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>
21 #ifdef CONFIG_COMPACTION
22 static inline void count_compact_event(enum vm_event_item item)
27 static inline void count_compact_events(enum vm_event_item item, long delta)
29 count_vm_events(item, delta);
32 #define count_compact_event(item) do { } while (0)
33 #define count_compact_events(item, delta) do { } while (0)
36 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
37 #ifdef CONFIG_TRACEPOINTS
38 static const char *const compaction_status_string[] = {
49 #define CREATE_TRACE_POINTS
50 #include <trace/events/compaction.h>
52 static unsigned long release_freepages(struct list_head *freelist)
54 struct page *page, *next;
55 unsigned long high_pfn = 0;
57 list_for_each_entry_safe(page, next, freelist, lru) {
58 unsigned long pfn = page_to_pfn(page);
68 static void map_pages(struct list_head *list)
72 list_for_each_entry(page, list, lru) {
73 arch_alloc_page(page, 0);
74 kernel_map_pages(page, 1, 1);
78 static inline bool migrate_async_suitable(int migratetype)
80 return is_migrate_cma(migratetype) || migratetype == MIGRATE_MOVABLE;
84 * Check that the whole (or subset of) a pageblock given by the interval of
85 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
86 * with the migration of free compaction scanner. The scanners then need to
87 * use only pfn_valid_within() check for arches that allow holes within
90 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
92 * It's possible on some configurations to have a setup like node0 node1 node0
93 * i.e. it's possible that all pages within a zones range of pages do not
94 * belong to a single zone. We assume that a border between node0 and node1
95 * can occur within a single pageblock, but not a node0 node1 node0
96 * interleaving within a single pageblock. It is therefore sufficient to check
97 * the first and last page of a pageblock and avoid checking each individual
98 * page in a pageblock.
100 static struct page *pageblock_pfn_to_page(unsigned long start_pfn,
101 unsigned long end_pfn, struct zone *zone)
103 struct page *start_page;
104 struct page *end_page;
106 /* end_pfn is one past the range we are checking */
109 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
112 start_page = pfn_to_page(start_pfn);
114 if (page_zone(start_page) != zone)
117 end_page = pfn_to_page(end_pfn);
119 /* This gives a shorter code than deriving page_zone(end_page) */
120 if (page_zone_id(start_page) != page_zone_id(end_page))
126 #ifdef CONFIG_COMPACTION
128 /* Do not skip compaction more than 64 times */
129 #define COMPACT_MAX_DEFER_SHIFT 6
132 * Compaction is deferred when compaction fails to result in a page
133 * allocation success. 1 << compact_defer_limit compactions are skipped up
134 * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
136 void defer_compaction(struct zone *zone, int order)
138 zone->compact_considered = 0;
139 zone->compact_defer_shift++;
141 if (order < zone->compact_order_failed)
142 zone->compact_order_failed = order;
144 if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
145 zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
147 trace_mm_compaction_defer_compaction(zone, order);
150 /* Returns true if compaction should be skipped this time */
151 bool compaction_deferred(struct zone *zone, int order)
153 unsigned long defer_limit = 1UL << zone->compact_defer_shift;
155 if (order < zone->compact_order_failed)
158 /* Avoid possible overflow */
159 if (++zone->compact_considered > defer_limit)
160 zone->compact_considered = defer_limit;
162 if (zone->compact_considered >= defer_limit)
165 trace_mm_compaction_deferred(zone, order);
171 * Update defer tracking counters after successful compaction of given order,
172 * which means an allocation either succeeded (alloc_success == true) or is
173 * expected to succeed.
175 void compaction_defer_reset(struct zone *zone, int order,
179 zone->compact_considered = 0;
180 zone->compact_defer_shift = 0;
182 if (order >= zone->compact_order_failed)
183 zone->compact_order_failed = order + 1;
185 trace_mm_compaction_defer_reset(zone, order);
188 /* Returns true if restarting compaction after many failures */
189 bool compaction_restarting(struct zone *zone, int order)
191 if (order < zone->compact_order_failed)
194 return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
195 zone->compact_considered >= 1UL << zone->compact_defer_shift;
198 /* Returns true if the pageblock should be scanned for pages to isolate. */
199 static inline bool isolation_suitable(struct compact_control *cc,
202 if (cc->ignore_skip_hint)
205 return !get_pageblock_skip(page);
209 * This function is called to clear all cached information on pageblocks that
210 * should be skipped for page isolation when the migrate and free page scanner
213 static void __reset_isolation_suitable(struct zone *zone)
215 unsigned long start_pfn = zone->zone_start_pfn;
216 unsigned long end_pfn = zone_end_pfn(zone);
219 zone->compact_cached_migrate_pfn[0] = start_pfn;
220 zone->compact_cached_migrate_pfn[1] = start_pfn;
221 zone->compact_cached_free_pfn = end_pfn;
222 zone->compact_blockskip_flush = false;
224 /* Walk the zone and mark every pageblock as suitable for isolation */
225 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
233 page = pfn_to_page(pfn);
234 if (zone != page_zone(page))
237 clear_pageblock_skip(page);
241 void reset_isolation_suitable(pg_data_t *pgdat)
245 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
246 struct zone *zone = &pgdat->node_zones[zoneid];
247 if (!populated_zone(zone))
250 /* Only flush if a full compaction finished recently */
251 if (zone->compact_blockskip_flush)
252 __reset_isolation_suitable(zone);
257 * If no pages were isolated then mark this pageblock to be skipped in the
258 * future. The information is later cleared by __reset_isolation_suitable().
