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);
44 #define count_compact_event(item) do { } while (0)
45 #define count_compact_events(item, delta) do { } while (0)
48 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
50 #define CREATE_TRACE_POINTS
51 #include <trace/events/compaction.h>
53 #define block_start_pfn(pfn, order) round_down(pfn, 1UL << (order))
54 #define block_end_pfn(pfn, order) ALIGN((pfn) + 1, 1UL << (order))
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.
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
66 #define COMPACTION_HPAGE_ORDER (PMD_SHIFT - PAGE_SHIFT)
69 static unsigned long release_freepages(struct list_head *freelist)
71 struct page *page, *next;
72 unsigned long high_pfn = 0;
74 list_for_each_entry_safe(page, next, freelist, lru) {
75 unsigned long pfn = page_to_pfn(page);
85 static void split_map_pages(struct list_head *list)
87 unsigned int i, order, nr_pages;
88 struct page *page, *next;
91 list_for_each_entry_safe(page, next, list, lru) {
94 order = page_private(page);
95 nr_pages = 1 << order;
97 post_alloc_hook(page, order, __GFP_MOVABLE);
99 split_page(page, order);
101 for (i = 0; i < nr_pages; i++) {
102 list_add(&page->lru, &tmp_list);
107 list_splice(&tmp_list, list);
110 #ifdef CONFIG_COMPACTION
111 bool PageMovable(struct page *page)
113 const struct movable_operations *mops;
115 VM_BUG_ON_PAGE(!PageLocked(page), page);
116 if (!__PageMovable(page))
119 mops = page_movable_ops(page);
126 void __SetPageMovable(struct page *page, const struct movable_operations *mops)
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);
132 EXPORT_SYMBOL(__SetPageMovable);
134 void __ClearPageMovable(struct page *page)
136 VM_BUG_ON_PAGE(!PageMovable(page), page);
138 * This page still has the type of a movable page, but it's
139 * actually not movable any more.
141 page->mapping = (void *)PAGE_MAPPING_MOVABLE;
143 EXPORT_SYMBOL(__ClearPageMovable);
145 /* Do not skip compaction more than 64 times */
146 #define COMPACT_MAX_DEFER_SHIFT 6
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
153 static void defer_compaction(struct zone *zone, int order)
155 zone->compact_considered = 0;
156 zone->compact_defer_shift++;
158 if (order < zone->compact_order_failed)
159 zone->compact_order_failed = order;
161 if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
162 zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
164 trace_mm_compaction_defer_compaction(zone, order);
167 /* Returns true if compaction should be skipped this time */
168 static bool compaction_deferred(struct zone *zone, int order)
170 unsigned long defer_limit = 1UL << zone->compact_defer_shift;
172 if (order < zone->compact_order_failed)
175 /* Avoid possible overflow */
176 if (++zone->compact_considered >= defer_limit) {
177 zone->compact_considered = defer_limit;
181 trace_mm_compaction_deferred(zone, order);
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.
191 void compaction_defer_reset(struct zone *zone, int order,
195 zone->compact_considered = 0;
196 zone->compact_defer_shift = 0;
198 if (order >= zone->compact_order_failed)
199 zone->compact_order_failed = order + 1;
201 trace_mm_compaction_defer_reset(zone, order);
204 /* Returns true if restarting compaction after many failures */
205 static bool compaction_restarting(struct zone *zone, int order)
207 if (order < zone->compact_order_failed)
210 return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
211 zone->compact_considered >= 1UL << zone->compact_defer_shift;
214 /* Returns true if the pageblock should be scanned for pages to isolate. */
215 static inline bool isolation_suitable(struct compact_control *cc,
218 if (cc->ignore_skip_hint)
221 return !get_pageblock_skip(page);
224 static void reset_cached_positions(struct zone *zone)
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);
232 #ifdef CONFIG_SPARSEMEM
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.
238 static unsigned long skip_offline_sections(unsigned long start_pfn)
240 unsigned long start_nr = pfn_to_section_nr(start_pfn);
242 if (online_section_nr(start_nr))
245 while (++start_nr <= __highest_present_section_nr) {
246 if (online_section_nr(start_nr))
247 return section_nr_to_pfn(start_nr);
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.
258 static unsigned long skip_offline_sections_reverse(unsigned long start_pfn)
260 unsigned long start_nr = pfn_to_section_nr(start_pfn);
262 if (!start_nr || online_section_nr(start_nr))
265 while (start_nr-- > 0) {
266 if (online_section_nr(start_nr))
267 return section_nr_to_pfn(start_nr) + PAGES_PER_SECTION;
273 static unsigned long skip_offline_sections(unsigned long start_pfn)
278 static unsigned long skip_offline_sections_reverse(unsigned long start_pfn)
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.
289 static bool pageblock_skip_persistent(struct page *page)
291 if (!PageCompound(page))
294 page = compound_head(page);
296 if (compound_order(page) >= pageblock_order)
303 __reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source,
306 struct page *page = pfn_to_online_page(pfn);
307 struct page *block_page;
308 struct page *end_page;
309 unsigned long block_pfn;
313 if (zone != page_zone(page))
315 if (pageblock_skip_persistent(page))
319 * If skip is already cleared do no further checking once the
320 * restart points have been set.
322 if (check_source && check_target && !get_pageblock_skip(page))
326 * If clearing skip for the target scanner, do not select a
327 * non-movable pageblock as the starting point.
329 if (!check_source && check_target &&
330 get_pageblock_migratetype(page) != MIGRATE_MOVABLE)
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);
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);
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.
355 if (check_source && PageLRU(page)) {
356 clear_pageblock_skip(page);
360 if (check_target && PageBuddy(page)) {
361 clear_pageblock_skip(page);
365 page += (1 << PAGE_ALLOC_COSTLY_ORDER);
366 } while (page <= end_page);
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
376 static void __reset_isolation_suitable(struct zone *zone)
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;
385 /* Only flush if a full compaction finished recently */
386 if (!zone->compact_blockskip_flush)
389 zone->compact_blockskip_flush = false;
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.
397 for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages,
398 free_pfn -= pageblock_nr_pages) {
401 /* Update the migrate PFN */
402 if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) &&
403 migrate_pfn < reset_migrate) {
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;
411 /* Update the free PFN */
412 if (__reset_isolation_pfn(zone, free_pfn, free_set, true) &&
413 free_pfn > reset_free) {
415 reset_free = free_pfn;
416 zone->compact_init_free_pfn = reset_free;
417 zone->compact_cached_free_pfn = reset_free;
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;
429 void reset_isolation_suitable(pg_data_t *pgdat)
433 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
434 struct zone *zone = &pgdat->node_zones[zoneid];
435 if (!populated_zone(zone))
438 __reset_isolation_suitable(zone);
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.
446 static bool test_and_set_skip(struct compact_control *cc, struct page *page)
450 /* Do not update if skip hint is being ignored */
451 if (cc->ignore_skip_hint)
454 skip = get_pageblock_skip(page);
455 if (!skip && !cc->no_set_skip_hint)
456 set_pageblock_skip(page);
461 static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
463 struct zone *zone = cc->zone;
465 /* Set for isolation rather than compaction */
466 if (cc->no_set_skip_hint)
469 pfn = pageblock_end_pfn(pfn);
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;
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().
483 static void update_pageblock_skip(struct compact_control *cc,
484 struct page *page, unsigned long pfn)
486 struct zone *zone = cc->zone;
488 if (cc->no_set_skip_hint)
491 set_pageblock_skip(page);
493 if (pfn < zone->compact_cached_free_pfn)
494 zone->compact_cached_free_pfn = pfn;
497 static inline bool isolation_suitable(struct compact_control *cc,
503 static inline bool pageblock_skip_persistent(struct page *page)
508 static inline void update_pageblock_skip(struct compact_control *cc,
509 struct page *page, unsigned long pfn)
513 static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
517 static bool test_and_set_skip(struct compact_control *cc, struct page *page)
521 #endif /* CONFIG_COMPACTION */
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.
530 * Always returns true which makes it easier to track lock state in callers.
532 static bool compact_lock_irqsave(spinlock_t *lock, unsigned long *flags,
533 struct compact_control *cc)
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))
541 cc->contended = true;
544 spin_lock_irqsave(lock, *flags);
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.
557 * Returns true if compaction should abort due to fatal signal pending.
558 * Returns false when compaction can continue.
560 static bool compact_unlock_should_abort(spinlock_t *lock,
561 unsigned long flags, bool *locked, struct compact_control *cc)
564 spin_unlock_irqrestore(lock, flags);
568 if (fatal_signal_pending(current)) {
569 cc->contended = true;
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).
583 static unsigned long isolate_freepages_block(struct compact_control *cc,
584 unsigned long *start_pfn,
585 unsigned long end_pfn,
586 struct list_head *freelist,
590 int nr_scanned = 0, total_isolated = 0;
592 unsigned long flags = 0;
594 unsigned long blockpfn = *start_pfn;
597 /* Strict mode is for isolation, speed is secondary */
601 page = pfn_to_page(blockpfn);
603 /* Isolate free pages. */
604 for (; blockpfn < end_pfn; blockpfn += stride, page += stride) {
608 * Periodically drop the lock (if held) regardless of its
609 * contention, to give chance to IRQs. Abort if fatal signal
612 if (!(blockpfn % COMPACT_CLUSTER_MAX)
613 && compact_unlock_should_abort(&cc->zone->lock, flags,
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.
