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);
233 * Compound pages of >= pageblock_order should consistently be skipped until
234 * released. It is always pointless to compact pages of such order (if they are
235 * migratable), and the pageblocks they occupy cannot contain any free pages.
237 static bool pageblock_skip_persistent(struct page *page)
239 if (!PageCompound(page))
242 page = compound_head(page);
244 if (compound_order(page) >= pageblock_order)
251 __reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source,
254 struct page *page = pfn_to_online_page(pfn);
255 struct page *block_page;
256 struct page *end_page;
257 unsigned long block_pfn;
261 if (zone != page_zone(page))
263 if (pageblock_skip_persistent(page))
267 * If skip is already cleared do no further checking once the
268 * restart points have been set.
270 if (check_source && check_target && !get_pageblock_skip(page))
274 * If clearing skip for the target scanner, do not select a
275 * non-movable pageblock as the starting point.
277 if (!check_source && check_target &&
278 get_pageblock_migratetype(page) != MIGRATE_MOVABLE)
281 /* Ensure the start of the pageblock or zone is online and valid */
282 block_pfn = pageblock_start_pfn(pfn);
283 block_pfn = max(block_pfn, zone->zone_start_pfn);
284 block_page = pfn_to_online_page(block_pfn);
290 /* Ensure the end of the pageblock or zone is online and valid */
291 block_pfn = pageblock_end_pfn(pfn) - 1;
292 block_pfn = min(block_pfn, zone_end_pfn(zone) - 1);
293 end_page = pfn_to_online_page(block_pfn);
298 * Only clear the hint if a sample indicates there is either a
299 * free page or an LRU page in the block. One or other condition
300 * is necessary for the block to be a migration source/target.
303 if (check_source && PageLRU(page)) {
304 clear_pageblock_skip(page);
308 if (check_target && PageBuddy(page)) {
309 clear_pageblock_skip(page);
313 page += (1 << PAGE_ALLOC_COSTLY_ORDER);
314 } while (page <= end_page);
320 * This function is called to clear all cached information on pageblocks that
321 * should be skipped for page isolation when the migrate and free page scanner
324 static void __reset_isolation_suitable(struct zone *zone)
326 unsigned long migrate_pfn = zone->zone_start_pfn;
327 unsigned long free_pfn = zone_end_pfn(zone) - 1;
328 unsigned long reset_migrate = free_pfn;
329 unsigned long reset_free = migrate_pfn;
330 bool source_set = false;
331 bool free_set = false;
333 if (!zone->compact_blockskip_flush)
336 zone->compact_blockskip_flush = false;
339 * Walk the zone and update pageblock skip information. Source looks
340 * for PageLRU while target looks for PageBuddy. When the scanner
341 * is found, both PageBuddy and PageLRU are checked as the pageblock
342 * is suitable as both source and target.
344 for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages,
345 free_pfn -= pageblock_nr_pages) {
348 /* Update the migrate PFN */
349 if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) &&
350 migrate_pfn < reset_migrate) {
352 reset_migrate = migrate_pfn;
353 zone->compact_init_migrate_pfn = reset_migrate;
354 zone->compact_cached_migrate_pfn[0] = reset_migrate;
355 zone->compact_cached_migrate_pfn[1] = reset_migrate;
358 /* Update the free PFN */
359 if (__reset_isolation_pfn(zone, free_pfn, free_set, true) &&
360 free_pfn > reset_free) {
362 reset_free = free_pfn;
363 zone->compact_init_free_pfn = reset_free;
364 zone->compact_cached_free_pfn = reset_free;
368 /* Leave no distance if no suitable block was reset */
369 if (reset_migrate >= reset_free) {
370 zone->compact_cached_migrate_pfn[0] = migrate_pfn;
371 zone->compact_cached_migrate_pfn[1] = migrate_pfn;
372 zone->compact_cached_free_pfn = free_pfn;
376 void reset_isolation_suitable(pg_data_t *pgdat)
380 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
381 struct zone *zone = &pgdat->node_zones[zoneid];
382 if (!populated_zone(zone))
385 /* Only flush if a full compaction finished recently */
386 if (zone->compact_blockskip_flush)
387 __reset_isolation_suitable(zone);
392 * Sets the pageblock skip bit if it was clear. Note that this is a hint as
393 * locks are not required for read/writers. Returns true if it was already set.
395 static bool test_and_set_skip(struct compact_control *cc, struct page *page)
399 /* Do not update if skip hint is being ignored */
400 if (cc->ignore_skip_hint)
403 skip = get_pageblock_skip(page);
404 if (!skip && !cc->no_set_skip_hint)
405 set_pageblock_skip(page);
410 static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
412 struct zone *zone = cc->zone;
414 pfn = pageblock_end_pfn(pfn);
416 /* Set for isolation rather than compaction */
417 if (cc->no_set_skip_hint)
420 if (pfn > zone->compact_cached_migrate_pfn[0])
421 zone->compact_cached_migrate_pfn[0] = pfn;
422 if (cc->mode != MIGRATE_ASYNC &&
423 pfn > zone->compact_cached_migrate_pfn[1])
424 zone->compact_cached_migrate_pfn[1] = pfn;
428 * If no pages were isolated then mark this pageblock to be skipped in the
429 * future. The information is later cleared by __reset_isolation_suitable().
431 static void update_pageblock_skip(struct compact_control *cc,
432 struct page *page, unsigned long pfn)
434 struct zone *zone = cc->zone;
436 if (cc->no_set_skip_hint)
442 set_pageblock_skip(page);
444 /* Update where async and sync compaction should restart */
445 if (pfn < zone->compact_cached_free_pfn)
446 zone->compact_cached_free_pfn = pfn;
449 static inline bool isolation_suitable(struct compact_control *cc,
455 static inline bool pageblock_skip_persistent(struct page *page)
460 static inline void update_pageblock_skip(struct compact_control *cc,
461 struct page *page, unsigned long pfn)
465 static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
469 static bool test_and_set_skip(struct compact_control *cc, struct page *page)
473 #endif /* CONFIG_COMPACTION */
476 * Compaction requires the taking of some coarse locks that are potentially
477 * very heavily contended. For async compaction, trylock and record if the
478 * lock is contended. The lock will still be acquired but compaction will
479 * abort when the current block is finished regardless of success rate.
480 * Sync compaction acquires the lock.
482 * Always returns true which makes it easier to track lock state in callers.
484 static bool compact_lock_irqsave(spinlock_t *lock, unsigned long *flags,
485 struct compact_control *cc)
488 /* Track if the lock is contended in async mode */
489 if (cc->mode == MIGRATE_ASYNC && !cc->contended) {
490 if (spin_trylock_irqsave(lock, *flags))
493 cc->contended = true;
496 spin_lock_irqsave(lock, *flags);
501 * Compaction requires the taking of some coarse locks that are potentially
502 * very heavily contended. The lock should be periodically unlocked to avoid
503 * having disabled IRQs for a long time, even when there is nobody waiting on
504 * the lock. It might also be that allowing the IRQs will result in
505 * need_resched() becoming true. If scheduling is needed, compaction schedules.
506 * Either compaction type will also abort if a fatal signal is pending.
507 * In either case if the lock was locked, it is dropped and not regained.
509 * Returns true if compaction should abort due to fatal signal pending.
510 * Returns false when compaction can continue.
512 static bool compact_unlock_should_abort(spinlock_t *lock,
513 unsigned long flags, bool *locked, struct compact_control *cc)
516 spin_unlock_irqrestore(lock, flags);
520 if (fatal_signal_pending(current)) {
521 cc->contended = true;
531 * Isolate free pages onto a private freelist. If @strict is true, will abort
532 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
533 * (even though it may still end up isolating some pages).
535 static unsigned long isolate_freepages_block(struct compact_control *cc,
536 unsigned long *start_pfn,
537 unsigned long end_pfn,
538 struct list_head *freelist,
542 int nr_scanned = 0, total_isolated = 0;
544 unsigned long flags = 0;
546 unsigned long blockpfn = *start_pfn;
549 /* Strict mode is for isolation, speed is secondary */
553 cursor = pfn_to_page(blockpfn);
555 /* Isolate free pages. */
556 for (; blockpfn < end_pfn; blockpfn += stride, cursor += stride) {
558 struct page *page = cursor;
561 * Periodically drop the lock (if held) regardless of its
562 * contention, to give chance to IRQs. Abort if fatal signal
565 if (!(blockpfn % COMPACT_CLUSTER_MAX)
566 && compact_unlock_should_abort(&cc->zone->lock, flags,
573 * For compound pages such as THP and hugetlbfs, we can save
574 * potentially a lot of iterations if we skip them at once.
575 * The check is racy, but we can consider only valid values
576 * and the only danger is skipping too much.
578 if (PageCompound(page)) {
579 const unsigned int order = compound_order(page);
581 if (likely(order <= MAX_ORDER)) {
582 blockpfn += (1UL << order) - 1;
583 cursor += (1UL << order) - 1;
584 nr_scanned += (1UL << order) - 1;
589 if (!PageBuddy(page))
592 /* If we already hold the lock, we can skip some rechecking. */
594 locked = compact_lock_irqsave(&cc->zone->lock,
597 /* Recheck this is a buddy page under lock */
598 if (!PageBuddy(page))
602 /* Found a free page, will break it into order-0 pages */
603 order = buddy_order(page);
604 isolated = __isolate_free_page(page, order);
607 set_page_private(page, order);
609 nr_scanned += isolated - 1;
610 total_isolated += isolated;
611 cc->nr_freepages += isolated;
612 list_add_tail(&page->lru, freelist);
614 if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
615 blockpfn += isolated;
618 /* Advance to the end of split page */
619 blockpfn += isolated - 1;
620 cursor += isolated - 1;
632 spin_unlock_irqrestore(&cc->zone->lock, flags);
635 * There is a tiny chance that we have read bogus compound_order(),
636 * so be careful to not go outside of the pageblock.
638 if (unlikely(blockpfn > end_pfn))
641 trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
642 nr_scanned, total_isolated);
644 /* Record how far we have got within the block */
645 *start_pfn = blockpfn;
648 * If strict isolation is requested by CMA then check that all the
649 * pages requested were isolated. If there were any failures, 0 is
650 * returned and CMA will fail.