260 static void update_pageblock_skip(struct compact_control *cc,
261 struct page *page, unsigned long nr_isolated,
262 bool migrate_scanner)
264 struct zone *zone = cc->zone;
267 if (cc->ignore_skip_hint)
276 set_pageblock_skip(page);
278 pfn = page_to_pfn(page);
280 /* Update where async and sync compaction should restart */
281 if (migrate_scanner) {
282 if (pfn > zone->compact_cached_migrate_pfn[0])
283 zone->compact_cached_migrate_pfn[0] = pfn;
284 if (cc->mode != MIGRATE_ASYNC &&
285 pfn > zone->compact_cached_migrate_pfn[1])
286 zone->compact_cached_migrate_pfn[1] = pfn;
288 if (pfn < zone->compact_cached_free_pfn)
289 zone->compact_cached_free_pfn = pfn;
293 static inline bool isolation_suitable(struct compact_control *cc,
299 static void update_pageblock_skip(struct compact_control *cc,
300 struct page *page, unsigned long nr_isolated,
301 bool migrate_scanner)
304 #endif /* CONFIG_COMPACTION */
307 * Compaction requires the taking of some coarse locks that are potentially
308 * very heavily contended. For async compaction, back out if the lock cannot
309 * be taken immediately. For sync compaction, spin on the lock if needed.
311 * Returns true if the lock is held
312 * Returns false if the lock is not held and compaction should abort
314 static bool compact_trylock_irqsave(spinlock_t *lock, unsigned long *flags,
315 struct compact_control *cc)
317 if (cc->mode == MIGRATE_ASYNC) {
318 if (!spin_trylock_irqsave(lock, *flags)) {
319 cc->contended = COMPACT_CONTENDED_LOCK;
323 spin_lock_irqsave(lock, *flags);
330 * Compaction requires the taking of some coarse locks that are potentially
331 * very heavily contended. The lock should be periodically unlocked to avoid
332 * having disabled IRQs for a long time, even when there is nobody waiting on
333 * the lock. It might also be that allowing the IRQs will result in
334 * need_resched() becoming true. If scheduling is needed, async compaction
335 * aborts. Sync compaction schedules.
336 * Either compaction type will also abort if a fatal signal is pending.
337 * In either case if the lock was locked, it is dropped and not regained.
339 * Returns true if compaction should abort due to fatal signal pending, or
340 * async compaction due to need_resched()
341 * Returns false when compaction can continue (sync compaction might have
344 static bool compact_unlock_should_abort(spinlock_t *lock,
345 unsigned long flags, bool *locked, struct compact_control *cc)
348 spin_unlock_irqrestore(lock, flags);
352 if (fatal_signal_pending(current)) {
353 cc->contended = COMPACT_CONTENDED_SCHED;
357 if (need_resched()) {
358 if (cc->mode == MIGRATE_ASYNC) {
359 cc->contended = COMPACT_CONTENDED_SCHED;
369 * Aside from avoiding lock contention, compaction also periodically checks
370 * need_resched() and either schedules in sync compaction or aborts async
371 * compaction. This is similar to what compact_unlock_should_abort() does, but
372 * is used where no lock is concerned.
374 * Returns false when no scheduling was needed, or sync compaction scheduled.
375 * Returns true when async compaction should abort.
377 static inline bool compact_should_abort(struct compact_control *cc)
379 /* async compaction aborts if contended */
380 if (need_resched()) {
381 if (cc->mode == MIGRATE_ASYNC) {
382 cc->contended = COMPACT_CONTENDED_SCHED;
392 /* Returns true if the page is within a block suitable for migration to */
393 static bool suitable_migration_target(struct page *page)
395 /* If the page is a large free page, then disallow migration */
396 if (PageBuddy(page)) {
398 * We are checking page_order without zone->lock taken. But
399 * the only small danger is that we skip a potentially suitable
400 * pageblock, so it's not worth to check order for valid range.
402 if (page_order_unsafe(page) >= pageblock_order)
406 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
407 if (migrate_async_suitable(get_pageblock_migratetype(page)))
410 /* Otherwise skip the block */
415 * Isolate free pages onto a private freelist. If @strict is true, will abort
416 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
417 * (even though it may still end up isolating some pages).
419 static unsigned long isolate_freepages_block(struct compact_control *cc,
420 unsigned long *start_pfn,
421 unsigned long end_pfn,
422 struct list_head *freelist,
425 int nr_scanned = 0, total_isolated = 0;
426 struct page *cursor, *valid_page = NULL;
427 unsigned long flags = 0;
429 unsigned long blockpfn = *start_pfn;
431 cursor = pfn_to_page(blockpfn);
433 /* Isolate free pages. */
434 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
436 struct page *page = cursor;
439 * Periodically drop the lock (if held) regardless of its
440 * contention, to give chance to IRQs. Abort if fatal signal
441 * pending or async compaction detects need_resched()
443 if (!(blockpfn % SWAP_CLUSTER_MAX)
444 && compact_unlock_should_abort(&cc->zone->lock, flags,
449 if (!pfn_valid_within(blockpfn))
454 if (!PageBuddy(page))
458 * If we already hold the lock, we can skip some rechecking.
459 * Note that if we hold the lock now, checked_pageblock was
460 * already set in some previous iteration (or strict is true),
461 * so it is correct to skip the suitable migration target
466 * The zone lock must be held to isolate freepages.
