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/swap.h>
11 #include <linux/migrate.h>
12 #include <linux/compaction.h>
13 #include <linux/mm_inline.h>
14 #include <linux/backing-dev.h>
15 #include <linux/sysctl.h>
16 #include <linux/sysfs.h>
17 #include <linux/balloon_compaction.h>
18 #include <linux/page-isolation.h>
19 #include <linux/kasan.h>
22 #ifdef CONFIG_COMPACTION
23 static inline void count_compact_event(enum vm_event_item item)
28 static inline void count_compact_events(enum vm_event_item item, long delta)
30 count_vm_events(item, delta);
33 #define count_compact_event(item) do { } while (0)
34 #define count_compact_events(item, delta) do { } while (0)
37 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
38 #ifdef CONFIG_TRACEPOINTS
39 static const char *const compaction_status_string[] = {
50 #define CREATE_TRACE_POINTS
51 #include <trace/events/compaction.h>
53 static unsigned long release_freepages(struct list_head *freelist)
55 struct page *page, *next;
56 unsigned long high_pfn = 0;
58 list_for_each_entry_safe(page, next, freelist, lru) {
59 unsigned long pfn = page_to_pfn(page);
69 static void map_pages(struct list_head *list)
73 list_for_each_entry(page, list, lru) {
74 arch_alloc_page(page, 0);
75 kernel_map_pages(page, 1, 1);
76 kasan_alloc_pages(page, 0);
80 static inline bool migrate_async_suitable(int migratetype)
82 return is_migrate_cma(migratetype) || migratetype == MIGRATE_MOVABLE;
86 * Check that the whole (or subset of) a pageblock given by the interval of
87 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
88 * with the migration of free compaction scanner. The scanners then need to
89 * use only pfn_valid_within() check for arches that allow holes within
92 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
94 * It's possible on some configurations to have a setup like node0 node1 node0
95 * i.e. it's possible that all pages within a zones range of pages do not
96 * belong to a single zone. We assume that a border between node0 and node1
97 * can occur within a single pageblock, but not a node0 node1 node0
98 * interleaving within a single pageblock. It is therefore sufficient to check
99 * the first and last page of a pageblock and avoid checking each individual
100 * page in a pageblock.
102 static struct page *pageblock_pfn_to_page(unsigned long start_pfn,
103 unsigned long end_pfn, struct zone *zone)
105 struct page *start_page;
106 struct page *end_page;
108 /* end_pfn is one past the range we are checking */
111 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
114 start_page = pfn_to_page(start_pfn);
116 if (page_zone(start_page) != zone)
119 end_page = pfn_to_page(end_pfn);
121 /* This gives a shorter code than deriving page_zone(end_page) */
122 if (page_zone_id(start_page) != page_zone_id(end_page))
128 #ifdef CONFIG_COMPACTION
130 /* Do not skip compaction more than 64 times */
131 #define COMPACT_MAX_DEFER_SHIFT 6
134 * Compaction is deferred when compaction fails to result in a page
135 * allocation success. 1 << compact_defer_limit compactions are skipped up
136 * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
138 void defer_compaction(struct zone *zone, int order)
140 zone->compact_considered = 0;
141 zone->compact_defer_shift++;
143 if (order < zone->compact_order_failed)
144 zone->compact_order_failed = order;
146 if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
147 zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
149 trace_mm_compaction_defer_compaction(zone, order);
152 /* Returns true if compaction should be skipped this time */
153 bool compaction_deferred(struct zone *zone, int order)
155 unsigned long defer_limit = 1UL << zone->compact_defer_shift;
157 if (order < zone->compact_order_failed)
160 /* Avoid possible overflow */
161 if (++zone->compact_considered > defer_limit)
162 zone->compact_considered = defer_limit;
164 if (zone->compact_considered >= defer_limit)
167 trace_mm_compaction_deferred(zone, order);
173 * Update defer tracking counters after successful compaction of given order,
174 * which means an allocation either succeeded (alloc_success == true) or is
175 * expected to succeed.
177 void compaction_defer_reset(struct zone *zone, int order,
181 zone->compact_considered = 0;
182 zone->compact_defer_shift = 0;
184 if (order >= zone->compact_order_failed)
185 zone->compact_order_failed = order + 1;
187 trace_mm_compaction_defer_reset(zone, order);
190 /* Returns true if restarting compaction after many failures */
191 bool compaction_restarting(struct zone *zone, int order)
193 if (order < zone->compact_order_failed)
196 return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
197 zone->compact_considered >= 1UL << zone->compact_defer_shift;
200 /* Returns true if the pageblock should be scanned for pages to isolate. */
201 static inline bool isolation_suitable(struct compact_control *cc,
204 if (cc->ignore_skip_hint)
207 return !get_pageblock_skip(page);
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_cached_migrate_pfn[0] = start_pfn;
222 zone->compact_cached_migrate_pfn[1] = start_pfn;
223 zone->compact_cached_free_pfn = end_pfn;
224 zone->compact_blockskip_flush = false;
226 /* Walk the zone and mark every pageblock as suitable for isolation */
227 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
235 page = pfn_to_page(pfn);
236 if (zone != page_zone(page))
239 clear_pageblock_skip(page);
243 void reset_isolation_suitable(pg_data_t *pgdat)
247 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
248 struct zone *zone = &pgdat->node_zones[zoneid];
249 if (!populated_zone(zone))
252 /* Only flush if a full compaction finished recently */
253 if (zone->compact_blockskip_flush)
254 __reset_isolation_suitable(zone);
259 * If no pages were isolated then mark this pageblock to be skipped in the
260 * future. The information is later cleared by __reset_isolation_suitable().
