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 /* Update the number of anon and file isolated pages in the zone */
638 static void acct_isolated(struct zone *zone, struct compact_control *cc)
641 unsigned int count[2] = { 0, };
643 if (list_empty(&cc->migratepages))
646 list_for_each_entry(page, &cc->migratepages, lru)
647 count[!!page_is_file_cache(page)]++;
649 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON, count[0]);
650 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, count[1]);
653 /* Similar to reclaim, but different enough that they don't share logic */
654 static bool too_many_isolated(struct zone *zone)
656 unsigned long active, inactive, isolated;
658 inactive = node_page_state(zone->zone_pgdat, NR_INACTIVE_FILE) +
659 node_page_state(zone->zone_pgdat, NR_INACTIVE_ANON);
660 active = node_page_state(zone->zone_pgdat, NR_ACTIVE_FILE) +
661 node_page_state(zone->zone_pgdat, NR_ACTIVE_ANON);
662 isolated = node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE) +
663 node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON);
665 return isolated > (inactive + active) / 2;
669 * isolate_migratepages_block() - isolate all migrate-able pages within
671 * @cc: Compaction control structure.
672 * @low_pfn: The first PFN to isolate
673 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
674 * @isolate_mode: Isolation mode to be used.
676 * Isolate all pages that can be migrated from the range specified by
677 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
678 * Returns zero if there is a fatal signal pending, otherwise PFN of the
679 * first page that was not scanned (which may be both less, equal to or more
682 * The pages are isolated on cc->migratepages list (not required to be empty),
683 * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
684 * is neither read nor updated.
687 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
688 unsigned long end_pfn, isolate_mode_t isolate_mode)
690 struct zone *zone = cc->zone;
691 unsigned long nr_scanned = 0, nr_isolated = 0;
692 struct lruvec *lruvec;
693 unsigned long flags = 0;
695 struct page *page = NULL, *valid_page = NULL;
696 unsigned long start_pfn = low_pfn;
697 bool skip_on_failure = false;
698 unsigned long next_skip_pfn = 0;
701 * Ensure that there are not too many pages isolated from the LRU
702 * list by either parallel reclaimers or compaction. If there are,
703 * delay for some time until fewer pages are isolated
705 while (unlikely(too_many_isolated(zone))) {
706 /* async migration should just abort */
707 if (cc->mode == MIGRATE_ASYNC)
710 congestion_wait(BLK_RW_ASYNC, HZ/10);
712 if (fatal_signal_pending(current))
716 if (compact_should_abort(cc))
719 if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
720 skip_on_failure = true;
721 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
724 /* Time to isolate some pages for migration */
725 for (; low_pfn < end_pfn; low_pfn++) {
727 if (skip_on_failure && low_pfn >= next_skip_pfn) {
729 * We have isolated all migration candidates in the
730 * previous order-aligned block, and did not skip it due
731 * to failure. We should migrate the pages now and
732 * hopefully succeed compaction.
738 * We failed to isolate in the previous order-aligned
739 * block. Set the new boundary to the end of the
740 * current block. Note we can't simply increase
741 * next_skip_pfn by 1 << order, as low_pfn might have
742 * been incremented by a higher number due to skipping
743 * a compound or a high-order buddy page in the
744 * previous loop iteration.
746 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
750 * Periodically drop the lock (if held) regardless of its
751 * contention, to give chance to IRQs. Abort async compaction
754 if (!(low_pfn % SWAP_CLUSTER_MAX)
755 && compact_unlock_should_abort(zone_lru_lock(zone), flags,
759 if (!pfn_valid_within(low_pfn))
763 page = pfn_to_page(low_pfn);
769 * Skip if free. We read page order here without zone lock
770 * which is generally unsafe, but the race window is small and
771 * the worst thing that can happen is that we skip some
772 * potential isolation targets.
774 if (PageBuddy(page)) {
775 unsigned long freepage_order = page_order_unsafe(page);
778 * Without lock, we cannot be sure that what we got is
779 * a valid page order. Consider only values in the
780 * valid order range to prevent low_pfn overflow.
782 if (freepage_order > 0 && freepage_order < MAX_ORDER)
783 low_pfn += (1UL << freepage_order) - 1;
788 * Regardless of being on LRU, compound pages such as THP and
789 * hugetlbfs are not to be compacted. We can potentially save
790 * a lot of iterations if we skip them at once. The check is
791 * racy, but we can consider only valid values and the only
792 * danger is skipping too much.
794 if (PageCompound(page)) {
795 unsigned int comp_order = compound_order(page);
797 if (likely(comp_order < MAX_ORDER))
798 low_pfn += (1UL << comp_order) - 1;
804 * Check may be lockless but that's ok as we recheck later.
805 * It's possible to migrate LRU and non-lru movable pages.