467 * Unfortunately this is a very coarse lock and can be
468 * heavily contended if there are parallel allocations
469 * or parallel compactions. For async compaction do not
470 * spin on the lock and we acquire the lock as late as
473 locked = compact_trylock_irqsave(&cc->zone->lock,
478 /* Recheck this is a buddy page under lock */
479 if (!PageBuddy(page))
483 /* Found a free page, break it into order-0 pages */
484 isolated = split_free_page(page);
485 total_isolated += isolated;
486 for (i = 0; i < isolated; i++) {
487 list_add(&page->lru, freelist);
491 /* If a page was split, advance to the end of it */
493 blockpfn += isolated - 1;
494 cursor += isolated - 1;
506 trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
507 nr_scanned, total_isolated);
509 /* Record how far we have got within the block */
510 *start_pfn = blockpfn;
513 * If strict isolation is requested by CMA then check that all the
514 * pages requested were isolated. If there were any failures, 0 is
515 * returned and CMA will fail.
517 if (strict && blockpfn < end_pfn)
521 spin_unlock_irqrestore(&cc->zone->lock, flags);
523 /* Update the pageblock-skip if the whole pageblock was scanned */
524 if (blockpfn == end_pfn)
525 update_pageblock_skip(cc, valid_page, total_isolated, false);
527 count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
529 count_compact_events(COMPACTISOLATED, total_isolated);
530 return total_isolated;
534 * isolate_freepages_range() - isolate free pages.
535 * @start_pfn: The first PFN to start isolating.
536 * @end_pfn: The one-past-last PFN.
538 * Non-free pages, invalid PFNs, or zone boundaries within the
539 * [start_pfn, end_pfn) range are considered errors, cause function to
540 * undo its actions and return zero.
542 * Otherwise, function returns one-past-the-last PFN of isolated page
543 * (which may be greater then end_pfn if end fell in a middle of
547 isolate_freepages_range(struct compact_control *cc,
548 unsigned long start_pfn, unsigned long end_pfn)
550 unsigned long isolated, pfn, block_end_pfn;
554 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
556 for (; pfn < end_pfn; pfn += isolated,
557 block_end_pfn += pageblock_nr_pages) {
558 /* Protect pfn from changing by isolate_freepages_block */
559 unsigned long isolate_start_pfn = pfn;
561 block_end_pfn = min(block_end_pfn, end_pfn);
564 * pfn could pass the block_end_pfn if isolated freepage
565 * is more than pageblock order. In this case, we adjust
566 * scanning range to right one.
568 if (pfn >= block_end_pfn) {
569 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
570 block_end_pfn = min(block_end_pfn, end_pfn);
573 if (!pageblock_pfn_to_page(pfn, block_end_pfn, cc->zone))
576 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
577 block_end_pfn, &freelist, true);
580 * In strict mode, isolate_freepages_block() returns 0 if
581 * there are any holes in the block (ie. invalid PFNs or
588 * If we managed to isolate pages, it is always (1 << n) *
589 * pageblock_nr_pages for some non-negative n. (Max order
590 * page may span two pageblocks).
594 /* split_free_page does not map the pages */
595 map_pages(&freelist);
598 /* Loop terminated early, cleanup. */
599 release_freepages(&freelist);
603 /* We don't use freelists for anything. */
607 /* Update the number of anon and file isolated pages in the zone */
608 static void acct_isolated(struct zone *zone, struct compact_control *cc)
611 unsigned int count[2] = { 0, };
613 if (list_empty(&cc->migratepages))
616 list_for_each_entry(page, &cc->migratepages, lru)
617 count[!!page_is_file_cache(page)]++;
619 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
620 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
623 /* Similar to reclaim, but different enough that they don't share logic */
624 static bool too_many_isolated(struct zone *zone)
626 unsigned long active, inactive, isolated;
628 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
629 zone_page_state(zone, NR_INACTIVE_ANON);
630 active = zone_page_state(zone, NR_ACTIVE_FILE) +
631 zone_page_state(zone, NR_ACTIVE_ANON);
632 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
633 zone_page_state(zone, NR_ISOLATED_ANON);
635 return isolated > (inactive + active) / 2;
639 * isolate_migratepages_block() - isolate all migrate-able pages within
641 * @cc: Compaction control structure.
642 * @low_pfn: The first PFN to isolate
643 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
644 * @isolate_mode: Isolation mode to be used.
646 * Isolate all pages that can be migrated from the range specified by
647 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
648 * Returns zero if there is a fatal signal pending, otherwise PFN of the
649 * first page that was not scanned (which may be both less, equal to or more
652 * The pages are isolated on cc->migratepages list (not required to be empty),
653 * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
654 * is neither read nor updated.
657 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
658 unsigned long end_pfn, isolate_mode_t isolate_mode)
660 struct zone *zone = cc->zone;
661 unsigned long nr_scanned = 0, nr_isolated = 0;
662 struct list_head *migratelist = &cc->migratepages;
663 struct lruvec *lruvec;
664 unsigned long flags = 0;
666 struct page *page = NULL, *valid_page = NULL;
667 unsigned long start_pfn = low_pfn;
670 * Ensure that there are not too many pages isolated from the LRU
671 * list by either parallel reclaimers or compaction. If there are,
672 * delay for some time until fewer pages are isolated
674 while (unlikely(too_many_isolated(zone))) {
675 /* async migration should just abort */
676 if (cc->mode == MIGRATE_ASYNC)
679 congestion_wait(BLK_RW_ASYNC, HZ/10);
681 if (fatal_signal_pending(current))
685 if (compact_should_abort(cc))
688 /* Time to isolate some pages for migration */
689 for (; low_pfn < end_pfn; low_pfn++) {
691 * Periodically drop the lock (if held) regardless of its
692 * contention, to give chance to IRQs. Abort async compaction
695 if (!(low_pfn % SWAP_CLUSTER_MAX)
696 && compact_unlock_should_abort(&zone->lru_lock, flags,
700 if (!pfn_valid_within(low_pfn))
704 page = pfn_to_page(low_pfn);
710 * Skip if free. We read page order here without zone lock
711 * which is generally unsafe, but the race window is small and
712 * the worst thing that can happen is that we skip some
713 * potential isolation targets.
