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
38 #define CREATE_TRACE_POINTS
39 #include <trace/events/compaction.h>
41 static unsigned long release_freepages(struct list_head *freelist)
43 struct page *page, *next;
44 unsigned long count = 0;
46 list_for_each_entry_safe(page, next, freelist, lru) {
55 static void map_pages(struct list_head *list)
59 list_for_each_entry(page, list, lru) {
60 arch_alloc_page(page, 0);
61 kernel_map_pages(page, 1, 1);
65 static inline bool migrate_async_suitable(int migratetype)
67 return is_migrate_cma(migratetype) || migratetype == MIGRATE_MOVABLE;
70 #ifdef CONFIG_COMPACTION
71 /* Returns true if the pageblock should be scanned for pages to isolate. */
72 static inline bool isolation_suitable(struct compact_control *cc,
75 if (cc->ignore_skip_hint)
78 return !get_pageblock_skip(page);
82 * This function is called to clear all cached information on pageblocks that
83 * should be skipped for page isolation when the migrate and free page scanner
86 static void __reset_isolation_suitable(struct zone *zone)
88 unsigned long start_pfn = zone->zone_start_pfn;
89 unsigned long end_pfn = zone_end_pfn(zone);
92 zone->compact_cached_migrate_pfn[0] = start_pfn;
93 zone->compact_cached_migrate_pfn[1] = start_pfn;
94 zone->compact_cached_free_pfn = end_pfn;
95 zone->compact_blockskip_flush = false;
97 /* Walk the zone and mark every pageblock as suitable for isolation */
98 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
106 page = pfn_to_page(pfn);
107 if (zone != page_zone(page))
110 clear_pageblock_skip(page);
114 void reset_isolation_suitable(pg_data_t *pgdat)
118 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
119 struct zone *zone = &pgdat->node_zones[zoneid];
120 if (!populated_zone(zone))
123 /* Only flush if a full compaction finished recently */
124 if (zone->compact_blockskip_flush)
125 __reset_isolation_suitable(zone);
130 * If no pages were isolated then mark this pageblock to be skipped in the
131 * future. The information is later cleared by __reset_isolation_suitable().
133 static void update_pageblock_skip(struct compact_control *cc,
134 struct page *page, unsigned long nr_isolated,
135 bool set_unsuitable, bool migrate_scanner)
137 struct zone *zone = cc->zone;
140 if (cc->ignore_skip_hint)
150 * Only skip pageblocks when all forms of compaction will be known to
151 * fail in the near future.
154 set_pageblock_skip(page);
156 pfn = page_to_pfn(page);
158 /* Update where async and sync compaction should restart */
159 if (migrate_scanner) {
160 if (cc->finished_update_migrate)
162 if (pfn > zone->compact_cached_migrate_pfn[0])
163 zone->compact_cached_migrate_pfn[0] = pfn;
164 if (cc->sync && pfn > zone->compact_cached_migrate_pfn[1])
165 zone->compact_cached_migrate_pfn[1] = pfn;
167 if (cc->finished_update_free)
169 if (pfn < zone->compact_cached_free_pfn)
170 zone->compact_cached_free_pfn = pfn;
174 static inline bool isolation_suitable(struct compact_control *cc,
180 static void update_pageblock_skip(struct compact_control *cc,
181 struct page *page, unsigned long nr_isolated,
182 bool set_unsuitable, bool migrate_scanner)
185 #endif /* CONFIG_COMPACTION */
187 static inline bool should_release_lock(spinlock_t *lock)
189 return need_resched() || spin_is_contended(lock);
193 * Compaction requires the taking of some coarse locks that are potentially
194 * very heavily contended. Check if the process needs to be scheduled or
195 * if the lock is contended. For async compaction, back out in the event
196 * if contention is severe. For sync compaction, schedule.
198 * Returns true if the lock is held.
199 * Returns false if the lock is released and compaction should abort
201 static bool compact_checklock_irqsave(spinlock_t *lock, unsigned long *flags,
202 bool locked, struct compact_control *cc)
204 if (should_release_lock(lock)) {
206 spin_unlock_irqrestore(lock, *flags);
210 /* async aborts if taking too long or contended */
212 cc->contended = true;
220 spin_lock_irqsave(lock, *flags);
224 /* Returns true if the page is within a block suitable for migration to */
225 static bool suitable_migration_target(struct page *page)
227 /* If the page is a large free page, then disallow migration */
228 if (PageBuddy(page) && page_order(page) >= pageblock_order)
231 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
232 if (migrate_async_suitable(get_pageblock_migratetype(page)))
235 /* Otherwise skip the block */
240 * Isolate free pages onto a private freelist. If @strict is true, will abort
241 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
242 * (even though it may still end up isolating some pages).
