free_pfn, free_set, true) &&
418 free_pfn > reset_free) {
419 free_set = true;
420 reset_free = free_pfn;
421 zone->compact_init_free_pfn = reset_free;
422 zone->compact_cached_free_pfn = reset_free;
423 }
424 }
425
426 /* Leave no distance if no suitable block was reset */
427 if (reset_migrate >= reset_free) {
428 zone->compact_cached_migrate_pfn[0] = migrate_pfn;
429 zone->compact_cached_migrate_pfn[1] = migrate_pfn;
430 zone->compact_cached_free_pfn = free_pfn;
431 }
432}
433
434void reset_isolation_suitable(pg_data_t *pgdat)
435{
436 int zoneid;
437
438 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
439 struct zone *zone = &pgdat->node_zones[zoneid];
440 if (!populated_zone(zone))
441 continue;
442
443 __reset_isolation_suitable(zone);
444 }
445}
446
447/*
448 * Sets the pageblock skip bit if it was clear. Note that this is a hint as
449 * locks are not required for read/writers. Returns true if it was already set.
450 */
451static bool test_and_set_skip(struct compact_control *cc, struct page *page)
452{
453 bool skip;
454
455 /* Do not update if skip hint is being ignored */
456 if (cc->ignore_skip_hint)
457 return false;
458
459 skip = get_pageblock_skip(page);
460 if (!skip && !cc->no_set_skip_hint)
461 set_pageblock_skip(page);
462
463 return skip;
464}
465
466static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
467{
468 struct zone *zone = cc->zone;
469
470 /* Set for isolation rather than compaction */
471 if (cc->no_set_skip_hint)
472 return;
473
474 pfn = pageblock_end_pfn(pfn);
475
476 /* Update where async and sync compaction should restart */
477 if (pfn > zone->compact_cached_migrate_pfn[0])
478 zone->compact_cached_migrate_pfn[0] = pfn;
479 if (cc->mode != MIGRATE_ASYNC &&
480 pfn > zone->compact_cached_migrate_pfn[1])
481 zone->compact_cached_migrate_pfn[1] = pfn;
482}
483
484/*
485 * If no pages were isolated then mark this pageblock to be skipped in the
486 * future. The information is later cleared by __reset_isolation_suitable().
487 */
488static void update_pageblock_skip(struct compact_control *cc,
489 struct page *page, unsigned long pfn)
490{
491 struct zone *zone = cc->zone;
492
493 if (cc->no_set_skip_hint)
494 return;
495
496 set_pageblock_skip(page);
497
498 if (pfn < zone->compact_cached_free_pfn)
499 zone->compact_cached_free_pfn = pfn;
500}
501#else
502static inline bool isolation_suitable(struct compact_control *cc,
503 struct page *page)
504{
505 return true;
506}
507
508static inline bool pageblock_skip_persistent(struct page *page)
509{
510 return false;
511}
512
513static inline void update_pageblock_skip(struct compact_control *cc,
514 struct page *page, unsigned long pfn)
515{
516}
517
518static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
519{
520}
521
522static bool test_and_set_skip(struct compact_control *cc, struct page *page)
523{
524 return false;
525}
526#endif /* CONFIG_COMPACTION */
527
528/*
529 * Compaction requires the taking of some coarse locks that are potentially
530 * very heavily contended. For async compaction, trylock and record if the
531 * lock is contended. The lock will still be acquired but compaction will
532 * abort when the current block is finished regardless of success rate.
533 * Sync compaction acquires the lock.
534 *
535 * Always returns true which makes it easier to track lock state in callers.
536 */
537static bool compact_lock_irqsave(spinlock_t *lock, unsigned long *flags,
538 struct compact_control *cc)
539 __acquires(lock)
540{
541 /* Track if the lock is contended in async mode */
542 if (cc->mode == MIGRATE_ASYNC && !cc->contended) {
543 if (spin_trylock_irqsave(lock, *flags))
544 return true;
545
546 cc->contended = true;
547 }
548
549 spin_lock_irqsave(lock, *flags);
550 return true;
551}
552
553/*
554 * Compaction requires the taking of some coarse locks that are potentially
555 * very heavily contended. The lock should be periodically unlocked to avoid
556 * having disabled IRQs for a long time, even when there is nobody waiting on
557 * the lock. It might also be that allowing the IRQs will result in
558 * need_resched() becoming true. If scheduling is needed, compaction schedules.
559 * Either compaction type will also abort if a fatal signal is pending.
560 * In either case if the lock was locked, it is dropped and not regained.
561 *
562 * Returns true if compaction should abort due to fatal signal pending.
563 * Returns false when compaction can continue.
564 */
565static bool compact_unlock_should_abort(spinlock_t *lock,
566 unsigned long flags, bool *locked, struct compact_control *cc)
567{
568 if (*locked) {
569 spin_unlock_irqrestore(lock, flags);
570 *locked = false;
571 }
572
573 if (fatal_signal_pending(current)) {
574 cc->contended = true;
575 return true;
576 }
577
578 cond_resched();
579
580 return false;
581}
582
583/*
584 * Isolate free pages onto a private freelist. If @strict is true, will abort
585 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
586 * (even though it may still end up isolating some pages).
587 */
588static unsigned long isolate_freepages_block(struct compact_control *cc,
589 unsigned long *start_pfn,
590 unsigned long end_pfn,
591 struct list_head *freelist,
592 unsigned int stride,
593 bool strict)
594{
595 int nr_scanned = 0, total_isolated = 0;
596 struct page *page;
597 unsigned long flags = 0;
598 bool locked = false;
599 unsigned long blockpfn = *start_pfn;
600 unsigned int order;
601
602 /* Strict mode is for isolation, speed is secondary */
603 if (strict)
604 stride = 1;
605
606 page = pfn_to_page(blockpfn);
607
608 /* Isolate free pages. */
609 for (; blockpfn < end_pfn; blockpfn += stride, page += stride) {
610 int isolated;
611
612 /*
613 * Periodically drop the lock (if held) regardless of its
614 * contention, to give chance to IRQs. Abort if fatal signal
615 * pending.
616 */
617 if (!(blockpfn % COMPACT_CLUSTER_MAX)
618 && compact_unlock_should_abort(&cc->zone->lock, flags,
619 &locked, cc))
620 break;
621
622 nr_scanned++;
623
624 /*
625 * For compound pages such as THP and hugetlbfs, we can save
626 * potentially a lot of iterations if we skip them at once.
627 * The check is racy, but we can consider only valid values
628 * and the only danger is skipping too much.
629 */
630 if (PageCompound(page)) {
631 const unsigned int order = compound_order(page);
632
633 if ((order <= MAX_PAGE_ORDER) &&
634 (blockpfn + (1UL << order) <= end_pfn)) {
635 blockpfn += (1UL << order) - 1;
636 page += (1UL << order) - 1;
637 nr_scanned += (1UL << order) - 1;
638 }
639
640 goto isolate_fail;
641 }
642
643 if (!PageBuddy(page))
644 goto isolate_fail;
645
646 /* If we already hold the lock, we can skip some rechecking. */
647 if (!locked) {
648 locked = compact_lock_irqsave(&cc->zone->lock,
649 &flags, cc);
650
651 /* Recheck this is a buddy page under lock */
652 if (!PageBuddy(page))
653 goto isolate_fail;
654 }
655
656 /* Found a free page, will break it into order-0 pages */
657 order = buddy_order(page);
658 isolated = __isolate_free_page(page, order);
659 if (!isolated)
660 break;
661 set_page_private(page, order);
662
663 nr_scanned += isolated - 1;
664 total_isolated += isolated;
665 cc->nr_freepages += isolated;
666 list_add_tail(&page->lru, &freelist[order]);
667
668 if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
669 blockpfn += isolated;
670 break;
671 }
672 /* Advance to the end of split page */
673 blockpfn += isolated - 1;
674 page += isolated - 1;
675 continue;
676
677isolate_fail:
678 if (strict)
679 break;
680
681 }
682
683 if (locked)
684 spin_unlock_irqrestore(&cc->zone->lock, flags);
685
686 /*
687 * Be careful to not go outside of the pageblock.
688 */
689 if (unlikely(blockpfn > end_pfn))
690 blockpfn = end_pfn;
691
692 trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
693 nr_scanned, total_isolated);
694
695 /* Record how far we have got within the block */
696 *start_pfn = blockpfn;
697
698 /*
699 * If strict isolation is requested by CMA then check that all the
700 * pages requested were isolated. If there were any failures, 0 is
701 * returned and CMA will fail.
702 */
703 if (strict && blockpfn < end_pfn)
704 total_isolated = 0;
705
706 cc->total_free_scanned += nr_scanned;
707 if (total_isolated)
708 count_compact_events(COMPACTISOLATED, total_isolated);
709 return total_isolated;
710}
711
712/**
713 * isolate_freepages_range() - isolate free pages.
714 * @cc: Compaction control structure.
715 * @start_pfn: The first PFN to start isolating.
716 * @end_pfn: The one-past-last PFN.
717 *
718 * Non-free pages, invalid PFNs, or zone boundaries within the
719 * [start_pfn, end_pfn) range are considered errors, cause function to
720 * undo its actions and return zero. cc->freepages[] are empty.
721 *
722 * Otherwise, function returns one-past-the-last PFN of isolated page
723 * (which may be greater then end_pfn if end fell in a middle of
724 * a free page). cc->freepages[] contain free pages isolated.
725 */
726unsigned long
727isolate_freepages_range(struct compact_control *cc,
728 unsigned long start_pfn, unsigned long end_pfn)
729{
730 unsigned long isolated, pfn, block_start_pfn, block_end_pfn;
731 int order;
732
733 for (order = 0; order < NR_PAGE_ORDERS; order++)
734 INIT_LIST_HEAD(&cc->freepages[order]);
735
736 pfn = start_pfn;
737 block_start_pfn = pageblock_start_pfn(pfn);
738 if (block_start_pfn < cc->zone->zone_start_pfn)
739 block_start_pfn = cc->zone->zone_start_pfn;
740 block_end_pfn = pageblock_end_pfn(pfn);
741
742 for (; pfn < end_pfn; pfn += isolated,
743 block_start_pfn = block_end_pfn,
744 block_end_pfn += pageblock_nr_pages) {
745 /* Protect pfn from changing by isolate_freepages_block */
746 unsigned long isolate_start_pfn = pfn;
747
748 /*
749 * pfn could pass the block_end_pfn if isolated freepage
750 * is more than pageblock order. In this case, we adjust
751 * scanning range to right one.
752 */
753 if (pfn >= block_end_pfn) {
754 block_start_pfn = pageblock_start_pfn(pfn);
755 block_end_pfn = pageblock_end_pfn(pfn);
756 }
757
758 block_end_pfn = min(block_end_pfn, end_pfn);
759
760 if (!pageblock_pfn_to_page(block_start_pfn,
761 block_end_pfn, cc->zone))
762 break;
763
764 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
765 block_end_pfn, cc->freepages, 0, true);
766
767 /*
768 * In strict mode, isolate_freepages_block() returns 0 if
769 * there are any holes in the block (ie. invalid PFNs or
770 * non-free pages).
771 */
772 if (!isolated)
773 break;
774
775 /*
776 * If we managed to isolate pages, it is always (1 << n) *
777 * pageblock_nr_pages for some non-negative n. (Max order
778 * page may span two pageblocks).
779 */
780 }
781
782 if (pfn < end_pfn) {
783 /* Loop terminated early, cleanup. */
784 release_free_list(cc->freepages);
785 return 0;
786 }
787
788 /* We don't use freelists for anything. */
789 return pfn;
790}
791
792/* Similar to reclaim, but different enough that they don't share logic */
793static bool too_many_isolated(struct compact_control *cc)
794{
795 pg_data_t *pgdat = cc->zone->zone_pgdat;
796 bool too_many;
797
798 unsigned long active, inactive, isolated;
799
800 inactive = node_page_state(pgdat, NR_INACTIVE_FILE) +
801 node_page_state(pgdat, NR_INACTIVE_ANON);
802 active = node_page_state(pgdat, NR_ACTIVE_FILE) +
803 node_page_state(pgdat, NR_ACTIVE_ANON);
804 isolated = node_page_state(pgdat, NR_ISOLATED_FILE) +
805 node_page_state(pgdat, NR_ISOLATED_ANON);
806
807 /*
808 * Allow GFP_NOFS to isolate past the limit set for regular
809 * compaction runs. This prevents an ABBA deadlock when other
810 * compactors have already isolated to the limit, but are
811 * blocked on filesystem locks held by the GFP_NOFS thread.
