1 // SPDX-License-Identifier: GPL-2.0
3 * linux/mm/compaction.c
5 * Memory compaction for the reduction of external fragmentation. Note that
6 * this heavily depends upon page migration to do all the real heavy
9 * Copyright IBM Corp. 2007-2010 Mel Gorman <mel@csn.ul.ie>
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>
28 #ifdef CONFIG_COMPACTION
30 * Fragmentation score check interval for proactive compaction purposes.
32 #define HPAGE_FRAG_CHECK_INTERVAL_MSEC (500)
34 static inline void count_compact_event(enum vm_event_item item)
39 static inline void count_compact_events(enum vm_event_item item, long delta)
41 count_vm_events(item, delta);
45 * order == -1 is expected when compacting proactively via
46 * 1. /proc/sys/vm/compact_memory
47 * 2. /sys/devices/system/node/nodex/compact
48 * 3. /proc/sys/vm/compaction_proactiveness
50 static inline bool is_via_compact_memory(int order)
56 #define count_compact_event(item) do { } while (0)
57 #define count_compact_events(item, delta) do { } while (0)
58 static inline bool is_via_compact_memory(int order) { return false; }
61 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
63 #define CREATE_TRACE_POINTS
64 #include <trace/events/compaction.h>
66 #define block_start_pfn(pfn, order) round_down(pfn, 1UL << (order))
67 #define block_end_pfn(pfn, order) ALIGN((pfn) + 1, 1UL << (order))
70 * Page order with-respect-to which proactive compaction
71 * calculates external fragmentation, which is used as
72 * the "fragmentation score" of a node/zone.
74 #if defined CONFIG_TRANSPARENT_HUGEPAGE
75 #define COMPACTION_HPAGE_ORDER HPAGE_PMD_ORDER
76 #elif defined CONFIG_HUGETLBFS
77 #define COMPACTION_HPAGE_ORDER HUGETLB_PAGE_ORDER
79 #define COMPACTION_HPAGE_ORDER (PMD_SHIFT - PAGE_SHIFT)
82 static void split_map_pages(struct list_head *freepages)
84 unsigned int i, order;
85 struct page *page, *next;
88 for (order = 0; order < NR_PAGE_ORDERS; order++) {
89 list_for_each_entry_safe(page, next, &freepages[order], lru) {
90 unsigned int nr_pages;
94 nr_pages = 1 << order;
96 post_alloc_hook(page, order, __GFP_MOVABLE);
98 split_page(page, order);
100 for (i = 0; i < nr_pages; i++) {
101 list_add(&page->lru, &tmp_list);
105 list_splice_init(&tmp_list, &freepages[0]);
109 static unsigned long release_free_list(struct list_head *freepages)
112 unsigned long high_pfn = 0;
114 for (order = 0; order < NR_PAGE_ORDERS; order++) {
115 struct page *page, *next;
117 list_for_each_entry_safe(page, next, &freepages[order], lru) {
118 unsigned long pfn = page_to_pfn(page);
120 list_del(&page->lru);
122 * Convert free pages into post allocation pages, so
123 * that we can free them via __free_page.
125 post_alloc_hook(page, order, __GFP_MOVABLE);
126 __free_pages(page, order);
134 #ifdef CONFIG_COMPACTION
135 bool PageMovable(struct page *page)
137 const struct movable_operations *mops;
139 VM_BUG_ON_PAGE(!PageLocked(page), page);
140 if (!__PageMovable(page))
143 mops = page_movable_ops(page);
150 void __SetPageMovable(struct page *page, const struct movable_operations *mops)
152 VM_BUG_ON_PAGE(!PageLocked(page), page);
153 VM_BUG_ON_PAGE((unsigned long)mops & PAGE_MAPPING_MOVABLE, page);
154 page->mapping = (void *)((unsigned long)mops | PAGE_MAPPING_MOVABLE);
156 EXPORT_SYMBOL(__SetPageMovable);
158 void __ClearPageMovable(struct page *page)
160 VM_BUG_ON_PAGE(!PageMovable(page), page);
162 * This page still has the type of a movable page, but it's
163 * actually not movable any more.
165 page->mapping = (void *)PAGE_MAPPING_MOVABLE;
167 EXPORT_SYMBOL(__ClearPageMovable);
169 /* Do not skip compaction more than 64 times */
170 #define COMPACT_MAX_DEFER_SHIFT 6
173 * Compaction is deferred when compaction fails to result in a page
174 * allocation success. 1 << compact_defer_shift, compactions are skipped up
175 * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
177 static void defer_compaction(struct zone *zone, int order)
179 zone->compact_considered = 0;
180 zone->compact_defer_shift++;
182 if (order < zone->compact_order_failed)
183 zone->compact_order_failed = order;
185 if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
186 zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
188 trace_mm_compaction_defer_compaction(zone, order);
191 /* Returns true if compaction should be skipped this time */
192 static bool compaction_deferred(struct zone *zone, int order)
194 unsigned long defer_limit = 1UL << zone->compact_defer_shift;
196 if (order < zone->compact_order_failed)
199 /* Avoid possible overflow */
200 if (++zone->compact_considered >= defer_limit) {
201 zone->compact_considered = defer_limit;
205 trace_mm_compaction_deferred(zone, order);
211 * Update defer tracking counters after successful compaction of given order,
212 * which means an allocation either succeeded (alloc_success == true) or is
213 * expected to succeed.
215 void compaction_defer_reset(struct zone *zone, int order,
219 zone->compact_considered = 0;
220 zone->compact_defer_shift = 0;
222 if (order >= zone->compact_order_failed)
223 zone->compact_order_failed = order + 1;
225 trace_mm_compaction_defer_reset(zone, order);
228 /* Returns true if restarting compaction after many failures */
229 static bool compaction_restarting(struct zone *zone, int order)
231 if (order < zone->compact_order_failed)
234 return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
235 zone->compact_considered >= 1UL << zone->compact_defer_shift;
238 /* Returns true if the pageblock should be scanned for pages to isolate. */
239 static inline bool isolation_suitable(struct compact_control *cc,
242 if (cc->ignore_skip_hint)
245 return !get_pageblock_skip(page);
248 static void reset_cached_positions(struct zone *zone)
250 zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
251 zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
252 zone->compact_cached_free_pfn =
253 pageblock_start_pfn(zone_end_pfn(zone) - 1);
256 #ifdef CONFIG_SPARSEMEM
258 * If the PFN falls into an offline section, return the start PFN of the
259 * next online section. If the PFN falls into an online section or if
260 * there is no next online section, return 0.
262 static unsigned long skip_offline_sections(unsigned long start_pfn)
264 unsigned long start_nr = pfn_to_section_nr(start_pfn);
266 if (online_section_nr(start_nr))
269 while (++start_nr <= __highest_present_section_nr) {
270 if (online_section_nr(start_nr))
271 return section_nr_to_pfn(start_nr);
278 * If the PFN falls into an offline section, return the end PFN of the
279 * next online section in reverse. If the PFN falls into an online section
280 * or if there is no next online section in reverse, return 0.
282 static unsigned long skip_offline_sections_reverse(unsigned long start_pfn)
284 unsigned long start_nr = pfn_to_section_nr(start_pfn);
286 if (!start_nr || online_section_nr(start_nr))
289 while (start_nr-- > 0) {
290 if (online_section_nr(start_nr))
291 return section_nr_to_pfn(start_nr) + PAGES_PER_SECTION;
297 static unsigned long skip_offline_sections(unsigned long start_pfn)
302 static unsigned long skip_offline_sections_reverse(unsigned long start_pfn)
309 * Compound pages of >= pageblock_order should consistently be skipped until
310 * released. It is always pointless to compact pages of such order (if they are
311 * migratable), and the pageblocks they occupy cannot contain any free pages.
313 static bool pageblock_skip_persistent(struct page *page)
315 if (!PageCompound(page))
318 page = compound_head(page);
320 if (compound_order(page) >= pageblock_order)
327 __reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source,
330 struct page *page = pfn_to_online_page(pfn);
331 struct page *block_page;
332 struct page *end_page;
333 unsigned long block_pfn;
337 if (zone != page_zone(page))
339 if (pageblock_skip_persistent(page))
343 * If skip is already cleared do no further checking once the
344 * restart points have been set.
346 if (check_source && check_target && !get_pageblock_skip(page))
350 * If clearing skip for the target scanner, do not select a
351 * non-movable pageblock as the starting point.
353 if (!check_source && check_target &&
354 get_pageblock_migratetype(page) != MIGRATE_MOVABLE)
357 /* Ensure the start of the pageblock or zone is online and valid */
358 block_pfn = pageblock_start_pfn(pfn);
359 block_pfn = max(block_pfn, zone->zone_start_pfn);
360 block_page = pfn_to_online_page(block_pfn);
366 /* Ensure the end of the pageblock or zone is online and valid */
367 block_pfn = pageblock_end_pfn(pfn) - 1;
368 block_pfn = min(block_pfn, zone_end_pfn(zone) - 1);
369 end_page = pfn_to_online_page(block_pfn);
374 * Only clear the hint if a sample indicates there is either a
375 * free page or an LRU page in the block. One or other condition
376 * is necessary for the block to be a migration source/target.
379 if (check_source && PageLRU(page)) {
380 clear_pageblock_skip(page);
384 if (check_target && PageBuddy(page)) {
385 clear_pageblock_skip(page);
389 page += (1 << PAGE_ALLOC_COSTLY_ORDER);
390 } while (page <= end_page);
396 * This function is called to clear all cached information on pageblocks that
397 * should be skipped for page isolation when the migrate and free page scanner
400 static void __reset_isolation_suitable(struct zone *zone)
402 unsigned long migrate_pfn = zone->zone_start_pfn;
403 unsigned long free_pfn = zone_end_pfn(zone) - 1;
404 unsigned long reset_migrate = free_pfn;
405 unsigned long reset_free = migrate_pfn;
406 bool source_set = false;
407 bool free_set = false;
409 /* Only flush if a full compaction finished recently */
410 if (!zone->compact_blockskip_flush)
413 zone->compact_blockskip_flush = false;
416 * Walk the zone and update pageblock skip information. Source looks
417 * for PageLRU while target looks for PageBuddy. When the scanner
418 * is found, both PageBuddy and PageLRU are checked as the pageblock
419 * is suitable as both source and target.
421 for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages,
422 free_pfn -= pageblock_nr_pages) {
425 /* Update the migrate PFN */
426 if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) &&
427 migrate_pfn < reset_migrate) {
429 reset_migrate = migrate_pfn;
430 zone->compact_init_migrate_pfn = reset_migrate;
431 zone->compact_cached_migrate_pfn[0] = reset_migrate;
432 zone->compact_cached_migrate_pfn[1] = reset_migrate;
435 /* Update the free PFN */
436 if (__reset_isolation_pfn(zone, free_pfn, free_set, true) &&
437 free_pfn > reset_free) {
439 reset_free = free_pfn;
440 zone->compact_init_free_pfn = reset_free;
441 zone->compact_cached_free_pfn = reset_free;
445 /* Leave no distance if no suitable block was reset */
446 if (reset_migrate >= reset_free) {
447 zone->compact_cached_migrate_pfn[0] = migrate_pfn;
448 zone->compact_cached_migrate_pfn[1] = migrate_pfn;
449 zone->compact_cached_free_pfn = free_pfn;
453 void reset_isolation_suitable(pg_data_t *pgdat)
457 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
458 struct zone *zone = &pgdat->node_zones[zoneid];
459 if (!populated_zone(zone))
462 __reset_isolation_suitable(zone);
467 * Sets the pageblock skip bit if it was clear. Note that this is a hint as
468 * locks are not required for read/writers. Returns true if it was already set.
