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(!PageMovable(page), page);
142 * Clear registered address_space val with keeping PAGE_MAPPING_MOVABLE
143 * flag so that VM can catch up released page by driver after isolation.
144 * With it, VM migration doesn't try to put it back.
146 page->mapping = (void *)((unsigned long)page->mapping &
147 PAGE_MAPPING_MOVABLE);
149 EXPORT_SYMBOL(__ClearPageMovable);
151 /* Do not skip compaction more than 64 times */
152 #define COMPACT_MAX_DEFER_SHIFT 6
155 * Compaction is deferred when compaction fails to result in a page
156 * allocation success. 1 << compact_defer_shift, compactions are skipped up
157 * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
159 static void defer_compaction(struct zone *zone, int order)
161 zone->compact_considered = 0;
162 zone->compact_defer_shift++;
164 if (order < zone->compact_order_failed)
165 zone->compact_order_failed = order;
167 if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
168 zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
170 trace_mm_compaction_defer_compaction(zone, order);
173 /* Returns true if compaction should be skipped this time */
174 static bool compaction_deferred(struct zone *zone, int order)
176 unsigned long defer_limit = 1UL << zone->compact_defer_shift;
178 if (order < zone->compact_order_failed)
181 /* Avoid possible overflow */
182 if (++zone->compact_considered >= defer_limit) {
183 zone->compact_considered = defer_limit;
187 trace_mm_compaction_deferred(zone, order);
193 * Update defer tracking counters after successful compaction of given order,
194 * which means an allocation either succeeded (alloc_success == true) or is
195 * expected to succeed.
197 void compaction_defer_reset(struct zone *zone, int order,
201 zone->compact_considered = 0;
202 zone->compact_defer_shift = 0;
204 if (order >= zone->compact_order_failed)
205 zone->compact_order_failed = order + 1;
207 trace_mm_compaction_defer_reset(zone, order);
210 /* Returns true if restarting compaction after many failures */
211 static bool compaction_restarting(struct zone *zone, int order)
213 if (order < zone->compact_order_failed)
216 return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
217 zone->compact_considered >= 1UL << zone->compact_defer_shift;
220 /* Returns true if the pageblock should be scanned for pages to isolate. */
221 static inline bool isolation_suitable(struct compact_control *cc,
224 if (cc->ignore_skip_hint)
227 return !get_pageblock_skip(page);
230 static void reset_cached_positions(struct zone *zone)
232 zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
233 zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
234 zone->compact_cached_free_pfn =
235 pageblock_start_pfn(zone_end_pfn(zone) - 1);
239 * Compound pages of >= pageblock_order should consistently be skipped until
240 * released. It is always pointless to compact pages of such order (if they are
241 * migratable), and the pageblocks they occupy cannot contain any free pages.
243 static bool pageblock_skip_persistent(struct page *page)
245 if (!PageCompound(page))
248 page = compound_head(page);
250 if (compound_order(page) >= pageblock_order)
257 __reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source,
260 struct page *page = pfn_to_online_page(pfn);
261 struct page *block_page;
262 struct page *end_page;
263 unsigned long block_pfn;
267 if (zone != page_zone(page))
269 if (pageblock_skip_persistent(page))
273 * If skip is already cleared do no further checking once the
274 * restart points have been set.
276 if (check_source && check_target && !get_pageblock_skip(page))
280 * If clearing skip for the target scanner, do not select a
281 * non-movable pageblock as the starting point.
283 if (!check_source && check_target &&
284 get_pageblock_migratetype(page) != MIGRATE_MOVABLE)
287 /* Ensure the start of the pageblock or zone is online and valid */
288 block_pfn = pageblock_start_pfn(pfn);
289 block_pfn = max(block_pfn, zone->zone_start_pfn);
290 block_page = pfn_to_online_page(block_pfn);
296 /* Ensure the end of the pageblock or zone is online and valid */
297 block_pfn = pageblock_end_pfn(pfn) - 1;
298 block_pfn = min(block_pfn, zone_end_pfn(zone) - 1);
299 end_page = pfn_to_online_page(block_pfn);
304 * Only clear the hint if a sample indicates there is either a
305 * free page or an LRU page in the block. One or other condition
306 * is necessary for the block to be a migration source/target.
309 if (pfn_valid_within(pfn)) {
310 if (check_source && PageLRU(page)) {
311 clear_pageblock_skip(page);
315 if (check_target && PageBuddy(page)) {
316 clear_pageblock_skip(page);
321 page += (1 << PAGE_ALLOC_COSTLY_ORDER);
322 pfn += (1 << PAGE_ALLOC_COSTLY_ORDER);
323 } while (page <= end_page);
329 * This function is called to clear all cached information on pageblocks that
330 * should be skipped for page isolation when the migrate and free page scanner
333 static void __reset_isolation_suitable(struct zone *zone)
335 unsigned long migrate_pfn = zone->zone_start_pfn;
336 unsigned long free_pfn = zone_end_pfn(zone) - 1;
337 unsigned long reset_migrate = free_pfn;
338 unsigned long reset_free = migrate_pfn;
339 bool source_set = false;
340 bool free_set = false;
342 if (!zone->compact_blockskip_flush)
345 zone->compact_blockskip_flush = false;
348 * Walk the zone and update pageblock skip information. Source looks
349 * for PageLRU while target looks for PageBuddy. When the scanner
350 * is found, both PageBuddy and PageLRU are checked as the pageblock
351 * is suitable as both source and target.
353 for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages,
354 free_pfn -= pageblock_nr_pages) {
357 /* Update the migrate PFN */
358 if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) &&
359 migrate_pfn < reset_migrate) {
361 reset_migrate = migrate_pfn;
362 zone->compact_init_migrate_pfn = reset_migrate;
363 zone->compact_cached_migrate_pfn[0] = reset_migrate;
364 zone->compact_cached_migrate_pfn[1] = reset_migrate;
367 /* Update the free PFN */
368 if (__reset_isolation_pfn(zone, free_pfn, free_set, true) &&
369 free_pfn > reset_free) {
371 reset_free = free_pfn;
372 zone->compact_init_free_pfn = reset_free;
373 zone->compact_cached_free_pfn = reset_free;
377 /* Leave no distance if no suitable block was reset */
378 if (reset_migrate >= reset_free) {
379 zone->compact_cached_migrate_pfn[0] = migrate_pfn;
380 zone->compact_cached_migrate_pfn[1] = migrate_pfn;
381 zone->compact_cached_free_pfn = free_pfn;
385 void reset_isolation_suitable(pg_data_t *pgdat)
389 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
390 struct zone *zone = &pgdat->node_zones[zoneid];
391 if (!populated_zone(zone))
394 /* Only flush if a full compaction finished recently */
395 if (zone->compact_blockskip_flush)
396 __reset_isolation_suitable(zone);
401 * Sets the pageblock skip bit if it was clear. Note that this is a hint as
402 * locks are not required for read/writers. Returns true if it was already set.
404 static bool test_and_set_skip(struct compact_control *cc, struct page *page,
409 /* Do no update if skip hint is being ignored */
410 if (cc->ignore_skip_hint)
413 if (!IS_ALIGNED(pfn, pageblock_nr_pages))
416 skip = get_pageblock_skip(page);
417 if (!skip && !cc->no_set_skip_hint)
418 set_pageblock_skip(page);
423 static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
425 struct zone *zone = cc->zone;
427 pfn = pageblock_end_pfn(pfn);
429 /* Set for isolation rather than compaction */
430 if (cc->no_set_skip_hint)
433 if (pfn > zone->compact_cached_migrate_pfn[0])
434 zone->compact_cached_migrate_pfn[0] = pfn;
435 if (cc->mode != MIGRATE_ASYNC &&
436 pfn > zone->compact_cached_migrate_pfn[1])
437 zone->compact_cached_migrate_pfn[1] = pfn;
441 * If no pages were isolated then mark this pageblock to be skipped in the
442 * future. The information is later cleared by __reset_isolation_suitable().
444 static void update_pageblock_skip(struct compact_control *cc,
445 struct page *page, unsigned long pfn)
447 struct zone *zone = cc->zone;
449 if (cc->no_set_skip_hint)
455 set_pageblock_skip(page);
457 /* Update where async and sync compaction should restart */
458 if (pfn < zone->compact_cached_free_pfn)
459 zone->compact_cached_free_pfn = pfn;
462 static inline bool isolation_suitable(struct compact_control *cc,
468 static inline bool pageblock_skip_persistent(struct page *page)
473 static inline void update_pageblock_skip(struct compact_control *cc,
474 struct page *page, unsigned long pfn)
478 static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
482 static bool test_and_set_skip(struct compact_control *cc, struct page *page,
487 #endif /* CONFIG_COMPACTION */
490 * Compaction requires the taking of some coarse locks that are potentially
491 * very heavily contended. For async compaction, trylock and record if the
492 * lock is contended. The lock will still be acquired but compaction will
493 * abort when the current block is finished regardless of success rate.
494 * Sync compaction acquires the lock.
496 * Always returns true which makes it easier to track lock state in callers.
498 static bool compact_lock_irqsave(spinlock_t *lock, unsigned long *flags,
499 struct compact_control *cc)
502 /* Track if the lock is contended in async mode */
503 if (cc->mode == MIGRATE_ASYNC && !cc->contended) {
504 if (spin_trylock_irqsave(lock, *flags))
507 cc->contended = true;
510 spin_lock_irqsave(lock, *flags);
515 * Compaction requires the taking of some coarse locks that are potentially
516 * very heavily contended. The lock should be periodically unlocked to avoid
517 * having disabled IRQs for a long time, even when there is nobody waiting on
518 * the lock. It might also be that allowing the IRQs will result in
519 * need_resched() becoming true. If scheduling is needed, async compaction
520 * aborts. Sync compaction schedules.
