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>
27 #ifdef CONFIG_COMPACTION
28 static inline void count_compact_event(enum vm_event_item item)
33 static inline void count_compact_events(enum vm_event_item item, long delta)
35 count_vm_events(item, delta);
38 #define count_compact_event(item) do { } while (0)
39 #define count_compact_events(item, delta) do { } while (0)
42 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
44 #define CREATE_TRACE_POINTS
45 #include <trace/events/compaction.h>
47 #define block_start_pfn(pfn, order) round_down(pfn, 1UL << (order))
48 #define block_end_pfn(pfn, order) ALIGN((pfn) + 1, 1UL << (order))
49 #define pageblock_start_pfn(pfn) block_start_pfn(pfn, pageblock_order)
50 #define pageblock_end_pfn(pfn) block_end_pfn(pfn, pageblock_order)
52 static unsigned long release_freepages(struct list_head *freelist)
54 struct page *page, *next;
55 unsigned long high_pfn = 0;
57 list_for_each_entry_safe(page, next, freelist, lru) {
58 unsigned long pfn = page_to_pfn(page);
68 static void map_pages(struct list_head *list)
70 unsigned int i, order, nr_pages;
71 struct page *page, *next;
74 list_for_each_entry_safe(page, next, list, lru) {
77 order = page_private(page);
78 nr_pages = 1 << order;
80 post_alloc_hook(page, order, __GFP_MOVABLE);
82 split_page(page, order);
84 for (i = 0; i < nr_pages; i++) {
85 list_add(&page->lru, &tmp_list);
90 list_splice(&tmp_list, list);
93 #ifdef CONFIG_COMPACTION
95 int PageMovable(struct page *page)
97 struct address_space *mapping;
99 VM_BUG_ON_PAGE(!PageLocked(page), page);
100 if (!__PageMovable(page))
103 mapping = page_mapping(page);
104 if (mapping && mapping->a_ops && mapping->a_ops->isolate_page)
109 EXPORT_SYMBOL(PageMovable);
111 void __SetPageMovable(struct page *page, struct address_space *mapping)
113 VM_BUG_ON_PAGE(!PageLocked(page), page);
114 VM_BUG_ON_PAGE((unsigned long)mapping & PAGE_MAPPING_MOVABLE, page);
115 page->mapping = (void *)((unsigned long)mapping | PAGE_MAPPING_MOVABLE);
117 EXPORT_SYMBOL(__SetPageMovable);
119 void __ClearPageMovable(struct page *page)
121 VM_BUG_ON_PAGE(!PageLocked(page), page);
122 VM_BUG_ON_PAGE(!PageMovable(page), page);
124 * Clear registered address_space val with keeping PAGE_MAPPING_MOVABLE
125 * flag so that VM can catch up released page by driver after isolation.
126 * With it, VM migration doesn't try to put it back.
128 page->mapping = (void *)((unsigned long)page->mapping &
129 PAGE_MAPPING_MOVABLE);
131 EXPORT_SYMBOL(__ClearPageMovable);
133 /* Do not skip compaction more than 64 times */
134 #define COMPACT_MAX_DEFER_SHIFT 6
137 * Compaction is deferred when compaction fails to result in a page
138 * allocation success. 1 << compact_defer_limit compactions are skipped up
139 * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
141 void defer_compaction(struct zone *zone, int order)
143 zone->compact_considered = 0;
144 zone->compact_defer_shift++;
146 if (order < zone->compact_order_failed)
147 zone->compact_order_failed = order;
149 if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
150 zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
152 trace_mm_compaction_defer_compaction(zone, order);
155 /* Returns true if compaction should be skipped this time */
156 bool compaction_deferred(struct zone *zone, int order)
158 unsigned long defer_limit = 1UL << zone->compact_defer_shift;
160 if (order < zone->compact_order_failed)
163 /* Avoid possible overflow */
164 if (++zone->compact_considered > defer_limit)
165 zone->compact_considered = defer_limit;
167 if (zone->compact_considered >= defer_limit)
170 trace_mm_compaction_deferred(zone, order);
176 * Update defer tracking counters after successful compaction of given order,
177 * which means an allocation either succeeded (alloc_success == true) or is
178 * expected to succeed.
180 void compaction_defer_reset(struct zone *zone, int order,
184 zone->compact_considered = 0;
185 zone->compact_defer_shift = 0;
187 if (order >= zone->compact_order_failed)
188 zone->compact_order_failed = order + 1;
190 trace_mm_compaction_defer_reset(zone, order);
193 /* Returns true if restarting compaction after many failures */
194 bool compaction_restarting(struct zone *zone, int order)
196 if (order < zone->compact_order_failed)
199 return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
200 zone->compact_considered >= 1UL << zone->compact_defer_shift;
203 /* Returns true if the pageblock should be scanned for pages to isolate. */
204 static inline bool isolation_suitable(struct compact_control *cc,
207 if (cc->ignore_skip_hint)
210 return !get_pageblock_skip(page);
213 static void reset_cached_positions(struct zone *zone)
215 zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
216 zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
217 zone->compact_cached_free_pfn =
218 pageblock_start_pfn(zone_end_pfn(zone) - 1);
222 * Hugetlbfs pages should consistenly be skipped until updated by the hugetlb
223 * subsystem. It is always pointless to compact pages of pageblock_order and
224 * the free scanner can reconsider when no longer huge.
226 static bool pageblock_skip_persistent(struct page *page, unsigned int order)
230 if (order != pageblock_order)
236 * This function is called to clear all cached information on pageblocks that
237 * should be skipped for page isolation when the migrate and free page scanner
240 static void __reset_isolation_suitable(struct zone *zone)
242 unsigned long start_pfn = zone->zone_start_pfn;
243 unsigned long end_pfn = zone_end_pfn(zone);
246 zone->compact_blockskip_flush = false;
248 /* Walk the zone and mark every pageblock as suitable for isolation */
249 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
254 page = pfn_to_online_page(pfn);
257 if (zone != page_zone(page))
259 if (pageblock_skip_persistent(page, compound_order(page)))
262 clear_pageblock_skip(page);
265 reset_cached_positions(zone);
268 void reset_isolation_suitable(pg_data_t *pgdat)
272 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
273 struct zone *zone = &pgdat->node_zones[zoneid];
274 if (!populated_zone(zone))
277 /* Only flush if a full compaction finished recently */
278 if (zone->compact_blockskip_flush)
279 __reset_isolation_suitable(zone);
284 * If no pages were isolated then mark this pageblock to be skipped in the
285 * future. The information is later cleared by __reset_isolation_suitable().
