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 /* Do not skip compaction more than 64 times */
85 #define COMPACT_MAX_DEFER_SHIFT 6
88 * Compaction is deferred when compaction fails to result in a page
89 * allocation success. 1 << compact_defer_limit compactions are skipped up
90 * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
92 void defer_compaction(struct zone *zone, int order)
94 zone->compact_considered = 0;
95 zone->compact_defer_shift++;
97 if (order < zone->compact_order_failed)
98 zone->compact_order_failed = order;
100 if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
101 zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
103 trace_mm_compaction_defer_compaction(zone, order);
106 /* Returns true if compaction should be skipped this time */
107 bool compaction_deferred(struct zone *zone, int order)
109 unsigned long defer_limit = 1UL << zone->compact_defer_shift;
111 if (order < zone->compact_order_failed)
114 /* Avoid possible overflow */
115 if (++zone->compact_considered > defer_limit)
116 zone->compact_considered = defer_limit;
118 if (zone->compact_considered >= defer_limit)
121 trace_mm_compaction_deferred(zone, order);
127 * Update defer tracking counters after successful compaction of given order,
128 * which means an allocation either succeeded (alloc_success == true) or is
129 * expected to succeed.
131 void compaction_defer_reset(struct zone *zone, int order,
135 zone->compact_considered = 0;
136 zone->compact_defer_shift = 0;
138 if (order >= zone->compact_order_failed)
139 zone->compact_order_failed = order + 1;
141 trace_mm_compaction_defer_reset(zone, order);
144 /* Returns true if restarting compaction after many failures */
145 bool compaction_restarting(struct zone *zone, int order)
147 if (order < zone->compact_order_failed)
150 return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
151 zone->compact_considered >= 1UL << zone->compact_defer_shift;
154 /* Returns true if the pageblock should be scanned for pages to isolate. */
155 static inline bool isolation_suitable(struct compact_control *cc,
158 if (cc->ignore_skip_hint)
161 return !get_pageblock_skip(page);
164 static void reset_cached_positions(struct zone *zone)
166 zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
167 zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
168 zone->compact_cached_free_pfn =
169 pageblock_start_pfn(zone_end_pfn(zone) - 1);
173 * This function is called to clear all cached information on pageblocks that
174 * should be skipped for page isolation when the migrate and free page scanner
177 static void __reset_isolation_suitable(struct zone *zone)
179 unsigned long start_pfn = zone->zone_start_pfn;
180 unsigned long end_pfn = zone_end_pfn(zone);
183 zone->compact_blockskip_flush = false;
185 /* Walk the zone and mark every pageblock as suitable for isolation */
186 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
194 page = pfn_to_page(pfn);
195 if (zone != page_zone(page))
198 clear_pageblock_skip(page);
201 reset_cached_positions(zone);
204 void reset_isolation_suitable(pg_data_t *pgdat)
208 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
209 struct zone *zone = &pgdat->node_zones[zoneid];
210 if (!populated_zone(zone))
213 /* Only flush if a full compaction finished recently */
214 if (zone->compact_blockskip_flush)
215 __reset_isolation_suitable(zone);
220 * If no pages were isolated then mark this pageblock to be skipped in the
221 * future. The information is later cleared by __reset_isolation_suitable().
223 static void update_pageblock_skip(struct compact_control *cc,
224 struct page *page, unsigned long nr_isolated,
225 bool migrate_scanner)
227 struct zone *zone = cc->zone;
230 if (cc->ignore_skip_hint)
239 set_pageblock_skip(page);
241 pfn = page_to_pfn(page);
243 /* Update where async and sync compaction should restart */
244 if (migrate_scanner) {
245 if (pfn > zone->compact_cached_migrate_pfn[0])
246 zone->compact_cached_migrate_pfn[0] = pfn;
247 if (cc->mode != MIGRATE_ASYNC &&
248 pfn > zone->compact_cached_migrate_pfn[1])
249 zone->compact_cached_migrate_pfn[1] = pfn;
251 if (pfn < zone->compact_cached_free_pfn)
252 zone->compact_cached_free_pfn = pfn;
256 static inline bool isolation_suitable(struct compact_control *cc,
262 static void update_pageblock_skip(struct compact_control *cc,
263 struct page *page, unsigned long nr_isolated,
264 bool migrate_scanner)
267 #endif /* CONFIG_COMPACTION */
270 * Compaction requires the taking of some coarse locks that are potentially
271 * very heavily contended. For async compaction, back out if the lock cannot
272 * be taken immediately. For sync compaction, spin on the lock if needed.
274 * Returns true if the lock is held
275 * Returns false if the lock is not held and compaction should abort
277 static bool compact_trylock_irqsave(spinlock_t *lock, unsigned long *flags,
278 struct compact_control *cc)
280 if (cc->mode == MIGRATE_ASYNC) {
281 if (!spin_trylock_irqsave(lock, *flags)) {
282 cc->contended = COMPACT_CONTENDED_LOCK;
286 spin_lock_irqsave(lock, *flags);
293 * Compaction requires the taking of some coarse locks that are potentially
294 * very heavily contended. The lock should be periodically unlocked to avoid
295 * having disabled IRQs for a long time, even when there is nobody waiting on
296 * the lock. It might also be that allowing the IRQs will result in
297 * need_resched() becoming true. If scheduling is needed, async compaction
298 * aborts. Sync compaction schedules.
299 * Either compaction type will also abort if a fatal signal is pending.
300 * In either case if the lock was locked, it is dropped and not regained.
302 * Returns true if compaction should abort due to fatal signal pending, or
303 * async compaction due to need_resched()
304 * Returns false when compaction can continue (sync compaction might have
307 static bool compact_unlock_should_abort(spinlock_t *lock,
308 unsigned long flags, bool *locked, struct compact_control *cc)
311 spin_unlock_irqrestore(lock, flags);
315 if (fatal_signal_pending(current)) {
316 cc->contended = COMPACT_CONTENDED_SCHED;
320 if (need_resched()) {
321 if (cc->mode == MIGRATE_ASYNC) {
322 cc->contended = COMPACT_CONTENDED_SCHED;
332 * Aside from avoiding lock contention, compaction also periodically checks
333 * need_resched() and either schedules in sync compaction or aborts async
334 * compaction. This is similar to what compact_unlock_should_abort() does, but
335 * is used where no lock is concerned.