625 if (PageCompound(page)) {
626 const unsigned int order = compound_order(page);
628 if (blockpfn + (1UL << order) <= end_pfn) {
629 blockpfn += (1UL << order) - 1;
630 page += (1UL << order) - 1;
631 nr_scanned += (1UL << order) - 1;
637 if (!PageBuddy(page))
640 /* If we already hold the lock, we can skip some rechecking. */
642 locked = compact_lock_irqsave(&cc->zone->lock,
645 /* Recheck this is a buddy page under lock */
646 if (!PageBuddy(page))
650 /* Found a free page, will break it into order-0 pages */
651 order = buddy_order(page);
652 isolated = __isolate_free_page(page, order);
655 set_page_private(page, order);
657 nr_scanned += isolated - 1;
658 total_isolated += isolated;
659 cc->nr_freepages += isolated;
660 list_add_tail(&page->lru, freelist);
662 if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
663 blockpfn += isolated;
666 /* Advance to the end of split page */
667 blockpfn += isolated - 1;
668 page += isolated - 1;
678 spin_unlock_irqrestore(&cc->zone->lock, flags);
681 * Be careful to not go outside of the pageblock.
683 if (unlikely(blockpfn > end_pfn))
686 trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
687 nr_scanned, total_isolated);
689 /* Record how far we have got within the block */
690 *start_pfn = blockpfn;
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.
697 if (strict && blockpfn < end_pfn)
700 cc->total_free_scanned += nr_scanned;
702 count_compact_events(COMPACTISOLATED, total_isolated);
703 return total_isolated;
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.
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.
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
721 isolate_freepages_range(struct compact_control *cc,
722 unsigned long start_pfn, unsigned long end_pfn)
724 unsigned long isolated, pfn, block_start_pfn, block_end_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);
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;
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.
744 if (pfn >= block_end_pfn) {
745 block_start_pfn = pageblock_start_pfn(pfn);
746 block_end_pfn = pageblock_end_pfn(pfn);
749 block_end_pfn = min(block_end_pfn, end_pfn);
751 if (!pageblock_pfn_to_page(block_start_pfn,
752 block_end_pfn, cc->zone))
755 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
756 block_end_pfn, &freelist, 0, true);
759 * In strict mode, isolate_freepages_block() returns 0 if
760 * there are any holes in the block (ie. invalid PFNs or
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).
773 /* __isolate_free_page() does not map the pages */
774 split_map_pages(&freelist);
777 /* Loop terminated early, cleanup. */
778 release_freepages(&freelist);
782 /* We don't use freelists for anything. */
786 /* Similar to reclaim, but different enough that they don't share logic */
787 static bool too_many_isolated(struct compact_control *cc)
789 pg_data_t *pgdat = cc->zone->zone_pgdat;
792 unsigned long active, inactive, isolated;
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);
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.
807 if (cc->gfp_mask & __GFP_FS) {
812 too_many = isolated > (inactive + active) / 2;
814 wake_throttle_isolated(pgdat);
820 * isolate_migratepages_block() - isolate all migrate-able pages within
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.
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.
833 * The pages are isolated on cc->migratepages list (not required to be empty),
834 * and cc->nr_migratepages is updated accordingly.
837 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
838 unsigned long end_pfn, isolate_mode_t mode)
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;
854 cc->migrate_pfn = low_pfn;
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
861 while (unlikely(too_many_isolated(cc))) {
862 /* stop isolation if there are still pages not migrated */
863 if (cc->nr_migratepages)
866 /* async migration should just abort */
867 if (cc->mode == MIGRATE_ASYNC)
870 reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED);
872 if (fatal_signal_pending(current))
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);
883 /* Time to isolate some pages for migration */
884 for (; low_pfn < end_pfn; low_pfn++) {
886 if (skip_on_failure && low_pfn >= next_skip_pfn) {
888 * We have isolated all migration candidates in the
889 * previous order-aligned block, and did not skip it due
890 * to failure. We should migrate the pages now and
891 * hopefully succeed compaction.
897 * We failed to isolate in the previous order-aligned
898 * block. Set the new boundary to the end of the
899 * current block. Note we can't simply increase
900 * next_skip_pfn by 1 << order, as low_pfn might have
901 * been incremented by a higher number due to skipping
902 * a compound or a high-order buddy page in the
903 * previous loop iteration.
905 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
909 * Periodically drop the lock (if held) regardless of its
910 * contention, to give chance to IRQs. Abort completely if
911 * a fatal signal is pending.
913 if (!(low_pfn % COMPACT_CLUSTER_MAX)) {
915 unlock_page_lruvec_irqrestore(locked, flags);
919 if (fatal_signal_pending(current)) {
920 cc->contended = true;
931 page = pfn_to_page(low_pfn);
934 * Check if the pageblock has already been marked skipped.
935 * Only the first PFN is checked as the caller isolates
936 * COMPACT_CLUSTER_MAX at a time so the second call must
937 * not falsely conclude that the block should be skipped.
939 if (!valid_page && (pageblock_aligned(low_pfn) ||
940 low_pfn == cc->zone->zone_start_pfn)) {
941 if (!isolation_suitable(cc, page)) {
949 if (PageHuge(page) && cc->alloc_contig) {
951 unlock_page_lruvec_irqrestore(locked, flags);
955 ret = isolate_or_dissolve_huge_page(page, &cc->migratepages);
958 * Fail isolation in case isolate_or_dissolve_huge_page()
959 * reports an error. In case of -ENOMEM, abort right away.
962 /* Do not report -EBUSY down the chain */
965 low_pfn += compound_nr(page) - 1;
966 nr_scanned += compound_nr(page) - 1;
970 if (PageHuge(page)) {
972 * Hugepage was successfully isolated and placed
973 * on the cc->migratepages list.
975 folio = page_folio(page);
976 low_pfn += folio_nr_pages(folio) - 1;
977 goto isolate_success_no_list;
981 * Ok, the hugepage was dissolved. Now these pages are
982 * Buddy and cannot be re-allocated because they are
983 * isolated. Fall-through as the check below handles
989 * Skip if free. We read page order here without zone lock
990 * which is generally unsafe, but the race window is small and
991 * the worst thing that can happen is that we skip some
992 * potential isolation targets.
994 if (PageBuddy(page)) {
995 unsigned long freepage_order = buddy_order_unsafe(page);
998 * Without lock, we cannot be sure that what we got is
999 * a valid page order. Consider only values in the
1000 * valid order range to prevent low_pfn overflow.
1002 if (freepage_order > 0 && freepage_order <= MAX_PAGE_ORDER) {
1003 low_pfn += (1UL << freepage_order) - 1;
1004 nr_scanned += (1UL << freepage_order) - 1;
1010 * Regardless of being on LRU, compound pages such as THP and
1011 * hugetlbfs are not to be compacted unless we are attempting
1012 * an allocation much larger than the huge page size (eg CMA).
1013 * We can potentially save a lot of iterations if we skip them
1014 * at once. The check is racy, but we can consider only valid
1015 * values and the only danger is skipping too much.
1017 if (PageCompound(page) && !cc->alloc_contig) {
1018 const unsigned int order = compound_order(page);
1020 if (likely(order <= MAX_PAGE_ORDER)) {
1021 low_pfn += (1UL << order) - 1;
1022 nr_scanned += (1UL << order) - 1;
1028 * Check may be lockless but that's ok as we recheck later.
1029 * It's possible to migrate LRU and non-lru movable pages.
1030 * Skip any other type of page
1032 if (!PageLRU(page)) {
1034 * __PageMovable can return false positive so we need
1035 * to verify it under page_lock.
1037 if (unlikely(__PageMovable(page)) &&
1038 !PageIsolated(page)) {
1040 unlock_page_lruvec_irqrestore(locked, flags);
1044 if (isolate_movable_page(page, mode)) {
1045 folio = page_folio(page);
1046 goto isolate_success;
1054 * Be careful not to clear PageLRU until after we're
1055 * sure the page is not being freed elsewhere -- the
1056 * page release code relies on it.
1058 folio = folio_get_nontail_page(page);
1059 if (unlikely(!folio))
1063 * Migration will fail if an anonymous page is pinned in memory,
1064 * so avoid taking lru_lock and isolating it unnecessarily in an
1065 * admittedly racy check.
1067 mapping = folio_mapping(folio);
1068 if (!mapping && (folio_ref_count(folio) - 1) > folio_mapcount(folio))
1069 goto isolate_fail_put;
1072 * Only allow to migrate anonymous pages in GFP_NOFS context
1073 * because those do not depend on fs locks.
1075 if (!(cc->gfp_mask & __GFP_FS) && mapping)
1076 goto isolate_fail_put;
1078 /* Only take pages on LRU: a check now makes later tests safe */
1079 if (!folio_test_lru(folio))
1080 goto isolate_fail_put;
1082 /* Compaction might skip unevictable pages but CMA takes them */
1083 if (!(mode & ISOLATE_UNEVICTABLE) && folio_test_unevictable(folio))
1084 goto isolate_fail_put;
1087 * To minimise LRU disruption, the caller can indicate with
1088 * ISOLATE_ASYNC_MIGRATE that it only wants to isolate pages
1089 * it will be able to migrate without blocking - clean pages
1090 * for the most part. PageWriteback would require blocking.
1092 if ((mode & ISOLATE_ASYNC_MIGRATE) && folio_test_writeback(folio))
1093 goto isolate_fail_put;
1095 if ((mode & ISOLATE_ASYNC_MIGRATE) && folio_test_dirty(folio)) {
1099 * Only folios without mappings or that have
1100 * a ->migrate_folio callback are possible to
1101 * migrate without blocking. However, we may
1102 * be racing with truncation, which can free
1103 * the mapping. Truncation holds the folio lock
1104 * until after the folio is removed from the page
1105 * cache so holding it ourselves is sufficient.
1107 if (!folio_trylock(folio))
1108 goto isolate_fail_put;
1110 mapping = folio_mapping(folio);
1111 migrate_dirty = !mapping ||
1112 mapping->a_ops->migrate_folio;
1113 folio_unlock(folio);
1115 goto isolate_fail_put;
1118 /* Try isolate the folio */
1119 if (!folio_test_clear_lru(folio))
1120 goto isolate_fail_put;
1122 lruvec = folio_lruvec(folio);
1124 /* If we already hold the lock, we can skip some rechecking */
1125 if (lruvec != locked) {
1127 unlock_page_lruvec_irqrestore(locked, flags);
1129 compact_lock_irqsave(&lruvec->lru_lock, &flags, cc);
1132 lruvec_memcg_debug(lruvec, folio);
1135 * Try get exclusive access under lock. If marked for
1136 * skip, the scan is aborted unless the current context
1137 * is a rescan to reach the end of the pageblock.
