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
29 static inline void count_compact_event(enum vm_event_item item)
34 static inline void count_compact_events(enum vm_event_item item, long delta)
36 count_vm_events(item, delta);
39 #define count_compact_event(item) do { } while (0)
40 #define count_compact_events(item, delta) do { } while (0)
43 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
45 #define CREATE_TRACE_POINTS
46 #include <trace/events/compaction.h>
48 #define block_start_pfn(pfn, order) round_down(pfn, 1UL << (order))
49 #define block_end_pfn(pfn, order) ALIGN((pfn) + 1, 1UL << (order))
50 #define pageblock_start_pfn(pfn) block_start_pfn(pfn, pageblock_order)
51 #define pageblock_end_pfn(pfn) block_end_pfn(pfn, pageblock_order)
53 static unsigned long release_freepages(struct list_head *freelist)
55 struct page *page, *next;
56 unsigned long high_pfn = 0;
58 list_for_each_entry_safe(page, next, freelist, lru) {
59 unsigned long pfn = page_to_pfn(page);
69 static void split_map_pages(struct list_head *list)
71 unsigned int i, order, nr_pages;
72 struct page *page, *next;
75 list_for_each_entry_safe(page, next, list, lru) {
78 order = page_private(page);
79 nr_pages = 1 << order;
81 post_alloc_hook(page, order, __GFP_MOVABLE);
83 split_page(page, order);
85 for (i = 0; i < nr_pages; i++) {
86 list_add(&page->lru, &tmp_list);
91 list_splice(&tmp_list, list);
94 #ifdef CONFIG_COMPACTION
96 int PageMovable(struct page *page)
98 struct address_space *mapping;
100 VM_BUG_ON_PAGE(!PageLocked(page), page);
101 if (!__PageMovable(page))
104 mapping = page_mapping(page);
105 if (mapping && mapping->a_ops && mapping->a_ops->isolate_page)
110 EXPORT_SYMBOL(PageMovable);
112 void __SetPageMovable(struct page *page, struct address_space *mapping)
114 VM_BUG_ON_PAGE(!PageLocked(page), page);
115 VM_BUG_ON_PAGE((unsigned long)mapping & PAGE_MAPPING_MOVABLE, page);
116 page->mapping = (void *)((unsigned long)mapping | PAGE_MAPPING_MOVABLE);
118 EXPORT_SYMBOL(__SetPageMovable);
120 void __ClearPageMovable(struct page *page)
122 VM_BUG_ON_PAGE(!PageLocked(page), page);
123 VM_BUG_ON_PAGE(!PageMovable(page), page);
125 * Clear registered address_space val with keeping PAGE_MAPPING_MOVABLE
126 * flag so that VM can catch up released page by driver after isolation.
127 * With it, VM migration doesn't try to put it back.
129 page->mapping = (void *)((unsigned long)page->mapping &
130 PAGE_MAPPING_MOVABLE);
132 EXPORT_SYMBOL(__ClearPageMovable);
134 /* Do not skip compaction more than 64 times */
135 #define COMPACT_MAX_DEFER_SHIFT 6
138 * Compaction is deferred when compaction fails to result in a page
139 * allocation success. 1 << compact_defer_limit compactions are skipped up
140 * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
142 void defer_compaction(struct zone *zone, int order)
144 zone->compact_considered = 0;
145 zone->compact_defer_shift++;
147 if (order < zone->compact_order_failed)
148 zone->compact_order_failed = order;
150 if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
151 zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
153 trace_mm_compaction_defer_compaction(zone, order);
156 /* Returns true if compaction should be skipped this time */
157 bool compaction_deferred(struct zone *zone, int order)
159 unsigned long defer_limit = 1UL << zone->compact_defer_shift;
161 if (order < zone->compact_order_failed)
164 /* Avoid possible overflow */
165 if (++zone->compact_considered > defer_limit)
166 zone->compact_considered = defer_limit;
168 if (zone->compact_considered >= defer_limit)
171 trace_mm_compaction_deferred(zone, order);
177 * Update defer tracking counters after successful compaction of given order,
178 * which means an allocation either succeeded (alloc_success == true) or is
179 * expected to succeed.
181 void compaction_defer_reset(struct zone *zone, int order,
185 zone->compact_considered = 0;
186 zone->compact_defer_shift = 0;
188 if (order >= zone->compact_order_failed)
189 zone->compact_order_failed = order + 1;
191 trace_mm_compaction_defer_reset(zone, order);
194 /* Returns true if restarting compaction after many failures */
195 bool compaction_restarting(struct zone *zone, int order)
197 if (order < zone->compact_order_failed)
200 return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
201 zone->compact_considered >= 1UL << zone->compact_defer_shift;
204 /* Returns true if the pageblock should be scanned for pages to isolate. */
205 static inline bool isolation_suitable(struct compact_control *cc,
208 if (cc->ignore_skip_hint)
211 return !get_pageblock_skip(page);
214 static void reset_cached_positions(struct zone *zone)
216 zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
217 zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
218 zone->compact_cached_free_pfn =
219 pageblock_start_pfn(zone_end_pfn(zone) - 1);
223 * Compound pages of >= pageblock_order should consistenly be skipped until
224 * released. It is always pointless to compact pages of such order (if they are
225 * migratable), and the pageblocks they occupy cannot contain any free pages.
227 static bool pageblock_skip_persistent(struct page *page)
229 if (!PageCompound(page))
232 page = compound_head(page);
234 if (compound_order(page) >= pageblock_order)
241 __reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source,
244 struct page *page = pfn_to_online_page(pfn);
245 struct page *block_page;
246 struct page *end_page;
247 unsigned long block_pfn;
251 if (zone != page_zone(page))
253 if (pageblock_skip_persistent(page))
257 * If skip is already cleared do no further checking once the
258 * restart points have been set.
260 if (check_source && check_target && !get_pageblock_skip(page))
264 * If clearing skip for the target scanner, do not select a
265 * non-movable pageblock as the starting point.
267 if (!check_source && check_target &&
268 get_pageblock_migratetype(page) != MIGRATE_MOVABLE)
271 /* Ensure the start of the pageblock or zone is online and valid */
272 block_pfn = pageblock_start_pfn(pfn);
273 block_page = pfn_to_online_page(max(block_pfn, zone->zone_start_pfn));
279 /* Ensure the end of the pageblock or zone is online and valid */
280 block_pfn += pageblock_nr_pages;
281 block_pfn = min(block_pfn, zone_end_pfn(zone) - 1);
282 end_page = pfn_to_online_page(block_pfn);
287 * Only clear the hint if a sample indicates there is either a
288 * free page or an LRU page in the block. One or other condition
289 * is necessary for the block to be a migration source/target.
292 if (pfn_valid_within(pfn)) {
293 if (check_source && PageLRU(page)) {
294 clear_pageblock_skip(page);
298 if (check_target && PageBuddy(page)) {
299 clear_pageblock_skip(page);
304 page += (1 << PAGE_ALLOC_COSTLY_ORDER);
305 pfn += (1 << PAGE_ALLOC_COSTLY_ORDER);
306 } while (page < end_page);
312 * This function is called to clear all cached information on pageblocks that
313 * should be skipped for page isolation when the migrate and free page scanner
316 static void __reset_isolation_suitable(struct zone *zone)
318 unsigned long migrate_pfn = zone->zone_start_pfn;
319 unsigned long free_pfn = zone_end_pfn(zone) - 1;
320 unsigned long reset_migrate = free_pfn;
321 unsigned long reset_free = migrate_pfn;
322 bool source_set = false;
323 bool free_set = false;
325 if (!zone->compact_blockskip_flush)
328 zone->compact_blockskip_flush = false;
331 * Walk the zone and update pageblock skip information. Source looks
332 * for PageLRU while target looks for PageBuddy. When the scanner
333 * is found, both PageBuddy and PageLRU are checked as the pageblock
334 * is suitable as both source and target.
336 for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages,
337 free_pfn -= pageblock_nr_pages) {
340 /* Update the migrate PFN */
341 if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) &&
342 migrate_pfn < reset_migrate) {
344 reset_migrate = migrate_pfn;
345 zone->compact_init_migrate_pfn = reset_migrate;
346 zone->compact_cached_migrate_pfn[0] = reset_migrate;
347 zone->compact_cached_migrate_pfn[1] = reset_migrate;
350 /* Update the free PFN */
351 if (__reset_isolation_pfn(zone, free_pfn, free_set, true) &&
352 free_pfn > reset_free) {
354 reset_free = free_pfn;
355 zone->compact_init_free_pfn = reset_free;
356 zone->compact_cached_free_pfn = reset_free;
360 /* Leave no distance if no suitable block was reset */
361 if (reset_migrate >= reset_free) {
362 zone->compact_cached_migrate_pfn[0] = migrate_pfn;
363 zone->compact_cached_migrate_pfn[1] = migrate_pfn;
364 zone->compact_cached_free_pfn = free_pfn;
368 void reset_isolation_suitable(pg_data_t *pgdat)
372 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
373 struct zone *zone = &pgdat->node_zones[zoneid];
374 if (!populated_zone(zone))
377 /* Only flush if a full compaction finished recently */
378 if (zone->compact_blockskip_flush)
379 __reset_isolation_suitable(zone);
384 * Sets the pageblock skip bit if it was clear. Note that this is a hint as
385 * locks are not required for read/writers. Returns true if it was already set.