337 * Returns false when no scheduling was needed, or sync compaction scheduled.
338 * Returns true when async compaction should abort.
340 static inline bool compact_should_abort(struct compact_control *cc)
342 /* async compaction aborts if contended */
343 if (need_resched()) {
344 if (cc->mode == MIGRATE_ASYNC) {
345 cc->contended = COMPACT_CONTENDED_SCHED;
356 * Isolate free pages onto a private freelist. If @strict is true, will abort
357 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
358 * (even though it may still end up isolating some pages).
360 static unsigned long isolate_freepages_block(struct compact_control *cc,
361 unsigned long *start_pfn,
362 unsigned long end_pfn,
363 struct list_head *freelist,
366 int nr_scanned = 0, total_isolated = 0;
367 struct page *cursor, *valid_page = NULL;
368 unsigned long flags = 0;
370 unsigned long blockpfn = *start_pfn;
372 cursor = pfn_to_page(blockpfn);
374 /* Isolate free pages. */
375 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
377 struct page *page = cursor;
380 * Periodically drop the lock (if held) regardless of its
381 * contention, to give chance to IRQs. Abort if fatal signal
382 * pending or async compaction detects need_resched()
384 if (!(blockpfn % SWAP_CLUSTER_MAX)
385 && compact_unlock_should_abort(&cc->zone->lock, flags,
390 if (!pfn_valid_within(blockpfn))
397 * For compound pages such as THP and hugetlbfs, we can save
398 * potentially a lot of iterations if we skip them at once.
399 * The check is racy, but we can consider only valid values
400 * and the only danger is skipping too much.
402 if (PageCompound(page)) {
403 unsigned int comp_order = compound_order(page);
405 if (likely(comp_order < MAX_ORDER)) {
406 blockpfn += (1UL << comp_order) - 1;
407 cursor += (1UL << comp_order) - 1;
413 if (!PageBuddy(page))
417 * If we already hold the lock, we can skip some rechecking.
418 * Note that if we hold the lock now, checked_pageblock was
419 * already set in some previous iteration (or strict is true),
420 * so it is correct to skip the suitable migration target
425 * The zone lock must be held to isolate freepages.
426 * Unfortunately this is a very coarse lock and can be
427 * heavily contended if there are parallel allocations
428 * or parallel compactions. For async compaction do not
429 * spin on the lock and we acquire the lock as late as
432 locked = compact_trylock_irqsave(&cc->zone->lock,
437 /* Recheck this is a buddy page under lock */
438 if (!PageBuddy(page))
442 /* Found a free page, break it into order-0 pages */
443 isolated = split_free_page(page);
444 total_isolated += isolated;
445 for (i = 0; i < isolated; i++) {
446 list_add(&page->lru, freelist);
450 /* If a page was split, advance to the end of it */
452 cc->nr_freepages += isolated;
454 cc->nr_migratepages <= cc->nr_freepages) {
455 blockpfn += isolated;
459 blockpfn += isolated - 1;
460 cursor += isolated - 1;
473 * There is a tiny chance that we have read bogus compound_order(),
474 * so be careful to not go outside of the pageblock.
476 if (unlikely(blockpfn > end_pfn))
479 trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
480 nr_scanned, total_isolated);
482 /* Record how far we have got within the block */
483 *start_pfn = blockpfn;
486 * If strict isolation is requested by CMA then check that all the
487 * pages requested were isolated. If there were any failures, 0 is
488 * returned and CMA will fail.
490 if (strict && blockpfn < end_pfn)
494 spin_unlock_irqrestore(&cc->zone->lock, flags);
496 /* Update the pageblock-skip if the whole pageblock was scanned */
497 if (blockpfn == end_pfn)
498 update_pageblock_skip(cc, valid_page, total_isolated, false);
500 count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
502 count_compact_events(COMPACTISOLATED, total_isolated);
503 return total_isolated;
507 * isolate_freepages_range() - isolate free pages.
508 * @start_pfn: The first PFN to start isolating.
509 * @end_pfn: The one-past-last PFN.
511 * Non-free pages, invalid PFNs, or zone boundaries within the
512 * [start_pfn, end_pfn) range are considered errors, cause function to
513 * undo its actions and return zero.
515 * Otherwise, function returns one-past-the-last PFN of isolated page
516 * (which may be greater then end_pfn if end fell in a middle of
520 isolate_freepages_range(struct compact_control *cc,
521 unsigned long start_pfn, unsigned long end_pfn)
523 unsigned long isolated, pfn, block_start_pfn, block_end_pfn;
527 block_start_pfn = pageblock_start_pfn(pfn);
528 if (block_start_pfn < cc->zone->zone_start_pfn)
529 block_start_pfn = cc->zone->zone_start_pfn;
530 block_end_pfn = pageblock_end_pfn(pfn);
532 for (; pfn < end_pfn; pfn += isolated,
533 block_start_pfn = block_end_pfn,
534 block_end_pfn += pageblock_nr_pages) {
535 /* Protect pfn from changing by isolate_freepages_block */
536 unsigned long isolate_start_pfn = pfn;
538 block_end_pfn = min(block_end_pfn, end_pfn);
541 * pfn could pass the block_end_pfn if isolated freepage
542 * is more than pageblock order. In this case, we adjust
543 * scanning range to right one.
545 if (pfn >= block_end_pfn) {
546 block_start_pfn = pageblock_start_pfn(pfn);
547 block_end_pfn = pageblock_end_pfn(pfn);
548 block_end_pfn = min(block_end_pfn, end_pfn);
551 if (!pageblock_pfn_to_page(block_start_pfn,
552 block_end_pfn, cc->zone))
555 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
556 block_end_pfn, &freelist, true);
559 * In strict mode, isolate_freepages_block() returns 0 if
560 * there are any holes in the block (ie. invalid PFNs or
567 * If we managed to isolate pages, it is always (1 << n) *
568 * pageblock_nr_pages for some non-negative n. (Max order
569 * page may span two pageblocks).