1139 if (!skip_updated && valid_page) {
1140 skip_updated = true;
1141 if (test_and_set_skip(cc, valid_page) &&
1142 !cc->finish_pageblock) {
1149 * folio become large since the non-locked check,
1152 if (unlikely(folio_test_large(folio) && !cc->alloc_contig)) {
1153 low_pfn += folio_nr_pages(folio) - 1;
1154 nr_scanned += folio_nr_pages(folio) - 1;
1155 folio_set_lru(folio);
1156 goto isolate_fail_put;
1160 /* The folio is taken off the LRU */
1161 if (folio_test_large(folio))
1162 low_pfn += folio_nr_pages(folio) - 1;
1164 /* Successfully isolated */
1165 lruvec_del_folio(lruvec, folio);
1166 node_stat_mod_folio(folio,
1167 NR_ISOLATED_ANON + folio_is_file_lru(folio),
1168 folio_nr_pages(folio));
1171 list_add(&folio->lru, &cc->migratepages);
1172 isolate_success_no_list:
1173 cc->nr_migratepages += folio_nr_pages(folio);
1174 nr_isolated += folio_nr_pages(folio);
1175 nr_scanned += folio_nr_pages(folio) - 1;
1178 * Avoid isolating too much unless this block is being
1179 * fully scanned (e.g. dirty/writeback pages, parallel allocation)
1180 * or a lock is contended. For contention, isolate quickly to
1181 * potentially remove one source of contention.
1183 if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX &&
1184 !cc->finish_pageblock && !cc->contended) {
1192 /* Avoid potential deadlock in freeing page under lru_lock */
1194 unlock_page_lruvec_irqrestore(locked, flags);
1200 if (!skip_on_failure && ret != -ENOMEM)
1204 * We have isolated some pages, but then failed. Release them
1205 * instead of migrating, as we cannot form the cc->order buddy
1210 unlock_page_lruvec_irqrestore(locked, flags);
1213 putback_movable_pages(&cc->migratepages);
1214 cc->nr_migratepages = 0;
1218 if (low_pfn < next_skip_pfn) {
1219 low_pfn = next_skip_pfn - 1;
1221 * The check near the loop beginning would have updated
1222 * next_skip_pfn too, but this is a bit simpler.
1224 next_skip_pfn += 1UL << cc->order;
1232 * The PageBuddy() check could have potentially brought us outside
1233 * the range to be scanned.
1235 if (unlikely(low_pfn > end_pfn))
1242 unlock_page_lruvec_irqrestore(locked, flags);
1244 folio_set_lru(folio);
1249 * Update the cached scanner pfn once the pageblock has been scanned.
1250 * Pages will either be migrated in which case there is no point
1251 * scanning in the near future or migration failed in which case the
1252 * failure reason may persist. The block is marked for skipping if
1253 * there were no pages isolated in the block or if the block is
1254 * rescanned twice in a row.
1256 if (low_pfn == end_pfn && (!nr_isolated || cc->finish_pageblock)) {
1257 if (!cc->no_set_skip_hint && valid_page && !skip_updated)
1258 set_pageblock_skip(valid_page);
1259 update_cached_migrate(cc, low_pfn);
1262 trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
1263 nr_scanned, nr_isolated);
1266 cc->total_migrate_scanned += nr_scanned;
1268 count_compact_events(COMPACTISOLATED, nr_isolated);
1270 cc->migrate_pfn = low_pfn;
1276 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
1277 * @cc: Compaction control structure.
1278 * @start_pfn: The first PFN to start isolating.
1279 * @end_pfn: The one-past-last PFN.
1281 * Returns -EAGAIN when contented, -EINTR in case of a signal pending, -ENOMEM
1282 * in case we could not allocate a page, or 0.
1285 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
1286 unsigned long end_pfn)
1288 unsigned long pfn, block_start_pfn, block_end_pfn;
1291 /* Scan block by block. First and last block may be incomplete */
1293 block_start_pfn = pageblock_start_pfn(pfn);
1294 if (block_start_pfn < cc->zone->zone_start_pfn)
1295 block_start_pfn = cc->zone->zone_start_pfn;
1296 block_end_pfn = pageblock_end_pfn(pfn);
1298 for (; pfn < end_pfn; pfn = block_end_pfn,
1299 block_start_pfn = block_end_pfn,
1300 block_end_pfn += pageblock_nr_pages) {
1302 block_end_pfn = min(block_end_pfn, end_pfn);
1304 if (!pageblock_pfn_to_page(block_start_pfn,
1305 block_end_pfn, cc->zone))
1308 ret = isolate_migratepages_block(cc, pfn, block_end_pfn,
1309 ISOLATE_UNEVICTABLE);
1314 if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX)
1321 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
1322 #ifdef CONFIG_COMPACTION
1324 static bool suitable_migration_source(struct compact_control *cc,
1329 if (pageblock_skip_persistent(page))
1332 if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction)
1335 block_mt = get_pageblock_migratetype(page);
1337 if (cc->migratetype == MIGRATE_MOVABLE)
1338 return is_migrate_movable(block_mt);
1340 return block_mt == cc->migratetype;
1343 /* Returns true if the page is within a block suitable for migration to */
1344 static bool suitable_migration_target(struct compact_control *cc,
1347 /* If the page is a large free page, then disallow migration */
1348 if (PageBuddy(page)) {
1350 * We are checking page_order without zone->lock taken. But
1351 * the only small danger is that we skip a potentially suitable
1352 * pageblock, so it's not worth to check order for valid range.
1354 if (buddy_order_unsafe(page) >= pageblock_order)
1358 if (cc->ignore_block_suitable)
1361 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1362 if (is_migrate_movable(get_pageblock_migratetype(page)))
1365 /* Otherwise skip the block */
1369 static inline unsigned int
1370 freelist_scan_limit(struct compact_control *cc)
1372 unsigned short shift = BITS_PER_LONG - 1;
1374 return (COMPACT_CLUSTER_MAX >> min(shift, cc->fast_search_fail)) + 1;
1378 * Test whether the free scanner has reached the same or lower pageblock than
1379 * the migration scanner, and compaction should thus terminate.
1381 static inline bool compact_scanners_met(struct compact_control *cc)
1383 return (cc->free_pfn >> pageblock_order)
1384 <= (cc->migrate_pfn >> pageblock_order);
1388 * Used when scanning for a suitable migration target which scans freelists
1389 * in reverse. Reorders the list such as the unscanned pages are scanned
1390 * first on the next iteration of the free scanner
1393 move_freelist_head(struct list_head *freelist, struct page *freepage)
1397 if (!list_is_first(&freepage->buddy_list, freelist)) {
1398 list_cut_before(&sublist, freelist, &freepage->buddy_list);
1399 list_splice_tail(&sublist, freelist);
1404 * Similar to move_freelist_head except used by the migration scanner
1405 * when scanning forward. It's possible for these list operations to
1406 * move against each other if they search the free list exactly in
1410 move_freelist_tail(struct list_head *freelist, struct page *freepage)
1414 if (!list_is_last(&freepage->buddy_list, freelist)) {
1415 list_cut_position(&sublist, freelist, &freepage->buddy_list);
1416 list_splice_tail(&sublist, freelist);
1421 fast_isolate_around(struct compact_control *cc, unsigned long pfn)
1423 unsigned long start_pfn, end_pfn;
1426 /* Do not search around if there are enough pages already */
1427 if (cc->nr_freepages >= cc->nr_migratepages)
1430 /* Minimise scanning during async compaction */
1431 if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC)
1434 /* Pageblock boundaries */
1435 start_pfn = max(pageblock_start_pfn(pfn), cc->zone->zone_start_pfn);
1436 end_pfn = min(pageblock_end_pfn(pfn), zone_end_pfn(cc->zone));
1438 page = pageblock_pfn_to_page(start_pfn, end_pfn, cc->zone);
1442 isolate_freepages_block(cc, &start_pfn, end_pfn, &cc->freepages, 1, false);
1444 /* Skip this pageblock in the future as it's full or nearly full */
1445 if (start_pfn == end_pfn && !cc->no_set_skip_hint)
1446 set_pageblock_skip(page);
1449 /* Search orders in round-robin fashion */
1450 static int next_search_order(struct compact_control *cc, int order)
1454 order = cc->order - 1;
1456 /* Search wrapped around? */
1457 if (order == cc->search_order) {
1459 if (cc->search_order < 0)
1460 cc->search_order = cc->order - 1;
1467 static void fast_isolate_freepages(struct compact_control *cc)
1469 unsigned int limit = max(1U, freelist_scan_limit(cc) >> 1);
1470 unsigned int nr_scanned = 0, total_isolated = 0;
1471 unsigned long low_pfn, min_pfn, highest = 0;
1472 unsigned long nr_isolated = 0;
1473 unsigned long distance;
1474 struct page *page = NULL;
1475 bool scan_start = false;
1478 /* Full compaction passes in a negative order */
1483 * If starting the scan, use a deeper search and use the highest
1484 * PFN found if a suitable one is not found.
1486 if (cc->free_pfn >= cc->zone->compact_init_free_pfn) {
1487 limit = pageblock_nr_pages >> 1;
1492 * Preferred point is in the top quarter of the scan space but take
1493 * a pfn from the top half if the search is problematic.