521 * Either compaction type will also abort if a fatal signal is pending.
522 * In either case if the lock was locked, it is dropped and not regained.
524 * Returns true if compaction should abort due to fatal signal pending, or
525 * async compaction due to need_resched()
526 * Returns false when compaction can continue (sync compaction might have
529 static bool compact_unlock_should_abort(spinlock_t *lock,
530 unsigned long flags, bool *locked, struct compact_control *cc)
533 spin_unlock_irqrestore(lock, flags);
537 if (fatal_signal_pending(current)) {
538 cc->contended = true;
548 * Isolate free pages onto a private freelist. If @strict is true, will abort
549 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
550 * (even though it may still end up isolating some pages).
552 static unsigned long isolate_freepages_block(struct compact_control *cc,
553 unsigned long *start_pfn,
554 unsigned long end_pfn,
555 struct list_head *freelist,
559 int nr_scanned = 0, total_isolated = 0;
561 unsigned long flags = 0;
563 unsigned long blockpfn = *start_pfn;
566 /* Strict mode is for isolation, speed is secondary */
570 cursor = pfn_to_page(blockpfn);
572 /* Isolate free pages. */
573 for (; blockpfn < end_pfn; blockpfn += stride, cursor += stride) {
575 struct page *page = cursor;
578 * Periodically drop the lock (if held) regardless of its
579 * contention, to give chance to IRQs. Abort if fatal signal
580 * pending or async compaction detects need_resched()
582 if (!(blockpfn % SWAP_CLUSTER_MAX)
583 && compact_unlock_should_abort(&cc->zone->lock, flags,
588 if (!pfn_valid_within(blockpfn))
592 * For compound pages such as THP and hugetlbfs, we can save
593 * potentially a lot of iterations if we skip them at once.
594 * The check is racy, but we can consider only valid values
595 * and the only danger is skipping too much.
597 if (PageCompound(page)) {
598 const unsigned int order = compound_order(page);
600 if (likely(order < MAX_ORDER)) {
601 blockpfn += (1UL << order) - 1;
602 cursor += (1UL << order) - 1;
607 if (!PageBuddy(page))
611 * If we already hold the lock, we can skip some rechecking.
612 * Note that if we hold the lock now, checked_pageblock was
613 * already set in some previous iteration (or strict is true),
614 * so it is correct to skip the suitable migration target
618 locked = compact_lock_irqsave(&cc->zone->lock,
621 /* Recheck this is a buddy page under lock */
622 if (!PageBuddy(page))
626 /* Found a free page, will break it into order-0 pages */
627 order = buddy_order(page);
628 isolated = __isolate_free_page(page, order);
631 set_page_private(page, order);
633 total_isolated += isolated;
634 cc->nr_freepages += isolated;
635 list_add_tail(&page->lru, freelist);
637 if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
638 blockpfn += isolated;
641 /* Advance to the end of split page */
642 blockpfn += isolated - 1;
643 cursor += isolated - 1;
655 spin_unlock_irqrestore(&cc->zone->lock, flags);
658 * There is a tiny chance that we have read bogus compound_order(),
659 * so be careful to not go outside of the pageblock.
661 if (unlikely(blockpfn > end_pfn))
664 trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
665 nr_scanned, total_isolated);
667 /* Record how far we have got within the block */
668 *start_pfn = blockpfn;
671 * If strict isolation is requested by CMA then check that all the
672 * pages requested were isolated. If there were any failures, 0 is
673 * returned and CMA will fail.
675 if (strict && blockpfn < end_pfn)
678 cc->total_free_scanned += nr_scanned;
680 count_compact_events(COMPACTISOLATED, total_isolated);
681 return total_isolated;
685 * isolate_freepages_range() - isolate free pages.
686 * @cc: Compaction control structure.
687 * @start_pfn: The first PFN to start isolating.
688 * @end_pfn: The one-past-last PFN.
690 * Non-free pages, invalid PFNs, or zone boundaries within the
691 * [start_pfn, end_pfn) range are considered errors, cause function to
692 * undo its actions and return zero.
694 * Otherwise, function returns one-past-the-last PFN of isolated page
695 * (which may be greater then end_pfn if end fell in a middle of
699 isolate_freepages_range(struct compact_control *cc,
700 unsigned long start_pfn, unsigned long end_pfn)
702 unsigned long isolated, pfn, block_start_pfn, block_end_pfn;
706 block_start_pfn = pageblock_start_pfn(pfn);
707 if (block_start_pfn < cc->zone->zone_start_pfn)
708 block_start_pfn = cc->zone->zone_start_pfn;
709 block_end_pfn = pageblock_end_pfn(pfn);
711 for (; pfn < end_pfn; pfn += isolated,
712 block_start_pfn = block_end_pfn,
713 block_end_pfn += pageblock_nr_pages) {
714 /* Protect pfn from changing by isolate_freepages_block */
715 unsigned long isolate_start_pfn = pfn;
717 block_end_pfn = min(block_end_pfn, end_pfn);
720 * pfn could pass the block_end_pfn if isolated freepage
721 * is more than pageblock order. In this case, we adjust
722 * scanning range to right one.
724 if (pfn >= block_end_pfn) {
725 block_start_pfn = pageblock_start_pfn(pfn);
726 block_end_pfn = pageblock_end_pfn(pfn);
727 block_end_pfn = min(block_end_pfn, end_pfn);
730 if (!pageblock_pfn_to_page(block_start_pfn,
731 block_end_pfn, cc->zone))
734 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
735 block_end_pfn, &freelist, 0, true);
738 * In strict mode, isolate_freepages_block() returns 0 if
739 * there are any holes in the block (ie. invalid PFNs or
746 * If we managed to isolate pages, it is always (1 << n) *
747 * pageblock_nr_pages for some non-negative n. (Max order
748 * page may span two pageblocks).
752 /* __isolate_free_page() does not map the pages */
753 split_map_pages(&freelist);
756 /* Loop terminated early, cleanup. */
757 release_freepages(&freelist);
761 /* We don't use freelists for anything. */
765 /* Similar to reclaim, but different enough that they don't share logic */
766 static bool too_many_isolated(pg_data_t *pgdat)
768 unsigned long active, inactive, isolated;
770 inactive = node_page_state(pgdat, NR_INACTIVE_FILE) +
771 node_page_state(pgdat, NR_INACTIVE_ANON);
772 active = node_page_state(pgdat, NR_ACTIVE_FILE) +
773 node_page_state(pgdat, NR_ACTIVE_ANON);
774 isolated = node_page_state(pgdat, NR_ISOLATED_FILE) +
775 node_page_state(pgdat, NR_ISOLATED_ANON);
777 return isolated > (inactive + active) / 2;
781 * isolate_migratepages_block() - isolate all migrate-able pages within
783 * @cc: Compaction control structure.
784 * @low_pfn: The first PFN to isolate
785 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
786 * @isolate_mode: Isolation mode to be used.
788 * Isolate all pages that can be migrated from the range specified by
789 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
790 * Returns zero if there is a fatal signal pending, otherwise PFN of the
791 * first page that was not scanned (which may be both less, equal to or more
794 * The pages are isolated on cc->migratepages list (not required to be empty),
795 * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
796 * is neither read nor updated.
799 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
800 unsigned long end_pfn, isolate_mode_t isolate_mode)
802 pg_data_t *pgdat = cc->zone->zone_pgdat;
803 unsigned long nr_scanned = 0, nr_isolated = 0;
804 struct lruvec *lruvec;
805 unsigned long flags = 0;
806 struct lruvec *locked = NULL;
807 struct page *page = NULL, *valid_page = NULL;
808 unsigned long start_pfn = low_pfn;
809 bool skip_on_failure = false;
810 unsigned long next_skip_pfn = 0;
811 bool skip_updated = false;
814 * Ensure that there are not too many pages isolated from the LRU
815 * list by either parallel reclaimers or compaction. If there are,
816 * delay for some time until fewer pages are isolated
818 while (unlikely(too_many_isolated(pgdat))) {
819 /* stop isolation if there are still pages not migrated */
820 if (cc->nr_migratepages)
823 /* async migration should just abort */
824 if (cc->mode == MIGRATE_ASYNC)
827 congestion_wait(BLK_RW_ASYNC, HZ/10);
829 if (fatal_signal_pending(current))
835 if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
836 skip_on_failure = true;
837 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
840 /* Time to isolate some pages for migration */
841 for (; low_pfn < end_pfn; low_pfn++) {
843 if (skip_on_failure && low_pfn >= next_skip_pfn) {
845 * We have isolated all migration candidates in the
846 * previous order-aligned block, and did not skip it due
847 * to failure. We should migrate the pages now and
848 * hopefully succeed compaction.
854 * We failed to isolate in the previous order-aligned
855 * block. Set the new boundary to the end of the
856 * current block. Note we can't simply increase
857 * next_skip_pfn by 1 << order, as low_pfn might have
858 * been incremented by a higher number due to skipping
859 * a compound or a high-order buddy page in the
860 * previous loop iteration.
862 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
866 * Periodically drop the lock (if held) regardless of its
867 * contention, to give chance to IRQs. Abort completely if
868 * a fatal signal is pending.
870 if (!(low_pfn % SWAP_CLUSTER_MAX)) {
872 unlock_page_lruvec_irqrestore(locked, flags);
876 if (fatal_signal_pending(current)) {
877 cc->contended = true;
886 if (!pfn_valid_within(low_pfn))
890 page = pfn_to_page(low_pfn);
893 * Check if the pageblock has already been marked skipped.