262 static void update_pageblock_skip(struct compact_control *cc,
263 struct page *page, unsigned long nr_isolated,
264 bool migrate_scanner)
266 struct zone *zone = cc->zone;
269 if (cc->ignore_skip_hint)
278 set_pageblock_skip(page);
280 pfn = page_to_pfn(page);
282 /* Update where async and sync compaction should restart */
283 if (migrate_scanner) {
284 if (pfn > zone->compact_cached_migrate_pfn[0])
285 zone->compact_cached_migrate_pfn[0] = pfn;
286 if (cc->mode != MIGRATE_ASYNC &&
287 pfn > zone->compact_cached_migrate_pfn[1])
288 zone->compact_cached_migrate_pfn[1] = pfn;
290 if (pfn < zone->compact_cached_free_pfn)
291 zone->compact_cached_free_pfn = pfn;
295 static inline bool isolation_suitable(struct compact_control *cc,
301 static void update_pageblock_skip(struct compact_control *cc,
302 struct page *page, unsigned long nr_isolated,
303 bool migrate_scanner)
306 #endif /* CONFIG_COMPACTION */
309 * Compaction requires the taking of some coarse locks that are potentially
310 * very heavily contended. For async compaction, back out if the lock cannot
311 * be taken immediately. For sync compaction, spin on the lock if needed.
313 * Returns true if the lock is held
314 * Returns false if the lock is not held and compaction should abort
316 static bool compact_trylock_irqsave(spinlock_t *lock, unsigned long *flags,
317 struct compact_control *cc)
319 if (cc->mode == MIGRATE_ASYNC) {
320 if (!spin_trylock_irqsave(lock, *flags)) {
321 cc->contended = COMPACT_CONTENDED_LOCK;
325 spin_lock_irqsave(lock, *flags);
332 * Compaction requires the taking of some coarse locks that are potentially
333 * very heavily contended. The lock should be periodically unlocked to avoid
334 * having disabled IRQs for a long time, even when there is nobody waiting on
335 * the lock. It might also be that allowing the IRQs will result in
336 * need_resched() becoming true. If scheduling is needed, async compaction
337 * aborts. Sync compaction schedules.
338 * Either compaction type will also abort if a fatal signal is pending.
339 * In either case if the lock was locked, it is dropped and not regained.
341 * Returns true if compaction should abort due to fatal signal pending, or
342 * async compaction due to need_resched()
343 * Returns false when compaction can continue (sync compaction might have
346 static bool compact_unlock_should_abort(spinlock_t *lock,
347 unsigned long flags, bool *locked, struct compact_control *cc)
350 spin_unlock_irqrestore(lock, flags);
354 if (fatal_signal_pending(current)) {
355 cc->contended = COMPACT_CONTENDED_SCHED;
359 if (need_resched()) {
360 if (cc->mode == MIGRATE_ASYNC) {
361 cc->contended = COMPACT_CONTENDED_SCHED;
371 * Aside from avoiding lock contention, compaction also periodically checks
372 * need_resched() and either schedules in sync compaction or aborts async
373 * compaction. This is similar to what compact_unlock_should_abort() does, but
374 * is used where no lock is concerned.
376 * Returns false when no scheduling was needed, or sync compaction scheduled.
377 * Returns true when async compaction should abort.
379 static inline bool compact_should_abort(struct compact_control *cc)
381 /* async compaction aborts if contended */
382 if (need_resched()) {
383 if (cc->mode == MIGRATE_ASYNC) {
384 cc->contended = COMPACT_CONTENDED_SCHED;
395 * Isolate free pages onto a private freelist. If @strict is true, will abort
396 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
397 * (even though it may still end up isolating some pages).
399 static unsigned long isolate_freepages_block(struct compact_control *cc,
400 unsigned long *start_pfn,
401 unsigned long end_pfn,
402 struct list_head *freelist,
405 int nr_scanned = 0, total_isolated = 0;
406 struct page *cursor, *valid_page = NULL;
407 unsigned long flags = 0;
409 unsigned long blockpfn = *start_pfn;
411 cursor = pfn_to_page(blockpfn);
413 /* Isolate free pages. */
414 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
416 struct page *page = cursor;
419 * Periodically drop the lock (if held) regardless of its
420 * contention, to give chance to IRQs. Abort if fatal signal
421 * pending or async compaction detects need_resched()
423 if (!(blockpfn % SWAP_CLUSTER_MAX)
424 && compact_unlock_should_abort(&cc->zone->lock, flags,
429 if (!pfn_valid_within(blockpfn))
434 if (!PageBuddy(page))
438 * If we already hold the lock, we can skip some rechecking.
439 * Note that if we hold the lock now, checked_pageblock was
440 * already set in some previous iteration (or strict is true),
441 * so it is correct to skip the suitable migration target
446 * The zone lock must be held to isolate freepages.