387 static bool test_and_set_skip(struct compact_control *cc, struct page *page,
392 /* Do no update if skip hint is being ignored */
393 if (cc->ignore_skip_hint)
396 if (!IS_ALIGNED(pfn, pageblock_nr_pages))
399 skip = get_pageblock_skip(page);
400 if (!skip && !cc->no_set_skip_hint)
401 set_pageblock_skip(page);
406 static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
408 struct zone *zone = cc->zone;
410 pfn = pageblock_end_pfn(pfn);
412 /* Set for isolation rather than compaction */
413 if (cc->no_set_skip_hint)
416 if (pfn > zone->compact_cached_migrate_pfn[0])
417 zone->compact_cached_migrate_pfn[0] = pfn;
418 if (cc->mode != MIGRATE_ASYNC &&
419 pfn > zone->compact_cached_migrate_pfn[1])
420 zone->compact_cached_migrate_pfn[1] = pfn;
424 * If no pages were isolated then mark this pageblock to be skipped in the
425 * future. The information is later cleared by __reset_isolation_suitable().
427 static void update_pageblock_skip(struct compact_control *cc,
428 struct page *page, unsigned long pfn)
430 struct zone *zone = cc->zone;
432 if (cc->no_set_skip_hint)
438 set_pageblock_skip(page);
440 /* Update where async and sync compaction should restart */
441 if (pfn < zone->compact_cached_free_pfn)
442 zone->compact_cached_free_pfn = pfn;
445 static inline bool isolation_suitable(struct compact_control *cc,
451 static inline bool pageblock_skip_persistent(struct page *page)
456 static inline void update_pageblock_skip(struct compact_control *cc,
457 struct page *page, unsigned long pfn)
461 static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
465 static bool test_and_set_skip(struct compact_control *cc, struct page *page,
470 #endif /* CONFIG_COMPACTION */
473 * Compaction requires the taking of some coarse locks that are potentially
474 * very heavily contended. For async compaction, trylock and record if the
475 * lock is contended. The lock will still be acquired but compaction will
476 * abort when the current block is finished regardless of success rate.
477 * Sync compaction acquires the lock.
479 * Always returns true which makes it easier to track lock state in callers.
481 static bool compact_lock_irqsave(spinlock_t *lock, unsigned long *flags,
482 struct compact_control *cc)
484 /* Track if the lock is contended in async mode */
485 if (cc->mode == MIGRATE_ASYNC && !cc->contended) {
486 if (spin_trylock_irqsave(lock, *flags))
489 cc->contended = true;
492 spin_lock_irqsave(lock, *flags);
497 * Compaction requires the taking of some coarse locks that are potentially
498 * very heavily contended. The lock should be periodically unlocked to avoid
499 * having disabled IRQs for a long time, even when there is nobody waiting on
500 * the lock. It might also be that allowing the IRQs will result in
501 * need_resched() becoming true. If scheduling is needed, async compaction
502 * aborts. Sync compaction schedules.
503 * Either compaction type will also abort if a fatal signal is pending.
504 * In either case if the lock was locked, it is dropped and not regained.
506 * Returns true if compaction should abort due to fatal signal pending, or
507 * async compaction due to need_resched()
508 * Returns false when compaction can continue (sync compaction might have
511 static bool compact_unlock_should_abort(spinlock_t *lock,
512 unsigned long flags, bool *locked, struct compact_control *cc)
515 spin_unlock_irqrestore(lock, flags);
519 if (fatal_signal_pending(current)) {
520 cc->contended = true;
530 * Isolate free pages onto a private freelist. If @strict is true, will abort
531 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
532 * (even though it may still end up isolating some pages).
534 static unsigned long isolate_freepages_block(struct compact_control *cc,
535 unsigned long *start_pfn,
536 unsigned long end_pfn,
537 struct list_head *freelist,
541 int nr_scanned = 0, total_isolated = 0;
543 unsigned long flags = 0;
545 unsigned long blockpfn = *start_pfn;
548 /* Strict mode is for isolation, speed is secondary */
552 cursor = pfn_to_page(blockpfn);
554 /* Isolate free pages. */
555 for (; blockpfn < end_pfn; blockpfn += stride, cursor += stride) {
557 struct page *page = cursor;
560 * Periodically drop the lock (if held) regardless of its
561 * contention, to give chance to IRQs. Abort if fatal signal
562 * pending or async compaction detects need_resched()
564 if (!(blockpfn % SWAP_CLUSTER_MAX)
565 && compact_unlock_should_abort(&cc->zone->lock, flags,
570 if (!pfn_valid_within(blockpfn))
574 * For compound pages such as THP and hugetlbfs, we can save
575 * potentially a lot of iterations if we skip them at once.
576 * The check is racy, but we can consider only valid values
577 * and the only danger is skipping too much.
579 if (PageCompound(page)) {
580 const unsigned int order = compound_order(page);
582 if (likely(order < MAX_ORDER)) {
583 blockpfn += (1UL << order) - 1;
584 cursor += (1UL << order) - 1;
589 if (!PageBuddy(page))
593 * If we already hold the lock, we can skip some rechecking.
594 * Note that if we hold the lock now, checked_pageblock was
595 * already set in some previous iteration (or strict is true),
596 * so it is correct to skip the suitable migration target
600 locked = compact_lock_irqsave(&cc->zone->lock,
603 /* Recheck this is a buddy page under lock */
604 if (!PageBuddy(page))
608 /* Found a free page, will break it into order-0 pages */
609 order = page_order(page);
610 isolated = __isolate_free_page(page, order);
613 set_page_private(page, order);
615 total_isolated += isolated;
616 cc->nr_freepages += isolated;
617 list_add_tail(&page->lru, freelist);
619 if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
620 blockpfn += isolated;
623 /* Advance to the end of split page */
624 blockpfn += isolated - 1;
625 cursor += isolated - 1;
637 spin_unlock_irqrestore(&cc->zone->lock, flags);
640 * There is a tiny chance that we have read bogus compound_order(),
641 * so be careful to not go outside of the pageblock.
643 if (unlikely(blockpfn > end_pfn))
646 trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
647 nr_scanned, total_isolated);
649 /* Record how far we have got within the block */
650 *start_pfn = blockpfn;
653 * If strict isolation is requested by CMA then check that all the
654 * pages requested were isolated. If there were any failures, 0 is
655 * returned and CMA will fail.
657 if (strict && blockpfn < end_pfn)
660 cc->total_free_scanned += nr_scanned;
662 count_compact_events(COMPACTISOLATED, total_isolated);
663 return total_isolated;
667 * isolate_freepages_range() - isolate free pages.
668 * @cc: Compaction control structure.
669 * @start_pfn: The first PFN to start isolating.
670 * @end_pfn: The one-past-last PFN.
672 * Non-free pages, invalid PFNs, or zone boundaries within the
673 * [start_pfn, end_pfn) range are considered errors, cause function to
674 * undo its actions and return zero.
676 * Otherwise, function returns one-past-the-last PFN of isolated page
677 * (which may be greater then end_pfn if end fell in a middle of
681 isolate_freepages_range(struct compact_control *cc,
682 unsigned long start_pfn, unsigned long end_pfn)
684 unsigned long isolated, pfn, block_start_pfn, block_end_pfn;
688 block_start_pfn = pageblock_start_pfn(pfn);
689 if (block_start_pfn < cc->zone->zone_start_pfn)
690 block_start_pfn = cc->zone->zone_start_pfn;
691 block_end_pfn = pageblock_end_pfn(pfn);
693 for (; pfn < end_pfn; pfn += isolated,
694 block_start_pfn = block_end_pfn,
695 block_end_pfn += pageblock_nr_pages) {
696 /* Protect pfn from changing by isolate_freepages_block */
697 unsigned long isolate_start_pfn = pfn;
699 block_end_pfn = min(block_end_pfn, end_pfn);
702 * pfn could pass the block_end_pfn if isolated freepage
703 * is more than pageblock order. In this case, we adjust
704 * scanning range to right one.
706 if (pfn >= block_end_pfn) {
707 block_start_pfn = pageblock_start_pfn(pfn);
708 block_end_pfn = pageblock_end_pfn(pfn);
709 block_end_pfn = min(block_end_pfn, end_pfn);
712 if (!pageblock_pfn_to_page(block_start_pfn,
713 block_end_pfn, cc->zone))
716 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
717 block_end_pfn, &freelist, 0, true);
720 * In strict mode, isolate_freepages_block() returns 0 if
721 * there are any holes in the block (ie. invalid PFNs or
728 * If we managed to isolate pages, it is always (1 << n) *
729 * pageblock_nr_pages for some non-negative n. (Max order
730 * page may span two pageblocks).
