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 if (!zone->compact_blockskip_flush)
388 zone->compact_blockskip_flush = false;
391 * Walk the zone and update pageblock skip information. Source looks
392 * for PageLRU while target looks for PageBuddy. When the scanner
393 * is found, both PageBuddy and PageLRU are checked as the pageblock
394 * is suitable as both source and target.
396 for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages,
397 free_pfn -= pageblock_nr_pages) {
400 /* Update the migrate PFN */
401 if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) &&
402 migrate_pfn < reset_migrate) {
404 reset_migrate = migrate_pfn;
405 zone->compact_init_migrate_pfn = reset_migrate;
406 zone->compact_cached_migrate_pfn[0] = reset_migrate;
407 zone->compact_cached_migrate_pfn[1] = reset_migrate;
410 /* Update the free PFN */
411 if (__reset_isolation_pfn(zone, free_pfn, free_set, true) &&
412 free_pfn > reset_free) {
414 reset_free = free_pfn;
415 zone->compact_init_free_pfn = reset_free;
416 zone->compact_cached_free_pfn = reset_free;
420 /* Leave no distance if no suitable block was reset */
421 if (reset_migrate >= reset_free) {
422 zone->compact_cached_migrate_pfn[0] = migrate_pfn;
423 zone->compact_cached_migrate_pfn[1] = migrate_pfn;
424 zone->compact_cached_free_pfn = free_pfn;
428 void reset_isolation_suitable(pg_data_t *pgdat)
432 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
433 struct zone *zone = &pgdat->node_zones[zoneid];
434 if (!populated_zone(zone))
437 /* Only flush if a full compaction finished recently */
438 if (zone->compact_blockskip_flush)
439 __reset_isolation_suitable(zone);
444 * Sets the pageblock skip bit if it was clear. Note that this is a hint as
445 * locks are not required for read/writers. Returns true if it was already set.
447 static bool test_and_set_skip(struct compact_control *cc, struct page *page)
451 /* Do not update if skip hint is being ignored */
452 if (cc->ignore_skip_hint)
455 skip = get_pageblock_skip(page);
456 if (!skip && !cc->no_set_skip_hint)
457 set_pageblock_skip(page);
462 static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
464 struct zone *zone = cc->zone;
466 pfn = pageblock_end_pfn(pfn);
468 /* Set for isolation rather than compaction */
469 if (cc->no_set_skip_hint)
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 /* Update where async and sync compaction should restart */
494 if (pfn < zone->compact_cached_free_pfn)
495 zone->compact_cached_free_pfn = pfn;
498 static inline bool isolation_suitable(struct compact_control *cc,
504 static inline bool pageblock_skip_persistent(struct page *page)
509 static inline void update_pageblock_skip(struct compact_control *cc,
510 struct page *page, unsigned long pfn)
514 static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
518 static bool test_and_set_skip(struct compact_control *cc, struct page *page)
522 #endif /* CONFIG_COMPACTION */
525 * Compaction requires the taking of some coarse locks that are potentially
526 * very heavily contended. For async compaction, trylock and record if the
527 * lock is contended. The lock will still be acquired but compaction will
528 * abort when the current block is finished regardless of success rate.
529 * Sync compaction acquires the lock.
531 * Always returns true which makes it easier to track lock state in callers.
533 static bool compact_lock_irqsave(spinlock_t *lock, unsigned long *flags,
534 struct compact_control *cc)
537 /* Track if the lock is contended in async mode */
538 if (cc->mode == MIGRATE_ASYNC && !cc->contended) {
539 if (spin_trylock_irqsave(lock, *flags))
542 cc->contended = true;
545 spin_lock_irqsave(lock, *flags);
550 * Compaction requires the taking of some coarse locks that are potentially
551 * very heavily contended. The lock should be periodically unlocked to avoid
552 * having disabled IRQs for a long time, even when there is nobody waiting on
553 * the lock. It might also be that allowing the IRQs will result in
554 * need_resched() becoming true. If scheduling is needed, compaction schedules.
555 * Either compaction type will also abort if a fatal signal is pending.
556 * In either case if the lock was locked, it is dropped and not regained.
558 * Returns true if compaction should abort due to fatal signal pending.
559 * Returns false when compaction can continue.
561 static bool compact_unlock_should_abort(spinlock_t *lock,
562 unsigned long flags, bool *locked, struct compact_control *cc)
565 spin_unlock_irqrestore(lock, flags);
569 if (fatal_signal_pending(current)) {
570 cc->contended = true;
580 * Isolate free pages onto a private freelist. If @strict is true, will abort
581 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
582 * (even though it may still end up isolating some pages).
584 static unsigned long isolate_freepages_block(struct compact_control *cc,
585 unsigned long *start_pfn,
586 unsigned long end_pfn,
587 struct list_head *freelist,
591 int nr_scanned = 0, total_isolated = 0;
593 unsigned long flags = 0;
595 unsigned long blockpfn = *start_pfn;
598 /* Strict mode is for isolation, speed is secondary */
602 cursor = pfn_to_page(blockpfn);
604 /* Isolate free pages. */
605 for (; blockpfn < end_pfn; blockpfn += stride, cursor += stride) {
607 struct page *page = cursor;
610 * Periodically drop the lock (if held) regardless of its
611 * contention, to give chance to IRQs. Abort if fatal signal
614 if (!(blockpfn % COMPACT_CLUSTER_MAX)
615 && compact_unlock_should_abort(&cc->zone->lock, flags,
622 * For compound pages such as THP and hugetlbfs, we can save
623 * potentially a lot of iterations if we skip them at once.
624 * The check is racy, but we can consider only valid values
625 * and the only danger is skipping too much.
627 if (PageCompound(page)) {
628 const unsigned int order = compound_order(page);
630 if (likely(order <= MAX_ORDER)) {
631 blockpfn += (1UL << order) - 1;
632 cursor += (1UL << order) - 1;
633 nr_scanned += (1UL << order) - 1;
638 if (!PageBuddy(page))
641 /* If we already hold the lock, we can skip some rechecking. */
643 locked = compact_lock_irqsave(&cc->zone->lock,
646 /* Recheck this is a buddy page under lock */
647 if (!PageBuddy(page))
651 /* Found a free page, will break it into order-0 pages */
652 order = buddy_order(page);
653 isolated = __isolate_free_page(page, order);
656 set_page_private(page, order);
658 nr_scanned += isolated - 1;
659 total_isolated += isolated;
660 cc->nr_freepages += isolated;
661 list_add_tail(&page->lru, freelist);
663 if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
664 blockpfn += isolated;
667 /* Advance to the end of split page */
668 blockpfn += isolated - 1;
669 cursor += isolated - 1;
681 spin_unlock_irqrestore(&cc->zone->lock, flags);
684 * There is a tiny chance that we have read bogus compound_order(),
685 * so be careful to not go outside of the pageblock.
687 if (unlikely(blockpfn > end_pfn))
690 trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
691 nr_scanned, total_isolated);
693 /* Record how far we have got within the block */
694 *start_pfn = blockpfn;
697 * If strict isolation is requested by CMA then check that all the
698 * pages requested were isolated. If there were any failures, 0 is
699 * returned and CMA will fail.
701 if (strict && blockpfn < end_pfn)
704 cc->total_free_scanned += nr_scanned;
706 count_compact_events(COMPACTISOLATED, total_isolated);
707 return total_isolated;
711 * isolate_freepages_range() - isolate free pages.
712 * @cc: Compaction control structure.
713 * @start_pfn: The first PFN to start isolating.
714 * @end_pfn: The one-past-last PFN.
716 * Non-free pages, invalid PFNs, or zone boundaries within the
717 * [start_pfn, end_pfn) range are considered errors, cause function to
718 * undo its actions and return zero.
720 * Otherwise, function returns one-past-the-last PFN of isolated page
721 * (which may be greater then end_pfn if end fell in a middle of
725 isolate_freepages_range(struct compact_control *cc,
726 unsigned long start_pfn, unsigned long end_pfn)
728 unsigned long isolated, pfn, block_start_pfn, block_end_pfn;
732 block_start_pfn = pageblock_start_pfn(pfn);
733 if (block_start_pfn < cc->zone->zone_start_pfn)
734 block_start_pfn = cc->zone->zone_start_pfn;
735 block_end_pfn = pageblock_end_pfn(pfn);
737 for (; pfn < end_pfn; pfn += isolated,
738 block_start_pfn = block_end_pfn,
739 block_end_pfn += pageblock_nr_pages) {
740 /* Protect pfn from changing by isolate_freepages_block */
741 unsigned long isolate_start_pfn = pfn;
743 block_end_pfn = min(block_end_pfn, end_pfn);
746 * pfn could pass the block_end_pfn if isolated freepage
747 * is more than pageblock order. In this case, we adjust
748 * scanning range to right one.
750 if (pfn >= block_end_pfn) {
751 block_start_pfn = pageblock_start_pfn(pfn);
752 block_end_pfn = pageblock_end_pfn(pfn);
753 block_end_pfn = min(block_end_pfn, end_pfn);
756 if (!pageblock_pfn_to_page(block_start_pfn,
757 block_end_pfn, cc->zone))
760 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
761 block_end_pfn, &freelist, 0, true);
764 * In strict mode, isolate_freepages_block() returns 0 if
765 * there are any holes in the block (ie. invalid PFNs or
772 * If we managed to isolate pages, it is always (1 << n) *
773 * pageblock_nr_pages for some non-negative n. (Max order
774 * page may span two pageblocks).
