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/balloon_compaction.h>
19 #include <linux/page-isolation.h>
20 #include <linux/kasan.h>
21 #include <linux/kthread.h>
22 #include <linux/freezer.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)
70 list_for_each_entry(page, list, lru) {
71 arch_alloc_page(page, 0);
72 kernel_map_pages(page, 1, 1);
73 kasan_alloc_pages(page, 0);
77 static inline bool migrate_async_suitable(int migratetype)
79 return is_migrate_cma(migratetype) || migratetype == MIGRATE_MOVABLE;
82 #ifdef CONFIG_COMPACTION
84 int PageMovable(struct page *page)
86 struct address_space *mapping;
88 VM_BUG_ON_PAGE(!PageLocked(page), page);
89 if (!__PageMovable(page))
92 mapping = page_mapping(page);
93 if (mapping && mapping->a_ops && mapping->a_ops->isolate_page)
98 EXPORT_SYMBOL(PageMovable);
100 void __SetPageMovable(struct page *page, struct address_space *mapping)
102 VM_BUG_ON_PAGE(!PageLocked(page), page);
103 VM_BUG_ON_PAGE((unsigned long)mapping & PAGE_MAPPING_MOVABLE, page);
104 page->mapping = (void *)((unsigned long)mapping | PAGE_MAPPING_MOVABLE);
106 EXPORT_SYMBOL(__SetPageMovable);
108 void __ClearPageMovable(struct page *page)
110 VM_BUG_ON_PAGE(!PageLocked(page), page);
111 VM_BUG_ON_PAGE(!PageMovable(page), page);
113 * Clear registered address_space val with keeping PAGE_MAPPING_MOVABLE
114 * flag so that VM can catch up released page by driver after isolation.
115 * With it, VM migration doesn't try to put it back.
117 page->mapping = (void *)((unsigned long)page->mapping &
118 PAGE_MAPPING_MOVABLE);
120 EXPORT_SYMBOL(__ClearPageMovable);
122 /* Do not skip compaction more than 64 times */
123 #define COMPACT_MAX_DEFER_SHIFT 6
126 * Compaction is deferred when compaction fails to result in a page
127 * allocation success. 1 << compact_defer_limit compactions are skipped up
128 * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
130 void defer_compaction(struct zone *zone, int order)
132 zone->compact_considered = 0;
133 zone->compact_defer_shift++;
135 if (order < zone->compact_order_failed)
136 zone->compact_order_failed = order;
138 if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
139 zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
141 trace_mm_compaction_defer_compaction(zone, order);
144 /* Returns true if compaction should be skipped this time */
145 bool compaction_deferred(struct zone *zone, int order)
147 unsigned long defer_limit = 1UL << zone->compact_defer_shift;
149 if (order < zone->compact_order_failed)
152 /* Avoid possible overflow */
153 if (++zone->compact_considered > defer_limit)
154 zone->compact_considered = defer_limit;
156 if (zone->compact_considered >= defer_limit)
159 trace_mm_compaction_deferred(zone, order);
165 * Update defer tracking counters after successful compaction of given order,
166 * which means an allocation either succeeded (alloc_success == true) or is
167 * expected to succeed.
169 void compaction_defer_reset(struct zone *zone, int order,
173 zone->compact_considered = 0;
174 zone->compact_defer_shift = 0;
176 if (order >= zone->compact_order_failed)
177 zone->compact_order_failed = order + 1;
179 trace_mm_compaction_defer_reset(zone, order);
182 /* Returns true if restarting compaction after many failures */
183 bool compaction_restarting(struct zone *zone, int order)
185 if (order < zone->compact_order_failed)
188 return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
189 zone->compact_considered >= 1UL << zone->compact_defer_shift;
192 /* Returns true if the pageblock should be scanned for pages to isolate. */
193 static inline bool isolation_suitable(struct compact_control *cc,
196 if (cc->ignore_skip_hint)
199 return !get_pageblock_skip(page);
202 static void reset_cached_positions(struct zone *zone)
204 zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
205 zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
206 zone->compact_cached_free_pfn =
207 pageblock_start_pfn(zone_end_pfn(zone) - 1);
211 * This function is called to clear all cached information on pageblocks that
212 * should be skipped for page isolation when the migrate and free page scanner
215 static void __reset_isolation_suitable(struct zone *zone)
217 unsigned long start_pfn = zone->zone_start_pfn;
218 unsigned long end_pfn = zone_end_pfn(zone);
221 zone->compact_blockskip_flush = false;
223 /* Walk the zone and mark every pageblock as suitable for isolation */
224 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
232 page = pfn_to_page(pfn);
233 if (zone != page_zone(page))
236 clear_pageblock_skip(page);
239 reset_cached_positions(zone);
242 void reset_isolation_suitable(pg_data_t *pgdat)
246 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
247 struct zone *zone = &pgdat->node_zones[zoneid];
248 if (!populated_zone(zone))
251 /* Only flush if a full compaction finished recently */
252 if (zone->compact_blockskip_flush)
253 __reset_isolation_suitable(zone);
258 * If no pages were isolated then mark this pageblock to be skipped in the
259 * future. The information is later cleared by __reset_isolation_suitable().
261 static void update_pageblock_skip(struct compact_control *cc,
262 struct page *page, unsigned long nr_isolated,
263 bool migrate_scanner)
265 struct zone *zone = cc->zone;
268 if (cc->ignore_skip_hint)
277 set_pageblock_skip(page);
279 pfn = page_to_pfn(page);
281 /* Update where async and sync compaction should restart */
282 if (migrate_scanner) {
283 if (pfn > zone->compact_cached_migrate_pfn[0])
284 zone->compact_cached_migrate_pfn[0] = pfn;
285 if (cc->mode != MIGRATE_ASYNC &&
286 pfn > zone->compact_cached_migrate_pfn[1])
287 zone->compact_cached_migrate_pfn[1] = pfn;
289 if (pfn < zone->compact_cached_free_pfn)
290 zone->compact_cached_free_pfn = pfn;
294 static inline bool isolation_suitable(struct compact_control *cc,
300 static void update_pageblock_skip(struct compact_control *cc,
301 struct page *page, unsigned long nr_isolated,
302 bool migrate_scanner)
305 #endif /* CONFIG_COMPACTION */
308 * Compaction requires the taking of some coarse locks that are potentially
309 * very heavily contended. For async compaction, back out if the lock cannot
310 * be taken immediately. For sync compaction, spin on the lock if needed.
