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 (check_source && PageLRU(page)) {
310 clear_pageblock_skip(page);
314 if (check_target && PageBuddy(page)) {
315 clear_pageblock_skip(page);
319 page += (1 << PAGE_ALLOC_COSTLY_ORDER);
320 pfn += (1 << PAGE_ALLOC_COSTLY_ORDER);
321 } while (page <= end_page);
327 * This function is called to clear all cached information on pageblocks that
328 * should be skipped for page isolation when the migrate and free page scanner
331 static void __reset_isolation_suitable(struct zone *zone)
333 unsigned long migrate_pfn = zone->zone_start_pfn;
334 unsigned long free_pfn = zone_end_pfn(zone) - 1;
335 unsigned long reset_migrate = free_pfn;
336 unsigned long reset_free = migrate_pfn;
337 bool source_set = false;
338 bool free_set = false;
340 if (!zone->compact_blockskip_flush)
343 zone->compact_blockskip_flush = false;
346 * Walk the zone and update pageblock skip information. Source looks
347 * for PageLRU while target looks for PageBuddy. When the scanner
348 * is found, both PageBuddy and PageLRU are checked as the pageblock
349 * is suitable as both source and target.
351 for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages,
352 free_pfn -= pageblock_nr_pages) {
355 /* Update the migrate PFN */
356 if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) &&
357 migrate_pfn < reset_migrate) {
359 reset_migrate = migrate_pfn;
360 zone->compact_init_migrate_pfn = reset_migrate;
361 zone->compact_cached_migrate_pfn[0] = reset_migrate;
362 zone->compact_cached_migrate_pfn[1] = reset_migrate;
365 /* Update the free PFN */
366 if (__reset_isolation_pfn(zone, free_pfn, free_set, true) &&
367 free_pfn > reset_free) {
369 reset_free = free_pfn;
370 zone->compact_init_free_pfn = reset_free;
371 zone->compact_cached_free_pfn = reset_free;
375 /* Leave no distance if no suitable block was reset */
376 if (reset_migrate >= reset_free) {
377 zone->compact_cached_migrate_pfn[0] = migrate_pfn;
378 zone->compact_cached_migrate_pfn[1] = migrate_pfn;
379 zone->compact_cached_free_pfn = free_pfn;
383 void reset_isolation_suitable(pg_data_t *pgdat)
387 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
388 struct zone *zone = &pgdat->node_zones[zoneid];
389 if (!populated_zone(zone))
392 /* Only flush if a full compaction finished recently */
393 if (zone->compact_blockskip_flush)
394 __reset_isolation_suitable(zone);
399 * Sets the pageblock skip bit if it was clear. Note that this is a hint as
400 * locks are not required for read/writers. Returns true if it was already set.
402 static bool test_and_set_skip(struct compact_control *cc, struct page *page,
407 /* Do no update if skip hint is being ignored */
408 if (cc->ignore_skip_hint)
411 if (!IS_ALIGNED(pfn, pageblock_nr_pages))
414 skip = get_pageblock_skip(page);
415 if (!skip && !cc->no_set_skip_hint)
416 set_pageblock_skip(page);
421 static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
423 struct zone *zone = cc->zone;
425 pfn = pageblock_end_pfn(pfn);
427 /* Set for isolation rather than compaction */
428 if (cc->no_set_skip_hint)
431 if (pfn > zone->compact_cached_migrate_pfn[0])
432 zone->compact_cached_migrate_pfn[0] = pfn;
433 if (cc->mode != MIGRATE_ASYNC &&
434 pfn > zone->compact_cached_migrate_pfn[1])
435 zone->compact_cached_migrate_pfn[1] = pfn;
439 * If no pages were isolated then mark this pageblock to be skipped in the
440 * future. The information is later cleared by __reset_isolation_suitable().
442 static void update_pageblock_skip(struct compact_control *cc,
443 struct page *page, unsigned long pfn)
445 struct zone *zone = cc->zone;
447 if (cc->no_set_skip_hint)
453 set_pageblock_skip(page);
455 /* Update where async and sync compaction should restart */
456 if (pfn < zone->compact_cached_free_pfn)
457 zone->compact_cached_free_pfn = pfn;
460 static inline bool isolation_suitable(struct compact_control *cc,
466 static inline bool pageblock_skip_persistent(struct page *page)
471 static inline void update_pageblock_skip(struct compact_control *cc,
472 struct page *page, unsigned long pfn)
476 static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
480 static bool test_and_set_skip(struct compact_control *cc, struct page *page,
485 #endif /* CONFIG_COMPACTION */
488 * Compaction requires the taking of some coarse locks that are potentially
489 * very heavily contended. For async compaction, trylock and record if the
490 * lock is contended. The lock will still be acquired but compaction will
491 * abort when the current block is finished regardless of success rate.
492 * Sync compaction acquires the lock.
494 * Always returns true which makes it easier to track lock state in callers.
496 static bool compact_lock_irqsave(spinlock_t *lock, unsigned long *flags,
497 struct compact_control *cc)
500 /* Track if the lock is contended in async mode */
501 if (cc->mode == MIGRATE_ASYNC && !cc->contended) {
502 if (spin_trylock_irqsave(lock, *flags))
505 cc->contended = true;
508 spin_lock_irqsave(lock, *flags);
513 * Compaction requires the taking of some coarse locks that are potentially
514 * very heavily contended. The lock should be periodically unlocked to avoid
515 * having disabled IRQs for a long time, even when there is nobody waiting on
516 * the lock. It might also be that allowing the IRQs will result in
517 * need_resched() becoming true. If scheduling is needed, async compaction
518 * aborts. Sync compaction schedules.
519 * Either compaction type will also abort if a fatal signal is pending.
520 * In either case if the lock was locked, it is dropped and not regained.
522 * Returns true if compaction should abort due to fatal signal pending, or
523 * async compaction due to need_resched()
524 * Returns false when compaction can continue (sync compaction might have
527 static bool compact_unlock_should_abort(spinlock_t *lock,
528 unsigned long flags, bool *locked, struct compact_control *cc)
531 spin_unlock_irqrestore(lock, flags);
535 if (fatal_signal_pending(current)) {
536 cc->contended = true;
546 * Isolate free pages onto a private freelist. If @strict is true, will abort
547 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
548 * (even though it may still end up isolating some pages).
550 static unsigned long isolate_freepages_block(struct compact_control *cc,
551 unsigned long *start_pfn,
552 unsigned long end_pfn,
553 struct list_head *freelist,
557 int nr_scanned = 0, total_isolated = 0;
559 unsigned long flags = 0;
561 unsigned long blockpfn = *start_pfn;
564 /* Strict mode is for isolation, speed is secondary */
568 cursor = pfn_to_page(blockpfn);
570 /* Isolate free pages. */
571 for (; blockpfn < end_pfn; blockpfn += stride, cursor += stride) {
573 struct page *page = cursor;
576 * Periodically drop the lock (if held) regardless of its
577 * contention, to give chance to IRQs. Abort if fatal signal
578 * pending or async compaction detects need_resched()
580 if (!(blockpfn % SWAP_CLUSTER_MAX)
581 && compact_unlock_should_abort(&cc->zone->lock, flags,
588 * For compound pages such as THP and hugetlbfs, we can save
589 * potentially a lot of iterations if we skip them at once.
590 * The check is racy, but we can consider only valid values
591 * and the only danger is skipping too much.
593 if (PageCompound(page)) {
594 const unsigned int order = compound_order(page);
596 if (likely(order < MAX_ORDER)) {
597 blockpfn += (1UL << order) - 1;
598 cursor += (1UL << order) - 1;
603 if (!PageBuddy(page))
607 * If we already hold the lock, we can skip some rechecking.
608 * Note that if we hold the lock now, checked_pageblock was
609 * already set in some previous iteration (or strict is true),
610 * so it is correct to skip the suitable migration target
614 locked = compact_lock_irqsave(&cc->zone->lock,
617 /* Recheck this is a buddy page under lock */
618 if (!PageBuddy(page))
622 /* Found a free page, will break it into order-0 pages */
623 order = buddy_order(page);
624 isolated = __isolate_free_page(page, order);
627 set_page_private(page, order);
629 total_isolated += isolated;
630 cc->nr_freepages += isolated;
631 list_add_tail(&page->lru, freelist);
633 if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
634 blockpfn += isolated;
637 /* Advance to the end of split page */
638 blockpfn += isolated - 1;
639 cursor += isolated - 1;
651 spin_unlock_irqrestore(&cc->zone->lock, flags);
654 * There is a tiny chance that we have read bogus compound_order(),
655 * so be careful to not go outside of the pageblock.
657 if (unlikely(blockpfn > end_pfn))
660 trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
661 nr_scanned, total_isolated);
663 /* Record how far we have got within the block */
664 *start_pfn = blockpfn;
667 * If strict isolation is requested by CMA then check that all the
668 * pages requested were isolated. If there were any failures, 0 is
669 * returned and CMA will fail.
671 if (strict && blockpfn < end_pfn)
674 cc->total_free_scanned += nr_scanned;
676 count_compact_events(COMPACTISOLATED, total_isolated);
677 return total_isolated;
681 * isolate_freepages_range() - isolate free pages.
682 * @cc: Compaction control structure.
683 * @start_pfn: The first PFN to start isolating.
684 * @end_pfn: The one-past-last PFN.