812 */
813 if (cc->gfp_mask & __GFP_FS) {
814 inactive >>= 3;
815 active >>= 3;
816 }
817
818 too_many = isolated > (inactive + active) / 2;
819 if (!too_many)
820 wake_throttle_isolated(pgdat);
821
822 return too_many;
823}
824
825/**
826 * skip_isolation_on_order() - determine when to skip folio isolation based on
827 * folio order and compaction target order
828 * @order: to-be-isolated folio order
829 * @target_order: compaction target order
830 *
831 * This avoids unnecessary folio isolations during compaction.
832 */
833static bool skip_isolation_on_order(int order, int target_order)
834{
835 /*
836 * Unless we are performing global compaction (i.e.,
837 * is_via_compact_memory), skip any folios that are larger than the
838 * target order: we wouldn't be here if we'd have a free folio with
839 * the desired target_order, so migrating this folio would likely fail
840 * later.
841 */
842 if (!is_via_compact_memory(target_order) && order >= target_order)
843 return true;
844 /*
845 * We limit memory compaction to pageblocks and won't try
846 * creating free blocks of memory that are larger than that.
847 */
848 return order >= pageblock_order;
849}
850
851/**
852 * isolate_migratepages_block() - isolate all migrate-able pages within
853 * a single pageblock
854 * @cc: Compaction control structure.
855 * @low_pfn: The first PFN to isolate
856 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
857 * @mode: Isolation mode to be used.
858 *
859 * Isolate all pages that can be migrated from the range specified by
860 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
861 * Returns errno, like -EAGAIN or -EINTR in case e.g signal pending or congestion,
862 * -ENOMEM in case we could not allocate a page, or 0.
863 * cc->migrate_pfn will contain the next pfn to scan.
864 *
865 * The pages are isolated on cc->migratepages list (not required to be empty),
866 * and cc->nr_migratepages is updated accordingly.
867 */
868static int
869isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
870 unsigned long end_pfn, isolate_mode_t mode)
871{
872 pg_data_t *pgdat = cc->zone->zone_pgdat;
873 unsigned long nr_scanned = 0, nr_isolated = 0;
874 struct lruvec *lruvec;
875 unsigned long flags = 0;
876 struct lruvec *locked = NULL;
877 struct folio *folio = NULL;
878 struct page *page = NULL, *valid_page = NULL;
879 struct address_space *mapping;
880 unsigned long start_pfn = low_pfn;
881 bool skip_on_failure = false;
882 unsigned long next_skip_pfn = 0;
883 bool skip_updated = false;
884 int ret = 0;
885
886 cc->migrate_pfn = low_pfn;
887
888 /*
889 * Ensure that there are not too many pages isolated from the LRU
890 * list by either parallel reclaimers or compaction. If there are,
891 * delay for some time until fewer pages are isolated
892 */
893 while (unlikely(too_many_isolated(cc))) {
894 /* stop isolation if there are still pages not migrated */
895 if (cc->nr_migratepages)
896 return -EAGAIN;
897
898 /* async migration should just abort */
899 if (cc->mode == MIGRATE_ASYNC)
900 return -EAGAIN;
901
902 reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED);
903
904 if (fatal_signal_pending(current))
905 return -EINTR;
906 }
907
908 cond_resched();
909
910 if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
911 skip_on_failure = true;
912 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
913 }
914
915 /* Time to isolate some pages for migration */
916 for (; low_pfn < end_pfn; low_pfn++) {
917 bool is_dirty, is_unevictable;
918
919 if (skip_on_failure && low_pfn >= next_skip_pfn) {
920 /*
921 * We have isolated all migration candidates in the
922 * previous order-aligned block, and did not skip it due
923 * to failure. We should migrate the pages now and
924 * hopefully succeed compaction.
925 */
926 if (nr_isolated)
927 break;
928
929 /*
930 * We failed to isolate in the previous order-aligned
931 * block. Set the new boundary to the end of the
932 * current block. Note we can't simply increase
933 * next_skip_pfn by 1 << order, as low_pfn might have
934 * been incremented by a higher number due to skipping
935 * a compound or a high-order buddy page in the
936 * previous loop iteration.
937 */
938 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
939 }
940
941 /*
942 * Periodically drop the lock (if held) regardless of its
943 * contention, to give chance to IRQs. Abort completely if
944 * a fatal signal is pending.
945 */
946 if (!(low_pfn % COMPACT_CLUSTER_MAX)) {
947 if (locked) {
948 unlock_page_lruvec_irqrestore(locked, flags);
949 locked = NULL;
950 }
951
952 if (fatal_signal_pending(current)) {
953 cc->contended = true;
954 ret = -EINTR;
955
956 goto fatal_pending;
957 }
958
959 cond_resched();
960 }
961
962 nr_scanned++;
963
964 page = pfn_to_page(low_pfn);
965
966 /*
967 * Check if the pageblock has already been marked skipped.
968 * Only the first PFN is checked as the caller isolates
969 * COMPACT_CLUSTER_MAX at a time so the second call must
970 * not falsely conclude that the block should be skipped.
971 */
972 if (!valid_page && (pageblock_aligned(low_pfn) ||
973 low_pfn == cc->zone->zone_start_pfn)) {
974 if (!isolation_suitable(cc, page)) {
975 low_pfn = end_pfn;
976 folio = NULL;
977 goto isolate_abort;
978 }
979 valid_page = page;
980 }
981
982 if (PageHuge(page)) {
983 /*
984 * skip hugetlbfs if we are not compacting for pages
985 * bigger than its order. THPs and other compound pages
986 * are handled below.
987 */
988 if (!cc->alloc_contig) {
989 const unsigned int order = compound_order(page);
990
991 if (order <= MAX_PAGE_ORDER) {
992 low_pfn += (1UL << order) - 1;
993 nr_scanned += (1UL << order) - 1;
994 }
995 goto isolate_fail;
996 }
997 /* for alloc_contig case */
998 if (locked) {
999 unlock_page_lruvec_irqrestore(locked, flags);
1000 locked = NULL;
1001 }
1002
1003 ret = isolate_or_dissolve_huge_page(page, &cc->migratepages);
1004
1005 /*
1006 * Fail isolation in case isolate_or_dissolve_huge_page()
1007 * reports an error. In case of -ENOMEM, abort right away.
1008 */
1009 if (ret < 0) {
1010 /* Do not report -EBUSY down the chain */
1011 if (ret == -EBUSY)
1012 ret = 0;
1013 low_pfn += compound_nr(page) - 1;
1014 nr_scanned += compound_nr(page) - 1;
1015 goto isolate_fail;
1016 }
1017
1018 if (PageHuge(page)) {
1019 /*
1020 * Hugepage was successfully isolated and placed
1021 * on the cc->migratepages list.
1022 */
1023 folio = page_folio(page);
1024 low_pfn += folio_nr_pages(folio) - 1;
1025 goto isolate_success_no_list;
1026 }
1027
1028 /*
1029 * Ok, the hugepage was dissolved. Now these pages are
1030 * Buddy and cannot be re-allocated because they are
1031 * isolated. Fall-through as the check below handles
1032 * Buddy pages.
1033 */
1034 }
1035
1036 /*
1037 * Skip if free. We read page order here without zone lock
1038 * which is generally unsafe, but the race window is small and
1039 * the worst thing that can happen is that we skip some
1040 * potential isolation targets.
1041 */
1042 if (PageBuddy(page)) {
1043 unsigned long freepage_order = buddy_order_unsafe(page);
1044
1045 /*
1046 * Without lock, we cannot be sure that what we got is
1047 * a valid page order. Consider only values in the
1048 * valid order range to prevent low_pfn overflow.
1049 */
1050 if (freepage_order > 0 && freepage_order <= MAX_PAGE_ORDER) {
1051 low_pfn += (1UL << freepage_order) - 1;
1052 nr_scanned += (1UL << freepage_order) - 1;
1053 }
1054 continue;
1055 }
1056
1057 /*
1058 * Regardless of being on LRU, compound pages such as THP
1059 * (hugetlbfs is handled above) are not to be compacted unless
1060 * we are attempting an allocation larger than the compound
1061 * page size. We can potentially save a lot of iterations if we
1062 * skip them at once. The check is racy, but we can consider
1063 * only valid values and the only danger is skipping too much.
1064 */
1065 if (PageCompound(page) && !cc->alloc_contig) {
1066 const unsigned int order = compound_order(page);
1067
1068 /* Skip based on page order and compaction target order. */
1069 if (skip_isolation_on_order(order, cc->order)) {
1070 if (order <= MAX_PAGE_ORDER) {
1071 low_pfn += (1UL << order) - 1;
1072 nr_scanned += (1UL << order) - 1;
1073 }
1074 goto isolate_fail;
1075 }
1076 }
1077
1078 /*
1079 * Check may be lockless but that's ok as we recheck later.
1080 * It's possible to migrate LRU and non-lru movable pages.
1081 * Skip any other type of page
1082 */
1083 if (!PageLRU(page)) {
1084 /*
1085 * __PageMovable can return false positive so we need
1086 * to verify it under page_lock.
1087 */
1088 if (unlikely(__PageMovable(page)) &&
1089 !PageIsolated(page)) {
1090 if (locked) {
1091 unlock_page_lruvec_irqrestore(locked, flags);
1092 locked = NULL;
1093 }
1094
1095 if (isolate_movable_page(page, mode)) {
1096 folio = page_folio(page);
1097 goto isolate_success;
1098 }
1099 }
1100
1101 goto isolate_fail;
1102 }
1103
1104 /*
1105 * Be careful not to clear PageLRU until after we're
1106 * sure the page is not being freed elsewhere -- the
1107 * page release code relies on it.
1108 */
1109 folio = folio_get_nontail_page(page);
1110 if (unlikely(!folio))
1111 goto isolate_fail;
1112
1113 /*
1114 * Migration will fail if an anonymous page is pinned in memory,
1115 * so avoid taking lru_lock and isolating it unnecessarily in an
1116 * admittedly racy check.
1117 */
1118 mapping = folio_mapping(folio);
1119 if (!mapping && (folio_ref_count(folio) - 1) > folio_mapcount(folio))
1120 goto isolate_fail_put;
1121
1122 /*
1123 * Only allow to migrate anonymous pages in GFP_NOFS context
1124 * because those do not depend on fs locks.
1125 */
1126 if (!(cc->gfp_mask & __GFP_FS) && mapping)
1127 goto isolate_fail_put;
1128
1129 /* Only take pages on LRU: a check now makes later tests safe */
1130 if (!folio_test_lru(folio))
1131 goto isolate_fail_put;
1132
1133 is_unevictable = folio_test_unevictable(folio);
1134
1135 /* Compaction might skip unevictable pages but CMA takes them */
1136 if (!(mode & ISOLATE_UNEVICTABLE) && is_unevictable)
1137 goto isolate_fail_put;
1138
1139 /*
1140 * To minimise LRU disruption, the caller can indicate with
1141 * ISOLATE_ASYNC_MIGRATE that it only wants to isolate pages
1142 * it will be able to migrate without blocking - clean pages
1143 * for the most part. PageWriteback would require blocking.
1144 */
1145 if ((mode & ISOLATE_ASYNC_MIGRATE) && folio_test_writeback(folio))
1146 goto isolate_fail_put;
1147
1148 is_dirty = folio_test_dirty(folio);
1149
1150 if (((mode & ISOLATE_ASYNC_MIGRATE) && is_dirty) ||
1151 (mapping && is_unevictable)) {
1152 bool migrate_dirty = true;
1153 bool is_inaccessible;
1154
1155 /*
1156 * Only folios without mappings or that have
1157 * a ->migrate_folio callback are possible to migrate
1158 * without blocking.
1159 *
1160 * Folios from inaccessible mappings are not migratable.
1161 *
1162 * However, we can be racing with truncation, which can
1163 * free the mapping that we need to check. Truncation
1164 * holds the folio lock until after the folio is removed
1165 * from the page so holding it ourselves is sufficient.
1166 *
1167 * To avoid locking the folio just to check inaccessible,
1168 * assume every inaccessible folio is also unevictable,
1169 * which is a cheaper test. If our assumption goes
1170 * wrong, it's not a correctness bug, just potentially
1171 * wasted cycles.