778 /* __isolate_free_page() does not map the pages */
779 split_map_pages(&freelist);
782 /* Loop terminated early, cleanup. */
783 release_freepages(&freelist);
787 /* We don't use freelists for anything. */
791 /* Similar to reclaim, but different enough that they don't share logic */
792 static bool too_many_isolated(struct compact_control *cc)
794 pg_data_t *pgdat = cc->zone->zone_pgdat;
797 unsigned long active, inactive, isolated;
799 inactive = node_page_state(pgdat, NR_INACTIVE_FILE) +
800 node_page_state(pgdat, NR_INACTIVE_ANON);
801 active = node_page_state(pgdat, NR_ACTIVE_FILE) +
802 node_page_state(pgdat, NR_ACTIVE_ANON);
803 isolated = node_page_state(pgdat, NR_ISOLATED_FILE) +
804 node_page_state(pgdat, NR_ISOLATED_ANON);
807 * Allow GFP_NOFS to isolate past the limit set for regular
808 * compaction runs. This prevents an ABBA deadlock when other
809 * compactors have already isolated to the limit, but are
810 * blocked on filesystem locks held by the GFP_NOFS thread.
812 if (cc->gfp_mask & __GFP_FS) {
817 too_many = isolated > (inactive + active) / 2;
819 wake_throttle_isolated(pgdat);
825 * isolate_migratepages_block() - isolate all migrate-able pages within
827 * @cc: Compaction control structure.
828 * @low_pfn: The first PFN to isolate
829 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
830 * @mode: Isolation mode to be used.
832 * Isolate all pages that can be migrated from the range specified by
833 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
834 * Returns errno, like -EAGAIN or -EINTR in case e.g signal pending or congestion,
835 * -ENOMEM in case we could not allocate a page, or 0.
836 * cc->migrate_pfn will contain the next pfn to scan.
838 * The pages are isolated on cc->migratepages list (not required to be empty),
839 * and cc->nr_migratepages is updated accordingly.
842 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
843 unsigned long end_pfn, isolate_mode_t mode)
845 pg_data_t *pgdat = cc->zone->zone_pgdat;
846 unsigned long nr_scanned = 0, nr_isolated = 0;
847 struct lruvec *lruvec;
848 unsigned long flags = 0;
849 struct lruvec *locked = NULL;
850 struct folio *folio = NULL;
851 struct page *page = NULL, *valid_page = NULL;
852 struct address_space *mapping;
853 unsigned long start_pfn = low_pfn;
854 bool skip_on_failure = false;
855 unsigned long next_skip_pfn = 0;
856 bool skip_updated = false;
859 cc->migrate_pfn = low_pfn;
862 * Ensure that there are not too many pages isolated from the LRU
863 * list by either parallel reclaimers or compaction. If there are,
864 * delay for some time until fewer pages are isolated
866 while (unlikely(too_many_isolated(cc))) {
867 /* stop isolation if there are still pages not migrated */
868 if (cc->nr_migratepages)
871 /* async migration should just abort */
872 if (cc->mode == MIGRATE_ASYNC)
875 reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED);
877 if (fatal_signal_pending(current))
883 if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
884 skip_on_failure = true;
885 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
888 /* Time to isolate some pages for migration */
889 for (; low_pfn < end_pfn; low_pfn++) {
891 if (skip_on_failure && low_pfn >= next_skip_pfn) {
893 * We have isolated all migration candidates in the
894 * previous order-aligned block, and did not skip it due
895 * to failure. We should migrate the pages now and
896 * hopefully succeed compaction.
902 * We failed to isolate in the previous order-aligned
903 * block. Set the new boundary to the end of the
904 * current block. Note we can't simply increase
905 * next_skip_pfn by 1 << order, as low_pfn might have
906 * been incremented by a higher number due to skipping
907 * a compound or a high-order buddy page in the
908 * previous loop iteration.
910 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
914 * Periodically drop the lock (if held) regardless of its
915 * contention, to give chance to IRQs. Abort completely if
916 * a fatal signal is pending.
918 if (!(low_pfn % COMPACT_CLUSTER_MAX)) {
920 unlock_page_lruvec_irqrestore(locked, flags);
924 if (fatal_signal_pending(current)) {
925 cc->contended = true;
936 page = pfn_to_page(low_pfn);
939 * Check if the pageblock has already been marked skipped.
940 * Only the aligned PFN is checked as the caller isolates
941 * COMPACT_CLUSTER_MAX at a time so the second call must
942 * not falsely conclude that the block should be skipped.
944 if (!valid_page && pageblock_aligned(low_pfn)) {
945 if (!isolation_suitable(cc, page)) {
953 if (PageHuge(page) && cc->alloc_contig) {
955 unlock_page_lruvec_irqrestore(locked, flags);
959 ret = isolate_or_dissolve_huge_page(page, &cc->migratepages);
962 * Fail isolation in case isolate_or_dissolve_huge_page()
963 * reports an error. In case of -ENOMEM, abort right away.
966 /* Do not report -EBUSY down the chain */
969 low_pfn += compound_nr(page) - 1;
970 nr_scanned += compound_nr(page) - 1;
974 if (PageHuge(page)) {
976 * Hugepage was successfully isolated and placed
977 * on the cc->migratepages list.
979 folio = page_folio(page);
980 low_pfn += folio_nr_pages(folio) - 1;
981 goto isolate_success_no_list;
985 * Ok, the hugepage was dissolved. Now these pages are
986 * Buddy and cannot be re-allocated because they are
987 * isolated. Fall-through as the check below handles
993 * Skip if free. We read page order here without zone lock
994 * which is generally unsafe, but the race window is small and
995 * the worst thing that can happen is that we skip some
996 * potential isolation targets.
998 if (PageBuddy(page)) {
999 unsigned long freepage_order = buddy_order_unsafe(page);
1002 * Without lock, we cannot be sure that what we got is
1003 * a valid page order. Consider only values in the
1004 * valid order range to prevent low_pfn overflow.
1006 if (freepage_order > 0 && freepage_order <= MAX_ORDER) {
1007 low_pfn += (1UL << freepage_order) - 1;
1008 nr_scanned += (1UL << freepage_order) - 1;
1014 * Regardless of being on LRU, compound pages such as THP and
1015 * hugetlbfs are not to be compacted unless we are attempting
1016 * an allocation much larger than the huge page size (eg CMA).
1017 * We can potentially save a lot of iterations if we skip them
1018 * at once. The check is racy, but we can consider only valid
1019 * values and the only danger is skipping too much.
1021 if (PageCompound(page) && !cc->alloc_contig) {
1022 const unsigned int order = compound_order(page);
1024 if (likely(order <= MAX_ORDER)) {
1025 low_pfn += (1UL << order) - 1;
1026 nr_scanned += (1UL << order) - 1;
1032 * Check may be lockless but that's ok as we recheck later.
1033 * It's possible to migrate LRU and non-lru movable pages.
1034 * Skip any other type of page
1036 if (!PageLRU(page)) {
1038 * __PageMovable can return false positive so we need
1039 * to verify it under page_lock.
1041 if (unlikely(__PageMovable(page)) &&
1042 !PageIsolated(page)) {
1044 unlock_page_lruvec_irqrestore(locked, flags);
1048 if (isolate_movable_page(page, mode)) {
1049 folio = page_folio(page);
1050 goto isolate_success;
1058 * Be careful not to clear PageLRU until after we're
1059 * sure the page is not being freed elsewhere -- the
1060 * page release code relies on it.
1062 folio = folio_get_nontail_page(page);
1063 if (unlikely(!folio))
1067 * Migration will fail if an anonymous page is pinned in memory,
1068 * so avoid taking lru_lock and isolating it unnecessarily in an
1069 * admittedly racy check.
1071 mapping = folio_mapping(folio);
1072 if (!mapping && (folio_ref_count(folio) - 1) > folio_mapcount(folio))
1073 goto isolate_fail_put;
1076 * Only allow to migrate anonymous pages in GFP_NOFS context
1077 * because those do not depend on fs locks.
1079 if (!(cc->gfp_mask & __GFP_FS) && mapping)
1080 goto isolate_fail_put;
1082 /* Only take pages on LRU: a check now makes later tests safe */
1083 if (!folio_test_lru(folio))
1084 goto isolate_fail_put;
1086 /* Compaction might skip unevictable pages but CMA takes them */
1087 if (!(mode & ISOLATE_UNEVICTABLE) && folio_test_unevictable(folio))
1088 goto isolate_fail_put;
1091 * To minimise LRU disruption, the caller can indicate with
1092 * ISOLATE_ASYNC_MIGRATE that it only wants to isolate pages
1093 * it will be able to migrate without blocking - clean pages
1094 * for the most part. PageWriteback would require blocking.
1096 if ((mode & ISOLATE_ASYNC_MIGRATE) && folio_test_writeback(folio))
1097 goto isolate_fail_put;
1099 if ((mode & ISOLATE_ASYNC_MIGRATE) && folio_test_dirty(folio)) {
1103 * Only pages without mappings or that have a
1104 * ->migrate_folio callback are possible to migrate
1105 * without blocking. However, we can be racing with
1106 * truncation so it's necessary to lock the page
1107 * to stabilise the mapping as truncation holds
1108 * the page lock until after the page is removed
1109 * from the page cache.
1111 if (!folio_trylock(folio))
1112 goto isolate_fail_put;
1114 mapping = folio_mapping(folio);
1115 migrate_dirty = !mapping ||
1116 mapping->a_ops->migrate_folio;
1117 folio_unlock(folio);
1119 goto isolate_fail_put;
1122 /* Try isolate the folio */
1123 if (!folio_test_clear_lru(folio))
1124 goto isolate_fail_put;
1126 lruvec = folio_lruvec(folio);
1128 /* If we already hold the lock, we can skip some rechecking */
1129 if (lruvec != locked) {
1131 unlock_page_lruvec_irqrestore(locked, flags);
1133 compact_lock_irqsave(&lruvec->lru_lock, &flags, cc);
1136 lruvec_memcg_debug(lruvec, folio);
1139 * Try get exclusive access under lock. If marked for
1140 * skip, the scan is aborted unless the current context
1141 * is a rescan to reach the end of the pageblock.