806 * Skip any other type of page
808 if (!PageLRU(page)) {
810 * __PageMovable can return false positive so we need
811 * to verify it under page_lock.
813 if (unlikely(__PageMovable(page)) &&
814 !PageIsolated(page)) {
816 spin_unlock_irqrestore(zone_lru_lock(zone),
821 if (isolate_movable_page(page, isolate_mode))
822 goto isolate_success;
829 * Migration will fail if an anonymous page is pinned in memory,
830 * so avoid taking lru_lock and isolating it unnecessarily in an
831 * admittedly racy check.
833 if (!page_mapping(page) &&
834 page_count(page) > page_mapcount(page))
837 /* If we already hold the lock, we can skip some rechecking */
839 locked = compact_trylock_irqsave(zone_lru_lock(zone),
844 /* Recheck PageLRU and PageCompound under lock */
849 * Page become compound since the non-locked check,
850 * and it's on LRU. It can only be a THP so the order
851 * is safe to read and it's 0 for tail pages.
853 if (unlikely(PageCompound(page))) {
854 low_pfn += (1UL << compound_order(page)) - 1;
859 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
861 /* Try isolate the page */
862 if (__isolate_lru_page(page, isolate_mode) != 0)
865 VM_BUG_ON_PAGE(PageCompound(page), page);
867 /* Successfully isolated */
868 del_page_from_lru_list(page, lruvec, page_lru(page));
871 list_add(&page->lru, &cc->migratepages);
872 cc->nr_migratepages++;
876 * Record where we could have freed pages by migration and not
877 * yet flushed them to buddy allocator.
878 * - this is the lowest page that was isolated and likely be
879 * then freed by migration.
881 if (!cc->last_migrated_pfn)
882 cc->last_migrated_pfn = low_pfn;
884 /* Avoid isolating too much */
885 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
892 if (!skip_on_failure)
896 * We have isolated some pages, but then failed. Release them
897 * instead of migrating, as we cannot form the cc->order buddy
902 spin_unlock_irqrestore(zone_lru_lock(zone), flags);
905 acct_isolated(zone, cc);
906 putback_movable_pages(&cc->migratepages);
907 cc->nr_migratepages = 0;
908 cc->last_migrated_pfn = 0;
912 if (low_pfn < next_skip_pfn) {
913 low_pfn = next_skip_pfn - 1;
915 * The check near the loop beginning would have updated
916 * next_skip_pfn too, but this is a bit simpler.
918 next_skip_pfn += 1UL << cc->order;
923 * The PageBuddy() check could have potentially brought us outside
924 * the range to be scanned.
926 if (unlikely(low_pfn > end_pfn))
930 spin_unlock_irqrestore(zone_lru_lock(zone), flags);
933 * Update the pageblock-skip information and cached scanner pfn,
934 * if the whole pageblock was scanned without isolating any page.
936 if (low_pfn == end_pfn)
937 update_pageblock_skip(cc, valid_page, nr_isolated, true);
939 trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
940 nr_scanned, nr_isolated);
942 count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
944 count_compact_events(COMPACTISOLATED, nr_isolated);
950 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
951 * @cc: Compaction control structure.
952 * @start_pfn: The first PFN to start isolating.
953 * @end_pfn: The one-past-last PFN.
955 * Returns zero if isolation fails fatally due to e.g. pending signal.
956 * Otherwise, function returns one-past-the-last PFN of isolated page
957 * (which may be greater than end_pfn if end fell in a middle of a THP page).
960 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
961 unsigned long end_pfn)
963 unsigned long pfn, block_start_pfn, block_end_pfn;
965 /* Scan block by block. First and last block may be incomplete */
967 block_start_pfn = pageblock_start_pfn(pfn);
968 if (block_start_pfn < cc->zone->zone_start_pfn)
969 block_start_pfn = cc->zone->zone_start_pfn;
970 block_end_pfn = pageblock_end_pfn(pfn);
972 for (; pfn < end_pfn; pfn = block_end_pfn,
973 block_start_pfn = block_end_pfn,
974 block_end_pfn += pageblock_nr_pages) {
976 block_end_pfn = min(block_end_pfn, end_pfn);
978 if (!pageblock_pfn_to_page(block_start_pfn,
979 block_end_pfn, cc->zone))
982 pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
983 ISOLATE_UNEVICTABLE);
988 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
991 acct_isolated(cc->zone, cc);
996 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
997 #ifdef CONFIG_COMPACTION
999 /* Returns true if the page is within a block suitable for migration to */
1000 static bool suitable_migration_target(struct page *page)
1002 /* If the page is a large free page, then disallow migration */
1003 if (PageBuddy(page)) {
1005 * We are checking page_order without zone->lock taken. But
1006 * the only small danger is that we skip a potentially suitable
1007 * pageblock, so it's not worth to check order for valid range.