287 static void update_pageblock_skip(struct compact_control *cc,
288 struct page *page, unsigned long nr_isolated,
289 bool migrate_scanner)
291 struct zone *zone = cc->zone;
294 if (cc->ignore_skip_hint)
303 set_pageblock_skip(page);
305 pfn = page_to_pfn(page);
307 /* Update where async and sync compaction should restart */
308 if (migrate_scanner) {
309 if (pfn > zone->compact_cached_migrate_pfn[0])
310 zone->compact_cached_migrate_pfn[0] = pfn;
311 if (cc->mode != MIGRATE_ASYNC &&
312 pfn > zone->compact_cached_migrate_pfn[1])
313 zone->compact_cached_migrate_pfn[1] = pfn;
315 if (pfn < zone->compact_cached_free_pfn)
316 zone->compact_cached_free_pfn = pfn;
320 static inline bool isolation_suitable(struct compact_control *cc,
326 static inline bool pageblock_skip_persistent(struct page *page,
332 static inline void update_pageblock_skip(struct compact_control *cc,
333 struct page *page, unsigned long nr_isolated,
334 bool migrate_scanner)
337 #endif /* CONFIG_COMPACTION */
340 * Compaction requires the taking of some coarse locks that are potentially
341 * very heavily contended. For async compaction, back out if the lock cannot
342 * be taken immediately. For sync compaction, spin on the lock if needed.
344 * Returns true if the lock is held
345 * Returns false if the lock is not held and compaction should abort
347 static bool compact_trylock_irqsave(spinlock_t *lock, unsigned long *flags,
348 struct compact_control *cc)
350 if (cc->mode == MIGRATE_ASYNC) {
351 if (!spin_trylock_irqsave(lock, *flags)) {
352 cc->contended = true;
356 spin_lock_irqsave(lock, *flags);
363 * Compaction requires the taking of some coarse locks that are potentially
364 * very heavily contended. The lock should be periodically unlocked to avoid
365 * having disabled IRQs for a long time, even when there is nobody waiting on
366 * the lock. It might also be that allowing the IRQs will result in
367 * need_resched() becoming true. If scheduling is needed, async compaction
368 * aborts. Sync compaction schedules.
369 * Either compaction type will also abort if a fatal signal is pending.
370 * In either case if the lock was locked, it is dropped and not regained.
372 * Returns true if compaction should abort due to fatal signal pending, or
373 * async compaction due to need_resched()
374 * Returns false when compaction can continue (sync compaction might have
377 static bool compact_unlock_should_abort(spinlock_t *lock,
378 unsigned long flags, bool *locked, struct compact_control *cc)
381 spin_unlock_irqrestore(lock, flags);
385 if (fatal_signal_pending(current)) {
386 cc->contended = true;
390 if (need_resched()) {
391 if (cc->mode == MIGRATE_ASYNC) {
392 cc->contended = true;
402 * Aside from avoiding lock contention, compaction also periodically checks
403 * need_resched() and either schedules in sync compaction or aborts async
404 * compaction. This is similar to what compact_unlock_should_abort() does, but
405 * is used where no lock is concerned.
407 * Returns false when no scheduling was needed, or sync compaction scheduled.
408 * Returns true when async compaction should abort.
410 static inline bool compact_should_abort(struct compact_control *cc)
412 /* async compaction aborts if contended */
413 if (need_resched()) {
414 if (cc->mode == MIGRATE_ASYNC) {
415 cc->contended = true;
426 * Isolate free pages onto a private freelist. If @strict is true, will abort
427 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
428 * (even though it may still end up isolating some pages).
430 static unsigned long isolate_freepages_block(struct compact_control *cc,
431 unsigned long *start_pfn,
432 unsigned long end_pfn,
433 struct list_head *freelist,
436 int nr_scanned = 0, total_isolated = 0;
437 struct page *cursor, *valid_page = NULL;
438 unsigned long flags = 0;
440 unsigned long blockpfn = *start_pfn;
443 cursor = pfn_to_page(blockpfn);
445 /* Isolate free pages. */
446 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
448 struct page *page = cursor;
451 * Periodically drop the lock (if held) regardless of its
452 * contention, to give chance to IRQs. Abort if fatal signal
453 * pending or async compaction detects need_resched()
455 if (!(blockpfn % SWAP_CLUSTER_MAX)
456 && compact_unlock_should_abort(&cc->zone->lock, flags,
461 if (!pfn_valid_within(blockpfn))
468 * For compound pages such as THP and hugetlbfs, we can save
469 * potentially a lot of iterations if we skip them at once.
470 * The check is racy, but we can consider only valid values
471 * and the only danger is skipping too much.
473 if (PageCompound(page)) {
474 const unsigned int order = compound_order(page);
476 if (pageblock_skip_persistent(page, order)) {
477 set_pageblock_skip(page);
479 } else if (likely(order < MAX_ORDER)) {
480 blockpfn += (1UL << order) - 1;
481 cursor += (1UL << order) - 1;
486 if (!PageBuddy(page))
490 * If we already hold the lock, we can skip some rechecking.
491 * Note that if we hold the lock now, checked_pageblock was
492 * already set in some previous iteration (or strict is true),
493 * so it is correct to skip the suitable migration target
498 * The zone lock must be held to isolate freepages.
499 * Unfortunately this is a very coarse lock and can be
500 * heavily contended if there are parallel allocations
501 * or parallel compactions. For async compaction do not
502 * spin on the lock and we acquire the lock as late as
505 locked = compact_trylock_irqsave(&cc->zone->lock,
510 /* Recheck this is a buddy page under lock */
511 if (!PageBuddy(page))
515 /* Found a free page, will break it into order-0 pages */
516 order = page_order(page);
517 isolated = __isolate_free_page(page, order);
520 set_page_private(page, order);
522 total_isolated += isolated;
523 cc->nr_freepages += isolated;
524 list_add_tail(&page->lru, freelist);
526 if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
527 blockpfn += isolated;
530 /* Advance to the end of split page */
531 blockpfn += isolated - 1;
532 cursor += isolated - 1;
544 spin_unlock_irqrestore(&cc->zone->lock, flags);
547 * There is a tiny chance that we have read bogus compound_order(),
548 * so be careful to not go outside of the pageblock.
550 if (unlikely(blockpfn > end_pfn))
553 trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
554 nr_scanned, total_isolated);
556 /* Record how far we have got within the block */
557 *start_pfn = blockpfn;
560 * If strict isolation is requested by CMA then check that all the
561 * pages requested were isolated. If there were any failures, 0 is
562 * returned and CMA will fail.
564 if (strict && blockpfn < end_pfn)
567 /* Update the pageblock-skip if the whole pageblock was scanned */
568 if (blockpfn == end_pfn)
569 update_pageblock_skip(cc, valid_page, total_isolated, false);
571 cc->total_free_scanned += nr_scanned;
573 count_compact_events(COMPACTISOLATED, total_isolated);
574 return total_isolated;
578 * isolate_freepages_range() - isolate free pages.
579 * @start_pfn: The first PFN to start isolating.
580 * @end_pfn: The one-past-last PFN.
582 * Non-free pages, invalid PFNs, or zone boundaries within the
583 * [start_pfn, end_pfn) range are considered errors, cause function to
584 * undo its actions and return zero.