734 /* __isolate_free_page() does not map the pages */
735 split_map_pages(&freelist);
738 /* Loop terminated early, cleanup. */
739 release_freepages(&freelist);
743 /* We don't use freelists for anything. */
747 /* Similar to reclaim, but different enough that they don't share logic */
748 static bool too_many_isolated(pg_data_t *pgdat)
750 unsigned long active, inactive, isolated;
752 inactive = node_page_state(pgdat, NR_INACTIVE_FILE) +
753 node_page_state(pgdat, NR_INACTIVE_ANON);
754 active = node_page_state(pgdat, NR_ACTIVE_FILE) +
755 node_page_state(pgdat, NR_ACTIVE_ANON);
756 isolated = node_page_state(pgdat, NR_ISOLATED_FILE) +
757 node_page_state(pgdat, NR_ISOLATED_ANON);
759 return isolated > (inactive + active) / 2;
763 * isolate_migratepages_block() - isolate all migrate-able pages within
765 * @cc: Compaction control structure.
766 * @low_pfn: The first PFN to isolate
767 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
768 * @isolate_mode: Isolation mode to be used.
770 * Isolate all pages that can be migrated from the range specified by
771 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
772 * Returns zero if there is a fatal signal pending, otherwise PFN of the
773 * first page that was not scanned (which may be both less, equal to or more
776 * The pages are isolated on cc->migratepages list (not required to be empty),
777 * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
778 * is neither read nor updated.
781 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
782 unsigned long end_pfn, isolate_mode_t isolate_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;
789 struct page *page = NULL, *valid_page = NULL;
790 unsigned long start_pfn = low_pfn;
791 bool skip_on_failure = false;
792 unsigned long next_skip_pfn = 0;
793 bool skip_updated = false;
796 * Ensure that there are not too many pages isolated from the LRU
797 * list by either parallel reclaimers or compaction. If there are,
798 * delay for some time until fewer pages are isolated
800 while (unlikely(too_many_isolated(pgdat))) {
801 /* async migration should just abort */
802 if (cc->mode == MIGRATE_ASYNC)
805 congestion_wait(BLK_RW_ASYNC, HZ/10);
807 if (fatal_signal_pending(current))
813 if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
814 skip_on_failure = true;
815 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
818 /* Time to isolate some pages for migration */
819 for (; low_pfn < end_pfn; low_pfn++) {
821 if (skip_on_failure && low_pfn >= next_skip_pfn) {
823 * We have isolated all migration candidates in the
824 * previous order-aligned block, and did not skip it due
825 * to failure. We should migrate the pages now and
826 * hopefully succeed compaction.
832 * We failed to isolate in the previous order-aligned
833 * block. Set the new boundary to the end of the
834 * current block. Note we can't simply increase
835 * next_skip_pfn by 1 << order, as low_pfn might have
836 * been incremented by a higher number due to skipping
837 * a compound or a high-order buddy page in the
838 * previous loop iteration.
840 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
844 * Periodically drop the lock (if held) regardless of its
845 * contention, to give chance to IRQs. Abort completely if
846 * a fatal signal is pending.
848 if (!(low_pfn % SWAP_CLUSTER_MAX)
849 && compact_unlock_should_abort(&pgdat->lru_lock,
850 flags, &locked, cc)) {
855 if (!pfn_valid_within(low_pfn))
859 page = pfn_to_page(low_pfn);
862 * Check if the pageblock has already been marked skipped.
863 * Only the aligned PFN is checked as the caller isolates
864 * COMPACT_CLUSTER_MAX at a time so the second call must
865 * not falsely conclude that the block should be skipped.
867 if (!valid_page && IS_ALIGNED(low_pfn, pageblock_nr_pages)) {
868 if (!cc->ignore_skip_hint && get_pageblock_skip(page)) {
876 * Skip if free. We read page order here without zone lock
877 * which is generally unsafe, but the race window is small and
878 * the worst thing that can happen is that we skip some
879 * potential isolation targets.
881 if (PageBuddy(page)) {
882 unsigned long freepage_order = page_order_unsafe(page);
885 * Without lock, we cannot be sure that what we got is
886 * a valid page order. Consider only values in the
887 * valid order range to prevent low_pfn overflow.
889 if (freepage_order > 0 && freepage_order < MAX_ORDER)
890 low_pfn += (1UL << freepage_order) - 1;
895 * Regardless of being on LRU, compound pages such as THP and
896 * hugetlbfs are not to be compacted. We can potentially save
897 * a lot of iterations if we skip them at once. The check is
898 * racy, but we can consider only valid values and the only
899 * danger is skipping too much.
901 if (PageCompound(page)) {
902 const unsigned int order = compound_order(page);
904 if (likely(order < MAX_ORDER))
905 low_pfn += (1UL << order) - 1;
910 * Check may be lockless but that's ok as we recheck later.
911 * It's possible to migrate LRU and non-lru movable pages.
912 * Skip any other type of page
914 if (!PageLRU(page)) {
916 * __PageMovable can return false positive so we need
917 * to verify it under page_lock.
919 if (unlikely(__PageMovable(page)) &&
920 !PageIsolated(page)) {
922 spin_unlock_irqrestore(&pgdat->lru_lock,
927 if (!isolate_movable_page(page, isolate_mode))
928 goto isolate_success;
935 * Migration will fail if an anonymous page is pinned in memory,
936 * so avoid taking lru_lock and isolating it unnecessarily in an
937 * admittedly racy check.
939 if (!page_mapping(page) &&
940 page_count(page) > page_mapcount(page))
944 * Only allow to migrate anonymous pages in GFP_NOFS context
945 * because those do not depend on fs locks.
947 if (!(cc->gfp_mask & __GFP_FS) && page_mapping(page))
950 /* If we already hold the lock, we can skip some rechecking */
952 locked = compact_lock_irqsave(&pgdat->lru_lock,
955 /* Try get exclusive access under lock */
958 if (test_and_set_skip(cc, page, low_pfn))
962 /* Recheck PageLRU and PageCompound under lock */
967 * Page become compound since the non-locked check,
968 * and it's on LRU. It can only be a THP so the order
969 * is safe to read and it's 0 for tail pages.
971 if (unlikely(PageCompound(page))) {
972 low_pfn += compound_nr(page) - 1;
977 lruvec = mem_cgroup_page_lruvec(page, pgdat);
979 /* Try isolate the page */
980 if (__isolate_lru_page(page, isolate_mode) != 0)
983 VM_BUG_ON_PAGE(PageCompound(page), page);
985 /* Successfully isolated */
986 del_page_from_lru_list(page, lruvec, page_lru(page));
987 inc_node_page_state(page,
988 NR_ISOLATED_ANON + page_is_file_cache(page));
991 list_add(&page->lru, &cc->migratepages);
992 cc->nr_migratepages++;
996 * Avoid isolating too much unless this block is being
997 * rescanned (e.g. dirty/writeback pages, parallel allocation)
998 * or a lock is contended. For contention, isolate quickly to
999 * potentially remove one source of contention.
1001 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX &&
1002 !cc->rescan && !cc->contended) {
1009 if (!skip_on_failure)
1013 * We have isolated some pages, but then failed. Release them
1014 * instead of migrating, as we cannot form the cc->order buddy
1019 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
1022 putback_movable_pages(&cc->migratepages);
1023 cc->nr_migratepages = 0;
1027 if (low_pfn < next_skip_pfn) {
1028 low_pfn = next_skip_pfn - 1;
1030 * The check near the loop beginning would have updated
1031 * next_skip_pfn too, but this is a bit simpler.
1033 next_skip_pfn += 1UL << cc->order;
1038 * The PageBuddy() check could have potentially brought us outside
1039 * the range to be scanned.
1041 if (unlikely(low_pfn > end_pfn))
1046 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
1049 * Updated the cached scanner pfn once the pageblock has been scanned
1050 * Pages will either be migrated in which case there is no point
1051 * scanning in the near future or migration failed in which case the
1052 * failure reason may persist. The block is marked for skipping if
1053 * there were no pages isolated in the block or if the block is
1054 * rescanned twice in a row.
1056 if (low_pfn == end_pfn && (!nr_isolated || cc->rescan)) {
1057 if (valid_page && !skip_updated)
1058 set_pageblock_skip(valid_page);
1059 update_cached_migrate(cc, low_pfn);
1062 trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
1063 nr_scanned, nr_isolated);
1066 cc->total_migrate_scanned += nr_scanned;
1068 count_compact_events(COMPACTISOLATED, nr_isolated);
1074 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
1075 * @cc: Compaction control structure.
1076 * @start_pfn: The first PFN to start isolating.
1077 * @end_pfn: The one-past-last PFN.
1079 * Returns zero if isolation fails fatally due to e.g. pending signal.
1080 * Otherwise, function returns one-past-the-last PFN of isolated page
1081 * (which may be greater than end_pfn if end fell in a middle of a THP page).