894 * Only the aligned PFN is checked as the caller isolates
895 * COMPACT_CLUSTER_MAX at a time so the second call must
896 * not falsely conclude that the block should be skipped.
898 if (!valid_page && IS_ALIGNED(low_pfn, pageblock_nr_pages)) {
899 if (!cc->ignore_skip_hint && get_pageblock_skip(page)) {
908 * Skip if free. We read page order here without zone lock
909 * which is generally unsafe, but the race window is small and
910 * the worst thing that can happen is that we skip some
911 * potential isolation targets.
913 if (PageBuddy(page)) {
914 unsigned long freepage_order = buddy_order_unsafe(page);
917 * Without lock, we cannot be sure that what we got is
918 * a valid page order. Consider only values in the
919 * valid order range to prevent low_pfn overflow.
921 if (freepage_order > 0 && freepage_order < MAX_ORDER)
922 low_pfn += (1UL << freepage_order) - 1;
927 * Regardless of being on LRU, compound pages such as THP and
928 * hugetlbfs are not to be compacted unless we are attempting
929 * an allocation much larger than the huge page size (eg CMA).
930 * We can potentially save a lot of iterations if we skip them
931 * at once. The check is racy, but we can consider only valid
932 * values and the only danger is skipping too much.
934 if (PageCompound(page) && !cc->alloc_contig) {
935 const unsigned int order = compound_order(page);
937 if (likely(order < MAX_ORDER))
938 low_pfn += (1UL << order) - 1;
943 * Check may be lockless but that's ok as we recheck later.
944 * It's possible to migrate LRU and non-lru movable pages.
945 * Skip any other type of page
947 if (!PageLRU(page)) {
949 * __PageMovable can return false positive so we need
950 * to verify it under page_lock.
952 if (unlikely(__PageMovable(page)) &&
953 !PageIsolated(page)) {
955 unlock_page_lruvec_irqrestore(locked, flags);
959 if (!isolate_movable_page(page, isolate_mode))
960 goto isolate_success;
967 * Migration will fail if an anonymous page is pinned in memory,
968 * so avoid taking lru_lock and isolating it unnecessarily in an
969 * admittedly racy check.
971 if (!page_mapping(page) &&
972 page_count(page) > page_mapcount(page))
976 * Only allow to migrate anonymous pages in GFP_NOFS context
977 * because those do not depend on fs locks.
979 if (!(cc->gfp_mask & __GFP_FS) && page_mapping(page))
983 * Be careful not to clear PageLRU until after we're
984 * sure the page is not being freed elsewhere -- the
985 * page release code relies on it.
987 if (unlikely(!get_page_unless_zero(page)))
990 if (!__isolate_lru_page_prepare(page, isolate_mode))
991 goto isolate_fail_put;
993 /* Try isolate the page */
994 if (!TestClearPageLRU(page))
995 goto isolate_fail_put;
997 lruvec = mem_cgroup_page_lruvec(page, pgdat);
999 /* If we already hold the lock, we can skip some rechecking */
1000 if (lruvec != locked) {
1002 unlock_page_lruvec_irqrestore(locked, flags);
1004 compact_lock_irqsave(&lruvec->lru_lock, &flags, cc);
1007 lruvec_memcg_debug(lruvec, page);
1009 /* Try get exclusive access under lock */
1010 if (!skip_updated) {
1011 skip_updated = true;
1012 if (test_and_set_skip(cc, page, low_pfn))
1017 * Page become compound since the non-locked check,
1018 * and it's on LRU. It can only be a THP so the order
1019 * is safe to read and it's 0 for tail pages.
1021 if (unlikely(PageCompound(page) && !cc->alloc_contig)) {
1022 low_pfn += compound_nr(page) - 1;
1024 goto isolate_fail_put;
1028 /* The whole page is taken off the LRU; skip the tail pages. */
1029 if (PageCompound(page))
1030 low_pfn += compound_nr(page) - 1;
1032 /* Successfully isolated */
1033 del_page_from_lru_list(page, lruvec);
1034 mod_node_page_state(page_pgdat(page),
1035 NR_ISOLATED_ANON + page_is_file_lru(page),
1036 thp_nr_pages(page));
1039 list_add(&page->lru, &cc->migratepages);
1040 cc->nr_migratepages += compound_nr(page);
1041 nr_isolated += compound_nr(page);
1044 * Avoid isolating too much unless this block is being
1045 * rescanned (e.g. dirty/writeback pages, parallel allocation)
1046 * or a lock is contended. For contention, isolate quickly to
1047 * potentially remove one source of contention.
1049 if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX &&
1050 !cc->rescan && !cc->contended) {
1058 /* Avoid potential deadlock in freeing page under lru_lock */
1060 unlock_page_lruvec_irqrestore(locked, flags);
1066 if (!skip_on_failure)
1070 * We have isolated some pages, but then failed. Release them
1071 * instead of migrating, as we cannot form the cc->order buddy
1076 unlock_page_lruvec_irqrestore(locked, flags);
1079 putback_movable_pages(&cc->migratepages);
1080 cc->nr_migratepages = 0;
1084 if (low_pfn < next_skip_pfn) {
1085 low_pfn = next_skip_pfn - 1;
1087 * The check near the loop beginning would have updated
1088 * next_skip_pfn too, but this is a bit simpler.
1090 next_skip_pfn += 1UL << cc->order;
1095 * The PageBuddy() check could have potentially brought us outside
1096 * the range to be scanned.
1098 if (unlikely(low_pfn > end_pfn))
1105 unlock_page_lruvec_irqrestore(locked, flags);
1112 * Updated the cached scanner pfn once the pageblock has been scanned
1113 * Pages will either be migrated in which case there is no point
1114 * scanning in the near future or migration failed in which case the
1115 * failure reason may persist. The block is marked for skipping if
1116 * there were no pages isolated in the block or if the block is
1117 * rescanned twice in a row.
1119 if (low_pfn == end_pfn && (!nr_isolated || cc->rescan)) {
1120 if (valid_page && !skip_updated)
1121 set_pageblock_skip(valid_page);
1122 update_cached_migrate(cc, low_pfn);
1125 trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
1126 nr_scanned, nr_isolated);
1129 cc->total_migrate_scanned += nr_scanned;
1131 count_compact_events(COMPACTISOLATED, nr_isolated);
1137 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
1138 * @cc: Compaction control structure.
1139 * @start_pfn: The first PFN to start isolating.
1140 * @end_pfn: The one-past-last PFN.
1142 * Returns zero if isolation fails fatally due to e.g. pending signal.
1143 * Otherwise, function returns one-past-the-last PFN of isolated page
1144 * (which may be greater than end_pfn if end fell in a middle of a THP page).
1147 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
1148 unsigned long end_pfn)
1150 unsigned long pfn, block_start_pfn, block_end_pfn;
1152 /* Scan block by block. First and last block may be incomplete */
1154 block_start_pfn = pageblock_start_pfn(pfn);
1155 if (block_start_pfn < cc->zone->zone_start_pfn)
1156 block_start_pfn = cc->zone->zone_start_pfn;
1157 block_end_pfn = pageblock_end_pfn(pfn);
1159 for (; pfn < end_pfn; pfn = block_end_pfn,
1160 block_start_pfn = block_end_pfn,
1161 block_end_pfn += pageblock_nr_pages) {
1163 block_end_pfn = min(block_end_pfn, end_pfn);
1165 if (!pageblock_pfn_to_page(block_start_pfn,
1166 block_end_pfn, cc->zone))
1169 pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
1170 ISOLATE_UNEVICTABLE);
1175 if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX)
1182 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
1183 #ifdef CONFIG_COMPACTION
1185 static bool suitable_migration_source(struct compact_control *cc,
1190 if (pageblock_skip_persistent(page))
1193 if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction)
1196 block_mt = get_pageblock_migratetype(page);
1198 if (cc->migratetype == MIGRATE_MOVABLE)
1199 return is_migrate_movable(block_mt);
1201 return block_mt == cc->migratetype;
1204 /* Returns true if the page is within a block suitable for migration to */
1205 static bool suitable_migration_target(struct compact_control *cc,
1208 /* If the page is a large free page, then disallow migration */
1209 if (PageBuddy(page)) {
1211 * We are checking page_order without zone->lock taken. But
1212 * the only small danger is that we skip a potentially suitable
1213 * pageblock, so it's not worth to check order for valid range.
1215 if (buddy_order_unsafe(page) >= pageblock_order)
1219 if (cc->ignore_block_suitable)
1222 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1223 if (is_migrate_movable(get_pageblock_migratetype(page)))
1226 /* Otherwise skip the block */
1230 static inline unsigned int
1231 freelist_scan_limit(struct compact_control *cc)
1233 unsigned short shift = BITS_PER_LONG - 1;
1235 return (COMPACT_CLUSTER_MAX >> min(shift, cc->fast_search_fail)) + 1;
1239 * Test whether the free scanner has reached the same or lower pageblock than
1240 * the migration scanner, and compaction should thus terminate.