686 * Non-free pages, invalid PFNs, or zone boundaries within the
687 * [start_pfn, end_pfn) range are considered errors, cause function to
688 * undo its actions and return zero.
690 * Otherwise, function returns one-past-the-last PFN of isolated page
691 * (which may be greater then end_pfn if end fell in a middle of
695 isolate_freepages_range(struct compact_control *cc,
696 unsigned long start_pfn, unsigned long end_pfn)
698 unsigned long isolated, pfn, block_start_pfn, block_end_pfn;
702 block_start_pfn = pageblock_start_pfn(pfn);
703 if (block_start_pfn < cc->zone->zone_start_pfn)
704 block_start_pfn = cc->zone->zone_start_pfn;
705 block_end_pfn = pageblock_end_pfn(pfn);
707 for (; pfn < end_pfn; pfn += isolated,
708 block_start_pfn = block_end_pfn,
709 block_end_pfn += pageblock_nr_pages) {
710 /* Protect pfn from changing by isolate_freepages_block */
711 unsigned long isolate_start_pfn = pfn;
713 block_end_pfn = min(block_end_pfn, end_pfn);
716 * pfn could pass the block_end_pfn if isolated freepage
717 * is more than pageblock order. In this case, we adjust
718 * scanning range to right one.
720 if (pfn >= block_end_pfn) {
721 block_start_pfn = pageblock_start_pfn(pfn);
722 block_end_pfn = pageblock_end_pfn(pfn);
723 block_end_pfn = min(block_end_pfn, end_pfn);
726 if (!pageblock_pfn_to_page(block_start_pfn,
727 block_end_pfn, cc->zone))
730 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
731 block_end_pfn, &freelist, 0, true);
734 * In strict mode, isolate_freepages_block() returns 0 if
735 * there are any holes in the block (ie. invalid PFNs or
742 * If we managed to isolate pages, it is always (1 << n) *
743 * pageblock_nr_pages for some non-negative n. (Max order
744 * page may span two pageblocks).
748 /* __isolate_free_page() does not map the pages */
749 split_map_pages(&freelist);
752 /* Loop terminated early, cleanup. */
753 release_freepages(&freelist);
757 /* We don't use freelists for anything. */
761 /* Similar to reclaim, but different enough that they don't share logic */
762 static bool too_many_isolated(pg_data_t *pgdat)
764 unsigned long active, inactive, isolated;
766 inactive = node_page_state(pgdat, NR_INACTIVE_FILE) +
767 node_page_state(pgdat, NR_INACTIVE_ANON);
768 active = node_page_state(pgdat, NR_ACTIVE_FILE) +
769 node_page_state(pgdat, NR_ACTIVE_ANON);
770 isolated = node_page_state(pgdat, NR_ISOLATED_FILE) +
771 node_page_state(pgdat, NR_ISOLATED_ANON);
773 return isolated > (inactive + active) / 2;
777 * isolate_migratepages_block() - isolate all migrate-able pages within
779 * @cc: Compaction control structure.
780 * @low_pfn: The first PFN to isolate
781 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
782 * @isolate_mode: Isolation mode to be used.
784 * Isolate all pages that can be migrated from the range specified by
785 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
786 * Returns errno, like -EAGAIN or -EINTR in case e.g signal pending or congestion,
787 * -ENOMEM in case we could not allocate a page, or 0.
788 * cc->migrate_pfn will contain the next pfn to scan.
790 * The pages are isolated on cc->migratepages list (not required to be empty),
791 * and cc->nr_migratepages is updated accordingly.
794 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
795 unsigned long end_pfn, isolate_mode_t isolate_mode)
797 pg_data_t *pgdat = cc->zone->zone_pgdat;
798 unsigned long nr_scanned = 0, nr_isolated = 0;
799 struct lruvec *lruvec;
800 unsigned long flags = 0;
801 struct lruvec *locked = NULL;
802 struct page *page = NULL, *valid_page = NULL;
803 unsigned long start_pfn = low_pfn;
804 bool skip_on_failure = false;
805 unsigned long next_skip_pfn = 0;
806 bool skip_updated = false;
809 cc->migrate_pfn = low_pfn;
812 * Ensure that there are not too many pages isolated from the LRU
813 * list by either parallel reclaimers or compaction. If there are,
814 * delay for some time until fewer pages are isolated
816 while (unlikely(too_many_isolated(pgdat))) {
817 /* stop isolation if there are still pages not migrated */
818 if (cc->nr_migratepages)
821 /* async migration should just abort */
822 if (cc->mode == MIGRATE_ASYNC)
825 congestion_wait(BLK_RW_ASYNC, HZ/10);
827 if (fatal_signal_pending(current))
833 if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
834 skip_on_failure = true;
835 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
838 /* Time to isolate some pages for migration */
839 for (; low_pfn < end_pfn; low_pfn++) {
841 if (skip_on_failure && low_pfn >= next_skip_pfn) {
843 * We have isolated all migration candidates in the
844 * previous order-aligned block, and did not skip it due
845 * to failure. We should migrate the pages now and
846 * hopefully succeed compaction.
852 * We failed to isolate in the previous order-aligned
853 * block. Set the new boundary to the end of the
854 * current block. Note we can't simply increase
855 * next_skip_pfn by 1 << order, as low_pfn might have
856 * been incremented by a higher number due to skipping
857 * a compound or a high-order buddy page in the
858 * previous loop iteration.
860 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
864 * Periodically drop the lock (if held) regardless of its
865 * contention, to give chance to IRQs. Abort completely if
866 * a fatal signal is pending.
868 if (!(low_pfn % SWAP_CLUSTER_MAX)) {
870 unlock_page_lruvec_irqrestore(locked, flags);
874 if (fatal_signal_pending(current)) {
875 cc->contended = true;
886 page = pfn_to_page(low_pfn);
889 * Check if the pageblock has already been marked skipped.
890 * Only the aligned PFN is checked as the caller isolates
891 * COMPACT_CLUSTER_MAX at a time so the second call must
892 * not falsely conclude that the block should be skipped.
894 if (!valid_page && IS_ALIGNED(low_pfn, pageblock_nr_pages)) {
895 if (!cc->ignore_skip_hint && get_pageblock_skip(page)) {
903 if (PageHuge(page) && cc->alloc_contig) {
904 ret = isolate_or_dissolve_huge_page(page, &cc->migratepages);
907 * Fail isolation in case isolate_or_dissolve_huge_page()
908 * reports an error. In case of -ENOMEM, abort right away.
911 /* Do not report -EBUSY down the chain */
914 low_pfn += (1UL << compound_order(page)) - 1;
918 if (PageHuge(page)) {
920 * Hugepage was successfully isolated and placed
921 * on the cc->migratepages list.
923 low_pfn += compound_nr(page) - 1;
924 goto isolate_success_no_list;
928 * Ok, the hugepage was dissolved. Now these pages are
929 * Buddy and cannot be re-allocated because they are
930 * isolated. Fall-through as the check below handles
936 * Skip if free. We read page order here without zone lock
937 * which is generally unsafe, but the race window is small and
938 * the worst thing that can happen is that we skip some
939 * potential isolation targets.
941 if (PageBuddy(page)) {
942 unsigned long freepage_order = buddy_order_unsafe(page);
945 * Without lock, we cannot be sure that what we got is
946 * a valid page order. Consider only values in the
947 * valid order range to prevent low_pfn overflow.
949 if (freepage_order > 0 && freepage_order < MAX_ORDER)
950 low_pfn += (1UL << freepage_order) - 1;
955 * Regardless of being on LRU, compound pages such as THP and
956 * hugetlbfs are not to be compacted unless we are attempting
957 * an allocation much larger than the huge page size (eg CMA).
958 * We can potentially save a lot of iterations if we skip them
959 * at once. The check is racy, but we can consider only valid
960 * values and the only danger is skipping too much.
962 if (PageCompound(page) && !cc->alloc_contig) {
963 const unsigned int order = compound_order(page);
965 if (likely(order < MAX_ORDER))
966 low_pfn += (1UL << order) - 1;
971 * Check may be lockless but that's ok as we recheck later.
972 * It's possible to migrate LRU and non-lru movable pages.
973 * Skip any other type of page
975 if (!PageLRU(page)) {
977 * __PageMovable can return false positive so we need
978 * to verify it under page_lock.
980 if (unlikely(__PageMovable(page)) &&
981 !PageIsolated(page)) {
983 unlock_page_lruvec_irqrestore(locked, flags);
987 if (!isolate_movable_page(page, isolate_mode))
988 goto isolate_success;
995 * Migration will fail if an anonymous page is pinned in memory,
996 * so avoid taking lru_lock and isolating it unnecessarily in an
997 * admittedly racy check.
999 if (!page_mapping(page) &&
1000 page_count(page) > page_mapcount(page))
1004 * Only allow to migrate anonymous pages in GFP_NOFS context
1005 * because those do not depend on fs locks.
1007 if (!(cc->gfp_mask & __GFP_FS) && page_mapping(page))
1011 * Be careful not to clear PageLRU until after we're
1012 * sure the page is not being freed elsewhere -- the
1013 * page release code relies on it.