1495 distance = (cc->free_pfn - cc->migrate_pfn);
1496 low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2));
1497 min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1));
1499 if (WARN_ON_ONCE(min_pfn > low_pfn))
1503 * Search starts from the last successful isolation order or the next
1504 * order to search after a previous failure
1506 cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order);
1508 for (order = cc->search_order;
1509 !page && order >= 0;
1510 order = next_search_order(cc, order)) {
1511 struct free_area *area = &cc->zone->free_area[order];
1512 struct list_head *freelist;
1513 struct page *freepage;
1514 unsigned long flags;
1515 unsigned int order_scanned = 0;
1516 unsigned long high_pfn = 0;
1521 spin_lock_irqsave(&cc->zone->lock, flags);
1522 freelist = &area->free_list[MIGRATE_MOVABLE];
1523 list_for_each_entry_reverse(freepage, freelist, buddy_list) {
1528 pfn = page_to_pfn(freepage);
1531 highest = max(pageblock_start_pfn(pfn),
1532 cc->zone->zone_start_pfn);
1534 if (pfn >= low_pfn) {
1535 cc->fast_search_fail = 0;
1536 cc->search_order = order;
1541 if (pfn >= min_pfn && pfn > high_pfn) {
1544 /* Shorten the scan if a candidate is found */
1548 if (order_scanned >= limit)
1552 /* Use a maximum candidate pfn if a preferred one was not found */
1553 if (!page && high_pfn) {
1554 page = pfn_to_page(high_pfn);
1556 /* Update freepage for the list reorder below */
1560 /* Reorder to so a future search skips recent pages */
1561 move_freelist_head(freelist, freepage);
1563 /* Isolate the page if available */
1565 if (__isolate_free_page(page, order)) {
1566 set_page_private(page, order);
1567 nr_isolated = 1 << order;
1568 nr_scanned += nr_isolated - 1;
1569 total_isolated += nr_isolated;
1570 cc->nr_freepages += nr_isolated;
1571 list_add_tail(&page->lru, &cc->freepages);
1572 count_compact_events(COMPACTISOLATED, nr_isolated);
1574 /* If isolation fails, abort the search */
1575 order = cc->search_order + 1;
1580 spin_unlock_irqrestore(&cc->zone->lock, flags);
1582 /* Skip fast search if enough freepages isolated */
1583 if (cc->nr_freepages >= cc->nr_migratepages)
1587 * Smaller scan on next order so the total scan is related
1588 * to freelist_scan_limit.
1590 if (order_scanned >= limit)
1591 limit = max(1U, limit >> 1);
1594 trace_mm_compaction_fast_isolate_freepages(min_pfn, cc->free_pfn,
1595 nr_scanned, total_isolated);
1598 cc->fast_search_fail++;
1601 * Use the highest PFN found above min. If one was
1602 * not found, be pessimistic for direct compaction
1603 * and use the min mark.
1605 if (highest >= min_pfn) {
1606 page = pfn_to_page(highest);
1607 cc->free_pfn = highest;
1609 if (cc->direct_compaction && pfn_valid(min_pfn)) {
1610 page = pageblock_pfn_to_page(min_pfn,
1611 min(pageblock_end_pfn(min_pfn),
1612 zone_end_pfn(cc->zone)),
1614 if (page && !suitable_migration_target(cc, page))
1617 cc->free_pfn = min_pfn;
1623 if (highest && highest >= cc->zone->compact_cached_free_pfn) {
1624 highest -= pageblock_nr_pages;
1625 cc->zone->compact_cached_free_pfn = highest;
1628 cc->total_free_scanned += nr_scanned;
1632 low_pfn = page_to_pfn(page);
1633 fast_isolate_around(cc, low_pfn);
1637 * Based on information in the current compact_control, find blocks
1638 * suitable for isolating free pages from and then isolate them.
1640 static void isolate_freepages(struct compact_control *cc)
1642 struct zone *zone = cc->zone;
1644 unsigned long block_start_pfn; /* start of current pageblock */
1645 unsigned long isolate_start_pfn; /* exact pfn we start at */
1646 unsigned long block_end_pfn; /* end of current pageblock */
1647 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
1648 struct list_head *freelist = &cc->freepages;
1649 unsigned int stride;
1651 /* Try a small search of the free lists for a candidate */
1652 fast_isolate_freepages(cc);
1653 if (cc->nr_freepages)
1657 * Initialise the free scanner. The starting point is where we last
1658 * successfully isolated from, zone-cached value, or the end of the
1659 * zone when isolating for the first time. For looping we also need
1660 * this pfn aligned down to the pageblock boundary, because we do
1661 * block_start_pfn -= pageblock_nr_pages in the for loop.
1662 * For ending point, take care when isolating in last pageblock of a
1663 * zone which ends in the middle of a pageblock.
1664 * The low boundary is the end of the pageblock the migration scanner
1667 isolate_start_pfn = cc->free_pfn;
1668 block_start_pfn = pageblock_start_pfn(isolate_start_pfn);
1669 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1670 zone_end_pfn(zone));
1671 low_pfn = pageblock_end_pfn(cc->migrate_pfn);
1672 stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1;
1675 * Isolate free pages until enough are available to migrate the
1676 * pages on cc->migratepages. We stop searching if the migrate
1677 * and free page scanners meet or enough free pages are isolated.
1679 for (; block_start_pfn >= low_pfn;
1680 block_end_pfn = block_start_pfn,
1681 block_start_pfn -= pageblock_nr_pages,
1682 isolate_start_pfn = block_start_pfn) {
1683 unsigned long nr_isolated;
1686 * This can iterate a massively long zone without finding any
1687 * suitable migration targets, so periodically check resched.
1689 if (!(block_start_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages)))
1692 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1695 unsigned long next_pfn;
1697 next_pfn = skip_offline_sections_reverse(block_start_pfn);
1699 block_start_pfn = max(next_pfn, low_pfn);
1704 /* Check the block is suitable for migration */
1705 if (!suitable_migration_target(cc, page))
1708 /* If isolation recently failed, do not retry */
1709 if (!isolation_suitable(cc, page))
1712 /* Found a block suitable for isolating free pages from. */
1713 nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn,
1714 block_end_pfn, freelist, stride, false);
1716 /* Update the skip hint if the full pageblock was scanned */
1717 if (isolate_start_pfn == block_end_pfn)
1718 update_pageblock_skip(cc, page, block_start_pfn -
1719 pageblock_nr_pages);
1721 /* Are enough freepages isolated? */
1722 if (cc->nr_freepages >= cc->nr_migratepages) {
1723 if (isolate_start_pfn >= block_end_pfn) {
1725 * Restart at previous pageblock if more
1726 * freepages can be isolated next time.
1729 block_start_pfn - pageblock_nr_pages;
1732 } else if (isolate_start_pfn < block_end_pfn) {
1734 * If isolation failed early, do not continue
1740 /* Adjust stride depending on isolation */
1745 stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1);
1749 * Record where the free scanner will restart next time. Either we
1750 * broke from the loop and set isolate_start_pfn based on the last
1751 * call to isolate_freepages_block(), or we met the migration scanner
1752 * and the loop terminated due to isolate_start_pfn < low_pfn
1754 cc->free_pfn = isolate_start_pfn;
1757 /* __isolate_free_page() does not map the pages */
1758 split_map_pages(freelist);
1762 * This is a migrate-callback that "allocates" freepages by taking pages
1763 * from the isolated freelists in the block we are migrating to.
1765 static struct folio *compaction_alloc(struct folio *src, unsigned long data)
1767 struct compact_control *cc = (struct compact_control *)data;
1770 if (list_empty(&cc->freepages)) {
1771 isolate_freepages(cc);
1773 if (list_empty(&cc->freepages))
1777 dst = list_entry(cc->freepages.next, struct folio, lru);
1778 list_del(&dst->lru);
1785 * This is a migrate-callback that "frees" freepages back to the isolated
1786 * freelist. All pages on the freelist are from the same zone, so there is no
1787 * special handling needed for NUMA.
1789 static void compaction_free(struct folio *dst, unsigned long data)
1791 struct compact_control *cc = (struct compact_control *)data;
1793 list_add(&dst->lru, &cc->freepages);
1797 /* possible outcome of isolate_migratepages */
1799 ISOLATE_ABORT, /* Abort compaction now */
1800 ISOLATE_NONE, /* No pages isolated, continue scanning */
1801 ISOLATE_SUCCESS, /* Pages isolated, migrate */
1802 } isolate_migrate_t;
1805 * Allow userspace to control policy on scanning the unevictable LRU for
1806 * compactable pages.
1808 static int sysctl_compact_unevictable_allowed __read_mostly = CONFIG_COMPACT_UNEVICTABLE_DEFAULT;
1810 * Tunable for proactive compaction. It determines how
1811 * aggressively the kernel should compact memory in the
1812 * background. It takes values in the range [0, 100].
1814 static unsigned int __read_mostly sysctl_compaction_proactiveness = 20;
1815 static int sysctl_extfrag_threshold = 500;
1816 static int __read_mostly sysctl_compact_memory;
1819 update_fast_start_pfn(struct compact_control *cc, unsigned long pfn)
1821 if (cc->fast_start_pfn == ULONG_MAX)
1824 if (!cc->fast_start_pfn)
1825 cc->fast_start_pfn = pfn;
1827 cc->fast_start_pfn = min(cc->fast_start_pfn, pfn);
1830 static inline unsigned long
1831 reinit_migrate_pfn(struct compact_control *cc)
1833 if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX)
1834 return cc->migrate_pfn;
1836 cc->migrate_pfn = cc->fast_start_pfn;
1837 cc->fast_start_pfn = ULONG_MAX;
1839 return cc->migrate_pfn;
1843 * Briefly search the free lists for a migration source that already has
1844 * some free pages to reduce the number of pages that need migration
1845 * before a pageblock is free.
1847 static unsigned long fast_find_migrateblock(struct compact_control *cc)
1849 unsigned int limit = freelist_scan_limit(cc);
1850 unsigned int nr_scanned = 0;
1851 unsigned long distance;
1852 unsigned long pfn = cc->migrate_pfn;
1853 unsigned long high_pfn;
1855 bool found_block = false;
1857 /* Skip hints are relied on to avoid repeats on the fast search */
1858 if (cc->ignore_skip_hint)
1862 * If the pageblock should be finished then do not select a different
1865 if (cc->finish_pageblock)
1869 * If the migrate_pfn is not at the start of a zone or the start
1870 * of a pageblock then assume this is a continuation of a previous
1871 * scan restarted due to COMPACT_CLUSTER_MAX.