715 if (PageBuddy(page)) {
716 unsigned long freepage_order = page_order_unsafe(page);
719 * Without lock, we cannot be sure that what we got is
720 * a valid page order. Consider only values in the
721 * valid order range to prevent low_pfn overflow.
723 if (freepage_order > 0 && freepage_order < MAX_ORDER)
724 low_pfn += (1UL << freepage_order) - 1;
729 * Check may be lockless but that's ok as we recheck later.
730 * It's possible to migrate LRU pages and balloon pages
731 * Skip any other type of page
733 if (!PageLRU(page)) {
734 if (unlikely(balloon_page_movable(page))) {
735 if (balloon_page_isolate(page)) {
736 /* Successfully isolated */
737 goto isolate_success;
744 * PageLRU is set. lru_lock normally excludes isolation
745 * splitting and collapsing (collapsing has already happened
746 * if PageLRU is set) but the lock is not necessarily taken
747 * here and it is wasteful to take it just to check transhuge.
748 * Check TransHuge without lock and skip the whole pageblock if
749 * it's either a transhuge or hugetlbfs page, as calling
750 * compound_order() without preventing THP from splitting the
751 * page underneath us may return surprising results.
753 if (PageTransHuge(page)) {
755 low_pfn = ALIGN(low_pfn + 1,
756 pageblock_nr_pages) - 1;
758 low_pfn += (1 << compound_order(page)) - 1;
764 * Migration will fail if an anonymous page is pinned in memory,
765 * so avoid taking lru_lock and isolating it unnecessarily in an
766 * admittedly racy check.
768 if (!page_mapping(page) &&
769 page_count(page) > page_mapcount(page))
772 /* If we already hold the lock, we can skip some rechecking */
774 locked = compact_trylock_irqsave(&zone->lru_lock,
779 /* Recheck PageLRU and PageTransHuge under lock */
782 if (PageTransHuge(page)) {
783 low_pfn += (1 << compound_order(page)) - 1;
788 lruvec = mem_cgroup_page_lruvec(page, zone);
790 /* Try isolate the page */
791 if (__isolate_lru_page(page, isolate_mode) != 0)
794 VM_BUG_ON_PAGE(PageTransCompound(page), page);
796 /* Successfully isolated */
797 del_page_from_lru_list(page, lruvec, page_lru(page));
800 list_add(&page->lru, migratelist);
801 cc->nr_migratepages++;
804 /* Avoid isolating too much */
805 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
812 * The PageBuddy() check could have potentially brought us outside
813 * the range to be scanned.
815 if (unlikely(low_pfn > end_pfn))
819 spin_unlock_irqrestore(&zone->lru_lock, flags);
822 * Update the pageblock-skip information and cached scanner pfn,
823 * if the whole pageblock was scanned without isolating any page.
825 if (low_pfn == end_pfn)
826 update_pageblock_skip(cc, valid_page, nr_isolated, true);
828 trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
829 nr_scanned, nr_isolated);
831 count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
833 count_compact_events(COMPACTISOLATED, nr_isolated);
839 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
840 * @cc: Compaction control structure.
841 * @start_pfn: The first PFN to start isolating.
842 * @end_pfn: The one-past-last PFN.
844 * Returns zero if isolation fails fatally due to e.g. pending signal.
845 * Otherwise, function returns one-past-the-last PFN of isolated page
846 * (which may be greater than end_pfn if end fell in a middle of a THP page).
849 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
850 unsigned long end_pfn)
852 unsigned long pfn, block_end_pfn;
854 /* Scan block by block. First and last block may be incomplete */
856 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
858 for (; pfn < end_pfn; pfn = block_end_pfn,
859 block_end_pfn += pageblock_nr_pages) {
861 block_end_pfn = min(block_end_pfn, end_pfn);
863 if (!pageblock_pfn_to_page(pfn, block_end_pfn, cc->zone))
866 pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
867 ISOLATE_UNEVICTABLE);
870 * In case of fatal failure, release everything that might
871 * have been isolated in the previous iteration, and signal
872 * the failure back to caller.
875 putback_movable_pages(&cc->migratepages);
876 cc->nr_migratepages = 0;
880 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
883 acct_isolated(cc->zone, cc);
888 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
889 #ifdef CONFIG_COMPACTION
891 * Based on information in the current compact_control, find blocks
892 * suitable for isolating free pages from and then isolate them.
894 static void isolate_freepages(struct compact_control *cc)
896 struct zone *zone = cc->zone;
898 unsigned long block_start_pfn; /* start of current pageblock */
899 unsigned long isolate_start_pfn; /* exact pfn we start at */
900 unsigned long block_end_pfn; /* end of current pageblock */
901 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
902 int nr_freepages = cc->nr_freepages;
903 struct list_head *freelist = &cc->freepages;
906 * Initialise the free scanner. The starting point is where we last
907 * successfully isolated from, zone-cached value, or the end of the
908 * zone when isolating for the first time. For looping we also need
909 * this pfn aligned down to the pageblock boundary, because we do
910 * block_start_pfn -= pageblock_nr_pages in the for loop.