470 static bool test_and_set_skip(struct compact_control *cc, struct page *page)
474 /* Do not update if skip hint is being ignored */
475 if (cc->ignore_skip_hint)
478 skip = get_pageblock_skip(page);
479 if (!skip && !cc->no_set_skip_hint)
480 set_pageblock_skip(page);
485 static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
487 struct zone *zone = cc->zone;
489 /* Set for isolation rather than compaction */
490 if (cc->no_set_skip_hint)
493 pfn = pageblock_end_pfn(pfn);
495 /* Update where async and sync compaction should restart */
496 if (pfn > zone->compact_cached_migrate_pfn[0])
497 zone->compact_cached_migrate_pfn[0] = pfn;
498 if (cc->mode != MIGRATE_ASYNC &&
499 pfn > zone->compact_cached_migrate_pfn[1])
500 zone->compact_cached_migrate_pfn[1] = pfn;
504 * If no pages were isolated then mark this pageblock to be skipped in the
505 * future. The information is later cleared by __reset_isolation_suitable().
507 static void update_pageblock_skip(struct compact_control *cc,
508 struct page *page, unsigned long pfn)
510 struct zone *zone = cc->zone;
512 if (cc->no_set_skip_hint)
515 set_pageblock_skip(page);
517 if (pfn < zone->compact_cached_free_pfn)
518 zone->compact_cached_free_pfn = pfn;
521 static inline bool isolation_suitable(struct compact_control *cc,
527 static inline bool pageblock_skip_persistent(struct page *page)
532 static inline void update_pageblock_skip(struct compact_control *cc,
533 struct page *page, unsigned long pfn)
537 static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
541 static bool test_and_set_skip(struct compact_control *cc, struct page *page)
545 #endif /* CONFIG_COMPACTION */
548 * Compaction requires the taking of some coarse locks that are potentially
549 * very heavily contended. For async compaction, trylock and record if the
550 * lock is contended. The lock will still be acquired but compaction will
551 * abort when the current block is finished regardless of success rate.
552 * Sync compaction acquires the lock.
554 * Always returns true which makes it easier to track lock state in callers.
556 static bool compact_lock_irqsave(spinlock_t *lock, unsigned long *flags,
557 struct compact_control *cc)
560 /* Track if the lock is contended in async mode */
561 if (cc->mode == MIGRATE_ASYNC && !cc->contended) {
562 if (spin_trylock_irqsave(lock, *flags))
565 cc->contended = true;
568 spin_lock_irqsave(lock, *flags);
573 * Compaction requires the taking of some coarse locks that are potentially
574 * very heavily contended. The lock should be periodically unlocked to avoid
575 * having disabled IRQs for a long time, even when there is nobody waiting on
576 * the lock. It might also be that allowing the IRQs will result in
577 * need_resched() becoming true. If scheduling is needed, compaction schedules.
578 * Either compaction type will also abort if a fatal signal is pending.
579 * In either case if the lock was locked, it is dropped and not regained.
581 * Returns true if compaction should abort due to fatal signal pending.
582 * Returns false when compaction can continue.
584 static bool compact_unlock_should_abort(spinlock_t *lock,
585 unsigned long flags, bool *locked, struct compact_control *cc)
588 spin_unlock_irqrestore(lock, flags);
592 if (fatal_signal_pending(current)) {
593 cc->contended = true;
603 * Isolate free pages onto a private freelist. If @strict is true, will abort
604 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
605 * (even though it may still end up isolating some pages).
607 static unsigned long isolate_freepages_block(struct compact_control *cc,
608 unsigned long *start_pfn,
609 unsigned long end_pfn,
610 struct list_head *freelist,
614 int nr_scanned = 0, total_isolated = 0;
616 unsigned long flags = 0;
618 unsigned long blockpfn = *start_pfn;
621 /* Strict mode is for isolation, speed is secondary */
625 page = pfn_to_page(blockpfn);
627 /* Isolate free pages. */
628 for (; blockpfn < end_pfn; blockpfn += stride, page += stride) {
632 * Periodically drop the lock (if held) regardless of its
633 * contention, to give chance to IRQs. Abort if fatal signal
636 if (!(blockpfn % COMPACT_CLUSTER_MAX)
637 && compact_unlock_should_abort(&cc->zone->lock, flags,
644 * For compound pages such as THP and hugetlbfs, we can save
645 * potentially a lot of iterations if we skip them at once.
646 * The check is racy, but we can consider only valid values
647 * and the only danger is skipping too much.
649 if (PageCompound(page)) {
650 const unsigned int order = compound_order(page);
652 if (blockpfn + (1UL << order) <= end_pfn) {
653 blockpfn += (1UL << order) - 1;
654 page += (1UL << order) - 1;
655 nr_scanned += (1UL << order) - 1;
661 if (!PageBuddy(page))
664 /* If we already hold the lock, we can skip some rechecking. */
666 locked = compact_lock_irqsave(&cc->zone->lock,
669 /* Recheck this is a buddy page under lock */
670 if (!PageBuddy(page))
674 /* Found a free page, will break it into order-0 pages */
675 order = buddy_order(page);
676 isolated = __isolate_free_page(page, order);
679 set_page_private(page, order);
681 nr_scanned += isolated - 1;
682 total_isolated += isolated;
683 cc->nr_freepages += isolated;
684 list_add_tail(&page->lru, &freelist[order]);
686 if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
687 blockpfn += isolated;
690 /* Advance to the end of split page */
691 blockpfn += isolated - 1;
692 page += isolated - 1;
702 spin_unlock_irqrestore(&cc->zone->lock, flags);
705 * Be careful to not go outside of the pageblock.
707 if (unlikely(blockpfn > end_pfn))
710 trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
711 nr_scanned, total_isolated);
713 /* Record how far we have got within the block */
714 *start_pfn = blockpfn;
717 * If strict isolation is requested by CMA then check that all the
718 * pages requested were isolated. If there were any failures, 0 is
719 * returned and CMA will fail.
721 if (strict && blockpfn < end_pfn)
724 cc->total_free_scanned += nr_scanned;
726 count_compact_events(COMPACTISOLATED, total_isolated);
727 return total_isolated;
731 * isolate_freepages_range() - isolate free pages.
732 * @cc: Compaction control structure.
733 * @start_pfn: The first PFN to start isolating.
734 * @end_pfn: The one-past-last PFN.
736 * Non-free pages, invalid PFNs, or zone boundaries within the
737 * [start_pfn, end_pfn) range are considered errors, cause function to
738 * undo its actions and return zero.
740 * Otherwise, function returns one-past-the-last PFN of isolated page
741 * (which may be greater then end_pfn if end fell in a middle of
745 isolate_freepages_range(struct compact_control *cc,
746 unsigned long start_pfn, unsigned long end_pfn)
748 unsigned long isolated, pfn, block_start_pfn, block_end_pfn;
750 struct list_head tmp_freepages[NR_PAGE_ORDERS];
752 for (order = 0; order < NR_PAGE_ORDERS; order++)
753 INIT_LIST_HEAD(&tmp_freepages[order]);
756 block_start_pfn = pageblock_start_pfn(pfn);
757 if (block_start_pfn < cc->zone->zone_start_pfn)
758 block_start_pfn = cc->zone->zone_start_pfn;
759 block_end_pfn = pageblock_end_pfn(pfn);
761 for (; pfn < end_pfn; pfn += isolated,
762 block_start_pfn = block_end_pfn,
763 block_end_pfn += pageblock_nr_pages) {
764 /* Protect pfn from changing by isolate_freepages_block */
765 unsigned long isolate_start_pfn = pfn;
768 * pfn could pass the block_end_pfn if isolated freepage
769 * is more than pageblock order. In this case, we adjust
770 * scanning range to right one.
772 if (pfn >= block_end_pfn) {
773 block_start_pfn = pageblock_start_pfn(pfn);
774 block_end_pfn = pageblock_end_pfn(pfn);
777 block_end_pfn = min(block_end_pfn, end_pfn);
779 if (!pageblock_pfn_to_page(block_start_pfn,
780 block_end_pfn, cc->zone))
783 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
784 block_end_pfn, tmp_freepages, 0, true);
787 * In strict mode, isolate_freepages_block() returns 0 if
788 * there are any holes in the block (ie. invalid PFNs or
795 * If we managed to isolate pages, it is always (1 << n) *
796 * pageblock_nr_pages for some non-negative n. (Max order
797 * page may span two pageblocks).
802 /* Loop terminated early, cleanup. */
803 release_free_list(tmp_freepages);
807 /* __isolate_free_page() does not map the pages */
808 split_map_pages(tmp_freepages);
810 /* We don't use freelists for anything. */
814 /* Similar to reclaim, but different enough that they don't share logic */
815 static bool too_many_isolated(struct compact_control *cc)
817 pg_data_t *pgdat = cc->zone->zone_pgdat;
820 unsigned long active, inactive, isolated;
822 inactive = node_page_state(pgdat, NR_INACTIVE_FILE) +
823 node_page_state(pgdat, NR_INACTIVE_ANON);
824 active = node_page_state(pgdat, NR_ACTIVE_FILE) +
825 node_page_state(pgdat, NR_ACTIVE_ANON);
826 isolated = node_page_state(pgdat, NR_ISOLATED_FILE) +
827 node_page_state(pgdat, NR_ISOLATED_ANON);
830 * Allow GFP_NOFS to isolate past the limit set for regular
831 * compaction runs. This prevents an ABBA deadlock when other
832 * compactors have already isolated to the limit, but are
833 * blocked on filesystem locks held by the GFP_NOFS thread.
835 if (cc->gfp_mask & __GFP_FS) {
840 too_many = isolated > (inactive + active) / 2;
842 wake_throttle_isolated(pgdat);
848 * skip_isolation_on_order() - determine when to skip folio isolation based on
849 * folio order and compaction target order
850 * @order: to-be-isolated folio order
851 * @target_order: compaction target order
853 * This avoids unnecessary folio isolations during compaction.
855 static bool skip_isolation_on_order(int order, int target_order)
858 * Unless we are performing global compaction (i.e.,
859 * is_via_compact_memory), skip any folios that are larger than the
860 * target order: we wouldn't be here if we'd have a free folio with
861 * the desired target_order, so migrating this folio would likely fail
864 if (!is_via_compact_memory(target_order) && order >= target_order)
867 * We limit memory compaction to pageblocks and won't try
868 * creating free blocks of memory that are larger than that.
870 return order >= pageblock_order;
874 * isolate_migratepages_block() - isolate all migrate-able pages within
876 * @cc: Compaction control structure.
877 * @low_pfn: The first PFN to isolate
878 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
879 * @mode: Isolation mode to be used.
881 * Isolate all pages that can be migrated from the range specified by
882 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
883 * Returns errno, like -EAGAIN or -EINTR in case e.g signal pending or congestion,
884 * -ENOMEM in case we could not allocate a page, or 0.
885 * cc->migrate_pfn will contain the next pfn to scan.
887 * The pages are isolated on cc->migratepages list (not required to be empty),
888 * and cc->nr_migratepages is updated accordingly.