1172 */
1173 if (!folio_trylock(folio))
1174 goto isolate_fail_put;
1175
1176 mapping = folio_mapping(folio);
1177 if ((mode & ISOLATE_ASYNC_MIGRATE) && is_dirty) {
1178 migrate_dirty = !mapping ||
1179 mapping->a_ops->migrate_folio;
1180 }
1181 is_inaccessible = mapping && mapping_inaccessible(mapping);
1182 folio_unlock(folio);
1183 if (!migrate_dirty || is_inaccessible)
1184 goto isolate_fail_put;
1185 }
1186
1187 /* Try isolate the folio */
1188 if (!folio_test_clear_lru(folio))
1189 goto isolate_fail_put;
1190
1191 lruvec = folio_lruvec(folio);
1192
1193 /* If we already hold the lock, we can skip some rechecking */
1194 if (lruvec != locked) {
1195 if (locked)
1196 unlock_page_lruvec_irqrestore(locked, flags);
1197
1198 compact_lock_irqsave(&lruvec->lru_lock, &flags, cc);
1199 locked = lruvec;
1200
1201 lruvec_memcg_debug(lruvec, folio);
1202
1203 /*
1204 * Try get exclusive access under lock. If marked for
1205 * skip, the scan is aborted unless the current context
1206 * is a rescan to reach the end of the pageblock.
1207 */
1208 if (!skip_updated && valid_page) {
1209 skip_updated = true;
1210 if (test_and_set_skip(cc, valid_page) &&
1211 !cc->finish_pageblock) {
1212 low_pfn = end_pfn;
1213 goto isolate_abort;
1214 }
1215 }
1216
1217 /*
1218 * Check LRU folio order under the lock
1219 */
1220 if (unlikely(skip_isolation_on_order(folio_order(folio),
1221 cc->order) &&
1222 !cc->alloc_contig)) {
1223 low_pfn += folio_nr_pages(folio) - 1;
1224 nr_scanned += folio_nr_pages(folio) - 1;
1225 folio_set_lru(folio);
1226 goto isolate_fail_put;
1227 }
1228 }
1229
1230 /* The folio is taken off the LRU */
1231 if (folio_test_large(folio))
1232 low_pfn += folio_nr_pages(folio) - 1;
1233
1234 /* Successfully isolated */
1235 lruvec_del_folio(lruvec, folio);
1236 node_stat_mod_folio(folio,
1237 NR_ISOLATED_ANON + folio_is_file_lru(folio),
1238 folio_nr_pages(folio));
1239
1240isolate_success:
1241 list_add(&folio->lru, &cc->migratepages);
1242isolate_success_no_list:
1243 cc->nr_migratepages += folio_nr_pages(folio);
1244 nr_isolated += folio_nr_pages(folio);
1245 nr_scanned += folio_nr_pages(folio) - 1;
1246
1247 /*
1248 * Avoid isolating too much unless this block is being
1249 * fully scanned (e.g. dirty/writeback pages, parallel allocation)
1250 * or a lock is contended. For contention, isolate quickly to
1251 * potentially remove one source of contention.
1252 */
1253 if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX &&
1254 !cc->finish_pageblock && !cc->contended) {
1255 ++low_pfn;
1256 break;
1257 }
1258
1259 continue;
1260
1261isolate_fail_put:
1262 /* Avoid potential deadlock in freeing page under lru_lock */
1263 if (locked) {
1264 unlock_page_lruvec_irqrestore(locked, flags);
1265 locked = NULL;
1266 }
1267 folio_put(folio);
1268
1269isolate_fail:
1270 if (!skip_on_failure && ret != -ENOMEM)
1271 continue;
1272
1273 /*
1274 * We have isolated some pages, but then failed. Release them
1275 * instead of migrating, as we cannot form the cc->order buddy
1276 * page anyway.
1277 */
1278 if (nr_isolated) {
1279 if (locked) {
1280 unlock_page_lruvec_irqrestore(locked, flags);
1281 locked = NULL;
1282 }
1283 putback_movable_pages(&cc->migratepages);
1284 cc->nr_migratepages = 0;
1285 nr_isolated = 0;
1286 }
1287
1288 if (low_pfn < next_skip_pfn) {
1289 low_pfn = next_skip_pfn - 1;
1290 /*
1291 * The check near the loop beginning would have updated
1292 * next_skip_pfn too, but this is a bit simpler.
1293 */
1294 next_skip_pfn += 1UL << cc->order;
1295 }
1296
1297 if (ret == -ENOMEM)
1298 break;
1299 }
1300
1301 /*
1302 * The PageBuddy() check could have potentially brought us outside
1303 * the range to be scanned.
1304 */
1305 if (unlikely(low_pfn > end_pfn))
1306 low_pfn = end_pfn;
1307
1308 folio = NULL;
1309
1310isolate_abort:
1311 if (locked)
1312 unlock_page_lruvec_irqrestore(locked, flags);
1313 if (folio) {
1314 folio_set_lru(folio);
1315 folio_put(folio);
1316 }
1317
1318 /*
1319 * Update the cached scanner pfn once the pageblock has been scanned.
1320 * Pages will either be migrated in which case there is no point
1321 * scanning in the near future or migration failed in which case the
1322 * failure reason may persist. The block is marked for skipping if
1323 * there were no pages isolated in the block or if the block is
1324 * rescanned twice in a row.
1325 */
1326 if (low_pfn == end_pfn && (!nr_isolated || cc->finish_pageblock)) {
1327 if (!cc->no_set_skip_hint && valid_page && !skip_updated)
1328 set_pageblock_skip(valid_page);
1329 update_cached_migrate(cc, low_pfn);
1330 }
1331
1332 trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
1333 nr_scanned, nr_isolated);
1334
1335fatal_pending:
1336 cc->total_migrate_scanned += nr_scanned;
1337 if (nr_isolated)
1338 count_compact_events(COMPACTISOLATED, nr_isolated);
1339
1340 cc->migrate_pfn = low_pfn;
1341
1342 return ret;
1343}
1344
1345/**
1346 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
1347 * @cc: Compaction control structure.
1348 * @start_pfn: The first PFN to start isolating.
1349 * @end_pfn: The one-past-last PFN.
1350 *
1351 * Returns -EAGAIN when contented, -EINTR in case of a signal pending, -ENOMEM
1352 * in case we could not allocate a page, or 0.
1353 */
1354int
1355isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
1356 unsigned long end_pfn)
1357{
1358 unsigned long pfn, block_start_pfn, block_end_pfn;
1359 int ret = 0;
1360
1361 /* Scan block by block. First and last block may be incomplete */
1362 pfn = start_pfn;
1363 block_start_pfn = pageblock_start_pfn(pfn);
1364 if (block_start_pfn < cc->zone->zone_start_pfn)
1365 block_start_pfn = cc->zone->zone_start_pfn;
1366 block_end_pfn = pageblock_end_pfn(pfn);
1367
1368 for (; pfn < end_pfn; pfn = block_end_pfn,
1369 block_start_pfn = block_end_pfn,
1370 block_end_pfn += pageblock_nr_pages) {
1371
1372 block_end_pfn = min(block_end_pfn, end_pfn);
1373
1374 if (!pageblock_pfn_to_page(block_start_pfn,
1375 block_end_pfn, cc->zone))
1376 continue;
1377
1378 ret = isolate_migratepages_block(cc, pfn, block_end_pfn,
1379 ISOLATE_UNEVICTABLE);
1380
1381 if (ret)
1382 break;
1383
1384 if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX)
1385 break;
1386 }
1387
1388 return ret;
1389}
1390
1391#endif /* CONFIG_COMPACTION || CONFIG_CMA */
1392#ifdef CONFIG_COMPACTION
1393
1394static bool suitable_migration_source(struct compact_control *cc,
1395 struct page *page)
1396{
1397 int block_mt;
1398
1399 if (pageblock_skip_persistent(page))
1400 return false;
1401
1402 if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction)
1403 return true;
1404
1405 block_mt = get_pageblock_migratetype(page);
1406
1407 if (cc->migratetype == MIGRATE_MOVABLE)
1408 return is_migrate_movable(block_mt);
1409 else
1410 return block_mt == cc->migratetype;
1411}
1412
1413/* Returns true if the page is within a block suitable for migration to */
1414static bool suitable_migration_target(struct compact_control *cc,
1415 struct page *page)
1416{
1417 /* If the page is a large free page, then disallow migration */
1418 if (PageBuddy(page)) {
1419 int order = cc->order > 0 ? cc->order : pageblock_order;
1420
1421 /*
1422 * We are checking page_order without zone->lock taken. But
1423 * the only small danger is that we skip a potentially suitable
1424 * pageblock, so it's not worth to check order for valid range.
1425 */
1426 if (buddy_order_unsafe(page) >= order)
1427 return false;
1428 }
1429
1430 if (cc->ignore_block_suitable)
1431 return true;
1432
1433 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1434 if (is_migrate_movable(get_pageblock_migratetype(page)))
1435 return true;
1436
1437 /* Otherwise skip the block */
1438 return false;
1439}
1440
1441static inline unsigned int
1442freelist_scan_limit(struct compact_control *cc)
1443{
1444 unsigned short shift = BITS_PER_LONG - 1;
1445
1446 return (COMPACT_CLUSTER_MAX >> min(shift, cc->fast_search_fail)) + 1;
1447}
1448
1449/*
1450 * Test whether the free scanner has reached the same or lower pageblock than
1451 * the migration scanner, and compaction should thus terminate.
1452 */
1453static inline bool compact_scanners_met(struct compact_control *cc)
1454{
1455 return (cc->free_pfn >> pageblock_order)
1456 <= (cc->migrate_pfn >> pageblock_order);
1457}
1458
1459/*
1460 * Used when scanning for a suitable migration target which scans freelists
1461 * in reverse. Reorders the list such as the unscanned pages are scanned
1462 * first on the next iteration of the free scanner
1463 */
1464static void
1465move_freelist_head(struct list_head *freelist, struct page *freepage)
1466{
1467 LIST_HEAD(sublist);
1468
1469 if (!list_is_first(&freepage->buddy_list, freelist)) {
1470 list_cut_before(&sublist, freelist, &freepage->buddy_list);
1471 list_splice_tail(&sublist, freelist);
1472 }
1473}
1474
1475/*
1476 * Similar to move_freelist_head except used by the migration scanner
1477 * when scanning forward. It's possible for these list operations to
1478 * move against each other if they search the free list exactly in
1479 * lockstep.
1480 */
1481static void
1482move_freelist_tail(struct list_head *freelist, struct page *freepage)
1483{
1484 LIST_HEAD(sublist);
1485
1486 if (!list_is_last(&freepage->buddy_list, freelist)) {
1487 list_cut_position(&sublist, freelist, &freepage->buddy_list);
1488 list_splice_tail(&sublist, freelist);
1489 }
1490}
1491
1492static void
1493fast_isolate_around(struct compact_control *cc, unsigned long pfn)
1494{
1495 unsigned long start_pfn, end_pfn;
1496 struct page *page;
1497
1498 /* Do not search around if there are enough pages already */
1499 if (cc->nr_freepages >= cc->nr_migratepages)
1500 return;
1501
1502 /* Minimise scanning during async compaction */
1503 if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC)
1504 return;
1505
1506 /* Pageblock boundaries */
1507 start_pfn = max(pageblock_start_pfn(pfn), cc->zone->zone_start_pfn);
1508 end_pfn = min(pageblock_end_pfn(pfn), zone_end_pfn(cc->zone));
1509
1510 page = pageblock_pfn_to_page(start_pfn, end_pfn, cc->zone);
1511 if (!page)
1512 return;
1513
1514 isolate_freepages_block(cc, &start_pfn, end_pfn, cc->freepages, 1, false);
1515
1516 /* Skip this pageblock in the future as it's full or nearly full */
1517 if (start_pfn == end_pfn && !cc->no_set_skip_hint)
1518 set_pageblock_skip(page);
1519}
1520
1521/* Search orders in round-robin fashion */
1522static int next_search_order(struct compact_control *cc, int order)
1523{
1524 order--;
1525 if (order < 0)
1526 order = cc->order - 1;
1527
1528 /* Search wrapped around? */
1529 if (order == cc->search_order) {
1530 cc->search_order--;
1531 if (cc->search_order < 0)
1532 cc->search_order = cc->order - 1;
1533 return -1;
1534 }
1535
1536 return order;
1537}
1538
1539static void fast_isolate_freepages(struct compact_control *cc)
1540{
1541 unsigned int limit = max(1U, freelist_scan_limit(cc) >> 1);
1542 unsigned int nr_scanned = 0, total_isolated = 0;
1543 unsigned long low_pfn, min_pfn, highest = 0;
1544 unsigned long nr_isolated = 0;
1545 unsigned long distance;
1546 struct page *page = NULL;
1547 bool scan_start = false;
1548 int order;
1549
1550 /* Full compaction passes in a negative order */
1551 if (cc->order <= 0)
1552 return;
1553
1554 /*
1555 * If starting the scan, use a deeper search and use the highest
1556 * PFN found if a suitable one is not found.