573 /* split_free_page does not map the pages */
574 map_pages(&freelist);
577 /* Loop terminated early, cleanup. */
578 release_freepages(&freelist);
582 /* We don't use freelists for anything. */
586 /* Update the number of anon and file isolated pages in the zone */
587 static void acct_isolated(struct zone *zone, struct compact_control *cc)
590 unsigned int count[2] = { 0, };
592 if (list_empty(&cc->migratepages))
595 list_for_each_entry(page, &cc->migratepages, lru)
596 count[!!page_is_file_cache(page)]++;
598 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
599 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
602 /* Similar to reclaim, but different enough that they don't share logic */
603 static bool too_many_isolated(struct zone *zone)
605 unsigned long active, inactive, isolated;
607 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
608 zone_page_state(zone, NR_INACTIVE_ANON);
609 active = zone_page_state(zone, NR_ACTIVE_FILE) +
610 zone_page_state(zone, NR_ACTIVE_ANON);
611 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
612 zone_page_state(zone, NR_ISOLATED_ANON);
614 return isolated > (inactive + active) / 2;
618 * isolate_migratepages_block() - isolate all migrate-able pages within
620 * @cc: Compaction control structure.
621 * @low_pfn: The first PFN to isolate
622 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
623 * @isolate_mode: Isolation mode to be used.
625 * Isolate all pages that can be migrated from the range specified by
626 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
627 * Returns zero if there is a fatal signal pending, otherwise PFN of the
628 * first page that was not scanned (which may be both less, equal to or more
631 * The pages are isolated on cc->migratepages list (not required to be empty),
632 * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
633 * is neither read nor updated.
636 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
637 unsigned long end_pfn, isolate_mode_t isolate_mode)
639 struct zone *zone = cc->zone;
640 unsigned long nr_scanned = 0, nr_isolated = 0;
641 struct list_head *migratelist = &cc->migratepages;
642 struct lruvec *lruvec;
643 unsigned long flags = 0;
645 struct page *page = NULL, *valid_page = NULL;
646 unsigned long start_pfn = low_pfn;
649 * Ensure that there are not too many pages isolated from the LRU
650 * list by either parallel reclaimers or compaction. If there are,
651 * delay for some time until fewer pages are isolated
653 while (unlikely(too_many_isolated(zone))) {
654 /* async migration should just abort */
655 if (cc->mode == MIGRATE_ASYNC)
658 congestion_wait(BLK_RW_ASYNC, HZ/10);
660 if (fatal_signal_pending(current))
664 if (compact_should_abort(cc))
667 /* Time to isolate some pages for migration */
668 for (; low_pfn < end_pfn; low_pfn++) {
672 * Periodically drop the lock (if held) regardless of its
673 * contention, to give chance to IRQs. Abort async compaction
676 if (!(low_pfn % SWAP_CLUSTER_MAX)
677 && compact_unlock_should_abort(&zone->lru_lock, flags,
681 if (!pfn_valid_within(low_pfn))
685 page = pfn_to_page(low_pfn);
691 * Skip if free. We read page order here without zone lock
692 * which is generally unsafe, but the race window is small and
693 * the worst thing that can happen is that we skip some
694 * potential isolation targets.
696 if (PageBuddy(page)) {
697 unsigned long freepage_order = page_order_unsafe(page);
700 * Without lock, we cannot be sure that what we got is
701 * a valid page order. Consider only values in the
702 * valid order range to prevent low_pfn overflow.
704 if (freepage_order > 0 && freepage_order < MAX_ORDER)
705 low_pfn += (1UL << freepage_order) - 1;
710 * Check may be lockless but that's ok as we recheck later.
711 * It's possible to migrate LRU pages and balloon pages
712 * Skip any other type of page
714 is_lru = PageLRU(page);
716 if (unlikely(balloon_page_movable(page))) {
717 if (balloon_page_isolate(page)) {
718 /* Successfully isolated */
719 goto isolate_success;
725 * Regardless of being on LRU, compound pages such as THP and
726 * hugetlbfs are not to be compacted. We can potentially save
727 * a lot of iterations if we skip them at once. The check is
728 * racy, but we can consider only valid values and the only
729 * danger is skipping too much.
731 if (PageCompound(page)) {
732 unsigned int comp_order = compound_order(page);
734 if (likely(comp_order < MAX_ORDER))
735 low_pfn += (1UL << comp_order) - 1;
744 * Migration will fail if an anonymous page is pinned in memory,
745 * so avoid taking lru_lock and isolating it unnecessarily in an
746 * admittedly racy check.
748 if (!page_mapping(page) &&
749 page_count(page) > page_mapcount(page))
752 /* If we already hold the lock, we can skip some rechecking */
754 locked = compact_trylock_irqsave(&zone->lru_lock,
759 /* Recheck PageLRU and PageCompound under lock */
764 * Page become compound since the non-locked check,
765 * and it's on LRU. It can only be a THP so the order
766 * is safe to read and it's 0 for tail pages.
768 if (unlikely(PageCompound(page))) {
769 low_pfn += (1UL << compound_order(page)) - 1;
774 lruvec = mem_cgroup_page_lruvec(page, zone);
776 /* Try isolate the page */
777 if (__isolate_lru_page(page, isolate_mode) != 0)
780 VM_BUG_ON_PAGE(PageCompound(page), page);
782 /* Successfully isolated */
783 del_page_from_lru_list(page, lruvec, page_lru(page));
786 list_add(&page->lru, migratelist);
787 cc->nr_migratepages++;
791 * Record where we could have freed pages by migration and not
792 * yet flushed them to buddy allocator.
793 * - this is the lowest page that was isolated and likely be
794 * then freed by migration.
796 if (!cc->last_migrated_pfn)
797 cc->last_migrated_pfn = low_pfn;
799 /* Avoid isolating too much */
800 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
807 * The PageBuddy() check could have potentially brought us outside
808 * the range to be scanned.
810 if (unlikely(low_pfn > end_pfn))
814 spin_unlock_irqrestore(&zone->lru_lock, flags);
817 * Update the pageblock-skip information and cached scanner pfn,
818 * if the whole pageblock was scanned without isolating any page.