1009 if (page_order_unsafe(page) >= pageblock_order)
1013 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1014 if (migrate_async_suitable(get_pageblock_migratetype(page)))
1017 /* Otherwise skip the block */
1022 * Test whether the free scanner has reached the same or lower pageblock than
1023 * the migration scanner, and compaction should thus terminate.
1025 static inline bool compact_scanners_met(struct compact_control *cc)
1027 return (cc->free_pfn >> pageblock_order)
1028 <= (cc->migrate_pfn >> pageblock_order);
1032 * Based on information in the current compact_control, find blocks
1033 * suitable for isolating free pages from and then isolate them.
1035 static void isolate_freepages(struct compact_control *cc)
1037 struct zone *zone = cc->zone;
1039 unsigned long block_start_pfn; /* start of current pageblock */
1040 unsigned long isolate_start_pfn; /* exact pfn we start at */
1041 unsigned long block_end_pfn; /* end of current pageblock */
1042 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
1043 struct list_head *freelist = &cc->freepages;
1046 * Initialise the free scanner. The starting point is where we last
1047 * successfully isolated from, zone-cached value, or the end of the
1048 * zone when isolating for the first time. For looping we also need
1049 * this pfn aligned down to the pageblock boundary, because we do
1050 * block_start_pfn -= pageblock_nr_pages in the for loop.
1051 * For ending point, take care when isolating in last pageblock of a
1052 * a zone which ends in the middle of a pageblock.
1053 * The low boundary is the end of the pageblock the migration scanner
1056 isolate_start_pfn = cc->free_pfn;
1057 block_start_pfn = pageblock_start_pfn(cc->free_pfn);
1058 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1059 zone_end_pfn(zone));
1060 low_pfn = pageblock_end_pfn(cc->migrate_pfn);
1063 * Isolate free pages until enough are available to migrate the
1064 * pages on cc->migratepages. We stop searching if the migrate
1065 * and free page scanners meet or enough free pages are isolated.
1067 for (; block_start_pfn >= low_pfn;
1068 block_end_pfn = block_start_pfn,
1069 block_start_pfn -= pageblock_nr_pages,
1070 isolate_start_pfn = block_start_pfn) {
1072 * This can iterate a massively long zone without finding any
1073 * suitable migration targets, so periodically check if we need
1074 * to schedule, or even abort async compaction.
1076 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1077 && compact_should_abort(cc))
1080 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1085 /* Check the block is suitable for migration */
1086 if (!suitable_migration_target(page))
1089 /* If isolation recently failed, do not retry */
1090 if (!isolation_suitable(cc, page))
1093 /* Found a block suitable for isolating free pages from. */
1094 isolate_freepages_block(cc, &isolate_start_pfn, block_end_pfn,
1098 * If we isolated enough freepages, or aborted due to lock
1099 * contention, terminate.
1101 if ((cc->nr_freepages >= cc->nr_migratepages)
1103 if (isolate_start_pfn >= block_end_pfn) {
1105 * Restart at previous pageblock if more
1106 * freepages can be isolated next time.
1109 block_start_pfn - pageblock_nr_pages;
1112 } else if (isolate_start_pfn < block_end_pfn) {
1114 * If isolation failed early, do not continue
1121 /* __isolate_free_page() does not map the pages */
1122 map_pages(freelist);
1125 * Record where the free scanner will restart next time. Either we
1126 * broke from the loop and set isolate_start_pfn based on the last
1127 * call to isolate_freepages_block(), or we met the migration scanner
1128 * and the loop terminated due to isolate_start_pfn < low_pfn
1130 cc->free_pfn = isolate_start_pfn;
1134 * This is a migrate-callback that "allocates" freepages by taking pages
1135 * from the isolated freelists in the block we are migrating to.
1137 static struct page *compaction_alloc(struct page *migratepage,
1141 struct compact_control *cc = (struct compact_control *)data;
1142 struct page *freepage;
1145 * Isolate free pages if necessary, and if we are not aborting due to
1148 if (list_empty(&cc->freepages)) {
1150 isolate_freepages(cc);
1152 if (list_empty(&cc->freepages))
1156 freepage = list_entry(cc->freepages.next, struct page, lru);
1157 list_del(&freepage->lru);
1164 * This is a migrate-callback that "frees" freepages back to the isolated
1165 * freelist. All pages on the freelist are from the same zone, so there is no
1166 * special handling needed for NUMA.