1242 static inline bool compact_scanners_met(struct compact_control *cc)
1244 return (cc->free_pfn >> pageblock_order)
1245 <= (cc->migrate_pfn >> pageblock_order);
1249 * Used when scanning for a suitable migration target which scans freelists
1250 * in reverse. Reorders the list such as the unscanned pages are scanned
1251 * first on the next iteration of the free scanner
1254 move_freelist_head(struct list_head *freelist, struct page *freepage)
1258 if (!list_is_last(freelist, &freepage->lru)) {
1259 list_cut_before(&sublist, freelist, &freepage->lru);
1260 if (!list_empty(&sublist))
1261 list_splice_tail(&sublist, freelist);
1266 * Similar to move_freelist_head except used by the migration scanner
1267 * when scanning forward. It's possible for these list operations to
1268 * move against each other if they search the free list exactly in
1272 move_freelist_tail(struct list_head *freelist, struct page *freepage)
1276 if (!list_is_first(freelist, &freepage->lru)) {
1277 list_cut_position(&sublist, freelist, &freepage->lru);
1278 if (!list_empty(&sublist))
1279 list_splice_tail(&sublist, freelist);
1284 fast_isolate_around(struct compact_control *cc, unsigned long pfn, unsigned long nr_isolated)
1286 unsigned long start_pfn, end_pfn;
1287 struct page *page = pfn_to_page(pfn);
1289 /* Do not search around if there are enough pages already */
1290 if (cc->nr_freepages >= cc->nr_migratepages)
1293 /* Minimise scanning during async compaction */
1294 if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC)
1297 /* Pageblock boundaries */
1298 start_pfn = pageblock_start_pfn(pfn);
1299 end_pfn = min(pageblock_end_pfn(pfn), zone_end_pfn(cc->zone)) - 1;
1302 if (start_pfn != pfn) {
1303 isolate_freepages_block(cc, &start_pfn, pfn, &cc->freepages, 1, false);
1304 if (cc->nr_freepages >= cc->nr_migratepages)
1309 start_pfn = pfn + nr_isolated;
1310 if (start_pfn < end_pfn)
1311 isolate_freepages_block(cc, &start_pfn, end_pfn, &cc->freepages, 1, false);
1313 /* Skip this pageblock in the future as it's full or nearly full */
1314 if (cc->nr_freepages < cc->nr_migratepages)
1315 set_pageblock_skip(page);
1318 /* Search orders in round-robin fashion */
1319 static int next_search_order(struct compact_control *cc, int order)
1323 order = cc->order - 1;
1325 /* Search wrapped around? */
1326 if (order == cc->search_order) {
1328 if (cc->search_order < 0)
1329 cc->search_order = cc->order - 1;
1336 static unsigned long
1337 fast_isolate_freepages(struct compact_control *cc)
1339 unsigned int limit = min(1U, freelist_scan_limit(cc) >> 1);
1340 unsigned int nr_scanned = 0;
1341 unsigned long low_pfn, min_pfn, highest = 0;
1342 unsigned long nr_isolated = 0;
1343 unsigned long distance;
1344 struct page *page = NULL;
1345 bool scan_start = false;
1348 /* Full compaction passes in a negative order */
1350 return cc->free_pfn;
1353 * If starting the scan, use a deeper search and use the highest
1354 * PFN found if a suitable one is not found.
1356 if (cc->free_pfn >= cc->zone->compact_init_free_pfn) {
1357 limit = pageblock_nr_pages >> 1;
1362 * Preferred point is in the top quarter of the scan space but take
1363 * a pfn from the top half if the search is problematic.
1365 distance = (cc->free_pfn - cc->migrate_pfn);
1366 low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2));
1367 min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1));
1369 if (WARN_ON_ONCE(min_pfn > low_pfn))
1373 * Search starts from the last successful isolation order or the next
1374 * order to search after a previous failure
1376 cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order);
1378 for (order = cc->search_order;
1379 !page && order >= 0;
1380 order = next_search_order(cc, order)) {
1381 struct free_area *area = &cc->zone->free_area[order];
1382 struct list_head *freelist;
1383 struct page *freepage;
1384 unsigned long flags;
1385 unsigned int order_scanned = 0;
1386 unsigned long high_pfn = 0;
1391 spin_lock_irqsave(&cc->zone->lock, flags);
1392 freelist = &area->free_list[MIGRATE_MOVABLE];
1393 list_for_each_entry_reverse(freepage, freelist, lru) {
1398 pfn = page_to_pfn(freepage);
1401 highest = pageblock_start_pfn(pfn);
1403 if (pfn >= low_pfn) {
1404 cc->fast_search_fail = 0;
1405 cc->search_order = order;
1410 if (pfn >= min_pfn && pfn > high_pfn) {
1413 /* Shorten the scan if a candidate is found */
1417 if (order_scanned >= limit)
1421 /* Use a minimum pfn if a preferred one was not found */
1422 if (!page && high_pfn) {
1423 page = pfn_to_page(high_pfn);
1425 /* Update freepage for the list reorder below */
1429 /* Reorder to so a future search skips recent pages */
1430 move_freelist_head(freelist, freepage);
1432 /* Isolate the page if available */
1434 if (__isolate_free_page(page, order)) {
1435 set_page_private(page, order);
1436 nr_isolated = 1 << order;
1437 cc->nr_freepages += nr_isolated;
1438 list_add_tail(&page->lru, &cc->freepages);
1439 count_compact_events(COMPACTISOLATED, nr_isolated);
1441 /* If isolation fails, abort the search */
1442 order = cc->search_order + 1;
1447 spin_unlock_irqrestore(&cc->zone->lock, flags);
1450 * Smaller scan on next order so the total scan ig related
1451 * to freelist_scan_limit.
1453 if (order_scanned >= limit)
1454 limit = min(1U, limit >> 1);
1458 cc->fast_search_fail++;
1461 * Use the highest PFN found above min. If one was
1462 * not found, be pessimistic for direct compaction
1463 * and use the min mark.
1466 page = pfn_to_page(highest);
1467 cc->free_pfn = highest;
1469 if (cc->direct_compaction && pfn_valid(min_pfn)) {
1470 page = pageblock_pfn_to_page(min_pfn,
1471 pageblock_end_pfn(min_pfn),
1473 cc->free_pfn = min_pfn;
1479 if (highest && highest >= cc->zone->compact_cached_free_pfn) {
1480 highest -= pageblock_nr_pages;
1481 cc->zone->compact_cached_free_pfn = highest;
1484 cc->total_free_scanned += nr_scanned;
1486 return cc->free_pfn;
1488 low_pfn = page_to_pfn(page);
1489 fast_isolate_around(cc, low_pfn, nr_isolated);
1494 * Based on information in the current compact_control, find blocks
1495 * suitable for isolating free pages from and then isolate them.
1497 static void isolate_freepages(struct compact_control *cc)
1499 struct zone *zone = cc->zone;
1501 unsigned long block_start_pfn; /* start of current pageblock */
1502 unsigned long isolate_start_pfn; /* exact pfn we start at */
1503 unsigned long block_end_pfn; /* end of current pageblock */
1504 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
1505 struct list_head *freelist = &cc->freepages;
1506 unsigned int stride;
1508 /* Try a small search of the free lists for a candidate */
1509 isolate_start_pfn = fast_isolate_freepages(cc);
1510 if (cc->nr_freepages)
1514 * Initialise the free scanner. The starting point is where we last
1515 * successfully isolated from, zone-cached value, or the end of the
1516 * zone when isolating for the first time. For looping we also need
1517 * this pfn aligned down to the pageblock boundary, because we do
1518 * block_start_pfn -= pageblock_nr_pages in the for loop.
1519 * For ending point, take care when isolating in last pageblock of a
1520 * zone which ends in the middle of a pageblock.
1521 * The low boundary is the end of the pageblock the migration scanner
1524 isolate_start_pfn = cc->free_pfn;
1525 block_start_pfn = pageblock_start_pfn(isolate_start_pfn);
1526 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1527 zone_end_pfn(zone));
1528 low_pfn = pageblock_end_pfn(cc->migrate_pfn);
1529 stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1;
1532 * Isolate free pages until enough are available to migrate the
1533 * pages on cc->migratepages. We stop searching if the migrate
1534 * and free page scanners meet or enough free pages are isolated.
1536 for (; block_start_pfn >= low_pfn;
1537 block_end_pfn = block_start_pfn,
1538 block_start_pfn -= pageblock_nr_pages,
1539 isolate_start_pfn = block_start_pfn) {
1540 unsigned long nr_isolated;
1543 * This can iterate a massively long zone without finding any
1544 * suitable migration targets, so periodically check resched.
1546 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)))
1549 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1554 /* Check the block is suitable for migration */
1555 if (!suitable_migration_target(cc, page))
1558 /* If isolation recently failed, do not retry */
1559 if (!isolation_suitable(cc, page))
1562 /* Found a block suitable for isolating free pages from. */
1563 nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn,
1564 block_end_pfn, freelist, stride, false);
1566 /* Update the skip hint if the full pageblock was scanned */
1567 if (isolate_start_pfn == block_end_pfn)
1568 update_pageblock_skip(cc, page, block_start_pfn);
1570 /* Are enough freepages isolated? */
1571 if (cc->nr_freepages >= cc->nr_migratepages) {
1572 if (isolate_start_pfn >= block_end_pfn) {
1574 * Restart at previous pageblock if more
1575 * freepages can be isolated next time.
1578 block_start_pfn - pageblock_nr_pages;
1581 } else if (isolate_start_pfn < block_end_pfn) {
1583 * If isolation failed early, do not continue
1589 /* Adjust stride depending on isolation */
1594 stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1);
1598 * Record where the free scanner will restart next time. Either we
1599 * broke from the loop and set isolate_start_pfn based on the last
1600 * call to isolate_freepages_block(), or we met the migration scanner
1601 * and the loop terminated due to isolate_start_pfn < low_pfn
1603 cc->free_pfn = isolate_start_pfn;
1606 /* __isolate_free_page() does not map the pages */
1607 split_map_pages(freelist);
1611 * This is a migrate-callback that "allocates" freepages by taking pages
1612 * from the isolated freelists in the block we are migrating to.