820 if (low_pfn == end_pfn)
821 update_pageblock_skip(cc, valid_page, nr_isolated, true);
823 trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
824 nr_scanned, nr_isolated);
826 count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
828 count_compact_events(COMPACTISOLATED, nr_isolated);
834 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
835 * @cc: Compaction control structure.
836 * @start_pfn: The first PFN to start isolating.
837 * @end_pfn: The one-past-last PFN.
839 * Returns zero if isolation fails fatally due to e.g. pending signal.
840 * Otherwise, function returns one-past-the-last PFN of isolated page
841 * (which may be greater than end_pfn if end fell in a middle of a THP page).
844 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
845 unsigned long end_pfn)
847 unsigned long pfn, block_start_pfn, block_end_pfn;
849 /* Scan block by block. First and last block may be incomplete */
851 block_start_pfn = pageblock_start_pfn(pfn);
852 if (block_start_pfn < cc->zone->zone_start_pfn)
853 block_start_pfn = cc->zone->zone_start_pfn;
854 block_end_pfn = pageblock_end_pfn(pfn);
856 for (; pfn < end_pfn; pfn = block_end_pfn,
857 block_start_pfn = block_end_pfn,
858 block_end_pfn += pageblock_nr_pages) {
860 block_end_pfn = min(block_end_pfn, end_pfn);
862 if (!pageblock_pfn_to_page(block_start_pfn,
863 block_end_pfn, cc->zone))
866 pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
867 ISOLATE_UNEVICTABLE);
872 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
875 acct_isolated(cc->zone, cc);
880 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
881 #ifdef CONFIG_COMPACTION
883 /* Returns true if the page is within a block suitable for migration to */
884 static bool suitable_migration_target(struct page *page)
886 /* If the page is a large free page, then disallow migration */
887 if (PageBuddy(page)) {
889 * We are checking page_order without zone->lock taken. But
890 * the only small danger is that we skip a potentially suitable
891 * pageblock, so it's not worth to check order for valid range.
893 if (page_order_unsafe(page) >= pageblock_order)
897 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
898 if (migrate_async_suitable(get_pageblock_migratetype(page)))
901 /* Otherwise skip the block */
906 * Test whether the free scanner has reached the same or lower pageblock than
907 * the migration scanner, and compaction should thus terminate.
909 static inline bool compact_scanners_met(struct compact_control *cc)
911 return (cc->free_pfn >> pageblock_order)
912 <= (cc->migrate_pfn >> pageblock_order);
916 * Based on information in the current compact_control, find blocks
917 * suitable for isolating free pages from and then isolate them.
919 static void isolate_freepages(struct compact_control *cc)
921 struct zone *zone = cc->zone;
923 unsigned long block_start_pfn; /* start of current pageblock */
924 unsigned long isolate_start_pfn; /* exact pfn we start at */
925 unsigned long block_end_pfn; /* end of current pageblock */
926 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
927 struct list_head *freelist = &cc->freepages;
930 * Initialise the free scanner. The starting point is where we last
931 * successfully isolated from, zone-cached value, or the end of the
932 * zone when isolating for the first time. For looping we also need
933 * this pfn aligned down to the pageblock boundary, because we do
934 * block_start_pfn -= pageblock_nr_pages in the for loop.
935 * For ending point, take care when isolating in last pageblock of a
936 * a zone which ends in the middle of a pageblock.
937 * The low boundary is the end of the pageblock the migration scanner
940 isolate_start_pfn = cc->free_pfn;
941 block_start_pfn = pageblock_start_pfn(cc->free_pfn);
942 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
944 low_pfn = pageblock_end_pfn(cc->migrate_pfn);
947 * Isolate free pages until enough are available to migrate the
948 * pages on cc->migratepages. We stop searching if the migrate
949 * and free page scanners meet or enough free pages are isolated.
951 for (; block_start_pfn >= low_pfn;
952 block_end_pfn = block_start_pfn,
953 block_start_pfn -= pageblock_nr_pages,
954 isolate_start_pfn = block_start_pfn) {
957 * This can iterate a massively long zone without finding any
958 * suitable migration targets, so periodically check if we need
959 * to schedule, or even abort async compaction.
961 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
962 && compact_should_abort(cc))
965 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
970 /* Check the block is suitable for migration */
971 if (!suitable_migration_target(page))
974 /* If isolation recently failed, do not retry */
975 if (!isolation_suitable(cc, page))
978 /* Found a block suitable for isolating free pages from. */
979 isolate_freepages_block(cc, &isolate_start_pfn,
980 block_end_pfn, freelist, false);
983 * If we isolated enough freepages, or aborted due to async
984 * compaction being contended, terminate the loop.
985 * Remember where the free scanner should restart next time,
986 * which is where isolate_freepages_block() left off.
987 * But if it scanned the whole pageblock, isolate_start_pfn
988 * now points at block_end_pfn, which is the start of the next
990 * In that case we will however want to restart at the start
991 * of the previous pageblock.
993 if ((cc->nr_freepages >= cc->nr_migratepages)
995 if (isolate_start_pfn >= block_end_pfn)
997 block_start_pfn - pageblock_nr_pages;
1001 * isolate_freepages_block() should not terminate
1002 * prematurely unless contended, or isolated enough
1004 VM_BUG_ON(isolate_start_pfn < block_end_pfn);
1008 /* split_free_page does not map the pages */
1009 map_pages(freelist);
1012 * Record where the free scanner will restart next time. Either we
1013 * broke from the loop and set isolate_start_pfn based on the last
1014 * call to isolate_freepages_block(), or we met the migration scanner
1015 * and the loop terminated due to isolate_start_pfn < low_pfn
1017 cc->free_pfn = isolate_start_pfn;
1021 * This is a migrate-callback that "allocates" freepages by taking pages
1022 * from the isolated freelists in the block we are migrating to.
1024 static struct page *compaction_alloc(struct page *migratepage,
1028 struct compact_control *cc = (struct compact_control *)data;
1029 struct page *freepage;
1032 * Isolate free pages if necessary, and if we are not aborting due to
1035 if (list_empty(&cc->freepages)) {
1037 isolate_freepages(cc);
1039 if (list_empty(&cc->freepages))
1043 freepage = list_entry(cc->freepages.next, struct page, lru);
1044 list_del(&freepage->lru);
1051 * This is a migrate-callback that "frees" freepages back to the isolated
1052 * freelist. All pages on the freelist are from the same zone, so there is no
1053 * special handling needed for NUMA.