911 * For ending point, take care when isolating in last pageblock of a
912 * a zone which ends in the middle of a pageblock.
913 * The low boundary is the end of the pageblock the migration scanner
916 isolate_start_pfn = cc->free_pfn;
917 block_start_pfn = cc->free_pfn & ~(pageblock_nr_pages-1);
918 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
920 low_pfn = ALIGN(cc->migrate_pfn + 1, pageblock_nr_pages);
923 * Isolate free pages until enough are available to migrate the
924 * pages on cc->migratepages. We stop searching if the migrate
925 * and free page scanners meet or enough free pages are isolated.
927 for (; block_start_pfn >= low_pfn && cc->nr_migratepages > nr_freepages;
928 block_end_pfn = block_start_pfn,
929 block_start_pfn -= pageblock_nr_pages,
930 isolate_start_pfn = block_start_pfn) {
931 unsigned long isolated;
934 * This can iterate a massively long zone without finding any
935 * suitable migration targets, so periodically check if we need
936 * to schedule, or even abort async compaction.
938 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
939 && compact_should_abort(cc))
942 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
947 /* Check the block is suitable for migration */
948 if (!suitable_migration_target(page))
951 /* If isolation recently failed, do not retry */
952 if (!isolation_suitable(cc, page))
955 /* Found a block suitable for isolating free pages from. */
956 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
957 block_end_pfn, freelist, false);
958 nr_freepages += isolated;
961 * Remember where the free scanner should restart next time,
962 * which is where isolate_freepages_block() left off.
963 * But if it scanned the whole pageblock, isolate_start_pfn
964 * now points at block_end_pfn, which is the start of the next
966 * In that case we will however want to restart at the start
967 * of the previous pageblock.
969 cc->free_pfn = (isolate_start_pfn < block_end_pfn) ?
971 block_start_pfn - pageblock_nr_pages;
974 * isolate_freepages_block() might have aborted due to async
975 * compaction being contended
981 /* split_free_page does not map the pages */
985 * If we crossed the migrate scanner, we want to keep it that way
986 * so that compact_finished() may detect this
988 if (block_start_pfn < low_pfn)
989 cc->free_pfn = cc->migrate_pfn;
991 cc->nr_freepages = nr_freepages;
995 * This is a migrate-callback that "allocates" freepages by taking pages
996 * from the isolated freelists in the block we are migrating to.
998 static struct page *compaction_alloc(struct page *migratepage,
1002 struct compact_control *cc = (struct compact_control *)data;
1003 struct page *freepage;
1006 * Isolate free pages if necessary, and if we are not aborting due to
1009 if (list_empty(&cc->freepages)) {
1011 isolate_freepages(cc);
1013 if (list_empty(&cc->freepages))
1017 freepage = list_entry(cc->freepages.next, struct page, lru);
1018 list_del(&freepage->lru);
1025 * This is a migrate-callback that "frees" freepages back to the isolated
1026 * freelist. All pages on the freelist are from the same zone, so there is no
1027 * special handling needed for NUMA.
1029 static void compaction_free(struct page *page, unsigned long data)
1031 struct compact_control *cc = (struct compact_control *)data;
1033 list_add(&page->lru, &cc->freepages);
1037 /* possible outcome of isolate_migratepages */
1039 ISOLATE_ABORT, /* Abort compaction now */
1040 ISOLATE_NONE, /* No pages isolated, continue scanning */
1041 ISOLATE_SUCCESS, /* Pages isolated, migrate */
1042 } isolate_migrate_t;
1045 * Isolate all pages that can be migrated from the first suitable block,
1046 * starting at the block pointed to by the migrate scanner pfn within
1049 static isolate_migrate_t isolate_migratepages(struct zone *zone,
1050 struct compact_control *cc)
1052 unsigned long low_pfn, end_pfn;
1054 const isolate_mode_t isolate_mode =
1055 (cc->mode == MIGRATE_ASYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1058 * Start at where we last stopped, or beginning of the zone as
1059 * initialized by compact_zone()
1061 low_pfn = cc->migrate_pfn;
1063 /* Only scan within a pageblock boundary */
1064 end_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages);
1067 * Iterate over whole pageblocks until we find the first suitable.
1068 * Do not cross the free scanner.
1070 for (; end_pfn <= cc->free_pfn;
1071 low_pfn = end_pfn, end_pfn += pageblock_nr_pages) {
1074 * This can potentially iterate a massively long zone with
1075 * many pageblocks unsuitable, so periodically check if we
1076 * need to schedule, or even abort async compaction.
1078 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1079 && compact_should_abort(cc))
1082 page = pageblock_pfn_to_page(low_pfn, end_pfn, zone);
1086 /* If isolation recently failed, do not retry */
1087 if (!isolation_suitable(cc, page))
1091 * For async compaction, also only scan in MOVABLE blocks.
1092 * Async compaction is optimistic to see if the minimum amount
1093 * of work satisfies the allocation.
