2 * linux/mm/compaction.c
4 * Memory compaction for the reduction of external fragmentation. Note that
5 * this heavily depends upon page migration to do all the real heavy
8 * Copyright IBM Corp. 2007-2010 Mel Gorman <mel@csn.ul.ie>
10 #include <linux/cpu.h>
11 #include <linux/swap.h>
12 #include <linux/migrate.h>
13 #include <linux/compaction.h>
14 #include <linux/mm_inline.h>
15 #include <linux/backing-dev.h>
16 #include <linux/sysctl.h>
17 #include <linux/sysfs.h>
18 #include <linux/page-isolation.h>
19 #include <linux/kasan.h>
20 #include <linux/kthread.h>
21 #include <linux/freezer.h>
22 #include <linux/page_owner.h>
25 #ifdef CONFIG_COMPACTION
26 static inline void count_compact_event(enum vm_event_item item)
31 static inline void count_compact_events(enum vm_event_item item, long delta)
33 count_vm_events(item, delta);
36 #define count_compact_event(item) do { } while (0)
37 #define count_compact_events(item, delta) do { } while (0)
40 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
42 #define CREATE_TRACE_POINTS
43 #include <trace/events/compaction.h>
45 #define block_start_pfn(pfn, order) round_down(pfn, 1UL << (order))
46 #define block_end_pfn(pfn, order) ALIGN((pfn) + 1, 1UL << (order))
47 #define pageblock_start_pfn(pfn) block_start_pfn(pfn, pageblock_order)
48 #define pageblock_end_pfn(pfn) block_end_pfn(pfn, pageblock_order)
50 static unsigned long release_freepages(struct list_head *freelist)
52 struct page *page, *next;
53 unsigned long high_pfn = 0;
55 list_for_each_entry_safe(page, next, freelist, lru) {
56 unsigned long pfn = page_to_pfn(page);
66 static void map_pages(struct list_head *list)
68 unsigned int i, order, nr_pages;
69 struct page *page, *next;
72 list_for_each_entry_safe(page, next, list, lru) {
75 order = page_private(page);
76 nr_pages = 1 << order;
78 post_alloc_hook(page, order, __GFP_MOVABLE);
80 split_page(page, order);
82 for (i = 0; i < nr_pages; i++) {
83 list_add(&page->lru, &tmp_list);
88 list_splice(&tmp_list, list);
91 static inline bool migrate_async_suitable(int migratetype)
93 return is_migrate_cma(migratetype) || migratetype == MIGRATE_MOVABLE;
96 #ifdef CONFIG_COMPACTION
98 int PageMovable(struct page *page)
100 struct address_space *mapping;
102 VM_BUG_ON_PAGE(!PageLocked(page), page);
103 if (!__PageMovable(page))
106 mapping = page_mapping(page);
107 if (mapping && mapping->a_ops && mapping->a_ops->isolate_page)
112 EXPORT_SYMBOL(PageMovable);
114 void __SetPageMovable(struct page *page, struct address_space *mapping)
116 VM_BUG_ON_PAGE(!PageLocked(page), page);
117 VM_BUG_ON_PAGE((unsigned long)mapping & PAGE_MAPPING_MOVABLE, page);
118 page->mapping = (void *)((unsigned long)mapping | PAGE_MAPPING_MOVABLE);
120 EXPORT_SYMBOL(__SetPageMovable);
122 void __ClearPageMovable(struct page *page)
124 VM_BUG_ON_PAGE(!PageLocked(page), page);
125 VM_BUG_ON_PAGE(!PageMovable(page), page);
127 * Clear registered address_space val with keeping PAGE_MAPPING_MOVABLE
128 * flag so that VM can catch up released page by driver after isolation.
129 * With it, VM migration doesn't try to put it back.
131 page->mapping = (void *)((unsigned long)page->mapping &
132 PAGE_MAPPING_MOVABLE);
134 EXPORT_SYMBOL(__ClearPageMovable);
136 /* Do not skip compaction more than 64 times */
137 #define COMPACT_MAX_DEFER_SHIFT 6
140 * Compaction is deferred when compaction fails to result in a page
141 * allocation success. 1 << compact_defer_limit compactions are skipped up
142 * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
144 void defer_compaction(struct zone *zone, int order)
146 zone->compact_considered = 0;
147 zone->compact_defer_shift++;
149 if (order < zone->compact_order_failed)
150 zone->compact_order_failed = order;
152 if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
153 zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
155 trace_mm_compaction_defer_compaction(zone, order);
158 /* Returns true if compaction should be skipped this time */
159 bool compaction_deferred(struct zone *zone, int order)
161 unsigned long defer_limit = 1UL << zone->compact_defer_shift;
163 if (order < zone->compact_order_failed)
166 /* Avoid possible overflow */
167 if (++zone->compact_considered > defer_limit)
168 zone->compact_considered = defer_limit;
170 if (zone->compact_considered >= defer_limit)
173 trace_mm_compaction_deferred(zone, order);
179 * Update defer tracking counters after successful compaction of given order,
180 * which means an allocation either succeeded (alloc_success == true) or is
181 * expected to succeed.
183 void compaction_defer_reset(struct zone *zone, int order,
187 zone->compact_considered = 0;
188 zone->compact_defer_shift = 0;
190 if (order >= zone->compact_order_failed)
191 zone->compact_order_failed = order + 1;
193 trace_mm_compaction_defer_reset(zone, order);
196 /* Returns true if restarting compaction after many failures */
197 bool compaction_restarting(struct zone *zone, int order)
199 if (order < zone->compact_order_failed)
202 return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
203 zone->compact_considered >= 1UL << zone->compact_defer_shift;
206 /* Returns true if the pageblock should be scanned for pages to isolate. */
207 static inline bool isolation_suitable(struct compact_control *cc,
210 if (cc->ignore_skip_hint)
213 return !get_pageblock_skip(page);
216 static void reset_cached_positions(struct zone *zone)
218 zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
219 zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
220 zone->compact_cached_free_pfn =
221 pageblock_start_pfn(zone_end_pfn(zone) - 1);
225 * This function is called to clear all cached information on pageblocks that
226 * should be skipped for page isolation when the migrate and free page scanner
229 static void __reset_isolation_suitable(struct zone *zone)
231 unsigned long start_pfn = zone->zone_start_pfn;
232 unsigned long end_pfn = zone_end_pfn(zone);
235 zone->compact_blockskip_flush = false;
237 /* Walk the zone and mark every pageblock as suitable for isolation */
238 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
246 page = pfn_to_page(pfn);
247 if (zone != page_zone(page))
250 clear_pageblock_skip(page);
253 reset_cached_positions(zone);
256 void reset_isolation_suitable(pg_data_t *pgdat)
260 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
261 struct zone *zone = &pgdat->node_zones[zoneid];
262 if (!populated_zone(zone))
265 /* Only flush if a full compaction finished recently */
266 if (zone->compact_blockskip_flush)
267 __reset_isolation_suitable(zone);
272 * If no pages were isolated then mark this pageblock to be skipped in the
273 * future. The information is later cleared by __reset_isolation_suitable().