1614 static struct page *compaction_alloc(struct page *migratepage,
1617 struct compact_control *cc = (struct compact_control *)data;
1618 struct page *freepage;
1620 if (list_empty(&cc->freepages)) {
1621 isolate_freepages(cc);
1623 if (list_empty(&cc->freepages))
1627 freepage = list_entry(cc->freepages.next, struct page, lru);
1628 list_del(&freepage->lru);
1635 * This is a migrate-callback that "frees" freepages back to the isolated
1636 * freelist. All pages on the freelist are from the same zone, so there is no
1637 * special handling needed for NUMA.
1639 static void compaction_free(struct page *page, unsigned long data)
1641 struct compact_control *cc = (struct compact_control *)data;
1643 list_add(&page->lru, &cc->freepages);
1647 /* possible outcome of isolate_migratepages */
1649 ISOLATE_ABORT, /* Abort compaction now */
1650 ISOLATE_NONE, /* No pages isolated, continue scanning */
1651 ISOLATE_SUCCESS, /* Pages isolated, migrate */
1652 } isolate_migrate_t;
1655 * Allow userspace to control policy on scanning the unevictable LRU for
1656 * compactable pages.
1658 #ifdef CONFIG_PREEMPT_RT
1659 int sysctl_compact_unevictable_allowed __read_mostly = 0;
1661 int sysctl_compact_unevictable_allowed __read_mostly = 1;
1665 update_fast_start_pfn(struct compact_control *cc, unsigned long pfn)
1667 if (cc->fast_start_pfn == ULONG_MAX)
1670 if (!cc->fast_start_pfn)
1671 cc->fast_start_pfn = pfn;
1673 cc->fast_start_pfn = min(cc->fast_start_pfn, pfn);
1676 static inline unsigned long
1677 reinit_migrate_pfn(struct compact_control *cc)
1679 if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX)
1680 return cc->migrate_pfn;
1682 cc->migrate_pfn = cc->fast_start_pfn;
1683 cc->fast_start_pfn = ULONG_MAX;
1685 return cc->migrate_pfn;
1689 * Briefly search the free lists for a migration source that already has
1690 * some free pages to reduce the number of pages that need migration
1691 * before a pageblock is free.
1693 static unsigned long fast_find_migrateblock(struct compact_control *cc)
1695 unsigned int limit = freelist_scan_limit(cc);
1696 unsigned int nr_scanned = 0;
1697 unsigned long distance;
1698 unsigned long pfn = cc->migrate_pfn;
1699 unsigned long high_pfn;
1701 bool found_block = false;
1703 /* Skip hints are relied on to avoid repeats on the fast search */
1704 if (cc->ignore_skip_hint)
1708 * If the migrate_pfn is not at the start of a zone or the start
1709 * of a pageblock then assume this is a continuation of a previous
1710 * scan restarted due to COMPACT_CLUSTER_MAX.
1712 if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn))
1716 * For smaller orders, just linearly scan as the number of pages
1717 * to migrate should be relatively small and does not necessarily
1718 * justify freeing up a large block for a small allocation.
1720 if (cc->order <= PAGE_ALLOC_COSTLY_ORDER)
1724 * Only allow kcompactd and direct requests for movable pages to
1725 * quickly clear out a MOVABLE pageblock for allocation. This
1726 * reduces the risk that a large movable pageblock is freed for
1727 * an unmovable/reclaimable small allocation.
1729 if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE)
1733 * When starting the migration scanner, pick any pageblock within the
1734 * first half of the search space. Otherwise try and pick a pageblock
1735 * within the first eighth to reduce the chances that a migration
1736 * target later becomes a source.
1738 distance = (cc->free_pfn - cc->migrate_pfn) >> 1;
1739 if (cc->migrate_pfn != cc->zone->zone_start_pfn)
1741 high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance);
1743 for (order = cc->order - 1;
1744 order >= PAGE_ALLOC_COSTLY_ORDER && !found_block && nr_scanned < limit;
1746 struct free_area *area = &cc->zone->free_area[order];
1747 struct list_head *freelist;
1748 unsigned long flags;
1749 struct page *freepage;
1754 spin_lock_irqsave(&cc->zone->lock, flags);
1755 freelist = &area->free_list[MIGRATE_MOVABLE];
1756 list_for_each_entry(freepage, freelist, lru) {
1757 unsigned long free_pfn;
1759 if (nr_scanned++ >= limit) {
1760 move_freelist_tail(freelist, freepage);
1764 free_pfn = page_to_pfn(freepage);
1765 if (free_pfn < high_pfn) {
1767 * Avoid if skipped recently. Ideally it would
1768 * move to the tail but even safe iteration of
1769 * the list assumes an entry is deleted, not
1772 if (get_pageblock_skip(freepage))
1775 /* Reorder to so a future search skips recent pages */
1776 move_freelist_tail(freelist, freepage);
1778 update_fast_start_pfn(cc, free_pfn);
1779 pfn = pageblock_start_pfn(free_pfn);
1780 cc->fast_search_fail = 0;
1782 set_pageblock_skip(freepage);
1786 spin_unlock_irqrestore(&cc->zone->lock, flags);
1789 cc->total_migrate_scanned += nr_scanned;
1792 * If fast scanning failed then use a cached entry for a page block
1793 * that had free pages as the basis for starting a linear scan.
1796 cc->fast_search_fail++;
1797 pfn = reinit_migrate_pfn(cc);
1803 * Isolate all pages that can be migrated from the first suitable block,
1804 * starting at the block pointed to by the migrate scanner pfn within
1807 static isolate_migrate_t isolate_migratepages(struct compact_control *cc)
1809 unsigned long block_start_pfn;
1810 unsigned long block_end_pfn;
1811 unsigned long low_pfn;
1813 const isolate_mode_t isolate_mode =
1814 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1815 (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1816 bool fast_find_block;
1819 * Start at where we last stopped, or beginning of the zone as
1820 * initialized by compact_zone(). The first failure will use
1821 * the lowest PFN as the starting point for linear scanning.
1823 low_pfn = fast_find_migrateblock(cc);
1824 block_start_pfn = pageblock_start_pfn(low_pfn);
1825 if (block_start_pfn < cc->zone->zone_start_pfn)
1826 block_start_pfn = cc->zone->zone_start_pfn;
1829 * fast_find_migrateblock marks a pageblock skipped so to avoid
1830 * the isolation_suitable check below, check whether the fast
1831 * search was successful.
1833 fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail;
1835 /* Only scan within a pageblock boundary */
1836 block_end_pfn = pageblock_end_pfn(low_pfn);
1839 * Iterate over whole pageblocks until we find the first suitable.
1840 * Do not cross the free scanner.
1842 for (; block_end_pfn <= cc->free_pfn;
1843 fast_find_block = false,
1844 low_pfn = block_end_pfn,
1845 block_start_pfn = block_end_pfn,
1846 block_end_pfn += pageblock_nr_pages) {
1849 * This can potentially iterate a massively long zone with
1850 * many pageblocks unsuitable, so periodically check if we
1853 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)))
1856 page = pageblock_pfn_to_page(block_start_pfn,
1857 block_end_pfn, cc->zone);
1862 * If isolation recently failed, do not retry. Only check the
1863 * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock
1864 * to be visited multiple times. Assume skip was checked
1865 * before making it "skip" so other compaction instances do
1866 * not scan the same block.
1868 if (IS_ALIGNED(low_pfn, pageblock_nr_pages) &&
1869 !fast_find_block && !isolation_suitable(cc, page))
1873 * For async compaction, also only scan in MOVABLE blocks
1874 * without huge pages. Async compaction is optimistic to see
1875 * if the minimum amount of work satisfies the allocation.
1876 * The cached PFN is updated as it's possible that all
1877 * remaining blocks between source and target are unsuitable
1878 * and the compaction scanners fail to meet.
1880 if (!suitable_migration_source(cc, page)) {
1881 update_cached_migrate(cc, block_end_pfn);
1885 /* Perform the isolation */
1886 low_pfn = isolate_migratepages_block(cc, low_pfn,
1887 block_end_pfn, isolate_mode);
1890 return ISOLATE_ABORT;
1893 * Either we isolated something and proceed with migration. Or
1894 * we failed and compact_zone should decide if we should
1900 /* Record where migration scanner will be restarted. */
1901 cc->migrate_pfn = low_pfn;
1903 return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
1907 * order == -1 is expected when compacting via
1908 * /proc/sys/vm/compact_memory
1910 static inline bool is_via_compact_memory(int order)
1915 static bool kswapd_is_running(pg_data_t *pgdat)
1917 return pgdat->kswapd && (pgdat->kswapd->state == TASK_RUNNING);
1921 * A zone's fragmentation score is the external fragmentation wrt to the
1922 * COMPACTION_HPAGE_ORDER. It returns a value in the range [0, 100].
1924 static unsigned int fragmentation_score_zone(struct zone *zone)
1926 return extfrag_for_order(zone, COMPACTION_HPAGE_ORDER);
1930 * A weighted zone's fragmentation score is the external fragmentation
1931 * wrt to the COMPACTION_HPAGE_ORDER scaled by the zone's size. It
1932 * returns a value in the range [0, 100].
1934 * The scaling factor ensures that proactive compaction focuses on larger
1935 * zones like ZONE_NORMAL, rather than smaller, specialized zones like
1936 * ZONE_DMA32. For smaller zones, the score value remains close to zero,
1937 * and thus never exceeds the high threshold for proactive compaction.