1143 if (!skip_updated && valid_page) {
1144 skip_updated = true;
1145 if (test_and_set_skip(cc, valid_page) &&
1146 !cc->finish_pageblock) {
1152 * folio become large since the non-locked check,
1155 if (unlikely(folio_test_large(folio) && !cc->alloc_contig)) {
1156 low_pfn += folio_nr_pages(folio) - 1;
1157 nr_scanned += folio_nr_pages(folio) - 1;
1158 folio_set_lru(folio);
1159 goto isolate_fail_put;
1163 /* The folio is taken off the LRU */
1164 if (folio_test_large(folio))
1165 low_pfn += folio_nr_pages(folio) - 1;
1167 /* Successfully isolated */
1168 lruvec_del_folio(lruvec, folio);
1169 node_stat_mod_folio(folio,
1170 NR_ISOLATED_ANON + folio_is_file_lru(folio),
1171 folio_nr_pages(folio));
1174 list_add(&folio->lru, &cc->migratepages);
1175 isolate_success_no_list:
1176 cc->nr_migratepages += folio_nr_pages(folio);
1177 nr_isolated += folio_nr_pages(folio);
1178 nr_scanned += folio_nr_pages(folio) - 1;
1181 * Avoid isolating too much unless this block is being
1182 * fully scanned (e.g. dirty/writeback pages, parallel allocation)
1183 * or a lock is contended. For contention, isolate quickly to
1184 * potentially remove one source of contention.
1186 if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX &&
1187 !cc->finish_pageblock && !cc->contended) {
1195 /* Avoid potential deadlock in freeing page under lru_lock */
1197 unlock_page_lruvec_irqrestore(locked, flags);
1203 if (!skip_on_failure && ret != -ENOMEM)
1207 * We have isolated some pages, but then failed. Release them
1208 * instead of migrating, as we cannot form the cc->order buddy
1213 unlock_page_lruvec_irqrestore(locked, flags);
1216 putback_movable_pages(&cc->migratepages);
1217 cc->nr_migratepages = 0;
1221 if (low_pfn < next_skip_pfn) {
1222 low_pfn = next_skip_pfn - 1;
1224 * The check near the loop beginning would have updated
1225 * next_skip_pfn too, but this is a bit simpler.
1227 next_skip_pfn += 1UL << cc->order;
1235 * The PageBuddy() check could have potentially brought us outside
1236 * the range to be scanned.
1238 if (unlikely(low_pfn > end_pfn))
1245 unlock_page_lruvec_irqrestore(locked, flags);
1247 folio_set_lru(folio);
1252 * Update the cached scanner pfn once the pageblock has been scanned.
1253 * Pages will either be migrated in which case there is no point
1254 * scanning in the near future or migration failed in which case the
1255 * failure reason may persist. The block is marked for skipping if
1256 * there were no pages isolated in the block or if the block is
1257 * rescanned twice in a row.
1259 if (low_pfn == end_pfn && (!nr_isolated || cc->finish_pageblock)) {
1260 if (!cc->no_set_skip_hint && valid_page && !skip_updated)
1261 set_pageblock_skip(valid_page);
1262 update_cached_migrate(cc, low_pfn);
1265 trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
1266 nr_scanned, nr_isolated);
1269 cc->total_migrate_scanned += nr_scanned;
1271 count_compact_events(COMPACTISOLATED, nr_isolated);
1273 cc->migrate_pfn = low_pfn;
1279 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
1280 * @cc: Compaction control structure.
1281 * @start_pfn: The first PFN to start isolating.
1282 * @end_pfn: The one-past-last PFN.
1284 * Returns -EAGAIN when contented, -EINTR in case of a signal pending, -ENOMEM
1285 * in case we could not allocate a page, or 0.
1288 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
1289 unsigned long end_pfn)
1291 unsigned long pfn, block_start_pfn, block_end_pfn;
1294 /* Scan block by block. First and last block may be incomplete */
1296 block_start_pfn = pageblock_start_pfn(pfn);
1297 if (block_start_pfn < cc->zone->zone_start_pfn)
1298 block_start_pfn = cc->zone->zone_start_pfn;
1299 block_end_pfn = pageblock_end_pfn(pfn);
1301 for (; pfn < end_pfn; pfn = block_end_pfn,
1302 block_start_pfn = block_end_pfn,
1303 block_end_pfn += pageblock_nr_pages) {
1305 block_end_pfn = min(block_end_pfn, end_pfn);
1307 if (!pageblock_pfn_to_page(block_start_pfn,
1308 block_end_pfn, cc->zone))
1311 ret = isolate_migratepages_block(cc, pfn, block_end_pfn,
1312 ISOLATE_UNEVICTABLE);
1317 if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX)
1324 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
1325 #ifdef CONFIG_COMPACTION
1327 static bool suitable_migration_source(struct compact_control *cc,
1332 if (pageblock_skip_persistent(page))
1335 if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction)
1338 block_mt = get_pageblock_migratetype(page);
1340 if (cc->migratetype == MIGRATE_MOVABLE)
1341 return is_migrate_movable(block_mt);
1343 return block_mt == cc->migratetype;
1346 /* Returns true if the page is within a block suitable for migration to */
1347 static bool suitable_migration_target(struct compact_control *cc,
1350 /* If the page is a large free page, then disallow migration */
1351 if (PageBuddy(page)) {
1353 * We are checking page_order without zone->lock taken. But
1354 * the only small danger is that we skip a potentially suitable
1355 * pageblock, so it's not worth to check order for valid range.
1357 if (buddy_order_unsafe(page) >= pageblock_order)
1361 if (cc->ignore_block_suitable)
1364 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1365 if (is_migrate_movable(get_pageblock_migratetype(page)))
1368 /* Otherwise skip the block */
1372 static inline unsigned int
1373 freelist_scan_limit(struct compact_control *cc)
1375 unsigned short shift = BITS_PER_LONG - 1;
1377 return (COMPACT_CLUSTER_MAX >> min(shift, cc->fast_search_fail)) + 1;
1381 * Test whether the free scanner has reached the same or lower pageblock than
1382 * the migration scanner, and compaction should thus terminate.
1384 static inline bool compact_scanners_met(struct compact_control *cc)
1386 return (cc->free_pfn >> pageblock_order)
1387 <= (cc->migrate_pfn >> pageblock_order);
1391 * Used when scanning for a suitable migration target which scans freelists
1392 * in reverse. Reorders the list such as the unscanned pages are scanned
1393 * first on the next iteration of the free scanner
1396 move_freelist_head(struct list_head *freelist, struct page *freepage)
1400 if (!list_is_last(freelist, &freepage->lru)) {
1401 list_cut_before(&sublist, freelist, &freepage->lru);
1402 list_splice_tail(&sublist, freelist);
1407 * Similar to move_freelist_head except used by the migration scanner
1408 * when scanning forward. It's possible for these list operations to
1409 * move against each other if they search the free list exactly in
1413 move_freelist_tail(struct list_head *freelist, struct page *freepage)
1417 if (!list_is_first(freelist, &freepage->lru)) {
1418 list_cut_position(&sublist, freelist, &freepage->lru);
1419 list_splice_tail(&sublist, freelist);
1424 fast_isolate_around(struct compact_control *cc, unsigned long pfn)
1426 unsigned long start_pfn, end_pfn;
1429 /* Do not search around if there are enough pages already */
1430 if (cc->nr_freepages >= cc->nr_migratepages)
1433 /* Minimise scanning during async compaction */
1434 if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC)
1437 /* Pageblock boundaries */
1438 start_pfn = max(pageblock_start_pfn(pfn), cc->zone->zone_start_pfn);
1439 end_pfn = min(pageblock_end_pfn(pfn), zone_end_pfn(cc->zone));
1441 page = pageblock_pfn_to_page(start_pfn, end_pfn, cc->zone);
1445 isolate_freepages_block(cc, &start_pfn, end_pfn, &cc->freepages, 1, false);
1447 /* Skip this pageblock in the future as it's full or nearly full */
1448 if (start_pfn == end_pfn)
1449 set_pageblock_skip(page);
1454 /* Search orders in round-robin fashion */
1455 static int next_search_order(struct compact_control *cc, int order)
1459 order = cc->order - 1;
1461 /* Search wrapped around? */
1462 if (order == cc->search_order) {
1464 if (cc->search_order < 0)
1465 cc->search_order = cc->order - 1;
1472 static void fast_isolate_freepages(struct compact_control *cc)
1474 unsigned int limit = max(1U, freelist_scan_limit(cc) >> 1);
1475 unsigned int nr_scanned = 0, total_isolated = 0;
1476 unsigned long low_pfn, min_pfn, highest = 0;
1477 unsigned long nr_isolated = 0;
1478 unsigned long distance;
1479 struct page *page = NULL;
1480 bool scan_start = false;
1483 /* Full compaction passes in a negative order */
1488 * If starting the scan, use a deeper search and use the highest
1489 * PFN found if a suitable one is not found.
1491 if (cc->free_pfn >= cc->zone->compact_init_free_pfn) {
1492 limit = pageblock_nr_pages >> 1;
1497 * Preferred point is in the top quarter of the scan space but take
1498 * a pfn from the top half if the search is problematic.