891 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
892 unsigned long end_pfn, isolate_mode_t mode)
894 pg_data_t *pgdat = cc->zone->zone_pgdat;
895 unsigned long nr_scanned = 0, nr_isolated = 0;
896 struct lruvec *lruvec;
897 unsigned long flags = 0;
898 struct lruvec *locked = NULL;
899 struct folio *folio = NULL;
900 struct page *page = NULL, *valid_page = NULL;
901 struct address_space *mapping;
902 unsigned long start_pfn = low_pfn;
903 bool skip_on_failure = false;
904 unsigned long next_skip_pfn = 0;
905 bool skip_updated = false;
908 cc->migrate_pfn = low_pfn;
911 * Ensure that there are not too many pages isolated from the LRU
912 * list by either parallel reclaimers or compaction. If there are,
913 * delay for some time until fewer pages are isolated
915 while (unlikely(too_many_isolated(cc))) {
916 /* stop isolation if there are still pages not migrated */
917 if (cc->nr_migratepages)
920 /* async migration should just abort */
921 if (cc->mode == MIGRATE_ASYNC)
924 reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED);
926 if (fatal_signal_pending(current))
932 if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
933 skip_on_failure = true;
934 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
937 /* Time to isolate some pages for migration */
938 for (; low_pfn < end_pfn; low_pfn++) {
939 bool is_dirty, is_unevictable;
941 if (skip_on_failure && low_pfn >= next_skip_pfn) {
943 * We have isolated all migration candidates in the
944 * previous order-aligned block, and did not skip it due
945 * to failure. We should migrate the pages now and
946 * hopefully succeed compaction.
952 * We failed to isolate in the previous order-aligned
953 * block. Set the new boundary to the end of the
954 * current block. Note we can't simply increase
955 * next_skip_pfn by 1 << order, as low_pfn might have
956 * been incremented by a higher number due to skipping
957 * a compound or a high-order buddy page in the
958 * previous loop iteration.
960 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
964 * Periodically drop the lock (if held) regardless of its
965 * contention, to give chance to IRQs. Abort completely if
966 * a fatal signal is pending.
968 if (!(low_pfn % COMPACT_CLUSTER_MAX)) {
970 unlock_page_lruvec_irqrestore(locked, flags);
974 if (fatal_signal_pending(current)) {
975 cc->contended = true;
986 page = pfn_to_page(low_pfn);
989 * Check if the pageblock has already been marked skipped.
990 * Only the first PFN is checked as the caller isolates
991 * COMPACT_CLUSTER_MAX at a time so the second call must
992 * not falsely conclude that the block should be skipped.
994 if (!valid_page && (pageblock_aligned(low_pfn) ||
995 low_pfn == cc->zone->zone_start_pfn)) {
996 if (!isolation_suitable(cc, page)) {
1004 if (PageHuge(page)) {
1006 * skip hugetlbfs if we are not compacting for pages
1007 * bigger than its order. THPs and other compound pages
1008 * are handled below.
1010 if (!cc->alloc_contig) {
1011 const unsigned int order = compound_order(page);
1013 if (order <= MAX_PAGE_ORDER) {
1014 low_pfn += (1UL << order) - 1;
1015 nr_scanned += (1UL << order) - 1;
1019 /* for alloc_contig case */
1021 unlock_page_lruvec_irqrestore(locked, flags);
1025 ret = isolate_or_dissolve_huge_page(page, &cc->migratepages);
1028 * Fail isolation in case isolate_or_dissolve_huge_page()
1029 * reports an error. In case of -ENOMEM, abort right away.
1032 /* Do not report -EBUSY down the chain */
1035 low_pfn += compound_nr(page) - 1;
1036 nr_scanned += compound_nr(page) - 1;
1040 if (PageHuge(page)) {
1042 * Hugepage was successfully isolated and placed
1043 * on the cc->migratepages list.
1045 folio = page_folio(page);
1046 low_pfn += folio_nr_pages(folio) - 1;
1047 goto isolate_success_no_list;
1051 * Ok, the hugepage was dissolved. Now these pages are
1052 * Buddy and cannot be re-allocated because they are
1053 * isolated. Fall-through as the check below handles
1059 * Skip if free. We read page order here without zone lock
1060 * which is generally unsafe, but the race window is small and
1061 * the worst thing that can happen is that we skip some
1062 * potential isolation targets.
1064 if (PageBuddy(page)) {
1065 unsigned long freepage_order = buddy_order_unsafe(page);
1068 * Without lock, we cannot be sure that what we got is
1069 * a valid page order. Consider only values in the
1070 * valid order range to prevent low_pfn overflow.
1072 if (freepage_order > 0 && freepage_order <= MAX_PAGE_ORDER) {
1073 low_pfn += (1UL << freepage_order) - 1;
1074 nr_scanned += (1UL << freepage_order) - 1;
1080 * Regardless of being on LRU, compound pages such as THP
1081 * (hugetlbfs is handled above) are not to be compacted unless
1082 * we are attempting an allocation larger than the compound
1083 * page size. We can potentially save a lot of iterations if we
1084 * skip them at once. The check is racy, but we can consider
1085 * only valid values and the only danger is skipping too much.
1087 if (PageCompound(page) && !cc->alloc_contig) {
1088 const unsigned int order = compound_order(page);
1090 /* Skip based on page order and compaction target order. */
1091 if (skip_isolation_on_order(order, cc->order)) {
1092 if (order <= MAX_PAGE_ORDER) {
1093 low_pfn += (1UL << order) - 1;
1094 nr_scanned += (1UL << order) - 1;
1101 * Check may be lockless but that's ok as we recheck later.
1102 * It's possible to migrate LRU and non-lru movable pages.
1103 * Skip any other type of page
1105 if (!PageLRU(page)) {
1107 * __PageMovable can return false positive so we need
1108 * to verify it under page_lock.
1110 if (unlikely(__PageMovable(page)) &&
1111 !PageIsolated(page)) {
1113 unlock_page_lruvec_irqrestore(locked, flags);
1117 if (isolate_movable_page(page, mode)) {
1118 folio = page_folio(page);
1119 goto isolate_success;
1127 * Be careful not to clear PageLRU until after we're
1128 * sure the page is not being freed elsewhere -- the
1129 * page release code relies on it.
1131 folio = folio_get_nontail_page(page);
1132 if (unlikely(!folio))
1136 * Migration will fail if an anonymous page is pinned in memory,
1137 * so avoid taking lru_lock and isolating it unnecessarily in an
1138 * admittedly racy check.
1140 mapping = folio_mapping(folio);
1141 if (!mapping && (folio_ref_count(folio) - 1) > folio_mapcount(folio))
1142 goto isolate_fail_put;
1145 * Only allow to migrate anonymous pages in GFP_NOFS context
1146 * because those do not depend on fs locks.
1148 if (!(cc->gfp_mask & __GFP_FS) && mapping)
1149 goto isolate_fail_put;
1151 /* Only take pages on LRU: a check now makes later tests safe */
1152 if (!folio_test_lru(folio))
1153 goto isolate_fail_put;
1155 is_unevictable = folio_test_unevictable(folio);
1157 /* Compaction might skip unevictable pages but CMA takes them */
1158 if (!(mode & ISOLATE_UNEVICTABLE) && is_unevictable)
1159 goto isolate_fail_put;
1162 * To minimise LRU disruption, the caller can indicate with
1163 * ISOLATE_ASYNC_MIGRATE that it only wants to isolate pages
1164 * it will be able to migrate without blocking - clean pages
1165 * for the most part. PageWriteback would require blocking.
1167 if ((mode & ISOLATE_ASYNC_MIGRATE) && folio_test_writeback(folio))
1168 goto isolate_fail_put;
1170 is_dirty = folio_test_dirty(folio);
1172 if (((mode & ISOLATE_ASYNC_MIGRATE) && is_dirty) ||
1173 (mapping && is_unevictable)) {
1174 bool migrate_dirty = true;
1178 * Only folios without mappings or that have
1179 * a ->migrate_folio callback are possible to migrate
1182 * Folios from unmovable mappings are not migratable.
1184 * However, we can be racing with truncation, which can
1185 * free the mapping that we need to check. Truncation
1186 * holds the folio lock until after the folio is removed
1187 * from the page so holding it ourselves is sufficient.
1189 * To avoid locking the folio just to check unmovable,
1190 * assume every unmovable folio is also unevictable,
1191 * which is a cheaper test. If our assumption goes
1192 * wrong, it's not a correctness bug, just potentially
1195 if (!folio_trylock(folio))
1196 goto isolate_fail_put;
1198 mapping = folio_mapping(folio);
1199 if ((mode & ISOLATE_ASYNC_MIGRATE) && is_dirty) {
1200 migrate_dirty = !mapping ||
1201 mapping->a_ops->migrate_folio;
1203 is_unmovable = mapping && mapping_unmovable(mapping);
1204 folio_unlock(folio);
1205 if (!migrate_dirty || is_unmovable)
1206 goto isolate_fail_put;
1209 /* Try isolate the folio */
1210 if (!folio_test_clear_lru(folio))
1211 goto isolate_fail_put;
1213 lruvec = folio_lruvec(folio);
1215 /* If we already hold the lock, we can skip some rechecking */
1216 if (lruvec != locked) {
1218 unlock_page_lruvec_irqrestore(locked, flags);
1220 compact_lock_irqsave(&lruvec->lru_lock, &flags, cc);
1223 lruvec_memcg_debug(lruvec, folio);
1226 * Try get exclusive access under lock. If marked for
1227 * skip, the scan is aborted unless the current context
1228 * is a rescan to reach the end of the pageblock.
1230 if (!skip_updated && valid_page) {
1231 skip_updated = true;
1232 if (test_and_set_skip(cc, valid_page) &&
1233 !cc->finish_pageblock) {
1240 * Check LRU folio order under the lock
1242 if (unlikely(skip_isolation_on_order(folio_order(folio),
1244 !cc->alloc_contig)) {
1245 low_pfn += folio_nr_pages(folio) - 1;
1246 nr_scanned += folio_nr_pages(folio) - 1;
1247 folio_set_lru(folio);
1248 goto isolate_fail_put;
1252 /* The folio is taken off the LRU */
1253 if (folio_test_large(folio))
1254 low_pfn += folio_nr_pages(folio) - 1;
1256 /* Successfully isolated */
1257 lruvec_del_folio(lruvec, folio);
1258 node_stat_mod_folio(folio,
1259 NR_ISOLATED_ANON + folio_is_file_lru(folio),
1260 folio_nr_pages(folio));
1263 list_add(&folio->lru, &cc->migratepages);
1264 isolate_success_no_list:
1265 cc->nr_migratepages += folio_nr_pages(folio);
1266 nr_isolated += folio_nr_pages(folio);
1267 nr_scanned += folio_nr_pages(folio) - 1;
1270 * Avoid isolating too much unless this block is being
1271 * fully scanned (e.g. dirty/writeback pages, parallel allocation)
1272 * or a lock is contended. For contention, isolate quickly to
1273 * potentially remove one source of contention.
1275 if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX &&
1276 !cc->finish_pageblock && !cc->contended) {
1284 /* Avoid potential deadlock in freeing page under lru_lock */
1286 unlock_page_lruvec_irqrestore(locked, flags);
1292 if (!skip_on_failure && ret != -ENOMEM)
1296 * We have isolated some pages, but then failed. Release them
1297 * instead of migrating, as we cannot form the cc->order buddy
1302 unlock_page_lruvec_irqrestore(locked, flags);
1305 putback_movable_pages(&cc->migratepages);
1306 cc->nr_migratepages = 0;
1310 if (low_pfn < next_skip_pfn) {
1311 low_pfn = next_skip_pfn - 1;
1313 * The check near the loop beginning would have updated
1314 * next_skip_pfn too, but this is a bit simpler.
1316 next_skip_pfn += 1UL << cc->order;
1324 * The PageBuddy() check could have potentially brought us outside
1325 * the range to be scanned.