1095 if (cc->mode == MIGRATE_ASYNC &&
1096 !migrate_async_suitable(get_pageblock_migratetype(page)))
1099 /* Perform the isolation */
1100 low_pfn = isolate_migratepages_block(cc, low_pfn, end_pfn,
1103 if (!low_pfn || cc->contended)
1104 return ISOLATE_ABORT;
1107 * Either we isolated something and proceed with migration. Or
1108 * we failed and compact_zone should decide if we should
1114 acct_isolated(zone, cc);
1116 * Record where migration scanner will be restarted. If we end up in
1117 * the same pageblock as the free scanner, make the scanners fully
1118 * meet so that compact_finished() terminates compaction.
1120 cc->migrate_pfn = (end_pfn <= cc->free_pfn) ? low_pfn : cc->free_pfn;
1122 return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
1125 static int __compact_finished(struct zone *zone, struct compact_control *cc,
1126 const int migratetype)
1129 unsigned long watermark;
1131 if (cc->contended || fatal_signal_pending(current))
1132 return COMPACT_PARTIAL;
1134 /* Compaction run completes if the migrate and free scanner meet */
1135 if (cc->free_pfn <= cc->migrate_pfn) {
1136 /* Let the next compaction start anew. */
1137 zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
1138 zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
1139 zone->compact_cached_free_pfn = zone_end_pfn(zone);
1142 * Mark that the PG_migrate_skip information should be cleared
1143 * by kswapd when it goes to sleep. kswapd does not set the
1144 * flag itself as the decision to be clear should be directly
1145 * based on an allocation request.
1147 if (!current_is_kswapd())
1148 zone->compact_blockskip_flush = true;
1150 return COMPACT_COMPLETE;
1154 * order == -1 is expected when compacting via
1155 * /proc/sys/vm/compact_memory
1157 if (cc->order == -1)
1158 return COMPACT_CONTINUE;
1160 /* Compaction run is not finished if the watermark is not met */
1161 watermark = low_wmark_pages(zone);
1163 if (!zone_watermark_ok(zone, cc->order, watermark, cc->classzone_idx,
1165 return COMPACT_CONTINUE;
1167 /* Direct compactor: Is a suitable page free? */
1168 for (order = cc->order; order < MAX_ORDER; order++) {
1169 struct free_area *area = &zone->free_area[order];
1171 /* Job done if page is free of the right migratetype */
1172 if (!list_empty(&area->free_list[migratetype]))
1173 return COMPACT_PARTIAL;
1175 /* Job done if allocation would set block type */
1176 if (cc->order >= pageblock_order && area->nr_free)
1177 return COMPACT_PARTIAL;
1180 return COMPACT_NO_SUITABLE_PAGE;
1183 static int compact_finished(struct zone *zone, struct compact_control *cc,
1184 const int migratetype)
1188 ret = __compact_finished(zone, cc, migratetype);
1189 trace_mm_compaction_finished(zone, cc->order, ret);
1190 if (ret == COMPACT_NO_SUITABLE_PAGE)
1191 ret = COMPACT_CONTINUE;
1197 * compaction_suitable: Is this suitable to run compaction on this zone now?
1199 * COMPACT_SKIPPED - If there are too few free pages for compaction
1200 * COMPACT_PARTIAL - If the allocation would succeed without compaction
1201 * COMPACT_CONTINUE - If compaction should run now
1203 static unsigned long __compaction_suitable(struct zone *zone, int order,
1204 int alloc_flags, int classzone_idx)
1207 unsigned long watermark;
1210 * order == -1 is expected when compacting via
1211 * /proc/sys/vm/compact_memory
1214 return COMPACT_CONTINUE;
1216 watermark = low_wmark_pages(zone);
1218 * If watermarks for high-order allocation are already met, there
1219 * should be no need for compaction at all.
1221 if (zone_watermark_ok(zone, order, watermark, classzone_idx,
1223 return COMPACT_PARTIAL;
1226 * Watermarks for order-0 must be met for compaction. Note the 2UL.
1227 * This is because during migration, copies of pages need to be
1228 * allocated and for a short time, the footprint is higher
1230 watermark += (2UL << order);
1231 if (!zone_watermark_ok(zone, 0, watermark, classzone_idx, alloc_flags))
1232 return COMPACT_SKIPPED;
1235 * fragmentation index determines if allocation failures are due to
1236 * low memory or external fragmentation
1238 * index of -1000 would imply allocations might succeed depending on
1239 * watermarks, but we already failed the high-order watermark check
1240 * index towards 0 implies failure is due to lack of memory
1241 * index towards 1000 implies failure is due to fragmentation
1243 * Only compact if a failure would be due to fragmentation.
1245 fragindex = fragmentation_index(zone, order);
1246 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
1247 return COMPACT_NOT_SUITABLE_ZONE;
1249 return COMPACT_CONTINUE;
1252 unsigned long compaction_suitable(struct zone *zone, int order,
1253 int alloc_flags, int classzone_idx)
1257 ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx);
1258 trace_mm_compaction_suitable(zone, order, ret);
1259 if (ret == COMPACT_NOT_SUITABLE_ZONE)
1260 ret = COMPACT_SKIPPED;
1265 static int compact_zone(struct zone *zone, struct compact_control *cc)
1268 unsigned long start_pfn = zone->zone_start_pfn;
1269 unsigned long end_pfn = zone_end_pfn(zone);
1270 const int migratetype = gfpflags_to_migratetype(cc->gfp_mask);
1271 const bool sync = cc->mode != MIGRATE_ASYNC;
1272 unsigned long last_migrated_pfn = 0;
1274 ret = compaction_suitable(zone, cc->order, cc->alloc_flags,
1277 case COMPACT_PARTIAL:
1278 case COMPACT_SKIPPED:
1279 /* Compaction is likely to fail */
1281 case COMPACT_CONTINUE:
1282 /* Fall through to compaction */
1287 * Clear pageblock skip if there were failures recently and compaction
1288 * is about to be retried after being deferred. kswapd does not do
1289 * this reset as it'll reset the cached information when going to sleep.
