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)
54 * Fragmentation score check interval for proactive compaction purposes.
56 static const unsigned int HPAGE_FRAG_CHECK_INTERVAL_MSEC = 500;
59 * Page order with-respect-to which proactive compaction
60 * calculates external fragmentation, which is used as
61 * the "fragmentation score" of a node/zone.
63 #if defined CONFIG_TRANSPARENT_HUGEPAGE
64 #define COMPACTION_HPAGE_ORDER HPAGE_PMD_ORDER
65 #elif defined CONFIG_HUGETLBFS
66 #define COMPACTION_HPAGE_ORDER HUGETLB_PAGE_ORDER
68 #define COMPACTION_HPAGE_ORDER (PMD_SHIFT - PAGE_SHIFT)
71 static unsigned long release_freepages(struct list_head *freelist)
73 struct page *page, *next;
74 unsigned long high_pfn = 0;
76 list_for_each_entry_safe(page, next, freelist, lru) {
77 unsigned long pfn = page_to_pfn(page);
87 static void split_map_pages(struct list_head *list)
89 unsigned int i, order, nr_pages;
90 struct page *page, *next;
93 list_for_each_entry_safe(page, next, list, lru) {
96 order = page_private(page);
97 nr_pages = 1 << order;
99 post_alloc_hook(page, order, __GFP_MOVABLE);
101 split_page(page, order);
103 for (i = 0; i < nr_pages; i++) {
104 list_add(&page->lru, &tmp_list);
109 list_splice(&tmp_list, list);
112 #ifdef CONFIG_COMPACTION
114 int PageMovable(struct page *page)
116 struct address_space *mapping;
118 VM_BUG_ON_PAGE(!PageLocked(page), page);
119 if (!__PageMovable(page))
122 mapping = page_mapping(page);
123 if (mapping && mapping->a_ops && mapping->a_ops->isolate_page)
128 EXPORT_SYMBOL(PageMovable);
130 void __SetPageMovable(struct page *page, struct address_space *mapping)
132 VM_BUG_ON_PAGE(!PageLocked(page), page);
133 VM_BUG_ON_PAGE((unsigned long)mapping & PAGE_MAPPING_MOVABLE, page);
134 page->mapping = (void *)((unsigned long)mapping | PAGE_MAPPING_MOVABLE);
136 EXPORT_SYMBOL(__SetPageMovable);
138 void __ClearPageMovable(struct page *page)
140 VM_BUG_ON_PAGE(!PageLocked(page), page);
141 VM_BUG_ON_PAGE(!PageMovable(page), page);
143 * Clear registered address_space val with keeping PAGE_MAPPING_MOVABLE
144 * flag so that VM can catch up released page by driver after isolation.
145 * With it, VM migration doesn't try to put it back.
147 page->mapping = (void *)((unsigned long)page->mapping &
148 PAGE_MAPPING_MOVABLE);
150 EXPORT_SYMBOL(__ClearPageMovable);
152 /* Do not skip compaction more than 64 times */
153 #define COMPACT_MAX_DEFER_SHIFT 6
156 * Compaction is deferred when compaction fails to result in a page
157 * allocation success. 1 << compact_defer_shift, compactions are skipped up
158 * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
160 void defer_compaction(struct zone *zone, int order)
162 zone->compact_considered = 0;
163 zone->compact_defer_shift++;
165 if (order < zone->compact_order_failed)
166 zone->compact_order_failed = order;
168 if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
169 zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
171 trace_mm_compaction_defer_compaction(zone, order);
174 /* Returns true if compaction should be skipped this time */
175 bool compaction_deferred(struct zone *zone, int order)
177 unsigned long defer_limit = 1UL << zone->compact_defer_shift;
179 if (order < zone->compact_order_failed)
182 /* Avoid possible overflow */
183 if (++zone->compact_considered >= defer_limit) {
184 zone->compact_considered = defer_limit;
188 trace_mm_compaction_deferred(zone, order);
194 * Update defer tracking counters after successful compaction of given order,
195 * which means an allocation either succeeded (alloc_success == true) or is
196 * expected to succeed.
198 void compaction_defer_reset(struct zone *zone, int order,
202 zone->compact_considered = 0;
203 zone->compact_defer_shift = 0;
205 if (order >= zone->compact_order_failed)
206 zone->compact_order_failed = order + 1;
208 trace_mm_compaction_defer_reset(zone, order);
211 /* Returns true if restarting compaction after many failures */
212 bool compaction_restarting(struct zone *zone, int order)
214 if (order < zone->compact_order_failed)
217 return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
218 zone->compact_considered >= 1UL << zone->compact_defer_shift;
221 /* Returns true if the pageblock should be scanned for pages to isolate. */
222 static inline bool isolation_suitable(struct compact_control *cc,
225 if (cc->ignore_skip_hint)
228 return !get_pageblock_skip(page);
231 static void reset_cached_positions(struct zone *zone)
233 zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
234 zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
235 zone->compact_cached_free_pfn =
236 pageblock_start_pfn(zone_end_pfn(zone) - 1);
240 * Compound pages of >= pageblock_order should consistenly be skipped until
241 * released. It is always pointless to compact pages of such order (if they are
242 * migratable), and the pageblocks they occupy cannot contain any free pages.
244 static bool pageblock_skip_persistent(struct page *page)
246 if (!PageCompound(page))
249 page = compound_head(page);
251 if (compound_order(page) >= pageblock_order)
258 __reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source,
261 struct page *page = pfn_to_online_page(pfn);
262 struct page *block_page;
263 struct page *end_page;
264 unsigned long block_pfn;
268 if (zone != page_zone(page))
270 if (pageblock_skip_persistent(page))
274 * If skip is already cleared do no further checking once the
275 * restart points have been set.
277 if (check_source && check_target && !get_pageblock_skip(page))
281 * If clearing skip for the target scanner, do not select a
282 * non-movable pageblock as the starting point.
284 if (!check_source && check_target &&
285 get_pageblock_migratetype(page) != MIGRATE_MOVABLE)
288 /* Ensure the start of the pageblock or zone is online and valid */
289 block_pfn = pageblock_start_pfn(pfn);
290 block_pfn = max(block_pfn, zone->zone_start_pfn);
291 block_page = pfn_to_online_page(block_pfn);
297 /* Ensure the end of the pageblock or zone is online and valid */
298 block_pfn = pageblock_end_pfn(pfn) - 1;
299 block_pfn = min(block_pfn, zone_end_pfn(zone) - 1);
300 end_page = pfn_to_online_page(block_pfn);
305 * Only clear the hint if a sample indicates there is either a
306 * free page or an LRU page in the block. One or other condition
307 * is necessary for the block to be a migration source/target.
310 if (pfn_valid_within(pfn)) {
311 if (check_source && PageLRU(page)) {
312 clear_pageblock_skip(page);
316 if (check_target && PageBuddy(page)) {
317 clear_pageblock_skip(page);
322 page += (1 << PAGE_ALLOC_COSTLY_ORDER);
323 pfn += (1 << PAGE_ALLOC_COSTLY_ORDER);
324 } while (page <= end_page);
330 * This function is called to clear all cached information on pageblocks that
331 * should be skipped for page isolation when the migrate and free page scanner
334 static void __reset_isolation_suitable(struct zone *zone)
336 unsigned long migrate_pfn = zone->zone_start_pfn;
337 unsigned long free_pfn = zone_end_pfn(zone) - 1;
338 unsigned long reset_migrate = free_pfn;
339 unsigned long reset_free = migrate_pfn;
340 bool source_set = false;
341 bool free_set = false;
343 if (!zone->compact_blockskip_flush)
346 zone->compact_blockskip_flush = false;
349 * Walk the zone and update pageblock skip information. Source looks
350 * for PageLRU while target looks for PageBuddy. When the scanner
351 * is found, both PageBuddy and PageLRU are checked as the pageblock
352 * is suitable as both source and target.
354 for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages,
355 free_pfn -= pageblock_nr_pages) {
358 /* Update the migrate PFN */
359 if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) &&
360 migrate_pfn < reset_migrate) {
362 reset_migrate = migrate_pfn;
363 zone->compact_init_migrate_pfn = reset_migrate;
364 zone->compact_cached_migrate_pfn[0] = reset_migrate;
365 zone->compact_cached_migrate_pfn[1] = reset_migrate;
368 /* Update the free PFN */
369 if (__reset_isolation_pfn(zone, free_pfn, free_set, true) &&
370 free_pfn > reset_free) {
372 reset_free = free_pfn;
373 zone->compact_init_free_pfn = reset_free;
374 zone->compact_cached_free_pfn = reset_free;
378 /* Leave no distance if no suitable block was reset */
379 if (reset_migrate >= reset_free) {
380 zone->compact_cached_migrate_pfn[0] = migrate_pfn;
381 zone->compact_cached_migrate_pfn[1] = migrate_pfn;
382 zone->compact_cached_free_pfn = free_pfn;
386 void reset_isolation_suitable(pg_data_t *pgdat)
390 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
391 struct zone *zone = &pgdat->node_zones[zoneid];
392 if (!populated_zone(zone))
395 /* Only flush if a full compaction finished recently */
396 if (zone->compact_blockskip_flush)
397 __reset_isolation_suitable(zone);
402 * Sets the pageblock skip bit if it was clear. Note that this is a hint as
403 * locks are not required for read/writers. Returns true if it was already set.