312 * Returns true if the lock is held
313 * Returns false if the lock is not held and compaction should abort
315 static bool compact_trylock_irqsave(spinlock_t *lock, unsigned long *flags,
316 struct compact_control *cc)
318 if (cc->mode == MIGRATE_ASYNC) {
319 if (!spin_trylock_irqsave(lock, *flags)) {
320 cc->contended = COMPACT_CONTENDED_LOCK;
324 spin_lock_irqsave(lock, *flags);
331 * Compaction requires the taking of some coarse locks that are potentially
332 * very heavily contended. The lock should be periodically unlocked to avoid
333 * having disabled IRQs for a long time, even when there is nobody waiting on
334 * the lock. It might also be that allowing the IRQs will result in
335 * need_resched() becoming true. If scheduling is needed, async compaction
336 * aborts. Sync compaction schedules.
337 * Either compaction type will also abort if a fatal signal is pending.
338 * In either case if the lock was locked, it is dropped and not regained.
340 * Returns true if compaction should abort due to fatal signal pending, or
341 * async compaction due to need_resched()
342 * Returns false when compaction can continue (sync compaction might have
345 static bool compact_unlock_should_abort(spinlock_t *lock,
346 unsigned long flags, bool *locked, struct compact_control *cc)
349 spin_unlock_irqrestore(lock, flags);
353 if (fatal_signal_pending(current)) {
354 cc->contended = COMPACT_CONTENDED_SCHED;
358 if (need_resched()) {
359 if (cc->mode == MIGRATE_ASYNC) {
360 cc->contended = COMPACT_CONTENDED_SCHED;
370 * Aside from avoiding lock contention, compaction also periodically checks
371 * need_resched() and either schedules in sync compaction or aborts async
372 * compaction. This is similar to what compact_unlock_should_abort() does, but
373 * is used where no lock is concerned.
375 * Returns false when no scheduling was needed, or sync compaction scheduled.
376 * Returns true when async compaction should abort.
378 static inline bool compact_should_abort(struct compact_control *cc)
380 /* async compaction aborts if contended */
381 if (need_resched()) {
382 if (cc->mode == MIGRATE_ASYNC) {
383 cc->contended = COMPACT_CONTENDED_SCHED;
394 * Isolate free pages onto a private freelist. If @strict is true, will abort
395 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
396 * (even though it may still end up isolating some pages).
398 static unsigned long isolate_freepages_block(struct compact_control *cc,
399 unsigned long *start_pfn,
400 unsigned long end_pfn,
401 struct list_head *freelist,
404 int nr_scanned = 0, total_isolated = 0;
405 struct page *cursor, *valid_page = NULL;
406 unsigned long flags = 0;
408 unsigned long blockpfn = *start_pfn;
410 cursor = pfn_to_page(blockpfn);
412 /* Isolate free pages. */
413 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
415 struct page *page = cursor;
418 * Periodically drop the lock (if held) regardless of its
419 * contention, to give chance to IRQs. Abort if fatal signal
420 * pending or async compaction detects need_resched()
422 if (!(blockpfn % SWAP_CLUSTER_MAX)
423 && compact_unlock_should_abort(&cc->zone->lock, flags,
428 if (!pfn_valid_within(blockpfn))
435 * For compound pages such as THP and hugetlbfs, we can save
436 * potentially a lot of iterations if we skip them at once.
437 * The check is racy, but we can consider only valid values
438 * and the only danger is skipping too much.
440 if (PageCompound(page)) {
441 unsigned int comp_order = compound_order(page);
443 if (likely(comp_order < MAX_ORDER)) {
444 blockpfn += (1UL << comp_order) - 1;
445 cursor += (1UL << comp_order) - 1;
451 if (!PageBuddy(page))
455 * If we already hold the lock, we can skip some rechecking.
456 * Note that if we hold the lock now, checked_pageblock was
457 * already set in some previous iteration (or strict is true),
458 * so it is correct to skip the suitable migration target
463 * The zone lock must be held to isolate freepages.
464 * Unfortunately this is a very coarse lock and can be
465 * heavily contended if there are parallel allocations
466 * or parallel compactions. For async compaction do not
467 * spin on the lock and we acquire the lock as late as
470 locked = compact_trylock_irqsave(&cc->zone->lock,
475 /* Recheck this is a buddy page under lock */
476 if (!PageBuddy(page))
480 /* Found a free page, break it into order-0 pages */
481 isolated = split_free_page(page);
485 total_isolated += isolated;
486 cc->nr_freepages += isolated;
487 for (i = 0; i < isolated; i++) {
488 list_add(&page->lru, freelist);
491 if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
492 blockpfn += isolated;
495 /* Advance to the end of split page */
496 blockpfn += isolated - 1;
497 cursor += isolated - 1;
509 spin_unlock_irqrestore(&cc->zone->lock, flags);
512 * There is a tiny chance that we have read bogus compound_order(),
513 * so be careful to not go outside of the pageblock.
515 if (unlikely(blockpfn > end_pfn))
518 trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
519 nr_scanned, total_isolated);
521 /* Record how far we have got within the block */
522 *start_pfn = blockpfn;
525 * If strict isolation is requested by CMA then check that all the
526 * pages requested were isolated. If there were any failures, 0 is
527 * returned and CMA will fail.
529 if (strict && blockpfn < end_pfn)
532 /* Update the pageblock-skip if the whole pageblock was scanned */
533 if (blockpfn == end_pfn)
534 update_pageblock_skip(cc, valid_page, total_isolated, false);
536 count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
538 count_compact_events(COMPACTISOLATED, total_isolated);
539 return total_isolated;
543 * isolate_freepages_range() - isolate free pages.
544 * @start_pfn: The first PFN to start isolating.
545 * @end_pfn: The one-past-last PFN.