1500 distance = (cc->free_pfn - cc->migrate_pfn);
1501 low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2));
1502 min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1));
1504 if (WARN_ON_ONCE(min_pfn > low_pfn))
1508 * Search starts from the last successful isolation order or the next
1509 * order to search after a previous failure
1511 cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order);
1513 for (order = cc->search_order;
1514 !page && order >= 0;
1515 order = next_search_order(cc, order)) {
1516 struct free_area *area = &cc->zone->free_area[order];
1517 struct list_head *freelist;
1518 struct page *freepage;
1519 unsigned long flags;
1520 unsigned int order_scanned = 0;
1521 unsigned long high_pfn = 0;
1526 spin_lock_irqsave(&cc->zone->lock, flags);
1527 freelist = &area->free_list[MIGRATE_MOVABLE];
1528 list_for_each_entry_reverse(freepage, freelist, buddy_list) {
1533 pfn = page_to_pfn(freepage);
1536 highest = max(pageblock_start_pfn(pfn),
1537 cc->zone->zone_start_pfn);
1539 if (pfn >= low_pfn) {
1540 cc->fast_search_fail = 0;
1541 cc->search_order = order;
1546 if (pfn >= min_pfn && pfn > high_pfn) {
1549 /* Shorten the scan if a candidate is found */
1553 if (order_scanned >= limit)
1557 /* Use a minimum pfn if a preferred one was not found */
1558 if (!page && high_pfn) {
1559 page = pfn_to_page(high_pfn);
1561 /* Update freepage for the list reorder below */
1565 /* Reorder to so a future search skips recent pages */
1566 move_freelist_head(freelist, freepage);
1568 /* Isolate the page if available */
1570 if (__isolate_free_page(page, order)) {
1571 set_page_private(page, order);
1572 nr_isolated = 1 << order;
1573 nr_scanned += nr_isolated - 1;
1574 total_isolated += nr_isolated;
1575 cc->nr_freepages += nr_isolated;
1576 list_add_tail(&page->lru, &cc->freepages);
1577 count_compact_events(COMPACTISOLATED, nr_isolated);
1579 /* If isolation fails, abort the search */
1580 order = cc->search_order + 1;
1585 spin_unlock_irqrestore(&cc->zone->lock, flags);
1587 /* Skip fast search if enough freepages isolated */
1588 if (cc->nr_freepages >= cc->nr_migratepages)
1592 * Smaller scan on next order so the total scan is related
1593 * to freelist_scan_limit.
1595 if (order_scanned >= limit)
1596 limit = max(1U, limit >> 1);
1599 trace_mm_compaction_fast_isolate_freepages(min_pfn, cc->free_pfn,
1600 nr_scanned, total_isolated);
1603 cc->fast_search_fail++;
1606 * Use the highest PFN found above min. If one was
1607 * not found, be pessimistic for direct compaction
1608 * and use the min mark.
1610 if (highest >= min_pfn) {
1611 page = pfn_to_page(highest);
1612 cc->free_pfn = highest;
1614 if (cc->direct_compaction && pfn_valid(min_pfn)) {
1615 page = pageblock_pfn_to_page(min_pfn,
1616 min(pageblock_end_pfn(min_pfn),
1617 zone_end_pfn(cc->zone)),
1619 cc->free_pfn = min_pfn;
1625 if (highest && highest >= cc->zone->compact_cached_free_pfn) {
1626 highest -= pageblock_nr_pages;
1627 cc->zone->compact_cached_free_pfn = highest;
1630 cc->total_free_scanned += nr_scanned;
1634 low_pfn = page_to_pfn(page);
1635 fast_isolate_around(cc, low_pfn);
1639 * Based on information in the current compact_control, find blocks
1640 * suitable for isolating free pages from and then isolate them.
1642 static void isolate_freepages(struct compact_control *cc)
1644 struct zone *zone = cc->zone;
1646 unsigned long block_start_pfn; /* start of current pageblock */
1647 unsigned long isolate_start_pfn; /* exact pfn we start at */
1648 unsigned long block_end_pfn; /* end of current pageblock */
1649 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
1650 struct list_head *freelist = &cc->freepages;
1651 unsigned int stride;
1653 /* Try a small search of the free lists for a candidate */
1654 fast_isolate_freepages(cc);
1655 if (cc->nr_freepages)
1659 * Initialise the free scanner. The starting point is where we last
1660 * successfully isolated from, zone-cached value, or the end of the
1661 * zone when isolating for the first time. For looping we also need
1662 * this pfn aligned down to the pageblock boundary, because we do
1663 * block_start_pfn -= pageblock_nr_pages in the for loop.
1664 * For ending point, take care when isolating in last pageblock of a
1665 * zone which ends in the middle of a pageblock.
1666 * The low boundary is the end of the pageblock the migration scanner
1669 isolate_start_pfn = cc->free_pfn;
1670 block_start_pfn = pageblock_start_pfn(isolate_start_pfn);
1671 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1672 zone_end_pfn(zone));
1673 low_pfn = pageblock_end_pfn(cc->migrate_pfn);
1674 stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1;
1677 * Isolate free pages until enough are available to migrate the
1678 * pages on cc->migratepages. We stop searching if the migrate
1679 * and free page scanners meet or enough free pages are isolated.
1681 for (; block_start_pfn >= low_pfn;
1682 block_end_pfn = block_start_pfn,
1683 block_start_pfn -= pageblock_nr_pages,
1684 isolate_start_pfn = block_start_pfn) {
1685 unsigned long nr_isolated;
1688 * This can iterate a massively long zone without finding any
1689 * suitable migration targets, so periodically check resched.
1691 if (!(block_start_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages)))
1694 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1697 unsigned long next_pfn;
1699 next_pfn = skip_offline_sections_reverse(block_start_pfn);
1701 block_start_pfn = max(next_pfn, low_pfn);
1706 /* Check the block is suitable for migration */
1707 if (!suitable_migration_target(cc, page))
1710 /* If isolation recently failed, do not retry */
1711 if (!isolation_suitable(cc, page))
1714 /* Found a block suitable for isolating free pages from. */
1715 nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn,
1716 block_end_pfn, freelist, stride, false);
1718 /* Update the skip hint if the full pageblock was scanned */
1719 if (isolate_start_pfn == block_end_pfn)
1720 update_pageblock_skip(cc, page, block_start_pfn);
1722 /* Are enough freepages isolated? */
1723 if (cc->nr_freepages >= cc->nr_migratepages) {
1724 if (isolate_start_pfn >= block_end_pfn) {
1726 * Restart at previous pageblock if more
1727 * freepages can be isolated next time.
1730 block_start_pfn - pageblock_nr_pages;
1733 } else if (isolate_start_pfn < block_end_pfn) {
1735 * If isolation failed early, do not continue
1741 /* Adjust stride depending on isolation */
1746 stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1);
1750 * Record where the free scanner will restart next time. Either we
1751 * broke from the loop and set isolate_start_pfn based on the last
1752 * call to isolate_freepages_block(), or we met the migration scanner
1753 * and the loop terminated due to isolate_start_pfn < low_pfn
1755 cc->free_pfn = isolate_start_pfn;
1758 /* __isolate_free_page() does not map the pages */
1759 split_map_pages(freelist);
1763 * This is a migrate-callback that "allocates" freepages by taking pages
1764 * from the isolated freelists in the block we are migrating to.
1766 static struct folio *compaction_alloc(struct folio *src, unsigned long data)
1768 struct compact_control *cc = (struct compact_control *)data;
1771 if (list_empty(&cc->freepages)) {
1772 isolate_freepages(cc);
1774 if (list_empty(&cc->freepages))
1778 dst = list_entry(cc->freepages.next, struct folio, lru);
1779 list_del(&dst->lru);
1786 * This is a migrate-callback that "frees" freepages back to the isolated
1787 * freelist. All pages on the freelist are from the same zone, so there is no
1788 * special handling needed for NUMA.
1790 static void compaction_free(struct folio *dst, unsigned long data)
1792 struct compact_control *cc = (struct compact_control *)data;
1794 list_add(&dst->lru, &cc->freepages);
1798 /* possible outcome of isolate_migratepages */
1800 ISOLATE_ABORT, /* Abort compaction now */
1801 ISOLATE_NONE, /* No pages isolated, continue scanning */
1802 ISOLATE_SUCCESS, /* Pages isolated, migrate */
1803 } isolate_migrate_t;
1806 * Allow userspace to control policy on scanning the unevictable LRU for
1807 * compactable pages.
1809 static int sysctl_compact_unevictable_allowed __read_mostly = CONFIG_COMPACT_UNEVICTABLE_DEFAULT;
1811 * Tunable for proactive compaction. It determines how
1812 * aggressively the kernel should compact memory in the
1813 * background. It takes values in the range [0, 100].
1815 static unsigned int __read_mostly sysctl_compaction_proactiveness = 20;
1816 static int sysctl_extfrag_threshold = 500;
1817 static int __read_mostly sysctl_compact_memory;
1820 update_fast_start_pfn(struct compact_control *cc, unsigned long pfn)
1822 if (cc->fast_start_pfn == ULONG_MAX)
1825 if (!cc->fast_start_pfn)
1826 cc->fast_start_pfn = pfn;
1828 cc->fast_start_pfn = min(cc->fast_start_pfn, pfn);
1831 static inline unsigned long
1832 reinit_migrate_pfn(struct compact_control *cc)
1834 if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX)
1835 return cc->migrate_pfn;
1837 cc->migrate_pfn = cc->fast_start_pfn;
1838 cc->fast_start_pfn = ULONG_MAX;
1840 return cc->migrate_pfn;
1844 * Briefly search the free lists for a migration source that already has
1845 * some free pages to reduce the number of pages that need migration
1846 * before a pageblock is free.