1015 if (unlikely(!get_page_unless_zero(page)))
1018 if (!__isolate_lru_page_prepare(page, isolate_mode))
1019 goto isolate_fail_put;
1021 /* Try isolate the page */
1022 if (!TestClearPageLRU(page))
1023 goto isolate_fail_put;
1025 lruvec = mem_cgroup_page_lruvec(page);
1027 /* If we already hold the lock, we can skip some rechecking */
1028 if (lruvec != locked) {
1030 unlock_page_lruvec_irqrestore(locked, flags);
1032 compact_lock_irqsave(&lruvec->lru_lock, &flags, cc);
1035 lruvec_memcg_debug(lruvec, page);
1037 /* Try get exclusive access under lock */
1038 if (!skip_updated) {
1039 skip_updated = true;
1040 if (test_and_set_skip(cc, page, low_pfn))
1045 * Page become compound since the non-locked check,
1046 * and it's on LRU. It can only be a THP so the order
1047 * is safe to read and it's 0 for tail pages.
1049 if (unlikely(PageCompound(page) && !cc->alloc_contig)) {
1050 low_pfn += compound_nr(page) - 1;
1052 goto isolate_fail_put;
1056 /* The whole page is taken off the LRU; skip the tail pages. */
1057 if (PageCompound(page))
1058 low_pfn += compound_nr(page) - 1;
1060 /* Successfully isolated */
1061 del_page_from_lru_list(page, lruvec);
1062 mod_node_page_state(page_pgdat(page),
1063 NR_ISOLATED_ANON + page_is_file_lru(page),
1064 thp_nr_pages(page));
1067 list_add(&page->lru, &cc->migratepages);
1068 isolate_success_no_list:
1069 cc->nr_migratepages += compound_nr(page);
1070 nr_isolated += compound_nr(page);
1073 * Avoid isolating too much unless this block is being
1074 * rescanned (e.g. dirty/writeback pages, parallel allocation)
1075 * or a lock is contended. For contention, isolate quickly to
1076 * potentially remove one source of contention.
1078 if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX &&
1079 !cc->rescan && !cc->contended) {
1087 /* Avoid potential deadlock in freeing page under lru_lock */
1089 unlock_page_lruvec_irqrestore(locked, flags);
1095 if (!skip_on_failure && ret != -ENOMEM)
1099 * We have isolated some pages, but then failed. Release them
1100 * instead of migrating, as we cannot form the cc->order buddy
1105 unlock_page_lruvec_irqrestore(locked, flags);
1108 putback_movable_pages(&cc->migratepages);
1109 cc->nr_migratepages = 0;
1113 if (low_pfn < next_skip_pfn) {
1114 low_pfn = next_skip_pfn - 1;
1116 * The check near the loop beginning would have updated
1117 * next_skip_pfn too, but this is a bit simpler.
1119 next_skip_pfn += 1UL << cc->order;
1127 * The PageBuddy() check could have potentially brought us outside
1128 * the range to be scanned.
1130 if (unlikely(low_pfn > end_pfn))
1137 unlock_page_lruvec_irqrestore(locked, flags);
1144 * Updated the cached scanner pfn once the pageblock has been scanned
1145 * Pages will either be migrated in which case there is no point
1146 * scanning in the near future or migration failed in which case the
1147 * failure reason may persist. The block is marked for skipping if
1148 * there were no pages isolated in the block or if the block is
1149 * rescanned twice in a row.
1151 if (low_pfn == end_pfn && (!nr_isolated || cc->rescan)) {
1152 if (valid_page && !skip_updated)
1153 set_pageblock_skip(valid_page);
1154 update_cached_migrate(cc, low_pfn);
1157 trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
1158 nr_scanned, nr_isolated);
1161 cc->total_migrate_scanned += nr_scanned;
1163 count_compact_events(COMPACTISOLATED, nr_isolated);
1165 cc->migrate_pfn = low_pfn;
1171 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
1172 * @cc: Compaction control structure.
1173 * @start_pfn: The first PFN to start isolating.
1174 * @end_pfn: The one-past-last PFN.
1176 * Returns -EAGAIN when contented, -EINTR in case of a signal pending, -ENOMEM
1177 * in case we could not allocate a page, or 0.
1180 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
1181 unsigned long end_pfn)
1183 unsigned long pfn, block_start_pfn, block_end_pfn;
1186 /* Scan block by block. First and last block may be incomplete */
1188 block_start_pfn = pageblock_start_pfn(pfn);
1189 if (block_start_pfn < cc->zone->zone_start_pfn)
1190 block_start_pfn = cc->zone->zone_start_pfn;
1191 block_end_pfn = pageblock_end_pfn(pfn);
1193 for (; pfn < end_pfn; pfn = block_end_pfn,
1194 block_start_pfn = block_end_pfn,
1195 block_end_pfn += pageblock_nr_pages) {
1197 block_end_pfn = min(block_end_pfn, end_pfn);
1199 if (!pageblock_pfn_to_page(block_start_pfn,
1200 block_end_pfn, cc->zone))
1203 ret = isolate_migratepages_block(cc, pfn, block_end_pfn,
1204 ISOLATE_UNEVICTABLE);
1209 if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX)
1216 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
1217 #ifdef CONFIG_COMPACTION
1219 static bool suitable_migration_source(struct compact_control *cc,
1224 if (pageblock_skip_persistent(page))
1227 if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction)
1230 block_mt = get_pageblock_migratetype(page);
1232 if (cc->migratetype == MIGRATE_MOVABLE)
1233 return is_migrate_movable(block_mt);
1235 return block_mt == cc->migratetype;
1238 /* Returns true if the page is within a block suitable for migration to */
1239 static bool suitable_migration_target(struct compact_control *cc,
1242 /* If the page is a large free page, then disallow migration */
1243 if (PageBuddy(page)) {
1245 * We are checking page_order without zone->lock taken. But
1246 * the only small danger is that we skip a potentially suitable
1247 * pageblock, so it's not worth to check order for valid range.
1249 if (buddy_order_unsafe(page) >= pageblock_order)
1253 if (cc->ignore_block_suitable)
1256 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1257 if (is_migrate_movable(get_pageblock_migratetype(page)))
1260 /* Otherwise skip the block */
1264 static inline unsigned int
1265 freelist_scan_limit(struct compact_control *cc)
1267 unsigned short shift = BITS_PER_LONG - 1;
1269 return (COMPACT_CLUSTER_MAX >> min(shift, cc->fast_search_fail)) + 1;
1273 * Test whether the free scanner has reached the same or lower pageblock than
1274 * the migration scanner, and compaction should thus terminate.
1276 static inline bool compact_scanners_met(struct compact_control *cc)
1278 return (cc->free_pfn >> pageblock_order)
1279 <= (cc->migrate_pfn >> pageblock_order);
1283 * Used when scanning for a suitable migration target which scans freelists
1284 * in reverse. Reorders the list such as the unscanned pages are scanned
1285 * first on the next iteration of the free scanner
1288 move_freelist_head(struct list_head *freelist, struct page *freepage)
1292 if (!list_is_last(freelist, &freepage->lru)) {
1293 list_cut_before(&sublist, freelist, &freepage->lru);
1294 list_splice_tail(&sublist, freelist);
1299 * Similar to move_freelist_head except used by the migration scanner
1300 * when scanning forward. It's possible for these list operations to
1301 * move against each other if they search the free list exactly in
1305 move_freelist_tail(struct list_head *freelist, struct page *freepage)
1309 if (!list_is_first(freelist, &freepage->lru)) {
1310 list_cut_position(&sublist, freelist, &freepage->lru);
1311 list_splice_tail(&sublist, freelist);
1316 fast_isolate_around(struct compact_control *cc, unsigned long pfn, unsigned long nr_isolated)
1318 unsigned long start_pfn, end_pfn;
1321 /* Do not search around if there are enough pages already */
1322 if (cc->nr_freepages >= cc->nr_migratepages)
1325 /* Minimise scanning during async compaction */
1326 if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC)
1329 /* Pageblock boundaries */
1330 start_pfn = max(pageblock_start_pfn(pfn), cc->zone->zone_start_pfn);
1331 end_pfn = min(pageblock_end_pfn(pfn), zone_end_pfn(cc->zone));
1333 page = pageblock_pfn_to_page(start_pfn, end_pfn, cc->zone);
1338 if (start_pfn != pfn) {
1339 isolate_freepages_block(cc, &start_pfn, pfn, &cc->freepages, 1, false);
1340 if (cc->nr_freepages >= cc->nr_migratepages)
1345 start_pfn = pfn + nr_isolated;
1346 if (start_pfn < end_pfn)
1347 isolate_freepages_block(cc, &start_pfn, end_pfn, &cc->freepages, 1, false);
1349 /* Skip this pageblock in the future as it's full or nearly full */
1350 if (cc->nr_freepages < cc->nr_migratepages)
1351 set_pageblock_skip(page);
1354 /* Search orders in round-robin fashion */
1355 static int next_search_order(struct compact_control *cc, int order)
1359 order = cc->order - 1;
1361 /* Search wrapped around? */
1362 if (order == cc->search_order) {
1364 if (cc->search_order < 0)
1365 cc->search_order = cc->order - 1;
1372 static unsigned long
1373 fast_isolate_freepages(struct compact_control *cc)
1375 unsigned int limit = max(1U, freelist_scan_limit(cc) >> 1);
1376 unsigned int nr_scanned = 0;
1377 unsigned long low_pfn, min_pfn, highest = 0;
1378 unsigned long nr_isolated = 0;
1379 unsigned long distance;
1380 struct page *page = NULL;
1381 bool scan_start = false;
1384 /* Full compaction passes in a negative order */
1386 return cc->free_pfn;
1389 * If starting the scan, use a deeper search and use the highest
1390 * PFN found if a suitable one is not found.