1939 static unsigned int fragmentation_score_zone_weighted(struct zone *zone)
1941 unsigned long score;
1943 score = zone->present_pages * fragmentation_score_zone(zone);
1944 return div64_ul(score, zone->zone_pgdat->node_present_pages + 1);
1948 * The per-node proactive (background) compaction process is started by its
1949 * corresponding kcompactd thread when the node's fragmentation score
1950 * exceeds the high threshold. The compaction process remains active till
1951 * the node's score falls below the low threshold, or one of the back-off
1952 * conditions is met.
1954 static unsigned int fragmentation_score_node(pg_data_t *pgdat)
1956 unsigned int score = 0;
1959 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1962 zone = &pgdat->node_zones[zoneid];
1963 score += fragmentation_score_zone_weighted(zone);
1969 static unsigned int fragmentation_score_wmark(pg_data_t *pgdat, bool low)
1971 unsigned int wmark_low;
1974 * Cap the low watermak to avoid excessive compaction
1975 * activity in case a user sets the proactivess tunable
1976 * close to 100 (maximum).
1978 wmark_low = max(100U - sysctl_compaction_proactiveness, 5U);
1979 return low ? wmark_low : min(wmark_low + 10, 100U);
1982 static bool should_proactive_compact_node(pg_data_t *pgdat)
1986 if (!sysctl_compaction_proactiveness || kswapd_is_running(pgdat))
1989 wmark_high = fragmentation_score_wmark(pgdat, false);
1990 return fragmentation_score_node(pgdat) > wmark_high;
1993 static enum compact_result __compact_finished(struct compact_control *cc)
1996 const int migratetype = cc->migratetype;
1999 /* Compaction run completes if the migrate and free scanner meet */
2000 if (compact_scanners_met(cc)) {
2001 /* Let the next compaction start anew. */
2002 reset_cached_positions(cc->zone);
2005 * Mark that the PG_migrate_skip information should be cleared
2006 * by kswapd when it goes to sleep. kcompactd does not set the
2007 * flag itself as the decision to be clear should be directly
2008 * based on an allocation request.
2010 if (cc->direct_compaction)
2011 cc->zone->compact_blockskip_flush = true;
2014 return COMPACT_COMPLETE;
2016 return COMPACT_PARTIAL_SKIPPED;
2019 if (cc->proactive_compaction) {
2020 int score, wmark_low;
2023 pgdat = cc->zone->zone_pgdat;
2024 if (kswapd_is_running(pgdat))
2025 return COMPACT_PARTIAL_SKIPPED;
2027 score = fragmentation_score_zone(cc->zone);
2028 wmark_low = fragmentation_score_wmark(pgdat, true);
2030 if (score > wmark_low)
2031 ret = COMPACT_CONTINUE;
2033 ret = COMPACT_SUCCESS;
2038 if (is_via_compact_memory(cc->order))
2039 return COMPACT_CONTINUE;
2042 * Always finish scanning a pageblock to reduce the possibility of
2043 * fallbacks in the future. This is particularly important when
2044 * migration source is unmovable/reclaimable but it's not worth
2047 if (!IS_ALIGNED(cc->migrate_pfn, pageblock_nr_pages))
2048 return COMPACT_CONTINUE;
2050 /* Direct compactor: Is a suitable page free? */
2051 ret = COMPACT_NO_SUITABLE_PAGE;
2052 for (order = cc->order; order < MAX_ORDER; order++) {
2053 struct free_area *area = &cc->zone->free_area[order];
2056 /* Job done if page is free of the right migratetype */
2057 if (!free_area_empty(area, migratetype))
2058 return COMPACT_SUCCESS;
2061 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
2062 if (migratetype == MIGRATE_MOVABLE &&
2063 !free_area_empty(area, MIGRATE_CMA))
2064 return COMPACT_SUCCESS;
2067 * Job done if allocation would steal freepages from
2068 * other migratetype buddy lists.
2070 if (find_suitable_fallback(area, order, migratetype,
2071 true, &can_steal) != -1) {
2073 /* movable pages are OK in any pageblock */
2074 if (migratetype == MIGRATE_MOVABLE)
2075 return COMPACT_SUCCESS;
2078 * We are stealing for a non-movable allocation. Make
2079 * sure we finish compacting the current pageblock
2080 * first so it is as free as possible and we won't
2081 * have to steal another one soon. This only applies
2082 * to sync compaction, as async compaction operates
2083 * on pageblocks of the same migratetype.
2085 if (cc->mode == MIGRATE_ASYNC ||
2086 IS_ALIGNED(cc->migrate_pfn,
2087 pageblock_nr_pages)) {
2088 return COMPACT_SUCCESS;
2091 ret = COMPACT_CONTINUE;
2097 if (cc->contended || fatal_signal_pending(current))
2098 ret = COMPACT_CONTENDED;
2103 static enum compact_result compact_finished(struct compact_control *cc)
2107 ret = __compact_finished(cc);
2108 trace_mm_compaction_finished(cc->zone, cc->order, ret);
2109 if (ret == COMPACT_NO_SUITABLE_PAGE)
2110 ret = COMPACT_CONTINUE;
2115 static enum compact_result __compaction_suitable(struct zone *zone, int order,
2116 unsigned int alloc_flags,
2117 int highest_zoneidx,
2118 unsigned long wmark_target)
2120 unsigned long watermark;
2122 if (is_via_compact_memory(order))
2123 return COMPACT_CONTINUE;
2125 watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
2127 * If watermarks for high-order allocation are already met, there
2128 * should be no need for compaction at all.
2130 if (zone_watermark_ok(zone, order, watermark, highest_zoneidx,
2132 return COMPACT_SUCCESS;
2135 * Watermarks for order-0 must be met for compaction to be able to
2136 * isolate free pages for migration targets. This means that the
2137 * watermark and alloc_flags have to match, or be more pessimistic than
2138 * the check in __isolate_free_page(). We don't use the direct
2139 * compactor's alloc_flags, as they are not relevant for freepage
2140 * isolation. We however do use the direct compactor's highest_zoneidx
2141 * to skip over zones where lowmem reserves would prevent allocation
2142 * even if compaction succeeds.
2143 * For costly orders, we require low watermark instead of min for
2144 * compaction to proceed to increase its chances.
2145 * ALLOC_CMA is used, as pages in CMA pageblocks are considered
2146 * suitable migration targets
2148 watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
2149 low_wmark_pages(zone) : min_wmark_pages(zone);
2150 watermark += compact_gap(order);
2151 if (!__zone_watermark_ok(zone, 0, watermark, highest_zoneidx,
2152 ALLOC_CMA, wmark_target))
2153 return COMPACT_SKIPPED;
2155 return COMPACT_CONTINUE;
2159 * compaction_suitable: Is this suitable to run compaction on this zone now?
2161 * COMPACT_SKIPPED - If there are too few free pages for compaction
2162 * COMPACT_SUCCESS - If the allocation would succeed without compaction
2163 * COMPACT_CONTINUE - If compaction should run now
2165 enum compact_result compaction_suitable(struct zone *zone, int order,
2166 unsigned int alloc_flags,
2167 int highest_zoneidx)
2169 enum compact_result ret;
2172 ret = __compaction_suitable(zone, order, alloc_flags, highest_zoneidx,
2173 zone_page_state(zone, NR_FREE_PAGES));
2175 * fragmentation index determines if allocation failures are due to
2176 * low memory or external fragmentation
2178 * index of -1000 would imply allocations might succeed depending on
2179 * watermarks, but we already failed the high-order watermark check
2180 * index towards 0 implies failure is due to lack of memory
2181 * index towards 1000 implies failure is due to fragmentation
2183 * Only compact if a failure would be due to fragmentation. Also
2184 * ignore fragindex for non-costly orders where the alternative to
2185 * a successful reclaim/compaction is OOM. Fragindex and the
2186 * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
2187 * excessive compaction for costly orders, but it should not be at the
2188 * expense of system stability.
2190 if (ret == COMPACT_CONTINUE && (order > PAGE_ALLOC_COSTLY_ORDER)) {
2191 fragindex = fragmentation_index(zone, order);
2192 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
2193 ret = COMPACT_NOT_SUITABLE_ZONE;
2196 trace_mm_compaction_suitable(zone, order, ret);
2197 if (ret == COMPACT_NOT_SUITABLE_ZONE)
2198 ret = COMPACT_SKIPPED;
2203 bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
2210 * Make sure at least one zone would pass __compaction_suitable if we continue
2211 * retrying the reclaim.
2213 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2214 ac->highest_zoneidx, ac->nodemask) {
2215 unsigned long available;
2216 enum compact_result compact_result;
2219 * Do not consider all the reclaimable memory because we do not
2220 * want to trash just for a single high order allocation which
2221 * is even not guaranteed to appear even if __compaction_suitable
2222 * is happy about the watermark check.
2224 available = zone_reclaimable_pages(zone) / order;
2225 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
2226 compact_result = __compaction_suitable(zone, order, alloc_flags,
2227 ac->highest_zoneidx, available);
2228 if (compact_result != COMPACT_SKIPPED)
2235 static enum compact_result
2236 compact_zone(struct compact_control *cc, struct capture_control *capc)
2238 enum compact_result ret;
2239 unsigned long start_pfn = cc->zone->zone_start_pfn;
2240 unsigned long end_pfn = zone_end_pfn(cc->zone);
2241 unsigned long last_migrated_pfn;
2242 const bool sync = cc->mode != MIGRATE_ASYNC;
2246 * These counters track activities during zone compaction. Initialize
2247 * them before compacting a new zone.