1873 if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn))
1877 * For smaller orders, just linearly scan as the number of pages
1878 * to migrate should be relatively small and does not necessarily
1879 * justify freeing up a large block for a small allocation.
1881 if (cc->order <= PAGE_ALLOC_COSTLY_ORDER)
1885 * Only allow kcompactd and direct requests for movable pages to
1886 * quickly clear out a MOVABLE pageblock for allocation. This
1887 * reduces the risk that a large movable pageblock is freed for
1888 * an unmovable/reclaimable small allocation.
1890 if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE)
1894 * When starting the migration scanner, pick any pageblock within the
1895 * first half of the search space. Otherwise try and pick a pageblock
1896 * within the first eighth to reduce the chances that a migration
1897 * target later becomes a source.
1899 distance = (cc->free_pfn - cc->migrate_pfn) >> 1;
1900 if (cc->migrate_pfn != cc->zone->zone_start_pfn)
1902 high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance);
1904 for (order = cc->order - 1;
1905 order >= PAGE_ALLOC_COSTLY_ORDER && !found_block && nr_scanned < limit;
1907 struct free_area *area = &cc->zone->free_area[order];
1908 struct list_head *freelist;
1909 unsigned long flags;
1910 struct page *freepage;
1915 spin_lock_irqsave(&cc->zone->lock, flags);
1916 freelist = &area->free_list[MIGRATE_MOVABLE];
1917 list_for_each_entry(freepage, freelist, buddy_list) {
1918 unsigned long free_pfn;
1920 if (nr_scanned++ >= limit) {
1921 move_freelist_tail(freelist, freepage);
1925 free_pfn = page_to_pfn(freepage);
1926 if (free_pfn < high_pfn) {
1928 * Avoid if skipped recently. Ideally it would
1929 * move to the tail but even safe iteration of
1930 * the list assumes an entry is deleted, not
1933 if (get_pageblock_skip(freepage))
1936 /* Reorder to so a future search skips recent pages */
1937 move_freelist_tail(freelist, freepage);
1939 update_fast_start_pfn(cc, free_pfn);
1940 pfn = pageblock_start_pfn(free_pfn);
1941 if (pfn < cc->zone->zone_start_pfn)
1942 pfn = cc->zone->zone_start_pfn;
1943 cc->fast_search_fail = 0;
1948 spin_unlock_irqrestore(&cc->zone->lock, flags);
1951 cc->total_migrate_scanned += nr_scanned;
1954 * If fast scanning failed then use a cached entry for a page block
1955 * that had free pages as the basis for starting a linear scan.
1958 cc->fast_search_fail++;
1959 pfn = reinit_migrate_pfn(cc);
1965 * Isolate all pages that can be migrated from the first suitable block,
1966 * starting at the block pointed to by the migrate scanner pfn within
1969 static isolate_migrate_t isolate_migratepages(struct compact_control *cc)
1971 unsigned long block_start_pfn;
1972 unsigned long block_end_pfn;
1973 unsigned long low_pfn;
1975 const isolate_mode_t isolate_mode =
1976 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1977 (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1978 bool fast_find_block;
1981 * Start at where we last stopped, or beginning of the zone as
1982 * initialized by compact_zone(). The first failure will use
1983 * the lowest PFN as the starting point for linear scanning.
1985 low_pfn = fast_find_migrateblock(cc);
1986 block_start_pfn = pageblock_start_pfn(low_pfn);
1987 if (block_start_pfn < cc->zone->zone_start_pfn)
1988 block_start_pfn = cc->zone->zone_start_pfn;
1991 * fast_find_migrateblock() has already ensured the pageblock is not
1992 * set with a skipped flag, so to avoid the isolation_suitable check
1993 * below again, check whether the fast search was successful.
1995 fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail;
1997 /* Only scan within a pageblock boundary */
1998 block_end_pfn = pageblock_end_pfn(low_pfn);
2001 * Iterate over whole pageblocks until we find the first suitable.
2002 * Do not cross the free scanner.
2004 for (; block_end_pfn <= cc->free_pfn;
2005 fast_find_block = false,
2006 cc->migrate_pfn = low_pfn = block_end_pfn,
2007 block_start_pfn = block_end_pfn,
2008 block_end_pfn += pageblock_nr_pages) {
2011 * This can potentially iterate a massively long zone with
2012 * many pageblocks unsuitable, so periodically check if we
2015 if (!(low_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages)))
2018 page = pageblock_pfn_to_page(block_start_pfn,
2019 block_end_pfn, cc->zone);
2021 unsigned long next_pfn;
2023 next_pfn = skip_offline_sections(block_start_pfn);
2025 block_end_pfn = min(next_pfn, cc->free_pfn);
2030 * If isolation recently failed, do not retry. Only check the
2031 * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock
2032 * to be visited multiple times. Assume skip was checked
2033 * before making it "skip" so other compaction instances do
2034 * not scan the same block.
2036 if ((pageblock_aligned(low_pfn) ||
2037 low_pfn == cc->zone->zone_start_pfn) &&
2038 !fast_find_block && !isolation_suitable(cc, page))
2042 * For async direct compaction, only scan the pageblocks of the
2043 * same migratetype without huge pages. Async direct compaction
2044 * is optimistic to see if the minimum amount of work satisfies
2045 * the allocation. The cached PFN is updated as it's possible
2046 * that all remaining blocks between source and target are
2047 * unsuitable and the compaction scanners fail to meet.
2049 if (!suitable_migration_source(cc, page)) {
2050 update_cached_migrate(cc, block_end_pfn);
2054 /* Perform the isolation */
2055 if (isolate_migratepages_block(cc, low_pfn, block_end_pfn,
2057 return ISOLATE_ABORT;
2060 * Either we isolated something and proceed with migration. Or
2061 * we failed and compact_zone should decide if we should
2067 return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
2071 * order == -1 is expected when compacting proactively via
2072 * 1. /proc/sys/vm/compact_memory
2073 * 2. /sys/devices/system/node/nodex/compact
2074 * 3. /proc/sys/vm/compaction_proactiveness
2076 static inline bool is_via_compact_memory(int order)
2082 * Determine whether kswapd is (or recently was!) running on this node.
2084 * pgdat_kswapd_lock() pins pgdat->kswapd, so a concurrent kswapd_stop() can't
2087 static bool kswapd_is_running(pg_data_t *pgdat)
2091 pgdat_kswapd_lock(pgdat);
2092 running = pgdat->kswapd && task_is_running(pgdat->kswapd);
2093 pgdat_kswapd_unlock(pgdat);
2099 * A zone's fragmentation score is the external fragmentation wrt to the
2100 * COMPACTION_HPAGE_ORDER. It returns a value in the range [0, 100].
2102 static unsigned int fragmentation_score_zone(struct zone *zone)
2104 return extfrag_for_order(zone, COMPACTION_HPAGE_ORDER);
2108 * A weighted zone's fragmentation score is the external fragmentation
2109 * wrt to the COMPACTION_HPAGE_ORDER scaled by the zone's size. It
2110 * returns a value in the range [0, 100].
2112 * The scaling factor ensures that proactive compaction focuses on larger
2113 * zones like ZONE_NORMAL, rather than smaller, specialized zones like
2114 * ZONE_DMA32. For smaller zones, the score value remains close to zero,
2115 * and thus never exceeds the high threshold for proactive compaction.
2117 static unsigned int fragmentation_score_zone_weighted(struct zone *zone)
2119 unsigned long score;
2121 score = zone->present_pages * fragmentation_score_zone(zone);
2122 return div64_ul(score, zone->zone_pgdat->node_present_pages + 1);
2126 * The per-node proactive (background) compaction process is started by its
2127 * corresponding kcompactd thread when the node's fragmentation score
2128 * exceeds the high threshold. The compaction process remains active till
2129 * the node's score falls below the low threshold, or one of the back-off
2130 * conditions is met.
2132 static unsigned int fragmentation_score_node(pg_data_t *pgdat)
2134 unsigned int score = 0;
2137 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2140 zone = &pgdat->node_zones[zoneid];
2141 if (!populated_zone(zone))
2143 score += fragmentation_score_zone_weighted(zone);
2149 static unsigned int fragmentation_score_wmark(bool low)
2151 unsigned int wmark_low;
2154 * Cap the low watermark to avoid excessive compaction
2155 * activity in case a user sets the proactiveness tunable
2156 * close to 100 (maximum).
2158 wmark_low = max(100U - sysctl_compaction_proactiveness, 5U);
2159 return low ? wmark_low : min(wmark_low + 10, 100U);
2162 static bool should_proactive_compact_node(pg_data_t *pgdat)
2166 if (!sysctl_compaction_proactiveness || kswapd_is_running(pgdat))
2169 wmark_high = fragmentation_score_wmark(false);
2170 return fragmentation_score_node(pgdat) > wmark_high;
2173 static enum compact_result __compact_finished(struct compact_control *cc)
2176 const int migratetype = cc->migratetype;
2179 /* Compaction run completes if the migrate and free scanner meet */
2180 if (compact_scanners_met(cc)) {
2181 /* Let the next compaction start anew. */
2182 reset_cached_positions(cc->zone);
2185 * Mark that the PG_migrate_skip information should be cleared
2186 * by kswapd when it goes to sleep. kcompactd does not set the
2187 * flag itself as the decision to be clear should be directly
2188 * based on an allocation request.