1084 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
1085 unsigned long end_pfn)
1087 unsigned long pfn, block_start_pfn, block_end_pfn;
1089 /* Scan block by block. First and last block may be incomplete */
1091 block_start_pfn = pageblock_start_pfn(pfn);
1092 if (block_start_pfn < cc->zone->zone_start_pfn)
1093 block_start_pfn = cc->zone->zone_start_pfn;
1094 block_end_pfn = pageblock_end_pfn(pfn);
1096 for (; pfn < end_pfn; pfn = block_end_pfn,
1097 block_start_pfn = block_end_pfn,
1098 block_end_pfn += pageblock_nr_pages) {
1100 block_end_pfn = min(block_end_pfn, end_pfn);
1102 if (!pageblock_pfn_to_page(block_start_pfn,
1103 block_end_pfn, cc->zone))
1106 pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
1107 ISOLATE_UNEVICTABLE);
1112 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
1119 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
1120 #ifdef CONFIG_COMPACTION
1122 static bool suitable_migration_source(struct compact_control *cc,
1127 if (pageblock_skip_persistent(page))
1130 if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction)
1133 block_mt = get_pageblock_migratetype(page);
1135 if (cc->migratetype == MIGRATE_MOVABLE)
1136 return is_migrate_movable(block_mt);
1138 return block_mt == cc->migratetype;
1141 /* Returns true if the page is within a block suitable for migration to */
1142 static bool suitable_migration_target(struct compact_control *cc,
1145 /* If the page is a large free page, then disallow migration */
1146 if (PageBuddy(page)) {
1148 * We are checking page_order without zone->lock taken. But
1149 * the only small danger is that we skip a potentially suitable
1150 * pageblock, so it's not worth to check order for valid range.
1152 if (page_order_unsafe(page) >= pageblock_order)
1156 if (cc->ignore_block_suitable)
1159 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1160 if (is_migrate_movable(get_pageblock_migratetype(page)))
1163 /* Otherwise skip the block */
1167 static inline unsigned int
1168 freelist_scan_limit(struct compact_control *cc)
1170 unsigned short shift = BITS_PER_LONG - 1;
1172 return (COMPACT_CLUSTER_MAX >> min(shift, cc->fast_search_fail)) + 1;
1176 * Test whether the free scanner has reached the same or lower pageblock than
1177 * the migration scanner, and compaction should thus terminate.
1179 static inline bool compact_scanners_met(struct compact_control *cc)
1181 return (cc->free_pfn >> pageblock_order)
1182 <= (cc->migrate_pfn >> pageblock_order);
1186 * Used when scanning for a suitable migration target which scans freelists
1187 * in reverse. Reorders the list such as the unscanned pages are scanned
1188 * first on the next iteration of the free scanner
1191 move_freelist_head(struct list_head *freelist, struct page *freepage)
1195 if (!list_is_last(freelist, &freepage->lru)) {
1196 list_cut_before(&sublist, freelist, &freepage->lru);
1197 if (!list_empty(&sublist))
1198 list_splice_tail(&sublist, freelist);
1203 * Similar to move_freelist_head except used by the migration scanner
1204 * when scanning forward. It's possible for these list operations to
1205 * move against each other if they search the free list exactly in
1209 move_freelist_tail(struct list_head *freelist, struct page *freepage)
1213 if (!list_is_first(freelist, &freepage->lru)) {
1214 list_cut_position(&sublist, freelist, &freepage->lru);
1215 if (!list_empty(&sublist))
1216 list_splice_tail(&sublist, freelist);
1221 fast_isolate_around(struct compact_control *cc, unsigned long pfn, unsigned long nr_isolated)
1223 unsigned long start_pfn, end_pfn;
1224 struct page *page = pfn_to_page(pfn);
1226 /* Do not search around if there are enough pages already */
1227 if (cc->nr_freepages >= cc->nr_migratepages)
1230 /* Minimise scanning during async compaction */
1231 if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC)
1234 /* Pageblock boundaries */
1235 start_pfn = pageblock_start_pfn(pfn);
1236 end_pfn = min(pageblock_end_pfn(pfn), zone_end_pfn(cc->zone)) - 1;
1239 if (start_pfn != pfn) {
1240 isolate_freepages_block(cc, &start_pfn, pfn, &cc->freepages, 1, false);
1241 if (cc->nr_freepages >= cc->nr_migratepages)
1246 start_pfn = pfn + nr_isolated;
1247 if (start_pfn < end_pfn)
1248 isolate_freepages_block(cc, &start_pfn, end_pfn, &cc->freepages, 1, false);
1250 /* Skip this pageblock in the future as it's full or nearly full */
1251 if (cc->nr_freepages < cc->nr_migratepages)
1252 set_pageblock_skip(page);
1255 /* Search orders in round-robin fashion */
1256 static int next_search_order(struct compact_control *cc, int order)
1260 order = cc->order - 1;
1262 /* Search wrapped around? */
1263 if (order == cc->search_order) {
1265 if (cc->search_order < 0)
1266 cc->search_order = cc->order - 1;
1273 static unsigned long
1274 fast_isolate_freepages(struct compact_control *cc)
1276 unsigned int limit = min(1U, freelist_scan_limit(cc) >> 1);
1277 unsigned int nr_scanned = 0;
1278 unsigned long low_pfn, min_pfn, high_pfn = 0, highest = 0;
1279 unsigned long nr_isolated = 0;
1280 unsigned long distance;
1281 struct page *page = NULL;
1282 bool scan_start = false;
1285 /* Full compaction passes in a negative order */
1287 return cc->free_pfn;
1290 * If starting the scan, use a deeper search and use the highest
1291 * PFN found if a suitable one is not found.
1293 if (cc->free_pfn >= cc->zone->compact_init_free_pfn) {
1294 limit = pageblock_nr_pages >> 1;
1299 * Preferred point is in the top quarter of the scan space but take
1300 * a pfn from the top half if the search is problematic.
1302 distance = (cc->free_pfn - cc->migrate_pfn);
1303 low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2));
1304 min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1));
1306 if (WARN_ON_ONCE(min_pfn > low_pfn))
1310 * Search starts from the last successful isolation order or the next
1311 * order to search after a previous failure
1313 cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order);
1315 for (order = cc->search_order;
1316 !page && order >= 0;
1317 order = next_search_order(cc, order)) {
1318 struct free_area *area = &cc->zone->free_area[order];
1319 struct list_head *freelist;
1320 struct page *freepage;
1321 unsigned long flags;
1322 unsigned int order_scanned = 0;
1327 spin_lock_irqsave(&cc->zone->lock, flags);
1328 freelist = &area->free_list[MIGRATE_MOVABLE];
1329 list_for_each_entry_reverse(freepage, freelist, lru) {
1334 pfn = page_to_pfn(freepage);
1337 highest = pageblock_start_pfn(pfn);
1339 if (pfn >= low_pfn) {
1340 cc->fast_search_fail = 0;
1341 cc->search_order = order;
1346 if (pfn >= min_pfn && pfn > high_pfn) {
1349 /* Shorten the scan if a candidate is found */
1353 if (order_scanned >= limit)
1357 /* Use a minimum pfn if a preferred one was not found */
1358 if (!page && high_pfn) {
1359 page = pfn_to_page(high_pfn);
1361 /* Update freepage for the list reorder below */
1365 /* Reorder to so a future search skips recent pages */
1366 move_freelist_head(freelist, freepage);
1368 /* Isolate the page if available */
1370 if (__isolate_free_page(page, order)) {
1371 set_page_private(page, order);
1372 nr_isolated = 1 << order;
1373 cc->nr_freepages += nr_isolated;
1374 list_add_tail(&page->lru, &cc->freepages);
1375 count_compact_events(COMPACTISOLATED, nr_isolated);
1377 /* If isolation fails, abort the search */
1378 order = cc->search_order + 1;
1383 spin_unlock_irqrestore(&cc->zone->lock, flags);
1386 * Smaller scan on next order so the total scan ig related
1387 * to freelist_scan_limit.
1389 if (order_scanned >= limit)
1390 limit = min(1U, limit >> 1);
1394 cc->fast_search_fail++;
1397 * Use the highest PFN found above min. If one was
1398 * not found, be pessemistic for direct compaction
1399 * and use the min mark.
1402 page = pfn_to_page(highest);
1403 cc->free_pfn = highest;
1405 if (cc->direct_compaction && pfn_valid(min_pfn)) {
1406 page = pfn_to_page(min_pfn);
1407 cc->free_pfn = min_pfn;
1413 if (highest && highest >= cc->zone->compact_cached_free_pfn) {
1414 highest -= pageblock_nr_pages;
1415 cc->zone->compact_cached_free_pfn = highest;
1418 cc->total_free_scanned += nr_scanned;
1420 return cc->free_pfn;
1422 low_pfn = page_to_pfn(page);
1423 fast_isolate_around(cc, low_pfn, nr_isolated);
1428 * Based on information in the current compact_control, find blocks
1429 * suitable for isolating free pages from and then isolate them.