1291 if (compaction_restarting(zone, cc->order) && !current_is_kswapd())
1292 __reset_isolation_suitable(zone);
1295 * Setup to move all movable pages to the end of the zone. Used cached
1296 * information on where the scanners should start but check that it
1297 * is initialised by ensuring the values are within zone boundaries.
1299 cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync];
1300 cc->free_pfn = zone->compact_cached_free_pfn;
1301 if (cc->free_pfn < start_pfn || cc->free_pfn > end_pfn) {
1302 cc->free_pfn = end_pfn & ~(pageblock_nr_pages-1);
1303 zone->compact_cached_free_pfn = cc->free_pfn;
1305 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn > end_pfn) {
1306 cc->migrate_pfn = start_pfn;
1307 zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
1308 zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
1311 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
1312 cc->free_pfn, end_pfn, sync);
1314 migrate_prep_local();
1316 while ((ret = compact_finished(zone, cc, migratetype)) ==
1319 unsigned long isolate_start_pfn = cc->migrate_pfn;
1321 switch (isolate_migratepages(zone, cc)) {
1323 ret = COMPACT_PARTIAL;
1324 putback_movable_pages(&cc->migratepages);
1325 cc->nr_migratepages = 0;
1329 * We haven't isolated and migrated anything, but
1330 * there might still be unflushed migrations from
1331 * previous cc->order aligned block.
1334 case ISOLATE_SUCCESS:
1338 err = migrate_pages(&cc->migratepages, compaction_alloc,
1339 compaction_free, (unsigned long)cc, cc->mode,
1342 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
1345 /* All pages were either migrated or will be released */
1346 cc->nr_migratepages = 0;
1348 putback_movable_pages(&cc->migratepages);
1350 * migrate_pages() may return -ENOMEM when scanners meet
1351 * and we want compact_finished() to detect it
1353 if (err == -ENOMEM && cc->free_pfn > cc->migrate_pfn) {
1354 ret = COMPACT_PARTIAL;
1360 * Record where we could have freed pages by migration and not
1361 * yet flushed them to buddy allocator. We use the pfn that
1362 * isolate_migratepages() started from in this loop iteration
1363 * - this is the lowest page that could have been isolated and
1364 * then freed by migration.
1366 if (!last_migrated_pfn)
1367 last_migrated_pfn = isolate_start_pfn;
1371 * Has the migration scanner moved away from the previous
1372 * cc->order aligned block where we migrated from? If yes,
1373 * flush the pages that were freed, so that they can merge and
1374 * compact_finished() can detect immediately if allocation
1377 if (cc->order > 0 && last_migrated_pfn) {
1379 unsigned long current_block_start =
1380 cc->migrate_pfn & ~((1UL << cc->order) - 1);
1382 if (last_migrated_pfn < current_block_start) {
1384 lru_add_drain_cpu(cpu);
1385 drain_local_pages(zone);
1387 /* No more flushing until we migrate again */
1388 last_migrated_pfn = 0;
1396 * Release free pages and update where the free scanner should restart,
1397 * so we don't leave any returned pages behind in the next attempt.
1399 if (cc->nr_freepages > 0) {
1400 unsigned long free_pfn = release_freepages(&cc->freepages);
1402 cc->nr_freepages = 0;
1403 VM_BUG_ON(free_pfn == 0);
1404 /* The cached pfn is always the first in a pageblock */
1405 free_pfn &= ~(pageblock_nr_pages-1);
1407 * Only go back, not forward. The cached pfn might have been
1408 * already reset to zone end in compact_finished()
1410 if (free_pfn > zone->compact_cached_free_pfn)
1411 zone->compact_cached_free_pfn = free_pfn;
1414 trace_mm_compaction_end(start_pfn, cc->migrate_pfn,
1415 cc->free_pfn, end_pfn, sync, ret);
1420 static unsigned long compact_zone_order(struct zone *zone, int order,
1421 gfp_t gfp_mask, enum migrate_mode mode, int *contended,
1422 int alloc_flags, int classzone_idx)
1425 struct compact_control cc = {
1427 .nr_migratepages = 0,
1429 .gfp_mask = gfp_mask,
1432 .alloc_flags = alloc_flags,
1433 .classzone_idx = classzone_idx,
1435 INIT_LIST_HEAD(&cc.freepages);
1436 INIT_LIST_HEAD(&cc.migratepages);
1438 ret = compact_zone(zone, &cc);
1440 VM_BUG_ON(!list_empty(&cc.freepages));
1441 VM_BUG_ON(!list_empty(&cc.migratepages));
1443 *contended = cc.contended;
1447 int sysctl_extfrag_threshold = 500;
1450 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1451 * @gfp_mask: The GFP mask of the current allocation
1452 * @order: The order of the current allocation
1453 * @alloc_flags: The allocation flags of the current allocation
1454 * @ac: The context of current allocation
1455 * @mode: The migration mode for async, sync light, or sync migration
1456 * @contended: Return value that determines if compaction was aborted due to
1457 * need_resched() or lock contention
1459 * This is the main entry point for direct page compaction.