2190 if (cc->direct_compaction)
2191 cc->zone->compact_blockskip_flush = true;
2194 return COMPACT_COMPLETE;
2196 return COMPACT_PARTIAL_SKIPPED;
2199 if (cc->proactive_compaction) {
2200 int score, wmark_low;
2203 pgdat = cc->zone->zone_pgdat;
2204 if (kswapd_is_running(pgdat))
2205 return COMPACT_PARTIAL_SKIPPED;
2207 score = fragmentation_score_zone(cc->zone);
2208 wmark_low = fragmentation_score_wmark(true);
2210 if (score > wmark_low)
2211 ret = COMPACT_CONTINUE;
2213 ret = COMPACT_SUCCESS;
2218 if (is_via_compact_memory(cc->order))
2219 return COMPACT_CONTINUE;
2222 * Always finish scanning a pageblock to reduce the possibility of
2223 * fallbacks in the future. This is particularly important when
2224 * migration source is unmovable/reclaimable but it's not worth
2227 if (!pageblock_aligned(cc->migrate_pfn))
2228 return COMPACT_CONTINUE;
2230 /* Direct compactor: Is a suitable page free? */
2231 ret = COMPACT_NO_SUITABLE_PAGE;
2232 for (order = cc->order; order < NR_PAGE_ORDERS; order++) {
2233 struct free_area *area = &cc->zone->free_area[order];
2236 /* Job done if page is free of the right migratetype */
2237 if (!free_area_empty(area, migratetype))
2238 return COMPACT_SUCCESS;
2241 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
2242 if (migratetype == MIGRATE_MOVABLE &&
2243 !free_area_empty(area, MIGRATE_CMA))
2244 return COMPACT_SUCCESS;
2247 * Job done if allocation would steal freepages from
2248 * other migratetype buddy lists.
2250 if (find_suitable_fallback(area, order, migratetype,
2251 true, &can_steal) != -1)
2253 * Movable pages are OK in any pageblock. If we are
2254 * stealing for a non-movable allocation, make sure
2255 * we finish compacting the current pageblock first
2256 * (which is assured by the above migrate_pfn align
2257 * check) so it is as free as possible and we won't
2258 * have to steal another one soon.
2260 return COMPACT_SUCCESS;
2264 if (cc->contended || fatal_signal_pending(current))
2265 ret = COMPACT_CONTENDED;
2270 static enum compact_result compact_finished(struct compact_control *cc)
2274 ret = __compact_finished(cc);
2275 trace_mm_compaction_finished(cc->zone, cc->order, ret);
2276 if (ret == COMPACT_NO_SUITABLE_PAGE)
2277 ret = COMPACT_CONTINUE;
2282 static bool __compaction_suitable(struct zone *zone, int order,
2283 int highest_zoneidx,
2284 unsigned long wmark_target)
2286 unsigned long watermark;
2288 * Watermarks for order-0 must be met for compaction to be able to
2289 * isolate free pages for migration targets. This means that the
2290 * watermark and alloc_flags have to match, or be more pessimistic than
2291 * the check in __isolate_free_page(). We don't use the direct
2292 * compactor's alloc_flags, as they are not relevant for freepage
2293 * isolation. We however do use the direct compactor's highest_zoneidx
2294 * to skip over zones where lowmem reserves would prevent allocation
2295 * even if compaction succeeds.
2296 * For costly orders, we require low watermark instead of min for
2297 * compaction to proceed to increase its chances.
2298 * ALLOC_CMA is used, as pages in CMA pageblocks are considered
2299 * suitable migration targets
2301 watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
2302 low_wmark_pages(zone) : min_wmark_pages(zone);
2303 watermark += compact_gap(order);
2304 return __zone_watermark_ok(zone, 0, watermark, highest_zoneidx,
2305 ALLOC_CMA, wmark_target);
2309 * compaction_suitable: Is this suitable to run compaction on this zone now?
2311 bool compaction_suitable(struct zone *zone, int order, int highest_zoneidx)
2313 enum compact_result compact_result;
2316 suitable = __compaction_suitable(zone, order, highest_zoneidx,
2317 zone_page_state(zone, NR_FREE_PAGES));
2319 * fragmentation index determines if allocation failures are due to
2320 * low memory or external fragmentation
2322 * index of -1000 would imply allocations might succeed depending on
2323 * watermarks, but we already failed the high-order watermark check
2324 * index towards 0 implies failure is due to lack of memory
2325 * index towards 1000 implies failure is due to fragmentation
2327 * Only compact if a failure would be due to fragmentation. Also
2328 * ignore fragindex for non-costly orders where the alternative to
2329 * a successful reclaim/compaction is OOM. Fragindex and the
2330 * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
2331 * excessive compaction for costly orders, but it should not be at the
2332 * expense of system stability.
2335 compact_result = COMPACT_CONTINUE;
2336 if (order > PAGE_ALLOC_COSTLY_ORDER) {
2337 int fragindex = fragmentation_index(zone, order);
2339 if (fragindex >= 0 &&
2340 fragindex <= sysctl_extfrag_threshold) {
2342 compact_result = COMPACT_NOT_SUITABLE_ZONE;
2346 compact_result = COMPACT_SKIPPED;
2349 trace_mm_compaction_suitable(zone, order, compact_result);
2354 bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
2361 * Make sure at least one zone would pass __compaction_suitable if we continue
2362 * retrying the reclaim.
2364 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2365 ac->highest_zoneidx, ac->nodemask) {
2366 unsigned long available;
2369 * Do not consider all the reclaimable memory because we do not
2370 * want to trash just for a single high order allocation which
2371 * is even not guaranteed to appear even if __compaction_suitable
2372 * is happy about the watermark check.
2374 available = zone_reclaimable_pages(zone) / order;
2375 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
2376 if (__compaction_suitable(zone, order, ac->highest_zoneidx,
2385 * Should we do compaction for target allocation order.
2386 * Return COMPACT_SUCCESS if allocation for target order can be already
2388 * Return COMPACT_SKIPPED if compaction for target order is likely to fail
2389 * Return COMPACT_CONTINUE if compaction for target order should be ran
2391 static enum compact_result
2392 compaction_suit_allocation_order(struct zone *zone, unsigned int order,
2393 int highest_zoneidx, unsigned int alloc_flags)
2395 unsigned long watermark;
2397 watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
2398 if (zone_watermark_ok(zone, order, watermark, highest_zoneidx,
2400 return COMPACT_SUCCESS;
2402 if (!compaction_suitable(zone, order, highest_zoneidx))
2403 return COMPACT_SKIPPED;
2405 return COMPACT_CONTINUE;
2408 static enum compact_result
2409 compact_zone(struct compact_control *cc, struct capture_control *capc)
2411 enum compact_result ret;
2412 unsigned long start_pfn = cc->zone->zone_start_pfn;
2413 unsigned long end_pfn = zone_end_pfn(cc->zone);
2414 unsigned long last_migrated_pfn;
2415 const bool sync = cc->mode != MIGRATE_ASYNC;
2417 unsigned int nr_succeeded = 0;
2420 * These counters track activities during zone compaction. Initialize
2421 * them before compacting a new zone.
2423 cc->total_migrate_scanned = 0;
2424 cc->total_free_scanned = 0;
2425 cc->nr_migratepages = 0;
2426 cc->nr_freepages = 0;
2427 INIT_LIST_HEAD(&cc->freepages);
2428 INIT_LIST_HEAD(&cc->migratepages);
2430 cc->migratetype = gfp_migratetype(cc->gfp_mask);
2432 if (!is_via_compact_memory(cc->order)) {
2433 ret = compaction_suit_allocation_order(cc->zone, cc->order,
2434 cc->highest_zoneidx,
2436 if (ret != COMPACT_CONTINUE)
2441 * Clear pageblock skip if there were failures recently and compaction
2442 * is about to be retried after being deferred.
2444 if (compaction_restarting(cc->zone, cc->order))
2445 __reset_isolation_suitable(cc->zone);
2448 * Setup to move all movable pages to the end of the zone. Used cached
2449 * information on where the scanners should start (unless we explicitly
2450 * want to compact the whole zone), but check that it is initialised
2451 * by ensuring the values are within zone boundaries.
2453 cc->fast_start_pfn = 0;
2454 if (cc->whole_zone) {
2455 cc->migrate_pfn = start_pfn;
2456 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2458 cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync];
2459 cc->free_pfn = cc->zone->compact_cached_free_pfn;
2460 if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
2461 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2462 cc->zone->compact_cached_free_pfn = cc->free_pfn;
2464 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
2465 cc->migrate_pfn = start_pfn;
2466 cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
2467 cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
2470 if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn)
2471 cc->whole_zone = true;
2474 last_migrated_pfn = 0;
2477 * Migrate has separate cached PFNs for ASYNC and SYNC* migration on
2478 * the basis that some migrations will fail in ASYNC mode. However,
2479 * if the cached PFNs match and pageblocks are skipped due to having
2480 * no isolation candidates, then the sync state does not matter.
2481 * Until a pageblock with isolation candidates is found, keep the
2482 * cached PFNs in sync to avoid revisiting the same blocks.
2484 update_cached = !sync &&
2485 cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1];
2487 trace_mm_compaction_begin(cc, start_pfn, end_pfn, sync);
2489 /* lru_add_drain_all could be expensive with involving other CPUs */
2492 while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) {
2494 unsigned long iteration_start_pfn = cc->migrate_pfn;
2497 * Avoid multiple rescans of the same pageblock which can
2498 * happen if a page cannot be isolated (dirty/writeback in
2499 * async mode) or if the migrated pages are being allocated
2500 * before the pageblock is cleared. The first rescan will
2501 * capture the entire pageblock for migration. If it fails,
2502 * it'll be marked skip and scanning will proceed as normal.
2504 cc->finish_pageblock = false;
2505 if (pageblock_start_pfn(last_migrated_pfn) ==
2506 pageblock_start_pfn(iteration_start_pfn)) {
2507 cc->finish_pageblock = true;
2511 switch (isolate_migratepages(cc)) {
2513 ret = COMPACT_CONTENDED;
2514 putback_movable_pages(&cc->migratepages);
2515 cc->nr_migratepages = 0;
2518 if (update_cached) {
2519 cc->zone->compact_cached_migrate_pfn[1] =
2520 cc->zone->compact_cached_migrate_pfn[0];
2524 * We haven't isolated and migrated anything, but
2525 * there might still be unflushed migrations from
2526 * previous cc->order aligned block.