586 * Otherwise, function returns one-past-the-last PFN of isolated page
587 * (which may be greater then end_pfn if end fell in a middle of
591 isolate_freepages_range(struct compact_control *cc,
592 unsigned long start_pfn, unsigned long end_pfn)
594 unsigned long isolated, pfn, block_start_pfn, block_end_pfn;
598 block_start_pfn = pageblock_start_pfn(pfn);
599 if (block_start_pfn < cc->zone->zone_start_pfn)
600 block_start_pfn = cc->zone->zone_start_pfn;
601 block_end_pfn = pageblock_end_pfn(pfn);
603 for (; pfn < end_pfn; pfn += isolated,
604 block_start_pfn = block_end_pfn,
605 block_end_pfn += pageblock_nr_pages) {
606 /* Protect pfn from changing by isolate_freepages_block */
607 unsigned long isolate_start_pfn = pfn;
609 block_end_pfn = min(block_end_pfn, end_pfn);
612 * pfn could pass the block_end_pfn if isolated freepage
613 * is more than pageblock order. In this case, we adjust
614 * scanning range to right one.
616 if (pfn >= block_end_pfn) {
617 block_start_pfn = pageblock_start_pfn(pfn);
618 block_end_pfn = pageblock_end_pfn(pfn);
619 block_end_pfn = min(block_end_pfn, end_pfn);
622 if (!pageblock_pfn_to_page(block_start_pfn,
623 block_end_pfn, cc->zone))
626 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
627 block_end_pfn, &freelist, true);
630 * In strict mode, isolate_freepages_block() returns 0 if
631 * there are any holes in the block (ie. invalid PFNs or
638 * If we managed to isolate pages, it is always (1 << n) *
639 * pageblock_nr_pages for some non-negative n. (Max order
640 * page may span two pageblocks).
644 /* __isolate_free_page() does not map the pages */
645 map_pages(&freelist);
648 /* Loop terminated early, cleanup. */
649 release_freepages(&freelist);
653 /* We don't use freelists for anything. */
657 /* Similar to reclaim, but different enough that they don't share logic */
658 static bool too_many_isolated(struct zone *zone)
660 unsigned long active, inactive, isolated;
662 inactive = node_page_state(zone->zone_pgdat, NR_INACTIVE_FILE) +
663 node_page_state(zone->zone_pgdat, NR_INACTIVE_ANON);
664 active = node_page_state(zone->zone_pgdat, NR_ACTIVE_FILE) +
665 node_page_state(zone->zone_pgdat, NR_ACTIVE_ANON);
666 isolated = node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE) +
667 node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON);
669 return isolated > (inactive + active) / 2;
673 * isolate_migratepages_block() - isolate all migrate-able pages within
675 * @cc: Compaction control structure.
676 * @low_pfn: The first PFN to isolate
677 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
678 * @isolate_mode: Isolation mode to be used.
680 * Isolate all pages that can be migrated from the range specified by
681 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
682 * Returns zero if there is a fatal signal pending, otherwise PFN of the
683 * first page that was not scanned (which may be both less, equal to or more
686 * The pages are isolated on cc->migratepages list (not required to be empty),
687 * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
688 * is neither read nor updated.
691 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
692 unsigned long end_pfn, isolate_mode_t isolate_mode)
694 struct zone *zone = cc->zone;
695 unsigned long nr_scanned = 0, nr_isolated = 0;
696 struct lruvec *lruvec;
697 unsigned long flags = 0;
699 struct page *page = NULL, *valid_page = NULL;
700 unsigned long start_pfn = low_pfn;
701 bool skip_on_failure = false;
702 unsigned long next_skip_pfn = 0;
705 * Ensure that there are not too many pages isolated from the LRU
706 * list by either parallel reclaimers or compaction. If there are,
707 * delay for some time until fewer pages are isolated
709 while (unlikely(too_many_isolated(zone))) {
710 /* async migration should just abort */
711 if (cc->mode == MIGRATE_ASYNC)
714 congestion_wait(BLK_RW_ASYNC, HZ/10);
716 if (fatal_signal_pending(current))
720 if (compact_should_abort(cc))
723 if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
724 skip_on_failure = true;
725 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
728 /* Time to isolate some pages for migration */
729 for (; low_pfn < end_pfn; low_pfn++) {
731 if (skip_on_failure && low_pfn >= next_skip_pfn) {
733 * We have isolated all migration candidates in the
734 * previous order-aligned block, and did not skip it due
735 * to failure. We should migrate the pages now and
736 * hopefully succeed compaction.
742 * We failed to isolate in the previous order-aligned
743 * block. Set the new boundary to the end of the
744 * current block. Note we can't simply increase
745 * next_skip_pfn by 1 << order, as low_pfn might have
746 * been incremented by a higher number due to skipping
747 * a compound or a high-order buddy page in the
748 * previous loop iteration.
750 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
754 * Periodically drop the lock (if held) regardless of its
755 * contention, to give chance to IRQs. Abort async compaction
758 if (!(low_pfn % SWAP_CLUSTER_MAX)
759 && compact_unlock_should_abort(zone_lru_lock(zone), flags,
763 if (!pfn_valid_within(low_pfn))
767 page = pfn_to_page(low_pfn);
773 * Skip if free. We read page order here without zone lock
774 * which is generally unsafe, but the race window is small and
775 * the worst thing that can happen is that we skip some
776 * potential isolation targets.
778 if (PageBuddy(page)) {
779 unsigned long freepage_order = page_order_unsafe(page);
782 * Without lock, we cannot be sure that what we got is
783 * a valid page order. Consider only values in the
784 * valid order range to prevent low_pfn overflow.
786 if (freepage_order > 0 && freepage_order < MAX_ORDER)
787 low_pfn += (1UL << freepage_order) - 1;
792 * Regardless of being on LRU, compound pages such as THP and
793 * hugetlbfs are not to be compacted. We can potentially save
794 * a lot of iterations if we skip them at once. The check is
795 * racy, but we can consider only valid values and the only
796 * danger is skipping too much.
798 if (PageCompound(page)) {
799 const unsigned int order = compound_order(page);
801 if (pageblock_skip_persistent(page, order)) {
802 set_pageblock_skip(page);
804 } else if (likely(order < MAX_ORDER))
805 low_pfn += (1UL << order) - 1;
810 * Check may be lockless but that's ok as we recheck later.
811 * It's possible to migrate LRU and non-lru movable pages.
812 * Skip any other type of page
814 if (!PageLRU(page)) {
816 * __PageMovable can return false positive so we need
817 * to verify it under page_lock.
819 if (unlikely(__PageMovable(page)) &&
820 !PageIsolated(page)) {
822 spin_unlock_irqrestore(zone_lru_lock(zone),
827 if (!isolate_movable_page(page, isolate_mode))
828 goto isolate_success;
835 * Migration will fail if an anonymous page is pinned in memory,
836 * so avoid taking lru_lock and isolating it unnecessarily in an
837 * admittedly racy check.
839 if (!page_mapping(page) &&
840 page_count(page) > page_mapcount(page))
844 * Only allow to migrate anonymous pages in GFP_NOFS context
845 * because those do not depend on fs locks.