1557 */
1558 if (cc->free_pfn >= cc->zone->compact_init_free_pfn) {
1559 limit = pageblock_nr_pages >> 1;
1560 scan_start = true;
1561 }
1562
1563 /*
1564 * Preferred point is in the top quarter of the scan space but take
1565 * a pfn from the top half if the search is problematic.
1566 */
1567 distance = (cc->free_pfn - cc->migrate_pfn);
1568 low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2));
1569 min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1));
1570
1571 if (WARN_ON_ONCE(min_pfn > low_pfn))
1572 low_pfn = min_pfn;
1573
1574 /*
1575 * Search starts from the last successful isolation order or the next
1576 * order to search after a previous failure
1577 */
1578 cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order);
1579
1580 for (order = cc->search_order;
1581 !page && order >= 0;
1582 order = next_search_order(cc, order)) {
1583 struct free_area *area = &cc->zone->free_area[order];
1584 struct list_head *freelist;
1585 struct page *freepage;
1586 unsigned long flags;
1587 unsigned int order_scanned = 0;
1588 unsigned long high_pfn = 0;
1589
1590 if (!area->nr_free)
1591 continue;
1592
1593 spin_lock_irqsave(&cc->zone->lock, flags);
1594 freelist = &area->free_list[MIGRATE_MOVABLE];
1595 list_for_each_entry_reverse(freepage, freelist, buddy_list) {
1596 unsigned long pfn;
1597
1598 order_scanned++;
1599 nr_scanned++;
1600 pfn = page_to_pfn(freepage);
1601
1602 if (pfn >= highest)
1603 highest = max(pageblock_start_pfn(pfn),
1604 cc->zone->zone_start_pfn);
1605
1606 if (pfn >= low_pfn) {
1607 cc->fast_search_fail = 0;
1608 cc->search_order = order;
1609 page = freepage;
1610 break;
1611 }
1612
1613 if (pfn >= min_pfn && pfn > high_pfn) {
1614 high_pfn = pfn;
1615
1616 /* Shorten the scan if a candidate is found */
1617 limit >>= 1;
1618 }
1619
1620 if (order_scanned >= limit)
1621 break;
1622 }
1623
1624 /* Use a maximum candidate pfn if a preferred one was not found */
1625 if (!page && high_pfn) {
1626 page = pfn_to_page(high_pfn);
1627
1628 /* Update freepage for the list reorder below */
1629 freepage = page;
1630 }
1631
1632 /* Reorder to so a future search skips recent pages */
1633 move_freelist_head(freelist, freepage);
1634
1635 /* Isolate the page if available */
1636 if (page) {
1637 if (__isolate_free_page(page, order)) {
1638 set_page_private(page, order);
1639 nr_isolated = 1 << order;
1640 nr_scanned += nr_isolated - 1;
1641 total_isolated += nr_isolated;
1642 cc->nr_freepages += nr_isolated;
1643 list_add_tail(&page->lru, &cc->freepages[order]);
1644 count_compact_events(COMPACTISOLATED, nr_isolated);
1645 } else {
1646 /* If isolation fails, abort the search */
1647 order = cc->search_order + 1;
1648 page = NULL;
1649 }
1650 }
1651
1652 spin_unlock_irqrestore(&cc->zone->lock, flags);
1653
1654 /* Skip fast search if enough freepages isolated */
1655 if (cc->nr_freepages >= cc->nr_migratepages)
1656 break;
1657
1658 /*
1659 * Smaller scan on next order so the total scan is related
1660 * to freelist_scan_limit.
1661 */
1662 if (order_scanned >= limit)
1663 limit = max(1U, limit >> 1);
1664 }
1665
1666 trace_mm_compaction_fast_isolate_freepages(min_pfn, cc->free_pfn,
1667 nr_scanned, total_isolated);
1668
1669 if (!page) {
1670 cc->fast_search_fail++;
1671 if (scan_start) {
1672 /*
1673 * Use the highest PFN found above min. If one was
1674 * not found, be pessimistic for direct compaction
1675 * and use the min mark.
1676 */
1677 if (highest >= min_pfn) {
1678 page = pfn_to_page(highest);
1679 cc->free_pfn = highest;
1680 } else {
1681 if (cc->direct_compaction && pfn_valid(min_pfn)) {
1682 page = pageblock_pfn_to_page(min_pfn,
1683 min(pageblock_end_pfn(min_pfn),
1684 zone_end_pfn(cc->zone)),
1685 cc->zone);
1686 if (page && !suitable_migration_target(cc, page))
1687 page = NULL;
1688
1689 cc->free_pfn = min_pfn;
1690 }
1691 }
1692 }
1693 }
1694
1695 if (highest && highest >= cc->zone->compact_cached_free_pfn) {
1696 highest -= pageblock_nr_pages;
1697 cc->zone->compact_cached_free_pfn = highest;
1698 }
1699
1700 cc->total_free_scanned += nr_scanned;
1701 if (!page)
1702 return;
1703
1704 low_pfn = page_to_pfn(page);
1705 fast_isolate_around(cc, low_pfn);
1706}
1707
1708/*
1709 * Based on information in the current compact_control, find blocks
1710 * suitable for isolating free pages from and then isolate them.
1711 */
1712static void isolate_freepages(struct compact_control *cc)
1713{
1714 struct zone *zone = cc->zone;
1715 struct page *page;
1716 unsigned long block_start_pfn; /* start of current pageblock */
1717 unsigned long isolate_start_pfn; /* exact pfn we start at */
1718 unsigned long block_end_pfn; /* end of current pageblock */
1719 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
1720 unsigned int stride;
1721
1722 /* Try a small search of the free lists for a candidate */
1723 fast_isolate_freepages(cc);
1724 if (cc->nr_freepages)
1725 return;
1726
1727 /*
1728 * Initialise the free scanner. The starting point is where we last
1729 * successfully isolated from, zone-cached value, or the end of the
1730 * zone when isolating for the first time. For looping we also need
1731 * this pfn aligned down to the pageblock boundary, because we do
1732 * block_start_pfn -= pageblock_nr_pages in the for loop.
1733 * For ending point, take care when isolating in last pageblock of a
1734 * zone which ends in the middle of a pageblock.
1735 * The low boundary is the end of the pageblock the migration scanner
1736 * is using.
1737 */
1738 isolate_start_pfn = cc->free_pfn;
1739 block_start_pfn = pageblock_start_pfn(isolate_start_pfn);
1740 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1741 zone_end_pfn(zone));
1742 low_pfn = pageblock_end_pfn(cc->migrate_pfn);
1743 stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1;
1744
1745 /*
1746 * Isolate free pages until enough are available to migrate the
1747 * pages on cc->migratepages. We stop searching if the migrate
1748 * and free page scanners meet or enough free pages are isolated.
1749 */
1750 for (; block_start_pfn >= low_pfn;
1751 block_end_pfn = block_start_pfn,
1752 block_start_pfn -= pageblock_nr_pages,
1753 isolate_start_pfn = block_start_pfn) {
1754 unsigned long nr_isolated;
1755
1756 /*
1757 * This can iterate a massively long zone without finding any
1758 * suitable migration targets, so periodically check resched.
1759 */
1760 if (!(block_start_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages)))
1761 cond_resched();
1762
1763 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1764 zone);
1765 if (!page) {
1766 unsigned long next_pfn;
1767
1768 next_pfn = skip_offline_sections_reverse(block_start_pfn);
1769 if (next_pfn)
1770 block_start_pfn = max(next_pfn, low_pfn);
1771
1772 continue;
1773 }
1774
1775 /* Check the block is suitable for migration */
1776 if (!suitable_migration_target(cc, page))
1777 continue;
1778
1779 /* If isolation recently failed, do not retry */
1780 if (!isolation_suitable(cc, page))
1781 continue;
1782
1783 /* Found a block suitable for isolating free pages from. */
1784 nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn,
1785 block_end_pfn, cc->freepages, stride, false);
1786
1787 /* Update the skip hint if the full pageblock was scanned */
1788 if (isolate_start_pfn == block_end_pfn)
1789 update_pageblock_skip(cc, page, block_start_pfn -
1790 pageblock_nr_pages);
1791
1792 /* Are enough freepages isolated? */
1793 if (cc->nr_freepages >= cc->nr_migratepages) {
1794 if (isolate_start_pfn >= block_end_pfn) {
1795 /*
1796 * Restart at previous pageblock if more
1797 * freepages can be isolated next time.
1798 */
1799 isolate_start_pfn =
1800 block_start_pfn - pageblock_nr_pages;
1801 }
1802 break;
1803 } else if (isolate_start_pfn < block_end_pfn) {
1804 /*
1805 * If isolation failed early, do not continue
1806 * needlessly.
1807 */
1808 break;
1809 }
1810
1811 /* Adjust stride depending on isolation */
1812 if (nr_isolated) {
1813 stride = 1;
1814 continue;
1815 }
1816 stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1);
1817 }
1818
1819 /*
1820 * Record where the free scanner will restart next time. Either we
1821 * broke from the loop and set isolate_start_pfn based on the last
1822 * call to isolate_freepages_block(), or we met the migration scanner
1823 * and the loop terminated due to isolate_start_pfn < low_pfn
1824 */
1825 cc->free_pfn = isolate_start_pfn;
1826}
1827
1828/*
1829 * This is a migrate-callback that "allocates" freepages by taking pages
1830 * from the isolated freelists in the block we are migrating to.
1831 */
1832static struct folio *compaction_alloc_noprof(struct folio *src, unsigned long data)
1833{
1834 struct compact_control *cc = (struct compact_control *)data;
1835 struct folio *dst;
1836 int order = folio_order(src);
1837 bool has_isolated_pages = false;
1838 int start_order;
1839 struct page *freepage;
1840 unsigned long size;
1841
1842again:
1843 for (start_order = order; start_order < NR_PAGE_ORDERS; start_order++)
1844 if (!list_empty(&cc->freepages[start_order]))
1845 break;
1846
1847 /* no free pages in the list */
1848 if (start_order == NR_PAGE_ORDERS) {
1849 if (has_isolated_pages)
1850 return NULL;
1851 isolate_freepages(cc);
1852 has_isolated_pages = true;
1853 goto again;
1854 }
1855
1856 freepage = list_first_entry(&cc->freepages[start_order], struct page,
1857 lru);
1858 size = 1 << start_order;
1859
1860 list_del(&freepage->lru);
1861
1862 while (start_order > order) {
1863 start_order--;
1864 size >>= 1;
1865
1866 list_add(&freepage[size].lru, &cc->freepages[start_order]);
1867 set_page_private(&freepage[size], start_order);
1868 }
1869 dst = (struct folio *)freepage;
1870
1871 post_alloc_hook(&dst->page, order, __GFP_MOVABLE);
1872 if (order)
1873 prep_compound_page(&dst->page, order);
1874 cc->nr_freepages -= 1 << order;
1875 cc->nr_migratepages -= 1 << order;
1876 return page_rmappable_folio(&dst->page);
1877}
1878
1879static struct folio *compaction_alloc(struct folio *src, unsigned long data)
1880{
1881 return alloc_hooks(compaction_alloc_noprof(src, data));
1882}
1883
1884/*
1885 * This is a migrate-callback that "frees" freepages back to the isolated
1886 * freelist. All pages on the freelist are from the same zone, so there is no
1887 * special handling needed for NUMA.
1888 */
1889static void compaction_free(struct folio *dst, unsigned long data)
1890{
1891 struct compact_control *cc = (struct compact_control *)data;
1892 int order = folio_order(dst);
1893 struct page *page = &dst->page;
1894
1895 if (folio_put_testzero(dst)) {
1896 free_pages_prepare(page, order);
1897 list_add(&dst->lru, &cc->freepages[order]);
1898 cc->nr_freepages += 1 << order;
1899 }
1900 cc->nr_migratepages += 1 << order;
1901 /*
1902 * someone else has referenced the page, we cannot take it back to our
1903 * free list.
1904 */
1905}
1906
1907/* possible outcome of isolate_migratepages */
1908typedef enum {
1909 ISOLATE_ABORT, /* Abort compaction now */
1910 ISOLATE_NONE, /* No pages isolated, continue scanning */
1911 ISOLATE_SUCCESS, /* Pages isolated, migrate */
1912} isolate_migrate_t;
1913
1914/*
1915 * Allow userspace to control policy on scanning the unevictable LRU for
1916 * compactable pages.