652 if (strict && blockpfn < end_pfn)
655 cc->total_free_scanned += nr_scanned;
657 count_compact_events(COMPACTISOLATED, total_isolated);
658 return total_isolated;
662 * isolate_freepages_range() - isolate free pages.
663 * @cc: Compaction control structure.
664 * @start_pfn: The first PFN to start isolating.
665 * @end_pfn: The one-past-last PFN.
667 * Non-free pages, invalid PFNs, or zone boundaries within the
668 * [start_pfn, end_pfn) range are considered errors, cause function to
669 * undo its actions and return zero.
671 * Otherwise, function returns one-past-the-last PFN of isolated page
672 * (which may be greater then end_pfn if end fell in a middle of
676 isolate_freepages_range(struct compact_control *cc,
677 unsigned long start_pfn, unsigned long end_pfn)
679 unsigned long isolated, pfn, block_start_pfn, block_end_pfn;
683 block_start_pfn = pageblock_start_pfn(pfn);
684 if (block_start_pfn < cc->zone->zone_start_pfn)
685 block_start_pfn = cc->zone->zone_start_pfn;
686 block_end_pfn = pageblock_end_pfn(pfn);
688 for (; pfn < end_pfn; pfn += isolated,
689 block_start_pfn = block_end_pfn,
690 block_end_pfn += pageblock_nr_pages) {
691 /* Protect pfn from changing by isolate_freepages_block */
692 unsigned long isolate_start_pfn = pfn;
694 block_end_pfn = min(block_end_pfn, end_pfn);
697 * pfn could pass the block_end_pfn if isolated freepage
698 * is more than pageblock order. In this case, we adjust
699 * scanning range to right one.
701 if (pfn >= block_end_pfn) {
702 block_start_pfn = pageblock_start_pfn(pfn);
703 block_end_pfn = pageblock_end_pfn(pfn);
704 block_end_pfn = min(block_end_pfn, end_pfn);
707 if (!pageblock_pfn_to_page(block_start_pfn,
708 block_end_pfn, cc->zone))
711 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
712 block_end_pfn, &freelist, 0, true);
715 * In strict mode, isolate_freepages_block() returns 0 if
716 * there are any holes in the block (ie. invalid PFNs or
723 * If we managed to isolate pages, it is always (1 << n) *
724 * pageblock_nr_pages for some non-negative n. (Max order
725 * page may span two pageblocks).
729 /* __isolate_free_page() does not map the pages */
730 split_map_pages(&freelist);
733 /* Loop terminated early, cleanup. */
734 release_freepages(&freelist);
738 /* We don't use freelists for anything. */
742 /* Similar to reclaim, but different enough that they don't share logic */
743 static bool too_many_isolated(pg_data_t *pgdat)
747 unsigned long active, inactive, isolated;
749 inactive = node_page_state(pgdat, NR_INACTIVE_FILE) +
750 node_page_state(pgdat, NR_INACTIVE_ANON);
751 active = node_page_state(pgdat, NR_ACTIVE_FILE) +
752 node_page_state(pgdat, NR_ACTIVE_ANON);
753 isolated = node_page_state(pgdat, NR_ISOLATED_FILE) +
754 node_page_state(pgdat, NR_ISOLATED_ANON);
756 too_many = isolated > (inactive + active) / 2;
758 wake_throttle_isolated(pgdat);
764 * isolate_migratepages_block() - isolate all migrate-able pages within
766 * @cc: Compaction control structure.
767 * @low_pfn: The first PFN to isolate
768 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
769 * @mode: Isolation mode to be used.
771 * Isolate all pages that can be migrated from the range specified by
772 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
773 * Returns errno, like -EAGAIN or -EINTR in case e.g signal pending or congestion,
774 * -ENOMEM in case we could not allocate a page, or 0.
775 * cc->migrate_pfn will contain the next pfn to scan.
777 * The pages are isolated on cc->migratepages list (not required to be empty),
778 * and cc->nr_migratepages is updated accordingly.
781 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
782 unsigned long end_pfn, isolate_mode_t mode)
784 pg_data_t *pgdat = cc->zone->zone_pgdat;
785 unsigned long nr_scanned = 0, nr_isolated = 0;
786 struct lruvec *lruvec;
787 unsigned long flags = 0;
788 struct lruvec *locked = NULL;
789 struct page *page = NULL, *valid_page = NULL;
790 struct address_space *mapping;
791 unsigned long start_pfn = low_pfn;
792 bool skip_on_failure = false;
793 unsigned long next_skip_pfn = 0;
794 bool skip_updated = false;
797 cc->migrate_pfn = low_pfn;
800 * Ensure that there are not too many pages isolated from the LRU
801 * list by either parallel reclaimers or compaction. If there are,
802 * delay for some time until fewer pages are isolated
804 while (unlikely(too_many_isolated(pgdat))) {
805 /* stop isolation if there are still pages not migrated */
806 if (cc->nr_migratepages)
809 /* async migration should just abort */
810 if (cc->mode == MIGRATE_ASYNC)
813 reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED);
815 if (fatal_signal_pending(current))
821 if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
822 skip_on_failure = true;
823 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
826 /* Time to isolate some pages for migration */
827 for (; low_pfn < end_pfn; low_pfn++) {
829 if (skip_on_failure && low_pfn >= next_skip_pfn) {
831 * We have isolated all migration candidates in the
832 * previous order-aligned block, and did not skip it due
833 * to failure. We should migrate the pages now and
834 * hopefully succeed compaction.
840 * We failed to isolate in the previous order-aligned
841 * block. Set the new boundary to the end of the
842 * current block. Note we can't simply increase
843 * next_skip_pfn by 1 << order, as low_pfn might have
844 * been incremented by a higher number due to skipping
845 * a compound or a high-order buddy page in the
846 * previous loop iteration.
848 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
852 * Periodically drop the lock (if held) regardless of its
853 * contention, to give chance to IRQs. Abort completely if
854 * a fatal signal is pending.
856 if (!(low_pfn % COMPACT_CLUSTER_MAX)) {
858 unlock_page_lruvec_irqrestore(locked, flags);
862 if (fatal_signal_pending(current)) {
863 cc->contended = true;
874 page = pfn_to_page(low_pfn);
877 * Check if the pageblock has already been marked skipped.
878 * Only the aligned PFN is checked as the caller isolates
879 * COMPACT_CLUSTER_MAX at a time so the second call must
880 * not falsely conclude that the block should be skipped.
882 if (!valid_page && pageblock_aligned(low_pfn)) {
883 if (!isolation_suitable(cc, page)) {
891 if (PageHuge(page) && cc->alloc_contig) {
893 unlock_page_lruvec_irqrestore(locked, flags);
897 ret = isolate_or_dissolve_huge_page(page, &cc->migratepages);
900 * Fail isolation in case isolate_or_dissolve_huge_page()
901 * reports an error. In case of -ENOMEM, abort right away.
904 /* Do not report -EBUSY down the chain */
907 low_pfn += compound_nr(page) - 1;
908 nr_scanned += compound_nr(page) - 1;
912 if (PageHuge(page)) {
914 * Hugepage was successfully isolated and placed
915 * on the cc->migratepages list.
917 low_pfn += compound_nr(page) - 1;
918 goto isolate_success_no_list;
922 * Ok, the hugepage was dissolved. Now these pages are
923 * Buddy and cannot be re-allocated because they are
924 * isolated. Fall-through as the check below handles
930 * Skip if free. We read page order here without zone lock
931 * which is generally unsafe, but the race window is small and
932 * the worst thing that can happen is that we skip some
933 * potential isolation targets.
935 if (PageBuddy(page)) {
936 unsigned long freepage_order = buddy_order_unsafe(page);
939 * Without lock, we cannot be sure that what we got is
940 * a valid page order. Consider only values in the
941 * valid order range to prevent low_pfn overflow.
943 if (freepage_order > 0 && freepage_order <= MAX_ORDER) {
944 low_pfn += (1UL << freepage_order) - 1;
945 nr_scanned += (1UL << freepage_order) - 1;
951 * Regardless of being on LRU, compound pages such as THP and
952 * hugetlbfs are not to be compacted unless we are attempting
953 * an allocation much larger than the huge page size (eg CMA).
954 * We can potentially save a lot of iterations if we skip them
955 * at once. The check is racy, but we can consider only valid
956 * values and the only danger is skipping too much.
958 if (PageCompound(page) && !cc->alloc_contig) {
959 const unsigned int order = compound_order(page);
961 if (likely(order <= MAX_ORDER)) {
962 low_pfn += (1UL << order) - 1;
963 nr_scanned += (1UL << order) - 1;
969 * Check may be lockless but that's ok as we recheck later.
970 * It's possible to migrate LRU and non-lru movable pages.
971 * Skip any other type of page
973 if (!PageLRU(page)) {
975 * __PageMovable can return false positive so we need
976 * to verify it under page_lock.
978 if (unlikely(__PageMovable(page)) &&
979 !PageIsolated(page)) {
981 unlock_page_lruvec_irqrestore(locked, flags);
985 if (isolate_movable_page(page, mode))
986 goto isolate_success;
993 * Be careful not to clear PageLRU until after we're
994 * sure the page is not being freed elsewhere -- the
995 * page release code relies on it.
997 if (unlikely(!get_page_unless_zero(page)))
1001 * Migration will fail if an anonymous page is pinned in memory,
1002 * so avoid taking lru_lock and isolating it unnecessarily in an
1003 * admittedly racy check.
1005 mapping = page_mapping(page);
1006 if (!mapping && (page_count(page) - 1) > total_mapcount(page))
1007 goto isolate_fail_put;
1010 * Only allow to migrate anonymous pages in GFP_NOFS context
1011 * because those do not depend on fs locks.