1327 if (unlikely(low_pfn > end_pfn))
1334 unlock_page_lruvec_irqrestore(locked, flags);
1336 folio_set_lru(folio);
1341 * Update the cached scanner pfn once the pageblock has been scanned.
1342 * Pages will either be migrated in which case there is no point
1343 * scanning in the near future or migration failed in which case the
1344 * failure reason may persist. The block is marked for skipping if
1345 * there were no pages isolated in the block or if the block is
1346 * rescanned twice in a row.
1348 if (low_pfn == end_pfn && (!nr_isolated || cc->finish_pageblock)) {
1349 if (!cc->no_set_skip_hint && valid_page && !skip_updated)
1350 set_pageblock_skip(valid_page);
1351 update_cached_migrate(cc, low_pfn);
1354 trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
1355 nr_scanned, nr_isolated);
1358 cc->total_migrate_scanned += nr_scanned;
1360 count_compact_events(COMPACTISOLATED, nr_isolated);
1362 cc->migrate_pfn = low_pfn;
1368 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
1369 * @cc: Compaction control structure.
1370 * @start_pfn: The first PFN to start isolating.
1371 * @end_pfn: The one-past-last PFN.
1373 * Returns -EAGAIN when contented, -EINTR in case of a signal pending, -ENOMEM
1374 * in case we could not allocate a page, or 0.
1377 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
1378 unsigned long end_pfn)
1380 unsigned long pfn, block_start_pfn, block_end_pfn;
1383 /* Scan block by block. First and last block may be incomplete */
1385 block_start_pfn = pageblock_start_pfn(pfn);
1386 if (block_start_pfn < cc->zone->zone_start_pfn)
1387 block_start_pfn = cc->zone->zone_start_pfn;
1388 block_end_pfn = pageblock_end_pfn(pfn);
1390 for (; pfn < end_pfn; pfn = block_end_pfn,
1391 block_start_pfn = block_end_pfn,
1392 block_end_pfn += pageblock_nr_pages) {
1394 block_end_pfn = min(block_end_pfn, end_pfn);
1396 if (!pageblock_pfn_to_page(block_start_pfn,
1397 block_end_pfn, cc->zone))
1400 ret = isolate_migratepages_block(cc, pfn, block_end_pfn,
1401 ISOLATE_UNEVICTABLE);
1406 if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX)
1413 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
1414 #ifdef CONFIG_COMPACTION
1416 static bool suitable_migration_source(struct compact_control *cc,
1421 if (pageblock_skip_persistent(page))
1424 if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction)
1427 block_mt = get_pageblock_migratetype(page);
1429 if (cc->migratetype == MIGRATE_MOVABLE)
1430 return is_migrate_movable(block_mt);
1432 return block_mt == cc->migratetype;
1435 /* Returns true if the page is within a block suitable for migration to */
1436 static bool suitable_migration_target(struct compact_control *cc,
1439 /* If the page is a large free page, then disallow migration */
1440 if (PageBuddy(page)) {
1441 int order = cc->order > 0 ? cc->order : pageblock_order;
1444 * We are checking page_order without zone->lock taken. But
1445 * the only small danger is that we skip a potentially suitable
1446 * pageblock, so it's not worth to check order for valid range.
1448 if (buddy_order_unsafe(page) >= order)
1452 if (cc->ignore_block_suitable)
1455 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1456 if (is_migrate_movable(get_pageblock_migratetype(page)))
1459 /* Otherwise skip the block */
1463 static inline unsigned int
1464 freelist_scan_limit(struct compact_control *cc)
1466 unsigned short shift = BITS_PER_LONG - 1;
1468 return (COMPACT_CLUSTER_MAX >> min(shift, cc->fast_search_fail)) + 1;
1472 * Test whether the free scanner has reached the same or lower pageblock than
1473 * the migration scanner, and compaction should thus terminate.
1475 static inline bool compact_scanners_met(struct compact_control *cc)
1477 return (cc->free_pfn >> pageblock_order)
1478 <= (cc->migrate_pfn >> pageblock_order);
1482 * Used when scanning for a suitable migration target which scans freelists
1483 * in reverse. Reorders the list such as the unscanned pages are scanned
1484 * first on the next iteration of the free scanner
1487 move_freelist_head(struct list_head *freelist, struct page *freepage)
1491 if (!list_is_first(&freepage->buddy_list, freelist)) {
1492 list_cut_before(&sublist, freelist, &freepage->buddy_list);
1493 list_splice_tail(&sublist, freelist);
1498 * Similar to move_freelist_head except used by the migration scanner
1499 * when scanning forward. It's possible for these list operations to
1500 * move against each other if they search the free list exactly in
1504 move_freelist_tail(struct list_head *freelist, struct page *freepage)
1508 if (!list_is_last(&freepage->buddy_list, freelist)) {
1509 list_cut_position(&sublist, freelist, &freepage->buddy_list);
1510 list_splice_tail(&sublist, freelist);
1515 fast_isolate_around(struct compact_control *cc, unsigned long pfn)
1517 unsigned long start_pfn, end_pfn;
1520 /* Do not search around if there are enough pages already */
1521 if (cc->nr_freepages >= cc->nr_migratepages)
1524 /* Minimise scanning during async compaction */
1525 if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC)
1528 /* Pageblock boundaries */
1529 start_pfn = max(pageblock_start_pfn(pfn), cc->zone->zone_start_pfn);
1530 end_pfn = min(pageblock_end_pfn(pfn), zone_end_pfn(cc->zone));
1532 page = pageblock_pfn_to_page(start_pfn, end_pfn, cc->zone);
1536 isolate_freepages_block(cc, &start_pfn, end_pfn, cc->freepages, 1, false);
1538 /* Skip this pageblock in the future as it's full or nearly full */
1539 if (start_pfn == end_pfn && !cc->no_set_skip_hint)
1540 set_pageblock_skip(page);
1543 /* Search orders in round-robin fashion */
1544 static int next_search_order(struct compact_control *cc, int order)
1548 order = cc->order - 1;
1550 /* Search wrapped around? */
1551 if (order == cc->search_order) {
1553 if (cc->search_order < 0)
1554 cc->search_order = cc->order - 1;
1561 static void fast_isolate_freepages(struct compact_control *cc)
1563 unsigned int limit = max(1U, freelist_scan_limit(cc) >> 1);
1564 unsigned int nr_scanned = 0, total_isolated = 0;
1565 unsigned long low_pfn, min_pfn, highest = 0;
1566 unsigned long nr_isolated = 0;
1567 unsigned long distance;
1568 struct page *page = NULL;
1569 bool scan_start = false;
1572 /* Full compaction passes in a negative order */
1577 * If starting the scan, use a deeper search and use the highest
1578 * PFN found if a suitable one is not found.
1580 if (cc->free_pfn >= cc->zone->compact_init_free_pfn) {
1581 limit = pageblock_nr_pages >> 1;
1586 * Preferred point is in the top quarter of the scan space but take
1587 * a pfn from the top half if the search is problematic.
1589 distance = (cc->free_pfn - cc->migrate_pfn);
1590 low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2));
1591 min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1));
1593 if (WARN_ON_ONCE(min_pfn > low_pfn))
1597 * Search starts from the last successful isolation order or the next
1598 * order to search after a previous failure
1600 cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order);
1602 for (order = cc->search_order;
1603 !page && order >= 0;
1604 order = next_search_order(cc, order)) {
1605 struct free_area *area = &cc->zone->free_area[order];
1606 struct list_head *freelist;
1607 struct page *freepage;
1608 unsigned long flags;
1609 unsigned int order_scanned = 0;
1610 unsigned long high_pfn = 0;
1615 spin_lock_irqsave(&cc->zone->lock, flags);
1616 freelist = &area->free_list[MIGRATE_MOVABLE];
1617 list_for_each_entry_reverse(freepage, freelist, buddy_list) {
1622 pfn = page_to_pfn(freepage);
1625 highest = max(pageblock_start_pfn(pfn),
1626 cc->zone->zone_start_pfn);
1628 if (pfn >= low_pfn) {
1629 cc->fast_search_fail = 0;
1630 cc->search_order = order;
1635 if (pfn >= min_pfn && pfn > high_pfn) {
1638 /* Shorten the scan if a candidate is found */
1642 if (order_scanned >= limit)
1646 /* Use a maximum candidate pfn if a preferred one was not found */
1647 if (!page && high_pfn) {
1648 page = pfn_to_page(high_pfn);
1650 /* Update freepage for the list reorder below */
1654 /* Reorder to so a future search skips recent pages */
1655 move_freelist_head(freelist, freepage);
1657 /* Isolate the page if available */
1659 if (__isolate_free_page(page, order)) {
1660 set_page_private(page, order);
1661 nr_isolated = 1 << order;
1662 nr_scanned += nr_isolated - 1;
1663 total_isolated += nr_isolated;
1664 cc->nr_freepages += nr_isolated;
1665 list_add_tail(&page->lru, &cc->freepages[order]);
1666 count_compact_events(COMPACTISOLATED, nr_isolated);
1668 /* If isolation fails, abort the search */
1669 order = cc->search_order + 1;
1674 spin_unlock_irqrestore(&cc->zone->lock, flags);
1676 /* Skip fast search if enough freepages isolated */
1677 if (cc->nr_freepages >= cc->nr_migratepages)
1681 * Smaller scan on next order so the total scan is related
1682 * to freelist_scan_limit.
1684 if (order_scanned >= limit)
1685 limit = max(1U, limit >> 1);
1688 trace_mm_compaction_fast_isolate_freepages(min_pfn, cc->free_pfn,
1689 nr_scanned, total_isolated);
1692 cc->fast_search_fail++;
1695 * Use the highest PFN found above min. If one was
1696 * not found, be pessimistic for direct compaction
1697 * and use the min mark.
1699 if (highest >= min_pfn) {
1700 page = pfn_to_page(highest);
1701 cc->free_pfn = highest;
1703 if (cc->direct_compaction && pfn_valid(min_pfn)) {
1704 page = pageblock_pfn_to_page(min_pfn,
1705 min(pageblock_end_pfn(min_pfn),
1706 zone_end_pfn(cc->zone)),
1708 if (page && !suitable_migration_target(cc, page))
1711 cc->free_pfn = min_pfn;
1717 if (highest && highest >= cc->zone->compact_cached_free_pfn) {
1718 highest -= pageblock_nr_pages;
1719 cc->zone->compact_cached_free_pfn = highest;
1722 cc->total_free_scanned += nr_scanned;
1726 low_pfn = page_to_pfn(page);
1727 fast_isolate_around(cc, low_pfn);
1731 * Based on information in the current compact_control, find blocks
1732 * suitable for isolating free pages from and then isolate them.
1734 static void isolate_freepages(struct compact_control *cc)
1736 struct zone *zone = cc->zone;
1738 unsigned long block_start_pfn; /* start of current pageblock */
1739 unsigned long isolate_start_pfn; /* exact pfn we start at */
1740 unsigned long block_end_pfn; /* end of current pageblock */
1741 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
1742 unsigned int stride;
1744 /* Try a small search of the free lists for a candidate */
1745 fast_isolate_freepages(cc);
1746 if (cc->nr_freepages)
1750 * Initialise the free scanner. The starting point is where we last
1751 * successfully isolated from, zone-cached value, or the end of the
1752 * zone when isolating for the first time. For looping we also need
1753 * this pfn aligned down to the pageblock boundary, because we do
1754 * block_start_pfn -= pageblock_nr_pages in the for loop.
1755 * For ending point, take care when isolating in last pageblock of a
1756 * zone which ends in the middle of a pageblock.