1917 */
1918static int sysctl_compact_unevictable_allowed __read_mostly = CONFIG_COMPACT_UNEVICTABLE_DEFAULT;
1919/*
1920 * Tunable for proactive compaction. It determines how
1921 * aggressively the kernel should compact memory in the
1922 * background. It takes values in the range [0, 100].
1923 */
1924static unsigned int __read_mostly sysctl_compaction_proactiveness = 20;
1925static int sysctl_extfrag_threshold = 500;
1926static int __read_mostly sysctl_compact_memory;
1927
1928static inline void
1929update_fast_start_pfn(struct compact_control *cc, unsigned long pfn)
1930{
1931 if (cc->fast_start_pfn == ULONG_MAX)
1932 return;
1933
1934 if (!cc->fast_start_pfn)
1935 cc->fast_start_pfn = pfn;
1936
1937 cc->fast_start_pfn = min(cc->fast_start_pfn, pfn);
1938}
1939
1940static inline unsigned long
1941reinit_migrate_pfn(struct compact_control *cc)
1942{
1943 if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX)
1944 return cc->migrate_pfn;
1945
1946 cc->migrate_pfn = cc->fast_start_pfn;
1947 cc->fast_start_pfn = ULONG_MAX;
1948
1949 return cc->migrate_pfn;
1950}
1951
1952/*
1953 * Briefly search the free lists for a migration source that already has
1954 * some free pages to reduce the number of pages that need migration
1955 * before a pageblock is free.
1956 */
1957static unsigned long fast_find_migrateblock(struct compact_control *cc)
1958{
1959 unsigned int limit = freelist_scan_limit(cc);
1960 unsigned int nr_scanned = 0;
1961 unsigned long distance;
1962 unsigned long pfn = cc->migrate_pfn;
1963 unsigned long high_pfn;
1964 int order;
1965 bool found_block = false;
1966
1967 /* Skip hints are relied on to avoid repeats on the fast search */
1968 if (cc->ignore_skip_hint)
1969 return pfn;
1970
1971 /*
1972 * If the pageblock should be finished then do not select a different
1973 * pageblock.
1974 */
1975 if (cc->finish_pageblock)
1976 return pfn;
1977
1978 /*
1979 * If the migrate_pfn is not at the start of a zone or the start
1980 * of a pageblock then assume this is a continuation of a previous
1981 * scan restarted due to COMPACT_CLUSTER_MAX.
1982 */
1983 if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn))
1984 return pfn;
1985
1986 /*
1987 * For smaller orders, just linearly scan as the number of pages
1988 * to migrate should be relatively small and does not necessarily
1989 * justify freeing up a large block for a small allocation.
1990 */
1991 if (cc->order <= PAGE_ALLOC_COSTLY_ORDER)
1992 return pfn;
1993
1994 /*
1995 * Only allow kcompactd and direct requests for movable pages to
1996 * quickly clear out a MOVABLE pageblock for allocation. This
1997 * reduces the risk that a large movable pageblock is freed for
1998 * an unmovable/reclaimable small allocation.
1999 */
2000 if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE)
2001 return pfn;
2002
2003 /*
2004 * When starting the migration scanner, pick any pageblock within the
2005 * first half of the search space. Otherwise try and pick a pageblock
2006 * within the first eighth to reduce the chances that a migration
2007 * target later becomes a source.
2008 */
2009 distance = (cc->free_pfn - cc->migrate_pfn) >> 1;
2010 if (cc->migrate_pfn != cc->zone->zone_start_pfn)
2011 distance >>= 2;
2012 high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance);
2013
2014 for (order = cc->order - 1;
2015 order >= PAGE_ALLOC_COSTLY_ORDER && !found_block && nr_scanned < limit;
2016 order--) {
2017 struct free_area *area = &cc->zone->free_area[order];
2018 struct list_head *freelist;
2019 unsigned long flags;
2020 struct page *freepage;
2021
2022 if (!area->nr_free)
2023 continue;
2024
2025 spin_lock_irqsave(&cc->zone->lock, flags);
2026 freelist = &area->free_list[MIGRATE_MOVABLE];
2027 list_for_each_entry(freepage, freelist, buddy_list) {
2028 unsigned long free_pfn;
2029
2030 if (nr_scanned++ >= limit) {
2031 move_freelist_tail(freelist, freepage);
2032 break;
2033 }
2034
2035 free_pfn = page_to_pfn(freepage);
2036 if (free_pfn < high_pfn) {
2037 /*
2038 * Avoid if skipped recently. Ideally it would
2039 * move to the tail but even safe iteration of
2040 * the list assumes an entry is deleted, not
2041 * reordered.
2042 */
2043 if (get_pageblock_skip(freepage))
2044 continue;
2045
2046 /* Reorder to so a future search skips recent pages */
2047 move_freelist_tail(freelist, freepage);
2048
2049 update_fast_start_pfn(cc, free_pfn);
2050 pfn = pageblock_start_pfn(free_pfn);
2051 if (pfn < cc->zone->zone_start_pfn)
2052 pfn = cc->zone->zone_start_pfn;
2053 cc->fast_search_fail = 0;
2054 found_block = true;
2055 break;
2056 }
2057 }
2058 spin_unlock_irqrestore(&cc->zone->lock, flags);
2059 }
2060
2061 cc->total_migrate_scanned += nr_scanned;
2062
2063 /*
2064 * If fast scanning failed then use a cached entry for a page block
2065 * that had free pages as the basis for starting a linear scan.
2066 */
2067 if (!found_block) {
2068 cc->fast_search_fail++;
2069 pfn = reinit_migrate_pfn(cc);
2070 }
2071 return pfn;
2072}
2073
2074/*
2075 * Isolate all pages that can be migrated from the first suitable block,
2076 * starting at the block pointed to by the migrate scanner pfn within
2077 * compact_control.
2078 */
2079static isolate_migrate_t isolate_migratepages(struct compact_control *cc)
2080{
2081 unsigned long block_start_pfn;
2082 unsigned long block_end_pfn;
2083 unsigned long low_pfn;
2084 struct page *page;
2085 const isolate_mode_t isolate_mode =
2086 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
2087 (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
2088 bool fast_find_block;
2089
2090 /*
2091 * Start at where we last stopped, or beginning of the zone as
2092 * initialized by compact_zone(). The first failure will use
2093 * the lowest PFN as the starting point for linear scanning.
2094 */
2095 low_pfn = fast_find_migrateblock(cc);
2096 block_start_pfn = pageblock_start_pfn(low_pfn);
2097 if (block_start_pfn < cc->zone->zone_start_pfn)
2098 block_start_pfn = cc->zone->zone_start_pfn;
2099
2100 /*
2101 * fast_find_migrateblock() has already ensured the pageblock is not
2102 * set with a skipped flag, so to avoid the isolation_suitable check
2103 * below again, check whether the fast search was successful.
2104 */
2105 fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail;
2106
2107 /* Only scan within a pageblock boundary */
2108 block_end_pfn = pageblock_end_pfn(low_pfn);
2109
2110 /*
2111 * Iterate over whole pageblocks until we find the first suitable.
2112 * Do not cross the free scanner.
2113 */
2114 for (; block_end_pfn <= cc->free_pfn;
2115 fast_find_block = false,
2116 cc->migrate_pfn = low_pfn = block_end_pfn,
2117 block_start_pfn = block_end_pfn,
2118 block_end_pfn += pageblock_nr_pages) {
2119
2120 /*
2121 * This can potentially iterate a massively long zone with
2122 * many pageblocks unsuitable, so periodically check if we
2123 * need to schedule.
2124 */
2125 if (!(low_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages)))
2126 cond_resched();
2127
2128 page = pageblock_pfn_to_page(block_start_pfn,
2129 block_end_pfn, cc->zone);
2130 if (!page) {
2131 unsigned long next_pfn;
2132
2133 next_pfn = skip_offline_sections(block_start_pfn);
2134 if (next_pfn)
2135 block_end_pfn = min(next_pfn, cc->free_pfn);
2136 continue;
2137 }
2138
2139 /*
2140 * If isolation recently failed, do not retry. Only check the
2141 * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock
2142 * to be visited multiple times. Assume skip was checked
2143 * before making it "skip" so other compaction instances do
2144 * not scan the same block.
2145 */
2146 if ((pageblock_aligned(low_pfn) ||
2147 low_pfn == cc->zone->zone_start_pfn) &&
2148 !fast_find_block && !isolation_suitable(cc, page))
2149 continue;
2150
2151 /*
2152 * For async direct compaction, only scan the pageblocks of the
2153 * same migratetype without huge pages. Async direct compaction
2154 * is optimistic to see if the minimum amount of work satisfies
2155 * the allocation. The cached PFN is updated as it's possible
2156 * that all remaining blocks between source and target are
2157 * unsuitable and the compaction scanners fail to meet.
2158 */
2159 if (!suitable_migration_source(cc, page)) {
2160 update_cached_migrate(cc, block_end_pfn);
2161 continue;
2162 }
2163
2164 /* Perform the isolation */
2165 if (isolate_migratepages_block(cc, low_pfn, block_end_pfn,
2166 isolate_mode))
2167 return ISOLATE_ABORT;
2168
2169 /*
2170 * Either we isolated something and proceed with migration. Or
2171 * we failed and compact_zone should decide if we should
2172 * continue or not.
2173 */
2174 break;
2175 }
2176
2177 return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
2178}
2179
2180/*
2181 * Determine whether kswapd is (or recently was!) running on this node.
2182 *
2183 * pgdat_kswapd_lock() pins pgdat->kswapd, so a concurrent kswapd_stop() can't
2184 * zero it.
2185 */
2186static bool kswapd_is_running(pg_data_t *pgdat)
2187{
2188 bool running;
2189
2190 pgdat_kswapd_lock(pgdat);
2191 running = pgdat->kswapd && task_is_running(pgdat->kswapd);
2192 pgdat_kswapd_unlock(pgdat);
2193
2194 return running;
2195}
2196
2197/*
2198 * A zone's fragmentation score is the external fragmentation wrt to the
2199 * COMPACTION_HPAGE_ORDER. It returns a value in the range [0, 100].
2200 */
2201static unsigned int fragmentation_score_zone(struct zone *zone)
2202{
2203 return extfrag_for_order(zone, COMPACTION_HPAGE_ORDER);
2204}
2205
2206/*
2207 * A weighted zone's fragmentation score is the external fragmentation
2208 * wrt to the COMPACTION_HPAGE_ORDER scaled by the zone's size. It
2209 * returns a value in the range [0, 100].
2210 *
2211 * The scaling factor ensures that proactive compaction focuses on larger
2212 * zones like ZONE_NORMAL, rather than smaller, specialized zones like
2213 * ZONE_DMA32. For smaller zones, the score value remains close to zero,
2214 * and thus never exceeds the high threshold for proactive compaction.
2215 */
2216static unsigned int fragmentation_score_zone_weighted(struct zone *zone)
2217{
2218 unsigned long score;
2219
2220 score = zone->present_pages * fragmentation_score_zone(zone);
2221 return div64_ul(score, zone->zone_pgdat->node_present_pages + 1);
2222}
2223
2224/*
2225 * The per-node proactive (background) compaction process is started by its
2226 * corresponding kcompactd thread when the node's fragmentation score
2227 * exceeds the high threshold. The compaction process remains active till
2228 * the node's score falls below the low threshold, or one of the back-off
2229 * conditions is met.
2230 */
2231static unsigned int fragmentation_score_node(pg_data_t *pgdat)
2232{
2233 unsigned int score = 0;
2234 int zoneid;
2235
2236 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2237 struct zone *zone;
2238
2239 zone = &pgdat->node_zones[zoneid];
2240 if (!populated_zone(zone))
2241 continue;
2242 score += fragmentation_score_zone_weighted(zone);
2243 }
2244
2245 return score;
2246}
2247
2248static unsigned int fragmentation_score_wmark(bool low)
2249{
2250 unsigned int wmark_low;
2251
2252 /*
2253 * Cap the low watermark to avoid excessive compaction
2254 * activity in case a user sets the proactiveness tunable
2255 * close to 100 (maximum).