547 * Non-free pages, invalid PFNs, or zone boundaries within the
548 * [start_pfn, end_pfn) range are considered errors, cause function to
549 * undo its actions and return zero.
551 * Otherwise, function returns one-past-the-last PFN of isolated page
552 * (which may be greater then end_pfn if end fell in a middle of
556 isolate_freepages_range(struct compact_control *cc,
557 unsigned long start_pfn, unsigned long end_pfn)
559 unsigned long isolated, pfn, block_start_pfn, block_end_pfn;
563 block_start_pfn = pageblock_start_pfn(pfn);
564 if (block_start_pfn < cc->zone->zone_start_pfn)
565 block_start_pfn = cc->zone->zone_start_pfn;
566 block_end_pfn = pageblock_end_pfn(pfn);
568 for (; pfn < end_pfn; pfn += isolated,
569 block_start_pfn = block_end_pfn,
570 block_end_pfn += pageblock_nr_pages) {
571 /* Protect pfn from changing by isolate_freepages_block */
572 unsigned long isolate_start_pfn = pfn;
574 block_end_pfn = min(block_end_pfn, end_pfn);
577 * pfn could pass the block_end_pfn if isolated freepage
578 * is more than pageblock order. In this case, we adjust
579 * scanning range to right one.
581 if (pfn >= block_end_pfn) {
582 block_start_pfn = pageblock_start_pfn(pfn);
583 block_end_pfn = pageblock_end_pfn(pfn);
584 block_end_pfn = min(block_end_pfn, end_pfn);
587 if (!pageblock_pfn_to_page(block_start_pfn,
588 block_end_pfn, cc->zone))
591 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
592 block_end_pfn, &freelist, true);
595 * In strict mode, isolate_freepages_block() returns 0 if
596 * there are any holes in the block (ie. invalid PFNs or
603 * If we managed to isolate pages, it is always (1 << n) *
604 * pageblock_nr_pages for some non-negative n. (Max order
605 * page may span two pageblocks).
609 /* split_free_page does not map the pages */
610 map_pages(&freelist);
613 /* Loop terminated early, cleanup. */
614 release_freepages(&freelist);
618 /* We don't use freelists for anything. */
622 /* Update the number of anon and file isolated pages in the zone */
623 static void acct_isolated(struct zone *zone, struct compact_control *cc)
626 unsigned int count[2] = { 0, };
628 if (list_empty(&cc->migratepages))
631 list_for_each_entry(page, &cc->migratepages, lru)
632 count[!!page_is_file_cache(page)]++;
634 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
635 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
638 /* Similar to reclaim, but different enough that they don't share logic */
639 static bool too_many_isolated(struct zone *zone)
641 unsigned long active, inactive, isolated;
643 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
644 zone_page_state(zone, NR_INACTIVE_ANON);
645 active = zone_page_state(zone, NR_ACTIVE_FILE) +
646 zone_page_state(zone, NR_ACTIVE_ANON);
647 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
648 zone_page_state(zone, NR_ISOLATED_ANON);
650 return isolated > (inactive + active) / 2;
654 * isolate_migratepages_block() - isolate all migrate-able pages within
656 * @cc: Compaction control structure.
657 * @low_pfn: The first PFN to isolate
658 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
659 * @isolate_mode: Isolation mode to be used.
661 * Isolate all pages that can be migrated from the range specified by
662 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
663 * Returns zero if there is a fatal signal pending, otherwise PFN of the
664 * first page that was not scanned (which may be both less, equal to or more
667 * The pages are isolated on cc->migratepages list (not required to be empty),
668 * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
669 * is neither read nor updated.
672 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
673 unsigned long end_pfn, isolate_mode_t isolate_mode)
675 struct zone *zone = cc->zone;
676 unsigned long nr_scanned = 0, nr_isolated = 0;
677 struct lruvec *lruvec;
678 unsigned long flags = 0;
680 struct page *page = NULL, *valid_page = NULL;
681 unsigned long start_pfn = low_pfn;
682 bool skip_on_failure = false;
683 unsigned long next_skip_pfn = 0;
686 * Ensure that there are not too many pages isolated from the LRU
687 * list by either parallel reclaimers or compaction. If there are,
688 * delay for some time until fewer pages are isolated
690 while (unlikely(too_many_isolated(zone))) {
691 /* async migration should just abort */
692 if (cc->mode == MIGRATE_ASYNC)
695 congestion_wait(BLK_RW_ASYNC, HZ/10);
697 if (fatal_signal_pending(current))
701 if (compact_should_abort(cc))
704 if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
705 skip_on_failure = true;
706 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
709 /* Time to isolate some pages for migration */
710 for (; low_pfn < end_pfn; low_pfn++) {
712 if (skip_on_failure && low_pfn >= next_skip_pfn) {
714 * We have isolated all migration candidates in the
715 * previous order-aligned block, and did not skip it due
716 * to failure. We should migrate the pages now and
717 * hopefully succeed compaction.
723 * We failed to isolate in the previous order-aligned
724 * block. Set the new boundary to the end of the
725 * current block. Note we can't simply increase
726 * next_skip_pfn by 1 << order, as low_pfn might have
727 * been incremented by a higher number due to skipping
728 * a compound or a high-order buddy page in the
729 * previous loop iteration.
731 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
735 * Periodically drop the lock (if held) regardless of its
736 * contention, to give chance to IRQs. Abort async compaction
739 if (!(low_pfn % SWAP_CLUSTER_MAX)
740 && compact_unlock_should_abort(&zone->lru_lock, flags,
744 if (!pfn_valid_within(low_pfn))
748 page = pfn_to_page(low_pfn);
754 * Skip if free. We read page order here without zone lock
755 * which is generally unsafe, but the race window is small and
756 * the worst thing that can happen is that we skip some
757 * potential isolation targets.
759 if (PageBuddy(page)) {
760 unsigned long freepage_order = page_order_unsafe(page);
763 * Without lock, we cannot be sure that what we got is
764 * a valid page order. Consider only values in the
765 * valid order range to prevent low_pfn overflow.