2249 cc->total_migrate_scanned = 0;
2250 cc->total_free_scanned = 0;
2251 cc->nr_migratepages = 0;
2252 cc->nr_freepages = 0;
2253 INIT_LIST_HEAD(&cc->freepages);
2254 INIT_LIST_HEAD(&cc->migratepages);
2256 cc->migratetype = gfp_migratetype(cc->gfp_mask);
2257 ret = compaction_suitable(cc->zone, cc->order, cc->alloc_flags,
2258 cc->highest_zoneidx);
2259 /* Compaction is likely to fail */
2260 if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED)
2263 /* huh, compaction_suitable is returning something unexpected */
2264 VM_BUG_ON(ret != COMPACT_CONTINUE);
2267 * Clear pageblock skip if there were failures recently and compaction
2268 * is about to be retried after being deferred.
2270 if (compaction_restarting(cc->zone, cc->order))
2271 __reset_isolation_suitable(cc->zone);
2274 * Setup to move all movable pages to the end of the zone. Used cached
2275 * information on where the scanners should start (unless we explicitly
2276 * want to compact the whole zone), but check that it is initialised
2277 * by ensuring the values are within zone boundaries.
2279 cc->fast_start_pfn = 0;
2280 if (cc->whole_zone) {
2281 cc->migrate_pfn = start_pfn;
2282 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2284 cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync];
2285 cc->free_pfn = cc->zone->compact_cached_free_pfn;
2286 if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
2287 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2288 cc->zone->compact_cached_free_pfn = cc->free_pfn;
2290 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
2291 cc->migrate_pfn = start_pfn;
2292 cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
2293 cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
2296 if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn)
2297 cc->whole_zone = true;
2300 last_migrated_pfn = 0;
2303 * Migrate has separate cached PFNs for ASYNC and SYNC* migration on
2304 * the basis that some migrations will fail in ASYNC mode. However,
2305 * if the cached PFNs match and pageblocks are skipped due to having
2306 * no isolation candidates, then the sync state does not matter.
2307 * Until a pageblock with isolation candidates is found, keep the
2308 * cached PFNs in sync to avoid revisiting the same blocks.
2310 update_cached = !sync &&
2311 cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1];
2313 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
2314 cc->free_pfn, end_pfn, sync);
2316 migrate_prep_local();
2318 while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) {
2320 unsigned long iteration_start_pfn = cc->migrate_pfn;
2323 * Avoid multiple rescans which can happen if a page cannot be
2324 * isolated (dirty/writeback in async mode) or if the migrated
2325 * pages are being allocated before the pageblock is cleared.
2326 * The first rescan will capture the entire pageblock for
2327 * migration. If it fails, it'll be marked skip and scanning
2328 * will proceed as normal.
2331 if (pageblock_start_pfn(last_migrated_pfn) ==
2332 pageblock_start_pfn(iteration_start_pfn)) {
2336 switch (isolate_migratepages(cc)) {
2338 ret = COMPACT_CONTENDED;
2339 putback_movable_pages(&cc->migratepages);
2340 cc->nr_migratepages = 0;
2343 if (update_cached) {
2344 cc->zone->compact_cached_migrate_pfn[1] =
2345 cc->zone->compact_cached_migrate_pfn[0];
2349 * We haven't isolated and migrated anything, but
2350 * there might still be unflushed migrations from
2351 * previous cc->order aligned block.
2354 case ISOLATE_SUCCESS:
2355 update_cached = false;
2356 last_migrated_pfn = iteration_start_pfn;
2359 err = migrate_pages(&cc->migratepages, compaction_alloc,
2360 compaction_free, (unsigned long)cc, cc->mode,
2363 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
2366 /* All pages were either migrated or will be released */
2367 cc->nr_migratepages = 0;
2369 putback_movable_pages(&cc->migratepages);
2371 * migrate_pages() may return -ENOMEM when scanners meet
2372 * and we want compact_finished() to detect it
2374 if (err == -ENOMEM && !compact_scanners_met(cc)) {
2375 ret = COMPACT_CONTENDED;
2379 * We failed to migrate at least one page in the current
2380 * order-aligned block, so skip the rest of it.
2382 if (cc->direct_compaction &&
2383 (cc->mode == MIGRATE_ASYNC)) {
2384 cc->migrate_pfn = block_end_pfn(
2385 cc->migrate_pfn - 1, cc->order);
2386 /* Draining pcplists is useless in this case */
2387 last_migrated_pfn = 0;
2393 * Has the migration scanner moved away from the previous
2394 * cc->order aligned block where we migrated from? If yes,
2395 * flush the pages that were freed, so that they can merge and
2396 * compact_finished() can detect immediately if allocation
2399 if (cc->order > 0 && last_migrated_pfn) {
2400 unsigned long current_block_start =
2401 block_start_pfn(cc->migrate_pfn, cc->order);
2403 if (last_migrated_pfn < current_block_start) {
2404 lru_add_drain_cpu_zone(cc->zone);
2405 /* No more flushing until we migrate again */
2406 last_migrated_pfn = 0;
2410 /* Stop if a page has been captured */
2411 if (capc && capc->page) {
2412 ret = COMPACT_SUCCESS;
2419 * Release free pages and update where the free scanner should restart,
2420 * so we don't leave any returned pages behind in the next attempt.
2422 if (cc->nr_freepages > 0) {
2423 unsigned long free_pfn = release_freepages(&cc->freepages);
2425 cc->nr_freepages = 0;
2426 VM_BUG_ON(free_pfn == 0);
2427 /* The cached pfn is always the first in a pageblock */
2428 free_pfn = pageblock_start_pfn(free_pfn);
2430 * Only go back, not forward. The cached pfn might have been
2431 * already reset to zone end in compact_finished()
2433 if (free_pfn > cc->zone->compact_cached_free_pfn)
2434 cc->zone->compact_cached_free_pfn = free_pfn;
2437 count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned);
2438 count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned);
2440 trace_mm_compaction_end(start_pfn, cc->migrate_pfn,
2441 cc->free_pfn, end_pfn, sync, ret);
2446 static enum compact_result compact_zone_order(struct zone *zone, int order,
2447 gfp_t gfp_mask, enum compact_priority prio,
2448 unsigned int alloc_flags, int highest_zoneidx,
2449 struct page **capture)
2451 enum compact_result ret;
2452 struct compact_control cc = {
2454 .search_order = order,
2455 .gfp_mask = gfp_mask,
2457 .mode = (prio == COMPACT_PRIO_ASYNC) ?
2458 MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT,
2459 .alloc_flags = alloc_flags,
2460 .highest_zoneidx = highest_zoneidx,
2461 .direct_compaction = true,
2462 .whole_zone = (prio == MIN_COMPACT_PRIORITY),
2463 .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
2464 .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
2466 struct capture_control capc = {
2472 * Make sure the structs are really initialized before we expose the
2473 * capture control, in case we are interrupted and the interrupt handler
2477 WRITE_ONCE(current->capture_control, &capc);
2479 ret = compact_zone(&cc, &capc);
2481 VM_BUG_ON(!list_empty(&cc.freepages));
2482 VM_BUG_ON(!list_empty(&cc.migratepages));
2485 * Make sure we hide capture control first before we read the captured
2486 * page pointer, otherwise an interrupt could free and capture a page
2487 * and we would leak it.
2489 WRITE_ONCE(current->capture_control, NULL);
2490 *capture = READ_ONCE(capc.page);
2495 int sysctl_extfrag_threshold = 500;
2498 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
2499 * @gfp_mask: The GFP mask of the current allocation
2500 * @order: The order of the current allocation
2501 * @alloc_flags: The allocation flags of the current allocation
2502 * @ac: The context of current allocation
2503 * @prio: Determines how hard direct compaction should try to succeed
2504 * @capture: Pointer to free page created by compaction will be stored here
2506 * This is the main entry point for direct page compaction.
2508 enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
2509 unsigned int alloc_flags, const struct alloc_context *ac,
2510 enum compact_priority prio, struct page **capture)
2512 int may_perform_io = gfp_mask & __GFP_IO;
2515 enum compact_result rc = COMPACT_SKIPPED;
2518 * Check if the GFP flags allow compaction - GFP_NOIO is really
2519 * tricky context because the migration might require IO
2521 if (!may_perform_io)
2522 return COMPACT_SKIPPED;
2524 trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
2526 /* Compact each zone in the list */
2527 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2528 ac->highest_zoneidx, ac->nodemask) {
2529 enum compact_result status;
2531 if (prio > MIN_COMPACT_PRIORITY
2532 && compaction_deferred(zone, order)) {
2533 rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
2537 status = compact_zone_order(zone, order, gfp_mask, prio,
2538 alloc_flags, ac->highest_zoneidx, capture);
2539 rc = max(status, rc);
2541 /* The allocation should succeed, stop compacting */
2542 if (status == COMPACT_SUCCESS) {
2544 * We think the allocation will succeed in this zone,
2545 * but it is not certain, hence the false. The caller
2546 * will repeat this with true if allocation indeed
2547 * succeeds in this zone.
2549 compaction_defer_reset(zone, order, false);
2554 if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
2555 status == COMPACT_PARTIAL_SKIPPED))
2557 * We think that allocation won't succeed in this zone
2558 * so we defer compaction there. If it ends up
2559 * succeeding after all, it will be reset.
2561 defer_compaction(zone, order);
2564 * We might have stopped compacting due to need_resched() in
2565 * async compaction, or due to a fatal signal detected. In that
2566 * case do not try further zones
2568 if ((prio == COMPACT_PRIO_ASYNC && need_resched())
2569 || fatal_signal_pending(current))
2577 * Compact all zones within a node till each zone's fragmentation score
2578 * reaches within proactive compaction thresholds (as determined by the
2579 * proactiveness tunable).