2256 */
2257 wmark_low = max(100U - sysctl_compaction_proactiveness, 5U);
2258 return low ? wmark_low : min(wmark_low + 10, 100U);
2259}
2260
2261static bool should_proactive_compact_node(pg_data_t *pgdat)
2262{
2263 int wmark_high;
2264
2265 if (!sysctl_compaction_proactiveness || kswapd_is_running(pgdat))
2266 return false;
2267
2268 wmark_high = fragmentation_score_wmark(false);
2269 return fragmentation_score_node(pgdat) > wmark_high;
2270}
2271
2272static enum compact_result __compact_finished(struct compact_control *cc)
2273{
2274 unsigned int order;
2275 const int migratetype = cc->migratetype;
2276 int ret;
2277
2278 /* Compaction run completes if the migrate and free scanner meet */
2279 if (compact_scanners_met(cc)) {
2280 /* Let the next compaction start anew. */
2281 reset_cached_positions(cc->zone);
2282
2283 /*
2284 * Mark that the PG_migrate_skip information should be cleared
2285 * by kswapd when it goes to sleep. kcompactd does not set the
2286 * flag itself as the decision to be clear should be directly
2287 * based on an allocation request.
2288 */
2289 if (cc->direct_compaction)
2290 cc->zone->compact_blockskip_flush = true;
2291
2292 if (cc->whole_zone)
2293 return COMPACT_COMPLETE;
2294 else
2295 return COMPACT_PARTIAL_SKIPPED;
2296 }
2297
2298 if (cc->proactive_compaction) {
2299 int score, wmark_low;
2300 pg_data_t *pgdat;
2301
2302 pgdat = cc->zone->zone_pgdat;
2303 if (kswapd_is_running(pgdat))
2304 return COMPACT_PARTIAL_SKIPPED;
2305
2306 score = fragmentation_score_zone(cc->zone);
2307 wmark_low = fragmentation_score_wmark(true);
2308
2309 if (score > wmark_low)
2310 ret = COMPACT_CONTINUE;
2311 else
2312 ret = COMPACT_SUCCESS;
2313
2314 goto out;
2315 }
2316
2317 if (is_via_compact_memory(cc->order))
2318 return COMPACT_CONTINUE;
2319
2320 /*
2321 * Always finish scanning a pageblock to reduce the possibility of
2322 * fallbacks in the future. This is particularly important when
2323 * migration source is unmovable/reclaimable but it's not worth
2324 * special casing.
2325 */
2326 if (!pageblock_aligned(cc->migrate_pfn))
2327 return COMPACT_CONTINUE;
2328
2329 /* Direct compactor: Is a suitable page free? */
2330 ret = COMPACT_NO_SUITABLE_PAGE;
2331 for (order = cc->order; order < NR_PAGE_ORDERS; order++) {
2332 struct free_area *area = &cc->zone->free_area[order];
2333 bool can_steal;
2334
2335 /* Job done if page is free of the right migratetype */
2336 if (!free_area_empty(area, migratetype))
2337 return COMPACT_SUCCESS;
2338
2339#ifdef CONFIG_CMA
2340 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
2341 if (migratetype == MIGRATE_MOVABLE &&
2342 !free_area_empty(area, MIGRATE_CMA))
2343 return COMPACT_SUCCESS;
2344#endif
2345 /*
2346 * Job done if allocation would steal freepages from
2347 * other migratetype buddy lists.
2348 */
2349 if (find_suitable_fallback(area, order, migratetype,
2350 true, &can_steal) != -1)
2351 /*
2352 * Movable pages are OK in any pageblock. If we are
2353 * stealing for a non-movable allocation, make sure
2354 * we finish compacting the current pageblock first
2355 * (which is assured by the above migrate_pfn align
2356 * check) so it is as free as possible and we won't
2357 * have to steal another one soon.
2358 */
2359 return COMPACT_SUCCESS;
2360 }
2361
2362out:
2363 if (cc->contended || fatal_signal_pending(current))
2364 ret = COMPACT_CONTENDED;
2365
2366 return ret;
2367}
2368
2369static enum compact_result compact_finished(struct compact_control *cc)
2370{
2371 int ret;
2372
2373 ret = __compact_finished(cc);
2374 trace_mm_compaction_finished(cc->zone, cc->order, ret);
2375 if (ret == COMPACT_NO_SUITABLE_PAGE)
2376 ret = COMPACT_CONTINUE;
2377
2378 return ret;
2379}
2380
2381static bool __compaction_suitable(struct zone *zone, int order,
2382 int highest_zoneidx,
2383 unsigned long wmark_target)
2384{
2385 unsigned long watermark;
2386 /*
2387 * Watermarks for order-0 must be met for compaction to be able to
2388 * isolate free pages for migration targets. This means that the
2389 * watermark and alloc_flags have to match, or be more pessimistic than
2390 * the check in __isolate_free_page(). We don't use the direct
2391 * compactor's alloc_flags, as they are not relevant for freepage
2392 * isolation. We however do use the direct compactor's highest_zoneidx
2393 * to skip over zones where lowmem reserves would prevent allocation
2394 * even if compaction succeeds.
2395 * For costly orders, we require low watermark instead of min for
2396 * compaction to proceed to increase its chances.
2397 * ALLOC_CMA is used, as pages in CMA pageblocks are considered
2398 * suitable migration targets
2399 */
2400 watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
2401 low_wmark_pages(zone) : min_wmark_pages(zone);
2402 watermark += compact_gap(order);
2403 return __zone_watermark_ok(zone, 0, watermark, highest_zoneidx,
2404 ALLOC_CMA, wmark_target);
2405}
2406
2407/*
2408 * compaction_suitable: Is this suitable to run compaction on this zone now?
2409 */
2410bool compaction_suitable(struct zone *zone, int order, int highest_zoneidx)
2411{
2412 enum compact_result compact_result;
2413 bool suitable;
2414
2415 suitable = __compaction_suitable(zone, order, highest_zoneidx,
2416 zone_page_state(zone, NR_FREE_PAGES));
2417 /*
2418 * fragmentation index determines if allocation failures are due to
2419 * low memory or external fragmentation
2420 *
2421 * index of -1000 would imply allocations might succeed depending on
2422 * watermarks, but we already failed the high-order watermark check
2423 * index towards 0 implies failure is due to lack of memory
2424 * index towards 1000 implies failure is due to fragmentation
2425 *
2426 * Only compact if a failure would be due to fragmentation. Also
2427 * ignore fragindex for non-costly orders where the alternative to
2428 * a successful reclaim/compaction is OOM. Fragindex and the
2429 * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
2430 * excessive compaction for costly orders, but it should not be at the
2431 * expense of system stability.
2432 */
2433 if (suitable) {
2434 compact_result = COMPACT_CONTINUE;
2435 if (order > PAGE_ALLOC_COSTLY_ORDER) {
2436 int fragindex = fragmentation_index(zone, order);
2437
2438 if (fragindex >= 0 &&
2439 fragindex <= sysctl_extfrag_threshold) {
2440 suitable = false;
2441 compact_result = COMPACT_NOT_SUITABLE_ZONE;
2442 }
2443 }
2444 } else {
2445 compact_result = COMPACT_SKIPPED;
2446 }
2447
2448 trace_mm_compaction_suitable(zone, order, compact_result);
2449
2450 return suitable;
2451}
2452
2453bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
2454 int alloc_flags)
2455{
2456 struct zone *zone;
2457 struct zoneref *z;
2458
2459 /*
2460 * Make sure at least one zone would pass __compaction_suitable if we continue
2461 * retrying the reclaim.
2462 */
2463 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2464 ac->highest_zoneidx, ac->nodemask) {
2465 unsigned long available;
2466
2467 /*
2468 * Do not consider all the reclaimable memory because we do not
2469 * want to trash just for a single high order allocation which
2470 * is even not guaranteed to appear even if __compaction_suitable
2471 * is happy about the watermark check.
2472 */
2473 available = zone_reclaimable_pages(zone) / order;
2474 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
2475 if (__compaction_suitable(zone, order, ac->highest_zoneidx,
2476 available))
2477 return true;
2478 }
2479
2480 return false;
2481}
2482
2483/*
2484 * Should we do compaction for target allocation order.
2485 * Return COMPACT_SUCCESS if allocation for target order can be already
2486 * satisfied
2487 * Return COMPACT_SKIPPED if compaction for target order is likely to fail
2488 * Return COMPACT_CONTINUE if compaction for target order should be ran
2489 */
2490static enum compact_result
2491compaction_suit_allocation_order(struct zone *zone, unsigned int order,
2492 int highest_zoneidx, unsigned int alloc_flags)
2493{
2494 unsigned long watermark;
2495
2496 watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
2497 if (zone_watermark_ok(zone, order, watermark, highest_zoneidx,
2498 alloc_flags))
2499 return COMPACT_SUCCESS;
2500
2501 if (!compaction_suitable(zone, order, highest_zoneidx))
2502 return COMPACT_SKIPPED;
2503
2504 return COMPACT_CONTINUE;
2505}
2506
2507static enum compact_result
2508compact_zone(struct compact_control *cc, struct capture_control *capc)
2509{
2510 enum compact_result ret;
2511 unsigned long start_pfn = cc->zone->zone_start_pfn;
2512 unsigned long end_pfn = zone_end_pfn(cc->zone);
2513 unsigned long last_migrated_pfn;
2514 const bool sync = cc->mode != MIGRATE_ASYNC;
2515 bool update_cached;
2516 unsigned int nr_succeeded = 0, nr_migratepages;
2517 int order;
2518
2519 /*
2520 * These counters track activities during zone compaction. Initialize
2521 * them before compacting a new zone.
2522 */
2523 cc->total_migrate_scanned = 0;
2524 cc->total_free_scanned = 0;
2525 cc->nr_migratepages = 0;
2526 cc->nr_freepages = 0;
2527 for (order = 0; order < NR_PAGE_ORDERS; order++)
2528 INIT_LIST_HEAD(&cc->freepages[order]);
2529 INIT_LIST_HEAD(&cc->migratepages);
2530
2531 cc->migratetype = gfp_migratetype(cc->gfp_mask);
2532
2533 if (!is_via_compact_memory(cc->order)) {
2534 ret = compaction_suit_allocation_order(cc->zone, cc->order,
2535 cc->highest_zoneidx,
2536 cc->alloc_flags);
2537 if (ret != COMPACT_CONTINUE)
2538 return ret;
2539 }
2540
2541 /*
2542 * Clear pageblock skip if there were failures recently and compaction
2543 * is about to be retried after being deferred.
2544 */
2545 if (compaction_restarting(cc->zone, cc->order))
2546 __reset_isolation_suitable(cc->zone);
2547
2548 /*
2549 * Setup to move all movable pages to the end of the zone. Used cached
2550 * information on where the scanners should start (unless we explicitly
2551 * want to compact the whole zone), but check that it is initialised
2552 * by ensuring the values are within zone boundaries.
2553 */
2554 cc->fast_start_pfn = 0;
2555 if (cc->whole_zone) {
2556 cc->migrate_pfn = start_pfn;
2557 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2558 } else {
2559 cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync];
2560 cc->free_pfn = cc->zone->compact_cached_free_pfn;
2561 if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
2562 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2563 cc->zone->compact_cached_free_pfn = cc->free_pfn;
2564 }
2565 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
2566 cc->migrate_pfn = start_pfn;
2567 cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
2568 cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
2569 }
2570
2571 if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn)
2572 cc->whole_zone = true;
2573 }
2574
2575 last_migrated_pfn = 0;
2576
2577 /*
2578 * Migrate has separate cached PFNs for ASYNC and SYNC* migration on
2579 * the basis that some migrations will fail in ASYNC mode. However,
2580 * if the cached PFNs match and pageblocks are skipped due to having
2581 * no isolation candidates, then the sync state does not matter.
2582 * Until a pageblock with isolation candidates is found, keep the
2583 * cached PFNs in sync to avoid revisiting the same blocks.
2584 */
2585 update_cached = !sync &&
2586 cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1];
2587
2588 trace_mm_compaction_begin(cc, start_pfn, end_pfn, sync);
2589
2590 /* lru_add_drain_all could be expensive with involving other CPUs */
2591 lru_add_drain();
2592
2593 while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) {
2594 int err;
2595 unsigned long iteration_start_pfn = cc->migrate_pfn;
2596
2597 /*
2598 * Avoid multiple rescans of the same pageblock which can
2599 * happen if a page cannot be isolated (dirty/writeback in
2600 * async mode) or if the migrated pages are being allocated
2601 * before the pageblock is cleared. The first rescan will
2602 * capture the entire pageblock for migration. If it fails,
2603 * it'll be marked skip and scanning will proceed as normal.