1168 static void compaction_free(struct page *page, unsigned long data)
1170 struct compact_control *cc = (struct compact_control *)data;
1172 list_add(&page->lru, &cc->freepages);
1176 /* possible outcome of isolate_migratepages */
1178 ISOLATE_ABORT, /* Abort compaction now */
1179 ISOLATE_NONE, /* No pages isolated, continue scanning */
1180 ISOLATE_SUCCESS, /* Pages isolated, migrate */
1181 } isolate_migrate_t;
1184 * Allow userspace to control policy on scanning the unevictable LRU for
1185 * compactable pages.
1187 int sysctl_compact_unevictable_allowed __read_mostly = 1;
1190 * Isolate all pages that can be migrated from the first suitable block,
1191 * starting at the block pointed to by the migrate scanner pfn within
1194 static isolate_migrate_t isolate_migratepages(struct zone *zone,
1195 struct compact_control *cc)
1197 unsigned long block_start_pfn;
1198 unsigned long block_end_pfn;
1199 unsigned long low_pfn;
1201 const isolate_mode_t isolate_mode =
1202 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1203 (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1206 * Start at where we last stopped, or beginning of the zone as
1207 * initialized by compact_zone()
1209 low_pfn = cc->migrate_pfn;
1210 block_start_pfn = pageblock_start_pfn(low_pfn);
1211 if (block_start_pfn < zone->zone_start_pfn)
1212 block_start_pfn = zone->zone_start_pfn;
1214 /* Only scan within a pageblock boundary */
1215 block_end_pfn = pageblock_end_pfn(low_pfn);
1218 * Iterate over whole pageblocks until we find the first suitable.
1219 * Do not cross the free scanner.
1221 for (; block_end_pfn <= cc->free_pfn;
1222 low_pfn = block_end_pfn,
1223 block_start_pfn = block_end_pfn,
1224 block_end_pfn += pageblock_nr_pages) {
1227 * This can potentially iterate a massively long zone with
1228 * many pageblocks unsuitable, so periodically check if we
1229 * need to schedule, or even abort async compaction.
1231 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1232 && compact_should_abort(cc))
1235 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1240 /* If isolation recently failed, do not retry */
1241 if (!isolation_suitable(cc, page))
1245 * For async compaction, also only scan in MOVABLE blocks.
1246 * Async compaction is optimistic to see if the minimum amount
1247 * of work satisfies the allocation.
1249 if (cc->mode == MIGRATE_ASYNC &&
1250 !migrate_async_suitable(get_pageblock_migratetype(page)))
1253 /* Perform the isolation */
1254 low_pfn = isolate_migratepages_block(cc, low_pfn,
1255 block_end_pfn, isolate_mode);
1257 if (!low_pfn || cc->contended) {
1258 acct_isolated(zone, cc);
1259 return ISOLATE_ABORT;
1263 * Either we isolated something and proceed with migration. Or
1264 * we failed and compact_zone should decide if we should
1270 acct_isolated(zone, cc);
1271 /* Record where migration scanner will be restarted. */
1272 cc->migrate_pfn = low_pfn;
1274 return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
1278 * order == -1 is expected when compacting via
1279 * /proc/sys/vm/compact_memory
1281 static inline bool is_via_compact_memory(int order)
1286 static enum compact_result __compact_finished(struct zone *zone, struct compact_control *cc,
1287 const int migratetype)
1290 unsigned long watermark;
1292 if (cc->contended || fatal_signal_pending(current))
1293 return COMPACT_CONTENDED;
1295 /* Compaction run completes if the migrate and free scanner meet */
1296 if (compact_scanners_met(cc)) {
1297 /* Let the next compaction start anew. */
1298 reset_cached_positions(zone);
1301 * Mark that the PG_migrate_skip information should be cleared
1302 * by kswapd when it goes to sleep. kcompactd does not set the
1303 * flag itself as the decision to be clear should be directly
1304 * based on an allocation request.
1306 if (cc->direct_compaction)
1307 zone->compact_blockskip_flush = true;
1310 return COMPACT_COMPLETE;
1312 return COMPACT_PARTIAL_SKIPPED;
1315 if (is_via_compact_memory(cc->order))
1316 return COMPACT_CONTINUE;
1318 /* Compaction run is not finished if the watermark is not met */
1319 watermark = low_wmark_pages(zone);
1321 if (!zone_watermark_ok(zone, cc->order, watermark, cc->classzone_idx,
1323 return COMPACT_CONTINUE;
1325 /* Direct compactor: Is a suitable page free? */
1326 for (order = cc->order; order < MAX_ORDER; order++) {
1327 struct free_area *area = &zone->free_area[order];
1330 /* Job done if page is free of the right migratetype */
1331 if (!list_empty(&area->free_list[migratetype]))
1332 return COMPACT_SUCCESS;
1335 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
1336 if (migratetype == MIGRATE_MOVABLE &&
1337 !list_empty(&area->free_list[MIGRATE_CMA]))
1338 return COMPACT_SUCCESS;
1341 * Job done if allocation would steal freepages from
1342 * other migratetype buddy lists.