447 * Unfortunately this is a very coarse lock and can be
448 * heavily contended if there are parallel allocations
449 * or parallel compactions. For async compaction do not
450 * spin on the lock and we acquire the lock as late as
453 locked = compact_trylock_irqsave(&cc->zone->lock,
458 /* Recheck this is a buddy page under lock */
459 if (!PageBuddy(page))
463 /* Found a free page, break it into order-0 pages */
464 isolated = split_free_page(page);
465 total_isolated += isolated;
466 for (i = 0; i < isolated; i++) {
467 list_add(&page->lru, freelist);
471 /* If a page was split, advance to the end of it */
473 cc->nr_freepages += isolated;
475 cc->nr_migratepages <= cc->nr_freepages) {
476 blockpfn += isolated;
480 blockpfn += isolated - 1;
481 cursor += isolated - 1;
493 trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
494 nr_scanned, total_isolated);
496 /* Record how far we have got within the block */
497 *start_pfn = blockpfn;
500 * If strict isolation is requested by CMA then check that all the
501 * pages requested were isolated. If there were any failures, 0 is
502 * returned and CMA will fail.
504 if (strict && blockpfn < end_pfn)
508 spin_unlock_irqrestore(&cc->zone->lock, flags);
510 /* Update the pageblock-skip if the whole pageblock was scanned */
511 if (blockpfn == end_pfn)
512 update_pageblock_skip(cc, valid_page, total_isolated, false);
514 count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
516 count_compact_events(COMPACTISOLATED, total_isolated);
517 return total_isolated;
521 * isolate_freepages_range() - isolate free pages.
522 * @start_pfn: The first PFN to start isolating.
523 * @end_pfn: The one-past-last PFN.
525 * Non-free pages, invalid PFNs, or zone boundaries within the
526 * [start_pfn, end_pfn) range are considered errors, cause function to
527 * undo its actions and return zero.
529 * Otherwise, function returns one-past-the-last PFN of isolated page
530 * (which may be greater then end_pfn if end fell in a middle of
534 isolate_freepages_range(struct compact_control *cc,
535 unsigned long start_pfn, unsigned long end_pfn)
537 unsigned long isolated, pfn, block_end_pfn;
541 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
543 for (; pfn < end_pfn; pfn += isolated,
544 block_end_pfn += pageblock_nr_pages) {
545 /* Protect pfn from changing by isolate_freepages_block */
546 unsigned long isolate_start_pfn = pfn;
548 block_end_pfn = min(block_end_pfn, end_pfn);
551 * pfn could pass the block_end_pfn if isolated freepage
552 * is more than pageblock order. In this case, we adjust
553 * scanning range to right one.
555 if (pfn >= block_end_pfn) {
556 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
557 block_end_pfn = min(block_end_pfn, end_pfn);
560 if (!pageblock_pfn_to_page(pfn, block_end_pfn, cc->zone))
563 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
564 block_end_pfn, &freelist, true);
567 * In strict mode, isolate_freepages_block() returns 0 if
568 * there are any holes in the block (ie. invalid PFNs or
575 * If we managed to isolate pages, it is always (1 << n) *
576 * pageblock_nr_pages for some non-negative n. (Max order
577 * page may span two pageblocks).
581 /* split_free_page does not map the pages */
582 map_pages(&freelist);
585 /* Loop terminated early, cleanup. */
586 release_freepages(&freelist);
590 /* We don't use freelists for anything. */
594 /* Update the number of anon and file isolated pages in the zone */
595 static void acct_isolated(struct zone *zone, struct compact_control *cc)
598 unsigned int count[2] = { 0, };
600 if (list_empty(&cc->migratepages))
603 list_for_each_entry(page, &cc->migratepages, lru)
604 count[!!page_is_file_cache(page)]++;
606 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
607 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
610 /* Similar to reclaim, but different enough that they don't share logic */
611 static bool too_many_isolated(struct zone *zone)
613 unsigned long active, inactive, isolated;
615 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
616 zone_page_state(zone, NR_INACTIVE_ANON);
617 active = zone_page_state(zone, NR_ACTIVE_FILE) +
618 zone_page_state(zone, NR_ACTIVE_ANON);
619 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
620 zone_page_state(zone, NR_ISOLATED_ANON);
622 return isolated > (inactive + active) / 2;
626 * isolate_migratepages_block() - isolate all migrate-able pages within
628 * @cc: Compaction control structure.
629 * @low_pfn: The first PFN to isolate
630 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
631 * @isolate_mode: Isolation mode to be used.
633 * Isolate all pages that can be migrated from the range specified by
634 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
635 * Returns zero if there is a fatal signal pending, otherwise PFN of the
636 * first page that was not scanned (which may be both less, equal to or more
639 * The pages are isolated on cc->migratepages list (not required to be empty),
640 * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
641 * is neither read nor updated.
644 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
645 unsigned long end_pfn, isolate_mode_t isolate_mode)
647 struct zone *zone = cc->zone;
648 unsigned long nr_scanned = 0, nr_isolated = 0;
649 struct list_head *migratelist = &cc->migratepages;
650 struct lruvec *lruvec;
651 unsigned long flags = 0;
653 struct page *page = NULL, *valid_page = NULL;
654 unsigned long start_pfn = low_pfn;
657 * Ensure that there are not too many pages isolated from the LRU
658 * list by either parallel reclaimers or compaction. If there are,
659 * delay for some time until fewer pages are isolated
661 while (unlikely(too_many_isolated(zone))) {
662 /* async migration should just abort */
663 if (cc->mode == MIGRATE_ASYNC)
666 congestion_wait(BLK_RW_ASYNC, HZ/10);
668 if (fatal_signal_pending(current))
672 if (compact_should_abort(cc))
675 /* Time to isolate some pages for migration */
676 for (; low_pfn < end_pfn; low_pfn++) {
678 * Periodically drop the lock (if held) regardless of its
679 * contention, to give chance to IRQs. Abort async compaction
682 if (!(low_pfn % SWAP_CLUSTER_MAX)
683 && compact_unlock_should_abort(&zone->lru_lock, flags,
687 if (!pfn_valid_within(low_pfn))
691 page = pfn_to_page(low_pfn);
697 * Skip if free. We read page order here without zone lock
698 * which is generally unsafe, but the race window is small and
699 * the worst thing that can happen is that we skip some
700 * potential isolation targets.