2604 */
2605 cc->finish_pageblock = false;
2606 if (pageblock_start_pfn(last_migrated_pfn) ==
2607 pageblock_start_pfn(iteration_start_pfn)) {
2608 cc->finish_pageblock = true;
2609 }
2610
2611rescan:
2612 switch (isolate_migratepages(cc)) {
2613 case ISOLATE_ABORT:
2614 ret = COMPACT_CONTENDED;
2615 putback_movable_pages(&cc->migratepages);
2616 cc->nr_migratepages = 0;
2617 goto out;
2618 case ISOLATE_NONE:
2619 if (update_cached) {
2620 cc->zone->compact_cached_migrate_pfn[1] =
2621 cc->zone->compact_cached_migrate_pfn[0];
2622 }
2623
2624 /*
2625 * We haven't isolated and migrated anything, but
2626 * there might still be unflushed migrations from
2627 * previous cc->order aligned block.
2628 */
2629 goto check_drain;
2630 case ISOLATE_SUCCESS:
2631 update_cached = false;
2632 last_migrated_pfn = max(cc->zone->zone_start_pfn,
2633 pageblock_start_pfn(cc->migrate_pfn - 1));
2634 }
2635
2636 /*
2637 * Record the number of pages to migrate since the
2638 * compaction_alloc/free() will update cc->nr_migratepages
2639 * properly.
2640 */
2641 nr_migratepages = cc->nr_migratepages;
2642 err = migrate_pages(&cc->migratepages, compaction_alloc,
2643 compaction_free, (unsigned long)cc, cc->mode,
2644 MR_COMPACTION, &nr_succeeded);
2645
2646 trace_mm_compaction_migratepages(nr_migratepages, nr_succeeded);
2647
2648 /* All pages were either migrated or will be released */
2649 cc->nr_migratepages = 0;
2650 if (err) {
2651 putback_movable_pages(&cc->migratepages);
2652 /*
2653 * migrate_pages() may return -ENOMEM when scanners meet
2654 * and we want compact_finished() to detect it
2655 */
2656 if (err == -ENOMEM && !compact_scanners_met(cc)) {
2657 ret = COMPACT_CONTENDED;
2658 goto out;
2659 }
2660 /*
2661 * If an ASYNC or SYNC_LIGHT fails to migrate a page
2662 * within the pageblock_order-aligned block and
2663 * fast_find_migrateblock may be used then scan the
2664 * remainder of the pageblock. This will mark the
2665 * pageblock "skip" to avoid rescanning in the near
2666 * future. This will isolate more pages than necessary
2667 * for the request but avoid loops due to
2668 * fast_find_migrateblock revisiting blocks that were
2669 * recently partially scanned.
2670 */
2671 if (!pageblock_aligned(cc->migrate_pfn) &&
2672 !cc->ignore_skip_hint && !cc->finish_pageblock &&
2673 (cc->mode < MIGRATE_SYNC)) {
2674 cc->finish_pageblock = true;
2675
2676 /*
2677 * Draining pcplists does not help THP if
2678 * any page failed to migrate. Even after
2679 * drain, the pageblock will not be free.
2680 */
2681 if (cc->order == COMPACTION_HPAGE_ORDER)
2682 last_migrated_pfn = 0;
2683
2684 goto rescan;
2685 }
2686 }
2687
2688 /* Stop if a page has been captured */
2689 if (capc && capc->page) {
2690 ret = COMPACT_SUCCESS;
2691 break;
2692 }
2693
2694check_drain:
2695 /*
2696 * Has the migration scanner moved away from the previous
2697 * cc->order aligned block where we migrated from? If yes,
2698 * flush the pages that were freed, so that they can merge and
2699 * compact_finished() can detect immediately if allocation
2700 * would succeed.
2701 */
2702 if (cc->order > 0 && last_migrated_pfn) {
2703 unsigned long current_block_start =
2704 block_start_pfn(cc->migrate_pfn, cc->order);
2705
2706 if (last_migrated_pfn < current_block_start) {
2707 lru_add_drain_cpu_zone(cc->zone);
2708 /* No more flushing until we migrate again */
2709 last_migrated_pfn = 0;
2710 }
2711 }
2712 }
2713
2714out:
2715 /*
2716 * Release free pages and update where the free scanner should restart,
2717 * so we don't leave any returned pages behind in the next attempt.
2718 */
2719 if (cc->nr_freepages > 0) {
2720 unsigned long free_pfn = release_free_list(cc->freepages);
2721
2722 cc->nr_freepages = 0;
2723 VM_BUG_ON(free_pfn == 0);
2724 /* The cached pfn is always the first in a pageblock */
2725 free_pfn = pageblock_start_pfn(free_pfn);
2726 /*
2727 * Only go back, not forward. The cached pfn might have been
2728 * already reset to zone end in compact_finished()
2729 */
2730 if (free_pfn > cc->zone->compact_cached_free_pfn)
2731 cc->zone->compact_cached_free_pfn = free_pfn;
2732 }
2733
2734 count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned);
2735 count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned);
2736
2737 trace_mm_compaction_end(cc, start_pfn, end_pfn, sync, ret);
2738
2739 VM_BUG_ON(!list_empty(&cc->migratepages));
2740
2741 return ret;
2742}
2743
2744static enum compact_result compact_zone_order(struct zone *zone, int order,
2745 gfp_t gfp_mask, enum compact_priority prio,
2746 unsigned int alloc_flags, int highest_zoneidx,
2747 struct page **capture)
2748{
2749 enum compact_result ret;
2750 struct compact_control cc = {
2751 .order = order,
2752 .search_order = order,
2753 .gfp_mask = gfp_mask,
2754 .zone = zone,
2755 .mode = (prio == COMPACT_PRIO_ASYNC) ?
2756 MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT,
2757 .alloc_flags = alloc_flags,
2758 .highest_zoneidx = highest_zoneidx,
2759 .direct_compaction = true,
2760 .whole_zone = (prio == MIN_COMPACT_PRIORITY),
2761 .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
2762 .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
2763 };
2764 struct capture_control capc = {
2765 .cc = &cc,
2766 .page = NULL,
2767 };
2768
2769 /*
2770 * Make sure the structs are really initialized before we expose the
2771 * capture control, in case we are interrupted and the interrupt handler
2772 * frees a page.
2773 */
2774 barrier();
2775 WRITE_ONCE(current->capture_control, &capc);
2776
2777 ret = compact_zone(&cc, &capc);
2778
2779 /*
2780 * Make sure we hide capture control first before we read the captured
2781 * page pointer, otherwise an interrupt could free and capture a page
2782 * and we would leak it.
2783 */
2784 WRITE_ONCE(current->capture_control, NULL);
2785 *capture = READ_ONCE(capc.page);
2786 /*
2787 * Technically, it is also possible that compaction is skipped but
2788 * the page is still captured out of luck(IRQ came and freed the page).
2789 * Returning COMPACT_SUCCESS in such cases helps in properly accounting
2790 * the COMPACT[STALL|FAIL] when compaction is skipped.
2791 */
2792 if (*capture)
2793 ret = COMPACT_SUCCESS;
2794
2795 return ret;
2796}
2797
2798/**
2799 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
2800 * @gfp_mask: The GFP mask of the current allocation
2801 * @order: The order of the current allocation
2802 * @alloc_flags: The allocation flags of the current allocation
2803 * @ac: The context of current allocation
2804 * @prio: Determines how hard direct compaction should try to succeed
2805 * @capture: Pointer to free page created by compaction will be stored here
2806 *
2807 * This is the main entry point for direct page compaction.
2808 */
2809enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
2810 unsigned int alloc_flags, const struct alloc_context *ac,
2811 enum compact_priority prio, struct page **capture)
2812{
2813 struct zoneref *z;
2814 struct zone *zone;
2815 enum compact_result rc = COMPACT_SKIPPED;
2816
2817 if (!gfp_compaction_allowed(gfp_mask))
2818 return COMPACT_SKIPPED;
2819
2820 trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
2821
2822 /* Compact each zone in the list */
2823 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2824 ac->highest_zoneidx, ac->nodemask) {
2825 enum compact_result status;
2826
2827 if (cpusets_enabled() &&
2828 (alloc_flags & ALLOC_CPUSET) &&
2829 !__cpuset_zone_allowed(zone, gfp_mask))
2830 continue;
2831
2832 if (prio > MIN_COMPACT_PRIORITY
2833 && compaction_deferred(zone, order)) {
2834 rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
2835 continue;
2836 }
2837
2838 status = compact_zone_order(zone, order, gfp_mask, prio,
2839 alloc_flags, ac->highest_zoneidx, capture);
2840 rc = max(status, rc);
2841
2842 /* The allocation should succeed, stop compacting */
2843 if (status == COMPACT_SUCCESS) {
2844 /*
2845 * We think the allocation will succeed in this zone,
2846 * but it is not certain, hence the false. The caller
2847 * will repeat this with true if allocation indeed
2848 * succeeds in this zone.
2849 */
2850 compaction_defer_reset(zone, order, false);
2851
2852 break;
2853 }
2854
2855 if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
2856 status == COMPACT_PARTIAL_SKIPPED))
2857 /*
2858 * We think that allocation won't succeed in this zone
2859 * so we defer compaction there. If it ends up
2860 * succeeding after all, it will be reset.
2861 */
2862 defer_compaction(zone, order);
2863
2864 /*
2865 * We might have stopped compacting due to need_resched() in
2866 * async compaction, or due to a fatal signal detected. In that
2867 * case do not try further zones
2868 */
2869 if ((prio == COMPACT_PRIO_ASYNC && need_resched())
2870 || fatal_signal_pending(current))
2871 break;
2872 }
2873
2874 return rc;
2875}
2876
2877/*
2878 * compact_node() - compact all zones within a node
2879 * @pgdat: The node page data
2880 * @proactive: Whether the compaction is proactive
2881 *
2882 * For proactive compaction, compact till each zone's fragmentation score
2883 * reaches within proactive compaction thresholds (as determined by the
2884 * proactiveness tunable), it is possible that the function returns before
2885 * reaching score targets due to various back-off conditions, such as,
2886 * contention on per-node or per-zone locks.
2887 */
2888static int compact_node(pg_data_t *pgdat, bool proactive)
2889{
2890 int zoneid;
2891 struct zone *zone;
2892 struct compact_control cc = {
2893 .order = -1,
2894 .mode = proactive ? MIGRATE_SYNC_LIGHT : MIGRATE_SYNC,
2895 .ignore_skip_hint = true,
2896 .whole_zone = true,
2897 .gfp_mask = GFP_KERNEL,
2898 .proactive_compaction = proactive,
2899 };
2900
2901 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2902 zone = &pgdat->node_zones[zoneid];
2903 if (!populated_zone(zone))
2904 continue;
2905
2906 if (fatal_signal_pending(current))
2907 return -EINTR;
2908
2909 cc.zone = zone;
2910
2911 compact_zone(&cc, NULL);
2912
2913 if (proactive) {
2914 count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
2915 cc.total_migrate_scanned);
2916 count_compact_events(KCOMPACTD_FREE_SCANNED,
2917 cc.total_free_scanned);
2918 }
2919 }
2920
2921 return 0;
2922}
2923
2924/* Compact all zones of all nodes in the system */
2925static int compact_nodes(void)
2926{
2927 int ret, nid;
2928
2929 /* Flush pending updates to the LRU lists */
2930 lru_add_drain_all();
2931
2932 for_each_online_node(nid) {
2933 ret = compact_node(NODE_DATA(nid), false);
2934 if (ret)
2935 return ret;
2936 }
2937
2938 return 0;
2939}
2940
2941static int compaction_proactiveness_sysctl_handler(const struct ctl_table *table, int write,
2942 void *buffer, size_t *length, loff_t *ppos)
2943{
2944 int rc, nid;
2945
2946 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
2947 if (rc)
2948 return rc;
2949
2950 if (write && sysctl_compaction_proactiveness) {
2951 for_each_online_node(nid) {
2952 pg_data_t *pgdat = NODE_DATA(nid);
2953
2954 if (pgdat->proactive_compact_trigger)
2955 continue;
2956
2957 pgdat->proactive_compact_trigger = true;
2958 trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, -1,
2959 pgdat->nr_zones - 1);
2960 wake_up_interruptible(&pgdat->kcompactd_wait);
2961 }
2962 }
2963
2964 return 0;
2965}
2966
2967/*
2968 * This is the entry point for compacting all nodes via
2969 * /proc/sys/vm/compact_memory
2970 */
2971static int sysctl_compaction_handler(const struct ctl_table *table, int write,
2972 void *buffer, size_t *length, loff_t *ppos)
2973{
2974 int ret;
2975
2976 ret = proc_dointvec(table, write, buffer, length, ppos);
2977 if (ret)
2978 return ret;
2979
2980 if (sysctl_compact_memory != 1)
2981 return -EINVAL;
2982
2983 if (write)
2984 ret = compact_nodes();
2985
2986 return ret;
2987}
2988
2989#if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
2990static ssize_t compact_store(struct device *dev,
2991 struct device_attribute *attr,
2992 const char *buf, size_t count)
2993{
2994 int nid = dev->id;
2995
2996 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
2997 /* Flush pending updates to the LRU lists */
2998 lru_add_drain_all();
2999
3000 compact_node(NODE_DATA(nid), false);
3001 }
3002
3003 return count;
3004}
3005static DEVICE_ATTR_WO(compact);
3006
3007int compaction_register_node(struct node *node)
3008{
3009 return device_create_file(&node->dev, &dev_attr_compact);
3010}
3011
3012void compaction_unregister_node(struct node *node)
3013{
3014 device_remove_file(&node->dev, &dev_attr_compact);
3015}
3016#endif /* CONFIG_SYSFS && CONFIG_NUMA */
3017
3018static inline bool kcompactd_work_requested(pg_data_t *pgdat)
3019{
3020 return pgdat->kcompactd_max_order > 0 || kthread_should_stop() ||
3021 pgdat->proactive_compact_trigger;
3022}
3023
3024static bool kcompactd_node_suitable(pg_data_t *pgdat)
3025{
3026 int zoneid;
3027 struct zone *zone;
3028 enum zone_type highest_zoneidx = pgdat->kcompactd_highest_zoneidx;
3029 enum compact_result ret;
3030
3031 for (zoneid = 0; zoneid <= highest_zoneidx; zoneid++) {
3032 zone = &pgdat->node_zones[zoneid];
3033
3034 if (!populated_zone(zone))
3035 continue;
3036
3037 ret = compaction_suit_allocation_order(zone,
3038 pgdat->kcompactd_max_order,
3039 highest_zoneidx, ALLOC_WMARK_MIN);
3040 if (ret == COMPACT_CONTINUE)
3041 return true;
3042 }
3043
3044 return false;
3045}
3046
3047static void kcompactd_do_work(pg_data_t *pgdat)
3048{
3049 /*
3050 * With no special task, compact all zones so that a page of requested
3051 * order is allocatable.