1848 static unsigned long fast_find_migrateblock(struct compact_control *cc)
1850 unsigned int limit = freelist_scan_limit(cc);
1851 unsigned int nr_scanned = 0;
1852 unsigned long distance;
1853 unsigned long pfn = cc->migrate_pfn;
1854 unsigned long high_pfn;
1856 bool found_block = false;
1858 /* Skip hints are relied on to avoid repeats on the fast search */
1859 if (cc->ignore_skip_hint)
1863 * If the pageblock should be finished then do not select a different
1866 if (cc->finish_pageblock)
1870 * If the migrate_pfn is not at the start of a zone or the start
1871 * of a pageblock then assume this is a continuation of a previous
1872 * scan restarted due to COMPACT_CLUSTER_MAX.
1874 if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn))
1878 * For smaller orders, just linearly scan as the number of pages
1879 * to migrate should be relatively small and does not necessarily
1880 * justify freeing up a large block for a small allocation.
1882 if (cc->order <= PAGE_ALLOC_COSTLY_ORDER)
1886 * Only allow kcompactd and direct requests for movable pages to
1887 * quickly clear out a MOVABLE pageblock for allocation. This
1888 * reduces the risk that a large movable pageblock is freed for
1889 * an unmovable/reclaimable small allocation.
1891 if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE)
1895 * When starting the migration scanner, pick any pageblock within the
1896 * first half of the search space. Otherwise try and pick a pageblock
1897 * within the first eighth to reduce the chances that a migration
1898 * target later becomes a source.
1900 distance = (cc->free_pfn - cc->migrate_pfn) >> 1;
1901 if (cc->migrate_pfn != cc->zone->zone_start_pfn)
1903 high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance);
1905 for (order = cc->order - 1;
1906 order >= PAGE_ALLOC_COSTLY_ORDER && !found_block && nr_scanned < limit;
1908 struct free_area *area = &cc->zone->free_area[order];
1909 struct list_head *freelist;
1910 unsigned long flags;
1911 struct page *freepage;
1916 spin_lock_irqsave(&cc->zone->lock, flags);
1917 freelist = &area->free_list[MIGRATE_MOVABLE];
1918 list_for_each_entry(freepage, freelist, buddy_list) {
1919 unsigned long free_pfn;
1921 if (nr_scanned++ >= limit) {
1922 move_freelist_tail(freelist, freepage);
1926 free_pfn = page_to_pfn(freepage);
1927 if (free_pfn < high_pfn) {
1929 * Avoid if skipped recently. Ideally it would
1930 * move to the tail but even safe iteration of
1931 * the list assumes an entry is deleted, not
1934 if (get_pageblock_skip(freepage))
1937 /* Reorder to so a future search skips recent pages */
1938 move_freelist_tail(freelist, freepage);
1940 update_fast_start_pfn(cc, free_pfn);
1941 pfn = pageblock_start_pfn(free_pfn);
1942 if (pfn < cc->zone->zone_start_pfn)
1943 pfn = cc->zone->zone_start_pfn;
1944 cc->fast_search_fail = 0;
1949 spin_unlock_irqrestore(&cc->zone->lock, flags);
1952 cc->total_migrate_scanned += nr_scanned;
1955 * If fast scanning failed then use a cached entry for a page block
1956 * that had free pages as the basis for starting a linear scan.
1959 cc->fast_search_fail++;
1960 pfn = reinit_migrate_pfn(cc);
1966 * Isolate all pages that can be migrated from the first suitable block,
1967 * starting at the block pointed to by the migrate scanner pfn within
1970 static isolate_migrate_t isolate_migratepages(struct compact_control *cc)
1972 unsigned long block_start_pfn;
1973 unsigned long block_end_pfn;
1974 unsigned long low_pfn;
1976 const isolate_mode_t isolate_mode =
1977 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1978 (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1979 bool fast_find_block;
1982 * Start at where we last stopped, or beginning of the zone as
1983 * initialized by compact_zone(). The first failure will use
1984 * the lowest PFN as the starting point for linear scanning.
1986 low_pfn = fast_find_migrateblock(cc);
1987 block_start_pfn = pageblock_start_pfn(low_pfn);
1988 if (block_start_pfn < cc->zone->zone_start_pfn)
1989 block_start_pfn = cc->zone->zone_start_pfn;
1992 * fast_find_migrateblock marks a pageblock skipped so to avoid
1993 * the isolation_suitable check below, check whether the fast
1994 * search was successful.
1996 fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail;
1998 /* Only scan within a pageblock boundary */
1999 block_end_pfn = pageblock_end_pfn(low_pfn);
2002 * Iterate over whole pageblocks until we find the first suitable.
2003 * Do not cross the free scanner.
2005 for (; block_end_pfn <= cc->free_pfn;
2006 fast_find_block = false,
2007 cc->migrate_pfn = low_pfn = block_end_pfn,
2008 block_start_pfn = block_end_pfn,
2009 block_end_pfn += pageblock_nr_pages) {
2012 * This can potentially iterate a massively long zone with
2013 * many pageblocks unsuitable, so periodically check if we
2016 if (!(low_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages)))
2019 page = pageblock_pfn_to_page(block_start_pfn,
2020 block_end_pfn, cc->zone);
2022 unsigned long next_pfn;
2024 next_pfn = skip_offline_sections(block_start_pfn);
2026 block_end_pfn = min(next_pfn, cc->free_pfn);
2031 * If isolation recently failed, do not retry. Only check the
2032 * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock
2033 * to be visited multiple times. Assume skip was checked
2034 * before making it "skip" so other compaction instances do
2035 * not scan the same block.
2037 if (pageblock_aligned(low_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 via
2072 * /proc/sys/vm/compact_memory
2074 static inline bool is_via_compact_memory(int order)
2080 * Determine whether kswapd is (or recently was!) running on this node.
2082 * pgdat_kswapd_lock() pins pgdat->kswapd, so a concurrent kswapd_stop() can't
2085 static bool kswapd_is_running(pg_data_t *pgdat)
2089 pgdat_kswapd_lock(pgdat);
2090 running = pgdat->kswapd && task_is_running(pgdat->kswapd);
2091 pgdat_kswapd_unlock(pgdat);
2097 * A zone's fragmentation score is the external fragmentation wrt to the
2098 * COMPACTION_HPAGE_ORDER. It returns a value in the range [0, 100].
2100 static unsigned int fragmentation_score_zone(struct zone *zone)
2102 return extfrag_for_order(zone, COMPACTION_HPAGE_ORDER);
2106 * A weighted zone's fragmentation score is the external fragmentation
2107 * wrt to the COMPACTION_HPAGE_ORDER scaled by the zone's size. It
2108 * returns a value in the range [0, 100].
2110 * The scaling factor ensures that proactive compaction focuses on larger
2111 * zones like ZONE_NORMAL, rather than smaller, specialized zones like
2112 * ZONE_DMA32. For smaller zones, the score value remains close to zero,
2113 * and thus never exceeds the high threshold for proactive compaction.
2115 static unsigned int fragmentation_score_zone_weighted(struct zone *zone)
2117 unsigned long score;
2119 score = zone->present_pages * fragmentation_score_zone(zone);
2120 return div64_ul(score, zone->zone_pgdat->node_present_pages + 1);
2124 * The per-node proactive (background) compaction process is started by its
2125 * corresponding kcompactd thread when the node's fragmentation score
2126 * exceeds the high threshold. The compaction process remains active till
2127 * the node's score falls below the low threshold, or one of the back-off
2128 * conditions is met.
2130 static unsigned int fragmentation_score_node(pg_data_t *pgdat)
2132 unsigned int score = 0;
2135 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2138 zone = &pgdat->node_zones[zoneid];
2139 if (!populated_zone(zone))
2141 score += fragmentation_score_zone_weighted(zone);
2147 static unsigned int fragmentation_score_wmark(pg_data_t *pgdat, bool low)
2149 unsigned int wmark_low;
2152 * Cap the low watermark to avoid excessive compaction
2153 * activity in case a user sets the proactiveness tunable
2154 * close to 100 (maximum).
2156 wmark_low = max(100U - sysctl_compaction_proactiveness, 5U);
2157 return low ? wmark_low : min(wmark_low + 10, 100U);
2160 static bool should_proactive_compact_node(pg_data_t *pgdat)
2164 if (!sysctl_compaction_proactiveness || kswapd_is_running(pgdat))
2167 wmark_high = fragmentation_score_wmark(pgdat, false);
2168 return fragmentation_score_node(pgdat) > wmark_high;
2171 static enum compact_result __compact_finished(struct compact_control *cc)
2174 const int migratetype = cc->migratetype;
2177 /* Compaction run completes if the migrate and free scanner meet */
2178 if (compact_scanners_met(cc)) {
2179 /* Let the next compaction start anew. */
2180 reset_cached_positions(cc->zone);
2183 * Mark that the PG_migrate_skip information should be cleared
2184 * by kswapd when it goes to sleep. kcompactd does not set the
2185 * flag itself as the decision to be clear should be directly
2186 * based on an allocation request.
2188 if (cc->direct_compaction)
2189 cc->zone->compact_blockskip_flush = true;
2192 return COMPACT_COMPLETE;
2194 return COMPACT_PARTIAL_SKIPPED;
2197 if (cc->proactive_compaction) {
2198 int score, wmark_low;
2201 pgdat = cc->zone->zone_pgdat;
2202 if (kswapd_is_running(pgdat))
2203 return COMPACT_PARTIAL_SKIPPED;
2205 score = fragmentation_score_zone(cc->zone);
2206 wmark_low = fragmentation_score_wmark(pgdat, true);
2208 if (score > wmark_low)
2209 ret = COMPACT_CONTINUE;
2211 ret = COMPACT_SUCCESS;
2216 if (is_via_compact_memory(cc->order))
2217 return COMPACT_CONTINUE;
2220 * Always finish scanning a pageblock to reduce the possibility of
2221 * fallbacks in the future. This is particularly important when
2222 * migration source is unmovable/reclaimable but it's not worth
2225 if (!pageblock_aligned(cc->migrate_pfn))
2226 return COMPACT_CONTINUE;
2228 /* Direct compactor: Is a suitable page free? */
2229 ret = COMPACT_NO_SUITABLE_PAGE;
2230 for (order = cc->order; order <= MAX_ORDER; order++) {
2231 struct free_area *area = &cc->zone->free_area[order];
2234 /* Job done if page is free of the right migratetype */
2235 if (!free_area_empty(area, migratetype))
2236 return COMPACT_SUCCESS;
2239 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
2240 if (migratetype == MIGRATE_MOVABLE &&
2241 !free_area_empty(area, MIGRATE_CMA))
2242 return COMPACT_SUCCESS;
2245 * Job done if allocation would steal freepages from
2246 * other migratetype buddy lists.