1461 unsigned long try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
1462 int alloc_flags, const struct alloc_context *ac,
1463 enum migrate_mode mode, int *contended)
1465 int may_enter_fs = gfp_mask & __GFP_FS;
1466 int may_perform_io = gfp_mask & __GFP_IO;
1469 int rc = COMPACT_DEFERRED;
1470 int all_zones_contended = COMPACT_CONTENDED_LOCK; /* init for &= op */
1472 *contended = COMPACT_CONTENDED_NONE;
1474 /* Check if the GFP flags allow compaction */
1475 if (!order || !may_enter_fs || !may_perform_io)
1476 return COMPACT_SKIPPED;
1478 trace_mm_compaction_try_to_compact_pages(order, gfp_mask, mode);
1480 /* Compact each zone in the list */
1481 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1486 if (compaction_deferred(zone, order))
1489 status = compact_zone_order(zone, order, gfp_mask, mode,
1490 &zone_contended, alloc_flags,
1492 rc = max(status, rc);
1494 * It takes at least one zone that wasn't lock contended
1495 * to clear all_zones_contended.
1497 all_zones_contended &= zone_contended;
1499 /* If a normal allocation would succeed, stop compacting */
1500 if (zone_watermark_ok(zone, order, low_wmark_pages(zone),
1501 ac->classzone_idx, alloc_flags)) {
1503 * We think the allocation will succeed in this zone,
1504 * but it is not certain, hence the false. The caller
1505 * will repeat this with true if allocation indeed
1506 * succeeds in this zone.
1508 compaction_defer_reset(zone, order, false);
1510 * It is possible that async compaction aborted due to
1511 * need_resched() and the watermarks were ok thanks to
1512 * somebody else freeing memory. The allocation can
1513 * however still fail so we better signal the
1514 * need_resched() contention anyway (this will not
1515 * prevent the allocation attempt).
1517 if (zone_contended == COMPACT_CONTENDED_SCHED)
1518 *contended = COMPACT_CONTENDED_SCHED;
1523 if (mode != MIGRATE_ASYNC && status == COMPACT_COMPLETE) {
1525 * We think that allocation won't succeed in this zone
1526 * so we defer compaction there. If it ends up
1527 * succeeding after all, it will be reset.
1529 defer_compaction(zone, order);
1533 * We might have stopped compacting due to need_resched() in
1534 * async compaction, or due to a fatal signal detected. In that
1535 * case do not try further zones and signal need_resched()
1538 if ((zone_contended == COMPACT_CONTENDED_SCHED)
1539 || fatal_signal_pending(current)) {
1540 *contended = COMPACT_CONTENDED_SCHED;
1547 * We might not have tried all the zones, so be conservative
1548 * and assume they are not all lock contended.
1550 all_zones_contended = 0;
1555 * If at least one zone wasn't deferred or skipped, we report if all
1556 * zones that were tried were lock contended.
1558 if (rc > COMPACT_SKIPPED && all_zones_contended)
1559 *contended = COMPACT_CONTENDED_LOCK;
1565 /* Compact all zones within a node */
1566 static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
1571 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1573 zone = &pgdat->node_zones[zoneid];
1574 if (!populated_zone(zone))
1577 cc->nr_freepages = 0;
1578 cc->nr_migratepages = 0;
1580 INIT_LIST_HEAD(&cc->freepages);
1581 INIT_LIST_HEAD(&cc->migratepages);
1583 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
1584 compact_zone(zone, cc);
1586 if (cc->order > 0) {
1587 if (zone_watermark_ok(zone, cc->order,
1588 low_wmark_pages(zone), 0, 0))
1589 compaction_defer_reset(zone, cc->order, false);
1592 VM_BUG_ON(!list_empty(&cc->freepages));
1593 VM_BUG_ON(!list_empty(&cc->migratepages));
1597 void compact_pgdat(pg_data_t *pgdat, int order)
1599 struct compact_control cc = {
1601 .mode = MIGRATE_ASYNC,
1607 __compact_pgdat(pgdat, &cc);
1610 static void compact_node(int nid)
1612 struct compact_control cc = {
1614 .mode = MIGRATE_SYNC,
1615 .ignore_skip_hint = true,
1618 __compact_pgdat(NODE_DATA(nid), &cc);
1621 /* Compact all nodes in the system */
1622 static void compact_nodes(void)
1626 /* Flush pending updates to the LRU lists */
1627 lru_add_drain_all();
1629 for_each_online_node(nid)
1633 /* The written value is actually unused, all memory is compacted */
1634 int sysctl_compact_memory;
1636 /* This is the entry point for compacting all nodes via /proc/sys/vm */
1637 int sysctl_compaction_handler(struct ctl_table *table, int write,
1638 void __user *buffer, size_t *length, loff_t *ppos)
1646 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1647 void __user *buffer, size_t *length, loff_t *ppos)
1649 proc_dointvec_minmax(table, write, buffer, length, ppos);
1654 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1655 static ssize_t sysfs_compact_node(struct device *dev,
1656 struct device_attribute *attr,
1657 const char *buf, size_t count)
1661 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1662 /* Flush pending updates to the LRU lists */
1663 lru_add_drain_all();
1670 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1672 int compaction_register_node(struct node *node)
1674 return device_create_file(&node->dev, &dev_attr_compact);
1677 void compaction_unregister_node(struct node *node)
1679 return device_remove_file(&node->dev, &dev_attr_compact);
1681 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1683 #endif /* CONFIG_COMPACTION */