1757 * The low boundary is the end of the pageblock the migration scanner
1760 isolate_start_pfn = cc->free_pfn;
1761 block_start_pfn = pageblock_start_pfn(isolate_start_pfn);
1762 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1763 zone_end_pfn(zone));
1764 low_pfn = pageblock_end_pfn(cc->migrate_pfn);
1765 stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1;
1768 * Isolate free pages until enough are available to migrate the
1769 * pages on cc->migratepages. We stop searching if the migrate
1770 * and free page scanners meet or enough free pages are isolated.
1772 for (; block_start_pfn >= low_pfn;
1773 block_end_pfn = block_start_pfn,
1774 block_start_pfn -= pageblock_nr_pages,
1775 isolate_start_pfn = block_start_pfn) {
1776 unsigned long nr_isolated;
1779 * This can iterate a massively long zone without finding any
1780 * suitable migration targets, so periodically check resched.
1782 if (!(block_start_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages)))
1785 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1788 unsigned long next_pfn;
1790 next_pfn = skip_offline_sections_reverse(block_start_pfn);
1792 block_start_pfn = max(next_pfn, low_pfn);
1797 /* Check the block is suitable for migration */
1798 if (!suitable_migration_target(cc, page))
1801 /* If isolation recently failed, do not retry */
1802 if (!isolation_suitable(cc, page))
1805 /* Found a block suitable for isolating free pages from. */
1806 nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn,
1807 block_end_pfn, cc->freepages, stride, false);
1809 /* Update the skip hint if the full pageblock was scanned */
1810 if (isolate_start_pfn == block_end_pfn)
1811 update_pageblock_skip(cc, page, block_start_pfn -
1812 pageblock_nr_pages);
1814 /* Are enough freepages isolated? */
1815 if (cc->nr_freepages >= cc->nr_migratepages) {
1816 if (isolate_start_pfn >= block_end_pfn) {
1818 * Restart at previous pageblock if more
1819 * freepages can be isolated next time.
1822 block_start_pfn - pageblock_nr_pages;
1825 } else if (isolate_start_pfn < block_end_pfn) {
1827 * If isolation failed early, do not continue
1833 /* Adjust stride depending on isolation */
1838 stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1);
1842 * Record where the free scanner will restart next time. Either we
1843 * broke from the loop and set isolate_start_pfn based on the last
1844 * call to isolate_freepages_block(), or we met the migration scanner
1845 * and the loop terminated due to isolate_start_pfn < low_pfn
1847 cc->free_pfn = isolate_start_pfn;
1851 * This is a migrate-callback that "allocates" freepages by taking pages
1852 * from the isolated freelists in the block we are migrating to.
1854 static struct folio *compaction_alloc(struct folio *src, unsigned long data)
1856 struct compact_control *cc = (struct compact_control *)data;
1858 int order = folio_order(src);
1859 bool has_isolated_pages = false;
1861 struct page *freepage;
1865 for (start_order = order; start_order < NR_PAGE_ORDERS; start_order++)
1866 if (!list_empty(&cc->freepages[start_order]))
1869 /* no free pages in the list */
1870 if (start_order == NR_PAGE_ORDERS) {
1871 if (has_isolated_pages)
1873 isolate_freepages(cc);
1874 has_isolated_pages = true;
1878 freepage = list_first_entry(&cc->freepages[start_order], struct page,
1880 size = 1 << start_order;
1882 list_del(&freepage->lru);
1884 while (start_order > order) {
1888 list_add(&freepage[size].lru, &cc->freepages[start_order]);
1889 set_page_private(&freepage[size], start_order);
1891 dst = (struct folio *)freepage;
1893 post_alloc_hook(&dst->page, order, __GFP_MOVABLE);
1895 prep_compound_page(&dst->page, order);
1896 cc->nr_freepages -= 1 << order;
1897 cc->nr_migratepages -= 1 << order;
1898 return page_rmappable_folio(&dst->page);
1902 * This is a migrate-callback that "frees" freepages back to the isolated
1903 * freelist. All pages on the freelist are from the same zone, so there is no
1904 * special handling needed for NUMA.
1906 static void compaction_free(struct folio *dst, unsigned long data)
1908 struct compact_control *cc = (struct compact_control *)data;
1909 int order = folio_order(dst);
1910 struct page *page = &dst->page;
1912 if (folio_put_testzero(dst)) {
1913 free_pages_prepare(page, order);
1914 list_add(&dst->lru, &cc->freepages[order]);
1915 cc->nr_freepages += 1 << order;
1917 cc->nr_migratepages += 1 << order;
1919 * someone else has referenced the page, we cannot take it back to our
1924 /* possible outcome of isolate_migratepages */
1926 ISOLATE_ABORT, /* Abort compaction now */
1927 ISOLATE_NONE, /* No pages isolated, continue scanning */
1928 ISOLATE_SUCCESS, /* Pages isolated, migrate */
1929 } isolate_migrate_t;
1932 * Allow userspace to control policy on scanning the unevictable LRU for
1933 * compactable pages.
1935 static int sysctl_compact_unevictable_allowed __read_mostly = CONFIG_COMPACT_UNEVICTABLE_DEFAULT;
1937 * Tunable for proactive compaction. It determines how
1938 * aggressively the kernel should compact memory in the
1939 * background. It takes values in the range [0, 100].
1941 static unsigned int __read_mostly sysctl_compaction_proactiveness = 20;
1942 static int sysctl_extfrag_threshold = 500;
1943 static int __read_mostly sysctl_compact_memory;
1946 update_fast_start_pfn(struct compact_control *cc, unsigned long pfn)
1948 if (cc->fast_start_pfn == ULONG_MAX)
1951 if (!cc->fast_start_pfn)
1952 cc->fast_start_pfn = pfn;
1954 cc->fast_start_pfn = min(cc->fast_start_pfn, pfn);
1957 static inline unsigned long
1958 reinit_migrate_pfn(struct compact_control *cc)
1960 if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX)
1961 return cc->migrate_pfn;
1963 cc->migrate_pfn = cc->fast_start_pfn;
1964 cc->fast_start_pfn = ULONG_MAX;
1966 return cc->migrate_pfn;
1970 * Briefly search the free lists for a migration source that already has
1971 * some free pages to reduce the number of pages that need migration
1972 * before a pageblock is free.
1974 static unsigned long fast_find_migrateblock(struct compact_control *cc)
1976 unsigned int limit = freelist_scan_limit(cc);
1977 unsigned int nr_scanned = 0;
1978 unsigned long distance;
1979 unsigned long pfn = cc->migrate_pfn;
1980 unsigned long high_pfn;
1982 bool found_block = false;
1984 /* Skip hints are relied on to avoid repeats on the fast search */
1985 if (cc->ignore_skip_hint)
1989 * If the pageblock should be finished then do not select a different
1992 if (cc->finish_pageblock)
1996 * If the migrate_pfn is not at the start of a zone or the start
1997 * of a pageblock then assume this is a continuation of a previous
1998 * scan restarted due to COMPACT_CLUSTER_MAX.
2000 if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn))
2004 * For smaller orders, just linearly scan as the number of pages
2005 * to migrate should be relatively small and does not necessarily
2006 * justify freeing up a large block for a small allocation.
2008 if (cc->order <= PAGE_ALLOC_COSTLY_ORDER)
2012 * Only allow kcompactd and direct requests for movable pages to
2013 * quickly clear out a MOVABLE pageblock for allocation. This
2014 * reduces the risk that a large movable pageblock is freed for
2015 * an unmovable/reclaimable small allocation.
2017 if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE)
2021 * When starting the migration scanner, pick any pageblock within the
2022 * first half of the search space. Otherwise try and pick a pageblock
2023 * within the first eighth to reduce the chances that a migration
2024 * target later becomes a source.
2026 distance = (cc->free_pfn - cc->migrate_pfn) >> 1;
2027 if (cc->migrate_pfn != cc->zone->zone_start_pfn)
2029 high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance);
2031 for (order = cc->order - 1;
2032 order >= PAGE_ALLOC_COSTLY_ORDER && !found_block && nr_scanned < limit;
2034 struct free_area *area = &cc->zone->free_area[order];
2035 struct list_head *freelist;
2036 unsigned long flags;
2037 struct page *freepage;
2042 spin_lock_irqsave(&cc->zone->lock, flags);
2043 freelist = &area->free_list[MIGRATE_MOVABLE];
2044 list_for_each_entry(freepage, freelist, buddy_list) {
2045 unsigned long free_pfn;
2047 if (nr_scanned++ >= limit) {
2048 move_freelist_tail(freelist, freepage);
2052 free_pfn = page_to_pfn(freepage);
2053 if (free_pfn < high_pfn) {
2055 * Avoid if skipped recently. Ideally it would
2056 * move to the tail but even safe iteration of
2057 * the list assumes an entry is deleted, not
2060 if (get_pageblock_skip(freepage))
2063 /* Reorder to so a future search skips recent pages */
2064 move_freelist_tail(freelist, freepage);
2066 update_fast_start_pfn(cc, free_pfn);
2067 pfn = pageblock_start_pfn(free_pfn);
2068 if (pfn < cc->zone->zone_start_pfn)
2069 pfn = cc->zone->zone_start_pfn;
2070 cc->fast_search_fail = 0;
2075 spin_unlock_irqrestore(&cc->zone->lock, flags);
2078 cc->total_migrate_scanned += nr_scanned;
2081 * If fast scanning failed then use a cached entry for a page block
2082 * that had free pages as the basis for starting a linear scan.
2085 cc->fast_search_fail++;
2086 pfn = reinit_migrate_pfn(cc);
2092 * Isolate all pages that can be migrated from the first suitable block,
2093 * starting at the block pointed to by the migrate scanner pfn within
2096 static isolate_migrate_t isolate_migratepages(struct compact_control *cc)
2098 unsigned long block_start_pfn;
2099 unsigned long block_end_pfn;
2100 unsigned long low_pfn;
2102 const isolate_mode_t isolate_mode =
2103 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
2104 (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
2105 bool fast_find_block;
2108 * Start at where we last stopped, or beginning of the zone as
2109 * initialized by compact_zone(). The first failure will use
2110 * the lowest PFN as the starting point for linear scanning.
2112 low_pfn = fast_find_migrateblock(cc);
2113 block_start_pfn = pageblock_start_pfn(low_pfn);
2114 if (block_start_pfn < cc->zone->zone_start_pfn)
2115 block_start_pfn = cc->zone->zone_start_pfn;
2118 * fast_find_migrateblock() has already ensured the pageblock is not
2119 * set with a skipped flag, so to avoid the isolation_suitable check
2120 * below again, check whether the fast search was successful.
2122 fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail;
2124 /* Only scan within a pageblock boundary */
2125 block_end_pfn = pageblock_end_pfn(low_pfn);
2128 * Iterate over whole pageblocks until we find the first suitable.
2129 * Do not cross the free scanner.
2131 for (; block_end_pfn <= cc->free_pfn;
2132 fast_find_block = false,
2133 cc->migrate_pfn = low_pfn = block_end_pfn,
2134 block_start_pfn = block_end_pfn,
2135 block_end_pfn += pageblock_nr_pages) {
2138 * This can potentially iterate a massively long zone with
2139 * many pageblocks unsuitable, so periodically check if we
2142 if (!(low_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages)))
2145 page = pageblock_pfn_to_page(block_start_pfn,
2146 block_end_pfn, cc->zone);
2148 unsigned long next_pfn;
2150 next_pfn = skip_offline_sections(block_start_pfn);
2152 block_end_pfn = min(next_pfn, cc->free_pfn);
2157 * If isolation recently failed, do not retry. Only check the
2158 * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock
2159 * to be visited multiple times. Assume skip was checked
2160 * before making it "skip" so other compaction instances do
2161 * not scan the same block.