405 static bool test_and_set_skip(struct compact_control *cc, struct page *page,
410 /* Do no update if skip hint is being ignored */
411 if (cc->ignore_skip_hint)
414 if (!IS_ALIGNED(pfn, pageblock_nr_pages))
417 skip = get_pageblock_skip(page);
418 if (!skip && !cc->no_set_skip_hint)
419 set_pageblock_skip(page);
424 static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
426 struct zone *zone = cc->zone;
428 pfn = pageblock_end_pfn(pfn);
430 /* Set for isolation rather than compaction */
431 if (cc->no_set_skip_hint)
434 if (pfn > zone->compact_cached_migrate_pfn[0])
435 zone->compact_cached_migrate_pfn[0] = pfn;
436 if (cc->mode != MIGRATE_ASYNC &&
437 pfn > zone->compact_cached_migrate_pfn[1])
438 zone->compact_cached_migrate_pfn[1] = pfn;
442 * If no pages were isolated then mark this pageblock to be skipped in the
443 * future. The information is later cleared by __reset_isolation_suitable().
445 static void update_pageblock_skip(struct compact_control *cc,
446 struct page *page, unsigned long pfn)
448 struct zone *zone = cc->zone;
450 if (cc->no_set_skip_hint)
456 set_pageblock_skip(page);
458 /* Update where async and sync compaction should restart */
459 if (pfn < zone->compact_cached_free_pfn)
460 zone->compact_cached_free_pfn = pfn;
463 static inline bool isolation_suitable(struct compact_control *cc,
469 static inline bool pageblock_skip_persistent(struct page *page)
474 static inline void update_pageblock_skip(struct compact_control *cc,
475 struct page *page, unsigned long pfn)
479 static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
483 static bool test_and_set_skip(struct compact_control *cc, struct page *page,
488 #endif /* CONFIG_COMPACTION */
491 * Compaction requires the taking of some coarse locks that are potentially
492 * very heavily contended. For async compaction, trylock and record if the
493 * lock is contended. The lock will still be acquired but compaction will
494 * abort when the current block is finished regardless of success rate.
495 * Sync compaction acquires the lock.
497 * Always returns true which makes it easier to track lock state in callers.
499 static bool compact_lock_irqsave(spinlock_t *lock, unsigned long *flags,
500 struct compact_control *cc)
503 /* Track if the lock is contended in async mode */
504 if (cc->mode == MIGRATE_ASYNC && !cc->contended) {
505 if (spin_trylock_irqsave(lock, *flags))
508 cc->contended = true;
511 spin_lock_irqsave(lock, *flags);
516 * Compaction requires the taking of some coarse locks that are potentially
517 * very heavily contended. The lock should be periodically unlocked to avoid
518 * having disabled IRQs for a long time, even when there is nobody waiting on
519 * the lock. It might also be that allowing the IRQs will result in
520 * need_resched() becoming true. If scheduling is needed, async compaction
521 * aborts. Sync compaction schedules.
522 * Either compaction type will also abort if a fatal signal is pending.
523 * In either case if the lock was locked, it is dropped and not regained.
525 * Returns true if compaction should abort due to fatal signal pending, or
526 * async compaction due to need_resched()
527 * Returns false when compaction can continue (sync compaction might have
530 static bool compact_unlock_should_abort(spinlock_t *lock,
531 unsigned long flags, bool *locked, struct compact_control *cc)
534 spin_unlock_irqrestore(lock, flags);
538 if (fatal_signal_pending(current)) {
539 cc->contended = true;
549 * Isolate free pages onto a private freelist. If @strict is true, will abort
550 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
551 * (even though it may still end up isolating some pages).
553 static unsigned long isolate_freepages_block(struct compact_control *cc,
554 unsigned long *start_pfn,
555 unsigned long end_pfn,
556 struct list_head *freelist,
560 int nr_scanned = 0, total_isolated = 0;
562 unsigned long flags = 0;
564 unsigned long blockpfn = *start_pfn;
567 /* Strict mode is for isolation, speed is secondary */
571 cursor = pfn_to_page(blockpfn);
573 /* Isolate free pages. */
574 for (; blockpfn < end_pfn; blockpfn += stride, cursor += stride) {
576 struct page *page = cursor;
579 * Periodically drop the lock (if held) regardless of its
580 * contention, to give chance to IRQs. Abort if fatal signal
581 * pending or async compaction detects need_resched()
583 if (!(blockpfn % SWAP_CLUSTER_MAX)
584 && compact_unlock_should_abort(&cc->zone->lock, flags,
589 if (!pfn_valid_within(blockpfn))
593 * For compound pages such as THP and hugetlbfs, we can save
594 * potentially a lot of iterations if we skip them at once.
595 * The check is racy, but we can consider only valid values
596 * and the only danger is skipping too much.
598 if (PageCompound(page)) {
599 const unsigned int order = compound_order(page);
601 if (likely(order < MAX_ORDER)) {
602 blockpfn += (1UL << order) - 1;
603 cursor += (1UL << order) - 1;
608 if (!PageBuddy(page))
612 * If we already hold the lock, we can skip some rechecking.
613 * Note that if we hold the lock now, checked_pageblock was
614 * already set in some previous iteration (or strict is true),
615 * so it is correct to skip the suitable migration target
619 locked = compact_lock_irqsave(&cc->zone->lock,
622 /* Recheck this is a buddy page under lock */
623 if (!PageBuddy(page))
627 /* Found a free page, will break it into order-0 pages */
628 order = buddy_order(page);
629 isolated = __isolate_free_page(page, order);
632 set_page_private(page, order);
634 total_isolated += isolated;
635 cc->nr_freepages += isolated;
636 list_add_tail(&page->lru, freelist);
638 if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
639 blockpfn += isolated;
642 /* Advance to the end of split page */
643 blockpfn += isolated - 1;
644 cursor += isolated - 1;
656 spin_unlock_irqrestore(&cc->zone->lock, flags);
659 * There is a tiny chance that we have read bogus compound_order(),
660 * so be careful to not go outside of the pageblock.
662 if (unlikely(blockpfn > end_pfn))
665 trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
666 nr_scanned, total_isolated);
668 /* Record how far we have got within the block */
669 *start_pfn = blockpfn;
672 * If strict isolation is requested by CMA then check that all the
673 * pages requested were isolated. If there were any failures, 0 is
674 * returned and CMA will fail.
676 if (strict && blockpfn < end_pfn)
679 cc->total_free_scanned += nr_scanned;
681 count_compact_events(COMPACTISOLATED, total_isolated);
682 return total_isolated;
686 * isolate_freepages_range() - isolate free pages.
687 * @cc: Compaction control structure.
688 * @start_pfn: The first PFN to start isolating.
689 * @end_pfn: The one-past-last PFN.
691 * Non-free pages, invalid PFNs, or zone boundaries within the
692 * [start_pfn, end_pfn) range are considered errors, cause function to
693 * undo its actions and return zero.
695 * Otherwise, function returns one-past-the-last PFN of isolated page
696 * (which may be greater then end_pfn if end fell in a middle of
700 isolate_freepages_range(struct compact_control *cc,
701 unsigned long start_pfn, unsigned long end_pfn)
703 unsigned long isolated, pfn, block_start_pfn, block_end_pfn;
707 block_start_pfn = pageblock_start_pfn(pfn);
708 if (block_start_pfn < cc->zone->zone_start_pfn)
709 block_start_pfn = cc->zone->zone_start_pfn;
710 block_end_pfn = pageblock_end_pfn(pfn);
712 for (; pfn < end_pfn; pfn += isolated,
713 block_start_pfn = block_end_pfn,
714 block_end_pfn += pageblock_nr_pages) {
715 /* Protect pfn from changing by isolate_freepages_block */
716 unsigned long isolate_start_pfn = pfn;
718 block_end_pfn = min(block_end_pfn, end_pfn);
721 * pfn could pass the block_end_pfn if isolated freepage
722 * is more than pageblock order. In this case, we adjust
723 * scanning range to right one.
725 if (pfn >= block_end_pfn) {
726 block_start_pfn = pageblock_start_pfn(pfn);
727 block_end_pfn = pageblock_end_pfn(pfn);
728 block_end_pfn = min(block_end_pfn, end_pfn);
731 if (!pageblock_pfn_to_page(block_start_pfn,
732 block_end_pfn, cc->zone))
735 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
736 block_end_pfn, &freelist, 0, true);
739 * In strict mode, isolate_freepages_block() returns 0 if
740 * there are any holes in the block (ie. invalid PFNs or
747 * If we managed to isolate pages, it is always (1 << n) *
748 * pageblock_nr_pages for some non-negative n. (Max order
749 * page may span two pageblocks).
753 /* __isolate_free_page() does not map the pages */
754 split_map_pages(&freelist);
757 /* Loop terminated early, cleanup. */
758 release_freepages(&freelist);
762 /* We don't use freelists for anything. */
766 /* Similar to reclaim, but different enough that they don't share logic */
767 static bool too_many_isolated(pg_data_t *pgdat)
769 unsigned long active, inactive, isolated;
771 inactive = node_page_state(pgdat, NR_INACTIVE_FILE) +
772 node_page_state(pgdat, NR_INACTIVE_ANON);
773 active = node_page_state(pgdat, NR_ACTIVE_FILE) +
774 node_page_state(pgdat, NR_ACTIVE_ANON);
775 isolated = node_page_state(pgdat, NR_ISOLATED_FILE) +
776 node_page_state(pgdat, NR_ISOLATED_ANON);
778 return isolated > (inactive + active) / 2;
782 * isolate_migratepages_block() - isolate all migrate-able pages within
784 * @cc: Compaction control structure.
785 * @low_pfn: The first PFN to isolate
786 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
787 * @isolate_mode: Isolation mode to be used.
789 * Isolate all pages that can be migrated from the range specified by
790 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
791 * Returns zero if there is a fatal signal pending, otherwise PFN of the
792 * first page that was not scanned (which may be both less, equal to or more
795 * The pages are isolated on cc->migratepages list (not required to be empty),
796 * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
797 * is neither read nor updated.