767 if (freepage_order > 0 && freepage_order < MAX_ORDER)
768 low_pfn += (1UL << freepage_order) - 1;
773 * Regardless of being on LRU, compound pages such as THP and
774 * hugetlbfs are not to be compacted. We can potentially save
775 * a lot of iterations if we skip them at once. The check is
776 * racy, but we can consider only valid values and the only
777 * danger is skipping too much.
779 if (PageCompound(page)) {
780 unsigned int comp_order = compound_order(page);
782 if (likely(comp_order < MAX_ORDER))
783 low_pfn += (1UL << comp_order) - 1;
789 * Check may be lockless but that's ok as we recheck later.
790 * It's possible to migrate LRU and non-lru movable pages.
791 * Skip any other type of page
793 if (!PageLRU(page)) {
795 * __PageMovable can return false positive so we need
796 * to verify it under page_lock.
798 if (unlikely(__PageMovable(page)) &&
799 !PageIsolated(page)) {
801 spin_unlock_irqrestore(&zone->lru_lock,
806 if (isolate_movable_page(page, isolate_mode))
807 goto isolate_success;
814 * Migration will fail if an anonymous page is pinned in memory,
815 * so avoid taking lru_lock and isolating it unnecessarily in an
816 * admittedly racy check.
818 if (!page_mapping(page) &&
819 page_count(page) > page_mapcount(page))
822 /* If we already hold the lock, we can skip some rechecking */
824 locked = compact_trylock_irqsave(&zone->lru_lock,
829 /* Recheck PageLRU and PageCompound under lock */
834 * Page become compound since the non-locked check,
835 * and it's on LRU. It can only be a THP so the order
836 * is safe to read and it's 0 for tail pages.
838 if (unlikely(PageCompound(page))) {
839 low_pfn += (1UL << compound_order(page)) - 1;
844 lruvec = mem_cgroup_page_lruvec(page, zone);
846 /* Try isolate the page */
847 if (__isolate_lru_page(page, isolate_mode) != 0)
850 VM_BUG_ON_PAGE(PageCompound(page), page);
852 /* Successfully isolated */
853 del_page_from_lru_list(page, lruvec, page_lru(page));
856 list_add(&page->lru, &cc->migratepages);
857 cc->nr_migratepages++;
861 * Record where we could have freed pages by migration and not
862 * yet flushed them to buddy allocator.
863 * - this is the lowest page that was isolated and likely be
864 * then freed by migration.
866 if (!cc->last_migrated_pfn)
867 cc->last_migrated_pfn = low_pfn;
869 /* Avoid isolating too much */
870 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
877 if (!skip_on_failure)
881 * We have isolated some pages, but then failed. Release them
882 * instead of migrating, as we cannot form the cc->order buddy
887 spin_unlock_irqrestore(&zone->lru_lock, flags);
890 acct_isolated(zone, cc);
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, 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)
976 acct_isolated(cc->zone, cc);
981 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
982 #ifdef CONFIG_COMPACTION
984 /* Returns true if the page is within a block suitable for migration to */
985 static bool suitable_migration_target(struct page *page)
987 /* If the page is a large free page, then disallow migration */
988 if (PageBuddy(page)) {
990 * We are checking page_order without zone->lock taken. But
991 * the only small danger is that we skip a potentially suitable
992 * pageblock, so it's not worth to check order for valid range.
994 if (page_order_unsafe(page) >= pageblock_order)
998 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
999 if (migrate_async_suitable(get_pageblock_migratetype(page)))
1002 /* Otherwise skip the block */
1007 * Test whether the free scanner has reached the same or lower pageblock than
1008 * the migration scanner, and compaction should thus terminate.
1010 static inline bool compact_scanners_met(struct compact_control *cc)
1012 return (cc->free_pfn >> pageblock_order)
1013 <= (cc->migrate_pfn >> pageblock_order);
1017 * Based on information in the current compact_control, find blocks
1018 * suitable for isolating free pages from and then isolate them.
1020 static void isolate_freepages(struct compact_control *cc)
1022 struct zone *zone = cc->zone;
1024 unsigned long block_start_pfn; /* start of current pageblock */
1025 unsigned long isolate_start_pfn; /* exact pfn we start at */
1026 unsigned long block_end_pfn; /* end of current pageblock */
1027 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
1028 struct list_head *freelist = &cc->freepages;
1031 * Initialise the free scanner. The starting point is where we last
1032 * successfully isolated from, zone-cached value, or the end of the
1033 * zone when isolating for the first time. For looping we also need
1034 * this pfn aligned down to the pageblock boundary, because we do
1035 * block_start_pfn -= pageblock_nr_pages in the for loop.
1036 * For ending point, take care when isolating in last pageblock of a
1037 * a zone which ends in the middle of a pageblock.
1038 * The low boundary is the end of the pageblock the migration scanner
1041 isolate_start_pfn = cc->free_pfn;
1042 block_start_pfn = pageblock_start_pfn(cc->free_pfn);
1043 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1044 zone_end_pfn(zone));
1045 low_pfn = pageblock_end_pfn(cc->migrate_pfn);
1048 * Isolate free pages until enough are available to migrate the
1049 * pages on cc->migratepages. We stop searching if the migrate
1050 * and free page scanners meet or enough free pages are isolated.
1052 for (; block_start_pfn >= low_pfn;
1053 block_end_pfn = block_start_pfn,
1054 block_start_pfn -= pageblock_nr_pages,
1055 isolate_start_pfn = block_start_pfn) {
1057 * This can iterate a massively long zone without finding any
1058 * suitable migration targets, so periodically check if we need
1059 * to schedule, or even abort async compaction.
1061 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1062 && compact_should_abort(cc))
1065 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1070 /* Check the block is suitable for migration */
1071 if (!suitable_migration_target(page))
1074 /* If isolation recently failed, do not retry */
1075 if (!isolation_suitable(cc, page))
1078 /* Found a block suitable for isolating free pages from. */
1079 isolate_freepages_block(cc, &isolate_start_pfn, block_end_pfn,
1083 * If we isolated enough freepages, or aborted due to lock
1084 * contention, terminate.
1086 if ((cc->nr_freepages >= cc->nr_migratepages)
1088 if (isolate_start_pfn >= block_end_pfn) {
1090 * Restart at previous pageblock if more
1091 * freepages can be isolated next time.