1392 if (cc->free_pfn >= cc->zone->compact_init_free_pfn) {
1393 limit = pageblock_nr_pages >> 1;
1398 * Preferred point is in the top quarter of the scan space but take
1399 * a pfn from the top half if the search is problematic.
1401 distance = (cc->free_pfn - cc->migrate_pfn);
1402 low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2));
1403 min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1));
1405 if (WARN_ON_ONCE(min_pfn > low_pfn))
1409 * Search starts from the last successful isolation order or the next
1410 * order to search after a previous failure
1412 cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order);
1414 for (order = cc->search_order;
1415 !page && order >= 0;
1416 order = next_search_order(cc, order)) {
1417 struct free_area *area = &cc->zone->free_area[order];
1418 struct list_head *freelist;
1419 struct page *freepage;
1420 unsigned long flags;
1421 unsigned int order_scanned = 0;
1422 unsigned long high_pfn = 0;
1427 spin_lock_irqsave(&cc->zone->lock, flags);
1428 freelist = &area->free_list[MIGRATE_MOVABLE];
1429 list_for_each_entry_reverse(freepage, freelist, lru) {
1434 pfn = page_to_pfn(freepage);
1437 highest = max(pageblock_start_pfn(pfn),
1438 cc->zone->zone_start_pfn);
1440 if (pfn >= low_pfn) {
1441 cc->fast_search_fail = 0;
1442 cc->search_order = order;
1447 if (pfn >= min_pfn && pfn > high_pfn) {
1450 /* Shorten the scan if a candidate is found */
1454 if (order_scanned >= limit)
1458 /* Use a minimum pfn if a preferred one was not found */
1459 if (!page && high_pfn) {
1460 page = pfn_to_page(high_pfn);
1462 /* Update freepage for the list reorder below */
1466 /* Reorder to so a future search skips recent pages */
1467 move_freelist_head(freelist, freepage);
1469 /* Isolate the page if available */
1471 if (__isolate_free_page(page, order)) {
1472 set_page_private(page, order);
1473 nr_isolated = 1 << order;
1474 cc->nr_freepages += nr_isolated;
1475 list_add_tail(&page->lru, &cc->freepages);
1476 count_compact_events(COMPACTISOLATED, nr_isolated);
1478 /* If isolation fails, abort the search */
1479 order = cc->search_order + 1;
1484 spin_unlock_irqrestore(&cc->zone->lock, flags);
1487 * Smaller scan on next order so the total scan is related
1488 * to freelist_scan_limit.
1490 if (order_scanned >= limit)
1491 limit = max(1U, limit >> 1);
1495 cc->fast_search_fail++;
1498 * Use the highest PFN found above min. If one was
1499 * not found, be pessimistic for direct compaction
1500 * and use the min mark.
1503 page = pfn_to_page(highest);
1504 cc->free_pfn = highest;
1506 if (cc->direct_compaction && pfn_valid(min_pfn)) {
1507 page = pageblock_pfn_to_page(min_pfn,
1508 min(pageblock_end_pfn(min_pfn),
1509 zone_end_pfn(cc->zone)),
1511 cc->free_pfn = min_pfn;
1517 if (highest && highest >= cc->zone->compact_cached_free_pfn) {
1518 highest -= pageblock_nr_pages;
1519 cc->zone->compact_cached_free_pfn = highest;
1522 cc->total_free_scanned += nr_scanned;
1524 return cc->free_pfn;
1526 low_pfn = page_to_pfn(page);
1527 fast_isolate_around(cc, low_pfn, nr_isolated);
1532 * Based on information in the current compact_control, find blocks
1533 * suitable for isolating free pages from and then isolate them.
1535 static void isolate_freepages(struct compact_control *cc)
1537 struct zone *zone = cc->zone;
1539 unsigned long block_start_pfn; /* start of current pageblock */
1540 unsigned long isolate_start_pfn; /* exact pfn we start at */
1541 unsigned long block_end_pfn; /* end of current pageblock */
1542 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
1543 struct list_head *freelist = &cc->freepages;
1544 unsigned int stride;
1546 /* Try a small search of the free lists for a candidate */
1547 isolate_start_pfn = fast_isolate_freepages(cc);
1548 if (cc->nr_freepages)
1552 * Initialise the free scanner. The starting point is where we last
1553 * successfully isolated from, zone-cached value, or the end of the
1554 * zone when isolating for the first time. For looping we also need
1555 * this pfn aligned down to the pageblock boundary, because we do
1556 * block_start_pfn -= pageblock_nr_pages in the for loop.
1557 * For ending point, take care when isolating in last pageblock of a
1558 * zone which ends in the middle of a pageblock.
1559 * The low boundary is the end of the pageblock the migration scanner
1562 isolate_start_pfn = cc->free_pfn;
1563 block_start_pfn = pageblock_start_pfn(isolate_start_pfn);
1564 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1565 zone_end_pfn(zone));
1566 low_pfn = pageblock_end_pfn(cc->migrate_pfn);
1567 stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1;
1570 * Isolate free pages until enough are available to migrate the
1571 * pages on cc->migratepages. We stop searching if the migrate
1572 * and free page scanners meet or enough free pages are isolated.
1574 for (; block_start_pfn >= low_pfn;
1575 block_end_pfn = block_start_pfn,
1576 block_start_pfn -= pageblock_nr_pages,
1577 isolate_start_pfn = block_start_pfn) {
1578 unsigned long nr_isolated;
1581 * This can iterate a massively long zone without finding any
1582 * suitable migration targets, so periodically check resched.
1584 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)))
1587 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1592 /* Check the block is suitable for migration */
1593 if (!suitable_migration_target(cc, page))
1596 /* If isolation recently failed, do not retry */
1597 if (!isolation_suitable(cc, page))
1600 /* Found a block suitable for isolating free pages from. */
1601 nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn,
1602 block_end_pfn, freelist, stride, false);
1604 /* Update the skip hint if the full pageblock was scanned */
1605 if (isolate_start_pfn == block_end_pfn)
1606 update_pageblock_skip(cc, page, block_start_pfn);
1608 /* Are enough freepages isolated? */
1609 if (cc->nr_freepages >= cc->nr_migratepages) {
1610 if (isolate_start_pfn >= block_end_pfn) {
1612 * Restart at previous pageblock if more
1613 * freepages can be isolated next time.
1616 block_start_pfn - pageblock_nr_pages;
1619 } else if (isolate_start_pfn < block_end_pfn) {
1621 * If isolation failed early, do not continue
1627 /* Adjust stride depending on isolation */
1632 stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1);
1636 * Record where the free scanner will restart next time. Either we
1637 * broke from the loop and set isolate_start_pfn based on the last
1638 * call to isolate_freepages_block(), or we met the migration scanner
1639 * and the loop terminated due to isolate_start_pfn < low_pfn
1641 cc->free_pfn = isolate_start_pfn;
1644 /* __isolate_free_page() does not map the pages */
1645 split_map_pages(freelist);
1649 * This is a migrate-callback that "allocates" freepages by taking pages
1650 * from the isolated freelists in the block we are migrating to.
1652 static struct page *compaction_alloc(struct page *migratepage,
1655 struct compact_control *cc = (struct compact_control *)data;
1656 struct page *freepage;
1658 if (list_empty(&cc->freepages)) {
1659 isolate_freepages(cc);
1661 if (list_empty(&cc->freepages))
1665 freepage = list_entry(cc->freepages.next, struct page, lru);
1666 list_del(&freepage->lru);
1673 * This is a migrate-callback that "frees" freepages back to the isolated
1674 * freelist. All pages on the freelist are from the same zone, so there is no
1675 * special handling needed for NUMA.
1677 static void compaction_free(struct page *page, unsigned long data)
1679 struct compact_control *cc = (struct compact_control *)data;
1681 list_add(&page->lru, &cc->freepages);
1685 /* possible outcome of isolate_migratepages */
1687 ISOLATE_ABORT, /* Abort compaction now */
1688 ISOLATE_NONE, /* No pages isolated, continue scanning */
1689 ISOLATE_SUCCESS, /* Pages isolated, migrate */
1690 } isolate_migrate_t;
1693 * Allow userspace to control policy on scanning the unevictable LRU for
1694 * compactable pages.
1696 #ifdef CONFIG_PREEMPT_RT
1697 int sysctl_compact_unevictable_allowed __read_mostly = 0;
1699 int sysctl_compact_unevictable_allowed __read_mostly = 1;
1703 update_fast_start_pfn(struct compact_control *cc, unsigned long pfn)
1705 if (cc->fast_start_pfn == ULONG_MAX)
1708 if (!cc->fast_start_pfn)
1709 cc->fast_start_pfn = pfn;
1711 cc->fast_start_pfn = min(cc->fast_start_pfn, pfn);
1714 static inline unsigned long
1715 reinit_migrate_pfn(struct compact_control *cc)
1717 if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX)
1718 return cc->migrate_pfn;
1720 cc->migrate_pfn = cc->fast_start_pfn;
1721 cc->fast_start_pfn = ULONG_MAX;
1723 return cc->migrate_pfn;
1727 * Briefly search the free lists for a migration source that already has
1728 * some free pages to reduce the number of pages that need migration
1729 * before a pageblock is free.