702 if (PageBuddy(page)) {
703 unsigned long freepage_order = page_order_unsafe(page);
706 * Without lock, we cannot be sure that what we got is
707 * a valid page order. Consider only values in the
708 * valid order range to prevent low_pfn overflow.
710 if (freepage_order > 0 && freepage_order < MAX_ORDER)
711 low_pfn += (1UL << freepage_order) - 1;
716 * Check may be lockless but that's ok as we recheck later.
717 * It's possible to migrate LRU pages and balloon pages
718 * Skip any other type of page
720 if (!PageLRU(page)) {
721 if (unlikely(balloon_page_movable(page))) {
722 if (balloon_page_isolate(page)) {
723 /* Successfully isolated */
724 goto isolate_success;
731 * PageLRU is set. lru_lock normally excludes isolation
732 * splitting and collapsing (collapsing has already happened
733 * if PageLRU is set) but the lock is not necessarily taken
734 * here and it is wasteful to take it just to check transhuge.
735 * Check TransHuge without lock and skip the whole pageblock if
736 * it's either a transhuge or hugetlbfs page, as calling
737 * compound_order() without preventing THP from splitting the
738 * page underneath us may return surprising results.
740 if (PageTransHuge(page)) {
742 low_pfn = ALIGN(low_pfn + 1,
743 pageblock_nr_pages) - 1;
745 low_pfn += (1 << compound_order(page)) - 1;
751 * Migration will fail if an anonymous page is pinned in memory,
752 * so avoid taking lru_lock and isolating it unnecessarily in an
753 * admittedly racy check.
755 if (!page_mapping(page) &&
756 page_count(page) > page_mapcount(page))
759 /* If we already hold the lock, we can skip some rechecking */
761 locked = compact_trylock_irqsave(&zone->lru_lock,
766 /* Recheck PageLRU and PageTransHuge under lock */
769 if (PageTransHuge(page)) {
770 low_pfn += (1 << compound_order(page)) - 1;
775 lruvec = mem_cgroup_page_lruvec(page, zone);
777 /* Try isolate the page */
778 if (__isolate_lru_page(page, isolate_mode) != 0)
781 VM_BUG_ON_PAGE(PageTransCompound(page), page);
783 /* Successfully isolated */
784 del_page_from_lru_list(page, lruvec, page_lru(page));
787 list_add(&page->lru, migratelist);
788 cc->nr_migratepages++;
791 /* Avoid isolating too much */
792 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
799 * The PageBuddy() check could have potentially brought us outside
800 * the range to be scanned.
802 if (unlikely(low_pfn > end_pfn))
806 spin_unlock_irqrestore(&zone->lru_lock, flags);
809 * Update the pageblock-skip information and cached scanner pfn,
810 * if the whole pageblock was scanned without isolating any page.
812 if (low_pfn == end_pfn)
813 update_pageblock_skip(cc, valid_page, nr_isolated, true);
815 trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
816 nr_scanned, nr_isolated);
818 count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
820 count_compact_events(COMPACTISOLATED, nr_isolated);
826 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
827 * @cc: Compaction control structure.
828 * @start_pfn: The first PFN to start isolating.
829 * @end_pfn: The one-past-last PFN.
831 * Returns zero if isolation fails fatally due to e.g. pending signal.
832 * Otherwise, function returns one-past-the-last PFN of isolated page
833 * (which may be greater than end_pfn if end fell in a middle of a THP page).
836 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
837 unsigned long end_pfn)
839 unsigned long pfn, block_end_pfn;
841 /* Scan block by block. First and last block may be incomplete */
843 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
845 for (; pfn < end_pfn; pfn = block_end_pfn,
846 block_end_pfn += pageblock_nr_pages) {
848 block_end_pfn = min(block_end_pfn, end_pfn);
850 if (!pageblock_pfn_to_page(pfn, block_end_pfn, cc->zone))
853 pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
854 ISOLATE_UNEVICTABLE);
857 * In case of fatal failure, release everything that might
858 * have been isolated in the previous iteration, and signal
859 * the failure back to caller.
862 putback_movable_pages(&cc->migratepages);
863 cc->nr_migratepages = 0;
867 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
870 acct_isolated(cc->zone, cc);
875 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
876 #ifdef CONFIG_COMPACTION
878 /* Returns true if the page is within a block suitable for migration to */
879 static bool suitable_migration_target(struct page *page)
881 /* If the page is a large free page, then disallow migration */
882 if (PageBuddy(page)) {
884 * We are checking page_order without zone->lock taken. But
885 * the only small danger is that we skip a potentially suitable
886 * pageblock, so it's not worth to check order for valid range.
888 if (page_order_unsafe(page) >= pageblock_order)
892 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
893 if (migrate_async_suitable(get_pageblock_migratetype(page)))
896 /* Otherwise skip the block */
901 * Test whether the free scanner has reached the same or lower pageblock than
902 * the migration scanner, and compaction should thus terminate.