1013 if (!(cc->gfp_mask & __GFP_FS) && mapping)
1014 goto isolate_fail_put;
1016 /* Only take pages on LRU: a check now makes later tests safe */
1018 goto isolate_fail_put;
1020 /* Compaction might skip unevictable pages but CMA takes them */
1021 if (!(mode & ISOLATE_UNEVICTABLE) && PageUnevictable(page))
1022 goto isolate_fail_put;
1025 * To minimise LRU disruption, the caller can indicate with
1026 * ISOLATE_ASYNC_MIGRATE that it only wants to isolate pages
1027 * it will be able to migrate without blocking - clean pages
1028 * for the most part. PageWriteback would require blocking.
1030 if ((mode & ISOLATE_ASYNC_MIGRATE) && PageWriteback(page))
1031 goto isolate_fail_put;
1033 if ((mode & ISOLATE_ASYNC_MIGRATE) && PageDirty(page)) {
1037 * Only pages without mappings or that have a
1038 * ->migrate_folio callback are possible to migrate
1039 * without blocking. However, we can be racing with
1040 * truncation so it's necessary to lock the page
1041 * to stabilise the mapping as truncation holds
1042 * the page lock until after the page is removed
1043 * from the page cache.
1045 if (!trylock_page(page))
1046 goto isolate_fail_put;
1048 mapping = page_mapping(page);
1049 migrate_dirty = !mapping ||
1050 mapping->a_ops->migrate_folio;
1053 goto isolate_fail_put;
1056 /* Try isolate the page */
1057 if (!TestClearPageLRU(page))
1058 goto isolate_fail_put;
1060 lruvec = folio_lruvec(page_folio(page));
1062 /* If we already hold the lock, we can skip some rechecking */
1063 if (lruvec != locked) {
1065 unlock_page_lruvec_irqrestore(locked, flags);
1067 compact_lock_irqsave(&lruvec->lru_lock, &flags, cc);
1070 lruvec_memcg_debug(lruvec, page_folio(page));
1073 * Try get exclusive access under lock. If marked for
1074 * skip, the scan is aborted unless the current context
1075 * is a rescan to reach the end of the pageblock.
1077 if (!skip_updated && valid_page) {
1078 skip_updated = true;
1079 if (test_and_set_skip(cc, valid_page) &&
1080 !cc->finish_pageblock) {
1086 * Page become compound since the non-locked check,
1087 * and it's on LRU. It can only be a THP so the order
1088 * is safe to read and it's 0 for tail pages.
1090 if (unlikely(PageCompound(page) && !cc->alloc_contig)) {
1091 low_pfn += compound_nr(page) - 1;
1092 nr_scanned += compound_nr(page) - 1;
1094 goto isolate_fail_put;
1098 /* The whole page is taken off the LRU; skip the tail pages. */
1099 if (PageCompound(page))
1100 low_pfn += compound_nr(page) - 1;
1102 /* Successfully isolated */
1103 del_page_from_lru_list(page, lruvec);
1104 mod_node_page_state(page_pgdat(page),
1105 NR_ISOLATED_ANON + page_is_file_lru(page),
1106 thp_nr_pages(page));
1109 list_add(&page->lru, &cc->migratepages);
1110 isolate_success_no_list:
1111 cc->nr_migratepages += compound_nr(page);
1112 nr_isolated += compound_nr(page);
1113 nr_scanned += compound_nr(page) - 1;
1116 * Avoid isolating too much unless this block is being
1117 * fully scanned (e.g. dirty/writeback pages, parallel allocation)
1118 * or a lock is contended. For contention, isolate quickly to
1119 * potentially remove one source of contention.
1121 if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX &&
1122 !cc->finish_pageblock && !cc->contended) {
1130 /* Avoid potential deadlock in freeing page under lru_lock */
1132 unlock_page_lruvec_irqrestore(locked, flags);
1138 if (!skip_on_failure && ret != -ENOMEM)
1142 * We have isolated some pages, but then failed. Release them
1143 * instead of migrating, as we cannot form the cc->order buddy
1148 unlock_page_lruvec_irqrestore(locked, flags);
1151 putback_movable_pages(&cc->migratepages);
1152 cc->nr_migratepages = 0;
1156 if (low_pfn < next_skip_pfn) {
1157 low_pfn = next_skip_pfn - 1;
1159 * The check near the loop beginning would have updated
1160 * next_skip_pfn too, but this is a bit simpler.
1162 next_skip_pfn += 1UL << cc->order;
1170 * The PageBuddy() check could have potentially brought us outside
1171 * the range to be scanned.
1173 if (unlikely(low_pfn > end_pfn))
1180 unlock_page_lruvec_irqrestore(locked, flags);
1187 * Update the cached scanner pfn once the pageblock has been scanned.
1188 * Pages will either be migrated in which case there is no point
1189 * scanning in the near future or migration failed in which case the
1190 * failure reason may persist. The block is marked for skipping if
1191 * there were no pages isolated in the block or if the block is
1192 * rescanned twice in a row.
1194 if (low_pfn == end_pfn && (!nr_isolated || cc->finish_pageblock)) {
1195 if (valid_page && !skip_updated)
1196 set_pageblock_skip(valid_page);
1197 update_cached_migrate(cc, low_pfn);
1200 trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
1201 nr_scanned, nr_isolated);
1204 cc->total_migrate_scanned += nr_scanned;
1206 count_compact_events(COMPACTISOLATED, nr_isolated);
1208 cc->migrate_pfn = low_pfn;
1214 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
1215 * @cc: Compaction control structure.
1216 * @start_pfn: The first PFN to start isolating.
1217 * @end_pfn: The one-past-last PFN.
1219 * Returns -EAGAIN when contented, -EINTR in case of a signal pending, -ENOMEM
1220 * in case we could not allocate a page, or 0.
1223 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
1224 unsigned long end_pfn)
1226 unsigned long pfn, block_start_pfn, block_end_pfn;
1229 /* Scan block by block. First and last block may be incomplete */
1231 block_start_pfn = pageblock_start_pfn(pfn);
1232 if (block_start_pfn < cc->zone->zone_start_pfn)
1233 block_start_pfn = cc->zone->zone_start_pfn;
1234 block_end_pfn = pageblock_end_pfn(pfn);
1236 for (; pfn < end_pfn; pfn = block_end_pfn,
1237 block_start_pfn = block_end_pfn,
1238 block_end_pfn += pageblock_nr_pages) {
1240 block_end_pfn = min(block_end_pfn, end_pfn);
1242 if (!pageblock_pfn_to_page(block_start_pfn,
1243 block_end_pfn, cc->zone))
1246 ret = isolate_migratepages_block(cc, pfn, block_end_pfn,
1247 ISOLATE_UNEVICTABLE);
1252 if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX)
1259 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
1260 #ifdef CONFIG_COMPACTION
1262 static bool suitable_migration_source(struct compact_control *cc,
1267 if (pageblock_skip_persistent(page))
1270 if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction)
1273 block_mt = get_pageblock_migratetype(page);
1275 if (cc->migratetype == MIGRATE_MOVABLE)
1276 return is_migrate_movable(block_mt);
1278 return block_mt == cc->migratetype;
1281 /* Returns true if the page is within a block suitable for migration to */
1282 static bool suitable_migration_target(struct compact_control *cc,
1285 /* If the page is a large free page, then disallow migration */
1286 if (PageBuddy(page)) {
1288 * We are checking page_order without zone->lock taken. But
1289 * the only small danger is that we skip a potentially suitable
1290 * pageblock, so it's not worth to check order for valid range.
1292 if (buddy_order_unsafe(page) >= pageblock_order)
1296 if (cc->ignore_block_suitable)
1299 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1300 if (is_migrate_movable(get_pageblock_migratetype(page)))
1303 /* Otherwise skip the block */
1307 static inline unsigned int
1308 freelist_scan_limit(struct compact_control *cc)
1310 unsigned short shift = BITS_PER_LONG - 1;
1312 return (COMPACT_CLUSTER_MAX >> min(shift, cc->fast_search_fail)) + 1;
1316 * Test whether the free scanner has reached the same or lower pageblock than
1317 * the migration scanner, and compaction should thus terminate.
1319 static inline bool compact_scanners_met(struct compact_control *cc)
1321 return (cc->free_pfn >> pageblock_order)
1322 <= (cc->migrate_pfn >> pageblock_order);
1326 * Used when scanning for a suitable migration target which scans freelists
1327 * in reverse. Reorders the list such as the unscanned pages are scanned
1328 * first on the next iteration of the free scanner
1331 move_freelist_head(struct list_head *freelist, struct page *freepage)
1335 if (!list_is_last(freelist, &freepage->lru)) {
1336 list_cut_before(&sublist, freelist, &freepage->lru);
1337 list_splice_tail(&sublist, freelist);
1342 * Similar to move_freelist_head except used by the migration scanner
1343 * when scanning forward. It's possible for these list operations to
1344 * move against each other if they search the free list exactly in
1348 move_freelist_tail(struct list_head *freelist, struct page *freepage)
1352 if (!list_is_first(freelist, &freepage->lru)) {
1353 list_cut_position(&sublist, freelist, &freepage->lru);
1354 list_splice_tail(&sublist, freelist);
1359 fast_isolate_around(struct compact_control *cc, unsigned long pfn)
1361 unsigned long start_pfn, end_pfn;
1364 /* Do not search around if there are enough pages already */
1365 if (cc->nr_freepages >= cc->nr_migratepages)
1368 /* Minimise scanning during async compaction */
1369 if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC)
1372 /* Pageblock boundaries */
1373 start_pfn = max(pageblock_start_pfn(pfn), cc->zone->zone_start_pfn);
1374 end_pfn = min(pageblock_end_pfn(pfn), zone_end_pfn(cc->zone));
1376 page = pageblock_pfn_to_page(start_pfn, end_pfn, cc->zone);
1380 isolate_freepages_block(cc, &start_pfn, end_pfn, &cc->freepages, 1, false);
1382 /* Skip this pageblock in the future as it's full or nearly full */
1383 if (cc->nr_freepages < cc->nr_migratepages)
1384 set_pageblock_skip(page);
1389 /* Search orders in round-robin fashion */
1390 static int next_search_order(struct compact_control *cc, int order)
1394 order = cc->order - 1;
1396 /* Search wrapped around? */
1397 if (order == cc->search_order) {
1399 if (cc->search_order < 0)
1400 cc->search_order = cc->order - 1;
1407 static unsigned long
1408 fast_isolate_freepages(struct compact_control *cc)
1410 unsigned int limit = max(1U, freelist_scan_limit(cc) >> 1);
1411 unsigned int nr_scanned = 0;
1412 unsigned long low_pfn, min_pfn, highest = 0;
1413 unsigned long nr_isolated = 0;
1414 unsigned long distance;
1415 struct page *page = NULL;
1416 bool scan_start = false;
1419 /* Full compaction passes in a negative order */
1421 return cc->free_pfn;
1424 * If starting the scan, use a deeper search and use the highest
1425 * PFN found if a suitable one is not found.