1731 static unsigned long fast_find_migrateblock(struct compact_control *cc)
1733 unsigned int limit = freelist_scan_limit(cc);
1734 unsigned int nr_scanned = 0;
1735 unsigned long distance;
1736 unsigned long pfn = cc->migrate_pfn;
1737 unsigned long high_pfn;
1739 bool found_block = false;
1741 /* Skip hints are relied on to avoid repeats on the fast search */
1742 if (cc->ignore_skip_hint)
1746 * If the migrate_pfn is not at the start of a zone or the start
1747 * of a pageblock then assume this is a continuation of a previous
1748 * scan restarted due to COMPACT_CLUSTER_MAX.
1750 if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn))
1754 * For smaller orders, just linearly scan as the number of pages
1755 * to migrate should be relatively small and does not necessarily
1756 * justify freeing up a large block for a small allocation.
1758 if (cc->order <= PAGE_ALLOC_COSTLY_ORDER)
1762 * Only allow kcompactd and direct requests for movable pages to
1763 * quickly clear out a MOVABLE pageblock for allocation. This
1764 * reduces the risk that a large movable pageblock is freed for
1765 * an unmovable/reclaimable small allocation.
1767 if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE)
1771 * When starting the migration scanner, pick any pageblock within the
1772 * first half of the search space. Otherwise try and pick a pageblock
1773 * within the first eighth to reduce the chances that a migration
1774 * target later becomes a source.
1776 distance = (cc->free_pfn - cc->migrate_pfn) >> 1;
1777 if (cc->migrate_pfn != cc->zone->zone_start_pfn)
1779 high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance);
1781 for (order = cc->order - 1;
1782 order >= PAGE_ALLOC_COSTLY_ORDER && !found_block && nr_scanned < limit;
1784 struct free_area *area = &cc->zone->free_area[order];
1785 struct list_head *freelist;
1786 unsigned long flags;
1787 struct page *freepage;
1792 spin_lock_irqsave(&cc->zone->lock, flags);
1793 freelist = &area->free_list[MIGRATE_MOVABLE];
1794 list_for_each_entry(freepage, freelist, lru) {
1795 unsigned long free_pfn;
1797 if (nr_scanned++ >= limit) {
1798 move_freelist_tail(freelist, freepage);
1802 free_pfn = page_to_pfn(freepage);
1803 if (free_pfn < high_pfn) {
1805 * Avoid if skipped recently. Ideally it would
1806 * move to the tail but even safe iteration of
1807 * the list assumes an entry is deleted, not
1810 if (get_pageblock_skip(freepage))
1813 /* Reorder to so a future search skips recent pages */
1814 move_freelist_tail(freelist, freepage);
1816 update_fast_start_pfn(cc, free_pfn);
1817 pfn = pageblock_start_pfn(free_pfn);
1818 if (pfn < cc->zone->zone_start_pfn)
1819 pfn = cc->zone->zone_start_pfn;
1820 cc->fast_search_fail = 0;
1822 set_pageblock_skip(freepage);
1826 spin_unlock_irqrestore(&cc->zone->lock, flags);
1829 cc->total_migrate_scanned += nr_scanned;
1832 * If fast scanning failed then use a cached entry for a page block
1833 * that had free pages as the basis for starting a linear scan.
1836 cc->fast_search_fail++;
1837 pfn = reinit_migrate_pfn(cc);
1843 * Isolate all pages that can be migrated from the first suitable block,
1844 * starting at the block pointed to by the migrate scanner pfn within
1847 static isolate_migrate_t isolate_migratepages(struct compact_control *cc)
1849 unsigned long block_start_pfn;
1850 unsigned long block_end_pfn;
1851 unsigned long low_pfn;
1853 const isolate_mode_t isolate_mode =
1854 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1855 (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1856 bool fast_find_block;
1859 * Start at where we last stopped, or beginning of the zone as
1860 * initialized by compact_zone(). The first failure will use
1861 * the lowest PFN as the starting point for linear scanning.
1863 low_pfn = fast_find_migrateblock(cc);
1864 block_start_pfn = pageblock_start_pfn(low_pfn);
1865 if (block_start_pfn < cc->zone->zone_start_pfn)
1866 block_start_pfn = cc->zone->zone_start_pfn;
1869 * fast_find_migrateblock marks a pageblock skipped so to avoid
1870 * the isolation_suitable check below, check whether the fast
1871 * search was successful.
1873 fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail;
1875 /* Only scan within a pageblock boundary */
1876 block_end_pfn = pageblock_end_pfn(low_pfn);
1879 * Iterate over whole pageblocks until we find the first suitable.
1880 * Do not cross the free scanner.
1882 for (; block_end_pfn <= cc->free_pfn;
1883 fast_find_block = false,
1884 cc->migrate_pfn = low_pfn = block_end_pfn,
1885 block_start_pfn = block_end_pfn,
1886 block_end_pfn += pageblock_nr_pages) {
1889 * This can potentially iterate a massively long zone with
1890 * many pageblocks unsuitable, so periodically check if we
1893 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)))
1896 page = pageblock_pfn_to_page(block_start_pfn,
1897 block_end_pfn, cc->zone);
1902 * If isolation recently failed, do not retry. Only check the
1903 * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock
1904 * to be visited multiple times. Assume skip was checked
1905 * before making it "skip" so other compaction instances do
1906 * not scan the same block.
1908 if (IS_ALIGNED(low_pfn, pageblock_nr_pages) &&
1909 !fast_find_block && !isolation_suitable(cc, page))
1913 * For async compaction, also only scan in MOVABLE blocks
1914 * without huge pages. Async compaction is optimistic to see
1915 * if the minimum amount of work satisfies the allocation.
1916 * The cached PFN is updated as it's possible that all
1917 * remaining blocks between source and target are unsuitable
1918 * and the compaction scanners fail to meet.
1920 if (!suitable_migration_source(cc, page)) {
1921 update_cached_migrate(cc, block_end_pfn);
1925 /* Perform the isolation */
1926 if (isolate_migratepages_block(cc, low_pfn, block_end_pfn,
1928 return ISOLATE_ABORT;
1931 * Either we isolated something and proceed with migration. Or
1932 * we failed and compact_zone should decide if we should
1938 return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
1942 * order == -1 is expected when compacting via
1943 * /proc/sys/vm/compact_memory
1945 static inline bool is_via_compact_memory(int order)
1950 static bool kswapd_is_running(pg_data_t *pgdat)
1952 return pgdat->kswapd && task_is_running(pgdat->kswapd);
1956 * A zone's fragmentation score is the external fragmentation wrt to the
1957 * COMPACTION_HPAGE_ORDER. It returns a value in the range [0, 100].
1959 static unsigned int fragmentation_score_zone(struct zone *zone)
1961 return extfrag_for_order(zone, COMPACTION_HPAGE_ORDER);
1965 * A weighted zone's fragmentation score is the external fragmentation
1966 * wrt to the COMPACTION_HPAGE_ORDER scaled by the zone's size. It
1967 * returns a value in the range [0, 100].
1969 * The scaling factor ensures that proactive compaction focuses on larger
1970 * zones like ZONE_NORMAL, rather than smaller, specialized zones like
1971 * ZONE_DMA32. For smaller zones, the score value remains close to zero,
1972 * and thus never exceeds the high threshold for proactive compaction.
1974 static unsigned int fragmentation_score_zone_weighted(struct zone *zone)
1976 unsigned long score;
1978 score = zone->present_pages * fragmentation_score_zone(zone);
1979 return div64_ul(score, zone->zone_pgdat->node_present_pages + 1);
1983 * The per-node proactive (background) compaction process is started by its
1984 * corresponding kcompactd thread when the node's fragmentation score
1985 * exceeds the high threshold. The compaction process remains active till
1986 * the node's score falls below the low threshold, or one of the back-off
1987 * conditions is met.
1989 static unsigned int fragmentation_score_node(pg_data_t *pgdat)
1991 unsigned int score = 0;
1994 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1997 zone = &pgdat->node_zones[zoneid];
1998 score += fragmentation_score_zone_weighted(zone);
2004 static unsigned int fragmentation_score_wmark(pg_data_t *pgdat, bool low)
2006 unsigned int wmark_low;
2009 * Cap the low watermark to avoid excessive compaction
2010 * activity in case a user sets the proactiveness tunable
2011 * close to 100 (maximum).
2013 wmark_low = max(100U - sysctl_compaction_proactiveness, 5U);
2014 return low ? wmark_low : min(wmark_low + 10, 100U);
2017 static bool should_proactive_compact_node(pg_data_t *pgdat)
2021 if (!sysctl_compaction_proactiveness || kswapd_is_running(pgdat))
2024 wmark_high = fragmentation_score_wmark(pgdat, false);
2025 return fragmentation_score_node(pgdat) > wmark_high;
2028 static enum compact_result __compact_finished(struct compact_control *cc)
2031 const int migratetype = cc->migratetype;
2034 /* Compaction run completes if the migrate and free scanner meet */
2035 if (compact_scanners_met(cc)) {
2036 /* Let the next compaction start anew. */
2037 reset_cached_positions(cc->zone);
2040 * Mark that the PG_migrate_skip information should be cleared
2041 * by kswapd when it goes to sleep. kcompactd does not set the
2042 * flag itself as the decision to be clear should be directly
2043 * based on an allocation request.