904 static inline bool compact_scanners_met(struct compact_control *cc)
906 return (cc->free_pfn >> pageblock_order)
907 <= (cc->migrate_pfn >> pageblock_order);
911 * Based on information in the current compact_control, find blocks
912 * suitable for isolating free pages from and then isolate them.
914 static void isolate_freepages(struct compact_control *cc)
916 struct zone *zone = cc->zone;
918 unsigned long block_start_pfn; /* start of current pageblock */
919 unsigned long isolate_start_pfn; /* exact pfn we start at */
920 unsigned long block_end_pfn; /* end of current pageblock */
921 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
922 struct list_head *freelist = &cc->freepages;
925 * Initialise the free scanner. The starting point is where we last
926 * successfully isolated from, zone-cached value, or the end of the
927 * zone when isolating for the first time. For looping we also need
928 * this pfn aligned down to the pageblock boundary, because we do
929 * block_start_pfn -= pageblock_nr_pages in the for loop.
930 * For ending point, take care when isolating in last pageblock of a
931 * a zone which ends in the middle of a pageblock.
932 * The low boundary is the end of the pageblock the migration scanner
935 isolate_start_pfn = cc->free_pfn;
936 block_start_pfn = cc->free_pfn & ~(pageblock_nr_pages-1);
937 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
939 low_pfn = ALIGN(cc->migrate_pfn + 1, pageblock_nr_pages);
942 * Isolate free pages until enough are available to migrate the
943 * pages on cc->migratepages. We stop searching if the migrate
944 * and free page scanners meet or enough free pages are isolated.
946 for (; block_start_pfn >= low_pfn;
947 block_end_pfn = block_start_pfn,
948 block_start_pfn -= pageblock_nr_pages,
949 isolate_start_pfn = block_start_pfn) {
952 * This can iterate a massively long zone without finding any
953 * suitable migration targets, so periodically check if we need
954 * to schedule, or even abort async compaction.
956 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
957 && compact_should_abort(cc))
960 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
965 /* Check the block is suitable for migration */
966 if (!suitable_migration_target(page))
969 /* If isolation recently failed, do not retry */
970 if (!isolation_suitable(cc, page))
973 /* Found a block suitable for isolating free pages from. */
974 isolate_freepages_block(cc, &isolate_start_pfn,
975 block_end_pfn, freelist, false);
978 * If we isolated enough freepages, or aborted due to async
979 * compaction being contended, terminate the loop.
980 * Remember where the free scanner should restart next time,
981 * which is where isolate_freepages_block() left off.
982 * But if it scanned the whole pageblock, isolate_start_pfn
983 * now points at block_end_pfn, which is the start of the next
985 * In that case we will however want to restart at the start
986 * of the previous pageblock.
988 if ((cc->nr_freepages >= cc->nr_migratepages)
990 if (isolate_start_pfn >= block_end_pfn)
992 block_start_pfn - pageblock_nr_pages;
996 * isolate_freepages_block() should not terminate
997 * prematurely unless contended, or isolated enough
999 VM_BUG_ON(isolate_start_pfn < block_end_pfn);
1003 /* split_free_page does not map the pages */
1004 map_pages(freelist);
1007 * Record where the free scanner will restart next time. Either we
1008 * broke from the loop and set isolate_start_pfn based on the last
1009 * call to isolate_freepages_block(), or we met the migration scanner
1010 * and the loop terminated due to isolate_start_pfn < low_pfn
1012 cc->free_pfn = isolate_start_pfn;
1016 * This is a migrate-callback that "allocates" freepages by taking pages
1017 * from the isolated freelists in the block we are migrating to.
1019 static struct page *compaction_alloc(struct page *migratepage,
1023 struct compact_control *cc = (struct compact_control *)data;
1024 struct page *freepage;
1027 * Isolate free pages if necessary, and if we are not aborting due to
1030 if (list_empty(&cc->freepages)) {
1032 isolate_freepages(cc);
1034 if (list_empty(&cc->freepages))
1038 freepage = list_entry(cc->freepages.next, struct page, lru);
1039 list_del(&freepage->lru);
1046 * This is a migrate-callback that "frees" freepages back to the isolated
1047 * freelist. All pages on the freelist are from the same zone, so there is no
1048 * special handling needed for NUMA.
1050 static void compaction_free(struct page *page, unsigned long data)
1052 struct compact_control *cc = (struct compact_control *)data;
1054 list_add(&page->lru, &cc->freepages);
1058 /* possible outcome of isolate_migratepages */
1060 ISOLATE_ABORT, /* Abort compaction now */
1061 ISOLATE_NONE, /* No pages isolated, continue scanning */
1062 ISOLATE_SUCCESS, /* Pages isolated, migrate */
1063 } isolate_migrate_t;
1066 * Allow userspace to control policy on scanning the unevictable LRU for
1067 * compactable pages.
1069 int sysctl_compact_unevictable_allowed __read_mostly = 1;
1072 * Isolate all pages that can be migrated from the first suitable block,
1073 * starting at the block pointed to by the migrate scanner pfn within
1076 static isolate_migrate_t isolate_migratepages(struct zone *zone,
1077 struct compact_control *cc)
1079 unsigned long low_pfn, end_pfn;
1081 const isolate_mode_t isolate_mode =
1082 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1083 (cc->mode == MIGRATE_ASYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1086 * Start at where we last stopped, or beginning of the zone as
1087 * initialized by compact_zone()
1089 low_pfn = cc->migrate_pfn;
1091 /* Only scan within a pageblock boundary */
1092 end_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages);
1095 * Iterate over whole pageblocks until we find the first suitable.