1094 block_start_pfn - pageblock_nr_pages;
1097 } else if (isolate_start_pfn < block_end_pfn) {
1099 * If isolation failed early, do not continue
1106 /* split_free_page does not map the pages */
1107 map_pages(freelist);
1110 * Record where the free scanner will restart next time. Either we
1111 * broke from the loop and set isolate_start_pfn based on the last
1112 * call to isolate_freepages_block(), or we met the migration scanner
1113 * and the loop terminated due to isolate_start_pfn < low_pfn
1115 cc->free_pfn = isolate_start_pfn;
1119 * This is a migrate-callback that "allocates" freepages by taking pages
1120 * from the isolated freelists in the block we are migrating to.
1122 static struct page *compaction_alloc(struct page *migratepage,
1126 struct compact_control *cc = (struct compact_control *)data;
1127 struct page *freepage;
1130 * Isolate free pages if necessary, and if we are not aborting due to
1133 if (list_empty(&cc->freepages)) {
1135 isolate_freepages(cc);
1137 if (list_empty(&cc->freepages))
1141 freepage = list_entry(cc->freepages.next, struct page, lru);
1142 list_del(&freepage->lru);
1149 * This is a migrate-callback that "frees" freepages back to the isolated
1150 * freelist. All pages on the freelist are from the same zone, so there is no
1151 * special handling needed for NUMA.
1153 static void compaction_free(struct page *page, unsigned long data)
1155 struct compact_control *cc = (struct compact_control *)data;
1157 list_add(&page->lru, &cc->freepages);
1161 /* possible outcome of isolate_migratepages */
1163 ISOLATE_ABORT, /* Abort compaction now */
1164 ISOLATE_NONE, /* No pages isolated, continue scanning */
1165 ISOLATE_SUCCESS, /* Pages isolated, migrate */
1166 } isolate_migrate_t;
1169 * Allow userspace to control policy on scanning the unevictable LRU for
1170 * compactable pages.
1172 int sysctl_compact_unevictable_allowed __read_mostly = 1;
1175 * Isolate all pages that can be migrated from the first suitable block,
1176 * starting at the block pointed to by the migrate scanner pfn within
1179 static isolate_migrate_t isolate_migratepages(struct zone *zone,
1180 struct compact_control *cc)
1182 unsigned long block_start_pfn;
1183 unsigned long block_end_pfn;
1184 unsigned long low_pfn;
1186 const isolate_mode_t isolate_mode =
1187 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1188 (cc->mode == MIGRATE_ASYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1191 * Start at where we last stopped, or beginning of the zone as
1192 * initialized by compact_zone()
1194 low_pfn = cc->migrate_pfn;
1195 block_start_pfn = pageblock_start_pfn(low_pfn);
1196 if (block_start_pfn < zone->zone_start_pfn)
1197 block_start_pfn = zone->zone_start_pfn;
1199 /* Only scan within a pageblock boundary */
1200 block_end_pfn = pageblock_end_pfn(low_pfn);
1203 * Iterate over whole pageblocks until we find the first suitable.
1204 * Do not cross the free scanner.
1206 for (; block_end_pfn <= cc->free_pfn;
1207 low_pfn = block_end_pfn,
1208 block_start_pfn = block_end_pfn,
1209 block_end_pfn += pageblock_nr_pages) {
1212 * This can potentially iterate a massively long zone with
1213 * many pageblocks unsuitable, so periodically check if we
1214 * need to schedule, or even abort async compaction.
1216 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1217 && compact_should_abort(cc))
1220 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1225 /* If isolation recently failed, do not retry */
1226 if (!isolation_suitable(cc, page))
1230 * For async compaction, also only scan in MOVABLE blocks.
1231 * Async compaction is optimistic to see if the minimum amount
1232 * of work satisfies the allocation.
1234 if (cc->mode == MIGRATE_ASYNC &&
1235 !migrate_async_suitable(get_pageblock_migratetype(page)))
1238 /* Perform the isolation */
1239 low_pfn = isolate_migratepages_block(cc, low_pfn,
1240 block_end_pfn, isolate_mode);
1242 if (!low_pfn || cc->contended) {
1243 acct_isolated(zone, cc);
1244 return ISOLATE_ABORT;
1248 * Either we isolated something and proceed with migration. Or
1249 * we failed and compact_zone should decide if we should
1255 acct_isolated(zone, cc);
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 = low_wmark_pages(zone);
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_PARTIAL;
1320 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
1321 if (migratetype == MIGRATE_MOVABLE &&
1322 !list_empty(&area->free_list[MIGRATE_CMA]))
1323 return COMPACT_PARTIAL;
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_PARTIAL;
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_PARTIAL - 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)
1364 unsigned long watermark;
1366 if (is_via_compact_memory(order))
1367 return COMPACT_CONTINUE;
1369 watermark = low_wmark_pages(zone);
1371 * If watermarks for high-order allocation are already met, there
1372 * should be no need for compaction at all.
1374 if (zone_watermark_ok(zone, order, watermark, classzone_idx,
1376 return COMPACT_PARTIAL;
1379 * Watermarks for order-0 must be met for compaction. Note the 2UL.
1380 * This is because during migration, copies of pages need to be
1381 * allocated and for a short time, the footprint is higher
1383 watermark += (2UL << order);
1384 if (!__zone_watermark_ok(zone, 0, watermark, classzone_idx,
1385 alloc_flags, wmark_target))
1386 return COMPACT_SKIPPED;
1389 * fragmentation index determines if allocation failures are due to
1390 * low memory or external fragmentation
1392 * index of -1000 would imply allocations might succeed depending on
1393 * watermarks, but we already failed the high-order watermark check
1394 * index towards 0 implies failure is due to lack of memory
1395 * index towards 1000 implies failure is due to fragmentation
1397 * Only compact if a failure would be due to fragmentation.