2248 if (find_suitable_fallback(area, order, migratetype,
2249 true, &can_steal) != -1)
2251 * Movable pages are OK in any pageblock. If we are
2252 * stealing for a non-movable allocation, make sure
2253 * we finish compacting the current pageblock first
2254 * (which is assured by the above migrate_pfn align
2255 * check) so it is as free as possible and we won't
2256 * have to steal another one soon.
2258 return COMPACT_SUCCESS;
2262 if (cc->contended || fatal_signal_pending(current))
2263 ret = COMPACT_CONTENDED;
2268 static enum compact_result compact_finished(struct compact_control *cc)
2272 ret = __compact_finished(cc);
2273 trace_mm_compaction_finished(cc->zone, cc->order, ret);
2274 if (ret == COMPACT_NO_SUITABLE_PAGE)
2275 ret = COMPACT_CONTINUE;
2280 static bool __compaction_suitable(struct zone *zone, int order,
2281 int highest_zoneidx,
2282 unsigned long wmark_target)
2284 unsigned long watermark;
2286 * Watermarks for order-0 must be met for compaction to be able to
2287 * isolate free pages for migration targets. This means that the
2288 * watermark and alloc_flags have to match, or be more pessimistic than
2289 * the check in __isolate_free_page(). We don't use the direct
2290 * compactor's alloc_flags, as they are not relevant for freepage
2291 * isolation. We however do use the direct compactor's highest_zoneidx
2292 * to skip over zones where lowmem reserves would prevent allocation
2293 * even if compaction succeeds.
2294 * For costly orders, we require low watermark instead of min for
2295 * compaction to proceed to increase its chances.
2296 * ALLOC_CMA is used, as pages in CMA pageblocks are considered
2297 * suitable migration targets
2299 watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
2300 low_wmark_pages(zone) : min_wmark_pages(zone);
2301 watermark += compact_gap(order);
2302 return __zone_watermark_ok(zone, 0, watermark, highest_zoneidx,
2303 ALLOC_CMA, wmark_target);
2307 * compaction_suitable: Is this suitable to run compaction on this zone now?
2309 bool compaction_suitable(struct zone *zone, int order, int highest_zoneidx)
2311 enum compact_result compact_result;
2314 suitable = __compaction_suitable(zone, order, highest_zoneidx,
2315 zone_page_state(zone, NR_FREE_PAGES));
2317 * fragmentation index determines if allocation failures are due to
2318 * low memory or external fragmentation
2320 * index of -1000 would imply allocations might succeed depending on
2321 * watermarks, but we already failed the high-order watermark check
2322 * index towards 0 implies failure is due to lack of memory
2323 * index towards 1000 implies failure is due to fragmentation
2325 * Only compact if a failure would be due to fragmentation. Also
2326 * ignore fragindex for non-costly orders where the alternative to
2327 * a successful reclaim/compaction is OOM. Fragindex and the
2328 * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
2329 * excessive compaction for costly orders, but it should not be at the
2330 * expense of system stability.
2333 compact_result = COMPACT_CONTINUE;
2334 if (order > PAGE_ALLOC_COSTLY_ORDER) {
2335 int fragindex = fragmentation_index(zone, order);
2337 if (fragindex >= 0 &&
2338 fragindex <= sysctl_extfrag_threshold) {
2340 compact_result = COMPACT_NOT_SUITABLE_ZONE;
2344 compact_result = COMPACT_SKIPPED;
2347 trace_mm_compaction_suitable(zone, order, compact_result);
2352 bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
2359 * Make sure at least one zone would pass __compaction_suitable if we continue
2360 * retrying the reclaim.
2362 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2363 ac->highest_zoneidx, ac->nodemask) {
2364 unsigned long available;
2367 * Do not consider all the reclaimable memory because we do not
2368 * want to trash just for a single high order allocation which
2369 * is even not guaranteed to appear even if __compaction_suitable
2370 * is happy about the watermark check.
2372 available = zone_reclaimable_pages(zone) / order;
2373 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
2374 if (__compaction_suitable(zone, order, ac->highest_zoneidx,
2382 static enum compact_result
2383 compact_zone(struct compact_control *cc, struct capture_control *capc)
2385 enum compact_result ret;
2386 unsigned long start_pfn = cc->zone->zone_start_pfn;
2387 unsigned long end_pfn = zone_end_pfn(cc->zone);
2388 unsigned long last_migrated_pfn;
2389 const bool sync = cc->mode != MIGRATE_ASYNC;
2391 unsigned int nr_succeeded = 0;
2394 * These counters track activities during zone compaction. Initialize
2395 * them before compacting a new zone.
2397 cc->total_migrate_scanned = 0;
2398 cc->total_free_scanned = 0;
2399 cc->nr_migratepages = 0;
2400 cc->nr_freepages = 0;
2401 INIT_LIST_HEAD(&cc->freepages);
2402 INIT_LIST_HEAD(&cc->migratepages);
2404 cc->migratetype = gfp_migratetype(cc->gfp_mask);
2406 if (!is_via_compact_memory(cc->order)) {
2407 unsigned long watermark;
2409 /* Allocation can already succeed, nothing to do */
2410 watermark = wmark_pages(cc->zone,
2411 cc->alloc_flags & ALLOC_WMARK_MASK);
2412 if (zone_watermark_ok(cc->zone, cc->order, watermark,
2413 cc->highest_zoneidx, cc->alloc_flags))
2414 return COMPACT_SUCCESS;
2416 /* Compaction is likely to fail */
2417 if (!compaction_suitable(cc->zone, cc->order,
2418 cc->highest_zoneidx))
2419 return COMPACT_SKIPPED;
2423 * Clear pageblock skip if there were failures recently and compaction
2424 * is about to be retried after being deferred.
2426 if (compaction_restarting(cc->zone, cc->order))
2427 __reset_isolation_suitable(cc->zone);
2430 * Setup to move all movable pages to the end of the zone. Used cached
2431 * information on where the scanners should start (unless we explicitly
2432 * want to compact the whole zone), but check that it is initialised
2433 * by ensuring the values are within zone boundaries.
2435 cc->fast_start_pfn = 0;
2436 if (cc->whole_zone) {
2437 cc->migrate_pfn = start_pfn;
2438 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2440 cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync];
2441 cc->free_pfn = cc->zone->compact_cached_free_pfn;
2442 if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
2443 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2444 cc->zone->compact_cached_free_pfn = cc->free_pfn;
2446 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
2447 cc->migrate_pfn = start_pfn;
2448 cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
2449 cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
2452 if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn)
2453 cc->whole_zone = true;
2456 last_migrated_pfn = 0;
2459 * Migrate has separate cached PFNs for ASYNC and SYNC* migration on
2460 * the basis that some migrations will fail in ASYNC mode. However,
2461 * if the cached PFNs match and pageblocks are skipped due to having
2462 * no isolation candidates, then the sync state does not matter.
2463 * Until a pageblock with isolation candidates is found, keep the
2464 * cached PFNs in sync to avoid revisiting the same blocks.
2466 update_cached = !sync &&
2467 cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1];
2469 trace_mm_compaction_begin(cc, start_pfn, end_pfn, sync);
2471 /* lru_add_drain_all could be expensive with involving other CPUs */
2474 while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) {
2476 unsigned long iteration_start_pfn = cc->migrate_pfn;
2479 * Avoid multiple rescans of the same pageblock which can
2480 * happen if a page cannot be isolated (dirty/writeback in
2481 * async mode) or if the migrated pages are being allocated
2482 * before the pageblock is cleared. The first rescan will
2483 * capture the entire pageblock for migration. If it fails,
2484 * it'll be marked skip and scanning will proceed as normal.
2486 cc->finish_pageblock = false;
2487 if (pageblock_start_pfn(last_migrated_pfn) ==
2488 pageblock_start_pfn(iteration_start_pfn)) {
2489 cc->finish_pageblock = true;
2493 switch (isolate_migratepages(cc)) {
2495 ret = COMPACT_CONTENDED;
2496 putback_movable_pages(&cc->migratepages);
2497 cc->nr_migratepages = 0;
2500 if (update_cached) {
2501 cc->zone->compact_cached_migrate_pfn[1] =
2502 cc->zone->compact_cached_migrate_pfn[0];
2506 * We haven't isolated and migrated anything, but
2507 * there might still be unflushed migrations from
2508 * previous cc->order aligned block.