1096 * Do not cross the free scanner.
1098 for (; end_pfn <= cc->free_pfn;
1099 low_pfn = end_pfn, end_pfn += pageblock_nr_pages) {
1102 * This can potentially iterate a massively long zone with
1103 * many pageblocks unsuitable, so periodically check if we
1104 * need to schedule, or even abort async compaction.
1106 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1107 && compact_should_abort(cc))
1110 page = pageblock_pfn_to_page(low_pfn, end_pfn, zone);
1114 /* If isolation recently failed, do not retry */
1115 if (!isolation_suitable(cc, page))
1119 * For async compaction, also only scan in MOVABLE blocks.
1120 * Async compaction is optimistic to see if the minimum amount
1121 * of work satisfies the allocation.
1123 if (cc->mode == MIGRATE_ASYNC &&
1124 !migrate_async_suitable(get_pageblock_migratetype(page)))
1127 /* Perform the isolation */
1128 low_pfn = isolate_migratepages_block(cc, low_pfn, end_pfn,
1131 if (!low_pfn || cc->contended) {
1132 acct_isolated(zone, cc);
1133 return ISOLATE_ABORT;
1137 * Either we isolated something and proceed with migration. Or
1138 * we failed and compact_zone should decide if we should
1144 acct_isolated(zone, cc);
1145 /* Record where migration scanner will be restarted. */
1146 cc->migrate_pfn = low_pfn;
1148 return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
1151 static int __compact_finished(struct zone *zone, struct compact_control *cc,
1152 const int migratetype)
1155 unsigned long watermark;
1157 if (cc->contended || fatal_signal_pending(current))
1158 return COMPACT_PARTIAL;
1160 /* Compaction run completes if the migrate and free scanner meet */
1161 if (compact_scanners_met(cc)) {
1162 /* Let the next compaction start anew. */
1163 zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
1164 zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
1165 zone->compact_cached_free_pfn = zone_end_pfn(zone);
1168 * Mark that the PG_migrate_skip information should be cleared
1169 * by kswapd when it goes to sleep. kswapd does not set the
1170 * flag itself as the decision to be clear should be directly
1171 * based on an allocation request.
1173 if (!current_is_kswapd())
1174 zone->compact_blockskip_flush = true;
1176 return COMPACT_COMPLETE;
1180 * order == -1 is expected when compacting via
1181 * /proc/sys/vm/compact_memory
1183 if (cc->order == -1)
1184 return COMPACT_CONTINUE;
1186 /* Compaction run is not finished if the watermark is not met */
1187 watermark = low_wmark_pages(zone);
1189 if (!zone_watermark_ok(zone, cc->order, watermark, cc->classzone_idx,
1191 return COMPACT_CONTINUE;
1193 /* Direct compactor: Is a suitable page free? */
1194 for (order = cc->order; order < MAX_ORDER; order++) {
1195 struct free_area *area = &zone->free_area[order];
1198 /* Job done if page is free of the right migratetype */
1199 if (!list_empty(&area->free_list[migratetype]))
1200 return COMPACT_PARTIAL;
1203 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
1204 if (migratetype == MIGRATE_MOVABLE &&
1205 !list_empty(&area->free_list[MIGRATE_CMA]))
1206 return COMPACT_PARTIAL;
1209 * Job done if allocation would steal freepages from
1210 * other migratetype buddy lists.
1212 if (find_suitable_fallback(area, order, migratetype,
1213 true, &can_steal) != -1)
1214 return COMPACT_PARTIAL;
1217 return COMPACT_NO_SUITABLE_PAGE;
1220 static int compact_finished(struct zone *zone, struct compact_control *cc,
1221 const int migratetype)
1225 ret = __compact_finished(zone, cc, migratetype);
1226 trace_mm_compaction_finished(zone, cc->order, ret);
1227 if (ret == COMPACT_NO_SUITABLE_PAGE)
1228 ret = COMPACT_CONTINUE;
1234 * compaction_suitable: Is this suitable to run compaction on this zone now?
1236 * COMPACT_SKIPPED - If there are too few free pages for compaction
1237 * COMPACT_PARTIAL - If the allocation would succeed without compaction
1238 * COMPACT_CONTINUE - If compaction should run now
1240 static unsigned long __compaction_suitable(struct zone *zone, int order,
1241 int alloc_flags, int classzone_idx)
1244 unsigned long watermark;
1247 * order == -1 is expected when compacting via
1248 * /proc/sys/vm/compact_memory
1251 return COMPACT_CONTINUE;
1253 watermark = low_wmark_pages(zone);
1255 * If watermarks for high-order allocation are already met, there
1256 * should be no need for compaction at all.
1258 if (zone_watermark_ok(zone, order, watermark, classzone_idx,
1260 return COMPACT_PARTIAL;
1263 * Watermarks for order-0 must be met for compaction. Note the 2UL.
1264 * This is because during migration, copies of pages need to be
1265 * allocated and for a short time, the footprint is higher
1267 watermark += (2UL << order);
1268 if (!zone_watermark_ok(zone, 0, watermark, classzone_idx, alloc_flags))
1269 return COMPACT_SKIPPED;
1272 * fragmentation index determines if allocation failures are due to
1273 * low memory or external fragmentation
1275 * index of -1000 would imply allocations might succeed depending on
1276 * watermarks, but we already failed the high-order watermark check
1277 * index towards 0 implies failure is due to lack of memory
1278 * index towards 1000 implies failure is due to fragmentation
1280 * Only compact if a failure would be due to fragmentation.