800 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
801 unsigned long end_pfn, isolate_mode_t isolate_mode)
803 pg_data_t *pgdat = cc->zone->zone_pgdat;
804 unsigned long nr_scanned = 0, nr_isolated = 0;
805 struct lruvec *lruvec;
806 unsigned long flags = 0;
808 struct page *page = NULL, *valid_page = NULL;
809 unsigned long start_pfn = low_pfn;
810 bool skip_on_failure = false;
811 unsigned long next_skip_pfn = 0;
812 bool skip_updated = false;
815 * Ensure that there are not too many pages isolated from the LRU
816 * list by either parallel reclaimers or compaction. If there are,
817 * delay for some time until fewer pages are isolated
819 while (unlikely(too_many_isolated(pgdat))) {
820 /* stop isolation if there are still pages not migrated */
821 if (cc->nr_migratepages)
824 /* async migration should just abort */
825 if (cc->mode == MIGRATE_ASYNC)
828 congestion_wait(BLK_RW_ASYNC, HZ/10);
830 if (fatal_signal_pending(current))
836 if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
837 skip_on_failure = true;
838 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
841 /* Time to isolate some pages for migration */
842 for (; low_pfn < end_pfn; low_pfn++) {
844 if (skip_on_failure && low_pfn >= next_skip_pfn) {
846 * We have isolated all migration candidates in the
847 * previous order-aligned block, and did not skip it due
848 * to failure. We should migrate the pages now and
849 * hopefully succeed compaction.
855 * We failed to isolate in the previous order-aligned
856 * block. Set the new boundary to the end of the
857 * current block. Note we can't simply increase
858 * next_skip_pfn by 1 << order, as low_pfn might have
859 * been incremented by a higher number due to skipping
860 * a compound or a high-order buddy page in the
861 * previous loop iteration.
863 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
867 * Periodically drop the lock (if held) regardless of its
868 * contention, to give chance to IRQs. Abort completely if
869 * a fatal signal is pending.
871 if (!(low_pfn % SWAP_CLUSTER_MAX)
872 && compact_unlock_should_abort(&pgdat->lru_lock,
873 flags, &locked, cc)) {
878 if (!pfn_valid_within(low_pfn))
882 page = pfn_to_page(low_pfn);
885 * Check if the pageblock has already been marked skipped.
886 * Only the aligned PFN is checked as the caller isolates
887 * COMPACT_CLUSTER_MAX at a time so the second call must
888 * not falsely conclude that the block should be skipped.
890 if (!valid_page && IS_ALIGNED(low_pfn, pageblock_nr_pages)) {
891 if (!cc->ignore_skip_hint && get_pageblock_skip(page)) {
899 * Skip if free. We read page order here without zone lock
900 * which is generally unsafe, but the race window is small and
901 * the worst thing that can happen is that we skip some
902 * potential isolation targets.
904 if (PageBuddy(page)) {
905 unsigned long freepage_order = buddy_order_unsafe(page);
908 * Without lock, we cannot be sure that what we got is
909 * a valid page order. Consider only values in the
910 * valid order range to prevent low_pfn overflow.
912 if (freepage_order > 0 && freepage_order < MAX_ORDER)
913 low_pfn += (1UL << freepage_order) - 1;
918 * Regardless of being on LRU, compound pages such as THP and
919 * hugetlbfs are not to be compacted unless we are attempting
920 * an allocation much larger than the huge page size (eg CMA).
921 * We can potentially save a lot of iterations if we skip them
922 * at once. The check is racy, but we can consider only valid
923 * values and the only danger is skipping too much.
925 if (PageCompound(page) && !cc->alloc_contig) {
926 const unsigned int order = compound_order(page);
928 if (likely(order < MAX_ORDER))
929 low_pfn += (1UL << order) - 1;
934 * Check may be lockless but that's ok as we recheck later.
935 * It's possible to migrate LRU and non-lru movable pages.
936 * Skip any other type of page
938 if (!PageLRU(page)) {
940 * __PageMovable can return false positive so we need
941 * to verify it under page_lock.
943 if (unlikely(__PageMovable(page)) &&
944 !PageIsolated(page)) {
946 spin_unlock_irqrestore(&pgdat->lru_lock,
951 if (!isolate_movable_page(page, isolate_mode))
952 goto isolate_success;
959 * Migration will fail if an anonymous page is pinned in memory,
960 * so avoid taking lru_lock and isolating it unnecessarily in an
961 * admittedly racy check.
963 if (!page_mapping(page) &&
964 page_count(page) > page_mapcount(page))
968 * Only allow to migrate anonymous pages in GFP_NOFS context
969 * because those do not depend on fs locks.
971 if (!(cc->gfp_mask & __GFP_FS) && page_mapping(page))
974 /* If we already hold the lock, we can skip some rechecking */
976 locked = compact_lock_irqsave(&pgdat->lru_lock,
979 /* Try get exclusive access under lock */
982 if (test_and_set_skip(cc, page, low_pfn))
986 /* Recheck PageLRU and PageCompound under lock */
991 * Page become compound since the non-locked check,
992 * and it's on LRU. It can only be a THP so the order
993 * is safe to read and it's 0 for tail pages.
995 if (unlikely(PageCompound(page) && !cc->alloc_contig)) {
996 low_pfn += compound_nr(page) - 1;
1001 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1003 /* Try isolate the page */
1004 if (__isolate_lru_page(page, isolate_mode) != 0)
1007 /* The whole page is taken off the LRU; skip the tail pages. */
1008 if (PageCompound(page))
1009 low_pfn += compound_nr(page) - 1;
1011 /* Successfully isolated */
1012 del_page_from_lru_list(page, lruvec, page_lru(page));
1013 mod_node_page_state(page_pgdat(page),
1014 NR_ISOLATED_ANON + page_is_file_lru(page),
1015 thp_nr_pages(page));
1018 list_add(&page->lru, &cc->migratepages);
1019 cc->nr_migratepages += compound_nr(page);
1020 nr_isolated += compound_nr(page);
1023 * Avoid isolating too much unless this block is being
1024 * rescanned (e.g. dirty/writeback pages, parallel allocation)
1025 * or a lock is contended. For contention, isolate quickly to
1026 * potentially remove one source of contention.
1028 if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX &&
1029 !cc->rescan && !cc->contended) {
1036 if (!skip_on_failure)
1040 * We have isolated some pages, but then failed. Release them
1041 * instead of migrating, as we cannot form the cc->order buddy
1046 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
1049 putback_movable_pages(&cc->migratepages);
1050 cc->nr_migratepages = 0;
1054 if (low_pfn < next_skip_pfn) {
1055 low_pfn = next_skip_pfn - 1;
1057 * The check near the loop beginning would have updated
1058 * next_skip_pfn too, but this is a bit simpler.
1060 next_skip_pfn += 1UL << cc->order;
1065 * The PageBuddy() check could have potentially brought us outside
1066 * the range to be scanned.
1068 if (unlikely(low_pfn > end_pfn))
1073 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
1076 * Updated the cached scanner pfn once the pageblock has been scanned
1077 * Pages will either be migrated in which case there is no point
1078 * scanning in the near future or migration failed in which case the
1079 * failure reason may persist. The block is marked for skipping if
1080 * there were no pages isolated in the block or if the block is
1081 * rescanned twice in a row.
1083 if (low_pfn == end_pfn && (!nr_isolated || cc->rescan)) {
1084 if (valid_page && !skip_updated)
1085 set_pageblock_skip(valid_page);
1086 update_cached_migrate(cc, low_pfn);
1089 trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
1090 nr_scanned, nr_isolated);
1093 cc->total_migrate_scanned += nr_scanned;
1095 count_compact_events(COMPACTISOLATED, nr_isolated);
1101 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
1102 * @cc: Compaction control structure.
1103 * @start_pfn: The first PFN to start isolating.
1104 * @end_pfn: The one-past-last PFN.
1106 * Returns zero if isolation fails fatally due to e.g. pending signal.
1107 * Otherwise, function returns one-past-the-last PFN of isolated page
1108 * (which may be greater than end_pfn if end fell in a middle of a THP page).
1111 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
1112 unsigned long end_pfn)
1114 unsigned long pfn, block_start_pfn, block_end_pfn;
1116 /* Scan block by block. First and last block may be incomplete */
1118 block_start_pfn = pageblock_start_pfn(pfn);
1119 if (block_start_pfn < cc->zone->zone_start_pfn)
1120 block_start_pfn = cc->zone->zone_start_pfn;
1121 block_end_pfn = pageblock_end_pfn(pfn);
1123 for (; pfn < end_pfn; pfn = block_end_pfn,
1124 block_start_pfn = block_end_pfn,
1125 block_end_pfn += pageblock_nr_pages) {
1127 block_end_pfn = min(block_end_pfn, end_pfn);
1129 if (!pageblock_pfn_to_page(block_start_pfn,
1130 block_end_pfn, cc->zone))
1133 pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
1134 ISOLATE_UNEVICTABLE);
1139 if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX)
1146 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
1147 #ifdef CONFIG_COMPACTION
1149 static bool suitable_migration_source(struct compact_control *cc,
1154 if (pageblock_skip_persistent(page))
1157 if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction)
1160 block_mt = get_pageblock_migratetype(page);
1162 if (cc->migratetype == MIGRATE_MOVABLE)
1163 return is_migrate_movable(block_mt);
1165 return block_mt == cc->migratetype;
1168 /* Returns true if the page is within a block suitable for migration to */
1169 static bool suitable_migration_target(struct compact_control *cc,
1172 /* If the page is a large free page, then disallow migration */
1173 if (PageBuddy(page)) {
1175 * We are checking page_order without zone->lock taken. But
1176 * the only small danger is that we skip a potentially suitable
1177 * pageblock, so it's not worth to check order for valid range.