275 static void update_pageblock_skip(struct compact_control *cc,
276 struct page *page, unsigned long nr_isolated,
277 bool migrate_scanner)
279 struct zone *zone = cc->zone;
282 if (cc->ignore_skip_hint)
291 set_pageblock_skip(page);
293 pfn = page_to_pfn(page);
295 /* Update where async and sync compaction should restart */
296 if (migrate_scanner) {
297 if (pfn > zone->compact_cached_migrate_pfn[0])
298 zone->compact_cached_migrate_pfn[0] = pfn;
299 if (cc->mode != MIGRATE_ASYNC &&
300 pfn > zone->compact_cached_migrate_pfn[1])
301 zone->compact_cached_migrate_pfn[1] = pfn;
303 if (pfn < zone->compact_cached_free_pfn)
304 zone->compact_cached_free_pfn = pfn;
308 static inline bool isolation_suitable(struct compact_control *cc,
314 static void update_pageblock_skip(struct compact_control *cc,
315 struct page *page, unsigned long nr_isolated,
316 bool migrate_scanner)
319 #endif /* CONFIG_COMPACTION */
322 * Compaction requires the taking of some coarse locks that are potentially
323 * very heavily contended. For async compaction, back out if the lock cannot
324 * be taken immediately. For sync compaction, spin on the lock if needed.
326 * Returns true if the lock is held
327 * Returns false if the lock is not held and compaction should abort
329 static bool compact_trylock_irqsave(spinlock_t *lock, unsigned long *flags,
330 struct compact_control *cc)
332 if (cc->mode == MIGRATE_ASYNC) {
333 if (!spin_trylock_irqsave(lock, *flags)) {
334 cc->contended = true;
338 spin_lock_irqsave(lock, *flags);
345 * Compaction requires the taking of some coarse locks that are potentially
346 * very heavily contended. The lock should be periodically unlocked to avoid
347 * having disabled IRQs for a long time, even when there is nobody waiting on
348 * the lock. It might also be that allowing the IRQs will result in
349 * need_resched() becoming true. If scheduling is needed, async compaction
350 * aborts. Sync compaction schedules.
351 * Either compaction type will also abort if a fatal signal is pending.
352 * In either case if the lock was locked, it is dropped and not regained.
354 * Returns true if compaction should abort due to fatal signal pending, or
355 * async compaction due to need_resched()
356 * Returns false when compaction can continue (sync compaction might have
359 static bool compact_unlock_should_abort(spinlock_t *lock,
360 unsigned long flags, bool *locked, struct compact_control *cc)
363 spin_unlock_irqrestore(lock, flags);
367 if (fatal_signal_pending(current)) {
368 cc->contended = true;
372 if (need_resched()) {
373 if (cc->mode == MIGRATE_ASYNC) {
374 cc->contended = true;
384 * Aside from avoiding lock contention, compaction also periodically checks
385 * need_resched() and either schedules in sync compaction or aborts async
386 * compaction. This is similar to what compact_unlock_should_abort() does, but
387 * is used where no lock is concerned.
389 * Returns false when no scheduling was needed, or sync compaction scheduled.
390 * Returns true when async compaction should abort.
392 static inline bool compact_should_abort(struct compact_control *cc)
394 /* async compaction aborts if contended */
395 if (need_resched()) {
396 if (cc->mode == MIGRATE_ASYNC) {
397 cc->contended = true;
408 * Isolate free pages onto a private freelist. If @strict is true, will abort
409 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
410 * (even though it may still end up isolating some pages).
412 static unsigned long isolate_freepages_block(struct compact_control *cc,
413 unsigned long *start_pfn,
414 unsigned long end_pfn,
415 struct list_head *freelist,
418 int nr_scanned = 0, total_isolated = 0;
419 struct page *cursor, *valid_page = NULL;
420 unsigned long flags = 0;
422 unsigned long blockpfn = *start_pfn;
425 cursor = pfn_to_page(blockpfn);
427 /* Isolate free pages. */
428 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
430 struct page *page = cursor;
433 * Periodically drop the lock (if held) regardless of its
434 * contention, to give chance to IRQs. Abort if fatal signal
435 * pending or async compaction detects need_resched()
437 if (!(blockpfn % SWAP_CLUSTER_MAX)
438 && compact_unlock_should_abort(&cc->zone->lock, flags,
443 if (!pfn_valid_within(blockpfn))
450 * For compound pages such as THP and hugetlbfs, we can save
451 * potentially a lot of iterations if we skip them at once.
452 * The check is racy, but we can consider only valid values
453 * and the only danger is skipping too much.
455 if (PageCompound(page)) {
456 unsigned int comp_order = compound_order(page);
458 if (likely(comp_order < MAX_ORDER)) {
459 blockpfn += (1UL << comp_order) - 1;
460 cursor += (1UL << comp_order) - 1;
466 if (!PageBuddy(page))
470 * If we already hold the lock, we can skip some rechecking.
471 * Note that if we hold the lock now, checked_pageblock was
472 * already set in some previous iteration (or strict is true),
473 * so it is correct to skip the suitable migration target
478 * The zone lock must be held to isolate freepages.
479 * Unfortunately this is a very coarse lock and can be
480 * heavily contended if there are parallel allocations
481 * or parallel compactions. For async compaction do not
482 * spin on the lock and we acquire the lock as late as
485 locked = compact_trylock_irqsave(&cc->zone->lock,
490 /* Recheck this is a buddy page under lock */
491 if (!PageBuddy(page))
495 /* Found a free page, will break it into order-0 pages */
496 order = page_order(page);
497 isolated = __isolate_free_page(page, order);
500 set_page_private(page, order);
502 total_isolated += isolated;
503 cc->nr_freepages += isolated;
504 list_add_tail(&page->lru, freelist);
506 if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
507 blockpfn += isolated;
510 /* Advance to the end of split page */
511 blockpfn += isolated - 1;
512 cursor += isolated - 1;
524 spin_unlock_irqrestore(&cc->zone->lock, flags);
527 * There is a tiny chance that we have read bogus compound_order(),
528 * so be careful to not go outside of the pageblock.
530 if (unlikely(blockpfn > end_pfn))
533 trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
534 nr_scanned, total_isolated);
536 /* Record how far we have got within the block */
537 *start_pfn = blockpfn;
540 * If strict isolation is requested by CMA then check that all the
541 * pages requested were isolated. If there were any failures, 0 is
542 * returned and CMA will fail.