1399 fragindex = fragmentation_index(zone, order);
1400 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
1401 return COMPACT_NOT_SUITABLE_ZONE;
1403 return COMPACT_CONTINUE;
1406 enum compact_result compaction_suitable(struct zone *zone, int order,
1407 unsigned int alloc_flags,
1410 enum compact_result ret;
1412 ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx,
1413 zone_page_state(zone, NR_FREE_PAGES));
1414 trace_mm_compaction_suitable(zone, order, ret);
1415 if (ret == COMPACT_NOT_SUITABLE_ZONE)
1416 ret = COMPACT_SKIPPED;
1421 bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
1428 * Make sure at least one zone would pass __compaction_suitable if we continue
1429 * retrying the reclaim.
1431 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1433 unsigned long available;
1434 enum compact_result compact_result;
1437 * Do not consider all the reclaimable memory because we do not
1438 * want to trash just for a single high order allocation which
1439 * is even not guaranteed to appear even if __compaction_suitable
1440 * is happy about the watermark check.
1442 available = zone_reclaimable_pages(zone) / order;
1443 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
1444 compact_result = __compaction_suitable(zone, order, alloc_flags,
1445 ac_classzone_idx(ac), available);
1446 if (compact_result != COMPACT_SKIPPED &&
1447 compact_result != COMPACT_NOT_SUITABLE_ZONE)
1454 static enum compact_result compact_zone(struct zone *zone, struct compact_control *cc)
1456 enum compact_result ret;
1457 unsigned long start_pfn = zone->zone_start_pfn;
1458 unsigned long end_pfn = zone_end_pfn(zone);
1459 const int migratetype = gfpflags_to_migratetype(cc->gfp_mask);
1460 const bool sync = cc->mode != MIGRATE_ASYNC;
1462 ret = compaction_suitable(zone, cc->order, cc->alloc_flags,
1464 /* Compaction is likely to fail */
1465 if (ret == COMPACT_PARTIAL || ret == COMPACT_SKIPPED)
1468 /* huh, compaction_suitable is returning something unexpected */
1469 VM_BUG_ON(ret != COMPACT_CONTINUE);
1472 * Clear pageblock skip if there were failures recently and compaction
1473 * is about to be retried after being deferred.
1475 if (compaction_restarting(zone, cc->order))
1476 __reset_isolation_suitable(zone);
1479 * Setup to move all movable pages to the end of the zone. Used cached
1480 * information on where the scanners should start but check that it
1481 * is initialised by ensuring the values are within zone boundaries.
1483 cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync];
1484 cc->free_pfn = zone->compact_cached_free_pfn;
1485 if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
1486 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
1487 zone->compact_cached_free_pfn = cc->free_pfn;
1489 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
1490 cc->migrate_pfn = start_pfn;
1491 zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
1492 zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
1495 if (cc->migrate_pfn == start_pfn)
1496 cc->whole_zone = true;
1498 cc->last_migrated_pfn = 0;
1500 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
1501 cc->free_pfn, end_pfn, sync);
1503 migrate_prep_local();
1505 while ((ret = compact_finished(zone, cc, migratetype)) ==
1509 switch (isolate_migratepages(zone, cc)) {
1511 ret = COMPACT_CONTENDED;
1512 putback_movable_pages(&cc->migratepages);
1513 cc->nr_migratepages = 0;
1517 * We haven't isolated and migrated anything, but
1518 * there might still be unflushed migrations from
1519 * previous cc->order aligned block.
1522 case ISOLATE_SUCCESS:
1526 err = migrate_pages(&cc->migratepages, compaction_alloc,
1527 compaction_free, (unsigned long)cc, cc->mode,
1530 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
1533 /* All pages were either migrated or will be released */
1534 cc->nr_migratepages = 0;
1536 putback_movable_pages(&cc->migratepages);
1538 * migrate_pages() may return -ENOMEM when scanners meet
1539 * and we want compact_finished() to detect it
1541 if (err == -ENOMEM && !compact_scanners_met(cc)) {
1542 ret = COMPACT_CONTENDED;
1546 * We failed to migrate at least one page in the current
1547 * order-aligned block, so skip the rest of it.
1549 if (cc->direct_compaction &&
1550 (cc->mode == MIGRATE_ASYNC)) {
1551 cc->migrate_pfn = block_end_pfn(
1552 cc->migrate_pfn - 1, cc->order);
1553 /* Draining pcplists is useless in this case */
1554 cc->last_migrated_pfn = 0;
1561 * Has the migration scanner moved away from the previous
1562 * cc->order aligned block where we migrated from? If yes,
1563 * flush the pages that were freed, so that they can merge and
1564 * compact_finished() can detect immediately if allocation
1567 if (cc->order > 0 && cc->last_migrated_pfn) {
1569 unsigned long current_block_start =
1570 block_start_pfn(cc->migrate_pfn, cc->order);
1572 if (cc->last_migrated_pfn < current_block_start) {
1574 lru_add_drain_cpu(cpu);
1575 drain_local_pages(zone);
1577 /* No more flushing until we migrate again */
1578 cc->last_migrated_pfn = 0;
1586 * Release free pages and update where the free scanner should restart,
1587 * so we don't leave any returned pages behind in the next attempt.
1589 if (cc->nr_freepages > 0) {
1590 unsigned long free_pfn = release_freepages(&cc->freepages);
1592 cc->nr_freepages = 0;
1593 VM_BUG_ON(free_pfn == 0);
1594 /* The cached pfn is always the first in a pageblock */
1595 free_pfn = pageblock_start_pfn(free_pfn);
1597 * Only go back, not forward. The cached pfn might have been
1598 * already reset to zone end in compact_finished()
1600 if (free_pfn > zone->compact_cached_free_pfn)
1601 zone->compact_cached_free_pfn = free_pfn;
1604 trace_mm_compaction_end(start_pfn, cc->migrate_pfn,
1605 cc->free_pfn, end_pfn, sync, ret);
1607 if (ret == COMPACT_CONTENDED)
1608 ret = COMPACT_PARTIAL;
1613 static enum compact_result compact_zone_order(struct zone *zone, int order,
1614 gfp_t gfp_mask, enum migrate_mode mode, int *contended,
1615 unsigned int alloc_flags, int classzone_idx)
1617 enum compact_result ret;
1618 struct compact_control cc = {
1620 .nr_migratepages = 0,
1622 .gfp_mask = gfp_mask,
1625 .alloc_flags = alloc_flags,
1626 .classzone_idx = classzone_idx,
1627 .direct_compaction = true,
1629 INIT_LIST_HEAD(&cc.freepages);
1630 INIT_LIST_HEAD(&cc.migratepages);
1632 ret = compact_zone(zone, &cc);
1634 VM_BUG_ON(!list_empty(&cc.freepages));
1635 VM_BUG_ON(!list_empty(&cc.migratepages));
1637 *contended = cc.contended;
1641 int sysctl_extfrag_threshold = 500;
1644 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1645 * @gfp_mask: The GFP mask of the current allocation
1646 * @order: The order of the current allocation
1647 * @alloc_flags: The allocation flags of the current allocation
1648 * @ac: The context of current allocation
1649 * @mode: The migration mode for async, sync light, or sync migration
1650 * @contended: Return value that determines if compaction was aborted due to
1651 * need_resched() or lock contention
1653 * This is the main entry point for direct page compaction.