1344 if (find_suitable_fallback(area, order, migratetype,
1345 true, &can_steal) != -1)
1346 return COMPACT_SUCCESS;
1349 return COMPACT_NO_SUITABLE_PAGE;
1352 static enum compact_result compact_finished(struct zone *zone,
1353 struct compact_control *cc,
1354 const int migratetype)
1358 ret = __compact_finished(zone, cc, migratetype);
1359 trace_mm_compaction_finished(zone, cc->order, ret);
1360 if (ret == COMPACT_NO_SUITABLE_PAGE)
1361 ret = COMPACT_CONTINUE;
1367 * compaction_suitable: Is this suitable to run compaction on this zone now?
1369 * COMPACT_SKIPPED - If there are too few free pages for compaction
1370 * COMPACT_SUCCESS - If the allocation would succeed without compaction
1371 * COMPACT_CONTINUE - If compaction should run now
1373 static enum compact_result __compaction_suitable(struct zone *zone, int order,
1374 unsigned int alloc_flags,
1376 unsigned long wmark_target)
1379 unsigned long watermark;
1381 if (is_via_compact_memory(order))
1382 return COMPACT_CONTINUE;
1384 watermark = low_wmark_pages(zone);
1386 * If watermarks for high-order allocation are already met, there
1387 * should be no need for compaction at all.
1389 if (zone_watermark_ok(zone, order, watermark, classzone_idx,
1391 return COMPACT_SUCCESS;
1394 * Watermarks for order-0 must be met for compaction. Note the 2UL.
1395 * This is because during migration, copies of pages need to be
1396 * allocated and for a short time, the footprint is higher
1398 watermark += (2UL << order);
1399 if (!__zone_watermark_ok(zone, 0, watermark, classzone_idx,
1400 alloc_flags, wmark_target))
1401 return COMPACT_SKIPPED;
1404 * fragmentation index determines if allocation failures are due to
1405 * low memory or external fragmentation
1407 * index of -1000 would imply allocations might succeed depending on
1408 * watermarks, but we already failed the high-order watermark check
1409 * index towards 0 implies failure is due to lack of memory
1410 * index towards 1000 implies failure is due to fragmentation
1412 * Only compact if a failure would be due to fragmentation.
1414 fragindex = fragmentation_index(zone, order);
1415 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
1416 return COMPACT_NOT_SUITABLE_ZONE;
1418 return COMPACT_CONTINUE;
1421 enum compact_result compaction_suitable(struct zone *zone, int order,
1422 unsigned int alloc_flags,
1425 enum compact_result ret;
1427 ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx,
1428 zone_page_state(zone, NR_FREE_PAGES));
1429 trace_mm_compaction_suitable(zone, order, ret);
1430 if (ret == COMPACT_NOT_SUITABLE_ZONE)
1431 ret = COMPACT_SKIPPED;
1436 bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
1443 * Make sure at least one zone would pass __compaction_suitable if we continue
1444 * retrying the reclaim.
1446 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1448 unsigned long available;
1449 enum compact_result compact_result;
1452 * Do not consider all the reclaimable memory because we do not
1453 * want to trash just for a single high order allocation which
1454 * is even not guaranteed to appear even if __compaction_suitable
1455 * is happy about the watermark check.
1457 available = zone_reclaimable_pages(zone) / order;
1458 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
1459 compact_result = __compaction_suitable(zone, order, alloc_flags,
1460 ac_classzone_idx(ac), available);
1461 if (compact_result != COMPACT_SKIPPED &&
1462 compact_result != COMPACT_NOT_SUITABLE_ZONE)
1469 static enum compact_result compact_zone(struct zone *zone, struct compact_control *cc)
1471 enum compact_result ret;
1472 unsigned long start_pfn = zone->zone_start_pfn;
1473 unsigned long end_pfn = zone_end_pfn(zone);
1474 const int migratetype = gfpflags_to_migratetype(cc->gfp_mask);
1475 const bool sync = cc->mode != MIGRATE_ASYNC;
1477 ret = compaction_suitable(zone, cc->order, cc->alloc_flags,
1479 /* Compaction is likely to fail */
1480 if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED)
1483 /* huh, compaction_suitable is returning something unexpected */
1484 VM_BUG_ON(ret != COMPACT_CONTINUE);
1487 * Clear pageblock skip if there were failures recently and compaction
1488 * is about to be retried after being deferred.
1490 if (compaction_restarting(zone, cc->order))
1491 __reset_isolation_suitable(zone);
1494 * Setup to move all movable pages to the end of the zone. Used cached
1495 * information on where the scanners should start (unless we explicitly
1496 * want to compact the whole zone), but check that it is initialised
1497 * by ensuring the values are within zone boundaries.