847 if (!(cc->gfp_mask & __GFP_FS) && page_mapping(page))
850 /* If we already hold the lock, we can skip some rechecking */
852 locked = compact_trylock_irqsave(zone_lru_lock(zone),
857 /* Recheck PageLRU and PageCompound under lock */
862 * Page become compound since the non-locked check,
863 * and it's on LRU. It can only be a THP so the order
864 * is safe to read and it's 0 for tail pages.
866 if (unlikely(PageCompound(page))) {
867 const unsigned int order = compound_order(page);
869 if (pageblock_skip_persistent(page, order)) {
870 set_pageblock_skip(page);
873 low_pfn += (1UL << order) - 1;
878 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
880 /* Try isolate the page */
881 if (__isolate_lru_page(page, isolate_mode) != 0)
884 VM_BUG_ON_PAGE(PageCompound(page), page);
886 /* Successfully isolated */
887 del_page_from_lru_list(page, lruvec, page_lru(page));
888 inc_node_page_state(page,
889 NR_ISOLATED_ANON + page_is_file_cache(page));
892 list_add(&page->lru, &cc->migratepages);
893 cc->nr_migratepages++;
897 * Record where we could have freed pages by migration and not
898 * yet flushed them to buddy allocator.
899 * - this is the lowest page that was isolated and likely be
900 * then freed by migration.
902 if (!cc->last_migrated_pfn)
903 cc->last_migrated_pfn = low_pfn;
905 /* Avoid isolating too much */
906 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
913 if (!skip_on_failure)
917 * We have isolated some pages, but then failed. Release them
918 * instead of migrating, as we cannot form the cc->order buddy
923 spin_unlock_irqrestore(zone_lru_lock(zone), flags);
926 putback_movable_pages(&cc->migratepages);
927 cc->nr_migratepages = 0;
928 cc->last_migrated_pfn = 0;
932 if (low_pfn < next_skip_pfn) {
933 low_pfn = next_skip_pfn - 1;
935 * The check near the loop beginning would have updated
936 * next_skip_pfn too, but this is a bit simpler.
938 next_skip_pfn += 1UL << cc->order;
943 * The PageBuddy() check could have potentially brought us outside
944 * the range to be scanned.
946 if (unlikely(low_pfn > end_pfn))
950 spin_unlock_irqrestore(zone_lru_lock(zone), flags);
953 * Update the pageblock-skip information and cached scanner pfn,
954 * if the whole pageblock was scanned without isolating any page.
956 if (low_pfn == end_pfn)
957 update_pageblock_skip(cc, valid_page, nr_isolated, true);
959 trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
960 nr_scanned, nr_isolated);
962 cc->total_migrate_scanned += nr_scanned;
964 count_compact_events(COMPACTISOLATED, nr_isolated);
970 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
971 * @cc: Compaction control structure.
972 * @start_pfn: The first PFN to start isolating.
973 * @end_pfn: The one-past-last PFN.
975 * Returns zero if isolation fails fatally due to e.g. pending signal.
976 * Otherwise, function returns one-past-the-last PFN of isolated page
977 * (which may be greater than end_pfn if end fell in a middle of a THP page).
980 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
981 unsigned long end_pfn)
983 unsigned long pfn, block_start_pfn, block_end_pfn;
985 /* Scan block by block. First and last block may be incomplete */
987 block_start_pfn = pageblock_start_pfn(pfn);
988 if (block_start_pfn < cc->zone->zone_start_pfn)
989 block_start_pfn = cc->zone->zone_start_pfn;
990 block_end_pfn = pageblock_end_pfn(pfn);
992 for (; pfn < end_pfn; pfn = block_end_pfn,
993 block_start_pfn = block_end_pfn,
994 block_end_pfn += pageblock_nr_pages) {
996 block_end_pfn = min(block_end_pfn, end_pfn);
998 if (!pageblock_pfn_to_page(block_start_pfn,
999 block_end_pfn, cc->zone))
1002 pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
1003 ISOLATE_UNEVICTABLE);
1008 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
1015 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
1016 #ifdef CONFIG_COMPACTION
1018 static bool suitable_migration_source(struct compact_control *cc,
1023 if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction)
1026 block_mt = get_pageblock_migratetype(page);
1028 if (cc->migratetype == MIGRATE_MOVABLE)
1029 return is_migrate_movable(block_mt);
1031 return block_mt == cc->migratetype;
1034 /* Returns true if the page is within a block suitable for migration to */
1035 static bool suitable_migration_target(struct compact_control *cc,
1038 /* If the page is a large free page, then disallow migration */
1039 if (PageBuddy(page)) {
1041 * We are checking page_order without zone->lock taken. But
1042 * the only small danger is that we skip a potentially suitable
1043 * pageblock, so it's not worth to check order for valid range.
1045 if (page_order_unsafe(page) >= pageblock_order)
1049 if (cc->ignore_block_suitable)
1052 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1053 if (is_migrate_movable(get_pageblock_migratetype(page)))
1056 /* Otherwise skip the block */
1061 * Test whether the free scanner has reached the same or lower pageblock than
1062 * the migration scanner, and compaction should thus terminate.
1064 static inline bool compact_scanners_met(struct compact_control *cc)
1066 return (cc->free_pfn >> pageblock_order)
1067 <= (cc->migrate_pfn >> pageblock_order);
1071 * Based on information in the current compact_control, find blocks
1072 * suitable for isolating free pages from and then isolate them.
1074 static void isolate_freepages(struct compact_control *cc)
1076 struct zone *zone = cc->zone;
1078 unsigned long block_start_pfn; /* start of current pageblock */
1079 unsigned long isolate_start_pfn; /* exact pfn we start at */
1080 unsigned long block_end_pfn; /* end of current pageblock */
1081 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
1082 struct list_head *freelist = &cc->freepages;
1085 * Initialise the free scanner. The starting point is where we last
1086 * successfully isolated from, zone-cached value, or the end of the
1087 * zone when isolating for the first time. For looping we also need
1088 * this pfn aligned down to the pageblock boundary, because we do
1089 * block_start_pfn -= pageblock_nr_pages in the for loop.
1090 * For ending point, take care when isolating in last pageblock of a
1091 * a zone which ends in the middle of a pageblock.
1092 * The low boundary is the end of the pageblock the migration scanner
1095 isolate_start_pfn = cc->free_pfn;
1096 block_start_pfn = pageblock_start_pfn(cc->free_pfn);
1097 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1098 zone_end_pfn(zone));
1099 low_pfn = pageblock_end_pfn(cc->migrate_pfn);
1102 * Isolate free pages until enough are available to migrate the
1103 * pages on cc->migratepages. We stop searching if the migrate
1104 * and free page scanners meet or enough free pages are isolated.
1106 for (; block_start_pfn >= low_pfn;
1107 block_end_pfn = block_start_pfn,
1108 block_start_pfn -= pageblock_nr_pages,
1109 isolate_start_pfn = block_start_pfn) {
1111 * This can iterate a massively long zone without finding any
1112 * suitable migration targets, so periodically check if we need
1113 * to schedule, or even abort async compaction.