2045 if (cc->direct_compaction)
2046 cc->zone->compact_blockskip_flush = true;
2049 return COMPACT_COMPLETE;
2051 return COMPACT_PARTIAL_SKIPPED;
2054 if (cc->proactive_compaction) {
2055 int score, wmark_low;
2058 pgdat = cc->zone->zone_pgdat;
2059 if (kswapd_is_running(pgdat))
2060 return COMPACT_PARTIAL_SKIPPED;
2062 score = fragmentation_score_zone(cc->zone);
2063 wmark_low = fragmentation_score_wmark(pgdat, true);
2065 if (score > wmark_low)
2066 ret = COMPACT_CONTINUE;
2068 ret = COMPACT_SUCCESS;
2073 if (is_via_compact_memory(cc->order))
2074 return COMPACT_CONTINUE;
2077 * Always finish scanning a pageblock to reduce the possibility of
2078 * fallbacks in the future. This is particularly important when
2079 * migration source is unmovable/reclaimable but it's not worth
2082 if (!IS_ALIGNED(cc->migrate_pfn, pageblock_nr_pages))
2083 return COMPACT_CONTINUE;
2085 /* Direct compactor: Is a suitable page free? */
2086 ret = COMPACT_NO_SUITABLE_PAGE;
2087 for (order = cc->order; order < MAX_ORDER; order++) {
2088 struct free_area *area = &cc->zone->free_area[order];
2091 /* Job done if page is free of the right migratetype */
2092 if (!free_area_empty(area, migratetype))
2093 return COMPACT_SUCCESS;
2096 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
2097 if (migratetype == MIGRATE_MOVABLE &&
2098 !free_area_empty(area, MIGRATE_CMA))
2099 return COMPACT_SUCCESS;
2102 * Job done if allocation would steal freepages from
2103 * other migratetype buddy lists.
2105 if (find_suitable_fallback(area, order, migratetype,
2106 true, &can_steal) != -1) {
2108 /* movable pages are OK in any pageblock */
2109 if (migratetype == MIGRATE_MOVABLE)
2110 return COMPACT_SUCCESS;
2113 * We are stealing for a non-movable allocation. Make
2114 * sure we finish compacting the current pageblock
2115 * first so it is as free as possible and we won't
2116 * have to steal another one soon. This only applies
2117 * to sync compaction, as async compaction operates
2118 * on pageblocks of the same migratetype.
2120 if (cc->mode == MIGRATE_ASYNC ||
2121 IS_ALIGNED(cc->migrate_pfn,
2122 pageblock_nr_pages)) {
2123 return COMPACT_SUCCESS;
2126 ret = COMPACT_CONTINUE;
2132 if (cc->contended || fatal_signal_pending(current))
2133 ret = COMPACT_CONTENDED;
2138 static enum compact_result compact_finished(struct compact_control *cc)
2142 ret = __compact_finished(cc);
2143 trace_mm_compaction_finished(cc->zone, cc->order, ret);
2144 if (ret == COMPACT_NO_SUITABLE_PAGE)
2145 ret = COMPACT_CONTINUE;
2150 static enum compact_result __compaction_suitable(struct zone *zone, int order,
2151 unsigned int alloc_flags,
2152 int highest_zoneidx,
2153 unsigned long wmark_target)
2155 unsigned long watermark;
2157 if (is_via_compact_memory(order))
2158 return COMPACT_CONTINUE;
2160 watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
2162 * If watermarks for high-order allocation are already met, there
2163 * should be no need for compaction at all.
2165 if (zone_watermark_ok(zone, order, watermark, highest_zoneidx,
2167 return COMPACT_SUCCESS;
2170 * Watermarks for order-0 must be met for compaction to be able to
2171 * isolate free pages for migration targets. This means that the
2172 * watermark and alloc_flags have to match, or be more pessimistic than
2173 * the check in __isolate_free_page(). We don't use the direct
2174 * compactor's alloc_flags, as they are not relevant for freepage
2175 * isolation. We however do use the direct compactor's highest_zoneidx
2176 * to skip over zones where lowmem reserves would prevent allocation
2177 * even if compaction succeeds.
2178 * For costly orders, we require low watermark instead of min for
2179 * compaction to proceed to increase its chances.
2180 * ALLOC_CMA is used, as pages in CMA pageblocks are considered
2181 * suitable migration targets
2183 watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
2184 low_wmark_pages(zone) : min_wmark_pages(zone);
2185 watermark += compact_gap(order);
2186 if (!__zone_watermark_ok(zone, 0, watermark, highest_zoneidx,
2187 ALLOC_CMA, wmark_target))
2188 return COMPACT_SKIPPED;
2190 return COMPACT_CONTINUE;
2194 * compaction_suitable: Is this suitable to run compaction on this zone now?
2196 * COMPACT_SKIPPED - If there are too few free pages for compaction
2197 * COMPACT_SUCCESS - If the allocation would succeed without compaction
2198 * COMPACT_CONTINUE - If compaction should run now
2200 enum compact_result compaction_suitable(struct zone *zone, int order,
2201 unsigned int alloc_flags,
2202 int highest_zoneidx)
2204 enum compact_result ret;
2207 ret = __compaction_suitable(zone, order, alloc_flags, highest_zoneidx,
2208 zone_page_state(zone, NR_FREE_PAGES));
2210 * fragmentation index determines if allocation failures are due to
2211 * low memory or external fragmentation
2213 * index of -1000 would imply allocations might succeed depending on
2214 * watermarks, but we already failed the high-order watermark check
2215 * index towards 0 implies failure is due to lack of memory
2216 * index towards 1000 implies failure is due to fragmentation
2218 * Only compact if a failure would be due to fragmentation. Also
2219 * ignore fragindex for non-costly orders where the alternative to
2220 * a successful reclaim/compaction is OOM. Fragindex and the
2221 * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
2222 * excessive compaction for costly orders, but it should not be at the
2223 * expense of system stability.
2225 if (ret == COMPACT_CONTINUE && (order > PAGE_ALLOC_COSTLY_ORDER)) {
2226 fragindex = fragmentation_index(zone, order);
2227 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
2228 ret = COMPACT_NOT_SUITABLE_ZONE;
2231 trace_mm_compaction_suitable(zone, order, ret);
2232 if (ret == COMPACT_NOT_SUITABLE_ZONE)
2233 ret = COMPACT_SKIPPED;
2238 bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
2245 * Make sure at least one zone would pass __compaction_suitable if we continue
2246 * retrying the reclaim.
2248 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2249 ac->highest_zoneidx, ac->nodemask) {
2250 unsigned long available;
2251 enum compact_result compact_result;
2254 * Do not consider all the reclaimable memory because we do not
2255 * want to trash just for a single high order allocation which
2256 * is even not guaranteed to appear even if __compaction_suitable
2257 * is happy about the watermark check.
2259 available = zone_reclaimable_pages(zone) / order;
2260 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
2261 compact_result = __compaction_suitable(zone, order, alloc_flags,
2262 ac->highest_zoneidx, available);
2263 if (compact_result != COMPACT_SKIPPED)
2270 static enum compact_result
2271 compact_zone(struct compact_control *cc, struct capture_control *capc)
2273 enum compact_result ret;
2274 unsigned long start_pfn = cc->zone->zone_start_pfn;
2275 unsigned long end_pfn = zone_end_pfn(cc->zone);
2276 unsigned long last_migrated_pfn;
2277 const bool sync = cc->mode != MIGRATE_ASYNC;
2281 * These counters track activities during zone compaction. Initialize
2282 * them before compacting a new zone.
2284 cc->total_migrate_scanned = 0;
2285 cc->total_free_scanned = 0;
2286 cc->nr_migratepages = 0;
2287 cc->nr_freepages = 0;
2288 INIT_LIST_HEAD(&cc->freepages);
2289 INIT_LIST_HEAD(&cc->migratepages);
2291 cc->migratetype = gfp_migratetype(cc->gfp_mask);
2292 ret = compaction_suitable(cc->zone, cc->order, cc->alloc_flags,
2293 cc->highest_zoneidx);
2294 /* Compaction is likely to fail */
2295 if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED)
2298 /* huh, compaction_suitable is returning something unexpected */
2299 VM_BUG_ON(ret != COMPACT_CONTINUE);
2302 * Clear pageblock skip if there were failures recently and compaction
2303 * is about to be retried after being deferred.
2305 if (compaction_restarting(cc->zone, cc->order))
2306 __reset_isolation_suitable(cc->zone);
2309 * Setup to move all movable pages to the end of the zone. Used cached
2310 * information on where the scanners should start (unless we explicitly
2311 * want to compact the whole zone), but check that it is initialised
2312 * by ensuring the values are within zone boundaries.
2314 cc->fast_start_pfn = 0;
2315 if (cc->whole_zone) {
2316 cc->migrate_pfn = start_pfn;
2317 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2319 cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync];
2320 cc->free_pfn = cc->zone->compact_cached_free_pfn;
2321 if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
2322 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2323 cc->zone->compact_cached_free_pfn = cc->free_pfn;
2325 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
2326 cc->migrate_pfn = start_pfn;
2327 cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
2328 cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
2331 if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn)
2332 cc->whole_zone = true;
2335 last_migrated_pfn = 0;
2338 * Migrate has separate cached PFNs for ASYNC and SYNC* migration on
2339 * the basis that some migrations will fail in ASYNC mode. However,
2340 * if the cached PFNs match and pageblocks are skipped due to having
2341 * no isolation candidates, then the sync state does not matter.
2342 * Until a pageblock with isolation candidates is found, keep the
2343 * cached PFNs in sync to avoid revisiting the same blocks.