1179 if (buddy_order_unsafe(page) >= pageblock_order)
1183 if (cc->ignore_block_suitable)
1186 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1187 if (is_migrate_movable(get_pageblock_migratetype(page)))
1190 /* Otherwise skip the block */
1194 static inline unsigned int
1195 freelist_scan_limit(struct compact_control *cc)
1197 unsigned short shift = BITS_PER_LONG - 1;
1199 return (COMPACT_CLUSTER_MAX >> min(shift, cc->fast_search_fail)) + 1;
1203 * Test whether the free scanner has reached the same or lower pageblock than
1204 * the migration scanner, and compaction should thus terminate.
1206 static inline bool compact_scanners_met(struct compact_control *cc)
1208 return (cc->free_pfn >> pageblock_order)
1209 <= (cc->migrate_pfn >> pageblock_order);
1213 * Used when scanning for a suitable migration target which scans freelists
1214 * in reverse. Reorders the list such as the unscanned pages are scanned
1215 * first on the next iteration of the free scanner
1218 move_freelist_head(struct list_head *freelist, struct page *freepage)
1222 if (!list_is_last(freelist, &freepage->lru)) {
1223 list_cut_before(&sublist, freelist, &freepage->lru);
1224 if (!list_empty(&sublist))
1225 list_splice_tail(&sublist, freelist);
1230 * Similar to move_freelist_head except used by the migration scanner
1231 * when scanning forward. It's possible for these list operations to
1232 * move against each other if they search the free list exactly in
1236 move_freelist_tail(struct list_head *freelist, struct page *freepage)
1240 if (!list_is_first(freelist, &freepage->lru)) {
1241 list_cut_position(&sublist, freelist, &freepage->lru);
1242 if (!list_empty(&sublist))
1243 list_splice_tail(&sublist, freelist);
1248 fast_isolate_around(struct compact_control *cc, unsigned long pfn)
1250 unsigned long start_pfn, end_pfn;
1253 /* Do not search around if there are enough pages already */
1254 if (cc->nr_freepages >= cc->nr_migratepages)
1257 /* Minimise scanning during async compaction */
1258 if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC)
1261 /* Pageblock boundaries */
1262 start_pfn = max(pageblock_start_pfn(pfn), cc->zone->zone_start_pfn);
1263 end_pfn = min(pageblock_end_pfn(pfn), zone_end_pfn(cc->zone));
1265 page = pageblock_pfn_to_page(start_pfn, end_pfn, cc->zone);
1269 isolate_freepages_block(cc, &start_pfn, end_pfn, &cc->freepages, 1, false);
1271 /* Skip this pageblock in the future as it's full or nearly full */
1272 if (cc->nr_freepages < cc->nr_migratepages)
1273 set_pageblock_skip(page);
1278 /* Search orders in round-robin fashion */
1279 static int next_search_order(struct compact_control *cc, int order)
1283 order = cc->order - 1;
1285 /* Search wrapped around? */
1286 if (order == cc->search_order) {
1288 if (cc->search_order < 0)
1289 cc->search_order = cc->order - 1;
1296 static unsigned long
1297 fast_isolate_freepages(struct compact_control *cc)
1299 unsigned int limit = min(1U, freelist_scan_limit(cc) >> 1);
1300 unsigned int nr_scanned = 0;
1301 unsigned long low_pfn, min_pfn, highest = 0;
1302 unsigned long nr_isolated = 0;
1303 unsigned long distance;
1304 struct page *page = NULL;
1305 bool scan_start = false;
1308 /* Full compaction passes in a negative order */
1310 return cc->free_pfn;
1313 * If starting the scan, use a deeper search and use the highest
1314 * PFN found if a suitable one is not found.
1316 if (cc->free_pfn >= cc->zone->compact_init_free_pfn) {
1317 limit = pageblock_nr_pages >> 1;
1322 * Preferred point is in the top quarter of the scan space but take
1323 * a pfn from the top half if the search is problematic.
1325 distance = (cc->free_pfn - cc->migrate_pfn);
1326 low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2));
1327 min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1));
1329 if (WARN_ON_ONCE(min_pfn > low_pfn))
1333 * Search starts from the last successful isolation order or the next
1334 * order to search after a previous failure
1336 cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order);
1338 for (order = cc->search_order;
1339 !page && order >= 0;
1340 order = next_search_order(cc, order)) {
1341 struct free_area *area = &cc->zone->free_area[order];
1342 struct list_head *freelist;
1343 struct page *freepage;
1344 unsigned long flags;
1345 unsigned int order_scanned = 0;
1346 unsigned long high_pfn = 0;
1351 spin_lock_irqsave(&cc->zone->lock, flags);
1352 freelist = &area->free_list[MIGRATE_MOVABLE];
1353 list_for_each_entry_reverse(freepage, freelist, lru) {
1358 pfn = page_to_pfn(freepage);
1361 highest = max(pageblock_start_pfn(pfn),
1362 cc->zone->zone_start_pfn);
1364 if (pfn >= low_pfn) {
1365 cc->fast_search_fail = 0;
1366 cc->search_order = order;
1371 if (pfn >= min_pfn && pfn > high_pfn) {
1374 /* Shorten the scan if a candidate is found */
1378 if (order_scanned >= limit)
1382 /* Use a minimum pfn if a preferred one was not found */
1383 if (!page && high_pfn) {
1384 page = pfn_to_page(high_pfn);
1386 /* Update freepage for the list reorder below */
1390 /* Reorder to so a future search skips recent pages */
1391 move_freelist_head(freelist, freepage);
1393 /* Isolate the page if available */
1395 if (__isolate_free_page(page, order)) {
1396 set_page_private(page, order);
1397 nr_isolated = 1 << order;
1398 cc->nr_freepages += nr_isolated;
1399 list_add_tail(&page->lru, &cc->freepages);
1400 count_compact_events(COMPACTISOLATED, nr_isolated);
1402 /* If isolation fails, abort the search */
1403 order = cc->search_order + 1;
1408 spin_unlock_irqrestore(&cc->zone->lock, flags);
1411 * Smaller scan on next order so the total scan ig related
1412 * to freelist_scan_limit.
1414 if (order_scanned >= limit)
1415 limit = min(1U, limit >> 1);
1419 cc->fast_search_fail++;
1422 * Use the highest PFN found above min. If one was
1423 * not found, be pessimistic for direct compaction
1424 * and use the min mark.
1427 page = pfn_to_page(highest);
1428 cc->free_pfn = highest;
1430 if (cc->direct_compaction && pfn_valid(min_pfn)) {
1431 page = pageblock_pfn_to_page(min_pfn,
1432 min(pageblock_end_pfn(min_pfn),
1433 zone_end_pfn(cc->zone)),
1435 cc->free_pfn = min_pfn;
1441 if (highest && highest >= cc->zone->compact_cached_free_pfn) {
1442 highest -= pageblock_nr_pages;
1443 cc->zone->compact_cached_free_pfn = highest;
1446 cc->total_free_scanned += nr_scanned;
1448 return cc->free_pfn;
1450 low_pfn = page_to_pfn(page);
1451 fast_isolate_around(cc, low_pfn);
1456 * Based on information in the current compact_control, find blocks
1457 * suitable for isolating free pages from and then isolate them.
1459 static void isolate_freepages(struct compact_control *cc)
1461 struct zone *zone = cc->zone;
1463 unsigned long block_start_pfn; /* start of current pageblock */
1464 unsigned long isolate_start_pfn; /* exact pfn we start at */
1465 unsigned long block_end_pfn; /* end of current pageblock */
1466 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
1467 struct list_head *freelist = &cc->freepages;
1468 unsigned int stride;
1470 /* Try a small search of the free lists for a candidate */
1471 isolate_start_pfn = fast_isolate_freepages(cc);
1472 if (cc->nr_freepages)
1476 * Initialise the free scanner. The starting point is where we last
1477 * successfully isolated from, zone-cached value, or the end of the
1478 * zone when isolating for the first time. For looping we also need
1479 * this pfn aligned down to the pageblock boundary, because we do
1480 * block_start_pfn -= pageblock_nr_pages in the for loop.
1481 * For ending point, take care when isolating in last pageblock of a
1482 * zone which ends in the middle of a pageblock.
1483 * The low boundary is the end of the pageblock the migration scanner
1486 isolate_start_pfn = cc->free_pfn;
1487 block_start_pfn = pageblock_start_pfn(isolate_start_pfn);
1488 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1489 zone_end_pfn(zone));
1490 low_pfn = pageblock_end_pfn(cc->migrate_pfn);
1491 stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1;
1494 * Isolate free pages until enough are available to migrate the
1495 * pages on cc->migratepages. We stop searching if the migrate
1496 * and free page scanners meet or enough free pages are isolated.
1498 for (; block_start_pfn >= low_pfn;
1499 block_end_pfn = block_start_pfn,
1500 block_start_pfn -= pageblock_nr_pages,
1501 isolate_start_pfn = block_start_pfn) {
1502 unsigned long nr_isolated;
1505 * This can iterate a massively long zone without finding any
1506 * suitable migration targets, so periodically check resched.
1508 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)))
1511 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1516 /* Check the block is suitable for migration */
1517 if (!suitable_migration_target(cc, page))
1520 /* If isolation recently failed, do not retry */
1521 if (!isolation_suitable(cc, page))
1524 /* Found a block suitable for isolating free pages from. */
1525 nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn,
1526 block_end_pfn, freelist, stride, false);
1528 /* Update the skip hint if the full pageblock was scanned */
1529 if (isolate_start_pfn == block_end_pfn)
1530 update_pageblock_skip(cc, page, block_start_pfn);
1532 /* Are enough freepages isolated? */
1533 if (cc->nr_freepages >= cc->nr_migratepages) {
1534 if (isolate_start_pfn >= block_end_pfn) {
1536 * Restart at previous pageblock if more
1537 * freepages can be isolated next time.