544 if (strict && blockpfn < end_pfn)
547 /* Update the pageblock-skip if the whole pageblock was scanned */
548 if (blockpfn == end_pfn)
549 update_pageblock_skip(cc, valid_page, total_isolated, false);
551 count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
553 count_compact_events(COMPACTISOLATED, total_isolated);
554 return total_isolated;
558 * isolate_freepages_range() - isolate free pages.
559 * @start_pfn: The first PFN to start isolating.
560 * @end_pfn: The one-past-last PFN.
562 * Non-free pages, invalid PFNs, or zone boundaries within the
563 * [start_pfn, end_pfn) range are considered errors, cause function to
564 * undo its actions and return zero.
566 * Otherwise, function returns one-past-the-last PFN of isolated page
567 * (which may be greater then end_pfn if end fell in a middle of
571 isolate_freepages_range(struct compact_control *cc,
572 unsigned long start_pfn, unsigned long end_pfn)
574 unsigned long isolated, pfn, block_start_pfn, block_end_pfn;
578 block_start_pfn = pageblock_start_pfn(pfn);
579 if (block_start_pfn < cc->zone->zone_start_pfn)
580 block_start_pfn = cc->zone->zone_start_pfn;
581 block_end_pfn = pageblock_end_pfn(pfn);
583 for (; pfn < end_pfn; pfn += isolated,
584 block_start_pfn = block_end_pfn,
585 block_end_pfn += pageblock_nr_pages) {
586 /* Protect pfn from changing by isolate_freepages_block */
587 unsigned long isolate_start_pfn = pfn;
589 block_end_pfn = min(block_end_pfn, end_pfn);
592 * pfn could pass the block_end_pfn if isolated freepage
593 * is more than pageblock order. In this case, we adjust
594 * scanning range to right one.
596 if (pfn >= block_end_pfn) {
597 block_start_pfn = pageblock_start_pfn(pfn);
598 block_end_pfn = pageblock_end_pfn(pfn);
599 block_end_pfn = min(block_end_pfn, end_pfn);
602 if (!pageblock_pfn_to_page(block_start_pfn,
603 block_end_pfn, cc->zone))
606 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
607 block_end_pfn, &freelist, true);
610 * In strict mode, isolate_freepages_block() returns 0 if
611 * there are any holes in the block (ie. invalid PFNs or
618 * If we managed to isolate pages, it is always (1 << n) *
619 * pageblock_nr_pages for some non-negative n. (Max order
620 * page may span two pageblocks).
624 /* __isolate_free_page() does not map the pages */
625 map_pages(&freelist);
628 /* Loop terminated early, cleanup. */
629 release_freepages(&freelist);
633 /* We don't use freelists for anything. */
637 /* Similar to reclaim, but different enough that they don't share logic */
638 static bool too_many_isolated(struct zone *zone)
640 unsigned long active, inactive, isolated;
642 inactive = node_page_state(zone->zone_pgdat, NR_INACTIVE_FILE) +
643 node_page_state(zone->zone_pgdat, NR_INACTIVE_ANON);
644 active = node_page_state(zone->zone_pgdat, NR_ACTIVE_FILE) +
645 node_page_state(zone->zone_pgdat, NR_ACTIVE_ANON);
646 isolated = node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE) +
647 node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON);
649 return isolated > (inactive + active) / 2;
653 * isolate_migratepages_block() - isolate all migrate-able pages within
655 * @cc: Compaction control structure.
656 * @low_pfn: The first PFN to isolate
657 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
658 * @isolate_mode: Isolation mode to be used.
660 * Isolate all pages that can be migrated from the range specified by
661 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
662 * Returns zero if there is a fatal signal pending, otherwise PFN of the
663 * first page that was not scanned (which may be both less, equal to or more
666 * The pages are isolated on cc->migratepages list (not required to be empty),
667 * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
668 * is neither read nor updated.
671 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
672 unsigned long end_pfn, isolate_mode_t isolate_mode)
674 struct zone *zone = cc->zone;
675 unsigned long nr_scanned = 0, nr_isolated = 0;
676 struct lruvec *lruvec;
677 unsigned long flags = 0;
679 struct page *page = NULL, *valid_page = NULL;
680 unsigned long start_pfn = low_pfn;
681 bool skip_on_failure = false;
682 unsigned long next_skip_pfn = 0;
685 * Ensure that there are not too many pages isolated from the LRU
686 * list by either parallel reclaimers or compaction. If there are,
687 * delay for some time until fewer pages are isolated
689 while (unlikely(too_many_isolated(zone))) {
690 /* async migration should just abort */
691 if (cc->mode == MIGRATE_ASYNC)
694 congestion_wait(BLK_RW_ASYNC, HZ/10);
696 if (fatal_signal_pending(current))
700 if (compact_should_abort(cc))
703 if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
704 skip_on_failure = true;
705 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
708 /* Time to isolate some pages for migration */
709 for (; low_pfn < end_pfn; low_pfn++) {
711 if (skip_on_failure && low_pfn >= next_skip_pfn) {
713 * We have isolated all migration candidates in the
714 * previous order-aligned block, and did not skip it due
715 * to failure. We should migrate the pages now and
716 * hopefully succeed compaction.
722 * We failed to isolate in the previous order-aligned
723 * block. Set the new boundary to the end of the
724 * current block. Note we can't simply increase
725 * next_skip_pfn by 1 << order, as low_pfn might have
726 * been incremented by a higher number due to skipping
727 * a compound or a high-order buddy page in the
728 * previous loop iteration.
730 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
734 * Periodically drop the lock (if held) regardless of its
735 * contention, to give chance to IRQs. Abort async compaction
738 if (!(low_pfn % SWAP_CLUSTER_MAX)
739 && compact_unlock_should_abort(zone_lru_lock(zone), flags,
743 if (!pfn_valid_within(low_pfn))
747 page = pfn_to_page(low_pfn);
753 * Skip if free. We read page order here without zone lock
754 * which is generally unsafe, but the race window is small and
755 * the worst thing that can happen is that we skip some
756 * potential isolation targets.
758 if (PageBuddy(page)) {
759 unsigned long freepage_order = page_order_unsafe(page);
762 * Without lock, we cannot be sure that what we got is
763 * a valid page order. Consider only values in the
764 * valid order range to prevent low_pfn overflow.
766 if (freepage_order > 0 && freepage_order < MAX_ORDER)
767 low_pfn += (1UL << freepage_order) - 1;
772 * Regardless of being on LRU, compound pages such as THP and
773 * hugetlbfs are not to be compacted. We can potentially save
774 * a lot of iterations if we skip them at once. The check is
775 * racy, but we can consider only valid values and the only
776 * danger is skipping too much.
778 if (PageCompound(page)) {
779 unsigned int comp_order = compound_order(page);
781 if (likely(comp_order < MAX_ORDER))
782 low_pfn += (1UL << comp_order) - 1;
788 * Check may be lockless but that's ok as we recheck later.