2511 case ISOLATE_SUCCESS:
2512 update_cached = false;
2513 last_migrated_pfn = iteration_start_pfn;
2516 err = migrate_pages(&cc->migratepages, compaction_alloc,
2517 compaction_free, (unsigned long)cc, cc->mode,
2518 MR_COMPACTION, &nr_succeeded);
2520 trace_mm_compaction_migratepages(cc, nr_succeeded);
2522 /* All pages were either migrated or will be released */
2523 cc->nr_migratepages = 0;
2525 putback_movable_pages(&cc->migratepages);
2527 * migrate_pages() may return -ENOMEM when scanners meet
2528 * and we want compact_finished() to detect it
2530 if (err == -ENOMEM && !compact_scanners_met(cc)) {
2531 ret = COMPACT_CONTENDED;
2535 * If an ASYNC or SYNC_LIGHT fails to migrate a page
2536 * within the current order-aligned block and
2537 * fast_find_migrateblock may be used then scan the
2538 * remainder of the pageblock. This will mark the
2539 * pageblock "skip" to avoid rescanning in the near
2540 * future. This will isolate more pages than necessary
2541 * for the request but avoid loops due to
2542 * fast_find_migrateblock revisiting blocks that were
2543 * recently partially scanned.
2545 if (!pageblock_aligned(cc->migrate_pfn) &&
2546 !cc->ignore_skip_hint && !cc->finish_pageblock &&
2547 (cc->mode < MIGRATE_SYNC)) {
2548 cc->finish_pageblock = true;
2551 * Draining pcplists does not help THP if
2552 * any page failed to migrate. Even after
2553 * drain, the pageblock will not be free.
2555 if (cc->order == COMPACTION_HPAGE_ORDER)
2556 last_migrated_pfn = 0;
2562 /* Stop if a page has been captured */
2563 if (capc && capc->page) {
2564 ret = COMPACT_SUCCESS;
2570 * Has the migration scanner moved away from the previous
2571 * cc->order aligned block where we migrated from? If yes,
2572 * flush the pages that were freed, so that they can merge and
2573 * compact_finished() can detect immediately if allocation
2576 if (cc->order > 0 && last_migrated_pfn) {
2577 unsigned long current_block_start =
2578 block_start_pfn(cc->migrate_pfn, cc->order);
2580 if (last_migrated_pfn < current_block_start) {
2581 lru_add_drain_cpu_zone(cc->zone);
2582 /* No more flushing until we migrate again */
2583 last_migrated_pfn = 0;
2590 * Release free pages and update where the free scanner should restart,
2591 * so we don't leave any returned pages behind in the next attempt.
2593 if (cc->nr_freepages > 0) {
2594 unsigned long free_pfn = release_freepages(&cc->freepages);
2596 cc->nr_freepages = 0;
2597 VM_BUG_ON(free_pfn == 0);
2598 /* The cached pfn is always the first in a pageblock */
2599 free_pfn = pageblock_start_pfn(free_pfn);
2601 * Only go back, not forward. The cached pfn might have been
2602 * already reset to zone end in compact_finished()
2604 if (free_pfn > cc->zone->compact_cached_free_pfn)
2605 cc->zone->compact_cached_free_pfn = free_pfn;
2608 count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned);
2609 count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned);
2611 trace_mm_compaction_end(cc, start_pfn, end_pfn, sync, ret);
2613 VM_BUG_ON(!list_empty(&cc->freepages));
2614 VM_BUG_ON(!list_empty(&cc->migratepages));
2619 static enum compact_result compact_zone_order(struct zone *zone, int order,
2620 gfp_t gfp_mask, enum compact_priority prio,
2621 unsigned int alloc_flags, int highest_zoneidx,
2622 struct page **capture)
2624 enum compact_result ret;
2625 struct compact_control cc = {
2627 .search_order = order,
2628 .gfp_mask = gfp_mask,
2630 .mode = (prio == COMPACT_PRIO_ASYNC) ?
2631 MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT,
2632 .alloc_flags = alloc_flags,
2633 .highest_zoneidx = highest_zoneidx,
2634 .direct_compaction = true,
2635 .whole_zone = (prio == MIN_COMPACT_PRIORITY),
2636 .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
2637 .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
2639 struct capture_control capc = {
2645 * Make sure the structs are really initialized before we expose the
2646 * capture control, in case we are interrupted and the interrupt handler
2650 WRITE_ONCE(current->capture_control, &capc);
2652 ret = compact_zone(&cc, &capc);
2655 * Make sure we hide capture control first before we read the captured
2656 * page pointer, otherwise an interrupt could free and capture a page
2657 * and we would leak it.
2659 WRITE_ONCE(current->capture_control, NULL);
2660 *capture = READ_ONCE(capc.page);
2662 * Technically, it is also possible that compaction is skipped but
2663 * the page is still captured out of luck(IRQ came and freed the page).
2664 * Returning COMPACT_SUCCESS in such cases helps in properly accounting
2665 * the COMPACT[STALL|FAIL] when compaction is skipped.
2668 ret = COMPACT_SUCCESS;
2674 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
2675 * @gfp_mask: The GFP mask of the current allocation
2676 * @order: The order of the current allocation
2677 * @alloc_flags: The allocation flags of the current allocation
2678 * @ac: The context of current allocation
2679 * @prio: Determines how hard direct compaction should try to succeed
2680 * @capture: Pointer to free page created by compaction will be stored here
2682 * This is the main entry point for direct page compaction.
2684 enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
2685 unsigned int alloc_flags, const struct alloc_context *ac,
2686 enum compact_priority prio, struct page **capture)
2688 int may_perform_io = (__force int)(gfp_mask & __GFP_IO);
2691 enum compact_result rc = COMPACT_SKIPPED;
2694 * Check if the GFP flags allow compaction - GFP_NOIO is really
2695 * tricky context because the migration might require IO
2697 if (!may_perform_io)
2698 return COMPACT_SKIPPED;
2700 trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
2702 /* Compact each zone in the list */
2703 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2704 ac->highest_zoneidx, ac->nodemask) {
2705 enum compact_result status;
2707 if (prio > MIN_COMPACT_PRIORITY
2708 && compaction_deferred(zone, order)) {
2709 rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
2713 status = compact_zone_order(zone, order, gfp_mask, prio,
2714 alloc_flags, ac->highest_zoneidx, capture);
2715 rc = max(status, rc);
2717 /* The allocation should succeed, stop compacting */
2718 if (status == COMPACT_SUCCESS) {
2720 * We think the allocation will succeed in this zone,
2721 * but it is not certain, hence the false. The caller
2722 * will repeat this with true if allocation indeed
2723 * succeeds in this zone.
2725 compaction_defer_reset(zone, order, false);
2730 if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
2731 status == COMPACT_PARTIAL_SKIPPED))
2733 * We think that allocation won't succeed in this zone
2734 * so we defer compaction there. If it ends up
2735 * succeeding after all, it will be reset.
2737 defer_compaction(zone, order);
2740 * We might have stopped compacting due to need_resched() in
2741 * async compaction, or due to a fatal signal detected. In that
2742 * case do not try further zones
2744 if ((prio == COMPACT_PRIO_ASYNC && need_resched())
2745 || fatal_signal_pending(current))
2753 * Compact all zones within a node till each zone's fragmentation score
2754 * reaches within proactive compaction thresholds (as determined by the
2755 * proactiveness tunable).
2757 * It is possible that the function returns before reaching score targets
2758 * due to various back-off conditions, such as, contention on per-node or
2761 static void proactive_compact_node(pg_data_t *pgdat)
2765 struct compact_control cc = {
2767 .mode = MIGRATE_SYNC_LIGHT,
2768 .ignore_skip_hint = true,
2770 .gfp_mask = GFP_KERNEL,
2771 .proactive_compaction = true,
2774 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2775 zone = &pgdat->node_zones[zoneid];
2776 if (!populated_zone(zone))
2781 compact_zone(&cc, NULL);
2783 count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
2784 cc.total_migrate_scanned);
2785 count_compact_events(KCOMPACTD_FREE_SCANNED,
2786 cc.total_free_scanned);
2790 /* Compact all zones within a node */
2791 static void compact_node(int nid)
2793 pg_data_t *pgdat = NODE_DATA(nid);
2796 struct compact_control cc = {
2798 .mode = MIGRATE_SYNC,
2799 .ignore_skip_hint = true,
2801 .gfp_mask = GFP_KERNEL,
2805 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2807 zone = &pgdat->node_zones[zoneid];
2808 if (!populated_zone(zone))
2813 compact_zone(&cc, NULL);
2817 /* Compact all nodes in the system */
2818 static void compact_nodes(void)
2822 /* Flush pending updates to the LRU lists */
2823 lru_add_drain_all();
2825 for_each_online_node(nid)
2829 static int compaction_proactiveness_sysctl_handler(struct ctl_table *table, int write,
2830 void *buffer, size_t *length, loff_t *ppos)
2834 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
2838 if (write && sysctl_compaction_proactiveness) {
2839 for_each_online_node(nid) {
2840 pg_data_t *pgdat = NODE_DATA(nid);
2842 if (pgdat->proactive_compact_trigger)
2845 pgdat->proactive_compact_trigger = true;
2846 trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, -1,
2847 pgdat->nr_zones - 1);
2848 wake_up_interruptible(&pgdat->kcompactd_wait);
2856 * This is the entry point for compacting all nodes via
2857 * /proc/sys/vm/compact_memory
2859 static int sysctl_compaction_handler(struct ctl_table *table, int write,
2860 void *buffer, size_t *length, loff_t *ppos)
2864 ret = proc_dointvec(table, write, buffer, length, ppos);
2868 if (sysctl_compact_memory != 1)
2877 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
2878 static ssize_t compact_store(struct device *dev,
2879 struct device_attribute *attr,
2880 const char *buf, size_t count)
2884 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
2885 /* Flush pending updates to the LRU lists */
2886 lru_add_drain_all();
2893 static DEVICE_ATTR_WO(compact);
2895 int compaction_register_node(struct node *node)
2897 return device_create_file(&node->dev, &dev_attr_compact);
2900 void compaction_unregister_node(struct node *node)
2902 return device_remove_file(&node->dev, &dev_attr_compact);
2904 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
2906 static inline bool kcompactd_work_requested(pg_data_t *pgdat)
2908 return pgdat->kcompactd_max_order > 0 || kthread_should_stop() ||
2909 pgdat->proactive_compact_trigger;
2912 static bool kcompactd_node_suitable(pg_data_t *pgdat)
2916 enum zone_type highest_zoneidx = pgdat->kcompactd_highest_zoneidx;
2918 for (zoneid = 0; zoneid <= highest_zoneidx; zoneid++) {
2919 zone = &pgdat->node_zones[zoneid];
2921 if (!populated_zone(zone))
2924 /* Allocation can already succeed, check other zones */
2925 if (zone_watermark_ok(zone, pgdat->kcompactd_max_order,
2926 min_wmark_pages(zone),
2927 highest_zoneidx, 0))
2930 if (compaction_suitable(zone, pgdat->kcompactd_max_order,
2938 static void kcompactd_do_work(pg_data_t *pgdat)
2941 * With no special task, compact all zones so that a page of requested
2942 * order is allocatable.