1499 if (cc->whole_zone) {
1500 cc->migrate_pfn = start_pfn;
1501 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
1503 cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync];
1504 cc->free_pfn = zone->compact_cached_free_pfn;
1505 if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
1506 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
1507 zone->compact_cached_free_pfn = cc->free_pfn;
1509 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
1510 cc->migrate_pfn = start_pfn;
1511 zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
1512 zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
1515 if (cc->migrate_pfn == start_pfn)
1516 cc->whole_zone = true;
1519 cc->last_migrated_pfn = 0;
1521 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
1522 cc->free_pfn, end_pfn, sync);
1524 migrate_prep_local();
1526 while ((ret = compact_finished(zone, cc, migratetype)) ==
1530 switch (isolate_migratepages(zone, cc)) {
1532 ret = COMPACT_CONTENDED;
1533 putback_movable_pages(&cc->migratepages);
1534 cc->nr_migratepages = 0;
1538 * We haven't isolated and migrated anything, but
1539 * there might still be unflushed migrations from
1540 * previous cc->order aligned block.
1543 case ISOLATE_SUCCESS:
1547 err = migrate_pages(&cc->migratepages, compaction_alloc,
1548 compaction_free, (unsigned long)cc, cc->mode,
1551 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
1554 /* All pages were either migrated or will be released */
1555 cc->nr_migratepages = 0;
1557 putback_movable_pages(&cc->migratepages);
1559 * migrate_pages() may return -ENOMEM when scanners meet
1560 * and we want compact_finished() to detect it
1562 if (err == -ENOMEM && !compact_scanners_met(cc)) {
1563 ret = COMPACT_CONTENDED;
1567 * We failed to migrate at least one page in the current
1568 * order-aligned block, so skip the rest of it.
1570 if (cc->direct_compaction &&
1571 (cc->mode == MIGRATE_ASYNC)) {
1572 cc->migrate_pfn = block_end_pfn(
1573 cc->migrate_pfn - 1, cc->order);
1574 /* Draining pcplists is useless in this case */
1575 cc->last_migrated_pfn = 0;
1582 * Has the migration scanner moved away from the previous
1583 * cc->order aligned block where we migrated from? If yes,
1584 * flush the pages that were freed, so that they can merge and
1585 * compact_finished() can detect immediately if allocation
1588 if (cc->order > 0 && cc->last_migrated_pfn) {
1590 unsigned long current_block_start =
1591 block_start_pfn(cc->migrate_pfn, cc->order);
1593 if (cc->last_migrated_pfn < current_block_start) {
1595 lru_add_drain_cpu(cpu);
1596 drain_local_pages(zone);
1598 /* No more flushing until we migrate again */
1599 cc->last_migrated_pfn = 0;
1607 * Release free pages and update where the free scanner should restart,
1608 * so we don't leave any returned pages behind in the next attempt.
1610 if (cc->nr_freepages > 0) {
1611 unsigned long free_pfn = release_freepages(&cc->freepages);
1613 cc->nr_freepages = 0;
1614 VM_BUG_ON(free_pfn == 0);
1615 /* The cached pfn is always the first in a pageblock */
1616 free_pfn = pageblock_start_pfn(free_pfn);
1618 * Only go back, not forward. The cached pfn might have been
1619 * already reset to zone end in compact_finished()
1621 if (free_pfn > zone->compact_cached_free_pfn)
1622 zone->compact_cached_free_pfn = free_pfn;
1625 trace_mm_compaction_end(start_pfn, cc->migrate_pfn,
1626 cc->free_pfn, end_pfn, sync, ret);
1631 static enum compact_result compact_zone_order(struct zone *zone, int order,
1632 gfp_t gfp_mask, enum compact_priority prio,
1633 unsigned int alloc_flags, int classzone_idx)
1635 enum compact_result ret;
1636 struct compact_control cc = {
1638 .nr_migratepages = 0,
1640 .gfp_mask = gfp_mask,
1642 .mode = (prio == COMPACT_PRIO_ASYNC) ?
1643 MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT,
1644 .alloc_flags = alloc_flags,
1645 .classzone_idx = classzone_idx,
1646 .direct_compaction = true,
1648 INIT_LIST_HEAD(&cc.freepages);
1649 INIT_LIST_HEAD(&cc.migratepages);
1651 ret = compact_zone(zone, &cc);
1653 VM_BUG_ON(!list_empty(&cc.freepages));
1654 VM_BUG_ON(!list_empty(&cc.migratepages));
1659 int sysctl_extfrag_threshold = 500;
1662 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1663 * @gfp_mask: The GFP mask of the current allocation
1664 * @order: The order of the current allocation
1665 * @alloc_flags: The allocation flags of the current allocation
1666 * @ac: The context of current allocation
1667 * @mode: The migration mode for async, sync light, or sync migration
1669 * This is the main entry point for direct page compaction.