1427 if (cc->free_pfn >= cc->zone->compact_init_free_pfn) {
1428 limit = pageblock_nr_pages >> 1;
1433 * Preferred point is in the top quarter of the scan space but take
1434 * a pfn from the top half if the search is problematic.
1436 distance = (cc->free_pfn - cc->migrate_pfn);
1437 low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2));
1438 min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1));
1440 if (WARN_ON_ONCE(min_pfn > low_pfn))
1444 * Search starts from the last successful isolation order or the next
1445 * order to search after a previous failure
1447 cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order);
1449 for (order = cc->search_order;
1450 !page && order >= 0;
1451 order = next_search_order(cc, order)) {
1452 struct free_area *area = &cc->zone->free_area[order];
1453 struct list_head *freelist;
1454 struct page *freepage;
1455 unsigned long flags;
1456 unsigned int order_scanned = 0;
1457 unsigned long high_pfn = 0;
1462 spin_lock_irqsave(&cc->zone->lock, flags);
1463 freelist = &area->free_list[MIGRATE_MOVABLE];
1464 list_for_each_entry_reverse(freepage, freelist, lru) {
1469 pfn = page_to_pfn(freepage);
1472 highest = max(pageblock_start_pfn(pfn),
1473 cc->zone->zone_start_pfn);
1475 if (pfn >= low_pfn) {
1476 cc->fast_search_fail = 0;
1477 cc->search_order = order;
1482 if (pfn >= min_pfn && pfn > high_pfn) {
1485 /* Shorten the scan if a candidate is found */
1489 if (order_scanned >= limit)
1493 /* Use a minimum pfn if a preferred one was not found */
1494 if (!page && high_pfn) {
1495 page = pfn_to_page(high_pfn);
1497 /* Update freepage for the list reorder below */
1501 /* Reorder to so a future search skips recent pages */
1502 move_freelist_head(freelist, freepage);
1504 /* Isolate the page if available */
1506 if (__isolate_free_page(page, order)) {
1507 set_page_private(page, order);
1508 nr_isolated = 1 << order;
1509 nr_scanned += nr_isolated - 1;
1510 cc->nr_freepages += nr_isolated;
1511 list_add_tail(&page->lru, &cc->freepages);
1512 count_compact_events(COMPACTISOLATED, nr_isolated);
1514 /* If isolation fails, abort the search */
1515 order = cc->search_order + 1;
1520 spin_unlock_irqrestore(&cc->zone->lock, flags);
1523 * Smaller scan on next order so the total scan is related
1524 * to freelist_scan_limit.
1526 if (order_scanned >= limit)
1527 limit = max(1U, limit >> 1);
1531 cc->fast_search_fail++;
1534 * Use the highest PFN found above min. If one was
1535 * not found, be pessimistic for direct compaction
1536 * and use the min mark.
1538 if (highest >= min_pfn) {
1539 page = pfn_to_page(highest);
1540 cc->free_pfn = highest;
1542 if (cc->direct_compaction && pfn_valid(min_pfn)) {
1543 page = pageblock_pfn_to_page(min_pfn,
1544 min(pageblock_end_pfn(min_pfn),
1545 zone_end_pfn(cc->zone)),
1547 cc->free_pfn = min_pfn;
1553 if (highest && highest >= cc->zone->compact_cached_free_pfn) {
1554 highest -= pageblock_nr_pages;
1555 cc->zone->compact_cached_free_pfn = highest;
1558 cc->total_free_scanned += nr_scanned;
1560 return cc->free_pfn;
1562 low_pfn = page_to_pfn(page);
1563 fast_isolate_around(cc, low_pfn);
1568 * Based on information in the current compact_control, find blocks
1569 * suitable for isolating free pages from and then isolate them.
1571 static void isolate_freepages(struct compact_control *cc)
1573 struct zone *zone = cc->zone;
1575 unsigned long block_start_pfn; /* start of current pageblock */
1576 unsigned long isolate_start_pfn; /* exact pfn we start at */
1577 unsigned long block_end_pfn; /* end of current pageblock */
1578 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
1579 struct list_head *freelist = &cc->freepages;
1580 unsigned int stride;
1582 /* Try a small search of the free lists for a candidate */
1583 fast_isolate_freepages(cc);
1584 if (cc->nr_freepages)
1588 * Initialise the free scanner. The starting point is where we last
1589 * successfully isolated from, zone-cached value, or the end of the
1590 * zone when isolating for the first time. For looping we also need
1591 * this pfn aligned down to the pageblock boundary, because we do
1592 * block_start_pfn -= pageblock_nr_pages in the for loop.
1593 * For ending point, take care when isolating in last pageblock of a
1594 * zone which ends in the middle of a pageblock.
1595 * The low boundary is the end of the pageblock the migration scanner
1598 isolate_start_pfn = cc->free_pfn;
1599 block_start_pfn = pageblock_start_pfn(isolate_start_pfn);
1600 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1601 zone_end_pfn(zone));
1602 low_pfn = pageblock_end_pfn(cc->migrate_pfn);
1603 stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1;
1606 * Isolate free pages until enough are available to migrate the
1607 * pages on cc->migratepages. We stop searching if the migrate
1608 * and free page scanners meet or enough free pages are isolated.
1610 for (; block_start_pfn >= low_pfn;
1611 block_end_pfn = block_start_pfn,
1612 block_start_pfn -= pageblock_nr_pages,
1613 isolate_start_pfn = block_start_pfn) {
1614 unsigned long nr_isolated;
1617 * This can iterate a massively long zone without finding any
1618 * suitable migration targets, so periodically check resched.
1620 if (!(block_start_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages)))
1623 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1628 /* Check the block is suitable for migration */
1629 if (!suitable_migration_target(cc, page))
1632 /* If isolation recently failed, do not retry */
1633 if (!isolation_suitable(cc, page))
1636 /* Found a block suitable for isolating free pages from. */
1637 nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn,
1638 block_end_pfn, freelist, stride, false);
1640 /* Update the skip hint if the full pageblock was scanned */
1641 if (isolate_start_pfn == block_end_pfn)
1642 update_pageblock_skip(cc, page, block_start_pfn);
1644 /* Are enough freepages isolated? */
1645 if (cc->nr_freepages >= cc->nr_migratepages) {
1646 if (isolate_start_pfn >= block_end_pfn) {
1648 * Restart at previous pageblock if more
1649 * freepages can be isolated next time.
1652 block_start_pfn - pageblock_nr_pages;
1655 } else if (isolate_start_pfn < block_end_pfn) {
1657 * If isolation failed early, do not continue
1663 /* Adjust stride depending on isolation */
1668 stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1);
1672 * Record where the free scanner will restart next time. Either we
1673 * broke from the loop and set isolate_start_pfn based on the last
1674 * call to isolate_freepages_block(), or we met the migration scanner
1675 * and the loop terminated due to isolate_start_pfn < low_pfn
1677 cc->free_pfn = isolate_start_pfn;
1680 /* __isolate_free_page() does not map the pages */
1681 split_map_pages(freelist);
1685 * This is a migrate-callback that "allocates" freepages by taking pages
1686 * from the isolated freelists in the block we are migrating to.
1688 static struct folio *compaction_alloc(struct folio *src, unsigned long data)
1690 struct compact_control *cc = (struct compact_control *)data;
1693 if (list_empty(&cc->freepages)) {
1694 isolate_freepages(cc);
1696 if (list_empty(&cc->freepages))
1700 dst = list_entry(cc->freepages.next, struct folio, lru);
1701 list_del(&dst->lru);
1708 * This is a migrate-callback that "frees" freepages back to the isolated
1709 * freelist. All pages on the freelist are from the same zone, so there is no
1710 * special handling needed for NUMA.
1712 static void compaction_free(struct folio *dst, unsigned long data)
1714 struct compact_control *cc = (struct compact_control *)data;
1716 list_add(&dst->lru, &cc->freepages);
1720 /* possible outcome of isolate_migratepages */
1722 ISOLATE_ABORT, /* Abort compaction now */
1723 ISOLATE_NONE, /* No pages isolated, continue scanning */
1724 ISOLATE_SUCCESS, /* Pages isolated, migrate */
1725 } isolate_migrate_t;
1728 * Allow userspace to control policy on scanning the unevictable LRU for
1729 * compactable pages.
1731 static int sysctl_compact_unevictable_allowed __read_mostly = CONFIG_COMPACT_UNEVICTABLE_DEFAULT;
1733 * Tunable for proactive compaction. It determines how
1734 * aggressively the kernel should compact memory in the
1735 * background. It takes values in the range [0, 100].
1737 static unsigned int __read_mostly sysctl_compaction_proactiveness = 20;
1738 static int sysctl_extfrag_threshold = 500;
1739 static int __read_mostly sysctl_compact_memory;
1742 update_fast_start_pfn(struct compact_control *cc, unsigned long pfn)
1744 if (cc->fast_start_pfn == ULONG_MAX)
1747 if (!cc->fast_start_pfn)
1748 cc->fast_start_pfn = pfn;
1750 cc->fast_start_pfn = min(cc->fast_start_pfn, pfn);
1753 static inline unsigned long
1754 reinit_migrate_pfn(struct compact_control *cc)
1756 if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX)
1757 return cc->migrate_pfn;
1759 cc->migrate_pfn = cc->fast_start_pfn;
1760 cc->fast_start_pfn = ULONG_MAX;
1762 return cc->migrate_pfn;
1766 * Briefly search the free lists for a migration source that already has
1767 * some free pages to reduce the number of pages that need migration
1768 * before a pageblock is free.