1282 fragindex = fragmentation_index(zone, order);
1283 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
1284 return COMPACT_NOT_SUITABLE_ZONE;
1286 return COMPACT_CONTINUE;
1289 unsigned long compaction_suitable(struct zone *zone, int order,
1290 int alloc_flags, int classzone_idx)
1294 ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx);
1295 trace_mm_compaction_suitable(zone, order, ret);
1296 if (ret == COMPACT_NOT_SUITABLE_ZONE)
1297 ret = COMPACT_SKIPPED;
1302 static int compact_zone(struct zone *zone, struct compact_control *cc)
1305 unsigned long start_pfn = zone->zone_start_pfn;
1306 unsigned long end_pfn = zone_end_pfn(zone);
1307 const int migratetype = gfpflags_to_migratetype(cc->gfp_mask);
1308 const bool sync = cc->mode != MIGRATE_ASYNC;
1309 unsigned long last_migrated_pfn = 0;
1311 ret = compaction_suitable(zone, cc->order, cc->alloc_flags,
1314 case COMPACT_PARTIAL:
1315 case COMPACT_SKIPPED:
1316 /* Compaction is likely to fail */
1318 case COMPACT_CONTINUE:
1319 /* Fall through to compaction */
1324 * Clear pageblock skip if there were failures recently and compaction
1325 * is about to be retried after being deferred. kswapd does not do
1326 * this reset as it'll reset the cached information when going to sleep.
1328 if (compaction_restarting(zone, cc->order) && !current_is_kswapd())
1329 __reset_isolation_suitable(zone);
1332 * Setup to move all movable pages to the end of the zone. Used cached
1333 * information on where the scanners should start but check that it
1334 * is initialised by ensuring the values are within zone boundaries.
1336 cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync];
1337 cc->free_pfn = zone->compact_cached_free_pfn;
1338 if (cc->free_pfn < start_pfn || cc->free_pfn > end_pfn) {
1339 cc->free_pfn = end_pfn & ~(pageblock_nr_pages-1);
1340 zone->compact_cached_free_pfn = cc->free_pfn;
1342 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn > end_pfn) {
1343 cc->migrate_pfn = start_pfn;
1344 zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
1345 zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
1348 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
1349 cc->free_pfn, end_pfn, sync);
1351 migrate_prep_local();
1353 while ((ret = compact_finished(zone, cc, migratetype)) ==
1356 unsigned long isolate_start_pfn = cc->migrate_pfn;
1358 switch (isolate_migratepages(zone, cc)) {
1360 ret = COMPACT_PARTIAL;
1361 putback_movable_pages(&cc->migratepages);
1362 cc->nr_migratepages = 0;
1366 * We haven't isolated and migrated anything, but
1367 * there might still be unflushed migrations from
1368 * previous cc->order aligned block.
1371 case ISOLATE_SUCCESS:
1375 err = migrate_pages(&cc->migratepages, compaction_alloc,
1376 compaction_free, (unsigned long)cc, cc->mode,
1379 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
1382 /* All pages were either migrated or will be released */
1383 cc->nr_migratepages = 0;
1385 putback_movable_pages(&cc->migratepages);
1387 * migrate_pages() may return -ENOMEM when scanners meet
1388 * and we want compact_finished() to detect it
1390 if (err == -ENOMEM && !compact_scanners_met(cc)) {
1391 ret = COMPACT_PARTIAL;
1397 * Record where we could have freed pages by migration and not
1398 * yet flushed them to buddy allocator. We use the pfn that
1399 * isolate_migratepages() started from in this loop iteration
1400 * - this is the lowest page that could have been isolated and
1401 * then freed by migration.
1403 if (!last_migrated_pfn)
1404 last_migrated_pfn = isolate_start_pfn;
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 && last_migrated_pfn) {
1416 unsigned long current_block_start =
1417 cc->migrate_pfn & ~((1UL << cc->order) - 1);
1419 if (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 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_nr_pages-1);
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);
1457 static unsigned long compact_zone_order(struct zone *zone, int order,
1458 gfp_t gfp_mask, enum migrate_mode mode, int *contended,
1459 int alloc_flags, int classzone_idx)
1462 struct compact_control cc = {
1464 .nr_migratepages = 0,
1466 .gfp_mask = gfp_mask,
1469 .alloc_flags = alloc_flags,
1470 .classzone_idx = classzone_idx,
1472 INIT_LIST_HEAD(&cc.freepages);
1473 INIT_LIST_HEAD(&cc.migratepages);
1475 ret = compact_zone(zone, &cc);
1477 VM_BUG_ON(!list_empty(&cc.freepages));
1478 VM_BUG_ON(!list_empty(&cc.migratepages));
1480 *contended = cc.contended;
1484 int sysctl_extfrag_threshold = 500;
1487 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1488 * @gfp_mask: The GFP mask of the current allocation
1489 * @order: The order of the current allocation
1490 * @alloc_flags: The allocation flags of the current allocation
1491 * @ac: The context of current allocation
1492 * @mode: The migration mode for async, sync light, or sync migration
1493 * @contended: Return value that determines if compaction was aborted due to
1494 * need_resched() or lock contention
1496 * This is the main entry point for direct page compaction.