789 * It's possible to migrate LRU and non-lru movable pages.
790 * Skip any other type of page
792 if (!PageLRU(page)) {
794 * __PageMovable can return false positive so we need
795 * to verify it under page_lock.
797 if (unlikely(__PageMovable(page)) &&
798 !PageIsolated(page)) {
800 spin_unlock_irqrestore(zone_lru_lock(zone),
805 if (isolate_movable_page(page, isolate_mode))
806 goto isolate_success;
813 * Migration will fail if an anonymous page is pinned in memory,
814 * so avoid taking lru_lock and isolating it unnecessarily in an
815 * admittedly racy check.
817 if (!page_mapping(page) &&
818 page_count(page) > page_mapcount(page))
821 /* If we already hold the lock, we can skip some rechecking */
823 locked = compact_trylock_irqsave(zone_lru_lock(zone),
828 /* Recheck PageLRU and PageCompound under lock */
833 * Page become compound since the non-locked check,
834 * and it's on LRU. It can only be a THP so the order
835 * is safe to read and it's 0 for tail pages.
837 if (unlikely(PageCompound(page))) {
838 low_pfn += (1UL << compound_order(page)) - 1;
843 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
845 /* Try isolate the page */
846 if (__isolate_lru_page(page, isolate_mode) != 0)
849 VM_BUG_ON_PAGE(PageCompound(page), page);
851 /* Successfully isolated */
852 del_page_from_lru_list(page, lruvec, page_lru(page));
853 inc_node_page_state(page,
854 NR_ISOLATED_ANON + page_is_file_cache(page));
857 list_add(&page->lru, &cc->migratepages);
858 cc->nr_migratepages++;
862 * Record where we could have freed pages by migration and not
863 * yet flushed them to buddy allocator.
864 * - this is the lowest page that was isolated and likely be
865 * then freed by migration.
867 if (!cc->last_migrated_pfn)
868 cc->last_migrated_pfn = low_pfn;
870 /* Avoid isolating too much */
871 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
878 if (!skip_on_failure)
882 * We have isolated some pages, but then failed. Release them
883 * instead of migrating, as we cannot form the cc->order buddy
888 spin_unlock_irqrestore(zone_lru_lock(zone), flags);
891 putback_movable_pages(&cc->migratepages);
892 cc->nr_migratepages = 0;
893 cc->last_migrated_pfn = 0;
897 if (low_pfn < next_skip_pfn) {
898 low_pfn = next_skip_pfn - 1;
900 * The check near the loop beginning would have updated
901 * next_skip_pfn too, but this is a bit simpler.
903 next_skip_pfn += 1UL << cc->order;
908 * The PageBuddy() check could have potentially brought us outside
909 * the range to be scanned.
911 if (unlikely(low_pfn > end_pfn))
915 spin_unlock_irqrestore(zone_lru_lock(zone), flags);
918 * Update the pageblock-skip information and cached scanner pfn,
919 * if the whole pageblock was scanned without isolating any page.
921 if (low_pfn == end_pfn)
922 update_pageblock_skip(cc, valid_page, nr_isolated, true);
924 trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
925 nr_scanned, nr_isolated);
927 count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
929 count_compact_events(COMPACTISOLATED, nr_isolated);
935 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
936 * @cc: Compaction control structure.
937 * @start_pfn: The first PFN to start isolating.
938 * @end_pfn: The one-past-last PFN.
940 * Returns zero if isolation fails fatally due to e.g. pending signal.
941 * Otherwise, function returns one-past-the-last PFN of isolated page
942 * (which may be greater than end_pfn if end fell in a middle of a THP page).
945 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
946 unsigned long end_pfn)
948 unsigned long pfn, block_start_pfn, block_end_pfn;
950 /* Scan block by block. First and last block may be incomplete */
952 block_start_pfn = pageblock_start_pfn(pfn);
953 if (block_start_pfn < cc->zone->zone_start_pfn)
954 block_start_pfn = cc->zone->zone_start_pfn;
955 block_end_pfn = pageblock_end_pfn(pfn);
957 for (; pfn < end_pfn; pfn = block_end_pfn,
958 block_start_pfn = block_end_pfn,
959 block_end_pfn += pageblock_nr_pages) {
961 block_end_pfn = min(block_end_pfn, end_pfn);
963 if (!pageblock_pfn_to_page(block_start_pfn,
964 block_end_pfn, cc->zone))
967 pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
968 ISOLATE_UNEVICTABLE);
973 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
980 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
981 #ifdef CONFIG_COMPACTION
983 /* Returns true if the page is within a block suitable for migration to */
984 static bool suitable_migration_target(struct compact_control *cc,
987 if (cc->ignore_block_suitable)
990 /* If the page is a large free page, then disallow migration */
991 if (PageBuddy(page)) {
993 * We are checking page_order without zone->lock taken. But
994 * the only small danger is that we skip a potentially suitable
995 * pageblock, so it's not worth to check order for valid range.
997 if (page_order_unsafe(page) >= pageblock_order)
1001 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1002 if (migrate_async_suitable(get_pageblock_migratetype(page)))
1005 /* Otherwise skip the block */
1010 * Test whether the free scanner has reached the same or lower pageblock than
1011 * the migration scanner, and compaction should thus terminate.
1013 static inline bool compact_scanners_met(struct compact_control *cc)
1015 return (cc->free_pfn >> pageblock_order)
1016 <= (cc->migrate_pfn >> pageblock_order);
1020 * Based on information in the current compact_control, find blocks
1021 * suitable for isolating free pages from and then isolate them.
1023 static void isolate_freepages(struct compact_control *cc)
1025 struct zone *zone = cc->zone;
1027 unsigned long block_start_pfn; /* start of current pageblock */
1028 unsigned long isolate_start_pfn; /* exact pfn we start at */
1029 unsigned long block_end_pfn; /* end of current pageblock */
1030 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
1031 struct list_head *freelist = &cc->freepages;
1034 * Initialise the free scanner. The starting point is where we last
1035 * successfully isolated from, zone-cached value, or the end of the
1036 * zone when isolating for the first time. For looping we also need
1037 * this pfn aligned down to the pageblock boundary, because we do
1038 * block_start_pfn -= pageblock_nr_pages in the for loop.
1039 * For ending point, take care when isolating in last pageblock of a
1040 * a zone which ends in the middle of a pageblock.
1041 * The low boundary is the end of the pageblock the migration scanner
1044 isolate_start_pfn = cc->free_pfn;
1045 block_start_pfn = pageblock_start_pfn(cc->free_pfn);
1046 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1047 zone_end_pfn(zone));
1048 low_pfn = pageblock_end_pfn(cc->migrate_pfn);
1051 * Isolate free pages until enough are available to migrate the
1052 * pages on cc->migratepages. We stop searching if the migrate
1053 * and free page scanners meet or enough free pages are isolated.