1540 block_start_pfn - pageblock_nr_pages;
1543 } else if (isolate_start_pfn < block_end_pfn) {
1545 * If isolation failed early, do not continue
1551 /* Adjust stride depending on isolation */
1556 stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1);
1560 * Record where the free scanner will restart next time. Either we
1561 * broke from the loop and set isolate_start_pfn based on the last
1562 * call to isolate_freepages_block(), or we met the migration scanner
1563 * and the loop terminated due to isolate_start_pfn < low_pfn
1565 cc->free_pfn = isolate_start_pfn;
1568 /* __isolate_free_page() does not map the pages */
1569 split_map_pages(freelist);
1573 * This is a migrate-callback that "allocates" freepages by taking pages
1574 * from the isolated freelists in the block we are migrating to.
1576 static struct page *compaction_alloc(struct page *migratepage,
1579 struct compact_control *cc = (struct compact_control *)data;
1580 struct page *freepage;
1582 if (list_empty(&cc->freepages)) {
1583 isolate_freepages(cc);
1585 if (list_empty(&cc->freepages))
1589 freepage = list_entry(cc->freepages.next, struct page, lru);
1590 list_del(&freepage->lru);
1597 * This is a migrate-callback that "frees" freepages back to the isolated
1598 * freelist. All pages on the freelist are from the same zone, so there is no
1599 * special handling needed for NUMA.
1601 static void compaction_free(struct page *page, unsigned long data)
1603 struct compact_control *cc = (struct compact_control *)data;
1605 list_add(&page->lru, &cc->freepages);
1609 /* possible outcome of isolate_migratepages */
1611 ISOLATE_ABORT, /* Abort compaction now */
1612 ISOLATE_NONE, /* No pages isolated, continue scanning */
1613 ISOLATE_SUCCESS, /* Pages isolated, migrate */
1614 } isolate_migrate_t;
1617 * Allow userspace to control policy on scanning the unevictable LRU for
1618 * compactable pages.
1620 #ifdef CONFIG_PREEMPT_RT
1621 int sysctl_compact_unevictable_allowed __read_mostly = 0;
1623 int sysctl_compact_unevictable_allowed __read_mostly = 1;
1627 update_fast_start_pfn(struct compact_control *cc, unsigned long pfn)
1629 if (cc->fast_start_pfn == ULONG_MAX)
1632 if (!cc->fast_start_pfn)
1633 cc->fast_start_pfn = pfn;
1635 cc->fast_start_pfn = min(cc->fast_start_pfn, pfn);
1638 static inline unsigned long
1639 reinit_migrate_pfn(struct compact_control *cc)
1641 if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX)
1642 return cc->migrate_pfn;
1644 cc->migrate_pfn = cc->fast_start_pfn;
1645 cc->fast_start_pfn = ULONG_MAX;
1647 return cc->migrate_pfn;
1651 * Briefly search the free lists for a migration source that already has
1652 * some free pages to reduce the number of pages that need migration
1653 * before a pageblock is free.
1655 static unsigned long fast_find_migrateblock(struct compact_control *cc)
1657 unsigned int limit = freelist_scan_limit(cc);
1658 unsigned int nr_scanned = 0;
1659 unsigned long distance;
1660 unsigned long pfn = cc->migrate_pfn;
1661 unsigned long high_pfn;
1663 bool found_block = false;
1665 /* Skip hints are relied on to avoid repeats on the fast search */
1666 if (cc->ignore_skip_hint)
1670 * If the migrate_pfn is not at the start of a zone or the start
1671 * of a pageblock then assume this is a continuation of a previous
1672 * scan restarted due to COMPACT_CLUSTER_MAX.
1674 if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn))
1678 * For smaller orders, just linearly scan as the number of pages
1679 * to migrate should be relatively small and does not necessarily
1680 * justify freeing up a large block for a small allocation.
1682 if (cc->order <= PAGE_ALLOC_COSTLY_ORDER)
1686 * Only allow kcompactd and direct requests for movable pages to
1687 * quickly clear out a MOVABLE pageblock for allocation. This
1688 * reduces the risk that a large movable pageblock is freed for
1689 * an unmovable/reclaimable small allocation.
1691 if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE)
1695 * When starting the migration scanner, pick any pageblock within the
1696 * first half of the search space. Otherwise try and pick a pageblock
1697 * within the first eighth to reduce the chances that a migration
1698 * target later becomes a source.
1700 distance = (cc->free_pfn - cc->migrate_pfn) >> 1;
1701 if (cc->migrate_pfn != cc->zone->zone_start_pfn)
1703 high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance);
1705 for (order = cc->order - 1;
1706 order >= PAGE_ALLOC_COSTLY_ORDER && !found_block && nr_scanned < limit;
1708 struct free_area *area = &cc->zone->free_area[order];
1709 struct list_head *freelist;
1710 unsigned long flags;
1711 struct page *freepage;
1716 spin_lock_irqsave(&cc->zone->lock, flags);
1717 freelist = &area->free_list[MIGRATE_MOVABLE];
1718 list_for_each_entry(freepage, freelist, lru) {
1719 unsigned long free_pfn;
1721 if (nr_scanned++ >= limit) {
1722 move_freelist_tail(freelist, freepage);
1726 free_pfn = page_to_pfn(freepage);
1727 if (free_pfn < high_pfn) {
1729 * Avoid if skipped recently. Ideally it would
1730 * move to the tail but even safe iteration of
1731 * the list assumes an entry is deleted, not
1734 if (get_pageblock_skip(freepage))
1737 /* Reorder to so a future search skips recent pages */
1738 move_freelist_tail(freelist, freepage);
1740 update_fast_start_pfn(cc, free_pfn);
1741 pfn = pageblock_start_pfn(free_pfn);
1742 if (pfn < cc->zone->zone_start_pfn)
1743 pfn = cc->zone->zone_start_pfn;
1744 cc->fast_search_fail = 0;
1746 set_pageblock_skip(freepage);
1750 spin_unlock_irqrestore(&cc->zone->lock, flags);
1753 cc->total_migrate_scanned += nr_scanned;
1756 * If fast scanning failed then use a cached entry for a page block
1757 * that had free pages as the basis for starting a linear scan.
1760 cc->fast_search_fail++;
1761 pfn = reinit_migrate_pfn(cc);
1767 * Isolate all pages that can be migrated from the first suitable block,
1768 * starting at the block pointed to by the migrate scanner pfn within
1771 static isolate_migrate_t isolate_migratepages(struct compact_control *cc)
1773 unsigned long block_start_pfn;
1774 unsigned long block_end_pfn;
1775 unsigned long low_pfn;
1777 const isolate_mode_t isolate_mode =
1778 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1779 (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1780 bool fast_find_block;
1783 * Start at where we last stopped, or beginning of the zone as
1784 * initialized by compact_zone(). The first failure will use
1785 * the lowest PFN as the starting point for linear scanning.
1787 low_pfn = fast_find_migrateblock(cc);
1788 block_start_pfn = pageblock_start_pfn(low_pfn);
1789 if (block_start_pfn < cc->zone->zone_start_pfn)
1790 block_start_pfn = cc->zone->zone_start_pfn;
1793 * fast_find_migrateblock marks a pageblock skipped so to avoid
1794 * the isolation_suitable check below, check whether the fast
1795 * search was successful.
1797 fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail;
1799 /* Only scan within a pageblock boundary */
1800 block_end_pfn = pageblock_end_pfn(low_pfn);
1803 * Iterate over whole pageblocks until we find the first suitable.
1804 * Do not cross the free scanner.
1806 for (; block_end_pfn <= cc->free_pfn;
1807 fast_find_block = false,
1808 low_pfn = block_end_pfn,
1809 block_start_pfn = block_end_pfn,
1810 block_end_pfn += pageblock_nr_pages) {
1813 * This can potentially iterate a massively long zone with
1814 * many pageblocks unsuitable, so periodically check if we
1817 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)))
1820 page = pageblock_pfn_to_page(block_start_pfn,
1821 block_end_pfn, cc->zone);
1826 * If isolation recently failed, do not retry. Only check the
1827 * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock
1828 * to be visited multiple times. Assume skip was checked
1829 * before making it "skip" so other compaction instances do
1830 * not scan the same block.
1832 if (IS_ALIGNED(low_pfn, pageblock_nr_pages) &&
1833 !fast_find_block && !isolation_suitable(cc, page))
1837 * For async compaction, also only scan in MOVABLE blocks
1838 * without huge pages. Async compaction is optimistic to see
1839 * if the minimum amount of work satisfies the allocation.
1840 * The cached PFN is updated as it's possible that all
1841 * remaining blocks between source and target are unsuitable
1842 * and the compaction scanners fail to meet.
1844 if (!suitable_migration_source(cc, page)) {
1845 update_cached_migrate(cc, block_end_pfn);
1849 /* Perform the isolation */
1850 low_pfn = isolate_migratepages_block(cc, low_pfn,
1851 block_end_pfn, isolate_mode);
1854 return ISOLATE_ABORT;
1857 * Either we isolated something and proceed with migration. Or
1858 * we failed and compact_zone should decide if we should
1864 /* Record where migration scanner will be restarted. */
1865 cc->migrate_pfn = low_pfn;
1867 return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
1871 * order == -1 is expected when compacting via
1872 * /proc/sys/vm/compact_memory
1874 static inline bool is_via_compact_memory(int order)
1879 static bool kswapd_is_running(pg_data_t *pgdat)
1881 return pgdat->kswapd && (pgdat->kswapd->state == TASK_RUNNING);
1885 * A zone's fragmentation score is the external fragmentation wrt to the
1886 * COMPACTION_HPAGE_ORDER scaled by the zone's size. It returns a value
1887 * in the range [0, 100].