1115 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1116 && compact_should_abort(cc))
1119 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1124 /* Check the block is suitable for migration */
1125 if (!suitable_migration_target(cc, page))
1128 /* If isolation recently failed, do not retry */
1129 if (!isolation_suitable(cc, page))
1132 /* Found a block suitable for isolating free pages from. */
1133 isolate_freepages_block(cc, &isolate_start_pfn, block_end_pfn,
1137 * If we isolated enough freepages, or aborted due to lock
1138 * contention, terminate.
1140 if ((cc->nr_freepages >= cc->nr_migratepages)
1142 if (isolate_start_pfn >= block_end_pfn) {
1144 * Restart at previous pageblock if more
1145 * freepages can be isolated next time.
1148 block_start_pfn - pageblock_nr_pages;
1151 } else if (isolate_start_pfn < block_end_pfn) {
1153 * If isolation failed early, do not continue
1160 /* __isolate_free_page() does not map the pages */
1161 map_pages(freelist);
1164 * Record where the free scanner will restart next time. Either we
1165 * broke from the loop and set isolate_start_pfn based on the last
1166 * call to isolate_freepages_block(), or we met the migration scanner
1167 * and the loop terminated due to isolate_start_pfn < low_pfn
1169 cc->free_pfn = isolate_start_pfn;
1173 * This is a migrate-callback that "allocates" freepages by taking pages
1174 * from the isolated freelists in the block we are migrating to.
1176 static struct page *compaction_alloc(struct page *migratepage,
1180 struct compact_control *cc = (struct compact_control *)data;
1181 struct page *freepage;
1184 * Isolate free pages if necessary, and if we are not aborting due to
1187 if (list_empty(&cc->freepages)) {
1189 isolate_freepages(cc);
1191 if (list_empty(&cc->freepages))
1195 freepage = list_entry(cc->freepages.next, struct page, lru);
1196 list_del(&freepage->lru);
1203 * This is a migrate-callback that "frees" freepages back to the isolated
1204 * freelist. All pages on the freelist are from the same zone, so there is no
1205 * special handling needed for NUMA.
1207 static void compaction_free(struct page *page, unsigned long data)
1209 struct compact_control *cc = (struct compact_control *)data;
1211 list_add(&page->lru, &cc->freepages);
1215 /* possible outcome of isolate_migratepages */
1217 ISOLATE_ABORT, /* Abort compaction now */
1218 ISOLATE_NONE, /* No pages isolated, continue scanning */
1219 ISOLATE_SUCCESS, /* Pages isolated, migrate */
1220 } isolate_migrate_t;
1223 * Allow userspace to control policy on scanning the unevictable LRU for
1224 * compactable pages.
1226 int sysctl_compact_unevictable_allowed __read_mostly = 1;
1229 * Isolate all pages that can be migrated from the first suitable block,
1230 * starting at the block pointed to by the migrate scanner pfn within
1233 static isolate_migrate_t isolate_migratepages(struct zone *zone,
1234 struct compact_control *cc)
1236 unsigned long block_start_pfn;
1237 unsigned long block_end_pfn;
1238 unsigned long low_pfn;
1240 const isolate_mode_t isolate_mode =
1241 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1242 (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1245 * Start at where we last stopped, or beginning of the zone as
1246 * initialized by compact_zone()
1248 low_pfn = cc->migrate_pfn;
1249 block_start_pfn = pageblock_start_pfn(low_pfn);
1250 if (block_start_pfn < zone->zone_start_pfn)
1251 block_start_pfn = zone->zone_start_pfn;
1253 /* Only scan within a pageblock boundary */
1254 block_end_pfn = pageblock_end_pfn(low_pfn);
1257 * Iterate over whole pageblocks until we find the first suitable.
1258 * Do not cross the free scanner.
1260 for (; block_end_pfn <= cc->free_pfn;
1261 low_pfn = block_end_pfn,
1262 block_start_pfn = block_end_pfn,
1263 block_end_pfn += pageblock_nr_pages) {
1266 * This can potentially iterate a massively long zone with
1267 * many pageblocks unsuitable, so periodically check if we
1268 * need to schedule, or even abort async compaction.
1270 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1271 && compact_should_abort(cc))
1274 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1279 /* If isolation recently failed, do not retry */
1280 if (!isolation_suitable(cc, page))
1284 * For async compaction, also only scan in MOVABLE blocks.
1285 * Async compaction is optimistic to see if the minimum amount
1286 * of work satisfies the allocation.
1288 if (!suitable_migration_source(cc, page))
1291 /* Perform the isolation */
1292 low_pfn = isolate_migratepages_block(cc, low_pfn,
1293 block_end_pfn, isolate_mode);
1295 if (!low_pfn || cc->contended)
1296 return ISOLATE_ABORT;
1299 * Either we isolated something and proceed with migration. Or
1300 * we failed and compact_zone should decide if we should
1306 /* Record where migration scanner will be restarted. */
1307 cc->migrate_pfn = low_pfn;
1309 return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
1313 * order == -1 is expected when compacting via
1314 * /proc/sys/vm/compact_memory
1316 static inline bool is_via_compact_memory(int order)
1321 static enum compact_result __compact_finished(struct zone *zone,
1322 struct compact_control *cc)
1325 const int migratetype = cc->migratetype;
1327 if (cc->contended || fatal_signal_pending(current))
1328 return COMPACT_CONTENDED;
1330 /* Compaction run completes if the migrate and free scanner meet */
1331 if (compact_scanners_met(cc)) {
1332 /* Let the next compaction start anew. */
1333 reset_cached_positions(zone);
1336 * Mark that the PG_migrate_skip information should be cleared
1337 * by kswapd when it goes to sleep. kcompactd does not set the
1338 * flag itself as the decision to be clear should be directly
1339 * based on an allocation request.
1341 if (cc->direct_compaction)
1342 zone->compact_blockskip_flush = true;
1345 return COMPACT_COMPLETE;
1347 return COMPACT_PARTIAL_SKIPPED;
1350 if (is_via_compact_memory(cc->order))
1351 return COMPACT_CONTINUE;
1353 if (cc->finishing_block) {
1355 * We have finished the pageblock, but better check again that
1356 * we really succeeded.
1358 if (IS_ALIGNED(cc->migrate_pfn, pageblock_nr_pages))
1359 cc->finishing_block = false;
1361 return COMPACT_CONTINUE;
1364 /* Direct compactor: Is a suitable page free? */
1365 for (order = cc->order; order < MAX_ORDER; order++) {
1366 struct free_area *area = &zone->free_area[order];
1369 /* Job done if page is free of the right migratetype */
1370 if (!list_empty(&area->free_list[migratetype]))
1371 return COMPACT_SUCCESS;
1374 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
1375 if (migratetype == MIGRATE_MOVABLE &&
1376 !list_empty(&area->free_list[MIGRATE_CMA]))
1377 return COMPACT_SUCCESS;
1380 * Job done if allocation would steal freepages from
1381 * other migratetype buddy lists.