1671 enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
1672 unsigned int alloc_flags, const struct alloc_context *ac,
1673 enum compact_priority prio)
1675 int may_enter_fs = gfp_mask & __GFP_FS;
1676 int may_perform_io = gfp_mask & __GFP_IO;
1679 enum compact_result rc = COMPACT_SKIPPED;
1681 /* Check if the GFP flags allow compaction */
1682 if (!may_enter_fs || !may_perform_io)
1683 return COMPACT_SKIPPED;
1685 trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
1687 /* Compact each zone in the list */
1688 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1690 enum compact_result status;
1692 if (compaction_deferred(zone, order)) {
1693 rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
1697 status = compact_zone_order(zone, order, gfp_mask, prio,
1698 alloc_flags, ac_classzone_idx(ac));
1699 rc = max(status, rc);
1701 /* If a normal allocation would succeed, stop compacting */
1702 if (zone_watermark_ok(zone, order, low_wmark_pages(zone),
1703 ac_classzone_idx(ac), alloc_flags)) {
1705 * We think the allocation will succeed in this zone,
1706 * but it is not certain, hence the false. The caller
1707 * will repeat this with true if allocation indeed
1708 * succeeds in this zone.
1710 compaction_defer_reset(zone, order, false);
1715 if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
1716 status == COMPACT_PARTIAL_SKIPPED))
1718 * We think that allocation won't succeed in this zone
1719 * so we defer compaction there. If it ends up
1720 * succeeding after all, it will be reset.
1722 defer_compaction(zone, order);
1725 * We might have stopped compacting due to need_resched() in
1726 * async compaction, or due to a fatal signal detected. In that
1727 * case do not try further zones
1729 if ((prio == COMPACT_PRIO_ASYNC && need_resched())
1730 || fatal_signal_pending(current))
1738 /* Compact all zones within a node */
1739 static void compact_node(int nid)
1741 pg_data_t *pgdat = NODE_DATA(nid);
1744 struct compact_control cc = {
1746 .mode = MIGRATE_SYNC,
1747 .ignore_skip_hint = true,
1752 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1754 zone = &pgdat->node_zones[zoneid];
1755 if (!populated_zone(zone))
1758 cc.nr_freepages = 0;
1759 cc.nr_migratepages = 0;
1761 INIT_LIST_HEAD(&cc.freepages);
1762 INIT_LIST_HEAD(&cc.migratepages);
1764 compact_zone(zone, &cc);
1766 VM_BUG_ON(!list_empty(&cc.freepages));
1767 VM_BUG_ON(!list_empty(&cc.migratepages));
1771 /* Compact all nodes in the system */
1772 static void compact_nodes(void)
1776 /* Flush pending updates to the LRU lists */
1777 lru_add_drain_all();
1779 for_each_online_node(nid)
1783 /* The written value is actually unused, all memory is compacted */
1784 int sysctl_compact_memory;
1787 * This is the entry point for compacting all nodes via
1788 * /proc/sys/vm/compact_memory
1790 int sysctl_compaction_handler(struct ctl_table *table, int write,
1791 void __user *buffer, size_t *length, loff_t *ppos)
1799 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1800 void __user *buffer, size_t *length, loff_t *ppos)
1802 proc_dointvec_minmax(table, write, buffer, length, ppos);
1807 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1808 static ssize_t sysfs_compact_node(struct device *dev,
1809 struct device_attribute *attr,
1810 const char *buf, size_t count)
1814 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1815 /* Flush pending updates to the LRU lists */
1816 lru_add_drain_all();
1823 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1825 int compaction_register_node(struct node *node)
1827 return device_create_file(&node->dev, &dev_attr_compact);
1830 void compaction_unregister_node(struct node *node)
1832 return device_remove_file(&node->dev, &dev_attr_compact);
1834 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1836 static inline bool kcompactd_work_requested(pg_data_t *pgdat)
1838 return pgdat->kcompactd_max_order > 0 || kthread_should_stop();
1841 static bool kcompactd_node_suitable(pg_data_t *pgdat)
1845 enum zone_type classzone_idx = pgdat->kcompactd_classzone_idx;
1847 for (zoneid = 0; zoneid <= classzone_idx; zoneid++) {
1848 zone = &pgdat->node_zones[zoneid];
1850 if (!populated_zone(zone))
1853 if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0,
1854 classzone_idx) == COMPACT_CONTINUE)
1861 static void kcompactd_do_work(pg_data_t *pgdat)
1864 * With no special task, compact all zones so that a page of requested
1865 * order is allocatable.