1770 static unsigned long fast_find_migrateblock(struct compact_control *cc)
1772 unsigned int limit = freelist_scan_limit(cc);
1773 unsigned int nr_scanned = 0;
1774 unsigned long distance;
1775 unsigned long pfn = cc->migrate_pfn;
1776 unsigned long high_pfn;
1778 bool found_block = false;
1780 /* Skip hints are relied on to avoid repeats on the fast search */
1781 if (cc->ignore_skip_hint)
1785 * If the pageblock should be finished then do not select a different
1788 if (cc->finish_pageblock)
1792 * If the migrate_pfn is not at the start of a zone or the start
1793 * of a pageblock then assume this is a continuation of a previous
1794 * scan restarted due to COMPACT_CLUSTER_MAX.
1796 if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn))
1800 * For smaller orders, just linearly scan as the number of pages
1801 * to migrate should be relatively small and does not necessarily
1802 * justify freeing up a large block for a small allocation.
1804 if (cc->order <= PAGE_ALLOC_COSTLY_ORDER)
1808 * Only allow kcompactd and direct requests for movable pages to
1809 * quickly clear out a MOVABLE pageblock for allocation. This
1810 * reduces the risk that a large movable pageblock is freed for
1811 * an unmovable/reclaimable small allocation.
1813 if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE)
1817 * When starting the migration scanner, pick any pageblock within the
1818 * first half of the search space. Otherwise try and pick a pageblock
1819 * within the first eighth to reduce the chances that a migration
1820 * target later becomes a source.
1822 distance = (cc->free_pfn - cc->migrate_pfn) >> 1;
1823 if (cc->migrate_pfn != cc->zone->zone_start_pfn)
1825 high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance);
1827 for (order = cc->order - 1;
1828 order >= PAGE_ALLOC_COSTLY_ORDER && !found_block && nr_scanned < limit;
1830 struct free_area *area = &cc->zone->free_area[order];
1831 struct list_head *freelist;
1832 unsigned long flags;
1833 struct page *freepage;
1838 spin_lock_irqsave(&cc->zone->lock, flags);
1839 freelist = &area->free_list[MIGRATE_MOVABLE];
1840 list_for_each_entry(freepage, freelist, lru) {
1841 unsigned long free_pfn;
1843 if (nr_scanned++ >= limit) {
1844 move_freelist_tail(freelist, freepage);
1848 free_pfn = page_to_pfn(freepage);
1849 if (free_pfn < high_pfn) {
1851 * Avoid if skipped recently. Ideally it would
1852 * move to the tail but even safe iteration of
1853 * the list assumes an entry is deleted, not
1856 if (get_pageblock_skip(freepage))
1859 /* Reorder to so a future search skips recent pages */
1860 move_freelist_tail(freelist, freepage);
1862 update_fast_start_pfn(cc, free_pfn);
1863 pfn = pageblock_start_pfn(free_pfn);
1864 if (pfn < cc->zone->zone_start_pfn)
1865 pfn = cc->zone->zone_start_pfn;
1866 cc->fast_search_fail = 0;
1871 spin_unlock_irqrestore(&cc->zone->lock, flags);
1874 cc->total_migrate_scanned += nr_scanned;
1877 * If fast scanning failed then use a cached entry for a page block
1878 * that had free pages as the basis for starting a linear scan.
1881 cc->fast_search_fail++;
1882 pfn = reinit_migrate_pfn(cc);
1888 * Isolate all pages that can be migrated from the first suitable block,
1889 * starting at the block pointed to by the migrate scanner pfn within
1892 static isolate_migrate_t isolate_migratepages(struct compact_control *cc)
1894 unsigned long block_start_pfn;
1895 unsigned long block_end_pfn;
1896 unsigned long low_pfn;
1898 const isolate_mode_t isolate_mode =
1899 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1900 (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1901 bool fast_find_block;
1904 * Start at where we last stopped, or beginning of the zone as
1905 * initialized by compact_zone(). The first failure will use
1906 * the lowest PFN as the starting point for linear scanning.
1908 low_pfn = fast_find_migrateblock(cc);
1909 block_start_pfn = pageblock_start_pfn(low_pfn);
1910 if (block_start_pfn < cc->zone->zone_start_pfn)
1911 block_start_pfn = cc->zone->zone_start_pfn;
1914 * fast_find_migrateblock marks a pageblock skipped so to avoid
1915 * the isolation_suitable check below, check whether the fast
1916 * search was successful.
1918 fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail;
1920 /* Only scan within a pageblock boundary */
1921 block_end_pfn = pageblock_end_pfn(low_pfn);
1924 * Iterate over whole pageblocks until we find the first suitable.
1925 * Do not cross the free scanner.
1927 for (; block_end_pfn <= cc->free_pfn;
1928 fast_find_block = false,
1929 cc->migrate_pfn = low_pfn = block_end_pfn,
1930 block_start_pfn = block_end_pfn,
1931 block_end_pfn += pageblock_nr_pages) {
1934 * This can potentially iterate a massively long zone with
1935 * many pageblocks unsuitable, so periodically check if we
1938 if (!(low_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages)))
1941 page = pageblock_pfn_to_page(block_start_pfn,
1942 block_end_pfn, cc->zone);
1947 * If isolation recently failed, do not retry. Only check the
1948 * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock
1949 * to be visited multiple times. Assume skip was checked
1950 * before making it "skip" so other compaction instances do
1951 * not scan the same block.
1953 if (pageblock_aligned(low_pfn) &&
1954 !fast_find_block && !isolation_suitable(cc, page))
1958 * For async direct compaction, only scan the pageblocks of the
1959 * same migratetype without huge pages. Async direct compaction
1960 * is optimistic to see if the minimum amount of work satisfies
1961 * the allocation. The cached PFN is updated as it's possible
1962 * that all remaining blocks between source and target are
1963 * unsuitable and the compaction scanners fail to meet.
1965 if (!suitable_migration_source(cc, page)) {
1966 update_cached_migrate(cc, block_end_pfn);
1970 /* Perform the isolation */
1971 if (isolate_migratepages_block(cc, low_pfn, block_end_pfn,
1973 return ISOLATE_ABORT;
1976 * Either we isolated something and proceed with migration. Or
1977 * we failed and compact_zone should decide if we should
1983 return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
1987 * order == -1 is expected when compacting via
1988 * /proc/sys/vm/compact_memory
1990 static inline bool is_via_compact_memory(int order)
1996 * Determine whether kswapd is (or recently was!) running on this node.
1998 * pgdat_kswapd_lock() pins pgdat->kswapd, so a concurrent kswapd_stop() can't
2001 static bool kswapd_is_running(pg_data_t *pgdat)
2005 pgdat_kswapd_lock(pgdat);
2006 running = pgdat->kswapd && task_is_running(pgdat->kswapd);
2007 pgdat_kswapd_unlock(pgdat);
2013 * A zone's fragmentation score is the external fragmentation wrt to the
2014 * COMPACTION_HPAGE_ORDER. It returns a value in the range [0, 100].
2016 static unsigned int fragmentation_score_zone(struct zone *zone)
2018 return extfrag_for_order(zone, COMPACTION_HPAGE_ORDER);
2022 * A weighted zone's fragmentation score is the external fragmentation
2023 * wrt to the COMPACTION_HPAGE_ORDER scaled by the zone's size. It
2024 * returns a value in the range [0, 100].
2026 * The scaling factor ensures that proactive compaction focuses on larger
2027 * zones like ZONE_NORMAL, rather than smaller, specialized zones like
2028 * ZONE_DMA32. For smaller zones, the score value remains close to zero,
2029 * and thus never exceeds the high threshold for proactive compaction.
2031 static unsigned int fragmentation_score_zone_weighted(struct zone *zone)
2033 unsigned long score;
2035 score = zone->present_pages * fragmentation_score_zone(zone);
2036 return div64_ul(score, zone->zone_pgdat->node_present_pages + 1);
2040 * The per-node proactive (background) compaction process is started by its
2041 * corresponding kcompactd thread when the node's fragmentation score
2042 * exceeds the high threshold. The compaction process remains active till
2043 * the node's score falls below the low threshold, or one of the back-off
2044 * conditions is met.
2046 static unsigned int fragmentation_score_node(pg_data_t *pgdat)
2048 unsigned int score = 0;
2051 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2054 zone = &pgdat->node_zones[zoneid];
2055 if (!populated_zone(zone))
2057 score += fragmentation_score_zone_weighted(zone);
2063 static unsigned int fragmentation_score_wmark(pg_data_t *pgdat, bool low)
2065 unsigned int wmark_low;
2068 * Cap the low watermark to avoid excessive compaction
2069 * activity in case a user sets the proactiveness tunable
2070 * close to 100 (maximum).
2072 wmark_low = max(100U - sysctl_compaction_proactiveness, 5U);
2073 return low ? wmark_low : min(wmark_low + 10, 100U);
2076 static bool should_proactive_compact_node(pg_data_t *pgdat)
2080 if (!sysctl_compaction_proactiveness || kswapd_is_running(pgdat))
2083 wmark_high = fragmentation_score_wmark(pgdat, false);
2084 return fragmentation_score_node(pgdat) > wmark_high;
2087 static enum compact_result __compact_finished(struct compact_control *cc)
2090 const int migratetype = cc->migratetype;
2093 /* Compaction run completes if the migrate and free scanner meet */
2094 if (compact_scanners_met(cc)) {
2095 /* Let the next compaction start anew. */
2096 reset_cached_positions(cc->zone);
2099 * Mark that the PG_migrate_skip information should be cleared
2100 * by kswapd when it goes to sleep. kcompactd does not set the
2101 * flag itself as the decision to be clear should be directly
2102 * based on an allocation request.