1383 if (find_suitable_fallback(area, order, migratetype,
1384 true, &can_steal) != -1) {
1386 /* movable pages are OK in any pageblock */
1387 if (migratetype == MIGRATE_MOVABLE)
1388 return COMPACT_SUCCESS;
1391 * We are stealing for a non-movable allocation. Make
1392 * sure we finish compacting the current pageblock
1393 * first so it is as free as possible and we won't
1394 * have to steal another one soon. This only applies
1395 * to sync compaction, as async compaction operates
1396 * on pageblocks of the same migratetype.
1398 if (cc->mode == MIGRATE_ASYNC ||
1399 IS_ALIGNED(cc->migrate_pfn,
1400 pageblock_nr_pages)) {
1401 return COMPACT_SUCCESS;
1404 cc->finishing_block = true;
1405 return COMPACT_CONTINUE;
1409 return COMPACT_NO_SUITABLE_PAGE;
1412 static enum compact_result compact_finished(struct zone *zone,
1413 struct compact_control *cc)
1417 ret = __compact_finished(zone, cc);
1418 trace_mm_compaction_finished(zone, cc->order, ret);
1419 if (ret == COMPACT_NO_SUITABLE_PAGE)
1420 ret = COMPACT_CONTINUE;
1426 * compaction_suitable: Is this suitable to run compaction on this zone now?
1428 * COMPACT_SKIPPED - If there are too few free pages for compaction
1429 * COMPACT_SUCCESS - If the allocation would succeed without compaction
1430 * COMPACT_CONTINUE - If compaction should run now
1432 static enum compact_result __compaction_suitable(struct zone *zone, int order,
1433 unsigned int alloc_flags,
1435 unsigned long wmark_target)
1437 unsigned long watermark;
1439 if (is_via_compact_memory(order))
1440 return COMPACT_CONTINUE;
1442 watermark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1444 * If watermarks for high-order allocation are already met, there
1445 * should be no need for compaction at all.
1447 if (zone_watermark_ok(zone, order, watermark, classzone_idx,
1449 return COMPACT_SUCCESS;
1452 * Watermarks for order-0 must be met for compaction to be able to
1453 * isolate free pages for migration targets. This means that the
1454 * watermark and alloc_flags have to match, or be more pessimistic than
1455 * the check in __isolate_free_page(). We don't use the direct
1456 * compactor's alloc_flags, as they are not relevant for freepage
1457 * isolation. We however do use the direct compactor's classzone_idx to
1458 * skip over zones where lowmem reserves would prevent allocation even
1459 * if compaction succeeds.
1460 * For costly orders, we require low watermark instead of min for
1461 * compaction to proceed to increase its chances.
1462 * ALLOC_CMA is used, as pages in CMA pageblocks are considered
1463 * suitable migration targets
1465 watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
1466 low_wmark_pages(zone) : min_wmark_pages(zone);
1467 watermark += compact_gap(order);
1468 if (!__zone_watermark_ok(zone, 0, watermark, classzone_idx,
1469 ALLOC_CMA, wmark_target))
1470 return COMPACT_SKIPPED;
1472 return COMPACT_CONTINUE;
1475 enum compact_result compaction_suitable(struct zone *zone, int order,
1476 unsigned int alloc_flags,
1479 enum compact_result ret;
1482 ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx,
1483 zone_page_state(zone, NR_FREE_PAGES));
1485 * fragmentation index determines if allocation failures are due to
1486 * low memory or external fragmentation
1488 * index of -1000 would imply allocations might succeed depending on
1489 * watermarks, but we already failed the high-order watermark check
1490 * index towards 0 implies failure is due to lack of memory
1491 * index towards 1000 implies failure is due to fragmentation
1493 * Only compact if a failure would be due to fragmentation. Also
1494 * ignore fragindex for non-costly orders where the alternative to
1495 * a successful reclaim/compaction is OOM. Fragindex and the
1496 * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
1497 * excessive compaction for costly orders, but it should not be at the
1498 * expense of system stability.
1500 if (ret == COMPACT_CONTINUE && (order > PAGE_ALLOC_COSTLY_ORDER)) {
1501 fragindex = fragmentation_index(zone, order);
1502 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
1503 ret = COMPACT_NOT_SUITABLE_ZONE;
1506 trace_mm_compaction_suitable(zone, order, ret);
1507 if (ret == COMPACT_NOT_SUITABLE_ZONE)
1508 ret = COMPACT_SKIPPED;
1513 bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
1520 * Make sure at least one zone would pass __compaction_suitable if we continue
1521 * retrying the reclaim.
1523 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1525 unsigned long available;
1526 enum compact_result compact_result;
1529 * Do not consider all the reclaimable memory because we do not
1530 * want to trash just for a single high order allocation which
1531 * is even not guaranteed to appear even if __compaction_suitable
1532 * is happy about the watermark check.
1534 available = zone_reclaimable_pages(zone) / order;
1535 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
1536 compact_result = __compaction_suitable(zone, order, alloc_flags,
1537 ac_classzone_idx(ac), available);
1538 if (compact_result != COMPACT_SKIPPED)
1545 static enum compact_result compact_zone(struct zone *zone, struct compact_control *cc)
1547 enum compact_result ret;
1548 unsigned long start_pfn = zone->zone_start_pfn;
1549 unsigned long end_pfn = zone_end_pfn(zone);
1550 const bool sync = cc->mode != MIGRATE_ASYNC;
1552 cc->migratetype = gfpflags_to_migratetype(cc->gfp_mask);
1553 ret = compaction_suitable(zone, cc->order, cc->alloc_flags,
1555 /* Compaction is likely to fail */
1556 if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED)
1559 /* huh, compaction_suitable is returning something unexpected */
1560 VM_BUG_ON(ret != COMPACT_CONTINUE);
1563 * Clear pageblock skip if there were failures recently and compaction
1564 * is about to be retried after being deferred.
1566 if (compaction_restarting(zone, cc->order))
1567 __reset_isolation_suitable(zone);
1570 * Setup to move all movable pages to the end of the zone. Used cached
1571 * information on where the scanners should start (unless we explicitly
1572 * want to compact the whole zone), but check that it is initialised
1573 * by ensuring the values are within zone boundaries.
1575 if (cc->whole_zone) {
1576 cc->migrate_pfn = start_pfn;
1577 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
1579 cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync];
1580 cc->free_pfn = zone->compact_cached_free_pfn;
1581 if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
1582 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
1583 zone->compact_cached_free_pfn = cc->free_pfn;
1585 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
1586 cc->migrate_pfn = start_pfn;
1587 zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
1588 zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
1591 if (cc->migrate_pfn == start_pfn)
1592 cc->whole_zone = true;
1595 cc->last_migrated_pfn = 0;
1597 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
1598 cc->free_pfn, end_pfn, sync);
1600 migrate_prep_local();
1602 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
1605 switch (isolate_migratepages(zone, cc)) {
1607 ret = COMPACT_CONTENDED;
1608 putback_movable_pages(&cc->migratepages);
1609 cc->nr_migratepages = 0;
1613 * We haven't isolated and migrated anything, but
1614 * there might still be unflushed migrations from
1615 * previous cc->order aligned block.