2946 struct compact_control cc = {
2947 .order = pgdat->kcompactd_max_order,
2948 .search_order = pgdat->kcompactd_max_order,
2949 .highest_zoneidx = pgdat->kcompactd_highest_zoneidx,
2950 .mode = MIGRATE_SYNC_LIGHT,
2951 .ignore_skip_hint = false,
2952 .gfp_mask = GFP_KERNEL,
2954 trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
2955 cc.highest_zoneidx);
2956 count_compact_event(KCOMPACTD_WAKE);
2958 for (zoneid = 0; zoneid <= cc.highest_zoneidx; zoneid++) {
2961 zone = &pgdat->node_zones[zoneid];
2962 if (!populated_zone(zone))
2965 if (compaction_deferred(zone, cc.order))
2968 /* Allocation can already succeed, nothing to do */
2969 if (zone_watermark_ok(zone, cc.order,
2970 min_wmark_pages(zone), zoneid, 0))
2973 if (!compaction_suitable(zone, cc.order, zoneid))
2976 if (kthread_should_stop())
2980 status = compact_zone(&cc, NULL);
2982 if (status == COMPACT_SUCCESS) {
2983 compaction_defer_reset(zone, cc.order, false);
2984 } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
2986 * Buddy pages may become stranded on pcps that could
2987 * otherwise coalesce on the zone's free area for
2988 * order >= cc.order. This is ratelimited by the
2989 * upcoming deferral.
2991 drain_all_pages(zone);
2994 * We use sync migration mode here, so we defer like
2995 * sync direct compaction does.
2997 defer_compaction(zone, cc.order);
3000 count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
3001 cc.total_migrate_scanned);
3002 count_compact_events(KCOMPACTD_FREE_SCANNED,
3003 cc.total_free_scanned);
3007 * Regardless of success, we are done until woken up next. But remember
3008 * the requested order/highest_zoneidx in case it was higher/tighter
3009 * than our current ones
3011 if (pgdat->kcompactd_max_order <= cc.order)
3012 pgdat->kcompactd_max_order = 0;
3013 if (pgdat->kcompactd_highest_zoneidx >= cc.highest_zoneidx)
3014 pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
3017 void wakeup_kcompactd(pg_data_t *pgdat, int order, int highest_zoneidx)
3022 if (pgdat->kcompactd_max_order < order)
3023 pgdat->kcompactd_max_order = order;
3025 if (pgdat->kcompactd_highest_zoneidx > highest_zoneidx)
3026 pgdat->kcompactd_highest_zoneidx = highest_zoneidx;
3029 * Pairs with implicit barrier in wait_event_freezable()
3030 * such that wakeups are not missed.
3032 if (!wq_has_sleeper(&pgdat->kcompactd_wait))
3035 if (!kcompactd_node_suitable(pgdat))
3038 trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
3040 wake_up_interruptible(&pgdat->kcompactd_wait);
3044 * The background compaction daemon, started as a kernel thread
3045 * from the init process.
3047 static int kcompactd(void *p)
3049 pg_data_t *pgdat = (pg_data_t *)p;
3050 struct task_struct *tsk = current;
3051 long default_timeout = msecs_to_jiffies(HPAGE_FRAG_CHECK_INTERVAL_MSEC);
3052 long timeout = default_timeout;
3054 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3056 if (!cpumask_empty(cpumask))
3057 set_cpus_allowed_ptr(tsk, cpumask);
3061 pgdat->kcompactd_max_order = 0;
3062 pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
3064 while (!kthread_should_stop()) {
3065 unsigned long pflags;
3068 * Avoid the unnecessary wakeup for proactive compaction
3069 * when it is disabled.
3071 if (!sysctl_compaction_proactiveness)
3072 timeout = MAX_SCHEDULE_TIMEOUT;
3073 trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
3074 if (wait_event_freezable_timeout(pgdat->kcompactd_wait,
3075 kcompactd_work_requested(pgdat), timeout) &&
3076 !pgdat->proactive_compact_trigger) {
3078 psi_memstall_enter(&pflags);
3079 kcompactd_do_work(pgdat);
3080 psi_memstall_leave(&pflags);
3082 * Reset the timeout value. The defer timeout from
3083 * proactive compaction is lost here but that is fine
3084 * as the condition of the zone changing substantionally
3085 * then carrying on with the previous defer interval is
3088 timeout = default_timeout;
3093 * Start the proactive work with default timeout. Based
3094 * on the fragmentation score, this timeout is updated.
3096 timeout = default_timeout;
3097 if (should_proactive_compact_node(pgdat)) {
3098 unsigned int prev_score, score;
3100 prev_score = fragmentation_score_node(pgdat);
3101 proactive_compact_node(pgdat);
3102 score = fragmentation_score_node(pgdat);
3104 * Defer proactive compaction if the fragmentation
3105 * score did not go down i.e. no progress made.
3107 if (unlikely(score >= prev_score))
3109 default_timeout << COMPACT_MAX_DEFER_SHIFT;
3111 if (unlikely(pgdat->proactive_compact_trigger))
3112 pgdat->proactive_compact_trigger = false;
3119 * This kcompactd start function will be called by init and node-hot-add.
3120 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
3122 void __meminit kcompactd_run(int nid)
3124 pg_data_t *pgdat = NODE_DATA(nid);
3126 if (pgdat->kcompactd)
3129 pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
3130 if (IS_ERR(pgdat->kcompactd)) {
3131 pr_err("Failed to start kcompactd on node %d\n", nid);
3132 pgdat->kcompactd = NULL;
3137 * Called by memory hotplug when all memory in a node is offlined. Caller must
3138 * be holding mem_hotplug_begin/done().
3140 void __meminit kcompactd_stop(int nid)
3142 struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
3145 kthread_stop(kcompactd);
3146 NODE_DATA(nid)->kcompactd = NULL;
3151 * It's optimal to keep kcompactd on the same CPUs as their memory, but
3152 * not required for correctness. So if the last cpu in a node goes
3153 * away, we get changed to run anywhere: as the first one comes back,
3154 * restore their cpu bindings.
3156 static int kcompactd_cpu_online(unsigned int cpu)
3160 for_each_node_state(nid, N_MEMORY) {
3161 pg_data_t *pgdat = NODE_DATA(nid);
3162 const struct cpumask *mask;
3164 mask = cpumask_of_node(pgdat->node_id);
3166 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3167 /* One of our CPUs online: restore mask */
3168 if (pgdat->kcompactd)
3169 set_cpus_allowed_ptr(pgdat->kcompactd, mask);
3174 static int proc_dointvec_minmax_warn_RT_change(struct ctl_table *table,
3175 int write, void *buffer, size_t *lenp, loff_t *ppos)
3179 if (!IS_ENABLED(CONFIG_PREEMPT_RT) || !write)
3180 return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
3182 old = *(int *)table->data;
3183 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
3186 if (old != *(int *)table->data)
3187 pr_warn_once("sysctl attribute %s changed by %s[%d]\n",
3188 table->procname, current->comm,
3189 task_pid_nr(current));
3193 static struct ctl_table vm_compaction[] = {
3195 .procname = "compact_memory",
3196 .data = &sysctl_compact_memory,
3197 .maxlen = sizeof(int),
3199 .proc_handler = sysctl_compaction_handler,
3202 .procname = "compaction_proactiveness",
3203 .data = &sysctl_compaction_proactiveness,
3204 .maxlen = sizeof(sysctl_compaction_proactiveness),
3206 .proc_handler = compaction_proactiveness_sysctl_handler,
3207 .extra1 = SYSCTL_ZERO,
3208 .extra2 = SYSCTL_ONE_HUNDRED,
3211 .procname = "extfrag_threshold",
3212 .data = &sysctl_extfrag_threshold,
3213 .maxlen = sizeof(int),
3215 .proc_handler = proc_dointvec_minmax,
3216 .extra1 = SYSCTL_ZERO,
3217 .extra2 = SYSCTL_ONE_THOUSAND,
3220 .procname = "compact_unevictable_allowed",
3221 .data = &sysctl_compact_unevictable_allowed,
3222 .maxlen = sizeof(int),
3224 .proc_handler = proc_dointvec_minmax_warn_RT_change,
3225 .extra1 = SYSCTL_ZERO,
3226 .extra2 = SYSCTL_ONE,
3231 static int __init kcompactd_init(void)
3236 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3237 "mm/compaction:online",
3238 kcompactd_cpu_online, NULL);
3240 pr_err("kcompactd: failed to register hotplug callbacks.\n");
3244 for_each_node_state(nid, N_MEMORY)
3246 register_sysctl_init("vm", vm_compaction);
3249 subsys_initcall(kcompactd_init)
3251 #endif /* CONFIG_COMPACTION */