2104 if (cc->direct_compaction)
2105 cc->zone->compact_blockskip_flush = true;
2108 return COMPACT_COMPLETE;
2110 return COMPACT_PARTIAL_SKIPPED;
2113 if (cc->proactive_compaction) {
2114 int score, wmark_low;
2117 pgdat = cc->zone->zone_pgdat;
2118 if (kswapd_is_running(pgdat))
2119 return COMPACT_PARTIAL_SKIPPED;
2121 score = fragmentation_score_zone(cc->zone);
2122 wmark_low = fragmentation_score_wmark(pgdat, true);
2124 if (score > wmark_low)
2125 ret = COMPACT_CONTINUE;
2127 ret = COMPACT_SUCCESS;
2132 if (is_via_compact_memory(cc->order))
2133 return COMPACT_CONTINUE;
2136 * Always finish scanning a pageblock to reduce the possibility of
2137 * fallbacks in the future. This is particularly important when
2138 * migration source is unmovable/reclaimable but it's not worth
2141 if (!pageblock_aligned(cc->migrate_pfn))
2142 return COMPACT_CONTINUE;
2144 /* Direct compactor: Is a suitable page free? */
2145 ret = COMPACT_NO_SUITABLE_PAGE;
2146 for (order = cc->order; order <= MAX_ORDER; order++) {
2147 struct free_area *area = &cc->zone->free_area[order];
2150 /* Job done if page is free of the right migratetype */
2151 if (!free_area_empty(area, migratetype))
2152 return COMPACT_SUCCESS;
2155 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
2156 if (migratetype == MIGRATE_MOVABLE &&
2157 !free_area_empty(area, MIGRATE_CMA))
2158 return COMPACT_SUCCESS;
2161 * Job done if allocation would steal freepages from
2162 * other migratetype buddy lists.
2164 if (find_suitable_fallback(area, order, migratetype,
2165 true, &can_steal) != -1)
2167 * Movable pages are OK in any pageblock. If we are
2168 * stealing for a non-movable allocation, make sure
2169 * we finish compacting the current pageblock first
2170 * (which is assured by the above migrate_pfn align
2171 * check) so it is as free as possible and we won't
2172 * have to steal another one soon.
2174 return COMPACT_SUCCESS;
2178 if (cc->contended || fatal_signal_pending(current))
2179 ret = COMPACT_CONTENDED;
2184 static enum compact_result compact_finished(struct compact_control *cc)
2188 ret = __compact_finished(cc);
2189 trace_mm_compaction_finished(cc->zone, cc->order, ret);
2190 if (ret == COMPACT_NO_SUITABLE_PAGE)
2191 ret = COMPACT_CONTINUE;
2196 static enum compact_result __compaction_suitable(struct zone *zone, int order,
2197 int highest_zoneidx,
2198 unsigned long wmark_target)
2200 unsigned long watermark;
2202 * Watermarks for order-0 must be met for compaction to be able to
2203 * isolate free pages for migration targets. This means that the
2204 * watermark and alloc_flags have to match, or be more pessimistic than
2205 * the check in __isolate_free_page(). We don't use the direct
2206 * compactor's alloc_flags, as they are not relevant for freepage
2207 * isolation. We however do use the direct compactor's highest_zoneidx
2208 * to skip over zones where lowmem reserves would prevent allocation
2209 * even if compaction succeeds.
2210 * For costly orders, we require low watermark instead of min for
2211 * compaction to proceed to increase its chances.
2212 * ALLOC_CMA is used, as pages in CMA pageblocks are considered
2213 * suitable migration targets
2215 watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
2216 low_wmark_pages(zone) : min_wmark_pages(zone);
2217 watermark += compact_gap(order);
2218 if (!__zone_watermark_ok(zone, 0, watermark, highest_zoneidx,
2219 ALLOC_CMA, wmark_target))
2220 return COMPACT_SKIPPED;
2222 return COMPACT_CONTINUE;
2226 * compaction_suitable: Is this suitable to run compaction on this zone now?
2228 * COMPACT_SKIPPED - If there are too few free pages for compaction
2229 * COMPACT_CONTINUE - If compaction should run now
2231 enum compact_result compaction_suitable(struct zone *zone, int order,
2232 int highest_zoneidx)
2234 enum compact_result ret;
2237 ret = __compaction_suitable(zone, order, highest_zoneidx,
2238 zone_page_state(zone, NR_FREE_PAGES));
2240 * fragmentation index determines if allocation failures are due to
2241 * low memory or external fragmentation
2243 * index of -1000 would imply allocations might succeed depending on
2244 * watermarks, but we already failed the high-order watermark check
2245 * index towards 0 implies failure is due to lack of memory
2246 * index towards 1000 implies failure is due to fragmentation
2248 * Only compact if a failure would be due to fragmentation. Also
2249 * ignore fragindex for non-costly orders where the alternative to
2250 * a successful reclaim/compaction is OOM. Fragindex and the
2251 * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
2252 * excessive compaction for costly orders, but it should not be at the
2253 * expense of system stability.
2255 if (ret == COMPACT_CONTINUE && (order > PAGE_ALLOC_COSTLY_ORDER)) {
2256 fragindex = fragmentation_index(zone, order);
2257 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
2258 ret = COMPACT_NOT_SUITABLE_ZONE;
2261 trace_mm_compaction_suitable(zone, order, ret);
2262 if (ret == COMPACT_NOT_SUITABLE_ZONE)
2263 ret = COMPACT_SKIPPED;
2268 bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
2275 * Make sure at least one zone would pass __compaction_suitable if we continue
2276 * retrying the reclaim.
2278 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2279 ac->highest_zoneidx, ac->nodemask) {
2280 unsigned long available;
2281 unsigned long watermark;
2283 /* Allocation can already succeed, nothing to do */
2284 watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
2285 if (zone_watermark_ok(zone, order, watermark,
2286 ac->highest_zoneidx, alloc_flags))
2290 * Do not consider all the reclaimable memory because we do not
2291 * want to trash just for a single high order allocation which
2292 * is even not guaranteed to appear even if __compaction_suitable
2293 * is happy about the watermark check.
2295 available = zone_reclaimable_pages(zone) / order;
2296 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
2297 if (__compaction_suitable(zone, order, ac->highest_zoneidx,
2298 available) == COMPACT_CONTINUE)
2305 static enum compact_result
2306 compact_zone(struct compact_control *cc, struct capture_control *capc)
2308 enum compact_result ret;
2309 unsigned long start_pfn = cc->zone->zone_start_pfn;
2310 unsigned long end_pfn = zone_end_pfn(cc->zone);
2311 unsigned long last_migrated_pfn;
2312 const bool sync = cc->mode != MIGRATE_ASYNC;
2314 unsigned int nr_succeeded = 0;
2317 * These counters track activities during zone compaction. Initialize
2318 * them before compacting a new zone.
2320 cc->total_migrate_scanned = 0;
2321 cc->total_free_scanned = 0;
2322 cc->nr_migratepages = 0;
2323 cc->nr_freepages = 0;
2324 INIT_LIST_HEAD(&cc->freepages);
2325 INIT_LIST_HEAD(&cc->migratepages);
2327 cc->migratetype = gfp_migratetype(cc->gfp_mask);
2329 if (!is_via_compact_memory(cc->order)) {
2330 unsigned long watermark;
2332 /* Allocation can already succeed, nothing to do */
2333 watermark = wmark_pages(cc->zone,
2334 cc->alloc_flags & ALLOC_WMARK_MASK);
2335 if (zone_watermark_ok(cc->zone, cc->order, watermark,
2336 cc->highest_zoneidx, cc->alloc_flags))
2337 return COMPACT_SUCCESS;
2339 ret = compaction_suitable(cc->zone, cc->order,
2340 cc->highest_zoneidx);
2341 /* Compaction is likely to fail */
2342 if (ret == COMPACT_SKIPPED)
2347 * Clear pageblock skip if there were failures recently and compaction
2348 * is about to be retried after being deferred.
2350 if (compaction_restarting(cc->zone, cc->order))
2351 __reset_isolation_suitable(cc->zone);
2354 * Setup to move all movable pages to the end of the zone. Used cached
2355 * information on where the scanners should start (unless we explicitly
2356 * want to compact the whole zone), but check that it is initialised
2357 * by ensuring the values are within zone boundaries.
2359 cc->fast_start_pfn = 0;
2360 if (cc->whole_zone) {
2361 cc->migrate_pfn = start_pfn;
2362 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2364 cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync];
2365 cc->free_pfn = cc->zone->compact_cached_free_pfn;
2366 if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
2367 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2368 cc->zone->compact_cached_free_pfn = cc->free_pfn;
2370 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
2371 cc->migrate_pfn = start_pfn;
2372 cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
2373 cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
2376 if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn)
2377 cc->whole_zone = true;
2380 last_migrated_pfn = 0;
2383 * Migrate has separate cached PFNs for ASYNC and SYNC* migration on
2384 * the basis that some migrations will fail in ASYNC mode. However,
2385 * if the cached PFNs match and pageblocks are skipped due to having
2386 * no isolation candidates, then the sync state does not matter.
2387 * Until a pageblock with isolation candidates is found, keep the
2388 * cached PFNs in sync to avoid revisiting the same blocks.
2390 update_cached = !sync &&
2391 cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1];
2393 trace_mm_compaction_begin(cc, start_pfn, end_pfn, sync);
2395 /* lru_add_drain_all could be expensive with involving other CPUs */
2398 while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) {
2400 unsigned long iteration_start_pfn = cc->migrate_pfn;
2403 * Avoid multiple rescans of the same pageblock which can
2404 * happen if a page cannot be isolated (dirty/writeback in
2405 * async mode) or if the migrated pages are being allocated
2406 * before the pageblock is cleared. The first rescan will
2407 * capture the entire pageblock for migration. If it fails,
2408 * it'll be marked skip and scanning will proceed as normal.
2410 cc->finish_pageblock = false;
2411 if (pageblock_start_pfn(last_migrated_pfn) ==
2412 pageblock_start_pfn(iteration_start_pfn)) {
2413 cc->finish_pageblock = true;
2417 switch (isolate_migratepages(cc)) {
2419 ret = COMPACT_CONTENDED;
2420 putback_movable_pages(&cc->migratepages);
2421 cc->nr_migratepages = 0;
2424 if (update_cached) {
2425 cc->zone->compact_cached_migrate_pfn[1] =
2426 cc->zone->compact_cached_migrate_pfn[0];
2430 * We haven't isolated and migrated anything, but
2431 * there might still be unflushed migrations from
2432 * previous cc->order aligned block.