1889 * The scaling factor ensures that proactive compaction focuses on larger
1890 * zones like ZONE_NORMAL, rather than smaller, specialized zones like
1891 * ZONE_DMA32. For smaller zones, the score value remains close to zero,
1892 * and thus never exceeds the high threshold for proactive compaction.
1894 static unsigned int fragmentation_score_zone(struct zone *zone)
1896 unsigned long score;
1898 score = zone->present_pages *
1899 extfrag_for_order(zone, COMPACTION_HPAGE_ORDER);
1900 return div64_ul(score, zone->zone_pgdat->node_present_pages + 1);
1904 * The per-node proactive (background) compaction process is started by its
1905 * corresponding kcompactd thread when the node's fragmentation score
1906 * exceeds the high threshold. The compaction process remains active till
1907 * the node's score falls below the low threshold, or one of the back-off
1908 * conditions is met.
1910 static unsigned int fragmentation_score_node(pg_data_t *pgdat)
1912 unsigned int score = 0;
1915 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1918 zone = &pgdat->node_zones[zoneid];
1919 score += fragmentation_score_zone(zone);
1925 static unsigned int fragmentation_score_wmark(pg_data_t *pgdat, bool low)
1927 unsigned int wmark_low;
1930 * Cap the low watermak to avoid excessive compaction
1931 * activity in case a user sets the proactivess tunable
1932 * close to 100 (maximum).
1934 wmark_low = max(100U - sysctl_compaction_proactiveness, 5U);
1935 return low ? wmark_low : min(wmark_low + 10, 100U);
1938 static bool should_proactive_compact_node(pg_data_t *pgdat)
1942 if (!sysctl_compaction_proactiveness || kswapd_is_running(pgdat))
1945 wmark_high = fragmentation_score_wmark(pgdat, false);
1946 return fragmentation_score_node(pgdat) > wmark_high;
1949 static enum compact_result __compact_finished(struct compact_control *cc)
1952 const int migratetype = cc->migratetype;
1955 /* Compaction run completes if the migrate and free scanner meet */
1956 if (compact_scanners_met(cc)) {
1957 /* Let the next compaction start anew. */
1958 reset_cached_positions(cc->zone);
1961 * Mark that the PG_migrate_skip information should be cleared
1962 * by kswapd when it goes to sleep. kcompactd does not set the
1963 * flag itself as the decision to be clear should be directly
1964 * based on an allocation request.
1966 if (cc->direct_compaction)
1967 cc->zone->compact_blockskip_flush = true;
1970 return COMPACT_COMPLETE;
1972 return COMPACT_PARTIAL_SKIPPED;
1975 if (cc->proactive_compaction) {
1976 int score, wmark_low;
1979 pgdat = cc->zone->zone_pgdat;
1980 if (kswapd_is_running(pgdat))
1981 return COMPACT_PARTIAL_SKIPPED;
1983 score = fragmentation_score_zone(cc->zone);
1984 wmark_low = fragmentation_score_wmark(pgdat, true);
1986 if (score > wmark_low)
1987 ret = COMPACT_CONTINUE;
1989 ret = COMPACT_SUCCESS;
1994 if (is_via_compact_memory(cc->order))
1995 return COMPACT_CONTINUE;
1998 * Always finish scanning a pageblock to reduce the possibility of
1999 * fallbacks in the future. This is particularly important when
2000 * migration source is unmovable/reclaimable but it's not worth
2003 if (!IS_ALIGNED(cc->migrate_pfn, pageblock_nr_pages))
2004 return COMPACT_CONTINUE;
2006 /* Direct compactor: Is a suitable page free? */
2007 ret = COMPACT_NO_SUITABLE_PAGE;
2008 for (order = cc->order; order < MAX_ORDER; order++) {
2009 struct free_area *area = &cc->zone->free_area[order];
2012 /* Job done if page is free of the right migratetype */
2013 if (!free_area_empty(area, migratetype))
2014 return COMPACT_SUCCESS;
2017 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
2018 if (migratetype == MIGRATE_MOVABLE &&
2019 !free_area_empty(area, MIGRATE_CMA))
2020 return COMPACT_SUCCESS;
2023 * Job done if allocation would steal freepages from
2024 * other migratetype buddy lists.
2026 if (find_suitable_fallback(area, order, migratetype,
2027 true, &can_steal) != -1) {
2029 /* movable pages are OK in any pageblock */
2030 if (migratetype == MIGRATE_MOVABLE)
2031 return COMPACT_SUCCESS;
2034 * We are stealing for a non-movable allocation. Make
2035 * sure we finish compacting the current pageblock
2036 * first so it is as free as possible and we won't
2037 * have to steal another one soon. This only applies
2038 * to sync compaction, as async compaction operates
2039 * on pageblocks of the same migratetype.
2041 if (cc->mode == MIGRATE_ASYNC ||
2042 IS_ALIGNED(cc->migrate_pfn,
2043 pageblock_nr_pages)) {
2044 return COMPACT_SUCCESS;
2047 ret = COMPACT_CONTINUE;
2053 if (cc->contended || fatal_signal_pending(current))
2054 ret = COMPACT_CONTENDED;
2059 static enum compact_result compact_finished(struct compact_control *cc)
2063 ret = __compact_finished(cc);
2064 trace_mm_compaction_finished(cc->zone, cc->order, ret);
2065 if (ret == COMPACT_NO_SUITABLE_PAGE)
2066 ret = COMPACT_CONTINUE;
2072 * compaction_suitable: Is this suitable to run compaction on this zone now?
2074 * COMPACT_SKIPPED - If there are too few free pages for compaction
2075 * COMPACT_SUCCESS - If the allocation would succeed without compaction
2076 * COMPACT_CONTINUE - If compaction should run now
2078 static enum compact_result __compaction_suitable(struct zone *zone, int order,
2079 unsigned int alloc_flags,
2080 int highest_zoneidx,
2081 unsigned long wmark_target)
2083 unsigned long watermark;
2085 if (is_via_compact_memory(order))
2086 return COMPACT_CONTINUE;
2088 watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
2090 * If watermarks for high-order allocation are already met, there
2091 * should be no need for compaction at all.
2093 if (zone_watermark_ok(zone, order, watermark, highest_zoneidx,
2095 return COMPACT_SUCCESS;
2098 * Watermarks for order-0 must be met for compaction to be able to
2099 * isolate free pages for migration targets. This means that the
2100 * watermark and alloc_flags have to match, or be more pessimistic than
2101 * the check in __isolate_free_page(). We don't use the direct
2102 * compactor's alloc_flags, as they are not relevant for freepage
2103 * isolation. We however do use the direct compactor's highest_zoneidx
2104 * to skip over zones where lowmem reserves would prevent allocation
2105 * even if compaction succeeds.
2106 * For costly orders, we require low watermark instead of min for
2107 * compaction to proceed to increase its chances.
2108 * ALLOC_CMA is used, as pages in CMA pageblocks are considered
2109 * suitable migration targets
2111 watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
2112 low_wmark_pages(zone) : min_wmark_pages(zone);
2113 watermark += compact_gap(order);
2114 if (!__zone_watermark_ok(zone, 0, watermark, highest_zoneidx,
2115 ALLOC_CMA, wmark_target))
2116 return COMPACT_SKIPPED;
2118 return COMPACT_CONTINUE;
2121 enum compact_result compaction_suitable(struct zone *zone, int order,
2122 unsigned int alloc_flags,
2123 int highest_zoneidx)
2125 enum compact_result ret;
2128 ret = __compaction_suitable(zone, order, alloc_flags, highest_zoneidx,
2129 zone_page_state(zone, NR_FREE_PAGES));
2131 * fragmentation index determines if allocation failures are due to
2132 * low memory or external fragmentation
2134 * index of -1000 would imply allocations might succeed depending on
2135 * watermarks, but we already failed the high-order watermark check
2136 * index towards 0 implies failure is due to lack of memory
2137 * index towards 1000 implies failure is due to fragmentation
2139 * Only compact if a failure would be due to fragmentation. Also
2140 * ignore fragindex for non-costly orders where the alternative to
2141 * a successful reclaim/compaction is OOM. Fragindex and the
2142 * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
2143 * excessive compaction for costly orders, but it should not be at the
2144 * expense of system stability.
2146 if (ret == COMPACT_CONTINUE && (order > PAGE_ALLOC_COSTLY_ORDER)) {
2147 fragindex = fragmentation_index(zone, order);
2148 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
2149 ret = COMPACT_NOT_SUITABLE_ZONE;
2152 trace_mm_compaction_suitable(zone, order, ret);
2153 if (ret == COMPACT_NOT_SUITABLE_ZONE)
2154 ret = COMPACT_SKIPPED;
2159 bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
2166 * Make sure at least one zone would pass __compaction_suitable if we continue
2167 * retrying the reclaim.
2169 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2170 ac->highest_zoneidx, ac->nodemask) {
2171 unsigned long available;
2172 enum compact_result compact_result;
2175 * Do not consider all the reclaimable memory because we do not
2176 * want to trash just for a single high order allocation which
2177 * is even not guaranteed to appear even if __compaction_suitable
2178 * is happy about the watermark check.
2180 available = zone_reclaimable_pages(zone) / order;
2181 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
2182 compact_result = __compaction_suitable(zone, order, alloc_flags,
2183 ac->highest_zoneidx, available);
2184 if (compact_result != COMPACT_SKIPPED)
2191 static enum compact_result
2192 compact_zone(struct compact_control *cc, struct capture_control *capc)
2194 enum compact_result ret;
2195 unsigned long start_pfn = cc->zone->zone_start_pfn;
2196 unsigned long end_pfn = zone_end_pfn(cc->zone);
2197 unsigned long last_migrated_pfn;
2198 const bool sync = cc->mode != MIGRATE_ASYNC;
2202 * These counters track activities during zone compaction. Initialize
2203 * them before compacting a new zone.