2345 update_cached = !sync &&
2346 cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1];
2348 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
2349 cc->free_pfn, end_pfn, sync);
2351 /* lru_add_drain_all could be expensive with involving other CPUs */
2354 while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) {
2356 unsigned long iteration_start_pfn = cc->migrate_pfn;
2359 * Avoid multiple rescans which can happen if a page cannot be
2360 * isolated (dirty/writeback in async mode) or if the migrated
2361 * pages are being allocated before the pageblock is cleared.
2362 * The first rescan will capture the entire pageblock for
2363 * migration. If it fails, it'll be marked skip and scanning
2364 * will proceed as normal.
2367 if (pageblock_start_pfn(last_migrated_pfn) ==
2368 pageblock_start_pfn(iteration_start_pfn)) {
2372 switch (isolate_migratepages(cc)) {
2374 ret = COMPACT_CONTENDED;
2375 putback_movable_pages(&cc->migratepages);
2376 cc->nr_migratepages = 0;
2379 if (update_cached) {
2380 cc->zone->compact_cached_migrate_pfn[1] =
2381 cc->zone->compact_cached_migrate_pfn[0];
2385 * We haven't isolated and migrated anything, but
2386 * there might still be unflushed migrations from
2387 * previous cc->order aligned block.
2390 case ISOLATE_SUCCESS:
2391 update_cached = false;
2392 last_migrated_pfn = iteration_start_pfn;
2395 err = migrate_pages(&cc->migratepages, compaction_alloc,
2396 compaction_free, (unsigned long)cc, cc->mode,
2397 MR_COMPACTION, NULL);
2399 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
2402 /* All pages were either migrated or will be released */
2403 cc->nr_migratepages = 0;
2405 putback_movable_pages(&cc->migratepages);
2407 * migrate_pages() may return -ENOMEM when scanners meet
2408 * and we want compact_finished() to detect it
2410 if (err == -ENOMEM && !compact_scanners_met(cc)) {
2411 ret = COMPACT_CONTENDED;
2415 * We failed to migrate at least one page in the current
2416 * order-aligned block, so skip the rest of it.
2418 if (cc->direct_compaction &&
2419 (cc->mode == MIGRATE_ASYNC)) {
2420 cc->migrate_pfn = block_end_pfn(
2421 cc->migrate_pfn - 1, cc->order);
2422 /* Draining pcplists is useless in this case */
2423 last_migrated_pfn = 0;
2429 * Has the migration scanner moved away from the previous
2430 * cc->order aligned block where we migrated from? If yes,
2431 * flush the pages that were freed, so that they can merge and
2432 * compact_finished() can detect immediately if allocation
2435 if (cc->order > 0 && last_migrated_pfn) {
2436 unsigned long current_block_start =
2437 block_start_pfn(cc->migrate_pfn, cc->order);
2439 if (last_migrated_pfn < current_block_start) {
2440 lru_add_drain_cpu_zone(cc->zone);
2441 /* No more flushing until we migrate again */
2442 last_migrated_pfn = 0;
2446 /* Stop if a page has been captured */
2447 if (capc && capc->page) {
2448 ret = COMPACT_SUCCESS;
2455 * Release free pages and update where the free scanner should restart,
2456 * so we don't leave any returned pages behind in the next attempt.
2458 if (cc->nr_freepages > 0) {
2459 unsigned long free_pfn = release_freepages(&cc->freepages);
2461 cc->nr_freepages = 0;
2462 VM_BUG_ON(free_pfn == 0);
2463 /* The cached pfn is always the first in a pageblock */
2464 free_pfn = pageblock_start_pfn(free_pfn);
2466 * Only go back, not forward. The cached pfn might have been
2467 * already reset to zone end in compact_finished()
2469 if (free_pfn > cc->zone->compact_cached_free_pfn)
2470 cc->zone->compact_cached_free_pfn = free_pfn;
2473 count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned);
2474 count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned);
2476 trace_mm_compaction_end(start_pfn, cc->migrate_pfn,
2477 cc->free_pfn, end_pfn, sync, ret);
2482 static enum compact_result compact_zone_order(struct zone *zone, int order,
2483 gfp_t gfp_mask, enum compact_priority prio,
2484 unsigned int alloc_flags, int highest_zoneidx,
2485 struct page **capture)
2487 enum compact_result ret;
2488 struct compact_control cc = {
2490 .search_order = order,
2491 .gfp_mask = gfp_mask,
2493 .mode = (prio == COMPACT_PRIO_ASYNC) ?
2494 MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT,
2495 .alloc_flags = alloc_flags,
2496 .highest_zoneidx = highest_zoneidx,
2497 .direct_compaction = true,
2498 .whole_zone = (prio == MIN_COMPACT_PRIORITY),
2499 .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
2500 .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
2502 struct capture_control capc = {
2508 * Make sure the structs are really initialized before we expose the
2509 * capture control, in case we are interrupted and the interrupt handler
2513 WRITE_ONCE(current->capture_control, &capc);
2515 ret = compact_zone(&cc, &capc);
2517 VM_BUG_ON(!list_empty(&cc.freepages));
2518 VM_BUG_ON(!list_empty(&cc.migratepages));
2521 * Make sure we hide capture control first before we read the captured
2522 * page pointer, otherwise an interrupt could free and capture a page
2523 * and we would leak it.
2525 WRITE_ONCE(current->capture_control, NULL);
2526 *capture = READ_ONCE(capc.page);
2528 * Technically, it is also possible that compaction is skipped but
2529 * the page is still captured out of luck(IRQ came and freed the page).
2530 * Returning COMPACT_SUCCESS in such cases helps in properly accounting
2531 * the COMPACT[STALL|FAIL] when compaction is skipped.
2534 ret = COMPACT_SUCCESS;
2539 int sysctl_extfrag_threshold = 500;
2542 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
2543 * @gfp_mask: The GFP mask of the current allocation
2544 * @order: The order of the current allocation
2545 * @alloc_flags: The allocation flags of the current allocation
2546 * @ac: The context of current allocation
2547 * @prio: Determines how hard direct compaction should try to succeed
2548 * @capture: Pointer to free page created by compaction will be stored here
2550 * This is the main entry point for direct page compaction.
2552 enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
2553 unsigned int alloc_flags, const struct alloc_context *ac,
2554 enum compact_priority prio, struct page **capture)
2556 int may_perform_io = gfp_mask & __GFP_IO;
2559 enum compact_result rc = COMPACT_SKIPPED;
2562 * Check if the GFP flags allow compaction - GFP_NOIO is really
2563 * tricky context because the migration might require IO
2565 if (!may_perform_io)
2566 return COMPACT_SKIPPED;
2568 trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
2570 /* Compact each zone in the list */
2571 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2572 ac->highest_zoneidx, ac->nodemask) {
2573 enum compact_result status;
2575 if (prio > MIN_COMPACT_PRIORITY
2576 && compaction_deferred(zone, order)) {
2577 rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
2581 status = compact_zone_order(zone, order, gfp_mask, prio,
2582 alloc_flags, ac->highest_zoneidx, capture);
2583 rc = max(status, rc);
2585 /* The allocation should succeed, stop compacting */
2586 if (status == COMPACT_SUCCESS) {
2588 * We think the allocation will succeed in this zone,
2589 * but it is not certain, hence the false. The caller
2590 * will repeat this with true if allocation indeed
2591 * succeeds in this zone.
2593 compaction_defer_reset(zone, order, false);
2598 if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
2599 status == COMPACT_PARTIAL_SKIPPED))
2601 * We think that allocation won't succeed in this zone
2602 * so we defer compaction there. If it ends up
2603 * succeeding after all, it will be reset.
2605 defer_compaction(zone, order);
2608 * We might have stopped compacting due to need_resched() in
2609 * async compaction, or due to a fatal signal detected. In that
2610 * case do not try further zones
2612 if ((prio == COMPACT_PRIO_ASYNC && need_resched())
2613 || fatal_signal_pending(current))
2621 * Compact all zones within a node till each zone's fragmentation score
2622 * reaches within proactive compaction thresholds (as determined by the
2623 * proactiveness tunable).
2625 * It is possible that the function returns before reaching score targets
2626 * due to various back-off conditions, such as, contention on per-node or
2629 static void proactive_compact_node(pg_data_t *pgdat)
2633 struct compact_control cc = {
2635 .mode = MIGRATE_SYNC_LIGHT,
2636 .ignore_skip_hint = true,
2638 .gfp_mask = GFP_KERNEL,
2639 .proactive_compaction = true,
2642 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2643 zone = &pgdat->node_zones[zoneid];
2644 if (!populated_zone(zone))
2649 compact_zone(&cc, NULL);
2651 VM_BUG_ON(!list_empty(&cc.freepages));
2652 VM_BUG_ON(!list_empty(&cc.migratepages));
2656 /* Compact all zones within a node */
2657 static void compact_node(int nid)
2659 pg_data_t *pgdat = NODE_DATA(nid);
2662 struct compact_control cc = {
2664 .mode = MIGRATE_SYNC,
2665 .ignore_skip_hint = true,
2667 .gfp_mask = GFP_KERNEL,
2671 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2673 zone = &pgdat->node_zones[zoneid];
2674 if (!populated_zone(zone))
2679 compact_zone(&cc, NULL);
2681 VM_BUG_ON(!list_empty(&cc.freepages));
2682 VM_BUG_ON(!list_empty(&cc.migratepages));
2686 /* Compact all nodes in the system */
2687 static void compact_nodes(void)
2691 /* Flush pending updates to the LRU lists */
2692 lru_add_drain_all();
2694 for_each_online_node(nid)
2699 * Tunable for proactive compaction. It determines how
2700 * aggressively the kernel should compact memory in the
2701 * background. It takes values in the range [0, 100].