1431 static void isolate_freepages(struct compact_control *cc)
1433 struct zone *zone = cc->zone;
1435 unsigned long block_start_pfn; /* start of current pageblock */
1436 unsigned long isolate_start_pfn; /* exact pfn we start at */
1437 unsigned long block_end_pfn; /* end of current pageblock */
1438 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
1439 struct list_head *freelist = &cc->freepages;
1440 unsigned int stride;
1442 /* Try a small search of the free lists for a candidate */
1443 isolate_start_pfn = fast_isolate_freepages(cc);
1444 if (cc->nr_freepages)
1448 * Initialise the free scanner. The starting point is where we last
1449 * successfully isolated from, zone-cached value, or the end of the
1450 * zone when isolating for the first time. For looping we also need
1451 * this pfn aligned down to the pageblock boundary, because we do
1452 * block_start_pfn -= pageblock_nr_pages in the for loop.
1453 * For ending point, take care when isolating in last pageblock of a
1454 * a zone which ends in the middle of a pageblock.
1455 * The low boundary is the end of the pageblock the migration scanner
1458 isolate_start_pfn = cc->free_pfn;
1459 block_start_pfn = pageblock_start_pfn(isolate_start_pfn);
1460 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1461 zone_end_pfn(zone));
1462 low_pfn = pageblock_end_pfn(cc->migrate_pfn);
1463 stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1;
1466 * Isolate free pages until enough are available to migrate the
1467 * pages on cc->migratepages. We stop searching if the migrate
1468 * and free page scanners meet or enough free pages are isolated.
1470 for (; block_start_pfn >= low_pfn;
1471 block_end_pfn = block_start_pfn,
1472 block_start_pfn -= pageblock_nr_pages,
1473 isolate_start_pfn = block_start_pfn) {
1474 unsigned long nr_isolated;
1477 * This can iterate a massively long zone without finding any
1478 * suitable migration targets, so periodically check resched.
1480 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)))
1483 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1488 /* Check the block is suitable for migration */
1489 if (!suitable_migration_target(cc, page))
1492 /* If isolation recently failed, do not retry */
1493 if (!isolation_suitable(cc, page))
1496 /* Found a block suitable for isolating free pages from. */
1497 nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn,
1498 block_end_pfn, freelist, stride, false);
1500 /* Update the skip hint if the full pageblock was scanned */
1501 if (isolate_start_pfn == block_end_pfn)
1502 update_pageblock_skip(cc, page, block_start_pfn);
1504 /* Are enough freepages isolated? */
1505 if (cc->nr_freepages >= cc->nr_migratepages) {
1506 if (isolate_start_pfn >= block_end_pfn) {
1508 * Restart at previous pageblock if more
1509 * freepages can be isolated next time.
1512 block_start_pfn - pageblock_nr_pages;
1515 } else if (isolate_start_pfn < block_end_pfn) {
1517 * If isolation failed early, do not continue
1523 /* Adjust stride depending on isolation */
1528 stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1);
1532 * Record where the free scanner will restart next time. Either we
1533 * broke from the loop and set isolate_start_pfn based on the last
1534 * call to isolate_freepages_block(), or we met the migration scanner
1535 * and the loop terminated due to isolate_start_pfn < low_pfn
1537 cc->free_pfn = isolate_start_pfn;
1540 /* __isolate_free_page() does not map the pages */
1541 split_map_pages(freelist);
1545 * This is a migrate-callback that "allocates" freepages by taking pages
1546 * from the isolated freelists in the block we are migrating to.
1548 static struct page *compaction_alloc(struct page *migratepage,
1551 struct compact_control *cc = (struct compact_control *)data;
1552 struct page *freepage;
1554 if (list_empty(&cc->freepages)) {
1555 isolate_freepages(cc);
1557 if (list_empty(&cc->freepages))
1561 freepage = list_entry(cc->freepages.next, struct page, lru);
1562 list_del(&freepage->lru);
1569 * This is a migrate-callback that "frees" freepages back to the isolated
1570 * freelist. All pages on the freelist are from the same zone, so there is no
1571 * special handling needed for NUMA.
1573 static void compaction_free(struct page *page, unsigned long data)
1575 struct compact_control *cc = (struct compact_control *)data;
1577 list_add(&page->lru, &cc->freepages);
1581 /* possible outcome of isolate_migratepages */
1583 ISOLATE_ABORT, /* Abort compaction now */
1584 ISOLATE_NONE, /* No pages isolated, continue scanning */
1585 ISOLATE_SUCCESS, /* Pages isolated, migrate */
1586 } isolate_migrate_t;
1589 * Allow userspace to control policy on scanning the unevictable LRU for
1590 * compactable pages.
1592 int sysctl_compact_unevictable_allowed __read_mostly = 1;
1595 update_fast_start_pfn(struct compact_control *cc, unsigned long pfn)
1597 if (cc->fast_start_pfn == ULONG_MAX)
1600 if (!cc->fast_start_pfn)
1601 cc->fast_start_pfn = pfn;
1603 cc->fast_start_pfn = min(cc->fast_start_pfn, pfn);
1606 static inline unsigned long
1607 reinit_migrate_pfn(struct compact_control *cc)
1609 if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX)
1610 return cc->migrate_pfn;
1612 cc->migrate_pfn = cc->fast_start_pfn;
1613 cc->fast_start_pfn = ULONG_MAX;
1615 return cc->migrate_pfn;
1619 * Briefly search the free lists for a migration source that already has
1620 * some free pages to reduce the number of pages that need migration
1621 * before a pageblock is free.
1623 static unsigned long fast_find_migrateblock(struct compact_control *cc)
1625 unsigned int limit = freelist_scan_limit(cc);
1626 unsigned int nr_scanned = 0;
1627 unsigned long distance;
1628 unsigned long pfn = cc->migrate_pfn;
1629 unsigned long high_pfn;
1632 /* Skip hints are relied on to avoid repeats on the fast search */
1633 if (cc->ignore_skip_hint)
1637 * If the migrate_pfn is not at the start of a zone or the start
1638 * of a pageblock then assume this is a continuation of a previous
1639 * scan restarted due to COMPACT_CLUSTER_MAX.
1641 if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn))
1645 * For smaller orders, just linearly scan as the number of pages
1646 * to migrate should be relatively small and does not necessarily
1647 * justify freeing up a large block for a small allocation.
1649 if (cc->order <= PAGE_ALLOC_COSTLY_ORDER)
1653 * Only allow kcompactd and direct requests for movable pages to
1654 * quickly clear out a MOVABLE pageblock for allocation. This
1655 * reduces the risk that a large movable pageblock is freed for
1656 * an unmovable/reclaimable small allocation.
1658 if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE)
1662 * When starting the migration scanner, pick any pageblock within the
1663 * first half of the search space. Otherwise try and pick a pageblock
1664 * within the first eighth to reduce the chances that a migration
1665 * target later becomes a source.
1667 distance = (cc->free_pfn - cc->migrate_pfn) >> 1;
1668 if (cc->migrate_pfn != cc->zone->zone_start_pfn)
1670 high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance);
1672 for (order = cc->order - 1;
1673 order >= PAGE_ALLOC_COSTLY_ORDER && pfn == cc->migrate_pfn && nr_scanned < limit;
1675 struct free_area *area = &cc->zone->free_area[order];
1676 struct list_head *freelist;
1677 unsigned long flags;
1678 struct page *freepage;
1683 spin_lock_irqsave(&cc->zone->lock, flags);
1684 freelist = &area->free_list[MIGRATE_MOVABLE];
1685 list_for_each_entry(freepage, freelist, lru) {
1686 unsigned long free_pfn;
1689 free_pfn = page_to_pfn(freepage);
1690 if (free_pfn < high_pfn) {
1692 * Avoid if skipped recently. Ideally it would
1693 * move to the tail but even safe iteration of
1694 * the list assumes an entry is deleted, not
1697 if (get_pageblock_skip(freepage)) {
1698 if (list_is_last(freelist, &freepage->lru))
1704 /* Reorder to so a future search skips recent pages */
1705 move_freelist_tail(freelist, freepage);
1707 update_fast_start_pfn(cc, free_pfn);
1708 pfn = pageblock_start_pfn(free_pfn);
1709 cc->fast_search_fail = 0;
1710 set_pageblock_skip(freepage);
1714 if (nr_scanned >= limit) {
1715 cc->fast_search_fail++;
1716 move_freelist_tail(freelist, freepage);
1720 spin_unlock_irqrestore(&cc->zone->lock, flags);
1723 cc->total_migrate_scanned += nr_scanned;
1726 * If fast scanning failed then use a cached entry for a page block
1727 * that had free pages as the basis for starting a linear scan.
1729 if (pfn == cc->migrate_pfn)
1730 pfn = reinit_migrate_pfn(cc);
1736 * Isolate all pages that can be migrated from the first suitable block,
1737 * starting at the block pointed to by the migrate scanner pfn within
1740 static isolate_migrate_t isolate_migratepages(struct zone *zone,
1741 struct compact_control *cc)
1743 unsigned long block_start_pfn;
1744 unsigned long block_end_pfn;
1745 unsigned long low_pfn;
1747 const isolate_mode_t isolate_mode =
1748 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1749 (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1750 bool fast_find_block;
1753 * Start at where we last stopped, or beginning of the zone as
1754 * initialized by compact_zone(). The first failure will use
1755 * the lowest PFN as the starting point for linear scanning.