2205 cc->total_migrate_scanned = 0;
2206 cc->total_free_scanned = 0;
2207 cc->nr_migratepages = 0;
2208 cc->nr_freepages = 0;
2209 INIT_LIST_HEAD(&cc->freepages);
2210 INIT_LIST_HEAD(&cc->migratepages);
2212 cc->migratetype = gfp_migratetype(cc->gfp_mask);
2213 ret = compaction_suitable(cc->zone, cc->order, cc->alloc_flags,
2214 cc->highest_zoneidx);
2215 /* Compaction is likely to fail */
2216 if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED)
2219 /* huh, compaction_suitable is returning something unexpected */
2220 VM_BUG_ON(ret != COMPACT_CONTINUE);
2223 * Clear pageblock skip if there were failures recently and compaction
2224 * is about to be retried after being deferred.
2226 if (compaction_restarting(cc->zone, cc->order))
2227 __reset_isolation_suitable(cc->zone);
2230 * Setup to move all movable pages to the end of the zone. Used cached
2231 * information on where the scanners should start (unless we explicitly
2232 * want to compact the whole zone), but check that it is initialised
2233 * by ensuring the values are within zone boundaries.
2235 cc->fast_start_pfn = 0;
2236 if (cc->whole_zone) {
2237 cc->migrate_pfn = start_pfn;
2238 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2240 cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync];
2241 cc->free_pfn = cc->zone->compact_cached_free_pfn;
2242 if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
2243 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2244 cc->zone->compact_cached_free_pfn = cc->free_pfn;
2246 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
2247 cc->migrate_pfn = start_pfn;
2248 cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
2249 cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
2252 if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn)
2253 cc->whole_zone = true;
2256 last_migrated_pfn = 0;
2259 * Migrate has separate cached PFNs for ASYNC and SYNC* migration on
2260 * the basis that some migrations will fail in ASYNC mode. However,
2261 * if the cached PFNs match and pageblocks are skipped due to having
2262 * no isolation candidates, then the sync state does not matter.
2263 * Until a pageblock with isolation candidates is found, keep the
2264 * cached PFNs in sync to avoid revisiting the same blocks.
2266 update_cached = !sync &&
2267 cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1];
2269 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
2270 cc->free_pfn, end_pfn, sync);
2272 migrate_prep_local();
2274 while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) {
2276 unsigned long start_pfn = cc->migrate_pfn;
2279 * Avoid multiple rescans which can happen if a page cannot be
2280 * isolated (dirty/writeback in async mode) or if the migrated
2281 * pages are being allocated before the pageblock is cleared.
2282 * The first rescan will capture the entire pageblock for
2283 * migration. If it fails, it'll be marked skip and scanning
2284 * will proceed as normal.
2287 if (pageblock_start_pfn(last_migrated_pfn) ==
2288 pageblock_start_pfn(start_pfn)) {
2292 switch (isolate_migratepages(cc)) {
2294 ret = COMPACT_CONTENDED;
2295 putback_movable_pages(&cc->migratepages);
2296 cc->nr_migratepages = 0;
2299 if (update_cached) {
2300 cc->zone->compact_cached_migrate_pfn[1] =
2301 cc->zone->compact_cached_migrate_pfn[0];
2305 * We haven't isolated and migrated anything, but
2306 * there might still be unflushed migrations from
2307 * previous cc->order aligned block.
2310 case ISOLATE_SUCCESS:
2311 update_cached = false;
2312 last_migrated_pfn = start_pfn;
2316 err = migrate_pages(&cc->migratepages, compaction_alloc,
2317 compaction_free, (unsigned long)cc, cc->mode,
2320 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
2323 /* All pages were either migrated or will be released */
2324 cc->nr_migratepages = 0;
2326 putback_movable_pages(&cc->migratepages);
2328 * migrate_pages() may return -ENOMEM when scanners meet
2329 * and we want compact_finished() to detect it
2331 if (err == -ENOMEM && !compact_scanners_met(cc)) {
2332 ret = COMPACT_CONTENDED;
2336 * We failed to migrate at least one page in the current
2337 * order-aligned block, so skip the rest of it.
2339 if (cc->direct_compaction &&
2340 (cc->mode == MIGRATE_ASYNC)) {
2341 cc->migrate_pfn = block_end_pfn(
2342 cc->migrate_pfn - 1, cc->order);
2343 /* Draining pcplists is useless in this case */
2344 last_migrated_pfn = 0;
2350 * Has the migration scanner moved away from the previous
2351 * cc->order aligned block where we migrated from? If yes,
2352 * flush the pages that were freed, so that they can merge and
2353 * compact_finished() can detect immediately if allocation
2356 if (cc->order > 0 && last_migrated_pfn) {
2357 unsigned long current_block_start =
2358 block_start_pfn(cc->migrate_pfn, cc->order);
2360 if (last_migrated_pfn < current_block_start) {
2361 lru_add_drain_cpu_zone(cc->zone);
2362 /* No more flushing until we migrate again */
2363 last_migrated_pfn = 0;
2367 /* Stop if a page has been captured */
2368 if (capc && capc->page) {
2369 ret = COMPACT_SUCCESS;
2376 * Release free pages and update where the free scanner should restart,
2377 * so we don't leave any returned pages behind in the next attempt.
2379 if (cc->nr_freepages > 0) {
2380 unsigned long free_pfn = release_freepages(&cc->freepages);
2382 cc->nr_freepages = 0;
2383 VM_BUG_ON(free_pfn == 0);
2384 /* The cached pfn is always the first in a pageblock */
2385 free_pfn = pageblock_start_pfn(free_pfn);
2387 * Only go back, not forward. The cached pfn might have been
2388 * already reset to zone end in compact_finished()
2390 if (free_pfn > cc->zone->compact_cached_free_pfn)
2391 cc->zone->compact_cached_free_pfn = free_pfn;
2394 count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned);
2395 count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned);
2397 trace_mm_compaction_end(start_pfn, cc->migrate_pfn,
2398 cc->free_pfn, end_pfn, sync, ret);
2403 static enum compact_result compact_zone_order(struct zone *zone, int order,
2404 gfp_t gfp_mask, enum compact_priority prio,
2405 unsigned int alloc_flags, int highest_zoneidx,
2406 struct page **capture)
2408 enum compact_result ret;
2409 struct compact_control cc = {
2411 .search_order = order,
2412 .gfp_mask = gfp_mask,
2414 .mode = (prio == COMPACT_PRIO_ASYNC) ?
2415 MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT,
2416 .alloc_flags = alloc_flags,
2417 .highest_zoneidx = highest_zoneidx,
2418 .direct_compaction = true,
2419 .whole_zone = (prio == MIN_COMPACT_PRIORITY),
2420 .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
2421 .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
2423 struct capture_control capc = {
2429 * Make sure the structs are really initialized before we expose the
2430 * capture control, in case we are interrupted and the interrupt handler
2434 WRITE_ONCE(current->capture_control, &capc);
2436 ret = compact_zone(&cc, &capc);
2438 VM_BUG_ON(!list_empty(&cc.freepages));
2439 VM_BUG_ON(!list_empty(&cc.migratepages));
2442 * Make sure we hide capture control first before we read the captured
2443 * page pointer, otherwise an interrupt could free and capture a page
2444 * and we would leak it.
2446 WRITE_ONCE(current->capture_control, NULL);
2447 *capture = READ_ONCE(capc.page);
2452 int sysctl_extfrag_threshold = 500;
2455 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
2456 * @gfp_mask: The GFP mask of the current allocation
2457 * @order: The order of the current allocation
2458 * @alloc_flags: The allocation flags of the current allocation
2459 * @ac: The context of current allocation
2460 * @prio: Determines how hard direct compaction should try to succeed
2461 * @capture: Pointer to free page created by compaction will be stored here
2463 * This is the main entry point for direct page compaction.
2465 enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
2466 unsigned int alloc_flags, const struct alloc_context *ac,
2467 enum compact_priority prio, struct page **capture)
2471 enum compact_result rc = COMPACT_SKIPPED;
2473 if (!gfp_compaction_allowed(gfp_mask))
2474 return COMPACT_SKIPPED;
2476 trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
2478 /* Compact each zone in the list */
2479 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2480 ac->highest_zoneidx, ac->nodemask) {
2481 enum compact_result status;
2483 if (prio > MIN_COMPACT_PRIORITY
2484 && compaction_deferred(zone, order)) {
2485 rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
2489 status = compact_zone_order(zone, order, gfp_mask, prio,
2490 alloc_flags, ac->highest_zoneidx, capture);
2491 rc = max(status, rc);
2493 /* The allocation should succeed, stop compacting */
2494 if (status == COMPACT_SUCCESS) {
2496 * We think the allocation will succeed in this zone,
2497 * but it is not certain, hence the false. The caller
2498 * will repeat this with true if allocation indeed
2499 * succeeds in this zone.
2501 compaction_defer_reset(zone, order, false);
2506 if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
2507 status == COMPACT_PARTIAL_SKIPPED))
2509 * We think that allocation won't succeed in this zone
2510 * so we defer compaction there. If it ends up
2511 * succeeding after all, it will be reset.
2513 defer_compaction(zone, order);
2516 * We might have stopped compacting due to need_resched() in
2517 * async compaction, or due to a fatal signal detected. In that
2518 * case do not try further zones
2520 if ((prio == COMPACT_PRIO_ASYNC && need_resched())
2521 || fatal_signal_pending(current))
2529 * Compact all zones within a node till each zone's fragmentation score
2530 * reaches within proactive compaction thresholds (as determined by the
2531 * proactiveness tunable).