2435 case ISOLATE_SUCCESS:
2436 update_cached = false;
2437 last_migrated_pfn = iteration_start_pfn;
2440 err = migrate_pages(&cc->migratepages, compaction_alloc,
2441 compaction_free, (unsigned long)cc, cc->mode,
2442 MR_COMPACTION, &nr_succeeded);
2444 trace_mm_compaction_migratepages(cc, nr_succeeded);
2446 /* All pages were either migrated or will be released */
2447 cc->nr_migratepages = 0;
2449 putback_movable_pages(&cc->migratepages);
2451 * migrate_pages() may return -ENOMEM when scanners meet
2452 * and we want compact_finished() to detect it
2454 if (err == -ENOMEM && !compact_scanners_met(cc)) {
2455 ret = COMPACT_CONTENDED;
2459 * If an ASYNC or SYNC_LIGHT fails to migrate a page
2460 * within the current order-aligned block and
2461 * fast_find_migrateblock may be used then scan the
2462 * remainder of the pageblock. This will mark the
2463 * pageblock "skip" to avoid rescanning in the near
2464 * future. This will isolate more pages than necessary
2465 * for the request but avoid loops due to
2466 * fast_find_migrateblock revisiting blocks that were
2467 * recently partially scanned.
2469 if (!pageblock_aligned(cc->migrate_pfn) &&
2470 !cc->ignore_skip_hint && !cc->finish_pageblock &&
2471 (cc->mode < MIGRATE_SYNC)) {
2472 cc->finish_pageblock = true;
2475 * Draining pcplists does not help THP if
2476 * any page failed to migrate. Even after
2477 * drain, the pageblock will not be free.
2479 if (cc->order == COMPACTION_HPAGE_ORDER)
2480 last_migrated_pfn = 0;
2486 /* Stop if a page has been captured */
2487 if (capc && capc->page) {
2488 ret = COMPACT_SUCCESS;
2494 * Has the migration scanner moved away from the previous
2495 * cc->order aligned block where we migrated from? If yes,
2496 * flush the pages that were freed, so that they can merge and
2497 * compact_finished() can detect immediately if allocation
2500 if (cc->order > 0 && last_migrated_pfn) {
2501 unsigned long current_block_start =
2502 block_start_pfn(cc->migrate_pfn, cc->order);
2504 if (last_migrated_pfn < current_block_start) {
2505 lru_add_drain_cpu_zone(cc->zone);
2506 /* No more flushing until we migrate again */
2507 last_migrated_pfn = 0;
2514 * Release free pages and update where the free scanner should restart,
2515 * so we don't leave any returned pages behind in the next attempt.
2517 if (cc->nr_freepages > 0) {
2518 unsigned long free_pfn = release_freepages(&cc->freepages);
2520 cc->nr_freepages = 0;
2521 VM_BUG_ON(free_pfn == 0);
2522 /* The cached pfn is always the first in a pageblock */
2523 free_pfn = pageblock_start_pfn(free_pfn);
2525 * Only go back, not forward. The cached pfn might have been
2526 * already reset to zone end in compact_finished()
2528 if (free_pfn > cc->zone->compact_cached_free_pfn)
2529 cc->zone->compact_cached_free_pfn = free_pfn;
2532 count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned);
2533 count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned);
2535 trace_mm_compaction_end(cc, start_pfn, end_pfn, sync, ret);
2537 VM_BUG_ON(!list_empty(&cc->freepages));
2538 VM_BUG_ON(!list_empty(&cc->migratepages));
2543 static enum compact_result compact_zone_order(struct zone *zone, int order,
2544 gfp_t gfp_mask, enum compact_priority prio,
2545 unsigned int alloc_flags, int highest_zoneidx,
2546 struct page **capture)
2548 enum compact_result ret;
2549 struct compact_control cc = {
2551 .search_order = order,
2552 .gfp_mask = gfp_mask,
2554 .mode = (prio == COMPACT_PRIO_ASYNC) ?
2555 MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT,
2556 .alloc_flags = alloc_flags,
2557 .highest_zoneidx = highest_zoneidx,
2558 .direct_compaction = true,
2559 .whole_zone = (prio == MIN_COMPACT_PRIORITY),
2560 .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
2561 .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
2563 struct capture_control capc = {
2569 * Make sure the structs are really initialized before we expose the
2570 * capture control, in case we are interrupted and the interrupt handler
2574 WRITE_ONCE(current->capture_control, &capc);
2576 ret = compact_zone(&cc, &capc);
2579 * Make sure we hide capture control first before we read the captured
2580 * page pointer, otherwise an interrupt could free and capture a page
2581 * and we would leak it.
2583 WRITE_ONCE(current->capture_control, NULL);
2584 *capture = READ_ONCE(capc.page);
2586 * Technically, it is also possible that compaction is skipped but
2587 * the page is still captured out of luck(IRQ came and freed the page).
2588 * Returning COMPACT_SUCCESS in such cases helps in properly accounting
2589 * the COMPACT[STALL|FAIL] when compaction is skipped.
2592 ret = COMPACT_SUCCESS;
2598 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
2599 * @gfp_mask: The GFP mask of the current allocation
2600 * @order: The order of the current allocation
2601 * @alloc_flags: The allocation flags of the current allocation
2602 * @ac: The context of current allocation
2603 * @prio: Determines how hard direct compaction should try to succeed
2604 * @capture: Pointer to free page created by compaction will be stored here
2606 * This is the main entry point for direct page compaction.
2608 enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
2609 unsigned int alloc_flags, const struct alloc_context *ac,
2610 enum compact_priority prio, struct page **capture)
2612 int may_perform_io = (__force int)(gfp_mask & __GFP_IO);
2615 enum compact_result rc = COMPACT_SKIPPED;
2618 * Check if the GFP flags allow compaction - GFP_NOIO is really
2619 * tricky context because the migration might require IO
2621 if (!may_perform_io)
2622 return COMPACT_SKIPPED;
2624 trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
2626 /* Compact each zone in the list */
2627 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2628 ac->highest_zoneidx, ac->nodemask) {
2629 enum compact_result status;
2631 if (prio > MIN_COMPACT_PRIORITY
2632 && compaction_deferred(zone, order)) {
2633 rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
2637 status = compact_zone_order(zone, order, gfp_mask, prio,
2638 alloc_flags, ac->highest_zoneidx, capture);
2639 rc = max(status, rc);
2641 /* The allocation should succeed, stop compacting */
2642 if (status == COMPACT_SUCCESS) {
2644 * We think the allocation will succeed in this zone,
2645 * but it is not certain, hence the false. The caller
2646 * will repeat this with true if allocation indeed
2647 * succeeds in this zone.
2649 compaction_defer_reset(zone, order, false);
2654 if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
2655 status == COMPACT_PARTIAL_SKIPPED))
2657 * We think that allocation won't succeed in this zone
2658 * so we defer compaction there. If it ends up
2659 * succeeding after all, it will be reset.
2661 defer_compaction(zone, order);
2664 * We might have stopped compacting due to need_resched() in
2665 * async compaction, or due to a fatal signal detected. In that
2666 * case do not try further zones
2668 if ((prio == COMPACT_PRIO_ASYNC && need_resched())
2669 || fatal_signal_pending(current))
2677 * Compact all zones within a node till each zone's fragmentation score
2678 * reaches within proactive compaction thresholds (as determined by the
2679 * proactiveness tunable).
2681 * It is possible that the function returns before reaching score targets
2682 * due to various back-off conditions, such as, contention on per-node or
2685 static void proactive_compact_node(pg_data_t *pgdat)
2689 struct compact_control cc = {
2691 .mode = MIGRATE_SYNC_LIGHT,
2692 .ignore_skip_hint = true,
2694 .gfp_mask = GFP_KERNEL,
2695 .proactive_compaction = true,
2698 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2699 zone = &pgdat->node_zones[zoneid];
2700 if (!populated_zone(zone))
2705 compact_zone(&cc, NULL);
2707 count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
2708 cc.total_migrate_scanned);
2709 count_compact_events(KCOMPACTD_FREE_SCANNED,
2710 cc.total_free_scanned);
2714 /* Compact all zones within a node */
2715 static void compact_node(int nid)
2717 pg_data_t *pgdat = NODE_DATA(nid);
2720 struct compact_control cc = {
2722 .mode = MIGRATE_SYNC,
2723 .ignore_skip_hint = true,
2725 .gfp_mask = GFP_KERNEL,
2729 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2731 zone = &pgdat->node_zones[zoneid];
2732 if (!populated_zone(zone))
2737 compact_zone(&cc, NULL);
2741 /* Compact all nodes in the system */
2742 static void compact_nodes(void)
2746 /* Flush pending updates to the LRU lists */
2747 lru_add_drain_all();
2749 for_each_online_node(nid)
2753 static int compaction_proactiveness_sysctl_handler(struct ctl_table *table, int write,
2754 void *buffer, size_t *length, loff_t *ppos)
2758 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
2762 if (write && sysctl_compaction_proactiveness) {
2763 for_each_online_node(nid) {
2764 pg_data_t *pgdat = NODE_DATA(nid);
2766 if (pgdat->proactive_compact_trigger)
2769 pgdat->proactive_compact_trigger = true;
2770 trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, -1,
2771 pgdat->nr_zones - 1);
2772 wake_up_interruptible(&pgdat->kcompactd_wait);
2780 * This is the entry point for compacting all nodes via
2781 * /proc/sys/vm/compact_memory
2783 static int sysctl_compaction_handler(struct ctl_table *table, int write,
2784 void *buffer, size_t *length, loff_t *ppos)
2788 ret = proc_dointvec(table, write, buffer, length, ppos);
2792 if (sysctl_compact_memory != 1)
2801 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
2802 static ssize_t compact_store(struct device *dev,
2803 struct device_attribute *attr,
2804 const char *buf, size_t count)
2808 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
2809 /* Flush pending updates to the LRU lists */
2810 lru_add_drain_all();
2817 static DEVICE_ATTR_WO(compact);
2819 int compaction_register_node(struct node *node)
2821 return device_create_file(&node->dev, &dev_attr_compact);
2824 void compaction_unregister_node(struct node *node)
2826 return device_remove_file(&node->dev, &dev_attr_compact);
2828 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
2830 static inline bool kcompactd_work_requested(pg_data_t *pgdat)
2832 return pgdat->kcompactd_max_order > 0 || kthread_should_stop() ||
2833 pgdat->proactive_compact_trigger;
2836 static bool kcompactd_node_suitable(pg_data_t *pgdat)
2840 enum zone_type highest_zoneidx = pgdat->kcompactd_highest_zoneidx;
2842 for (zoneid = 0; zoneid <= highest_zoneidx; zoneid++) {
2843 zone = &pgdat->node_zones[zoneid];
2845 if (!populated_zone(zone))
2848 /* Allocation can already succeed, check other zones */
2849 if (zone_watermark_ok(zone, pgdat->kcompactd_max_order,
2850 min_wmark_pages(zone),
2851 highest_zoneidx, 0))
2854 if (compaction_suitable(zone, pgdat->kcompactd_max_order,
2855 highest_zoneidx) == COMPACT_CONTINUE)
2862 static void kcompactd_do_work(pg_data_t *pgdat)
2865 * With no special task, compact all zones so that a page of requested
2866 * order is allocatable.