1618 case ISOLATE_SUCCESS:
1622 err = migrate_pages(&cc->migratepages, compaction_alloc,
1623 compaction_free, (unsigned long)cc, cc->mode,
1626 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
1629 /* All pages were either migrated or will be released */
1630 cc->nr_migratepages = 0;
1632 putback_movable_pages(&cc->migratepages);
1634 * migrate_pages() may return -ENOMEM when scanners meet
1635 * and we want compact_finished() to detect it
1637 if (err == -ENOMEM && !compact_scanners_met(cc)) {
1638 ret = COMPACT_CONTENDED;
1642 * We failed to migrate at least one page in the current
1643 * order-aligned block, so skip the rest of it.
1645 if (cc->direct_compaction &&
1646 (cc->mode == MIGRATE_ASYNC)) {
1647 cc->migrate_pfn = block_end_pfn(
1648 cc->migrate_pfn - 1, cc->order);
1649 /* Draining pcplists is useless in this case */
1650 cc->last_migrated_pfn = 0;
1657 * Has the migration scanner moved away from the previous
1658 * cc->order aligned block where we migrated from? If yes,
1659 * flush the pages that were freed, so that they can merge and
1660 * compact_finished() can detect immediately if allocation
1663 if (cc->order > 0 && cc->last_migrated_pfn) {
1665 unsigned long current_block_start =
1666 block_start_pfn(cc->migrate_pfn, cc->order);
1668 if (cc->last_migrated_pfn < current_block_start) {
1670 lru_add_drain_cpu(cpu);
1671 drain_local_pages(zone);
1673 /* No more flushing until we migrate again */
1674 cc->last_migrated_pfn = 0;
1682 * Release free pages and update where the free scanner should restart,
1683 * so we don't leave any returned pages behind in the next attempt.
1685 if (cc->nr_freepages > 0) {
1686 unsigned long free_pfn = release_freepages(&cc->freepages);
1688 cc->nr_freepages = 0;
1689 VM_BUG_ON(free_pfn == 0);
1690 /* The cached pfn is always the first in a pageblock */
1691 free_pfn = pageblock_start_pfn(free_pfn);
1693 * Only go back, not forward. The cached pfn might have been
1694 * already reset to zone end in compact_finished()
1696 if (free_pfn > zone->compact_cached_free_pfn)
1697 zone->compact_cached_free_pfn = free_pfn;
1700 count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned);
1701 count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned);
1703 trace_mm_compaction_end(start_pfn, cc->migrate_pfn,
1704 cc->free_pfn, end_pfn, sync, ret);
1709 static enum compact_result compact_zone_order(struct zone *zone, int order,
1710 gfp_t gfp_mask, enum compact_priority prio,
1711 unsigned int alloc_flags, int classzone_idx)
1713 enum compact_result ret;
1714 struct compact_control cc = {
1716 .nr_migratepages = 0,
1717 .total_migrate_scanned = 0,
1718 .total_free_scanned = 0,
1720 .gfp_mask = gfp_mask,
1722 .mode = (prio == COMPACT_PRIO_ASYNC) ?
1723 MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT,
1724 .alloc_flags = alloc_flags,
1725 .classzone_idx = classzone_idx,
1726 .direct_compaction = true,
1727 .whole_zone = (prio == MIN_COMPACT_PRIORITY),
1728 .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
1729 .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
1731 INIT_LIST_HEAD(&cc.freepages);
1732 INIT_LIST_HEAD(&cc.migratepages);
1734 ret = compact_zone(zone, &cc);
1736 VM_BUG_ON(!list_empty(&cc.freepages));
1737 VM_BUG_ON(!list_empty(&cc.migratepages));
1742 int sysctl_extfrag_threshold = 500;
1745 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1746 * @gfp_mask: The GFP mask of the current allocation
1747 * @order: The order of the current allocation
1748 * @alloc_flags: The allocation flags of the current allocation
1749 * @ac: The context of current allocation
1750 * @mode: The migration mode for async, sync light, or sync migration
1752 * This is the main entry point for direct page compaction.
1754 enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
1755 unsigned int alloc_flags, const struct alloc_context *ac,
1756 enum compact_priority prio)
1758 int may_perform_io = gfp_mask & __GFP_IO;
1761 enum compact_result rc = COMPACT_SKIPPED;
1764 * Check if the GFP flags allow compaction - GFP_NOIO is really
1765 * tricky context because the migration might require IO
1767 if (!may_perform_io)
1768 return COMPACT_SKIPPED;
1770 trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
1772 /* Compact each zone in the list */
1773 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1775 enum compact_result status;
1777 if (prio > MIN_COMPACT_PRIORITY
1778 && compaction_deferred(zone, order)) {
1779 rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
1783 status = compact_zone_order(zone, order, gfp_mask, prio,
1784 alloc_flags, ac_classzone_idx(ac));
1785 rc = max(status, rc);
1787 /* The allocation should succeed, stop compacting */
1788 if (status == COMPACT_SUCCESS) {
1790 * We think the allocation will succeed in this zone,
1791 * but it is not certain, hence the false. The caller
1792 * will repeat this with true if allocation indeed
1793 * succeeds in this zone.
1795 compaction_defer_reset(zone, order, false);
1800 if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
1801 status == COMPACT_PARTIAL_SKIPPED))
1803 * We think that allocation won't succeed in this zone
1804 * so we defer compaction there. If it ends up
1805 * succeeding after all, it will be reset.