1655 enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
1656 unsigned int alloc_flags, const struct alloc_context *ac,
1657 enum migrate_mode mode, int *contended)
1659 int may_enter_fs = gfp_mask & __GFP_FS;
1660 int may_perform_io = gfp_mask & __GFP_IO;
1663 enum compact_result rc = COMPACT_SKIPPED;
1664 int all_zones_contended = COMPACT_CONTENDED_LOCK; /* init for &= op */
1666 *contended = COMPACT_CONTENDED_NONE;
1668 /* Check if the GFP flags allow compaction */
1669 if (!order || !may_enter_fs || !may_perform_io)
1670 return COMPACT_SKIPPED;
1672 trace_mm_compaction_try_to_compact_pages(order, gfp_mask, mode);
1674 /* Compact each zone in the list */
1675 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1677 enum compact_result status;
1680 if (compaction_deferred(zone, order)) {
1681 rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
1685 status = compact_zone_order(zone, order, gfp_mask, mode,
1686 &zone_contended, alloc_flags,
1687 ac_classzone_idx(ac));
1688 rc = max(status, rc);
1690 * It takes at least one zone that wasn't lock contended
1691 * to clear all_zones_contended.
1693 all_zones_contended &= zone_contended;
1695 /* If a normal allocation would succeed, stop compacting */
1696 if (zone_watermark_ok(zone, order, low_wmark_pages(zone),
1697 ac_classzone_idx(ac), alloc_flags)) {
1699 * We think the allocation will succeed in this zone,
1700 * but it is not certain, hence the false. The caller
1701 * will repeat this with true if allocation indeed
1702 * succeeds in this zone.
1704 compaction_defer_reset(zone, order, false);
1706 * It is possible that async compaction aborted due to
1707 * need_resched() and the watermarks were ok thanks to
1708 * somebody else freeing memory. The allocation can
1709 * however still fail so we better signal the
1710 * need_resched() contention anyway (this will not
1711 * prevent the allocation attempt).
1713 if (zone_contended == COMPACT_CONTENDED_SCHED)
1714 *contended = COMPACT_CONTENDED_SCHED;
1719 if (mode != MIGRATE_ASYNC && (status == COMPACT_COMPLETE ||
1720 status == COMPACT_PARTIAL_SKIPPED)) {
1722 * We think that allocation won't succeed in this zone
1723 * so we defer compaction there. If it ends up
1724 * succeeding after all, it will be reset.
1726 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 and signal need_resched()
1735 if ((zone_contended == COMPACT_CONTENDED_SCHED)
1736 || fatal_signal_pending(current)) {
1737 *contended = COMPACT_CONTENDED_SCHED;
1744 * We might not have tried all the zones, so be conservative
1745 * and assume they are not all lock contended.
1747 all_zones_contended = 0;
1752 * If at least one zone wasn't deferred or skipped, we report if all
1753 * zones that were tried were lock contended.
1755 if (rc > COMPACT_INACTIVE && all_zones_contended)
1756 *contended = COMPACT_CONTENDED_LOCK;
1762 /* Compact all zones within a node */
1763 static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
1768 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1770 zone = &pgdat->node_zones[zoneid];
1771 if (!populated_zone(zone))
1774 cc->nr_freepages = 0;
1775 cc->nr_migratepages = 0;
1777 INIT_LIST_HEAD(&cc->freepages);
1778 INIT_LIST_HEAD(&cc->migratepages);
1781 * When called via /proc/sys/vm/compact_memory
1782 * this makes sure we compact the whole zone regardless of
1783 * cached scanner positions.
1785 if (is_via_compact_memory(cc->order))
1786 __reset_isolation_suitable(zone);
1788 if (is_via_compact_memory(cc->order) ||
1789 !compaction_deferred(zone, cc->order))
1790 compact_zone(zone, cc);
1792 VM_BUG_ON(!list_empty(&cc->freepages));
1793 VM_BUG_ON(!list_empty(&cc->migratepages));
1795 if (is_via_compact_memory(cc->order))
1798 if (zone_watermark_ok(zone, cc->order,
1799 low_wmark_pages(zone), 0, 0))
1800 compaction_defer_reset(zone, cc->order, false);
1804 void compact_pgdat(pg_data_t *pgdat, int order)
1806 struct compact_control cc = {
1808 .mode = MIGRATE_ASYNC,
1814 __compact_pgdat(pgdat, &cc);
1817 static void compact_node(int nid)
1819 struct compact_control cc = {
1821 .mode = MIGRATE_SYNC,
1822 .ignore_skip_hint = true,
1825 __compact_pgdat(NODE_DATA(nid), &cc);
1828 /* Compact all nodes in the system */
1829 static void compact_nodes(void)
1833 /* Flush pending updates to the LRU lists */
1834 lru_add_drain_all();
1836 for_each_online_node(nid)
1840 /* The written value is actually unused, all memory is compacted */
1841 int sysctl_compact_memory;
1844 * This is the entry point for compacting all nodes via
1845 * /proc/sys/vm/compact_memory
1847 int sysctl_compaction_handler(struct ctl_table *table, int write,
1848 void __user *buffer, size_t *length, loff_t *ppos)
1856 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1857 void __user *buffer, size_t *length, loff_t *ppos)
1859 proc_dointvec_minmax(table, write, buffer, length, ppos);
1864 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1865 static ssize_t sysfs_compact_node(struct device *dev,
1866 struct device_attribute *attr,
1867 const char *buf, size_t count)
1871 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1872 /* Flush pending updates to the LRU lists */
1873 lru_add_drain_all();
1880 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1882 int compaction_register_node(struct node *node)
1884 return device_create_file(&node->dev, &dev_attr_compact);
1887 void compaction_unregister_node(struct node *node)
1889 return device_remove_file(&node->dev, &dev_attr_compact);
1891 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1893 static inline bool kcompactd_work_requested(pg_data_t *pgdat)
1895 return pgdat->kcompactd_max_order > 0 || kthread_should_stop();
1898 static bool kcompactd_node_suitable(pg_data_t *pgdat)
1902 enum zone_type classzone_idx = pgdat->kcompactd_classzone_idx;
1904 for (zoneid = 0; zoneid <= classzone_idx; zoneid++) {
1905 zone = &pgdat->node_zones[zoneid];
1907 if (!populated_zone(zone))
1910 if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0,
1911 classzone_idx) == COMPACT_CONTINUE)
1918 static void kcompactd_do_work(pg_data_t *pgdat)
1921 * With no special task, compact all zones so that a page of requested
1922 * order is allocatable.