2163 if ((pageblock_aligned(low_pfn) ||
2164 low_pfn == cc->zone->zone_start_pfn) &&
2165 !fast_find_block && !isolation_suitable(cc, page))
2169 * For async direct compaction, only scan the pageblocks of the
2170 * same migratetype without huge pages. Async direct compaction
2171 * is optimistic to see if the minimum amount of work satisfies
2172 * the allocation. The cached PFN is updated as it's possible
2173 * that all remaining blocks between source and target are
2174 * unsuitable and the compaction scanners fail to meet.
2176 if (!suitable_migration_source(cc, page)) {
2177 update_cached_migrate(cc, block_end_pfn);
2181 /* Perform the isolation */
2182 if (isolate_migratepages_block(cc, low_pfn, block_end_pfn,
2184 return ISOLATE_ABORT;
2187 * Either we isolated something and proceed with migration. Or
2188 * we failed and compact_zone should decide if we should
2194 return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
2198 * Determine whether kswapd is (or recently was!) running on this node.
2200 * pgdat_kswapd_lock() pins pgdat->kswapd, so a concurrent kswapd_stop() can't
2203 static bool kswapd_is_running(pg_data_t *pgdat)
2207 pgdat_kswapd_lock(pgdat);
2208 running = pgdat->kswapd && task_is_running(pgdat->kswapd);
2209 pgdat_kswapd_unlock(pgdat);
2215 * A zone's fragmentation score is the external fragmentation wrt to the
2216 * COMPACTION_HPAGE_ORDER. It returns a value in the range [0, 100].
2218 static unsigned int fragmentation_score_zone(struct zone *zone)
2220 return extfrag_for_order(zone, COMPACTION_HPAGE_ORDER);
2224 * A weighted zone's fragmentation score is the external fragmentation
2225 * wrt to the COMPACTION_HPAGE_ORDER scaled by the zone's size. It
2226 * returns a value in the range [0, 100].
2228 * The scaling factor ensures that proactive compaction focuses on larger
2229 * zones like ZONE_NORMAL, rather than smaller, specialized zones like
2230 * ZONE_DMA32. For smaller zones, the score value remains close to zero,
2231 * and thus never exceeds the high threshold for proactive compaction.
2233 static unsigned int fragmentation_score_zone_weighted(struct zone *zone)
2235 unsigned long score;
2237 score = zone->present_pages * fragmentation_score_zone(zone);
2238 return div64_ul(score, zone->zone_pgdat->node_present_pages + 1);
2242 * The per-node proactive (background) compaction process is started by its
2243 * corresponding kcompactd thread when the node's fragmentation score
2244 * exceeds the high threshold. The compaction process remains active till
2245 * the node's score falls below the low threshold, or one of the back-off
2246 * conditions is met.
2248 static unsigned int fragmentation_score_node(pg_data_t *pgdat)
2250 unsigned int score = 0;
2253 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2256 zone = &pgdat->node_zones[zoneid];
2257 if (!populated_zone(zone))
2259 score += fragmentation_score_zone_weighted(zone);
2265 static unsigned int fragmentation_score_wmark(bool low)
2267 unsigned int wmark_low;
2270 * Cap the low watermark to avoid excessive compaction
2271 * activity in case a user sets the proactiveness tunable
2272 * close to 100 (maximum).
2274 wmark_low = max(100U - sysctl_compaction_proactiveness, 5U);
2275 return low ? wmark_low : min(wmark_low + 10, 100U);
2278 static bool should_proactive_compact_node(pg_data_t *pgdat)
2282 if (!sysctl_compaction_proactiveness || kswapd_is_running(pgdat))
2285 wmark_high = fragmentation_score_wmark(false);
2286 return fragmentation_score_node(pgdat) > wmark_high;
2289 static enum compact_result __compact_finished(struct compact_control *cc)
2292 const int migratetype = cc->migratetype;
2295 /* Compaction run completes if the migrate and free scanner meet */
2296 if (compact_scanners_met(cc)) {
2297 /* Let the next compaction start anew. */
2298 reset_cached_positions(cc->zone);
2301 * Mark that the PG_migrate_skip information should be cleared
2302 * by kswapd when it goes to sleep. kcompactd does not set the
2303 * flag itself as the decision to be clear should be directly
2304 * based on an allocation request.
2306 if (cc->direct_compaction)
2307 cc->zone->compact_blockskip_flush = true;
2310 return COMPACT_COMPLETE;
2312 return COMPACT_PARTIAL_SKIPPED;
2315 if (cc->proactive_compaction) {
2316 int score, wmark_low;
2319 pgdat = cc->zone->zone_pgdat;
2320 if (kswapd_is_running(pgdat))
2321 return COMPACT_PARTIAL_SKIPPED;
2323 score = fragmentation_score_zone(cc->zone);
2324 wmark_low = fragmentation_score_wmark(true);
2326 if (score > wmark_low)
2327 ret = COMPACT_CONTINUE;
2329 ret = COMPACT_SUCCESS;
2334 if (is_via_compact_memory(cc->order))
2335 return COMPACT_CONTINUE;
2338 * Always finish scanning a pageblock to reduce the possibility of
2339 * fallbacks in the future. This is particularly important when
2340 * migration source is unmovable/reclaimable but it's not worth
2343 if (!pageblock_aligned(cc->migrate_pfn))
2344 return COMPACT_CONTINUE;
2346 /* Direct compactor: Is a suitable page free? */
2347 ret = COMPACT_NO_SUITABLE_PAGE;
2348 for (order = cc->order; order < NR_PAGE_ORDERS; order++) {
2349 struct free_area *area = &cc->zone->free_area[order];
2352 /* Job done if page is free of the right migratetype */
2353 if (!free_area_empty(area, migratetype))
2354 return COMPACT_SUCCESS;
2357 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
2358 if (migratetype == MIGRATE_MOVABLE &&
2359 !free_area_empty(area, MIGRATE_CMA))
2360 return COMPACT_SUCCESS;
2363 * Job done if allocation would steal freepages from
2364 * other migratetype buddy lists.
2366 if (find_suitable_fallback(area, order, migratetype,
2367 true, &can_steal) != -1)
2369 * Movable pages are OK in any pageblock. If we are
2370 * stealing for a non-movable allocation, make sure
2371 * we finish compacting the current pageblock first
2372 * (which is assured by the above migrate_pfn align
2373 * check) so it is as free as possible and we won't
2374 * have to steal another one soon.
2376 return COMPACT_SUCCESS;
2380 if (cc->contended || fatal_signal_pending(current))
2381 ret = COMPACT_CONTENDED;
2386 static enum compact_result compact_finished(struct compact_control *cc)
2390 ret = __compact_finished(cc);
2391 trace_mm_compaction_finished(cc->zone, cc->order, ret);
2392 if (ret == COMPACT_NO_SUITABLE_PAGE)
2393 ret = COMPACT_CONTINUE;
2398 static bool __compaction_suitable(struct zone *zone, int order,
2399 int highest_zoneidx,
2400 unsigned long wmark_target)
2402 unsigned long watermark;
2404 * Watermarks for order-0 must be met for compaction to be able to
2405 * isolate free pages for migration targets. This means that the
2406 * watermark and alloc_flags have to match, or be more pessimistic than
2407 * the check in __isolate_free_page(). We don't use the direct
2408 * compactor's alloc_flags, as they are not relevant for freepage
2409 * isolation. We however do use the direct compactor's highest_zoneidx
2410 * to skip over zones where lowmem reserves would prevent allocation
2411 * even if compaction succeeds.
2412 * For costly orders, we require low watermark instead of min for
2413 * compaction to proceed to increase its chances.
2414 * ALLOC_CMA is used, as pages in CMA pageblocks are considered
2415 * suitable migration targets
2417 watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
2418 low_wmark_pages(zone) : min_wmark_pages(zone);
2419 watermark += compact_gap(order);
2420 return __zone_watermark_ok(zone, 0, watermark, highest_zoneidx,
2421 ALLOC_CMA, wmark_target);
2425 * compaction_suitable: Is this suitable to run compaction on this zone now?
2427 bool compaction_suitable(struct zone *zone, int order, int highest_zoneidx)
2429 enum compact_result compact_result;
2432 suitable = __compaction_suitable(zone, order, highest_zoneidx,
2433 zone_page_state(zone, NR_FREE_PAGES));
2435 * fragmentation index determines if allocation failures are due to
2436 * low memory or external fragmentation
2438 * index of -1000 would imply allocations might succeed depending on
2439 * watermarks, but we already failed the high-order watermark check
2440 * index towards 0 implies failure is due to lack of memory
2441 * index towards 1000 implies failure is due to fragmentation
2443 * Only compact if a failure would be due to fragmentation. Also
2444 * ignore fragindex for non-costly orders where the alternative to
2445 * a successful reclaim/compaction is OOM. Fragindex and the
2446 * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
2447 * excessive compaction for costly orders, but it should not be at the
2448 * expense of system stability.
2451 compact_result = COMPACT_CONTINUE;
2452 if (order > PAGE_ALLOC_COSTLY_ORDER) {
2453 int fragindex = fragmentation_index(zone, order);
2455 if (fragindex >= 0 &&
2456 fragindex <= sysctl_extfrag_threshold) {
2458 compact_result = COMPACT_NOT_SUITABLE_ZONE;
2462 compact_result = COMPACT_SKIPPED;
2465 trace_mm_compaction_suitable(zone, order, compact_result);
2470 bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
2477 * Make sure at least one zone would pass __compaction_suitable if we continue
2478 * retrying the reclaim.
2480 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2481 ac->highest_zoneidx, ac->nodemask) {
2482 unsigned long available;
2485 * Do not consider all the reclaimable memory because we do not
2486 * want to trash just for a single high order allocation which
2487 * is even not guaranteed to appear even if __compaction_suitable
2488 * is happy about the watermark check.
2490 available = zone_reclaimable_pages(zone) / order;
2491 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
2492 if (__compaction_suitable(zone, order, ac->highest_zoneidx,
2501 * Should we do compaction for target allocation order.
2502 * Return COMPACT_SUCCESS if allocation for target order can be already
2504 * Return COMPACT_SKIPPED if compaction for target order is likely to fail
2505 * Return COMPACT_CONTINUE if compaction for target order should be ran
2507 static enum compact_result
2508 compaction_suit_allocation_order(struct zone *zone, unsigned int order,
2509 int highest_zoneidx, unsigned int alloc_flags)
2511 unsigned long watermark;
2513 watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
2514 if (zone_watermark_ok(zone, order, watermark, highest_zoneidx,
2516 return COMPACT_SUCCESS;
2518 if (!compaction_suitable(zone, order, highest_zoneidx))
2519 return COMPACT_SKIPPED;
2521 return COMPACT_CONTINUE;
2524 static enum compact_result
2525 compact_zone(struct compact_control *cc, struct capture_control *capc)
2527 enum compact_result ret;
2528 unsigned long start_pfn = cc->zone->zone_start_pfn;
2529 unsigned long end_pfn = zone_end_pfn(cc->zone);
2530 unsigned long last_migrated_pfn;
2531 const bool sync = cc->mode != MIGRATE_ASYNC;
2533 unsigned int nr_succeeded = 0, nr_migratepages;
2537 * These counters track activities during zone compaction. Initialize
2538 * them before compacting a new zone.