1055 static void compaction_free(struct page *page, unsigned long data)
1057 struct compact_control *cc = (struct compact_control *)data;
1059 list_add(&page->lru, &cc->freepages);
1063 /* possible outcome of isolate_migratepages */
1065 ISOLATE_ABORT, /* Abort compaction now */
1066 ISOLATE_NONE, /* No pages isolated, continue scanning */
1067 ISOLATE_SUCCESS, /* Pages isolated, migrate */
1068 } isolate_migrate_t;
1071 * Allow userspace to control policy on scanning the unevictable LRU for
1072 * compactable pages.
1074 int sysctl_compact_unevictable_allowed __read_mostly = 1;
1077 * Isolate all pages that can be migrated from the first suitable block,
1078 * starting at the block pointed to by the migrate scanner pfn within
1081 static isolate_migrate_t isolate_migratepages(struct zone *zone,
1082 struct compact_control *cc)
1084 unsigned long block_start_pfn;
1085 unsigned long block_end_pfn;
1086 unsigned long low_pfn;
1088 const isolate_mode_t isolate_mode =
1089 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1090 (cc->mode == MIGRATE_ASYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1093 * Start at where we last stopped, or beginning of the zone as
1094 * initialized by compact_zone()
1096 low_pfn = cc->migrate_pfn;
1097 block_start_pfn = pageblock_start_pfn(low_pfn);
1098 if (block_start_pfn < zone->zone_start_pfn)
1099 block_start_pfn = zone->zone_start_pfn;
1101 /* Only scan within a pageblock boundary */
1102 block_end_pfn = pageblock_end_pfn(low_pfn);
1105 * Iterate over whole pageblocks until we find the first suitable.
1106 * Do not cross the free scanner.
1108 for (; block_end_pfn <= cc->free_pfn;
1109 low_pfn = block_end_pfn,
1110 block_start_pfn = block_end_pfn,
1111 block_end_pfn += pageblock_nr_pages) {
1114 * This can potentially iterate a massively long zone with
1115 * many pageblocks unsuitable, so periodically check if we
1116 * need to schedule, or even abort async compaction.
1118 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1119 && compact_should_abort(cc))
1122 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1127 /* If isolation recently failed, do not retry */
1128 if (!isolation_suitable(cc, page))
1132 * For async compaction, also only scan in MOVABLE blocks.
1133 * Async compaction is optimistic to see if the minimum amount
1134 * of work satisfies the allocation.
1136 if (cc->mode == MIGRATE_ASYNC &&
1137 !migrate_async_suitable(get_pageblock_migratetype(page)))
1140 /* Perform the isolation */
1141 low_pfn = isolate_migratepages_block(cc, low_pfn,
1142 block_end_pfn, isolate_mode);
1144 if (!low_pfn || cc->contended) {
1145 acct_isolated(zone, cc);
1146 return ISOLATE_ABORT;
1150 * Either we isolated something and proceed with migration. Or
1151 * we failed and compact_zone should decide if we should
1157 acct_isolated(zone, cc);
1158 /* Record where migration scanner will be restarted. */
1159 cc->migrate_pfn = low_pfn;
1161 return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
1165 * order == -1 is expected when compacting via
1166 * /proc/sys/vm/compact_memory
1168 static inline bool is_via_compact_memory(int order)
1173 static int __compact_finished(struct zone *zone, struct compact_control *cc,
1174 const int migratetype)
1177 unsigned long watermark;
1179 if (cc->contended || fatal_signal_pending(current))
1180 return COMPACT_CONTENDED;
1182 /* Compaction run completes if the migrate and free scanner meet */
1183 if (compact_scanners_met(cc)) {
1184 /* Let the next compaction start anew. */
1185 reset_cached_positions(zone);
1188 * Mark that the PG_migrate_skip information should be cleared
1189 * by kswapd when it goes to sleep. kcompactd does not set the
1190 * flag itself as the decision to be clear should be directly
1191 * based on an allocation request.
1193 if (cc->direct_compaction)
1194 zone->compact_blockskip_flush = true;
1196 return COMPACT_COMPLETE;
1199 if (is_via_compact_memory(cc->order))
1200 return COMPACT_CONTINUE;
1202 /* Compaction run is not finished if the watermark is not met */
1203 watermark = low_wmark_pages(zone);
1205 if (!zone_watermark_ok(zone, cc->order, watermark, cc->classzone_idx,
1207 return COMPACT_CONTINUE;
1209 /* Direct compactor: Is a suitable page free? */
1210 for (order = cc->order; order < MAX_ORDER; order++) {
1211 struct free_area *area = &zone->free_area[order];
1214 /* Job done if page is free of the right migratetype */
1215 if (!list_empty(&area->free_list[migratetype]))
1216 return COMPACT_PARTIAL;
1219 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
1220 if (migratetype == MIGRATE_MOVABLE &&
1221 !list_empty(&area->free_list[MIGRATE_CMA]))
1222 return COMPACT_PARTIAL;
1225 * Job done if allocation would steal freepages from
1226 * other migratetype buddy lists.
1228 if (find_suitable_fallback(area, order, migratetype,
1229 true, &can_steal) != -1)
1230 return COMPACT_PARTIAL;
1233 return COMPACT_NO_SUITABLE_PAGE;
1236 static int compact_finished(struct zone *zone, struct compact_control *cc,
1237 const int migratetype)
1241 ret = __compact_finished(zone, cc, migratetype);
1242 trace_mm_compaction_finished(zone, cc->order, ret);
1243 if (ret == COMPACT_NO_SUITABLE_PAGE)
1244 ret = COMPACT_CONTINUE;
1250 * compaction_suitable: Is this suitable to run compaction on this zone now?