2533 * It is possible that the function returns before reaching score targets
2534 * due to various back-off conditions, such as, contention on per-node or
2537 static void proactive_compact_node(pg_data_t *pgdat)
2541 struct compact_control cc = {
2543 .mode = MIGRATE_SYNC_LIGHT,
2544 .ignore_skip_hint = true,
2546 .gfp_mask = GFP_KERNEL,
2547 .proactive_compaction = true,
2550 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2551 zone = &pgdat->node_zones[zoneid];
2552 if (!populated_zone(zone))
2557 compact_zone(&cc, NULL);
2559 VM_BUG_ON(!list_empty(&cc.freepages));
2560 VM_BUG_ON(!list_empty(&cc.migratepages));
2564 /* Compact all zones within a node */
2565 static void compact_node(int nid)
2567 pg_data_t *pgdat = NODE_DATA(nid);
2570 struct compact_control cc = {
2572 .mode = MIGRATE_SYNC,
2573 .ignore_skip_hint = true,
2575 .gfp_mask = GFP_KERNEL,
2579 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2581 zone = &pgdat->node_zones[zoneid];
2582 if (!populated_zone(zone))
2587 compact_zone(&cc, NULL);
2589 VM_BUG_ON(!list_empty(&cc.freepages));
2590 VM_BUG_ON(!list_empty(&cc.migratepages));
2594 /* Compact all nodes in the system */
2595 static void compact_nodes(void)
2599 /* Flush pending updates to the LRU lists */
2600 lru_add_drain_all();
2602 for_each_online_node(nid)
2606 /* The written value is actually unused, all memory is compacted */
2607 int sysctl_compact_memory;
2610 * Tunable for proactive compaction. It determines how
2611 * aggressively the kernel should compact memory in the
2612 * background. It takes values in the range [0, 100].
2614 unsigned int __read_mostly sysctl_compaction_proactiveness = 20;
2617 * This is the entry point for compacting all nodes via
2618 * /proc/sys/vm/compact_memory
2620 int sysctl_compaction_handler(struct ctl_table *table, int write,
2621 void *buffer, size_t *length, loff_t *ppos)
2629 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
2630 static ssize_t sysfs_compact_node(struct device *dev,
2631 struct device_attribute *attr,
2632 const char *buf, size_t count)
2636 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
2637 /* Flush pending updates to the LRU lists */
2638 lru_add_drain_all();
2645 static DEVICE_ATTR(compact, 0200, NULL, sysfs_compact_node);
2647 int compaction_register_node(struct node *node)
2649 return device_create_file(&node->dev, &dev_attr_compact);
2652 void compaction_unregister_node(struct node *node)
2654 return device_remove_file(&node->dev, &dev_attr_compact);
2656 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
2658 static inline bool kcompactd_work_requested(pg_data_t *pgdat)
2660 return pgdat->kcompactd_max_order > 0 || kthread_should_stop();
2663 static bool kcompactd_node_suitable(pg_data_t *pgdat)
2667 enum zone_type highest_zoneidx = pgdat->kcompactd_highest_zoneidx;
2669 for (zoneid = 0; zoneid <= highest_zoneidx; zoneid++) {
2670 zone = &pgdat->node_zones[zoneid];
2672 if (!populated_zone(zone))
2675 if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0,
2676 highest_zoneidx) == COMPACT_CONTINUE)
2683 static void kcompactd_do_work(pg_data_t *pgdat)
2686 * With no special task, compact all zones so that a page of requested
2687 * order is allocatable.
2691 struct compact_control cc = {
2692 .order = pgdat->kcompactd_max_order,
2693 .search_order = pgdat->kcompactd_max_order,
2694 .highest_zoneidx = pgdat->kcompactd_highest_zoneidx,
2695 .mode = MIGRATE_SYNC_LIGHT,
2696 .ignore_skip_hint = false,
2697 .gfp_mask = GFP_KERNEL,
2699 trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
2700 cc.highest_zoneidx);
2701 count_compact_event(KCOMPACTD_WAKE);
2703 for (zoneid = 0; zoneid <= cc.highest_zoneidx; zoneid++) {
2706 zone = &pgdat->node_zones[zoneid];
2707 if (!populated_zone(zone))
2710 if (compaction_deferred(zone, cc.order))
2713 if (compaction_suitable(zone, cc.order, 0, zoneid) !=
2717 if (kthread_should_stop())
2721 status = compact_zone(&cc, NULL);
2723 if (status == COMPACT_SUCCESS) {
2724 compaction_defer_reset(zone, cc.order, false);
2725 } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
2727 * Buddy pages may become stranded on pcps that could
2728 * otherwise coalesce on the zone's free area for
2729 * order >= cc.order. This is ratelimited by the
2730 * upcoming deferral.
2732 drain_all_pages(zone);
2735 * We use sync migration mode here, so we defer like
2736 * sync direct compaction does.
2738 defer_compaction(zone, cc.order);
2741 count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
2742 cc.total_migrate_scanned);
2743 count_compact_events(KCOMPACTD_FREE_SCANNED,
2744 cc.total_free_scanned);
2746 VM_BUG_ON(!list_empty(&cc.freepages));
2747 VM_BUG_ON(!list_empty(&cc.migratepages));
2751 * Regardless of success, we are done until woken up next. But remember
2752 * the requested order/highest_zoneidx in case it was higher/tighter
2753 * than our current ones
2755 if (pgdat->kcompactd_max_order <= cc.order)
2756 pgdat->kcompactd_max_order = 0;
2757 if (pgdat->kcompactd_highest_zoneidx >= cc.highest_zoneidx)
2758 pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
2761 void wakeup_kcompactd(pg_data_t *pgdat, int order, int highest_zoneidx)
2766 if (pgdat->kcompactd_max_order < order)
2767 pgdat->kcompactd_max_order = order;
2769 if (pgdat->kcompactd_highest_zoneidx > highest_zoneidx)
2770 pgdat->kcompactd_highest_zoneidx = highest_zoneidx;
2773 * Pairs with implicit barrier in wait_event_freezable()
2774 * such that wakeups are not missed.
2776 if (!wq_has_sleeper(&pgdat->kcompactd_wait))
2779 if (!kcompactd_node_suitable(pgdat))
2782 trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
2784 wake_up_interruptible(&pgdat->kcompactd_wait);
2788 * The background compaction daemon, started as a kernel thread
2789 * from the init process.
2791 static int kcompactd(void *p)
2793 pg_data_t *pgdat = (pg_data_t*)p;
2794 struct task_struct *tsk = current;
2795 unsigned int proactive_defer = 0;
2797 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2799 if (!cpumask_empty(cpumask))
2800 set_cpus_allowed_ptr(tsk, cpumask);
2804 pgdat->kcompactd_max_order = 0;
2805 pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
2807 while (!kthread_should_stop()) {
2808 unsigned long pflags;
2810 trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
2811 if (wait_event_freezable_timeout(pgdat->kcompactd_wait,
2812 kcompactd_work_requested(pgdat),
2813 msecs_to_jiffies(HPAGE_FRAG_CHECK_INTERVAL_MSEC))) {
2815 psi_memstall_enter(&pflags);
2816 kcompactd_do_work(pgdat);
2817 psi_memstall_leave(&pflags);
2821 /* kcompactd wait timeout */
2822 if (should_proactive_compact_node(pgdat)) {
2823 unsigned int prev_score, score;
2825 if (proactive_defer) {
2829 prev_score = fragmentation_score_node(pgdat);
2830 proactive_compact_node(pgdat);
2831 score = fragmentation_score_node(pgdat);
2833 * Defer proactive compaction if the fragmentation
2834 * score did not go down i.e. no progress made.
2836 proactive_defer = score < prev_score ?
2837 0 : 1 << COMPACT_MAX_DEFER_SHIFT;
2845 * This kcompactd start function will be called by init and node-hot-add.
2846 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
2848 int kcompactd_run(int nid)
2850 pg_data_t *pgdat = NODE_DATA(nid);
2853 if (pgdat->kcompactd)
2856 pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
2857 if (IS_ERR(pgdat->kcompactd)) {
2858 pr_err("Failed to start kcompactd on node %d\n", nid);
2859 ret = PTR_ERR(pgdat->kcompactd);
2860 pgdat->kcompactd = NULL;
2866 * Called by memory hotplug when all memory in a node is offlined. Caller must
2867 * hold mem_hotplug_begin/end().
2869 void kcompactd_stop(int nid)
2871 struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
2874 kthread_stop(kcompactd);
2875 NODE_DATA(nid)->kcompactd = NULL;
2880 * It's optimal to keep kcompactd on the same CPUs as their memory, but
2881 * not required for correctness. So if the last cpu in a node goes
2882 * away, we get changed to run anywhere: as the first one comes back,
2883 * restore their cpu bindings.
2885 static int kcompactd_cpu_online(unsigned int cpu)
2889 for_each_node_state(nid, N_MEMORY) {
2890 pg_data_t *pgdat = NODE_DATA(nid);
2891 const struct cpumask *mask;
2893 mask = cpumask_of_node(pgdat->node_id);
2895 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2896 /* One of our CPUs online: restore mask */
2897 set_cpus_allowed_ptr(pgdat->kcompactd, mask);
2902 static int __init kcompactd_init(void)
2907 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
2908 "mm/compaction:online",
2909 kcompactd_cpu_online, NULL);
2911 pr_err("kcompactd: failed to register hotplug callbacks.\n");
2915 for_each_node_state(nid, N_MEMORY)
2919 subsys_initcall(kcompactd_init)
2921 #endif /* CONFIG_COMPACTION */