2540 cc->total_migrate_scanned = 0;
2541 cc->total_free_scanned = 0;
2542 cc->nr_migratepages = 0;
2543 cc->nr_freepages = 0;
2544 for (order = 0; order < NR_PAGE_ORDERS; order++)
2545 INIT_LIST_HEAD(&cc->freepages[order]);
2546 INIT_LIST_HEAD(&cc->migratepages);
2548 cc->migratetype = gfp_migratetype(cc->gfp_mask);
2550 if (!is_via_compact_memory(cc->order)) {
2551 ret = compaction_suit_allocation_order(cc->zone, cc->order,
2552 cc->highest_zoneidx,
2554 if (ret != COMPACT_CONTINUE)
2559 * Clear pageblock skip if there were failures recently and compaction
2560 * is about to be retried after being deferred.
2562 if (compaction_restarting(cc->zone, cc->order))
2563 __reset_isolation_suitable(cc->zone);
2566 * Setup to move all movable pages to the end of the zone. Used cached
2567 * information on where the scanners should start (unless we explicitly
2568 * want to compact the whole zone), but check that it is initialised
2569 * by ensuring the values are within zone boundaries.
2571 cc->fast_start_pfn = 0;
2572 if (cc->whole_zone) {
2573 cc->migrate_pfn = start_pfn;
2574 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2576 cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync];
2577 cc->free_pfn = cc->zone->compact_cached_free_pfn;
2578 if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
2579 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2580 cc->zone->compact_cached_free_pfn = cc->free_pfn;
2582 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
2583 cc->migrate_pfn = start_pfn;
2584 cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
2585 cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
2588 if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn)
2589 cc->whole_zone = true;
2592 last_migrated_pfn = 0;
2595 * Migrate has separate cached PFNs for ASYNC and SYNC* migration on
2596 * the basis that some migrations will fail in ASYNC mode. However,
2597 * if the cached PFNs match and pageblocks are skipped due to having
2598 * no isolation candidates, then the sync state does not matter.
2599 * Until a pageblock with isolation candidates is found, keep the
2600 * cached PFNs in sync to avoid revisiting the same blocks.
2602 update_cached = !sync &&
2603 cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1];
2605 trace_mm_compaction_begin(cc, start_pfn, end_pfn, sync);
2607 /* lru_add_drain_all could be expensive with involving other CPUs */
2610 while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) {
2612 unsigned long iteration_start_pfn = cc->migrate_pfn;
2615 * Avoid multiple rescans of the same pageblock which can
2616 * happen if a page cannot be isolated (dirty/writeback in
2617 * async mode) or if the migrated pages are being allocated
2618 * before the pageblock is cleared. The first rescan will
2619 * capture the entire pageblock for migration. If it fails,
2620 * it'll be marked skip and scanning will proceed as normal.
2622 cc->finish_pageblock = false;
2623 if (pageblock_start_pfn(last_migrated_pfn) ==
2624 pageblock_start_pfn(iteration_start_pfn)) {
2625 cc->finish_pageblock = true;
2629 switch (isolate_migratepages(cc)) {
2631 ret = COMPACT_CONTENDED;
2632 putback_movable_pages(&cc->migratepages);
2633 cc->nr_migratepages = 0;
2636 if (update_cached) {
2637 cc->zone->compact_cached_migrate_pfn[1] =
2638 cc->zone->compact_cached_migrate_pfn[0];
2642 * We haven't isolated and migrated anything, but
2643 * there might still be unflushed migrations from
2644 * previous cc->order aligned block.
2647 case ISOLATE_SUCCESS:
2648 update_cached = false;
2649 last_migrated_pfn = max(cc->zone->zone_start_pfn,
2650 pageblock_start_pfn(cc->migrate_pfn - 1));
2654 * Record the number of pages to migrate since the
2655 * compaction_alloc/free() will update cc->nr_migratepages
2658 nr_migratepages = cc->nr_migratepages;
2659 err = migrate_pages(&cc->migratepages, compaction_alloc,
2660 compaction_free, (unsigned long)cc, cc->mode,
2661 MR_COMPACTION, &nr_succeeded);
2663 trace_mm_compaction_migratepages(nr_migratepages, nr_succeeded);
2665 /* All pages were either migrated or will be released */
2666 cc->nr_migratepages = 0;
2668 putback_movable_pages(&cc->migratepages);
2670 * migrate_pages() may return -ENOMEM when scanners meet
2671 * and we want compact_finished() to detect it
2673 if (err == -ENOMEM && !compact_scanners_met(cc)) {
2674 ret = COMPACT_CONTENDED;
2678 * If an ASYNC or SYNC_LIGHT fails to migrate a page
2679 * within the pageblock_order-aligned block and
2680 * fast_find_migrateblock may be used then scan the
2681 * remainder of the pageblock. This will mark the
2682 * pageblock "skip" to avoid rescanning in the near
2683 * future. This will isolate more pages than necessary
2684 * for the request but avoid loops due to
2685 * fast_find_migrateblock revisiting blocks that were
2686 * recently partially scanned.
2688 if (!pageblock_aligned(cc->migrate_pfn) &&
2689 !cc->ignore_skip_hint && !cc->finish_pageblock &&
2690 (cc->mode < MIGRATE_SYNC)) {
2691 cc->finish_pageblock = true;
2694 * Draining pcplists does not help THP if
2695 * any page failed to migrate. Even after
2696 * drain, the pageblock will not be free.
2698 if (cc->order == COMPACTION_HPAGE_ORDER)
2699 last_migrated_pfn = 0;
2705 /* Stop if a page has been captured */
2706 if (capc && capc->page) {
2707 ret = COMPACT_SUCCESS;
2713 * Has the migration scanner moved away from the previous
2714 * cc->order aligned block where we migrated from? If yes,
2715 * flush the pages that were freed, so that they can merge and
2716 * compact_finished() can detect immediately if allocation
2719 if (cc->order > 0 && last_migrated_pfn) {
2720 unsigned long current_block_start =
2721 block_start_pfn(cc->migrate_pfn, cc->order);
2723 if (last_migrated_pfn < current_block_start) {
2724 lru_add_drain_cpu_zone(cc->zone);
2725 /* No more flushing until we migrate again */
2726 last_migrated_pfn = 0;
2733 * Release free pages and update where the free scanner should restart,
2734 * so we don't leave any returned pages behind in the next attempt.
2736 if (cc->nr_freepages > 0) {
2737 unsigned long free_pfn = release_free_list(cc->freepages);
2739 cc->nr_freepages = 0;
2740 VM_BUG_ON(free_pfn == 0);
2741 /* The cached pfn is always the first in a pageblock */
2742 free_pfn = pageblock_start_pfn(free_pfn);
2744 * Only go back, not forward. The cached pfn might have been
2745 * already reset to zone end in compact_finished()
2747 if (free_pfn > cc->zone->compact_cached_free_pfn)
2748 cc->zone->compact_cached_free_pfn = free_pfn;
2751 count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned);
2752 count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned);
2754 trace_mm_compaction_end(cc, start_pfn, end_pfn, sync, ret);
2756 VM_BUG_ON(!list_empty(&cc->migratepages));
2761 static enum compact_result compact_zone_order(struct zone *zone, int order,
2762 gfp_t gfp_mask, enum compact_priority prio,
2763 unsigned int alloc_flags, int highest_zoneidx,
2764 struct page **capture)
2766 enum compact_result ret;
2767 struct compact_control cc = {
2769 .search_order = order,
2770 .gfp_mask = gfp_mask,
2772 .mode = (prio == COMPACT_PRIO_ASYNC) ?
2773 MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT,
2774 .alloc_flags = alloc_flags,
2775 .highest_zoneidx = highest_zoneidx,
2776 .direct_compaction = true,
2777 .whole_zone = (prio == MIN_COMPACT_PRIORITY),
2778 .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
2779 .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
2781 struct capture_control capc = {
2787 * Make sure the structs are really initialized before we expose the
2788 * capture control, in case we are interrupted and the interrupt handler
2792 WRITE_ONCE(current->capture_control, &capc);
2794 ret = compact_zone(&cc, &capc);
2797 * Make sure we hide capture control first before we read the captured
2798 * page pointer, otherwise an interrupt could free and capture a page
2799 * and we would leak it.
2801 WRITE_ONCE(current->capture_control, NULL);
2802 *capture = READ_ONCE(capc.page);
2804 * Technically, it is also possible that compaction is skipped but
2805 * the page is still captured out of luck(IRQ came and freed the page).
2806 * Returning COMPACT_SUCCESS in such cases helps in properly accounting
2807 * the COMPACT[STALL|FAIL] when compaction is skipped.
2810 ret = COMPACT_SUCCESS;
2816 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
2817 * @gfp_mask: The GFP mask of the current allocation
2818 * @order: The order of the current allocation
2819 * @alloc_flags: The allocation flags of the current allocation
2820 * @ac: The context of current allocation
2821 * @prio: Determines how hard direct compaction should try to succeed
2822 * @capture: Pointer to free page created by compaction will be stored here
2824 * This is the main entry point for direct page compaction.
2826 enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
2827 unsigned int alloc_flags, const struct alloc_context *ac,
2828 enum compact_priority prio, struct page **capture)
2832 enum compact_result rc = COMPACT_SKIPPED;
2834 if (!gfp_compaction_allowed(gfp_mask))
2835 return COMPACT_SKIPPED;
2837 trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
2839 /* Compact each zone in the list */
2840 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2841 ac->highest_zoneidx, ac->nodemask) {
2842 enum compact_result status;
2844 if (prio > MIN_COMPACT_PRIORITY
2845 && compaction_deferred(zone, order)) {
2846 rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
2850 status = compact_zone_order(zone, order, gfp_mask, prio,
2851 alloc_flags, ac->highest_zoneidx, capture);
2852 rc = max(status, rc);
2854 /* The allocation should succeed, stop compacting */
2855 if (status == COMPACT_SUCCESS) {
2857 * We think the allocation will succeed in this zone,
2858 * but it is not certain, hence the false. The caller
2859 * will repeat this with true if allocation indeed
2860 * succeeds in this zone.
2862 compaction_defer_reset(zone, order, false);
2867 if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
2868 status == COMPACT_PARTIAL_SKIPPED))
2870 * We think that allocation won't succeed in this zone
2871 * so we defer compaction there. If it ends up
2872 * succeeding after all, it will be reset.
2874 defer_compaction(zone, order);
2877 * We might have stopped compacting due to need_resched() in
2878 * async compaction, or due to a fatal signal detected. In that
2879 * case do not try further zones
2881 if ((prio == COMPACT_PRIO_ASYNC && need_resched())
2882 || fatal_signal_pending(current))
2890 * compact_node() - compact all zones within a node
2891 * @pgdat: The node page data
2892 * @proactive: Whether the compaction is proactive
2894 * For proactive compaction, compact till each zone's fragmentation score
2895 * reaches within proactive compaction thresholds (as determined by the
2896 * proactiveness tunable), it is possible that the function returns before
2897 * reaching score targets due to various back-off conditions, such as,
2898 * contention on per-node or per-zone locks.