1869 struct compact_control cc = {
1870 .order = pgdat->kcompactd_max_order,
1871 .classzone_idx = pgdat->kcompactd_classzone_idx,
1872 .mode = MIGRATE_SYNC_LIGHT,
1873 .ignore_skip_hint = true,
1876 bool success = false;
1878 trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
1880 count_vm_event(KCOMPACTD_WAKE);
1882 for (zoneid = 0; zoneid <= cc.classzone_idx; zoneid++) {
1885 zone = &pgdat->node_zones[zoneid];
1886 if (!populated_zone(zone))
1889 if (compaction_deferred(zone, cc.order))
1892 if (compaction_suitable(zone, cc.order, 0, zoneid) !=
1896 cc.nr_freepages = 0;
1897 cc.nr_migratepages = 0;
1899 INIT_LIST_HEAD(&cc.freepages);
1900 INIT_LIST_HEAD(&cc.migratepages);
1902 if (kthread_should_stop())
1904 status = compact_zone(zone, &cc);
1906 if (zone_watermark_ok(zone, cc.order, low_wmark_pages(zone),
1907 cc.classzone_idx, 0)) {
1909 compaction_defer_reset(zone, cc.order, false);
1910 } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
1912 * We use sync migration mode here, so we defer like
1913 * sync direct compaction does.
1915 defer_compaction(zone, cc.order);
1918 VM_BUG_ON(!list_empty(&cc.freepages));
1919 VM_BUG_ON(!list_empty(&cc.migratepages));
1923 * Regardless of success, we are done until woken up next. But remember
1924 * the requested order/classzone_idx in case it was higher/tighter than
1927 if (pgdat->kcompactd_max_order <= cc.order)
1928 pgdat->kcompactd_max_order = 0;
1929 if (pgdat->kcompactd_classzone_idx >= cc.classzone_idx)
1930 pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
1933 void wakeup_kcompactd(pg_data_t *pgdat, int order, int classzone_idx)
1938 if (pgdat->kcompactd_max_order < order)
1939 pgdat->kcompactd_max_order = order;
1941 if (pgdat->kcompactd_classzone_idx > classzone_idx)
1942 pgdat->kcompactd_classzone_idx = classzone_idx;
1944 if (!waitqueue_active(&pgdat->kcompactd_wait))
1947 if (!kcompactd_node_suitable(pgdat))
1950 trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
1952 wake_up_interruptible(&pgdat->kcompactd_wait);
1956 * The background compaction daemon, started as a kernel thread
1957 * from the init process.
1959 static int kcompactd(void *p)
1961 pg_data_t *pgdat = (pg_data_t*)p;
1962 struct task_struct *tsk = current;
1964 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1966 if (!cpumask_empty(cpumask))
1967 set_cpus_allowed_ptr(tsk, cpumask);
1971 pgdat->kcompactd_max_order = 0;
1972 pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
1974 while (!kthread_should_stop()) {
1975 trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
1976 wait_event_freezable(pgdat->kcompactd_wait,
1977 kcompactd_work_requested(pgdat));
1979 kcompactd_do_work(pgdat);
1986 * This kcompactd start function will be called by init and node-hot-add.
1987 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
1989 int kcompactd_run(int nid)
1991 pg_data_t *pgdat = NODE_DATA(nid);
1994 if (pgdat->kcompactd)
1997 pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
1998 if (IS_ERR(pgdat->kcompactd)) {
1999 pr_err("Failed to start kcompactd on node %d\n", nid);
2000 ret = PTR_ERR(pgdat->kcompactd);
2001 pgdat->kcompactd = NULL;
2007 * Called by memory hotplug when all memory in a node is offlined. Caller must
2008 * hold mem_hotplug_begin/end().
2010 void kcompactd_stop(int nid)
2012 struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
2015 kthread_stop(kcompactd);
2016 NODE_DATA(nid)->kcompactd = NULL;
2021 * It's optimal to keep kcompactd on the same CPUs as their memory, but
2022 * not required for correctness. So if the last cpu in a node goes
2023 * away, we get changed to run anywhere: as the first one comes back,
2024 * restore their cpu bindings.
2026 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
2031 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2032 for_each_node_state(nid, N_MEMORY) {
2033 pg_data_t *pgdat = NODE_DATA(nid);
2034 const struct cpumask *mask;
2036 mask = cpumask_of_node(pgdat->node_id);
2038 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2039 /* One of our CPUs online: restore mask */
2040 set_cpus_allowed_ptr(pgdat->kcompactd, mask);
2046 static int __init kcompactd_init(void)
2050 for_each_node_state(nid, N_MEMORY)
2052 hotcpu_notifier(cpu_callback, 0);
2055 subsys_initcall(kcompactd_init)
2057 #endif /* CONFIG_COMPACTION */