1055 for (; block_start_pfn >= low_pfn;
1056 block_end_pfn = block_start_pfn,
1057 block_start_pfn -= pageblock_nr_pages,
1058 isolate_start_pfn = block_start_pfn) {
1060 * This can iterate a massively long zone without finding any
1061 * suitable migration targets, so periodically check if we need
1062 * to schedule, or even abort async compaction.
1064 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1065 && compact_should_abort(cc))
1068 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1073 /* Check the block is suitable for migration */
1074 if (!suitable_migration_target(cc, page))
1077 /* If isolation recently failed, do not retry */
1078 if (!isolation_suitable(cc, page))
1081 /* Found a block suitable for isolating free pages from. */
1082 isolate_freepages_block(cc, &isolate_start_pfn, block_end_pfn,
1086 * If we isolated enough freepages, or aborted due to lock
1087 * contention, terminate.
1089 if ((cc->nr_freepages >= cc->nr_migratepages)
1091 if (isolate_start_pfn >= block_end_pfn) {
1093 * Restart at previous pageblock if more
1094 * freepages can be isolated next time.
1097 block_start_pfn - pageblock_nr_pages;
1100 } else if (isolate_start_pfn < block_end_pfn) {
1102 * If isolation failed early, do not continue
1109 /* __isolate_free_page() does not map the pages */
1110 map_pages(freelist);
1113 * Record where the free scanner will restart next time. Either we
1114 * broke from the loop and set isolate_start_pfn based on the last
1115 * call to isolate_freepages_block(), or we met the migration scanner
1116 * and the loop terminated due to isolate_start_pfn < low_pfn
1118 cc->free_pfn = isolate_start_pfn;
1122 * This is a migrate-callback that "allocates" freepages by taking pages
1123 * from the isolated freelists in the block we are migrating to.
1125 static struct page *compaction_alloc(struct page *migratepage,
1129 struct compact_control *cc = (struct compact_control *)data;
1130 struct page *freepage;
1133 * Isolate free pages if necessary, and if we are not aborting due to
1136 if (list_empty(&cc->freepages)) {
1138 isolate_freepages(cc);
1140 if (list_empty(&cc->freepages))
1144 freepage = list_entry(cc->freepages.next, struct page, lru);
1145 list_del(&freepage->lru);
1152 * This is a migrate-callback that "frees" freepages back to the isolated
1153 * freelist. All pages on the freelist are from the same zone, so there is no
1154 * special handling needed for NUMA.
1156 static void compaction_free(struct page *page, unsigned long data)
1158 struct compact_control *cc = (struct compact_control *)data;
1160 list_add(&page->lru, &cc->freepages);
1164 /* possible outcome of isolate_migratepages */
1166 ISOLATE_ABORT, /* Abort compaction now */
1167 ISOLATE_NONE, /* No pages isolated, continue scanning */
1168 ISOLATE_SUCCESS, /* Pages isolated, migrate */
1169 } isolate_migrate_t;
1172 * Allow userspace to control policy on scanning the unevictable LRU for
1173 * compactable pages.
1175 int sysctl_compact_unevictable_allowed __read_mostly = 1;
1178 * Isolate all pages that can be migrated from the first suitable block,
1179 * starting at the block pointed to by the migrate scanner pfn within
1182 static isolate_migrate_t isolate_migratepages(struct zone *zone,
1183 struct compact_control *cc)
1185 unsigned long block_start_pfn;
1186 unsigned long block_end_pfn;
1187 unsigned long low_pfn;
1189 const isolate_mode_t isolate_mode =
1190 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1191 (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1194 * Start at where we last stopped, or beginning of the zone as
1195 * initialized by compact_zone()
1197 low_pfn = cc->migrate_pfn;
1198 block_start_pfn = pageblock_start_pfn(low_pfn);
1199 if (block_start_pfn < zone->zone_start_pfn)
1200 block_start_pfn = zone->zone_start_pfn;
1202 /* Only scan within a pageblock boundary */
1203 block_end_pfn = pageblock_end_pfn(low_pfn);
1206 * Iterate over whole pageblocks until we find the first suitable.
1207 * Do not cross the free scanner.
1209 for (; block_end_pfn <= cc->free_pfn;
1210 low_pfn = block_end_pfn,
1211 block_start_pfn = block_end_pfn,
1212 block_end_pfn += pageblock_nr_pages) {
1215 * This can potentially iterate a massively long zone with
1216 * many pageblocks unsuitable, so periodically check if we
1217 * need to schedule, or even abort async compaction.
1219 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1220 && compact_should_abort(cc))
1223 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1228 /* If isolation recently failed, do not retry */
1229 if (!isolation_suitable(cc, page))
1233 * For async compaction, also only scan in MOVABLE blocks.
1234 * Async compaction is optimistic to see if the minimum amount
1235 * of work satisfies the allocation.
1237 if (cc->mode == MIGRATE_ASYNC &&
1238 !migrate_async_suitable(get_pageblock_migratetype(page)))
1241 /* Perform the isolation */
1242 low_pfn = isolate_migratepages_block(cc, low_pfn,
1243 block_end_pfn, isolate_mode);
1245 if (!low_pfn || cc->contended)
1246 return ISOLATE_ABORT;
1249 * Either we isolated something and proceed with migration. Or
1250 * we failed and compact_zone should decide if we should
1256 /* Record where migration scanner will be restarted. */
1257 cc->migrate_pfn = low_pfn;
1259 return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
1263 * order == -1 is expected when compacting via
1264 * /proc/sys/vm/compact_memory
1266 static inline bool is_via_compact_memory(int order)
1271 static enum compact_result __compact_finished(struct zone *zone, struct compact_control *cc,
1272 const int migratetype)
1275 unsigned long watermark;
1277 if (cc->contended || fatal_signal_pending(current))
1278 return COMPACT_CONTENDED;
1280 /* Compaction run completes if the migrate and free scanner meet */
1281 if (compact_scanners_met(cc)) {
1282 /* Let the next compaction start anew. */
1283 reset_cached_positions(zone);
1286 * Mark that the PG_migrate_skip information should be cleared
1287 * by kswapd when it goes to sleep. kcompactd does not set the
1288 * flag itself as the decision to be clear should be directly
1289 * based on an allocation request.