1252 * COMPACT_SKIPPED - If there are too few free pages for compaction
1253 * COMPACT_PARTIAL - If the allocation would succeed without compaction
1254 * COMPACT_CONTINUE - If compaction should run now
1256 static unsigned long __compaction_suitable(struct zone *zone, int order,
1257 int alloc_flags, int classzone_idx)
1260 unsigned long watermark;
1262 if (is_via_compact_memory(order))
1263 return COMPACT_CONTINUE;
1265 watermark = low_wmark_pages(zone);
1267 * If watermarks for high-order allocation are already met, there
1268 * should be no need for compaction at all.
1270 if (zone_watermark_ok(zone, order, watermark, classzone_idx,
1272 return COMPACT_PARTIAL;
1275 * Watermarks for order-0 must be met for compaction. Note the 2UL.
1276 * This is because during migration, copies of pages need to be
1277 * allocated and for a short time, the footprint is higher
1279 watermark += (2UL << order);
1280 if (!zone_watermark_ok(zone, 0, watermark, classzone_idx, alloc_flags))
1281 return COMPACT_SKIPPED;
1284 * fragmentation index determines if allocation failures are due to
1285 * low memory or external fragmentation
1287 * index of -1000 would imply allocations might succeed depending on
1288 * watermarks, but we already failed the high-order watermark check
1289 * index towards 0 implies failure is due to lack of memory
1290 * index towards 1000 implies failure is due to fragmentation
1292 * Only compact if a failure would be due to fragmentation.
1294 fragindex = fragmentation_index(zone, order);
1295 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
1296 return COMPACT_NOT_SUITABLE_ZONE;
1298 return COMPACT_CONTINUE;
1301 unsigned long compaction_suitable(struct zone *zone, int order,
1302 int alloc_flags, int classzone_idx)
1306 ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx);
1307 trace_mm_compaction_suitable(zone, order, ret);
1308 if (ret == COMPACT_NOT_SUITABLE_ZONE)
1309 ret = COMPACT_SKIPPED;
1314 static int compact_zone(struct zone *zone, struct compact_control *cc)
1317 unsigned long start_pfn = zone->zone_start_pfn;
1318 unsigned long end_pfn = zone_end_pfn(zone);
1319 const int migratetype = gfpflags_to_migratetype(cc->gfp_mask);
1320 const bool sync = cc->mode != MIGRATE_ASYNC;
1322 ret = compaction_suitable(zone, cc->order, cc->alloc_flags,
1325 case COMPACT_PARTIAL:
1326 case COMPACT_SKIPPED:
1327 /* Compaction is likely to fail */
1329 case COMPACT_CONTINUE:
1330 /* Fall through to compaction */
1335 * Clear pageblock skip if there were failures recently and compaction
1336 * is about to be retried after being deferred.
1338 if (compaction_restarting(zone, cc->order))
1339 __reset_isolation_suitable(zone);
1342 * Setup to move all movable pages to the end of the zone. Used cached
1343 * information on where the scanners should start but check that it
1344 * is initialised by ensuring the values are within zone boundaries.
1346 cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync];
1347 cc->free_pfn = zone->compact_cached_free_pfn;
1348 if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
1349 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
1350 zone->compact_cached_free_pfn = cc->free_pfn;
1352 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
1353 cc->migrate_pfn = start_pfn;
1354 zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
1355 zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
1357 cc->last_migrated_pfn = 0;
1359 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
1360 cc->free_pfn, end_pfn, sync);
1362 migrate_prep_local();
1364 while ((ret = compact_finished(zone, cc, migratetype)) ==
1368 switch (isolate_migratepages(zone, cc)) {
1370 ret = COMPACT_CONTENDED;
1371 putback_movable_pages(&cc->migratepages);
1372 cc->nr_migratepages = 0;
1376 * We haven't isolated and migrated anything, but
1377 * there might still be unflushed migrations from
1378 * previous cc->order aligned block.
1381 case ISOLATE_SUCCESS:
1385 err = migrate_pages(&cc->migratepages, compaction_alloc,
1386 compaction_free, (unsigned long)cc, cc->mode,
1389 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
1392 /* All pages were either migrated or will be released */
1393 cc->nr_migratepages = 0;
1395 putback_movable_pages(&cc->migratepages);
1397 * migrate_pages() may return -ENOMEM when scanners meet
1398 * and we want compact_finished() to detect it
1400 if (err == -ENOMEM && !compact_scanners_met(cc)) {
1401 ret = COMPACT_CONTENDED;
1408 * Has the migration scanner moved away from the previous
1409 * cc->order aligned block where we migrated from? If yes,
1410 * flush the pages that were freed, so that they can merge and
1411 * compact_finished() can detect immediately if allocation
1414 if (cc->order > 0 && cc->last_migrated_pfn) {
1416 unsigned long current_block_start =
1417 block_start_pfn(cc->migrate_pfn, cc->order);
1419 if (cc->last_migrated_pfn < current_block_start) {
1421 lru_add_drain_cpu(cpu);
1422 drain_local_pages(zone);
1424 /* No more flushing until we migrate again */
1425 cc->last_migrated_pfn = 0;
1433 * Release free pages and update where the free scanner should restart,
1434 * so we don't leave any returned pages behind in the next attempt.
1436 if (cc->nr_freepages > 0) {
1437 unsigned long free_pfn = release_freepages(&cc->freepages);
1439 cc->nr_freepages = 0;
1440 VM_BUG_ON(free_pfn == 0);
1441 /* The cached pfn is always the first in a pageblock */
1442 free_pfn = pageblock_start_pfn(free_pfn);
1444 * Only go back, not forward. The cached pfn might have been
1445 * already reset to zone end in compact_finished()
1447 if (free_pfn > zone->compact_cached_free_pfn)
1448 zone->compact_cached_free_pfn = free_pfn;
1451 trace_mm_compaction_end(start_pfn, cc->migrate_pfn,
1452 cc->free_pfn, end_pfn, sync, ret);
1454 if (ret == COMPACT_CONTENDED)
1455 ret = COMPACT_PARTIAL;
1460 static unsigned long compact_zone_order(struct zone *zone, int order,
1461 gfp_t gfp_mask, enum migrate_mode mode, int *contended,
1462 int alloc_flags, int classzone_idx)
1465 struct compact_control cc = {
1467 .nr_migratepages = 0,
1469 .gfp_mask = gfp_mask,
1472 .alloc_flags = alloc_flags,
1473 .classzone_idx = classzone_idx,
1474 .direct_compaction = true,
1476 INIT_LIST_HEAD(&cc.freepages);
1477 INIT_LIST_HEAD(&cc.migratepages);
1479 ret = compact_zone(zone, &cc);
1481 VM_BUG_ON(!list_empty(&cc.freepages));
1482 VM_BUG_ON(!list_empty(&cc.migratepages));
1484 *contended = cc.contended;
1488 int sysctl_extfrag_threshold = 500;
1491 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1492 * @gfp_mask: The GFP mask of the current allocation
1493 * @order: The order of the current allocation
1494 * @alloc_flags: The allocation flags of the current allocation
1495 * @ac: The context of current allocation
1496 * @mode: The migration mode for async, sync light, or sync migration
1497 * @contended: Return value that determines if compaction was aborted due to
1498 * need_resched() or lock contention
1500 * This is the main entry point for direct page compaction.