1807 defer_compaction(zone, order);
1810 * We might have stopped compacting due to need_resched() in
1811 * async compaction, or due to a fatal signal detected. In that
1812 * case do not try further zones
1814 if ((prio == COMPACT_PRIO_ASYNC && need_resched())
1815 || fatal_signal_pending(current))
1823 /* Compact all zones within a node */
1824 static void compact_node(int nid)
1826 pg_data_t *pgdat = NODE_DATA(nid);
1829 struct compact_control cc = {
1831 .total_migrate_scanned = 0,
1832 .total_free_scanned = 0,
1833 .mode = MIGRATE_SYNC,
1834 .ignore_skip_hint = true,
1836 .gfp_mask = GFP_KERNEL,
1840 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1842 zone = &pgdat->node_zones[zoneid];
1843 if (!populated_zone(zone))
1846 cc.nr_freepages = 0;
1847 cc.nr_migratepages = 0;
1849 INIT_LIST_HEAD(&cc.freepages);
1850 INIT_LIST_HEAD(&cc.migratepages);
1852 compact_zone(zone, &cc);
1854 VM_BUG_ON(!list_empty(&cc.freepages));
1855 VM_BUG_ON(!list_empty(&cc.migratepages));
1859 /* Compact all nodes in the system */
1860 static void compact_nodes(void)
1864 /* Flush pending updates to the LRU lists */
1865 lru_add_drain_all();
1867 for_each_online_node(nid)
1871 /* The written value is actually unused, all memory is compacted */
1872 int sysctl_compact_memory;
1875 * This is the entry point for compacting all nodes via
1876 * /proc/sys/vm/compact_memory
1878 int sysctl_compaction_handler(struct ctl_table *table, int write,
1879 void __user *buffer, size_t *length, loff_t *ppos)
1887 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1888 void __user *buffer, size_t *length, loff_t *ppos)
1890 proc_dointvec_minmax(table, write, buffer, length, ppos);
1895 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1896 static ssize_t sysfs_compact_node(struct device *dev,
1897 struct device_attribute *attr,
1898 const char *buf, size_t count)
1902 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1903 /* Flush pending updates to the LRU lists */
1904 lru_add_drain_all();
1911 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1913 int compaction_register_node(struct node *node)
1915 return device_create_file(&node->dev, &dev_attr_compact);
1918 void compaction_unregister_node(struct node *node)
1920 return device_remove_file(&node->dev, &dev_attr_compact);
1922 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1924 static inline bool kcompactd_work_requested(pg_data_t *pgdat)
1926 return pgdat->kcompactd_max_order > 0 || kthread_should_stop();
1929 static bool kcompactd_node_suitable(pg_data_t *pgdat)
1933 enum zone_type classzone_idx = pgdat->kcompactd_classzone_idx;
1935 for (zoneid = 0; zoneid <= classzone_idx; zoneid++) {
1936 zone = &pgdat->node_zones[zoneid];
1938 if (!populated_zone(zone))
1941 if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0,
1942 classzone_idx) == COMPACT_CONTINUE)
1949 static void kcompactd_do_work(pg_data_t *pgdat)
1952 * With no special task, compact all zones so that a page of requested
1953 * order is allocatable.
1957 struct compact_control cc = {
1958 .order = pgdat->kcompactd_max_order,
1959 .total_migrate_scanned = 0,
1960 .total_free_scanned = 0,
1961 .classzone_idx = pgdat->kcompactd_classzone_idx,
1962 .mode = MIGRATE_SYNC_LIGHT,
1963 .ignore_skip_hint = false,
1964 .gfp_mask = GFP_KERNEL,
1966 trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
1968 count_compact_event(KCOMPACTD_WAKE);
1970 for (zoneid = 0; zoneid <= cc.classzone_idx; zoneid++) {
1973 zone = &pgdat->node_zones[zoneid];
1974 if (!populated_zone(zone))
1977 if (compaction_deferred(zone, cc.order))
1980 if (compaction_suitable(zone, cc.order, 0, zoneid) !=
1984 cc.nr_freepages = 0;
1985 cc.nr_migratepages = 0;
1986 cc.total_migrate_scanned = 0;
1987 cc.total_free_scanned = 0;
1989 INIT_LIST_HEAD(&cc.freepages);
1990 INIT_LIST_HEAD(&cc.migratepages);
1992 if (kthread_should_stop())
1994 status = compact_zone(zone, &cc);
1996 if (status == COMPACT_SUCCESS) {
1997 compaction_defer_reset(zone, cc.order, false);
1998 } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
2000 * We use sync migration mode here, so we defer like
2001 * sync direct compaction does.
2003 defer_compaction(zone, cc.order);
2006 count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
2007 cc.total_migrate_scanned);
2008 count_compact_events(KCOMPACTD_FREE_SCANNED,
2009 cc.total_free_scanned);
2011 VM_BUG_ON(!list_empty(&cc.freepages));
2012 VM_BUG_ON(!list_empty(&cc.migratepages));
2016 * Regardless of success, we are done until woken up next. But remember
2017 * the requested order/classzone_idx in case it was higher/tighter than
2020 if (pgdat->kcompactd_max_order <= cc.order)
2021 pgdat->kcompactd_max_order = 0;
2022 if (pgdat->kcompactd_classzone_idx >= cc.classzone_idx)
2023 pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
2026 void wakeup_kcompactd(pg_data_t *pgdat, int order, int classzone_idx)
2031 if (pgdat->kcompactd_max_order < order)
2032 pgdat->kcompactd_max_order = order;
2034 if (pgdat->kcompactd_classzone_idx > classzone_idx)
2035 pgdat->kcompactd_classzone_idx = classzone_idx;
2038 * Pairs with implicit barrier in wait_event_freezable()
2039 * such that wakeups are not missed.
2041 if (!wq_has_sleeper(&pgdat->kcompactd_wait))
2044 if (!kcompactd_node_suitable(pgdat))
2047 trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
2049 wake_up_interruptible(&pgdat->kcompactd_wait);
2053 * The background compaction daemon, started as a kernel thread
2054 * from the init process.
2056 static int kcompactd(void *p)
2058 pg_data_t *pgdat = (pg_data_t*)p;
2059 struct task_struct *tsk = current;
2061 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2063 if (!cpumask_empty(cpumask))
2064 set_cpus_allowed_ptr(tsk, cpumask);
2068 pgdat->kcompactd_max_order = 0;
2069 pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
2071 while (!kthread_should_stop()) {
2072 trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
2073 wait_event_freezable(pgdat->kcompactd_wait,
2074 kcompactd_work_requested(pgdat));
2076 kcompactd_do_work(pgdat);
2083 * This kcompactd start function will be called by init and node-hot-add.
2084 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
2086 int kcompactd_run(int nid)
2088 pg_data_t *pgdat = NODE_DATA(nid);
2091 if (pgdat->kcompactd)
2094 pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
2095 if (IS_ERR(pgdat->kcompactd)) {
2096 pr_err("Failed to start kcompactd on node %d\n", nid);
2097 ret = PTR_ERR(pgdat->kcompactd);
2098 pgdat->kcompactd = NULL;
2104 * Called by memory hotplug when all memory in a node is offlined. Caller must
2105 * hold mem_hotplug_begin/end().
2107 void kcompactd_stop(int nid)
2109 struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
2112 kthread_stop(kcompactd);
2113 NODE_DATA(nid)->kcompactd = NULL;
2118 * It's optimal to keep kcompactd on the same CPUs as their memory, but
2119 * not required for correctness. So if the last cpu in a node goes
2120 * away, we get changed to run anywhere: as the first one comes back,
2121 * restore their cpu bindings.
2123 static int kcompactd_cpu_online(unsigned int cpu)
2127 for_each_node_state(nid, N_MEMORY) {
2128 pg_data_t *pgdat = NODE_DATA(nid);
2129 const struct cpumask *mask;
2131 mask = cpumask_of_node(pgdat->node_id);
2133 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2134 /* One of our CPUs online: restore mask */
2135 set_cpus_allowed_ptr(pgdat->kcompactd, mask);
2140 static int __init kcompactd_init(void)
2145 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
2146 "mm/compaction:online",
2147 kcompactd_cpu_online, NULL);
2149 pr_err("kcompactd: failed to register hotplug callbacks.\n");
2153 for_each_node_state(nid, N_MEMORY)
2157 subsys_initcall(kcompactd_init)
2159 #endif /* CONFIG_COMPACTION */