2870 struct compact_control cc = {
2871 .order = pgdat->kcompactd_max_order,
2872 .search_order = pgdat->kcompactd_max_order,
2873 .highest_zoneidx = pgdat->kcompactd_highest_zoneidx,
2874 .mode = MIGRATE_SYNC_LIGHT,
2875 .ignore_skip_hint = false,
2876 .gfp_mask = GFP_KERNEL,
2878 trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
2879 cc.highest_zoneidx);
2880 count_compact_event(KCOMPACTD_WAKE);
2882 for (zoneid = 0; zoneid <= cc.highest_zoneidx; zoneid++) {
2885 zone = &pgdat->node_zones[zoneid];
2886 if (!populated_zone(zone))
2889 if (compaction_deferred(zone, cc.order))
2892 /* Allocation can already succeed, nothing to do */
2893 if (zone_watermark_ok(zone, cc.order,
2894 min_wmark_pages(zone), zoneid, 0))
2897 if (compaction_suitable(zone, cc.order,
2898 zoneid) != COMPACT_CONTINUE)
2901 if (kthread_should_stop())
2905 status = compact_zone(&cc, NULL);
2907 if (status == COMPACT_SUCCESS) {
2908 compaction_defer_reset(zone, cc.order, false);
2909 } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
2911 * Buddy pages may become stranded on pcps that could
2912 * otherwise coalesce on the zone's free area for
2913 * order >= cc.order. This is ratelimited by the
2914 * upcoming deferral.
2916 drain_all_pages(zone);
2919 * We use sync migration mode here, so we defer like
2920 * sync direct compaction does.
2922 defer_compaction(zone, cc.order);
2925 count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
2926 cc.total_migrate_scanned);
2927 count_compact_events(KCOMPACTD_FREE_SCANNED,
2928 cc.total_free_scanned);
2932 * Regardless of success, we are done until woken up next. But remember
2933 * the requested order/highest_zoneidx in case it was higher/tighter
2934 * than our current ones
2936 if (pgdat->kcompactd_max_order <= cc.order)
2937 pgdat->kcompactd_max_order = 0;
2938 if (pgdat->kcompactd_highest_zoneidx >= cc.highest_zoneidx)
2939 pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
2942 void wakeup_kcompactd(pg_data_t *pgdat, int order, int highest_zoneidx)
2947 if (pgdat->kcompactd_max_order < order)
2948 pgdat->kcompactd_max_order = order;
2950 if (pgdat->kcompactd_highest_zoneidx > highest_zoneidx)
2951 pgdat->kcompactd_highest_zoneidx = highest_zoneidx;
2954 * Pairs with implicit barrier in wait_event_freezable()
2955 * such that wakeups are not missed.
2957 if (!wq_has_sleeper(&pgdat->kcompactd_wait))
2960 if (!kcompactd_node_suitable(pgdat))
2963 trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
2965 wake_up_interruptible(&pgdat->kcompactd_wait);
2969 * The background compaction daemon, started as a kernel thread
2970 * from the init process.
2972 static int kcompactd(void *p)
2974 pg_data_t *pgdat = (pg_data_t *)p;
2975 struct task_struct *tsk = current;
2976 long default_timeout = msecs_to_jiffies(HPAGE_FRAG_CHECK_INTERVAL_MSEC);
2977 long timeout = default_timeout;
2979 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2981 if (!cpumask_empty(cpumask))
2982 set_cpus_allowed_ptr(tsk, cpumask);
2986 pgdat->kcompactd_max_order = 0;
2987 pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
2989 while (!kthread_should_stop()) {
2990 unsigned long pflags;
2993 * Avoid the unnecessary wakeup for proactive compaction
2994 * when it is disabled.
2996 if (!sysctl_compaction_proactiveness)
2997 timeout = MAX_SCHEDULE_TIMEOUT;
2998 trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
2999 if (wait_event_freezable_timeout(pgdat->kcompactd_wait,
3000 kcompactd_work_requested(pgdat), timeout) &&
3001 !pgdat->proactive_compact_trigger) {
3003 psi_memstall_enter(&pflags);
3004 kcompactd_do_work(pgdat);
3005 psi_memstall_leave(&pflags);
3007 * Reset the timeout value. The defer timeout from
3008 * proactive compaction is lost here but that is fine
3009 * as the condition of the zone changing substantionally
3010 * then carrying on with the previous defer interval is
3013 timeout = default_timeout;
3018 * Start the proactive work with default timeout. Based
3019 * on the fragmentation score, this timeout is updated.
3021 timeout = default_timeout;
3022 if (should_proactive_compact_node(pgdat)) {
3023 unsigned int prev_score, score;
3025 prev_score = fragmentation_score_node(pgdat);
3026 proactive_compact_node(pgdat);
3027 score = fragmentation_score_node(pgdat);
3029 * Defer proactive compaction if the fragmentation
3030 * score did not go down i.e. no progress made.
3032 if (unlikely(score >= prev_score))
3034 default_timeout << COMPACT_MAX_DEFER_SHIFT;
3036 if (unlikely(pgdat->proactive_compact_trigger))
3037 pgdat->proactive_compact_trigger = false;
3044 * This kcompactd start function will be called by init and node-hot-add.
3045 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
3047 void kcompactd_run(int nid)
3049 pg_data_t *pgdat = NODE_DATA(nid);
3051 if (pgdat->kcompactd)
3054 pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
3055 if (IS_ERR(pgdat->kcompactd)) {
3056 pr_err("Failed to start kcompactd on node %d\n", nid);
3057 pgdat->kcompactd = NULL;
3062 * Called by memory hotplug when all memory in a node is offlined. Caller must
3063 * be holding mem_hotplug_begin/done().
3065 void kcompactd_stop(int nid)
3067 struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
3070 kthread_stop(kcompactd);
3071 NODE_DATA(nid)->kcompactd = NULL;
3076 * It's optimal to keep kcompactd on the same CPUs as their memory, but
3077 * not required for correctness. So if the last cpu in a node goes
3078 * away, we get changed to run anywhere: as the first one comes back,
3079 * restore their cpu bindings.
3081 static int kcompactd_cpu_online(unsigned int cpu)
3085 for_each_node_state(nid, N_MEMORY) {
3086 pg_data_t *pgdat = NODE_DATA(nid);
3087 const struct cpumask *mask;
3089 mask = cpumask_of_node(pgdat->node_id);
3091 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3092 /* One of our CPUs online: restore mask */
3093 if (pgdat->kcompactd)
3094 set_cpus_allowed_ptr(pgdat->kcompactd, mask);
3099 static int proc_dointvec_minmax_warn_RT_change(struct ctl_table *table,
3100 int write, void *buffer, size_t *lenp, loff_t *ppos)
3104 if (!IS_ENABLED(CONFIG_PREEMPT_RT) || !write)
3105 return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
3107 old = *(int *)table->data;
3108 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
3111 if (old != *(int *)table->data)
3112 pr_warn_once("sysctl attribute %s changed by %s[%d]\n",
3113 table->procname, current->comm,
3114 task_pid_nr(current));
3118 static struct ctl_table vm_compaction[] = {
3120 .procname = "compact_memory",
3121 .data = &sysctl_compact_memory,
3122 .maxlen = sizeof(int),
3124 .proc_handler = sysctl_compaction_handler,
3127 .procname = "compaction_proactiveness",
3128 .data = &sysctl_compaction_proactiveness,
3129 .maxlen = sizeof(sysctl_compaction_proactiveness),
3131 .proc_handler = compaction_proactiveness_sysctl_handler,
3132 .extra1 = SYSCTL_ZERO,
3133 .extra2 = SYSCTL_ONE_HUNDRED,
3136 .procname = "extfrag_threshold",
3137 .data = &sysctl_extfrag_threshold,
3138 .maxlen = sizeof(int),
3140 .proc_handler = proc_dointvec_minmax,
3141 .extra1 = SYSCTL_ZERO,
3142 .extra2 = SYSCTL_ONE_THOUSAND,
3145 .procname = "compact_unevictable_allowed",
3146 .data = &sysctl_compact_unevictable_allowed,
3147 .maxlen = sizeof(int),
3149 .proc_handler = proc_dointvec_minmax_warn_RT_change,
3150 .extra1 = SYSCTL_ZERO,
3151 .extra2 = SYSCTL_ONE,
3156 static int __init kcompactd_init(void)
3161 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3162 "mm/compaction:online",
3163 kcompactd_cpu_online, NULL);
3165 pr_err("kcompactd: failed to register hotplug callbacks.\n");
3169 for_each_node_state(nid, N_MEMORY)
3171 register_sysctl_init("vm", vm_compaction);
3174 subsys_initcall(kcompactd_init)
3176 #endif /* CONFIG_COMPACTION */