1498 unsigned long try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
1499 int alloc_flags, const struct alloc_context *ac,
1500 enum migrate_mode mode, int *contended)
1502 int may_enter_fs = gfp_mask & __GFP_FS;
1503 int may_perform_io = gfp_mask & __GFP_IO;
1506 int rc = COMPACT_DEFERRED;
1507 int all_zones_contended = COMPACT_CONTENDED_LOCK; /* init for &= op */
1509 *contended = COMPACT_CONTENDED_NONE;
1511 /* Check if the GFP flags allow compaction */
1512 if (!order || !may_enter_fs || !may_perform_io)
1513 return COMPACT_SKIPPED;
1515 trace_mm_compaction_try_to_compact_pages(order, gfp_mask, mode);
1517 /* Compact each zone in the list */
1518 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1523 if (compaction_deferred(zone, order))
1526 status = compact_zone_order(zone, order, gfp_mask, mode,
1527 &zone_contended, alloc_flags,
1529 rc = max(status, rc);
1531 * It takes at least one zone that wasn't lock contended
1532 * to clear all_zones_contended.
1534 all_zones_contended &= zone_contended;
1536 /* If a normal allocation would succeed, stop compacting */
1537 if (zone_watermark_ok(zone, order, low_wmark_pages(zone),
1538 ac->classzone_idx, alloc_flags)) {
1540 * We think the allocation will succeed in this zone,
1541 * but it is not certain, hence the false. The caller
1542 * will repeat this with true if allocation indeed
1543 * succeeds in this zone.
1545 compaction_defer_reset(zone, order, false);
1547 * It is possible that async compaction aborted due to
1548 * need_resched() and the watermarks were ok thanks to
1549 * somebody else freeing memory. The allocation can
1550 * however still fail so we better signal the
1551 * need_resched() contention anyway (this will not
1552 * prevent the allocation attempt).
1554 if (zone_contended == COMPACT_CONTENDED_SCHED)
1555 *contended = COMPACT_CONTENDED_SCHED;
1560 if (mode != MIGRATE_ASYNC && status == COMPACT_COMPLETE) {
1562 * We think that allocation won't succeed in this zone
1563 * so we defer compaction there. If it ends up
1564 * succeeding after all, it will be reset.
1566 defer_compaction(zone, order);
1570 * We might have stopped compacting due to need_resched() in
1571 * async compaction, or due to a fatal signal detected. In that
1572 * case do not try further zones and signal need_resched()
1575 if ((zone_contended == COMPACT_CONTENDED_SCHED)
1576 || fatal_signal_pending(current)) {
1577 *contended = COMPACT_CONTENDED_SCHED;
1584 * We might not have tried all the zones, so be conservative
1585 * and assume they are not all lock contended.
1587 all_zones_contended = 0;
1592 * If at least one zone wasn't deferred or skipped, we report if all
1593 * zones that were tried were lock contended.
1595 if (rc > COMPACT_SKIPPED && all_zones_contended)
1596 *contended = COMPACT_CONTENDED_LOCK;
1602 /* Compact all zones within a node */
1603 static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
1608 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1610 zone = &pgdat->node_zones[zoneid];
1611 if (!populated_zone(zone))
1614 cc->nr_freepages = 0;
1615 cc->nr_migratepages = 0;
1617 INIT_LIST_HEAD(&cc->freepages);
1618 INIT_LIST_HEAD(&cc->migratepages);
1621 * When called via /proc/sys/vm/compact_memory
1622 * this makes sure we compact the whole zone regardless of
1623 * cached scanner positions.
1625 if (cc->order == -1)
1626 __reset_isolation_suitable(zone);
1628 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
1629 compact_zone(zone, cc);
1631 if (cc->order > 0) {
1632 if (zone_watermark_ok(zone, cc->order,
1633 low_wmark_pages(zone), 0, 0))
1634 compaction_defer_reset(zone, cc->order, false);
1637 VM_BUG_ON(!list_empty(&cc->freepages));
1638 VM_BUG_ON(!list_empty(&cc->migratepages));
1642 void compact_pgdat(pg_data_t *pgdat, int order)
1644 struct compact_control cc = {
1646 .mode = MIGRATE_ASYNC,
1652 __compact_pgdat(pgdat, &cc);
1655 static void compact_node(int nid)
1657 struct compact_control cc = {
1659 .mode = MIGRATE_SYNC,
1660 .ignore_skip_hint = true,
1663 __compact_pgdat(NODE_DATA(nid), &cc);
1666 /* Compact all nodes in the system */
1667 static void compact_nodes(void)
1671 /* Flush pending updates to the LRU lists */
1672 lru_add_drain_all();
1674 for_each_online_node(nid)
1678 /* The written value is actually unused, all memory is compacted */
1679 int sysctl_compact_memory;
1681 /* This is the entry point for compacting all nodes via /proc/sys/vm */
1682 int sysctl_compaction_handler(struct ctl_table *table, int write,
1683 void __user *buffer, size_t *length, loff_t *ppos)
1691 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1692 void __user *buffer, size_t *length, loff_t *ppos)
1694 proc_dointvec_minmax(table, write, buffer, length, ppos);
1699 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1700 static ssize_t sysfs_compact_node(struct device *dev,
1701 struct device_attribute *attr,
1702 const char *buf, size_t count)
1706 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1707 /* Flush pending updates to the LRU lists */
1708 lru_add_drain_all();
1715 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1717 int compaction_register_node(struct node *node)
1719 return device_create_file(&node->dev, &dev_attr_compact);
1722 void compaction_unregister_node(struct node *node)
1724 return device_remove_file(&node->dev, &dev_attr_compact);
1726 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1728 #endif /* CONFIG_COMPACTION */