1757 low_pfn = fast_find_migrateblock(cc);
1758 block_start_pfn = pageblock_start_pfn(low_pfn);
1759 if (block_start_pfn < zone->zone_start_pfn)
1760 block_start_pfn = zone->zone_start_pfn;
1763 * fast_find_migrateblock marks a pageblock skipped so to avoid
1764 * the isolation_suitable check below, check whether the fast
1765 * search was successful.
1767 fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail;
1769 /* Only scan within a pageblock boundary */
1770 block_end_pfn = pageblock_end_pfn(low_pfn);
1773 * Iterate over whole pageblocks until we find the first suitable.
1774 * Do not cross the free scanner.
1776 for (; block_end_pfn <= cc->free_pfn;
1777 fast_find_block = false,
1778 low_pfn = block_end_pfn,
1779 block_start_pfn = block_end_pfn,
1780 block_end_pfn += pageblock_nr_pages) {
1783 * This can potentially iterate a massively long zone with
1784 * many pageblocks unsuitable, so periodically check if we
1787 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)))
1790 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1796 * If isolation recently failed, do not retry. Only check the
1797 * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock
1798 * to be visited multiple times. Assume skip was checked
1799 * before making it "skip" so other compaction instances do
1800 * not scan the same block.
1802 if (IS_ALIGNED(low_pfn, pageblock_nr_pages) &&
1803 !fast_find_block && !isolation_suitable(cc, page))
1807 * For async compaction, also only scan in MOVABLE blocks
1808 * without huge pages. Async compaction is optimistic to see
1809 * if the minimum amount of work satisfies the allocation.
1810 * The cached PFN is updated as it's possible that all
1811 * remaining blocks between source and target are unsuitable
1812 * and the compaction scanners fail to meet.
1814 if (!suitable_migration_source(cc, page)) {
1815 update_cached_migrate(cc, block_end_pfn);
1819 /* Perform the isolation */
1820 low_pfn = isolate_migratepages_block(cc, low_pfn,
1821 block_end_pfn, isolate_mode);
1824 return ISOLATE_ABORT;
1827 * Either we isolated something and proceed with migration. Or
1828 * we failed and compact_zone should decide if we should
1834 /* Record where migration scanner will be restarted. */
1835 cc->migrate_pfn = low_pfn;
1837 return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
1841 * order == -1 is expected when compacting via
1842 * /proc/sys/vm/compact_memory
1844 static inline bool is_via_compact_memory(int order)
1849 static enum compact_result __compact_finished(struct compact_control *cc)
1852 const int migratetype = cc->migratetype;
1855 /* Compaction run completes if the migrate and free scanner meet */
1856 if (compact_scanners_met(cc)) {
1857 /* Let the next compaction start anew. */
1858 reset_cached_positions(cc->zone);
1861 * Mark that the PG_migrate_skip information should be cleared
1862 * by kswapd when it goes to sleep. kcompactd does not set the
1863 * flag itself as the decision to be clear should be directly
1864 * based on an allocation request.
1866 if (cc->direct_compaction)
1867 cc->zone->compact_blockskip_flush = true;
1870 return COMPACT_COMPLETE;
1872 return COMPACT_PARTIAL_SKIPPED;
1875 if (is_via_compact_memory(cc->order))
1876 return COMPACT_CONTINUE;
1879 * Always finish scanning a pageblock to reduce the possibility of
1880 * fallbacks in the future. This is particularly important when
1881 * migration source is unmovable/reclaimable but it's not worth
1884 if (!IS_ALIGNED(cc->migrate_pfn, pageblock_nr_pages))
1885 return COMPACT_CONTINUE;
1887 /* Direct compactor: Is a suitable page free? */
1888 ret = COMPACT_NO_SUITABLE_PAGE;
1889 for (order = cc->order; order < MAX_ORDER; order++) {
1890 struct free_area *area = &cc->zone->free_area[order];
1893 /* Job done if page is free of the right migratetype */
1894 if (!free_area_empty(area, migratetype))
1895 return COMPACT_SUCCESS;
1898 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
1899 if (migratetype == MIGRATE_MOVABLE &&
1900 !free_area_empty(area, MIGRATE_CMA))
1901 return COMPACT_SUCCESS;
1904 * Job done if allocation would steal freepages from
1905 * other migratetype buddy lists.
1907 if (find_suitable_fallback(area, order, migratetype,
1908 true, &can_steal) != -1) {
1910 /* movable pages are OK in any pageblock */
1911 if (migratetype == MIGRATE_MOVABLE)
1912 return COMPACT_SUCCESS;
1915 * We are stealing for a non-movable allocation. Make
1916 * sure we finish compacting the current pageblock
1917 * first so it is as free as possible and we won't
1918 * have to steal another one soon. This only applies
1919 * to sync compaction, as async compaction operates
1920 * on pageblocks of the same migratetype.
1922 if (cc->mode == MIGRATE_ASYNC ||
1923 IS_ALIGNED(cc->migrate_pfn,
1924 pageblock_nr_pages)) {
1925 return COMPACT_SUCCESS;
1928 ret = COMPACT_CONTINUE;
1933 if (cc->contended || fatal_signal_pending(current))
1934 ret = COMPACT_CONTENDED;
1939 static enum compact_result compact_finished(struct compact_control *cc)
1943 ret = __compact_finished(cc);
1944 trace_mm_compaction_finished(cc->zone, cc->order, ret);
1945 if (ret == COMPACT_NO_SUITABLE_PAGE)
1946 ret = COMPACT_CONTINUE;
1952 * compaction_suitable: Is this suitable to run compaction on this zone now?
1954 * COMPACT_SKIPPED - If there are too few free pages for compaction
1955 * COMPACT_SUCCESS - If the allocation would succeed without compaction
1956 * COMPACT_CONTINUE - If compaction should run now
1958 static enum compact_result __compaction_suitable(struct zone *zone, int order,
1959 unsigned int alloc_flags,
1961 unsigned long wmark_target)
1963 unsigned long watermark;
1965 if (is_via_compact_memory(order))
1966 return COMPACT_CONTINUE;
1968 watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
1970 * If watermarks for high-order allocation are already met, there
1971 * should be no need for compaction at all.
1973 if (zone_watermark_ok(zone, order, watermark, classzone_idx,
1975 return COMPACT_SUCCESS;
1978 * Watermarks for order-0 must be met for compaction to be able to
1979 * isolate free pages for migration targets. This means that the
1980 * watermark and alloc_flags have to match, or be more pessimistic than
1981 * the check in __isolate_free_page(). We don't use the direct
1982 * compactor's alloc_flags, as they are not relevant for freepage
1983 * isolation. We however do use the direct compactor's classzone_idx to
1984 * skip over zones where lowmem reserves would prevent allocation even
1985 * if compaction succeeds.
1986 * For costly orders, we require low watermark instead of min for
1987 * compaction to proceed to increase its chances.
1988 * ALLOC_CMA is used, as pages in CMA pageblocks are considered
1989 * suitable migration targets
1991 watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
1992 low_wmark_pages(zone) : min_wmark_pages(zone);
1993 watermark += compact_gap(order);
1994 if (!__zone_watermark_ok(zone, 0, watermark, classzone_idx,
1995 ALLOC_CMA, wmark_target))
1996 return COMPACT_SKIPPED;
1998 return COMPACT_CONTINUE;
2001 enum compact_result compaction_suitable(struct zone *zone, int order,
2002 unsigned int alloc_flags,
2005 enum compact_result ret;
2008 ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx,
2009 zone_page_state(zone, NR_FREE_PAGES));
2011 * fragmentation index determines if allocation failures are due to
2012 * low memory or external fragmentation
2014 * index of -1000 would imply allocations might succeed depending on
2015 * watermarks, but we already failed the high-order watermark check
2016 * index towards 0 implies failure is due to lack of memory
2017 * index towards 1000 implies failure is due to fragmentation
2019 * Only compact if a failure would be due to fragmentation. Also
2020 * ignore fragindex for non-costly orders where the alternative to
2021 * a successful reclaim/compaction is OOM. Fragindex and the
2022 * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
2023 * excessive compaction for costly orders, but it should not be at the
2024 * expense of system stability.
2026 if (ret == COMPACT_CONTINUE && (order > PAGE_ALLOC_COSTLY_ORDER)) {
2027 fragindex = fragmentation_index(zone, order);
2028 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
2029 ret = COMPACT_NOT_SUITABLE_ZONE;
2032 trace_mm_compaction_suitable(zone, order, ret);
2033 if (ret == COMPACT_NOT_SUITABLE_ZONE)
2034 ret = COMPACT_SKIPPED;
2039 bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
2046 * Make sure at least one zone would pass __compaction_suitable if we continue
2047 * retrying the reclaim.