1502 unsigned long try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
1503 int alloc_flags, const struct alloc_context *ac,
1504 enum migrate_mode mode, int *contended)
1506 int may_enter_fs = gfp_mask & __GFP_FS;
1507 int may_perform_io = gfp_mask & __GFP_IO;
1510 int rc = COMPACT_DEFERRED;
1511 int all_zones_contended = COMPACT_CONTENDED_LOCK; /* init for &= op */
1513 *contended = COMPACT_CONTENDED_NONE;
1515 /* Check if the GFP flags allow compaction */
1516 if (!order || !may_enter_fs || !may_perform_io)
1517 return COMPACT_SKIPPED;
1519 trace_mm_compaction_try_to_compact_pages(order, gfp_mask, mode);
1521 /* Compact each zone in the list */
1522 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1527 if (compaction_deferred(zone, order))
1530 status = compact_zone_order(zone, order, gfp_mask, mode,
1531 &zone_contended, alloc_flags,
1533 rc = max(status, rc);
1535 * It takes at least one zone that wasn't lock contended
1536 * to clear all_zones_contended.
1538 all_zones_contended &= zone_contended;
1540 /* If a normal allocation would succeed, stop compacting */
1541 if (zone_watermark_ok(zone, order, low_wmark_pages(zone),
1542 ac->classzone_idx, alloc_flags)) {
1544 * We think the allocation will succeed in this zone,
1545 * but it is not certain, hence the false. The caller
1546 * will repeat this with true if allocation indeed
1547 * succeeds in this zone.
1549 compaction_defer_reset(zone, order, false);
1551 * It is possible that async compaction aborted due to
1552 * need_resched() and the watermarks were ok thanks to
1553 * somebody else freeing memory. The allocation can
1554 * however still fail so we better signal the
1555 * need_resched() contention anyway (this will not
1556 * prevent the allocation attempt).
1558 if (zone_contended == COMPACT_CONTENDED_SCHED)
1559 *contended = COMPACT_CONTENDED_SCHED;
1564 if (mode != MIGRATE_ASYNC && status == COMPACT_COMPLETE) {
1566 * We think that allocation won't succeed in this zone
1567 * so we defer compaction there. If it ends up
1568 * succeeding after all, it will be reset.
1570 defer_compaction(zone, order);
1574 * We might have stopped compacting due to need_resched() in
1575 * async compaction, or due to a fatal signal detected. In that
1576 * case do not try further zones and signal need_resched()
1579 if ((zone_contended == COMPACT_CONTENDED_SCHED)
1580 || fatal_signal_pending(current)) {
1581 *contended = COMPACT_CONTENDED_SCHED;
1588 * We might not have tried all the zones, so be conservative
1589 * and assume they are not all lock contended.
1591 all_zones_contended = 0;
1596 * If at least one zone wasn't deferred or skipped, we report if all
1597 * zones that were tried were lock contended.
1599 if (rc > COMPACT_SKIPPED && all_zones_contended)
1600 *contended = COMPACT_CONTENDED_LOCK;
1606 /* Compact all zones within a node */
1607 static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
1612 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1614 zone = &pgdat->node_zones[zoneid];
1615 if (!populated_zone(zone))
1618 cc->nr_freepages = 0;
1619 cc->nr_migratepages = 0;
1621 INIT_LIST_HEAD(&cc->freepages);
1622 INIT_LIST_HEAD(&cc->migratepages);
1625 * When called via /proc/sys/vm/compact_memory
1626 * this makes sure we compact the whole zone regardless of
1627 * cached scanner positions.
1629 if (is_via_compact_memory(cc->order))
1630 __reset_isolation_suitable(zone);
1632 if (is_via_compact_memory(cc->order) ||
1633 !compaction_deferred(zone, cc->order))
1634 compact_zone(zone, cc);
1636 VM_BUG_ON(!list_empty(&cc->freepages));
1637 VM_BUG_ON(!list_empty(&cc->migratepages));
1639 if (is_via_compact_memory(cc->order))
1642 if (zone_watermark_ok(zone, cc->order,
1643 low_wmark_pages(zone), 0, 0))
1644 compaction_defer_reset(zone, cc->order, false);
1648 void compact_pgdat(pg_data_t *pgdat, int order)
1650 struct compact_control cc = {
1652 .mode = MIGRATE_ASYNC,
1658 __compact_pgdat(pgdat, &cc);
1661 static void compact_node(int nid)
1663 struct compact_control cc = {
1665 .mode = MIGRATE_SYNC,
1666 .ignore_skip_hint = true,
1669 __compact_pgdat(NODE_DATA(nid), &cc);
1672 /* Compact all nodes in the system */
1673 static void compact_nodes(void)
1677 /* Flush pending updates to the LRU lists */
1678 lru_add_drain_all();
1680 for_each_online_node(nid)
1684 /* The written value is actually unused, all memory is compacted */
1685 int sysctl_compact_memory;
1688 * This is the entry point for compacting all nodes via
1689 * /proc/sys/vm/compact_memory
1691 int sysctl_compaction_handler(struct ctl_table *table, int write,
1692 void __user *buffer, size_t *length, loff_t *ppos)
1700 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1701 void __user *buffer, size_t *length, loff_t *ppos)
1703 proc_dointvec_minmax(table, write, buffer, length, ppos);
1708 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1709 static ssize_t sysfs_compact_node(struct device *dev,
1710 struct device_attribute *attr,
1711 const char *buf, size_t count)
1715 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1716 /* Flush pending updates to the LRU lists */
1717 lru_add_drain_all();
1724 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1726 int compaction_register_node(struct node *node)
1728 return device_create_file(&node->dev, &dev_attr_compact);
1731 void compaction_unregister_node(struct node *node)
1733 return device_remove_file(&node->dev, &dev_attr_compact);
1735 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1737 static inline bool kcompactd_work_requested(pg_data_t *pgdat)
1739 return pgdat->kcompactd_max_order > 0 || kthread_should_stop();
1742 static bool kcompactd_node_suitable(pg_data_t *pgdat)
1746 enum zone_type classzone_idx = pgdat->kcompactd_classzone_idx;
1748 for (zoneid = 0; zoneid < classzone_idx; zoneid++) {
1749 zone = &pgdat->node_zones[zoneid];
1751 if (!populated_zone(zone))
1754 if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0,
1755 classzone_idx) == COMPACT_CONTINUE)
1762 static void kcompactd_do_work(pg_data_t *pgdat)
1765 * With no special task, compact all zones so that a page of requested
1766 * order is allocatable.