2703 unsigned int __read_mostly sysctl_compaction_proactiveness = 20;
2705 int compaction_proactiveness_sysctl_handler(struct ctl_table *table, int write,
2706 void *buffer, size_t *length, loff_t *ppos)
2710 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
2714 if (write && sysctl_compaction_proactiveness) {
2715 for_each_online_node(nid) {
2716 pg_data_t *pgdat = NODE_DATA(nid);
2718 if (pgdat->proactive_compact_trigger)
2721 pgdat->proactive_compact_trigger = true;
2722 wake_up_interruptible(&pgdat->kcompactd_wait);
2730 * This is the entry point for compacting all nodes via
2731 * /proc/sys/vm/compact_memory
2733 int sysctl_compaction_handler(struct ctl_table *table, int write,
2734 void *buffer, size_t *length, loff_t *ppos)
2742 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
2743 static ssize_t compact_store(struct device *dev,
2744 struct device_attribute *attr,
2745 const char *buf, size_t count)
2749 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
2750 /* Flush pending updates to the LRU lists */
2751 lru_add_drain_all();
2758 static DEVICE_ATTR_WO(compact);
2760 int compaction_register_node(struct node *node)
2762 return device_create_file(&node->dev, &dev_attr_compact);
2765 void compaction_unregister_node(struct node *node)
2767 return device_remove_file(&node->dev, &dev_attr_compact);
2769 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
2771 static inline bool kcompactd_work_requested(pg_data_t *pgdat)
2773 return pgdat->kcompactd_max_order > 0 || kthread_should_stop() ||
2774 pgdat->proactive_compact_trigger;
2777 static bool kcompactd_node_suitable(pg_data_t *pgdat)
2781 enum zone_type highest_zoneidx = pgdat->kcompactd_highest_zoneidx;
2783 for (zoneid = 0; zoneid <= highest_zoneidx; zoneid++) {
2784 zone = &pgdat->node_zones[zoneid];
2786 if (!populated_zone(zone))
2789 if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0,
2790 highest_zoneidx) == COMPACT_CONTINUE)
2797 static void kcompactd_do_work(pg_data_t *pgdat)
2800 * With no special task, compact all zones so that a page of requested
2801 * order is allocatable.
2805 struct compact_control cc = {
2806 .order = pgdat->kcompactd_max_order,
2807 .search_order = pgdat->kcompactd_max_order,
2808 .highest_zoneidx = pgdat->kcompactd_highest_zoneidx,
2809 .mode = MIGRATE_SYNC_LIGHT,
2810 .ignore_skip_hint = false,
2811 .gfp_mask = GFP_KERNEL,
2813 trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
2814 cc.highest_zoneidx);
2815 count_compact_event(KCOMPACTD_WAKE);
2817 for (zoneid = 0; zoneid <= cc.highest_zoneidx; zoneid++) {
2820 zone = &pgdat->node_zones[zoneid];
2821 if (!populated_zone(zone))
2824 if (compaction_deferred(zone, cc.order))
2827 if (compaction_suitable(zone, cc.order, 0, zoneid) !=
2831 if (kthread_should_stop())
2835 status = compact_zone(&cc, NULL);
2837 if (status == COMPACT_SUCCESS) {
2838 compaction_defer_reset(zone, cc.order, false);
2839 } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
2841 * Buddy pages may become stranded on pcps that could
2842 * otherwise coalesce on the zone's free area for
2843 * order >= cc.order. This is ratelimited by the
2844 * upcoming deferral.
2846 drain_all_pages(zone);
2849 * We use sync migration mode here, so we defer like
2850 * sync direct compaction does.
2852 defer_compaction(zone, cc.order);
2855 count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
2856 cc.total_migrate_scanned);
2857 count_compact_events(KCOMPACTD_FREE_SCANNED,
2858 cc.total_free_scanned);
2860 VM_BUG_ON(!list_empty(&cc.freepages));
2861 VM_BUG_ON(!list_empty(&cc.migratepages));
2865 * Regardless of success, we are done until woken up next. But remember
2866 * the requested order/highest_zoneidx in case it was higher/tighter
2867 * than our current ones
2869 if (pgdat->kcompactd_max_order <= cc.order)
2870 pgdat->kcompactd_max_order = 0;
2871 if (pgdat->kcompactd_highest_zoneidx >= cc.highest_zoneidx)
2872 pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
2875 void wakeup_kcompactd(pg_data_t *pgdat, int order, int highest_zoneidx)
2880 if (pgdat->kcompactd_max_order < order)
2881 pgdat->kcompactd_max_order = order;
2883 if (pgdat->kcompactd_highest_zoneidx > highest_zoneidx)
2884 pgdat->kcompactd_highest_zoneidx = highest_zoneidx;
2887 * Pairs with implicit barrier in wait_event_freezable()
2888 * such that wakeups are not missed.
2890 if (!wq_has_sleeper(&pgdat->kcompactd_wait))
2893 if (!kcompactd_node_suitable(pgdat))
2896 trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
2898 wake_up_interruptible(&pgdat->kcompactd_wait);
2902 * The background compaction daemon, started as a kernel thread
2903 * from the init process.
2905 static int kcompactd(void *p)
2907 pg_data_t *pgdat = (pg_data_t *)p;
2908 struct task_struct *tsk = current;
2909 long default_timeout = msecs_to_jiffies(HPAGE_FRAG_CHECK_INTERVAL_MSEC);
2910 long timeout = default_timeout;
2912 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2914 if (!cpumask_empty(cpumask))
2915 set_cpus_allowed_ptr(tsk, cpumask);
2919 pgdat->kcompactd_max_order = 0;
2920 pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
2922 while (!kthread_should_stop()) {
2923 unsigned long pflags;
2926 * Avoid the unnecessary wakeup for proactive compaction
2927 * when it is disabled.
2929 if (!sysctl_compaction_proactiveness)
2930 timeout = MAX_SCHEDULE_TIMEOUT;
2931 trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
2932 if (wait_event_freezable_timeout(pgdat->kcompactd_wait,
2933 kcompactd_work_requested(pgdat), timeout) &&
2934 !pgdat->proactive_compact_trigger) {
2936 psi_memstall_enter(&pflags);
2937 kcompactd_do_work(pgdat);
2938 psi_memstall_leave(&pflags);
2940 * Reset the timeout value. The defer timeout from
2941 * proactive compaction is lost here but that is fine
2942 * as the condition of the zone changing substantionally
2943 * then carrying on with the previous defer interval is
2946 timeout = default_timeout;
2951 * Start the proactive work with default timeout. Based
2952 * on the fragmentation score, this timeout is updated.
2954 timeout = default_timeout;
2955 if (should_proactive_compact_node(pgdat)) {
2956 unsigned int prev_score, score;
2958 prev_score = fragmentation_score_node(pgdat);
2959 proactive_compact_node(pgdat);
2960 score = fragmentation_score_node(pgdat);
2962 * Defer proactive compaction if the fragmentation
2963 * score did not go down i.e. no progress made.
2965 if (unlikely(score >= prev_score))
2967 default_timeout << COMPACT_MAX_DEFER_SHIFT;
2969 if (unlikely(pgdat->proactive_compact_trigger))
2970 pgdat->proactive_compact_trigger = false;
2977 * This kcompactd start function will be called by init and node-hot-add.
2978 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
2980 int kcompactd_run(int nid)
2982 pg_data_t *pgdat = NODE_DATA(nid);
2985 if (pgdat->kcompactd)
2988 pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
2989 if (IS_ERR(pgdat->kcompactd)) {
2990 pr_err("Failed to start kcompactd on node %d\n", nid);
2991 ret = PTR_ERR(pgdat->kcompactd);
2992 pgdat->kcompactd = NULL;
2998 * Called by memory hotplug when all memory in a node is offlined. Caller must
2999 * hold mem_hotplug_begin/end().
3001 void kcompactd_stop(int nid)
3003 struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
3006 kthread_stop(kcompactd);
3007 NODE_DATA(nid)->kcompactd = NULL;
3012 * It's optimal to keep kcompactd on the same CPUs as their memory, but
3013 * not required for correctness. So if the last cpu in a node goes
3014 * away, we get changed to run anywhere: as the first one comes back,
3015 * restore their cpu bindings.
3017 static int kcompactd_cpu_online(unsigned int cpu)
3021 for_each_node_state(nid, N_MEMORY) {
3022 pg_data_t *pgdat = NODE_DATA(nid);
3023 const struct cpumask *mask;
3025 mask = cpumask_of_node(pgdat->node_id);
3027 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3028 /* One of our CPUs online: restore mask */
3029 set_cpus_allowed_ptr(pgdat->kcompactd, mask);
3034 static int __init kcompactd_init(void)
3039 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3040 "mm/compaction:online",
3041 kcompactd_cpu_online, NULL);
3043 pr_err("kcompactd: failed to register hotplug callbacks.\n");
3047 for_each_node_state(nid, N_MEMORY)
3051 subsys_initcall(kcompactd_init)
3053 #endif /* CONFIG_COMPACTION */