1291 if (cc->direct_compaction)
1292 zone->compact_blockskip_flush = true;
1295 return COMPACT_COMPLETE;
1297 return COMPACT_PARTIAL_SKIPPED;
1300 if (is_via_compact_memory(cc->order))
1301 return COMPACT_CONTINUE;
1303 /* Compaction run is not finished if the watermark is not met */
1304 watermark = zone->watermark[cc->alloc_flags & ALLOC_WMARK_MASK];
1306 if (!zone_watermark_ok(zone, cc->order, watermark, cc->classzone_idx,
1308 return COMPACT_CONTINUE;
1310 /* Direct compactor: Is a suitable page free? */
1311 for (order = cc->order; order < MAX_ORDER; order++) {
1312 struct free_area *area = &zone->free_area[order];
1315 /* Job done if page is free of the right migratetype */
1316 if (!list_empty(&area->free_list[migratetype]))
1317 return COMPACT_SUCCESS;
1320 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
1321 if (migratetype == MIGRATE_MOVABLE &&
1322 !list_empty(&area->free_list[MIGRATE_CMA]))
1323 return COMPACT_SUCCESS;
1326 * Job done if allocation would steal freepages from
1327 * other migratetype buddy lists.
1329 if (find_suitable_fallback(area, order, migratetype,
1330 true, &can_steal) != -1)
1331 return COMPACT_SUCCESS;
1334 return COMPACT_NO_SUITABLE_PAGE;
1337 static enum compact_result compact_finished(struct zone *zone,
1338 struct compact_control *cc,
1339 const int migratetype)
1343 ret = __compact_finished(zone, cc, migratetype);
1344 trace_mm_compaction_finished(zone, cc->order, ret);
1345 if (ret == COMPACT_NO_SUITABLE_PAGE)
1346 ret = COMPACT_CONTINUE;
1352 * compaction_suitable: Is this suitable to run compaction on this zone now?
1354 * COMPACT_SKIPPED - If there are too few free pages for compaction
1355 * COMPACT_SUCCESS - If the allocation would succeed without compaction
1356 * COMPACT_CONTINUE - If compaction should run now
1358 static enum compact_result __compaction_suitable(struct zone *zone, int order,
1359 unsigned int alloc_flags,
1361 unsigned long wmark_target)
1363 unsigned long watermark;
1365 if (is_via_compact_memory(order))
1366 return COMPACT_CONTINUE;
1368 watermark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1370 * If watermarks for high-order allocation are already met, there
1371 * should be no need for compaction at all.
1373 if (zone_watermark_ok(zone, order, watermark, classzone_idx,
1375 return COMPACT_SUCCESS;
1378 * Watermarks for order-0 must be met for compaction to be able to
1379 * isolate free pages for migration targets. This means that the
1380 * watermark and alloc_flags have to match, or be more pessimistic than
1381 * the check in __isolate_free_page(). We don't use the direct
1382 * compactor's alloc_flags, as they are not relevant for freepage
1383 * isolation. We however do use the direct compactor's classzone_idx to
1384 * skip over zones where lowmem reserves would prevent allocation even
1385 * if compaction succeeds.
1386 * For costly orders, we require low watermark instead of min for
1387 * compaction to proceed to increase its chances.
1388 * ALLOC_CMA is used, as pages in CMA pageblocks are considered
1389 * suitable migration targets
1391 watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
1392 low_wmark_pages(zone) : min_wmark_pages(zone);
1393 watermark += compact_gap(order);
1394 if (!__zone_watermark_ok(zone, 0, watermark, classzone_idx,
1395 ALLOC_CMA, wmark_target))
1396 return COMPACT_SKIPPED;
1398 return COMPACT_CONTINUE;
1401 enum compact_result compaction_suitable(struct zone *zone, int order,
1402 unsigned int alloc_flags,
1405 enum compact_result ret;
1408 ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx,
1409 zone_page_state(zone, NR_FREE_PAGES));
1411 * fragmentation index determines if allocation failures are due to
1412 * low memory or external fragmentation
1414 * index of -1000 would imply allocations might succeed depending on
1415 * watermarks, but we already failed the high-order watermark check
1416 * index towards 0 implies failure is due to lack of memory
1417 * index towards 1000 implies failure is due to fragmentation
1419 * Only compact if a failure would be due to fragmentation. Also
1420 * ignore fragindex for non-costly orders where the alternative to
1421 * a successful reclaim/compaction is OOM. Fragindex and the
1422 * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
1423 * excessive compaction for costly orders, but it should not be at the
1424 * expense of system stability.
1426 if (ret == COMPACT_CONTINUE && (order > PAGE_ALLOC_COSTLY_ORDER)) {
1427 fragindex = fragmentation_index(zone, order);
1428 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
1429 ret = COMPACT_NOT_SUITABLE_ZONE;
1432 trace_mm_compaction_suitable(zone, order, ret);
1433 if (ret == COMPACT_NOT_SUITABLE_ZONE)
1434 ret = COMPACT_SKIPPED;
1439 bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
1446 * Make sure at least one zone would pass __compaction_suitable if we continue
1447 * retrying the reclaim.
1449 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1451 unsigned long available;
1452 enum compact_result compact_result;
1455 * Do not consider all the reclaimable memory because we do not
1456 * want to trash just for a single high order allocation which
1457 * is even not guaranteed to appear even if __compaction_suitable
1458 * is happy about the watermark check.
1460 available = zone_reclaimable_pages(zone) / order;
1461 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
1462 compact_result = __compaction_suitable(zone, order, alloc_flags,
1463 ac_classzone_idx(ac), available);
1464 if (compact_result != COMPACT_SKIPPED)
1471 static enum compact_result compact_zone(struct zone *zone, struct compact_control *cc)
1473 enum compact_result ret;
1474 unsigned long start_pfn = zone->zone_start_pfn;
1475 unsigned long end_pfn = zone_end_pfn(zone);
1476 const int migratetype = gfpflags_to_migratetype(cc->gfp_mask);
1477 const bool sync = cc->mode != MIGRATE_ASYNC;
1479 ret = compaction_suitable(zone, cc->order, cc->alloc_flags,
1481 /* Compaction is likely to fail */
1482 if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED)
1485 /* huh, compaction_suitable is returning something unexpected */
1486 VM_BUG_ON(ret != COMPACT_CONTINUE);
1489 * Clear pageblock skip if there were failures recently and compaction
1490 * is about to be retried after being deferred.
1492 if (compaction_restarting(zone, cc->order))
1493 __reset_isolation_suitable(zone);
1496 * Setup to move all movable pages to the end of the zone. Used cached
1497 * information on where the scanners should start (unless we explicitly
1498 * want to compact the whole zone), but check that it is initialised
1499 * by ensuring the values are within zone boundaries.
1501 if (cc->whole_zone) {
1502 cc->migrate_pfn = start_pfn;
1503 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
1505 cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync];
1506 cc->free_pfn = zone->compact_cached_free_pfn;
1507 if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
1508 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
1509 zone->compact_cached_free_pfn = cc->free_pfn;
1511 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
1512 cc->migrate_pfn = start_pfn;
1513 zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
1514 zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
1517 if (cc->migrate_pfn == start_pfn)
1518 cc->whole_zone = true;
1521 cc->last_migrated_pfn = 0;
1523 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
1524 cc->free_pfn, end_pfn, sync);
1526 migrate_prep_local();
1528 while ((ret = compact_finished(zone, cc, migratetype)) ==
1532 switch (isolate_migratepages(zone, cc)) {
1534 ret = COMPACT_CONTENDED;
1535 putback_movable_pages(&cc->migratepages);
1536 cc->nr_migratepages = 0;
1540 * We haven't isolated and migrated anything, but
1541 * there might still be unflushed migrations from
1542 * previous cc->order aligned block.