2900 static int compact_node(pg_data_t *pgdat, bool proactive)
2904 struct compact_control cc = {
2906 .mode = proactive ? MIGRATE_SYNC_LIGHT : MIGRATE_SYNC,
2907 .ignore_skip_hint = true,
2909 .gfp_mask = GFP_KERNEL,
2910 .proactive_compaction = proactive,
2913 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2914 zone = &pgdat->node_zones[zoneid];
2915 if (!populated_zone(zone))
2918 if (fatal_signal_pending(current))
2923 compact_zone(&cc, NULL);
2926 count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
2927 cc.total_migrate_scanned);
2928 count_compact_events(KCOMPACTD_FREE_SCANNED,
2929 cc.total_free_scanned);
2936 /* Compact all zones of all nodes in the system */
2937 static int compact_nodes(void)
2941 /* Flush pending updates to the LRU lists */
2942 lru_add_drain_all();
2944 for_each_online_node(nid) {
2945 ret = compact_node(NODE_DATA(nid), false);
2953 static int compaction_proactiveness_sysctl_handler(struct ctl_table *table, int write,
2954 void *buffer, size_t *length, loff_t *ppos)
2958 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
2962 if (write && sysctl_compaction_proactiveness) {
2963 for_each_online_node(nid) {
2964 pg_data_t *pgdat = NODE_DATA(nid);
2966 if (pgdat->proactive_compact_trigger)
2969 pgdat->proactive_compact_trigger = true;
2970 trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, -1,
2971 pgdat->nr_zones - 1);
2972 wake_up_interruptible(&pgdat->kcompactd_wait);
2980 * This is the entry point for compacting all nodes via
2981 * /proc/sys/vm/compact_memory
2983 static int sysctl_compaction_handler(struct ctl_table *table, int write,
2984 void *buffer, size_t *length, loff_t *ppos)
2988 ret = proc_dointvec(table, write, buffer, length, ppos);
2992 if (sysctl_compact_memory != 1)
2996 ret = compact_nodes();
3001 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
3002 static ssize_t compact_store(struct device *dev,
3003 struct device_attribute *attr,
3004 const char *buf, size_t count)
3008 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
3009 /* Flush pending updates to the LRU lists */
3010 lru_add_drain_all();
3012 compact_node(NODE_DATA(nid), false);
3017 static DEVICE_ATTR_WO(compact);
3019 int compaction_register_node(struct node *node)
3021 return device_create_file(&node->dev, &dev_attr_compact);
3024 void compaction_unregister_node(struct node *node)
3026 device_remove_file(&node->dev, &dev_attr_compact);
3028 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
3030 static inline bool kcompactd_work_requested(pg_data_t *pgdat)
3032 return pgdat->kcompactd_max_order > 0 || kthread_should_stop() ||
3033 pgdat->proactive_compact_trigger;
3036 static bool kcompactd_node_suitable(pg_data_t *pgdat)
3040 enum zone_type highest_zoneidx = pgdat->kcompactd_highest_zoneidx;
3041 enum compact_result ret;
3043 for (zoneid = 0; zoneid <= highest_zoneidx; zoneid++) {
3044 zone = &pgdat->node_zones[zoneid];
3046 if (!populated_zone(zone))
3049 ret = compaction_suit_allocation_order(zone,
3050 pgdat->kcompactd_max_order,
3051 highest_zoneidx, ALLOC_WMARK_MIN);
3052 if (ret == COMPACT_CONTINUE)
3059 static void kcompactd_do_work(pg_data_t *pgdat)
3062 * With no special task, compact all zones so that a page of requested
3063 * order is allocatable.
3067 struct compact_control cc = {
3068 .order = pgdat->kcompactd_max_order,
3069 .search_order = pgdat->kcompactd_max_order,
3070 .highest_zoneidx = pgdat->kcompactd_highest_zoneidx,
3071 .mode = MIGRATE_SYNC_LIGHT,
3072 .ignore_skip_hint = false,
3073 .gfp_mask = GFP_KERNEL,
3075 enum compact_result ret;
3077 trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
3078 cc.highest_zoneidx);
3079 count_compact_event(KCOMPACTD_WAKE);
3081 for (zoneid = 0; zoneid <= cc.highest_zoneidx; zoneid++) {
3084 zone = &pgdat->node_zones[zoneid];
3085 if (!populated_zone(zone))
3088 if (compaction_deferred(zone, cc.order))
3091 ret = compaction_suit_allocation_order(zone,
3092 cc.order, zoneid, ALLOC_WMARK_MIN);
3093 if (ret != COMPACT_CONTINUE)
3096 if (kthread_should_stop())
3100 status = compact_zone(&cc, NULL);
3102 if (status == COMPACT_SUCCESS) {
3103 compaction_defer_reset(zone, cc.order, false);
3104 } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
3106 * Buddy pages may become stranded on pcps that could
3107 * otherwise coalesce on the zone's free area for
3108 * order >= cc.order. This is ratelimited by the
3109 * upcoming deferral.
3111 drain_all_pages(zone);
3114 * We use sync migration mode here, so we defer like
3115 * sync direct compaction does.
3117 defer_compaction(zone, cc.order);
3120 count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
3121 cc.total_migrate_scanned);
3122 count_compact_events(KCOMPACTD_FREE_SCANNED,
3123 cc.total_free_scanned);
3127 * Regardless of success, we are done until woken up next. But remember
3128 * the requested order/highest_zoneidx in case it was higher/tighter
3129 * than our current ones
3131 if (pgdat->kcompactd_max_order <= cc.order)
3132 pgdat->kcompactd_max_order = 0;
3133 if (pgdat->kcompactd_highest_zoneidx >= cc.highest_zoneidx)
3134 pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
3137 void wakeup_kcompactd(pg_data_t *pgdat, int order, int highest_zoneidx)
3142 if (pgdat->kcompactd_max_order < order)
3143 pgdat->kcompactd_max_order = order;
3145 if (pgdat->kcompactd_highest_zoneidx > highest_zoneidx)
3146 pgdat->kcompactd_highest_zoneidx = highest_zoneidx;
3149 * Pairs with implicit barrier in wait_event_freezable()
3150 * such that wakeups are not missed.
3152 if (!wq_has_sleeper(&pgdat->kcompactd_wait))
3155 if (!kcompactd_node_suitable(pgdat))
3158 trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
3160 wake_up_interruptible(&pgdat->kcompactd_wait);
3164 * The background compaction daemon, started as a kernel thread
3165 * from the init process.
3167 static int kcompactd(void *p)
3169 pg_data_t *pgdat = (pg_data_t *)p;
3170 struct task_struct *tsk = current;
3171 long default_timeout = msecs_to_jiffies(HPAGE_FRAG_CHECK_INTERVAL_MSEC);
3172 long timeout = default_timeout;
3174 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3176 if (!cpumask_empty(cpumask))
3177 set_cpus_allowed_ptr(tsk, cpumask);
3181 pgdat->kcompactd_max_order = 0;
3182 pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
3184 while (!kthread_should_stop()) {
3185 unsigned long pflags;
3188 * Avoid the unnecessary wakeup for proactive compaction
3189 * when it is disabled.
3191 if (!sysctl_compaction_proactiveness)
3192 timeout = MAX_SCHEDULE_TIMEOUT;
3193 trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
3194 if (wait_event_freezable_timeout(pgdat->kcompactd_wait,
3195 kcompactd_work_requested(pgdat), timeout) &&
3196 !pgdat->proactive_compact_trigger) {
3198 psi_memstall_enter(&pflags);
3199 kcompactd_do_work(pgdat);
3200 psi_memstall_leave(&pflags);
3202 * Reset the timeout value. The defer timeout from
3203 * proactive compaction is lost here but that is fine
3204 * as the condition of the zone changing substantionally
3205 * then carrying on with the previous defer interval is
3208 timeout = default_timeout;
3213 * Start the proactive work with default timeout. Based
3214 * on the fragmentation score, this timeout is updated.
3216 timeout = default_timeout;
3217 if (should_proactive_compact_node(pgdat)) {
3218 unsigned int prev_score, score;
3220 prev_score = fragmentation_score_node(pgdat);
3221 compact_node(pgdat, true);
3222 score = fragmentation_score_node(pgdat);
3224 * Defer proactive compaction if the fragmentation
3225 * score did not go down i.e. no progress made.
3227 if (unlikely(score >= prev_score))
3229 default_timeout << COMPACT_MAX_DEFER_SHIFT;
3231 if (unlikely(pgdat->proactive_compact_trigger))
3232 pgdat->proactive_compact_trigger = false;
3239 * This kcompactd start function will be called by init and node-hot-add.
3240 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
3242 void __meminit kcompactd_run(int nid)
3244 pg_data_t *pgdat = NODE_DATA(nid);
3246 if (pgdat->kcompactd)
3249 pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
3250 if (IS_ERR(pgdat->kcompactd)) {
3251 pr_err("Failed to start kcompactd on node %d\n", nid);
3252 pgdat->kcompactd = NULL;
3257 * Called by memory hotplug when all memory in a node is offlined. Caller must
3258 * be holding mem_hotplug_begin/done().
3260 void __meminit kcompactd_stop(int nid)
3262 struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
3265 kthread_stop(kcompactd);
3266 NODE_DATA(nid)->kcompactd = NULL;
3271 * It's optimal to keep kcompactd on the same CPUs as their memory, but
3272 * not required for correctness. So if the last cpu in a node goes
3273 * away, we get changed to run anywhere: as the first one comes back,
3274 * restore their cpu bindings.
3276 static int kcompactd_cpu_online(unsigned int cpu)
3280 for_each_node_state(nid, N_MEMORY) {
3281 pg_data_t *pgdat = NODE_DATA(nid);
3282 const struct cpumask *mask;
3284 mask = cpumask_of_node(pgdat->node_id);
3286 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3287 /* One of our CPUs online: restore mask */
3288 if (pgdat->kcompactd)
3289 set_cpus_allowed_ptr(pgdat->kcompactd, mask);
3294 static int proc_dointvec_minmax_warn_RT_change(struct ctl_table *table,
3295 int write, void *buffer, size_t *lenp, loff_t *ppos)
3299 if (!IS_ENABLED(CONFIG_PREEMPT_RT) || !write)
3300 return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
3302 old = *(int *)table->data;
3303 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
3306 if (old != *(int *)table->data)
3307 pr_warn_once("sysctl attribute %s changed by %s[%d]\n",
3308 table->procname, current->comm,
3309 task_pid_nr(current));
3313 static struct ctl_table vm_compaction[] = {
3315 .procname = "compact_memory",
3316 .data = &sysctl_compact_memory,
3317 .maxlen = sizeof(int),
3319 .proc_handler = sysctl_compaction_handler,
3322 .procname = "compaction_proactiveness",
3323 .data = &sysctl_compaction_proactiveness,
3324 .maxlen = sizeof(sysctl_compaction_proactiveness),
3326 .proc_handler = compaction_proactiveness_sysctl_handler,
3327 .extra1 = SYSCTL_ZERO,
3328 .extra2 = SYSCTL_ONE_HUNDRED,
3331 .procname = "extfrag_threshold",
3332 .data = &sysctl_extfrag_threshold,
3333 .maxlen = sizeof(int),
3335 .proc_handler = proc_dointvec_minmax,
3336 .extra1 = SYSCTL_ZERO,
3337 .extra2 = SYSCTL_ONE_THOUSAND,
3340 .procname = "compact_unevictable_allowed",
3341 .data = &sysctl_compact_unevictable_allowed,
3342 .maxlen = sizeof(int),
3344 .proc_handler = proc_dointvec_minmax_warn_RT_change,
3345 .extra1 = SYSCTL_ZERO,
3346 .extra2 = SYSCTL_ONE,
3351 static int __init kcompactd_init(void)
3356 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3357 "mm/compaction:online",
3358 kcompactd_cpu_online, NULL);
3360 pr_err("kcompactd: failed to register hotplug callbacks.\n");
3364 for_each_node_state(nid, N_MEMORY)
3366 register_sysctl_init("vm", vm_compaction);
3369 subsys_initcall(kcompactd_init)
3371 #endif /* CONFIG_COMPACTION */