2529 case ISOLATE_SUCCESS:
2530 update_cached = false;
2531 last_migrated_pfn = max(cc->zone->zone_start_pfn,
2532 pageblock_start_pfn(cc->migrate_pfn - 1));
2535 err = migrate_pages(&cc->migratepages, compaction_alloc,
2536 compaction_free, (unsigned long)cc, cc->mode,
2537 MR_COMPACTION, &nr_succeeded);
2539 trace_mm_compaction_migratepages(cc, nr_succeeded);
2541 /* All pages were either migrated or will be released */
2542 cc->nr_migratepages = 0;
2544 putback_movable_pages(&cc->migratepages);
2546 * migrate_pages() may return -ENOMEM when scanners meet
2547 * and we want compact_finished() to detect it
2549 if (err == -ENOMEM && !compact_scanners_met(cc)) {
2550 ret = COMPACT_CONTENDED;
2554 * If an ASYNC or SYNC_LIGHT fails to migrate a page
2555 * within the pageblock_order-aligned block and
2556 * fast_find_migrateblock may be used then scan the
2557 * remainder of the pageblock. This will mark the
2558 * pageblock "skip" to avoid rescanning in the near
2559 * future. This will isolate more pages than necessary
2560 * for the request but avoid loops due to
2561 * fast_find_migrateblock revisiting blocks that were
2562 * recently partially scanned.
2564 if (!pageblock_aligned(cc->migrate_pfn) &&
2565 !cc->ignore_skip_hint && !cc->finish_pageblock &&
2566 (cc->mode < MIGRATE_SYNC)) {
2567 cc->finish_pageblock = true;
2570 * Draining pcplists does not help THP if
2571 * any page failed to migrate. Even after
2572 * drain, the pageblock will not be free.
2574 if (cc->order == COMPACTION_HPAGE_ORDER)
2575 last_migrated_pfn = 0;
2581 /* Stop if a page has been captured */
2582 if (capc && capc->page) {
2583 ret = COMPACT_SUCCESS;
2589 * Has the migration scanner moved away from the previous
2590 * cc->order aligned block where we migrated from? If yes,
2591 * flush the pages that were freed, so that they can merge and
2592 * compact_finished() can detect immediately if allocation
2595 if (cc->order > 0 && last_migrated_pfn) {
2596 unsigned long current_block_start =
2597 block_start_pfn(cc->migrate_pfn, cc->order);
2599 if (last_migrated_pfn < current_block_start) {
2600 lru_add_drain_cpu_zone(cc->zone);
2601 /* No more flushing until we migrate again */
2602 last_migrated_pfn = 0;
2609 * Release free pages and update where the free scanner should restart,
2610 * so we don't leave any returned pages behind in the next attempt.
2612 if (cc->nr_freepages > 0) {
2613 unsigned long free_pfn = release_freepages(&cc->freepages);
2615 cc->nr_freepages = 0;
2616 VM_BUG_ON(free_pfn == 0);
2617 /* The cached pfn is always the first in a pageblock */
2618 free_pfn = pageblock_start_pfn(free_pfn);
2620 * Only go back, not forward. The cached pfn might have been
2621 * already reset to zone end in compact_finished()
2623 if (free_pfn > cc->zone->compact_cached_free_pfn)
2624 cc->zone->compact_cached_free_pfn = free_pfn;
2627 count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned);
2628 count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned);
2630 trace_mm_compaction_end(cc, start_pfn, end_pfn, sync, ret);
2632 VM_BUG_ON(!list_empty(&cc->freepages));
2633 VM_BUG_ON(!list_empty(&cc->migratepages));
2638 static enum compact_result compact_zone_order(struct zone *zone, int order,
2639 gfp_t gfp_mask, enum compact_priority prio,
2640 unsigned int alloc_flags, int highest_zoneidx,
2641 struct page **capture)
2643 enum compact_result ret;
2644 struct compact_control cc = {
2646 .search_order = order,
2647 .gfp_mask = gfp_mask,
2649 .mode = (prio == COMPACT_PRIO_ASYNC) ?
2650 MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT,
2651 .alloc_flags = alloc_flags,
2652 .highest_zoneidx = highest_zoneidx,
2653 .direct_compaction = true,
2654 .whole_zone = (prio == MIN_COMPACT_PRIORITY),
2655 .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
2656 .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
2658 struct capture_control capc = {
2664 * Make sure the structs are really initialized before we expose the
2665 * capture control, in case we are interrupted and the interrupt handler
2669 WRITE_ONCE(current->capture_control, &capc);
2671 ret = compact_zone(&cc, &capc);
2674 * Make sure we hide capture control first before we read the captured
2675 * page pointer, otherwise an interrupt could free and capture a page
2676 * and we would leak it.
2678 WRITE_ONCE(current->capture_control, NULL);
2679 *capture = READ_ONCE(capc.page);
2681 * Technically, it is also possible that compaction is skipped but
2682 * the page is still captured out of luck(IRQ came and freed the page).
2683 * Returning COMPACT_SUCCESS in such cases helps in properly accounting
2684 * the COMPACT[STALL|FAIL] when compaction is skipped.
2687 ret = COMPACT_SUCCESS;
2693 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
2694 * @gfp_mask: The GFP mask of the current allocation
2695 * @order: The order of the current allocation
2696 * @alloc_flags: The allocation flags of the current allocation
2697 * @ac: The context of current allocation
2698 * @prio: Determines how hard direct compaction should try to succeed
2699 * @capture: Pointer to free page created by compaction will be stored here
2701 * This is the main entry point for direct page compaction.
2703 enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
2704 unsigned int alloc_flags, const struct alloc_context *ac,
2705 enum compact_priority prio, struct page **capture)
2707 int may_perform_io = (__force int)(gfp_mask & __GFP_IO);
2710 enum compact_result rc = COMPACT_SKIPPED;
2713 * Check if the GFP flags allow compaction - GFP_NOIO is really
2714 * tricky context because the migration might require IO
2716 if (!may_perform_io)
2717 return COMPACT_SKIPPED;
2719 trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
2721 /* Compact each zone in the list */
2722 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2723 ac->highest_zoneidx, ac->nodemask) {
2724 enum compact_result status;
2726 if (prio > MIN_COMPACT_PRIORITY
2727 && compaction_deferred(zone, order)) {
2728 rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
2732 status = compact_zone_order(zone, order, gfp_mask, prio,
2733 alloc_flags, ac->highest_zoneidx, capture);
2734 rc = max(status, rc);
2736 /* The allocation should succeed, stop compacting */
2737 if (status == COMPACT_SUCCESS) {
2739 * We think the allocation will succeed in this zone,
2740 * but it is not certain, hence the false. The caller
2741 * will repeat this with true if allocation indeed
2742 * succeeds in this zone.
2744 compaction_defer_reset(zone, order, false);
2749 if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
2750 status == COMPACT_PARTIAL_SKIPPED))
2752 * We think that allocation won't succeed in this zone
2753 * so we defer compaction there. If it ends up
2754 * succeeding after all, it will be reset.
2756 defer_compaction(zone, order);
2759 * We might have stopped compacting due to need_resched() in
2760 * async compaction, or due to a fatal signal detected. In that
2761 * case do not try further zones
2763 if ((prio == COMPACT_PRIO_ASYNC && need_resched())
2764 || fatal_signal_pending(current))
2772 * Compact all zones within a node till each zone's fragmentation score
2773 * reaches within proactive compaction thresholds (as determined by the
2774 * proactiveness tunable).
2776 * It is possible that the function returns before reaching score targets
2777 * due to various back-off conditions, such as, contention on per-node or
2780 static void proactive_compact_node(pg_data_t *pgdat)
2784 struct compact_control cc = {
2786 .mode = MIGRATE_SYNC_LIGHT,
2787 .ignore_skip_hint = true,
2789 .gfp_mask = GFP_KERNEL,
2790 .proactive_compaction = true,
2793 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2794 zone = &pgdat->node_zones[zoneid];
2795 if (!populated_zone(zone))
2800 compact_zone(&cc, NULL);
2802 count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
2803 cc.total_migrate_scanned);
2804 count_compact_events(KCOMPACTD_FREE_SCANNED,
2805 cc.total_free_scanned);
2809 /* Compact all zones within a node */
2810 static void compact_node(int nid)
2812 pg_data_t *pgdat = NODE_DATA(nid);
2815 struct compact_control cc = {
2817 .mode = MIGRATE_SYNC,
2818 .ignore_skip_hint = true,
2820 .gfp_mask = GFP_KERNEL,
2824 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2826 zone = &pgdat->node_zones[zoneid];
2827 if (!populated_zone(zone))
2832 compact_zone(&cc, NULL);
2836 /* Compact all nodes in the system */
2837 static void compact_nodes(void)
2841 /* Flush pending updates to the LRU lists */
2842 lru_add_drain_all();
2844 for_each_online_node(nid)
2848 static int compaction_proactiveness_sysctl_handler(struct ctl_table *table, int write,
2849 void *buffer, size_t *length, loff_t *ppos)
2853 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
2857 if (write && sysctl_compaction_proactiveness) {
2858 for_each_online_node(nid) {
2859 pg_data_t *pgdat = NODE_DATA(nid);
2861 if (pgdat->proactive_compact_trigger)
2864 pgdat->proactive_compact_trigger = true;
2865 trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, -1,
2866 pgdat->nr_zones - 1);
2867 wake_up_interruptible(&pgdat->kcompactd_wait);
2875 * This is the entry point for compacting all nodes via
2876 * /proc/sys/vm/compact_memory
2878 static int sysctl_compaction_handler(struct ctl_table *table, int write,
2879 void *buffer, size_t *length, loff_t *ppos)
2883 ret = proc_dointvec(table, write, buffer, length, ppos);
2887 if (sysctl_compact_memory != 1)
2896 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
2897 static ssize_t compact_store(struct device *dev,
2898 struct device_attribute *attr,
2899 const char *buf, size_t count)
2903 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
2904 /* Flush pending updates to the LRU lists */
2905 lru_add_drain_all();
2912 static DEVICE_ATTR_WO(compact);
2914 int compaction_register_node(struct node *node)
2916 return device_create_file(&node->dev, &dev_attr_compact);
2919 void compaction_unregister_node(struct node *node)
2921 device_remove_file(&node->dev, &dev_attr_compact);
2923 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
2925 static inline bool kcompactd_work_requested(pg_data_t *pgdat)
2927 return pgdat->kcompactd_max_order > 0 || kthread_should_stop() ||
2928 pgdat->proactive_compact_trigger;
2931 static bool kcompactd_node_suitable(pg_data_t *pgdat)
2935 enum zone_type highest_zoneidx = pgdat->kcompactd_highest_zoneidx;
2936 enum compact_result ret;
2938 for (zoneid = 0; zoneid <= highest_zoneidx; zoneid++) {
2939 zone = &pgdat->node_zones[zoneid];
2941 if (!populated_zone(zone))
2944 ret = compaction_suit_allocation_order(zone,
2945 pgdat->kcompactd_max_order,
2946 highest_zoneidx, ALLOC_WMARK_MIN);
2947 if (ret == COMPACT_CONTINUE)
2954 static void kcompactd_do_work(pg_data_t *pgdat)
2957 * With no special task, compact all zones so that a page of requested
2958 * order is allocatable.