1545 case ISOLATE_SUCCESS:
1549 err = migrate_pages(&cc->migratepages, compaction_alloc,
1550 compaction_free, (unsigned long)cc, cc->mode,
1553 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
1556 /* All pages were either migrated or will be released */
1557 cc->nr_migratepages = 0;
1559 putback_movable_pages(&cc->migratepages);
1561 * migrate_pages() may return -ENOMEM when scanners meet
1562 * and we want compact_finished() to detect it
1564 if (err == -ENOMEM && !compact_scanners_met(cc)) {
1565 ret = COMPACT_CONTENDED;
1569 * We failed to migrate at least one page in the current
1570 * order-aligned block, so skip the rest of it.
1572 if (cc->direct_compaction &&
1573 (cc->mode == MIGRATE_ASYNC)) {
1574 cc->migrate_pfn = block_end_pfn(
1575 cc->migrate_pfn - 1, cc->order);
1576 /* Draining pcplists is useless in this case */
1577 cc->last_migrated_pfn = 0;
1584 * Has the migration scanner moved away from the previous
1585 * cc->order aligned block where we migrated from? If yes,
1586 * flush the pages that were freed, so that they can merge and
1587 * compact_finished() can detect immediately if allocation
1590 if (cc->order > 0 && cc->last_migrated_pfn) {
1592 unsigned long current_block_start =
1593 block_start_pfn(cc->migrate_pfn, cc->order);
1595 if (cc->last_migrated_pfn < current_block_start) {
1597 lru_add_drain_cpu(cpu);
1598 drain_local_pages(zone);
1600 /* No more flushing until we migrate again */
1601 cc->last_migrated_pfn = 0;
1609 * Release free pages and update where the free scanner should restart,
1610 * so we don't leave any returned pages behind in the next attempt.
1612 if (cc->nr_freepages > 0) {
1613 unsigned long free_pfn = release_freepages(&cc->freepages);
1615 cc->nr_freepages = 0;
1616 VM_BUG_ON(free_pfn == 0);
1617 /* The cached pfn is always the first in a pageblock */
1618 free_pfn = pageblock_start_pfn(free_pfn);
1620 * Only go back, not forward. The cached pfn might have been
1621 * already reset to zone end in compact_finished()
1623 if (free_pfn > zone->compact_cached_free_pfn)
1624 zone->compact_cached_free_pfn = free_pfn;
1627 trace_mm_compaction_end(start_pfn, cc->migrate_pfn,
1628 cc->free_pfn, end_pfn, sync, ret);
1633 static enum compact_result compact_zone_order(struct zone *zone, int order,
1634 gfp_t gfp_mask, enum compact_priority prio,
1635 unsigned int alloc_flags, int classzone_idx)
1637 enum compact_result ret;
1638 struct compact_control cc = {
1640 .nr_migratepages = 0,
1642 .gfp_mask = gfp_mask,
1644 .mode = (prio == COMPACT_PRIO_ASYNC) ?
1645 MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT,
1646 .alloc_flags = alloc_flags,
1647 .classzone_idx = classzone_idx,
1648 .direct_compaction = true,
1649 .whole_zone = (prio == MIN_COMPACT_PRIORITY),
1650 .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
1651 .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
1653 INIT_LIST_HEAD(&cc.freepages);
1654 INIT_LIST_HEAD(&cc.migratepages);
1656 ret = compact_zone(zone, &cc);
1658 VM_BUG_ON(!list_empty(&cc.freepages));
1659 VM_BUG_ON(!list_empty(&cc.migratepages));
1664 int sysctl_extfrag_threshold = 500;
1667 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1668 * @gfp_mask: The GFP mask of the current allocation
1669 * @order: The order of the current allocation
1670 * @alloc_flags: The allocation flags of the current allocation
1671 * @ac: The context of current allocation
1672 * @mode: The migration mode for async, sync light, or sync migration
1674 * This is the main entry point for direct page compaction.
1676 enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
1677 unsigned int alloc_flags, const struct alloc_context *ac,
1678 enum compact_priority prio)
1680 int may_enter_fs = gfp_mask & __GFP_FS;
1681 int may_perform_io = gfp_mask & __GFP_IO;
1684 enum compact_result rc = COMPACT_SKIPPED;
1686 /* Check if the GFP flags allow compaction */
1687 if (!may_enter_fs || !may_perform_io)
1688 return COMPACT_SKIPPED;
1690 trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
1692 /* Compact each zone in the list */
1693 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1695 enum compact_result status;
1697 if (prio > MIN_COMPACT_PRIORITY
1698 && compaction_deferred(zone, order)) {
1699 rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
1703 status = compact_zone_order(zone, order, gfp_mask, prio,
1704 alloc_flags, ac_classzone_idx(ac));
1705 rc = max(status, rc);
1707 /* The allocation should succeed, stop compacting */
1708 if (status == COMPACT_SUCCESS) {
1710 * We think the allocation will succeed in this zone,
1711 * but it is not certain, hence the false. The caller
1712 * will repeat this with true if allocation indeed
1713 * succeeds in this zone.
1715 compaction_defer_reset(zone, order, false);
1720 if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
1721 status == COMPACT_PARTIAL_SKIPPED))
1723 * We think that allocation won't succeed in this zone
1724 * so we defer compaction there. If it ends up
1725 * succeeding after all, it will be reset.