1926 struct compact_control cc = {
1927 .order = pgdat->kcompactd_max_order,
1928 .classzone_idx = pgdat->kcompactd_classzone_idx,
1929 .mode = MIGRATE_SYNC_LIGHT,
1930 .ignore_skip_hint = true,
1933 bool success = false;
1935 trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
1937 count_vm_event(KCOMPACTD_WAKE);
1939 for (zoneid = 0; zoneid <= cc.classzone_idx; zoneid++) {
1942 zone = &pgdat->node_zones[zoneid];
1943 if (!populated_zone(zone))
1946 if (compaction_deferred(zone, cc.order))
1949 if (compaction_suitable(zone, cc.order, 0, zoneid) !=
1953 cc.nr_freepages = 0;
1954 cc.nr_migratepages = 0;
1956 INIT_LIST_HEAD(&cc.freepages);
1957 INIT_LIST_HEAD(&cc.migratepages);
1959 if (kthread_should_stop())
1961 status = compact_zone(zone, &cc);
1963 if (zone_watermark_ok(zone, cc.order, low_wmark_pages(zone),
1964 cc.classzone_idx, 0)) {
1966 compaction_defer_reset(zone, cc.order, false);
1967 } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
1969 * We use sync migration mode here, so we defer like
1970 * sync direct compaction does.
1972 defer_compaction(zone, cc.order);
1975 VM_BUG_ON(!list_empty(&cc.freepages));
1976 VM_BUG_ON(!list_empty(&cc.migratepages));
1980 * Regardless of success, we are done until woken up next. But remember
1981 * the requested order/classzone_idx in case it was higher/tighter than
1984 if (pgdat->kcompactd_max_order <= cc.order)
1985 pgdat->kcompactd_max_order = 0;
1986 if (pgdat->kcompactd_classzone_idx >= cc.classzone_idx)
1987 pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
1990 void wakeup_kcompactd(pg_data_t *pgdat, int order, int classzone_idx)
1995 if (pgdat->kcompactd_max_order < order)
1996 pgdat->kcompactd_max_order = order;
1998 if (pgdat->kcompactd_classzone_idx > classzone_idx)
1999 pgdat->kcompactd_classzone_idx = classzone_idx;
2001 if (!waitqueue_active(&pgdat->kcompactd_wait))
2004 if (!kcompactd_node_suitable(pgdat))
2007 trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
2009 wake_up_interruptible(&pgdat->kcompactd_wait);
2013 * The background compaction daemon, started as a kernel thread
2014 * from the init process.
2016 static int kcompactd(void *p)
2018 pg_data_t *pgdat = (pg_data_t*)p;
2019 struct task_struct *tsk = current;
2021 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2023 if (!cpumask_empty(cpumask))
2024 set_cpus_allowed_ptr(tsk, cpumask);
2028 pgdat->kcompactd_max_order = 0;
2029 pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
2031 while (!kthread_should_stop()) {
2032 trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
2033 wait_event_freezable(pgdat->kcompactd_wait,
2034 kcompactd_work_requested(pgdat));
2036 kcompactd_do_work(pgdat);
2043 * This kcompactd start function will be called by init and node-hot-add.
2044 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
2046 int kcompactd_run(int nid)
2048 pg_data_t *pgdat = NODE_DATA(nid);
2051 if (pgdat->kcompactd)
2054 pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
2055 if (IS_ERR(pgdat->kcompactd)) {
2056 pr_err("Failed to start kcompactd on node %d\n", nid);
2057 ret = PTR_ERR(pgdat->kcompactd);
2058 pgdat->kcompactd = NULL;
2064 * Called by memory hotplug when all memory in a node is offlined. Caller must
2065 * hold mem_hotplug_begin/end().
2067 void kcompactd_stop(int nid)
2069 struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
2072 kthread_stop(kcompactd);
2073 NODE_DATA(nid)->kcompactd = NULL;
2078 * It's optimal to keep kcompactd on the same CPUs as their memory, but
2079 * not required for correctness. So if the last cpu in a node goes
2080 * away, we get changed to run anywhere: as the first one comes back,
2081 * restore their cpu bindings.
2083 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
2088 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2089 for_each_node_state(nid, N_MEMORY) {
2090 pg_data_t *pgdat = NODE_DATA(nid);
2091 const struct cpumask *mask;
2093 mask = cpumask_of_node(pgdat->node_id);
2095 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2096 /* One of our CPUs online: restore mask */
2097 set_cpus_allowed_ptr(pgdat->kcompactd, mask);
2103 static int __init kcompactd_init(void)
2107 for_each_node_state(nid, N_MEMORY)
2109 hotcpu_notifier(cpu_callback, 0);
2112 subsys_initcall(kcompactd_init)
2114 #endif /* CONFIG_COMPACTION */