1770 struct compact_control cc = {
1771 .order = pgdat->kcompactd_max_order,
1772 .classzone_idx = pgdat->kcompactd_classzone_idx,
1773 .mode = MIGRATE_SYNC_LIGHT,
1774 .ignore_skip_hint = true,
1777 bool success = false;
1779 trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
1781 count_vm_event(KCOMPACTD_WAKE);
1783 for (zoneid = 0; zoneid < cc.classzone_idx; zoneid++) {
1786 zone = &pgdat->node_zones[zoneid];
1787 if (!populated_zone(zone))
1790 if (compaction_deferred(zone, cc.order))
1793 if (compaction_suitable(zone, cc.order, 0, zoneid) !=
1797 cc.nr_freepages = 0;
1798 cc.nr_migratepages = 0;
1800 INIT_LIST_HEAD(&cc.freepages);
1801 INIT_LIST_HEAD(&cc.migratepages);
1803 if (kthread_should_stop())
1805 status = compact_zone(zone, &cc);
1807 if (zone_watermark_ok(zone, cc.order, low_wmark_pages(zone),
1808 cc.classzone_idx, 0)) {
1810 compaction_defer_reset(zone, cc.order, false);
1811 } else if (status == COMPACT_COMPLETE) {
1813 * We use sync migration mode here, so we defer like
1814 * sync direct compaction does.
1816 defer_compaction(zone, cc.order);
1819 VM_BUG_ON(!list_empty(&cc.freepages));
1820 VM_BUG_ON(!list_empty(&cc.migratepages));
1824 * Regardless of success, we are done until woken up next. But remember
1825 * the requested order/classzone_idx in case it was higher/tighter than
1828 if (pgdat->kcompactd_max_order <= cc.order)
1829 pgdat->kcompactd_max_order = 0;
1830 if (pgdat->kcompactd_classzone_idx >= cc.classzone_idx)
1831 pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
1834 void wakeup_kcompactd(pg_data_t *pgdat, int order, int classzone_idx)
1839 if (pgdat->kcompactd_max_order < order)
1840 pgdat->kcompactd_max_order = order;
1842 if (pgdat->kcompactd_classzone_idx > classzone_idx)
1843 pgdat->kcompactd_classzone_idx = classzone_idx;
1845 if (!waitqueue_active(&pgdat->kcompactd_wait))
1848 if (!kcompactd_node_suitable(pgdat))
1851 trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
1853 wake_up_interruptible(&pgdat->kcompactd_wait);
1857 * The background compaction daemon, started as a kernel thread
1858 * from the init process.
1860 static int kcompactd(void *p)
1862 pg_data_t *pgdat = (pg_data_t*)p;
1863 struct task_struct *tsk = current;
1865 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1867 if (!cpumask_empty(cpumask))
1868 set_cpus_allowed_ptr(tsk, cpumask);
1872 pgdat->kcompactd_max_order = 0;
1873 pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
1875 while (!kthread_should_stop()) {
1876 trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
1877 wait_event_freezable(pgdat->kcompactd_wait,
1878 kcompactd_work_requested(pgdat));
1880 kcompactd_do_work(pgdat);
1887 * This kcompactd start function will be called by init and node-hot-add.
1888 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
1890 int kcompactd_run(int nid)
1892 pg_data_t *pgdat = NODE_DATA(nid);
1895 if (pgdat->kcompactd)
1898 pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
1899 if (IS_ERR(pgdat->kcompactd)) {
1900 pr_err("Failed to start kcompactd on node %d\n", nid);
1901 ret = PTR_ERR(pgdat->kcompactd);
1902 pgdat->kcompactd = NULL;
1908 * Called by memory hotplug when all memory in a node is offlined. Caller must
1909 * hold mem_hotplug_begin/end().
1911 void kcompactd_stop(int nid)
1913 struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
1916 kthread_stop(kcompactd);
1917 NODE_DATA(nid)->kcompactd = NULL;
1922 * It's optimal to keep kcompactd on the same CPUs as their memory, but
1923 * not required for correctness. So if the last cpu in a node goes
1924 * away, we get changed to run anywhere: as the first one comes back,
1925 * restore their cpu bindings.
1927 static int cpu_callback(struct notifier_block *nfb, unsigned long action,
1932 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
1933 for_each_node_state(nid, N_MEMORY) {
1934 pg_data_t *pgdat = NODE_DATA(nid);
1935 const struct cpumask *mask;
1937 mask = cpumask_of_node(pgdat->node_id);
1939 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
1940 /* One of our CPUs online: restore mask */
1941 set_cpus_allowed_ptr(pgdat->kcompactd, mask);
1947 static int __init kcompactd_init(void)
1951 for_each_node_state(nid, N_MEMORY)
1953 hotcpu_notifier(cpu_callback, 0);
1956 subsys_initcall(kcompactd_init)
1958 #endif /* CONFIG_COMPACTION */