1727 defer_compaction(zone, order);
1730 * We might have stopped compacting due to need_resched() in
1731 * async compaction, or due to a fatal signal detected. In that
1732 * case do not try further zones
1734 if ((prio == COMPACT_PRIO_ASYNC && need_resched())
1735 || fatal_signal_pending(current))
1743 /* Compact all zones within a node */
1744 static void compact_node(int nid)
1746 pg_data_t *pgdat = NODE_DATA(nid);
1749 struct compact_control cc = {
1751 .mode = MIGRATE_SYNC,
1752 .ignore_skip_hint = true,
1757 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1759 zone = &pgdat->node_zones[zoneid];
1760 if (!populated_zone(zone))
1763 cc.nr_freepages = 0;
1764 cc.nr_migratepages = 0;
1766 INIT_LIST_HEAD(&cc.freepages);
1767 INIT_LIST_HEAD(&cc.migratepages);
1769 compact_zone(zone, &cc);
1771 VM_BUG_ON(!list_empty(&cc.freepages));
1772 VM_BUG_ON(!list_empty(&cc.migratepages));
1776 /* Compact all nodes in the system */
1777 static void compact_nodes(void)
1781 /* Flush pending updates to the LRU lists */
1782 lru_add_drain_all();
1784 for_each_online_node(nid)
1788 /* The written value is actually unused, all memory is compacted */
1789 int sysctl_compact_memory;
1792 * This is the entry point for compacting all nodes via
1793 * /proc/sys/vm/compact_memory
1795 int sysctl_compaction_handler(struct ctl_table *table, int write,
1796 void __user *buffer, size_t *length, loff_t *ppos)
1804 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1805 void __user *buffer, size_t *length, loff_t *ppos)
1807 proc_dointvec_minmax(table, write, buffer, length, ppos);
1812 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1813 static ssize_t sysfs_compact_node(struct device *dev,
1814 struct device_attribute *attr,
1815 const char *buf, size_t count)
1819 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1820 /* Flush pending updates to the LRU lists */
1821 lru_add_drain_all();
1828 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1830 int compaction_register_node(struct node *node)
1832 return device_create_file(&node->dev, &dev_attr_compact);
1835 void compaction_unregister_node(struct node *node)
1837 return device_remove_file(&node->dev, &dev_attr_compact);
1839 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1841 static inline bool kcompactd_work_requested(pg_data_t *pgdat)
1843 return pgdat->kcompactd_max_order > 0 || kthread_should_stop();
1846 static bool kcompactd_node_suitable(pg_data_t *pgdat)
1850 enum zone_type classzone_idx = pgdat->kcompactd_classzone_idx;
1852 for (zoneid = 0; zoneid <= classzone_idx; zoneid++) {
1853 zone = &pgdat->node_zones[zoneid];
1855 if (!populated_zone(zone))
1858 if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0,
1859 classzone_idx) == COMPACT_CONTINUE)
1866 static void kcompactd_do_work(pg_data_t *pgdat)
1869 * With no special task, compact all zones so that a page of requested
1870 * order is allocatable.
1874 struct compact_control cc = {
1875 .order = pgdat->kcompactd_max_order,
1876 .classzone_idx = pgdat->kcompactd_classzone_idx,
1877 .mode = MIGRATE_SYNC_LIGHT,
1878 .ignore_skip_hint = true,
1881 trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
1883 count_vm_event(KCOMPACTD_WAKE);
1885 for (zoneid = 0; zoneid <= cc.classzone_idx; zoneid++) {
1888 zone = &pgdat->node_zones[zoneid];
1889 if (!populated_zone(zone))
1892 if (compaction_deferred(zone, cc.order))
1895 if (compaction_suitable(zone, cc.order, 0, zoneid) !=
1899 cc.nr_freepages = 0;
1900 cc.nr_migratepages = 0;
1902 INIT_LIST_HEAD(&cc.freepages);
1903 INIT_LIST_HEAD(&cc.migratepages);
1905 if (kthread_should_stop())
1907 status = compact_zone(zone, &cc);
1909 if (status == COMPACT_SUCCESS) {
1910 compaction_defer_reset(zone, cc.order, false);
1911 } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
1913 * We use sync migration mode here, so we defer like
1914 * sync direct compaction does.
1916 defer_compaction(zone, cc.order);
1919 VM_BUG_ON(!list_empty(&cc.freepages));
1920 VM_BUG_ON(!list_empty(&cc.migratepages));
1924 * Regardless of success, we are done until woken up next. But remember
1925 * the requested order/classzone_idx in case it was higher/tighter than
1928 if (pgdat->kcompactd_max_order <= cc.order)
1929 pgdat->kcompactd_max_order = 0;
1930 if (pgdat->kcompactd_classzone_idx >= cc.classzone_idx)
1931 pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
1934 void wakeup_kcompactd(pg_data_t *pgdat, int order, int classzone_idx)
1939 if (pgdat->kcompactd_max_order < order)
1940 pgdat->kcompactd_max_order = order;
1942 if (pgdat->kcompactd_classzone_idx > classzone_idx)
1943 pgdat->kcompactd_classzone_idx = classzone_idx;
1945 if (!waitqueue_active(&pgdat->kcompactd_wait))
1948 if (!kcompactd_node_suitable(pgdat))
1951 trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
1953 wake_up_interruptible(&pgdat->kcompactd_wait);
1957 * The background compaction daemon, started as a kernel thread
1958 * from the init process.
1960 static int kcompactd(void *p)
1962 pg_data_t *pgdat = (pg_data_t*)p;
1963 struct task_struct *tsk = current;
1965 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1967 if (!cpumask_empty(cpumask))
1968 set_cpus_allowed_ptr(tsk, cpumask);
1972 pgdat->kcompactd_max_order = 0;
1973 pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
1975 while (!kthread_should_stop()) {
1976 trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
1977 wait_event_freezable(pgdat->kcompactd_wait,
1978 kcompactd_work_requested(pgdat));
1980 kcompactd_do_work(pgdat);
1987 * This kcompactd start function will be called by init and node-hot-add.
1988 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
1990 int kcompactd_run(int nid)
1992 pg_data_t *pgdat = NODE_DATA(nid);
1995 if (pgdat->kcompactd)
1998 pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
1999 if (IS_ERR(pgdat->kcompactd)) {
2000 pr_err("Failed to start kcompactd on node %d\n", nid);
2001 ret = PTR_ERR(pgdat->kcompactd);
2002 pgdat->kcompactd = NULL;
2008 * Called by memory hotplug when all memory in a node is offlined. Caller must
2009 * hold mem_hotplug_begin/end().
2011 void kcompactd_stop(int nid)
2013 struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
2016 kthread_stop(kcompactd);
2017 NODE_DATA(nid)->kcompactd = NULL;
2022 * It's optimal to keep kcompactd on the same CPUs as their memory, but
2023 * not required for correctness. So if the last cpu in a node goes
2024 * away, we get changed to run anywhere: as the first one comes back,
2025 * restore their cpu bindings.
2027 static int kcompactd_cpu_online(unsigned int cpu)
2031 for_each_node_state(nid, N_MEMORY) {
2032 pg_data_t *pgdat = NODE_DATA(nid);
2033 const struct cpumask *mask;
2035 mask = cpumask_of_node(pgdat->node_id);
2037 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2038 /* One of our CPUs online: restore mask */
2039 set_cpus_allowed_ptr(pgdat->kcompactd, mask);
2044 static int __init kcompactd_init(void)
2049 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
2050 "mm/compaction:online",
2051 kcompactd_cpu_online, NULL);
2053 pr_err("kcompactd: failed to register hotplug callbacks.\n");
2057 for_each_node_state(nid, N_MEMORY)
2061 subsys_initcall(kcompactd_init)
2063 #endif /* CONFIG_COMPACTION */