2581 * It is possible that the function returns before reaching score targets
2582 * due to various back-off conditions, such as, contention on per-node or
2585 static void proactive_compact_node(pg_data_t *pgdat)
2589 struct compact_control cc = {
2591 .mode = MIGRATE_SYNC_LIGHT,
2592 .ignore_skip_hint = true,
2594 .gfp_mask = GFP_KERNEL,
2595 .proactive_compaction = true,
2598 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2599 zone = &pgdat->node_zones[zoneid];
2600 if (!populated_zone(zone))
2605 compact_zone(&cc, NULL);
2607 VM_BUG_ON(!list_empty(&cc.freepages));
2608 VM_BUG_ON(!list_empty(&cc.migratepages));
2612 /* Compact all zones within a node */
2613 static void compact_node(int nid)
2615 pg_data_t *pgdat = NODE_DATA(nid);
2618 struct compact_control cc = {
2620 .mode = MIGRATE_SYNC,
2621 .ignore_skip_hint = true,
2623 .gfp_mask = GFP_KERNEL,
2627 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2629 zone = &pgdat->node_zones[zoneid];
2630 if (!populated_zone(zone))
2635 compact_zone(&cc, NULL);
2637 VM_BUG_ON(!list_empty(&cc.freepages));
2638 VM_BUG_ON(!list_empty(&cc.migratepages));
2642 /* Compact all nodes in the system */
2643 static void compact_nodes(void)
2647 /* Flush pending updates to the LRU lists */
2648 lru_add_drain_all();
2650 for_each_online_node(nid)
2654 /* The written value is actually unused, all memory is compacted */
2655 int sysctl_compact_memory;
2658 * Tunable for proactive compaction. It determines how
2659 * aggressively the kernel should compact memory in the
2660 * background. It takes values in the range [0, 100].
2662 unsigned int __read_mostly sysctl_compaction_proactiveness = 20;
2665 * This is the entry point for compacting all nodes via
2666 * /proc/sys/vm/compact_memory
2668 int sysctl_compaction_handler(struct ctl_table *table, int write,
2669 void *buffer, size_t *length, loff_t *ppos)
2677 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
2678 static ssize_t sysfs_compact_node(struct device *dev,
2679 struct device_attribute *attr,
2680 const char *buf, size_t count)
2684 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
2685 /* Flush pending updates to the LRU lists */
2686 lru_add_drain_all();
2693 static DEVICE_ATTR(compact, 0200, NULL, sysfs_compact_node);
2695 int compaction_register_node(struct node *node)
2697 return device_create_file(&node->dev, &dev_attr_compact);
2700 void compaction_unregister_node(struct node *node)
2702 return device_remove_file(&node->dev, &dev_attr_compact);
2704 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
2706 static inline bool kcompactd_work_requested(pg_data_t *pgdat)
2708 return pgdat->kcompactd_max_order > 0 || kthread_should_stop();
2711 static bool kcompactd_node_suitable(pg_data_t *pgdat)
2715 enum zone_type highest_zoneidx = pgdat->kcompactd_highest_zoneidx;
2717 for (zoneid = 0; zoneid <= highest_zoneidx; zoneid++) {
2718 zone = &pgdat->node_zones[zoneid];
2720 if (!populated_zone(zone))
2723 if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0,
2724 highest_zoneidx) == COMPACT_CONTINUE)
2731 static void kcompactd_do_work(pg_data_t *pgdat)
2734 * With no special task, compact all zones so that a page of requested
2735 * order is allocatable.
2739 struct compact_control cc = {
2740 .order = pgdat->kcompactd_max_order,
2741 .search_order = pgdat->kcompactd_max_order,
2742 .highest_zoneidx = pgdat->kcompactd_highest_zoneidx,
2743 .mode = MIGRATE_SYNC_LIGHT,
2744 .ignore_skip_hint = false,
2745 .gfp_mask = GFP_KERNEL,
2747 trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
2748 cc.highest_zoneidx);
2749 count_compact_event(KCOMPACTD_WAKE);
2751 for (zoneid = 0; zoneid <= cc.highest_zoneidx; zoneid++) {
2754 zone = &pgdat->node_zones[zoneid];
2755 if (!populated_zone(zone))
2758 if (compaction_deferred(zone, cc.order))
2761 if (compaction_suitable(zone, cc.order, 0, zoneid) !=
2765 if (kthread_should_stop())
2769 status = compact_zone(&cc, NULL);
2771 if (status == COMPACT_SUCCESS) {
2772 compaction_defer_reset(zone, cc.order, false);
2773 } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
2775 * Buddy pages may become stranded on pcps that could
2776 * otherwise coalesce on the zone's free area for
2777 * order >= cc.order. This is ratelimited by the
2778 * upcoming deferral.
2780 drain_all_pages(zone);
2783 * We use sync migration mode here, so we defer like
2784 * sync direct compaction does.
2786 defer_compaction(zone, cc.order);
2789 count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
2790 cc.total_migrate_scanned);
2791 count_compact_events(KCOMPACTD_FREE_SCANNED,
2792 cc.total_free_scanned);
2794 VM_BUG_ON(!list_empty(&cc.freepages));
2795 VM_BUG_ON(!list_empty(&cc.migratepages));
2799 * Regardless of success, we are done until woken up next. But remember
2800 * the requested order/highest_zoneidx in case it was higher/tighter
2801 * than our current ones
2803 if (pgdat->kcompactd_max_order <= cc.order)
2804 pgdat->kcompactd_max_order = 0;
2805 if (pgdat->kcompactd_highest_zoneidx >= cc.highest_zoneidx)
2806 pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
2809 void wakeup_kcompactd(pg_data_t *pgdat, int order, int highest_zoneidx)
2814 if (pgdat->kcompactd_max_order < order)
2815 pgdat->kcompactd_max_order = order;
2817 if (pgdat->kcompactd_highest_zoneidx > highest_zoneidx)
2818 pgdat->kcompactd_highest_zoneidx = highest_zoneidx;
2821 * Pairs with implicit barrier in wait_event_freezable()
2822 * such that wakeups are not missed.
2824 if (!wq_has_sleeper(&pgdat->kcompactd_wait))
2827 if (!kcompactd_node_suitable(pgdat))
2830 trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
2832 wake_up_interruptible(&pgdat->kcompactd_wait);
2836 * The background compaction daemon, started as a kernel thread
2837 * from the init process.
2839 static int kcompactd(void *p)
2841 pg_data_t *pgdat = (pg_data_t*)p;
2842 struct task_struct *tsk = current;
2843 unsigned int proactive_defer = 0;
2845 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2847 if (!cpumask_empty(cpumask))
2848 set_cpus_allowed_ptr(tsk, cpumask);
2852 pgdat->kcompactd_max_order = 0;
2853 pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
2855 while (!kthread_should_stop()) {
2856 unsigned long pflags;
2858 trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
2859 if (wait_event_freezable_timeout(pgdat->kcompactd_wait,
2860 kcompactd_work_requested(pgdat),
2861 msecs_to_jiffies(HPAGE_FRAG_CHECK_INTERVAL_MSEC))) {
2863 psi_memstall_enter(&pflags);
2864 kcompactd_do_work(pgdat);
2865 psi_memstall_leave(&pflags);
2869 /* kcompactd wait timeout */
2870 if (should_proactive_compact_node(pgdat)) {
2871 unsigned int prev_score, score;
2873 if (proactive_defer) {
2877 prev_score = fragmentation_score_node(pgdat);
2878 proactive_compact_node(pgdat);
2879 score = fragmentation_score_node(pgdat);
2881 * Defer proactive compaction if the fragmentation
2882 * score did not go down i.e. no progress made.
2884 proactive_defer = score < prev_score ?
2885 0 : 1 << COMPACT_MAX_DEFER_SHIFT;
2893 * This kcompactd start function will be called by init and node-hot-add.
2894 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
2896 int kcompactd_run(int nid)
2898 pg_data_t *pgdat = NODE_DATA(nid);
2901 if (pgdat->kcompactd)
2904 pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
2905 if (IS_ERR(pgdat->kcompactd)) {
2906 pr_err("Failed to start kcompactd on node %d\n", nid);
2907 ret = PTR_ERR(pgdat->kcompactd);
2908 pgdat->kcompactd = NULL;
2914 * Called by memory hotplug when all memory in a node is offlined. Caller must
2915 * hold mem_hotplug_begin/end().
2917 void kcompactd_stop(int nid)
2919 struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
2922 kthread_stop(kcompactd);
2923 NODE_DATA(nid)->kcompactd = NULL;
2928 * It's optimal to keep kcompactd on the same CPUs as their memory, but
2929 * not required for correctness. So if the last cpu in a node goes
2930 * away, we get changed to run anywhere: as the first one comes back,
2931 * restore their cpu bindings.
2933 static int kcompactd_cpu_online(unsigned int cpu)
2937 for_each_node_state(nid, N_MEMORY) {
2938 pg_data_t *pgdat = NODE_DATA(nid);
2939 const struct cpumask *mask;
2941 mask = cpumask_of_node(pgdat->node_id);
2943 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2944 /* One of our CPUs online: restore mask */
2945 set_cpus_allowed_ptr(pgdat->kcompactd, mask);
2950 static int __init kcompactd_init(void)
2955 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
2956 "mm/compaction:online",
2957 kcompactd_cpu_online, NULL);
2959 pr_err("kcompactd: failed to register hotplug callbacks.\n");
2963 for_each_node_state(nid, N_MEMORY)
2967 subsys_initcall(kcompactd_init)
2969 #endif /* CONFIG_COMPACTION */