2049 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
2051 unsigned long available;
2052 enum compact_result compact_result;
2055 * Do not consider all the reclaimable memory because we do not
2056 * want to trash just for a single high order allocation which
2057 * is even not guaranteed to appear even if __compaction_suitable
2058 * is happy about the watermark check.
2060 available = zone_reclaimable_pages(zone) / order;
2061 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
2062 compact_result = __compaction_suitable(zone, order, alloc_flags,
2063 ac_classzone_idx(ac), available);
2064 if (compact_result != COMPACT_SKIPPED)
2071 static enum compact_result
2072 compact_zone(struct compact_control *cc, struct capture_control *capc)
2074 enum compact_result ret;
2075 unsigned long start_pfn = cc->zone->zone_start_pfn;
2076 unsigned long end_pfn = zone_end_pfn(cc->zone);
2077 unsigned long last_migrated_pfn;
2078 const bool sync = cc->mode != MIGRATE_ASYNC;
2082 * These counters track activities during zone compaction. Initialize
2083 * them before compacting a new zone.
2085 cc->total_migrate_scanned = 0;
2086 cc->total_free_scanned = 0;
2087 cc->nr_migratepages = 0;
2088 cc->nr_freepages = 0;
2089 INIT_LIST_HEAD(&cc->freepages);
2090 INIT_LIST_HEAD(&cc->migratepages);
2092 cc->migratetype = gfpflags_to_migratetype(cc->gfp_mask);
2093 ret = compaction_suitable(cc->zone, cc->order, cc->alloc_flags,
2095 /* Compaction is likely to fail */
2096 if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED)
2099 /* huh, compaction_suitable is returning something unexpected */
2100 VM_BUG_ON(ret != COMPACT_CONTINUE);
2103 * Clear pageblock skip if there were failures recently and compaction
2104 * is about to be retried after being deferred.
2106 if (compaction_restarting(cc->zone, cc->order))
2107 __reset_isolation_suitable(cc->zone);
2110 * Setup to move all movable pages to the end of the zone. Used cached
2111 * information on where the scanners should start (unless we explicitly
2112 * want to compact the whole zone), but check that it is initialised
2113 * by ensuring the values are within zone boundaries.
2115 cc->fast_start_pfn = 0;
2116 if (cc->whole_zone) {
2117 cc->migrate_pfn = start_pfn;
2118 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2120 cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync];
2121 cc->free_pfn = cc->zone->compact_cached_free_pfn;
2122 if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
2123 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2124 cc->zone->compact_cached_free_pfn = cc->free_pfn;
2126 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
2127 cc->migrate_pfn = start_pfn;
2128 cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
2129 cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
2132 if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn)
2133 cc->whole_zone = true;
2136 last_migrated_pfn = 0;
2139 * Migrate has separate cached PFNs for ASYNC and SYNC* migration on
2140 * the basis that some migrations will fail in ASYNC mode. However,
2141 * if the cached PFNs match and pageblocks are skipped due to having
2142 * no isolation candidates, then the sync state does not matter.
2143 * Until a pageblock with isolation candidates is found, keep the
2144 * cached PFNs in sync to avoid revisiting the same blocks.
2146 update_cached = !sync &&
2147 cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1];
2149 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
2150 cc->free_pfn, end_pfn, sync);
2152 migrate_prep_local();
2154 while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) {
2156 unsigned long start_pfn = cc->migrate_pfn;
2159 * Avoid multiple rescans which can happen if a page cannot be
2160 * isolated (dirty/writeback in async mode) or if the migrated
2161 * pages are being allocated before the pageblock is cleared.
2162 * The first rescan will capture the entire pageblock for
2163 * migration. If it fails, it'll be marked skip and scanning
2164 * will proceed as normal.
2167 if (pageblock_start_pfn(last_migrated_pfn) ==
2168 pageblock_start_pfn(start_pfn)) {
2172 switch (isolate_migratepages(cc->zone, cc)) {
2174 ret = COMPACT_CONTENDED;
2175 putback_movable_pages(&cc->migratepages);
2176 cc->nr_migratepages = 0;
2177 last_migrated_pfn = 0;
2180 if (update_cached) {
2181 cc->zone->compact_cached_migrate_pfn[1] =
2182 cc->zone->compact_cached_migrate_pfn[0];
2186 * We haven't isolated and migrated anything, but
2187 * there might still be unflushed migrations from
2188 * previous cc->order aligned block.
2191 case ISOLATE_SUCCESS:
2192 update_cached = false;
2193 last_migrated_pfn = start_pfn;
2197 err = migrate_pages(&cc->migratepages, compaction_alloc,
2198 compaction_free, (unsigned long)cc, cc->mode,
2201 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
2204 /* All pages were either migrated or will be released */
2205 cc->nr_migratepages = 0;
2207 putback_movable_pages(&cc->migratepages);
2209 * migrate_pages() may return -ENOMEM when scanners meet
2210 * and we want compact_finished() to detect it
2212 if (err == -ENOMEM && !compact_scanners_met(cc)) {
2213 ret = COMPACT_CONTENDED;
2217 * We failed to migrate at least one page in the current
2218 * order-aligned block, so skip the rest of it.
2220 if (cc->direct_compaction &&
2221 (cc->mode == MIGRATE_ASYNC)) {
2222 cc->migrate_pfn = block_end_pfn(
2223 cc->migrate_pfn - 1, cc->order);
2224 /* Draining pcplists is useless in this case */
2225 last_migrated_pfn = 0;
2231 * Has the migration scanner moved away from the previous
2232 * cc->order aligned block where we migrated from? If yes,
2233 * flush the pages that were freed, so that they can merge and
2234 * compact_finished() can detect immediately if allocation
2237 if (cc->order > 0 && last_migrated_pfn) {
2239 unsigned long current_block_start =
2240 block_start_pfn(cc->migrate_pfn, cc->order);
2242 if (last_migrated_pfn < current_block_start) {
2244 lru_add_drain_cpu(cpu);
2245 drain_local_pages(cc->zone);
2247 /* No more flushing until we migrate again */
2248 last_migrated_pfn = 0;
2252 /* Stop if a page has been captured */
2253 if (capc && capc->page) {
2254 ret = COMPACT_SUCCESS;
2261 * Release free pages and update where the free scanner should restart,
2262 * so we don't leave any returned pages behind in the next attempt.
2264 if (cc->nr_freepages > 0) {
2265 unsigned long free_pfn = release_freepages(&cc->freepages);
2267 cc->nr_freepages = 0;
2268 VM_BUG_ON(free_pfn == 0);
2269 /* The cached pfn is always the first in a pageblock */
2270 free_pfn = pageblock_start_pfn(free_pfn);
2272 * Only go back, not forward. The cached pfn might have been
2273 * already reset to zone end in compact_finished()
2275 if (free_pfn > cc->zone->compact_cached_free_pfn)
2276 cc->zone->compact_cached_free_pfn = free_pfn;
2279 count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned);
2280 count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned);
2282 trace_mm_compaction_end(start_pfn, cc->migrate_pfn,
2283 cc->free_pfn, end_pfn, sync, ret);
2288 static enum compact_result compact_zone_order(struct zone *zone, int order,
2289 gfp_t gfp_mask, enum compact_priority prio,
2290 unsigned int alloc_flags, int classzone_idx,
2291 struct page **capture)
2293 enum compact_result ret;
2294 struct compact_control cc = {
2296 .search_order = order,
2297 .gfp_mask = gfp_mask,
2299 .mode = (prio == COMPACT_PRIO_ASYNC) ?
2300 MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT,
2301 .alloc_flags = alloc_flags,
2302 .classzone_idx = classzone_idx,
2303 .direct_compaction = true,
2304 .whole_zone = (prio == MIN_COMPACT_PRIORITY),
2305 .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
2306 .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
2308 struct capture_control capc = {
2314 current->capture_control = &capc;
2316 ret = compact_zone(&cc, &capc);
2318 VM_BUG_ON(!list_empty(&cc.freepages));
2319 VM_BUG_ON(!list_empty(&cc.migratepages));
2321 *capture = capc.page;
2322 current->capture_control = NULL;
2327 int sysctl_extfrag_threshold = 500;
2330 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
2331 * @gfp_mask: The GFP mask of the current allocation
2332 * @order: The order of the current allocation
2333 * @alloc_flags: The allocation flags of the current allocation
2334 * @ac: The context of current allocation
2335 * @prio: Determines how hard direct compaction should try to succeed
2337 * This is the main entry point for direct page compaction.