3052 */
3053 int zoneid;
3054 struct zone *zone;
3055 struct compact_control cc = {
3056 .order = pgdat->kcompactd_max_order,
3057 .search_order = pgdat->kcompactd_max_order,
3058 .highest_zoneidx = pgdat->kcompactd_highest_zoneidx,
3059 .mode = MIGRATE_SYNC_LIGHT,
3060 .ignore_skip_hint = false,
3061 .gfp_mask = GFP_KERNEL,
3062 };
3063 enum compact_result ret;
3064
3065 trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
3066 cc.highest_zoneidx);
3067 count_compact_event(KCOMPACTD_WAKE);
3068
3069 for (zoneid = 0; zoneid <= cc.highest_zoneidx; zoneid++) {
3070 int status;
3071
3072 zone = &pgdat->node_zones[zoneid];
3073 if (!populated_zone(zone))
3074 continue;
3075
3076 if (compaction_deferred(zone, cc.order))
3077 continue;
3078
3079 ret = compaction_suit_allocation_order(zone,
3080 cc.order, zoneid, ALLOC_WMARK_MIN);
3081 if (ret != COMPACT_CONTINUE)
3082 continue;
3083
3084 if (kthread_should_stop())
3085 return;
3086
3087 cc.zone = zone;
3088 status = compact_zone(&cc, NULL);
3089
3090 if (status == COMPACT_SUCCESS) {
3091 compaction_defer_reset(zone, cc.order, false);
3092 } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
3093 /*
3094 * Buddy pages may become stranded on pcps that could
3095 * otherwise coalesce on the zone's free area for
3096 * order >= cc.order. This is ratelimited by the
3097 * upcoming deferral.
3098 */
3099 drain_all_pages(zone);
3100
3101 /*
3102 * We use sync migration mode here, so we defer like
3103 * sync direct compaction does.
3104 */
3105 defer_compaction(zone, cc.order);
3106 }
3107
3108 count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
3109 cc.total_migrate_scanned);
3110 count_compact_events(KCOMPACTD_FREE_SCANNED,
3111 cc.total_free_scanned);
3112 }
3113
3114 /*
3115 * Regardless of success, we are done until woken up next. But remember
3116 * the requested order/highest_zoneidx in case it was higher/tighter
3117 * than our current ones
3118 */
3119 if (pgdat->kcompactd_max_order <= cc.order)
3120 pgdat->kcompactd_max_order = 0;
3121 if (pgdat->kcompactd_highest_zoneidx >= cc.highest_zoneidx)
3122 pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
3123}
3124
3125void wakeup_kcompactd(pg_data_t *pgdat, int order, int highest_zoneidx)
3126{
3127 if (!order)
3128 return;
3129
3130 if (pgdat->kcompactd_max_order < order)
3131 pgdat->kcompactd_max_order = order;
3132
3133 if (pgdat->kcompactd_highest_zoneidx > highest_zoneidx)
3134 pgdat->kcompactd_highest_zoneidx = highest_zoneidx;
3135
3136 /*
3137 * Pairs with implicit barrier in wait_event_freezable()
3138 * such that wakeups are not missed.
3139 */
3140 if (!wq_has_sleeper(&pgdat->kcompactd_wait))
3141 return;
3142
3143 if (!kcompactd_node_suitable(pgdat))
3144 return;
3145
3146 trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
3147 highest_zoneidx);
3148 wake_up_interruptible(&pgdat->kcompactd_wait);
3149}
3150
3151/*
3152 * The background compaction daemon, started as a kernel thread
3153 * from the init process.
3154 */
3155static int kcompactd(void *p)
3156{
3157 pg_data_t *pgdat = (pg_data_t *)p;
3158 struct task_struct *tsk = current;
3159 long default_timeout = msecs_to_jiffies(HPAGE_FRAG_CHECK_INTERVAL_MSEC);
3160 long timeout = default_timeout;
3161
3162 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3163
3164 if (!cpumask_empty(cpumask))
3165 set_cpus_allowed_ptr(tsk, cpumask);
3166
3167 current->flags |= PF_KCOMPACTD;
3168 set_freezable();
3169
3170 pgdat->kcompactd_max_order = 0;
3171 pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
3172
3173 while (!kthread_should_stop()) {
3174 unsigned long pflags;
3175
3176 /*
3177 * Avoid the unnecessary wakeup for proactive compaction
3178 * when it is disabled.
3179 */
3180 if (!sysctl_compaction_proactiveness)
3181 timeout = MAX_SCHEDULE_TIMEOUT;
3182 trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
3183 if (wait_event_freezable_timeout(pgdat->kcompactd_wait,
3184 kcompactd_work_requested(pgdat), timeout) &&
3185 !pgdat->proactive_compact_trigger) {
3186
3187 psi_memstall_enter(&pflags);
3188 kcompactd_do_work(pgdat);
3189 psi_memstall_leave(&pflags);
3190 /*
3191 * Reset the timeout value. The defer timeout from
3192 * proactive compaction is lost here but that is fine
3193 * as the condition of the zone changing substantionally
3194 * then carrying on with the previous defer interval is
3195 * not useful.
3196 */
3197 timeout = default_timeout;
3198 continue;
3199 }
3200
3201 /*
3202 * Start the proactive work with default timeout. Based
3203 * on the fragmentation score, this timeout is updated.
3204 */
3205 timeout = default_timeout;
3206 if (should_proactive_compact_node(pgdat)) {
3207 unsigned int prev_score, score;
3208
3209 prev_score = fragmentation_score_node(pgdat);
3210 compact_node(pgdat, true);
3211 score = fragmentation_score_node(pgdat);
3212 /*
3213 * Defer proactive compaction if the fragmentation
3214 * score did not go down i.e. no progress made.
3215 */
3216 if (unlikely(score >= prev_score))
3217 timeout =
3218 default_timeout << COMPACT_MAX_DEFER_SHIFT;
3219 }
3220 if (unlikely(pgdat->proactive_compact_trigger))
3221 pgdat->proactive_compact_trigger = false;
3222 }
3223
3224 current->flags &= ~PF_KCOMPACTD;
3225
3226 return 0;
3227}
3228
3229/*
3230 * This kcompactd start function will be called by init and node-hot-add.
3231 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
3232 */
3233void __meminit kcompactd_run(int nid)
3234{
3235 pg_data_t *pgdat = NODE_DATA(nid);
3236
3237 if (pgdat->kcompactd)
3238 return;
3239
3240 pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
3241 if (IS_ERR(pgdat->kcompactd)) {
3242 pr_err("Failed to start kcompactd on node %d\n", nid);
3243 pgdat->kcompactd = NULL;
3244 }
3245}
3246
3247/*
3248 * Called by memory hotplug when all memory in a node is offlined. Caller must
3249 * be holding mem_hotplug_begin/done().
3250 */
3251void __meminit kcompactd_stop(int nid)
3252{
3253 struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
3254
3255 if (kcompactd) {
3256 kthread_stop(kcompactd);
3257 NODE_DATA(nid)->kcompactd = NULL;
3258 }
3259}
3260
3261/*
3262 * It's optimal to keep kcompactd on the same CPUs as their memory, but
3263 * not required for correctness. So if the last cpu in a node goes
3264 * away, we get changed to run anywhere: as the first one comes back,
3265 * restore their cpu bindings.
3266 */
3267static int kcompactd_cpu_online(unsigned int cpu)
3268{
3269 int nid;
3270
3271 for_each_node_state(nid, N_MEMORY) {
3272 pg_data_t *pgdat = NODE_DATA(nid);
3273 const struct cpumask *mask;
3274
3275 mask = cpumask_of_node(pgdat->node_id);
3276
3277 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3278 /* One of our CPUs online: restore mask */
3279 if (pgdat->kcompactd)
3280 set_cpus_allowed_ptr(pgdat->kcompactd, mask);
3281 }
3282 return 0;
3283}
3284
3285static int proc_dointvec_minmax_warn_RT_change(const struct ctl_table *table,
3286 int write, void *buffer, size_t *lenp, loff_t *ppos)
3287{
3288 int ret, old;
3289
3290 if (!IS_ENABLED(CONFIG_PREEMPT_RT) || !write)
3291 return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
3292
3293 old = *(int *)table->data;
3294 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
3295 if (ret)
3296 return ret;
3297 if (old != *(int *)table->data)
3298 pr_warn_once("sysctl attribute %s changed by %s[%d]\n",
3299 table->procname, current->comm,
3300 task_pid_nr(current));
3301 return ret;
3302}
3303
3304static struct ctl_table vm_compaction[] = {
3305 {
3306 .procname = "compact_memory",
3307 .data = &sysctl_compact_memory,
3308 .maxlen = sizeof(int),
3309 .mode = 0200,
3310 .proc_handler = sysctl_compaction_handler,
3311 },
3312 {
3313 .procname = "compaction_proactiveness",
3314 .data = &sysctl_compaction_proactiveness,
3315 .maxlen = sizeof(sysctl_compaction_proactiveness),
3316 .mode = 0644,
3317 .proc_handler = compaction_proactiveness_sysctl_handler,
3318 .extra1 = SYSCTL_ZERO,
3319 .extra2 = SYSCTL_ONE_HUNDRED,
3320 },
3321 {
3322 .procname = "extfrag_threshold",
3323 .data = &sysctl_extfrag_threshold,
3324 .maxlen = sizeof(int),
3325 .mode = 0644,
3326 .proc_handler = proc_dointvec_minmax,
3327 .extra1 = SYSCTL_ZERO,
3328 .extra2 = SYSCTL_ONE_THOUSAND,
3329 },
3330 {
3331 .procname = "compact_unevictable_allowed",
3332 .data = &sysctl_compact_unevictable_allowed,
3333 .maxlen = sizeof(int),
3334 .mode = 0644,
3335 .proc_handler = proc_dointvec_minmax_warn_RT_change,
3336 .extra1 = SYSCTL_ZERO,
3337 .extra2 = SYSCTL_ONE,
3338 },
3339};
3340
3341static int __init kcompactd_init(void)
3342{
3343 int nid;
3344 int ret;
3345
3346 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3347 "mm/compaction:online",
3348 kcompactd_cpu_online, NULL);
3349 if (ret < 0) {
3350 pr_err("kcompactd: failed to register hotplug callbacks.\n");
3351 return ret;
3352 }
3353
3354 for_each_node_state(nid, N_MEMORY)
3355 kcompactd_run(nid);
3356 register_sysctl_init("vm", vm_compaction);
3357 return 0;
3358}
3359subsys_initcall(kcompactd_init)
3360
3361#endif /* CONFIG_COMPACTION */