2962 struct compact_control cc = {
2963 .order = pgdat->kcompactd_max_order,
2964 .search_order = pgdat->kcompactd_max_order,
2965 .highest_zoneidx = pgdat->kcompactd_highest_zoneidx,
2966 .mode = MIGRATE_SYNC_LIGHT,
2967 .ignore_skip_hint = false,
2968 .gfp_mask = GFP_KERNEL,
2970 enum compact_result ret;
2972 trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
2973 cc.highest_zoneidx);
2974 count_compact_event(KCOMPACTD_WAKE);
2976 for (zoneid = 0; zoneid <= cc.highest_zoneidx; zoneid++) {
2979 zone = &pgdat->node_zones[zoneid];
2980 if (!populated_zone(zone))
2983 if (compaction_deferred(zone, cc.order))
2986 ret = compaction_suit_allocation_order(zone,
2987 cc.order, zoneid, ALLOC_WMARK_MIN);
2988 if (ret != COMPACT_CONTINUE)
2991 if (kthread_should_stop())
2995 status = compact_zone(&cc, NULL);
2997 if (status == COMPACT_SUCCESS) {
2998 compaction_defer_reset(zone, cc.order, false);
2999 } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
3001 * Buddy pages may become stranded on pcps that could
3002 * otherwise coalesce on the zone's free area for
3003 * order >= cc.order. This is ratelimited by the
3004 * upcoming deferral.
3006 drain_all_pages(zone);
3009 * We use sync migration mode here, so we defer like
3010 * sync direct compaction does.
3012 defer_compaction(zone, cc.order);
3015 count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
3016 cc.total_migrate_scanned);
3017 count_compact_events(KCOMPACTD_FREE_SCANNED,
3018 cc.total_free_scanned);
3022 * Regardless of success, we are done until woken up next. But remember
3023 * the requested order/highest_zoneidx in case it was higher/tighter
3024 * than our current ones
3026 if (pgdat->kcompactd_max_order <= cc.order)
3027 pgdat->kcompactd_max_order = 0;
3028 if (pgdat->kcompactd_highest_zoneidx >= cc.highest_zoneidx)
3029 pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
3032 void wakeup_kcompactd(pg_data_t *pgdat, int order, int highest_zoneidx)
3037 if (pgdat->kcompactd_max_order < order)
3038 pgdat->kcompactd_max_order = order;
3040 if (pgdat->kcompactd_highest_zoneidx > highest_zoneidx)
3041 pgdat->kcompactd_highest_zoneidx = highest_zoneidx;
3044 * Pairs with implicit barrier in wait_event_freezable()
3045 * such that wakeups are not missed.
3047 if (!wq_has_sleeper(&pgdat->kcompactd_wait))
3050 if (!kcompactd_node_suitable(pgdat))
3053 trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
3055 wake_up_interruptible(&pgdat->kcompactd_wait);
3059 * The background compaction daemon, started as a kernel thread
3060 * from the init process.
3062 static int kcompactd(void *p)
3064 pg_data_t *pgdat = (pg_data_t *)p;
3065 struct task_struct *tsk = current;
3066 long default_timeout = msecs_to_jiffies(HPAGE_FRAG_CHECK_INTERVAL_MSEC);
3067 long timeout = default_timeout;
3069 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3071 if (!cpumask_empty(cpumask))
3072 set_cpus_allowed_ptr(tsk, cpumask);
3076 pgdat->kcompactd_max_order = 0;
3077 pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
3079 while (!kthread_should_stop()) {
3080 unsigned long pflags;
3083 * Avoid the unnecessary wakeup for proactive compaction
3084 * when it is disabled.
3086 if (!sysctl_compaction_proactiveness)
3087 timeout = MAX_SCHEDULE_TIMEOUT;
3088 trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
3089 if (wait_event_freezable_timeout(pgdat->kcompactd_wait,
3090 kcompactd_work_requested(pgdat), timeout) &&
3091 !pgdat->proactive_compact_trigger) {
3093 psi_memstall_enter(&pflags);
3094 kcompactd_do_work(pgdat);
3095 psi_memstall_leave(&pflags);
3097 * Reset the timeout value. The defer timeout from
3098 * proactive compaction is lost here but that is fine
3099 * as the condition of the zone changing substantionally
3100 * then carrying on with the previous defer interval is
3103 timeout = default_timeout;
3108 * Start the proactive work with default timeout. Based
3109 * on the fragmentation score, this timeout is updated.
3111 timeout = default_timeout;
3112 if (should_proactive_compact_node(pgdat)) {
3113 unsigned int prev_score, score;
3115 prev_score = fragmentation_score_node(pgdat);
3116 proactive_compact_node(pgdat);
3117 score = fragmentation_score_node(pgdat);
3119 * Defer proactive compaction if the fragmentation
3120 * score did not go down i.e. no progress made.
3122 if (unlikely(score >= prev_score))
3124 default_timeout << COMPACT_MAX_DEFER_SHIFT;
3126 if (unlikely(pgdat->proactive_compact_trigger))
3127 pgdat->proactive_compact_trigger = false;
3134 * This kcompactd start function will be called by init and node-hot-add.
3135 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
3137 void __meminit kcompactd_run(int nid)
3139 pg_data_t *pgdat = NODE_DATA(nid);
3141 if (pgdat->kcompactd)
3144 pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
3145 if (IS_ERR(pgdat->kcompactd)) {
3146 pr_err("Failed to start kcompactd on node %d\n", nid);
3147 pgdat->kcompactd = NULL;
3152 * Called by memory hotplug when all memory in a node is offlined. Caller must
3153 * be holding mem_hotplug_begin/done().
3155 void __meminit kcompactd_stop(int nid)
3157 struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
3160 kthread_stop(kcompactd);
3161 NODE_DATA(nid)->kcompactd = NULL;
3166 * It's optimal to keep kcompactd on the same CPUs as their memory, but
3167 * not required for correctness. So if the last cpu in a node goes
3168 * away, we get changed to run anywhere: as the first one comes back,
3169 * restore their cpu bindings.
3171 static int kcompactd_cpu_online(unsigned int cpu)
3175 for_each_node_state(nid, N_MEMORY) {
3176 pg_data_t *pgdat = NODE_DATA(nid);
3177 const struct cpumask *mask;
3179 mask = cpumask_of_node(pgdat->node_id);
3181 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3182 /* One of our CPUs online: restore mask */
3183 if (pgdat->kcompactd)
3184 set_cpus_allowed_ptr(pgdat->kcompactd, mask);
3189 static int proc_dointvec_minmax_warn_RT_change(struct ctl_table *table,
3190 int write, void *buffer, size_t *lenp, loff_t *ppos)
3194 if (!IS_ENABLED(CONFIG_PREEMPT_RT) || !write)
3195 return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
3197 old = *(int *)table->data;
3198 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
3201 if (old != *(int *)table->data)
3202 pr_warn_once("sysctl attribute %s changed by %s[%d]\n",
3203 table->procname, current->comm,
3204 task_pid_nr(current));
3208 static struct ctl_table vm_compaction[] = {
3210 .procname = "compact_memory",
3211 .data = &sysctl_compact_memory,
3212 .maxlen = sizeof(int),
3214 .proc_handler = sysctl_compaction_handler,
3217 .procname = "compaction_proactiveness",
3218 .data = &sysctl_compaction_proactiveness,
3219 .maxlen = sizeof(sysctl_compaction_proactiveness),
3221 .proc_handler = compaction_proactiveness_sysctl_handler,
3222 .extra1 = SYSCTL_ZERO,
3223 .extra2 = SYSCTL_ONE_HUNDRED,
3226 .procname = "extfrag_threshold",
3227 .data = &sysctl_extfrag_threshold,
3228 .maxlen = sizeof(int),
3230 .proc_handler = proc_dointvec_minmax,
3231 .extra1 = SYSCTL_ZERO,
3232 .extra2 = SYSCTL_ONE_THOUSAND,
3235 .procname = "compact_unevictable_allowed",
3236 .data = &sysctl_compact_unevictable_allowed,
3237 .maxlen = sizeof(int),
3239 .proc_handler = proc_dointvec_minmax_warn_RT_change,
3240 .extra1 = SYSCTL_ZERO,
3241 .extra2 = SYSCTL_ONE,
3246 static int __init kcompactd_init(void)
3251 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3252 "mm/compaction:online",
3253 kcompactd_cpu_online, NULL);
3255 pr_err("kcompactd: failed to register hotplug callbacks.\n");
3259 for_each_node_state(nid, N_MEMORY)
3261 register_sysctl_init("vm", vm_compaction);
3264 subsys_initcall(kcompactd_init)
3266 #endif /* CONFIG_COMPACTION */