2339 enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
2340 unsigned int alloc_flags, const struct alloc_context *ac,
2341 enum compact_priority prio, struct page **capture)
2343 int may_perform_io = gfp_mask & __GFP_IO;
2346 enum compact_result rc = COMPACT_SKIPPED;
2349 * Check if the GFP flags allow compaction - GFP_NOIO is really
2350 * tricky context because the migration might require IO
2352 if (!may_perform_io)
2353 return COMPACT_SKIPPED;
2355 trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
2357 /* Compact each zone in the list */
2358 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
2360 enum compact_result status;
2362 if (prio > MIN_COMPACT_PRIORITY
2363 && compaction_deferred(zone, order)) {
2364 rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
2368 status = compact_zone_order(zone, order, gfp_mask, prio,
2369 alloc_flags, ac_classzone_idx(ac), capture);
2370 rc = max(status, rc);
2372 /* The allocation should succeed, stop compacting */
2373 if (status == COMPACT_SUCCESS) {
2375 * We think the allocation will succeed in this zone,
2376 * but it is not certain, hence the false. The caller
2377 * will repeat this with true if allocation indeed
2378 * succeeds in this zone.
2380 compaction_defer_reset(zone, order, false);
2385 if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
2386 status == COMPACT_PARTIAL_SKIPPED))
2388 * We think that allocation won't succeed in this zone
2389 * so we defer compaction there. If it ends up
2390 * succeeding after all, it will be reset.
2392 defer_compaction(zone, order);
2395 * We might have stopped compacting due to need_resched() in
2396 * async compaction, or due to a fatal signal detected. In that
2397 * case do not try further zones
2399 if ((prio == COMPACT_PRIO_ASYNC && need_resched())
2400 || fatal_signal_pending(current))
2408 /* Compact all zones within a node */
2409 static void compact_node(int nid)
2411 pg_data_t *pgdat = NODE_DATA(nid);
2414 struct compact_control cc = {
2416 .mode = MIGRATE_SYNC,
2417 .ignore_skip_hint = true,
2419 .gfp_mask = GFP_KERNEL,
2423 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2425 zone = &pgdat->node_zones[zoneid];
2426 if (!populated_zone(zone))
2431 compact_zone(&cc, NULL);
2433 VM_BUG_ON(!list_empty(&cc.freepages));
2434 VM_BUG_ON(!list_empty(&cc.migratepages));
2438 /* Compact all nodes in the system */
2439 static void compact_nodes(void)
2443 /* Flush pending updates to the LRU lists */
2444 lru_add_drain_all();
2446 for_each_online_node(nid)
2450 /* The written value is actually unused, all memory is compacted */
2451 int sysctl_compact_memory;
2454 * This is the entry point for compacting all nodes via
2455 * /proc/sys/vm/compact_memory
2457 int sysctl_compaction_handler(struct ctl_table *table, int write,
2458 void __user *buffer, size_t *length, loff_t *ppos)
2466 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
2467 static ssize_t sysfs_compact_node(struct device *dev,
2468 struct device_attribute *attr,
2469 const char *buf, size_t count)
2473 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
2474 /* Flush pending updates to the LRU lists */
2475 lru_add_drain_all();
2482 static DEVICE_ATTR(compact, 0200, NULL, sysfs_compact_node);
2484 int compaction_register_node(struct node *node)
2486 return device_create_file(&node->dev, &dev_attr_compact);
2489 void compaction_unregister_node(struct node *node)
2491 return device_remove_file(&node->dev, &dev_attr_compact);
2493 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
2495 static inline bool kcompactd_work_requested(pg_data_t *pgdat)
2497 return pgdat->kcompactd_max_order > 0 || kthread_should_stop();
2500 static bool kcompactd_node_suitable(pg_data_t *pgdat)
2504 enum zone_type classzone_idx = pgdat->kcompactd_classzone_idx;
2506 for (zoneid = 0; zoneid <= classzone_idx; zoneid++) {
2507 zone = &pgdat->node_zones[zoneid];
2509 if (!populated_zone(zone))
2512 if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0,
2513 classzone_idx) == COMPACT_CONTINUE)
2520 static void kcompactd_do_work(pg_data_t *pgdat)
2523 * With no special task, compact all zones so that a page of requested
2524 * order is allocatable.
2528 struct compact_control cc = {
2529 .order = pgdat->kcompactd_max_order,
2530 .search_order = pgdat->kcompactd_max_order,
2531 .classzone_idx = pgdat->kcompactd_classzone_idx,
2532 .mode = MIGRATE_SYNC_LIGHT,
2533 .ignore_skip_hint = false,
2534 .gfp_mask = GFP_KERNEL,
2536 trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
2538 count_compact_event(KCOMPACTD_WAKE);
2540 for (zoneid = 0; zoneid <= cc.classzone_idx; zoneid++) {
2543 zone = &pgdat->node_zones[zoneid];
2544 if (!populated_zone(zone))
2547 if (compaction_deferred(zone, cc.order))
2550 if (compaction_suitable(zone, cc.order, 0, zoneid) !=
2554 if (kthread_should_stop())
2558 status = compact_zone(&cc, NULL);
2560 if (status == COMPACT_SUCCESS) {
2561 compaction_defer_reset(zone, cc.order, false);
2562 } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
2564 * Buddy pages may become stranded on pcps that could
2565 * otherwise coalesce on the zone's free area for
2566 * order >= cc.order. This is ratelimited by the
2567 * upcoming deferral.
2569 drain_all_pages(zone);
2572 * We use sync migration mode here, so we defer like
2573 * sync direct compaction does.
2575 defer_compaction(zone, cc.order);
2578 count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
2579 cc.total_migrate_scanned);
2580 count_compact_events(KCOMPACTD_FREE_SCANNED,
2581 cc.total_free_scanned);
2583 VM_BUG_ON(!list_empty(&cc.freepages));
2584 VM_BUG_ON(!list_empty(&cc.migratepages));
2588 * Regardless of success, we are done until woken up next. But remember
2589 * the requested order/classzone_idx in case it was higher/tighter than
2592 if (pgdat->kcompactd_max_order <= cc.order)
2593 pgdat->kcompactd_max_order = 0;
2594 if (pgdat->kcompactd_classzone_idx >= cc.classzone_idx)
2595 pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
2598 void wakeup_kcompactd(pg_data_t *pgdat, int order, int classzone_idx)
2603 if (pgdat->kcompactd_max_order < order)
2604 pgdat->kcompactd_max_order = order;
2606 if (pgdat->kcompactd_classzone_idx > classzone_idx)
2607 pgdat->kcompactd_classzone_idx = classzone_idx;
2610 * Pairs with implicit barrier in wait_event_freezable()
2611 * such that wakeups are not missed.
2613 if (!wq_has_sleeper(&pgdat->kcompactd_wait))
2616 if (!kcompactd_node_suitable(pgdat))
2619 trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
2621 wake_up_interruptible(&pgdat->kcompactd_wait);
2625 * The background compaction daemon, started as a kernel thread
2626 * from the init process.
2628 static int kcompactd(void *p)
2630 pg_data_t *pgdat = (pg_data_t*)p;
2631 struct task_struct *tsk = current;
2633 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2635 if (!cpumask_empty(cpumask))
2636 set_cpus_allowed_ptr(tsk, cpumask);
2640 pgdat->kcompactd_max_order = 0;
2641 pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
2643 while (!kthread_should_stop()) {
2644 unsigned long pflags;
2646 trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
2647 wait_event_freezable(pgdat->kcompactd_wait,
2648 kcompactd_work_requested(pgdat));
2650 psi_memstall_enter(&pflags);
2651 kcompactd_do_work(pgdat);
2652 psi_memstall_leave(&pflags);
2659 * This kcompactd start function will be called by init and node-hot-add.
2660 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
2662 int kcompactd_run(int nid)
2664 pg_data_t *pgdat = NODE_DATA(nid);
2667 if (pgdat->kcompactd)
2670 pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
2671 if (IS_ERR(pgdat->kcompactd)) {
2672 pr_err("Failed to start kcompactd on node %d\n", nid);
2673 ret = PTR_ERR(pgdat->kcompactd);
2674 pgdat->kcompactd = NULL;
2680 * Called by memory hotplug when all memory in a node is offlined. Caller must
2681 * hold mem_hotplug_begin/end().
2683 void kcompactd_stop(int nid)
2685 struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
2688 kthread_stop(kcompactd);
2689 NODE_DATA(nid)->kcompactd = NULL;
2694 * It's optimal to keep kcompactd on the same CPUs as their memory, but
2695 * not required for correctness. So if the last cpu in a node goes
2696 * away, we get changed to run anywhere: as the first one comes back,
2697 * restore their cpu bindings.
2699 static int kcompactd_cpu_online(unsigned int cpu)
2703 for_each_node_state(nid, N_MEMORY) {
2704 pg_data_t *pgdat = NODE_DATA(nid);
2705 const struct cpumask *mask;
2707 mask = cpumask_of_node(pgdat->node_id);
2709 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2710 /* One of our CPUs online: restore mask */
2711 set_cpus_allowed_ptr(pgdat->kcompactd, mask);
2716 static int __init kcompactd_init(void)
2721 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
2722 "mm/compaction:online",
2723 kcompactd_cpu_online, NULL);
2725 pr_err("kcompactd: failed to register hotplug callbacks.\n");
2729 for_each_node_state(nid, N_MEMORY)
2733 subsys_initcall(kcompactd_init)
2735 #endif /* CONFIG_COMPACTION */