244 static unsigned long isolate_freepages_block(struct compact_control *cc,
245 unsigned long blockpfn,
246 unsigned long end_pfn,
247 struct list_head *freelist,
250 int nr_scanned = 0, total_isolated = 0;
251 struct page *cursor, *valid_page = NULL;
254 bool checked_pageblock = false;
256 cursor = pfn_to_page(blockpfn);
258 /* Isolate free pages. */
259 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
261 struct page *page = cursor;
264 if (!pfn_valid_within(blockpfn))
269 if (!PageBuddy(page))
273 * The zone lock must be held to isolate freepages.
274 * Unfortunately this is a very coarse lock and can be
275 * heavily contended if there are parallel allocations
276 * or parallel compactions. For async compaction do not
277 * spin on the lock and we acquire the lock as late as
280 locked = compact_checklock_irqsave(&cc->zone->lock, &flags,
285 /* Recheck this is a suitable migration target under lock */
286 if (!strict && !checked_pageblock) {
288 * We need to check suitability of pageblock only once
289 * and this isolate_freepages_block() is called with
290 * pageblock range, so just check once is sufficient.
292 checked_pageblock = true;
293 if (!suitable_migration_target(page))
297 /* Recheck this is a buddy page under lock */
298 if (!PageBuddy(page))
301 /* Found a free page, break it into order-0 pages */
302 isolated = split_free_page(page);
303 total_isolated += isolated;
304 for (i = 0; i < isolated; i++) {
305 list_add(&page->lru, freelist);
309 /* If a page was split, advance to the end of it */
311 blockpfn += isolated - 1;
312 cursor += isolated - 1;
324 trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
327 * If strict isolation is requested by CMA then check that all the
328 * pages requested were isolated. If there were any failures, 0 is
329 * returned and CMA will fail.
331 if (strict && blockpfn < end_pfn)
335 spin_unlock_irqrestore(&cc->zone->lock, flags);
337 /* Update the pageblock-skip if the whole pageblock was scanned */
338 if (blockpfn == end_pfn)
339 update_pageblock_skip(cc, valid_page, total_isolated, true,
342 count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
344 count_compact_events(COMPACTISOLATED, total_isolated);
345 return total_isolated;
349 * isolate_freepages_range() - isolate free pages.
350 * @start_pfn: The first PFN to start isolating.
351 * @end_pfn: The one-past-last PFN.
353 * Non-free pages, invalid PFNs, or zone boundaries within the
354 * [start_pfn, end_pfn) range are considered errors, cause function to
355 * undo its actions and return zero.
357 * Otherwise, function returns one-past-the-last PFN of isolated page
358 * (which may be greater then end_pfn if end fell in a middle of
362 isolate_freepages_range(struct compact_control *cc,
363 unsigned long start_pfn, unsigned long end_pfn)
365 unsigned long isolated, pfn, block_end_pfn;
368 for (pfn = start_pfn; pfn < end_pfn; pfn += isolated) {
369 if (!pfn_valid(pfn) || cc->zone != page_zone(pfn_to_page(pfn)))
373 * On subsequent iterations ALIGN() is actually not needed,
374 * but we keep it that we not to complicate the code.
376 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
377 block_end_pfn = min(block_end_pfn, end_pfn);
379 isolated = isolate_freepages_block(cc, pfn, block_end_pfn,
383 * In strict mode, isolate_freepages_block() returns 0 if
384 * there are any holes in the block (ie. invalid PFNs or
391 * If we managed to isolate pages, it is always (1 << n) *
392 * pageblock_nr_pages for some non-negative n. (Max order
393 * page may span two pageblocks).
397 /* split_free_page does not map the pages */
398 map_pages(&freelist);
401 /* Loop terminated early, cleanup. */
402 release_freepages(&freelist);
406 /* We don't use freelists for anything. */
410 /* Update the number of anon and file isolated pages in the zone */
411 static void acct_isolated(struct zone *zone, bool locked, struct compact_control *cc)
414 unsigned int count[2] = { 0, };
416 list_for_each_entry(page, &cc->migratepages, lru)
417 count[!!page_is_file_cache(page)]++;
419 /* If locked we can use the interrupt unsafe versions */
421 __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
422 __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
424 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
425 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
429 /* Similar to reclaim, but different enough that they don't share logic */
430 static bool too_many_isolated(struct zone *zone)
432 unsigned long active, inactive, isolated;
434 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
435 zone_page_state(zone, NR_INACTIVE_ANON);
436 active = zone_page_state(zone, NR_ACTIVE_FILE) +
437 zone_page_state(zone, NR_ACTIVE_ANON);
438 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
439 zone_page_state(zone, NR_ISOLATED_ANON);
441 return isolated > (inactive + active) / 2;
445 * isolate_migratepages_range() - isolate all migrate-able pages in range.
446 * @zone: Zone pages are in.
447 * @cc: Compaction control structure.
448 * @low_pfn: The first PFN of the range.
449 * @end_pfn: The one-past-the-last PFN of the range.
450 * @unevictable: true if it allows to isolate unevictable pages
452 * Isolate all pages that can be migrated from the range specified by
453 * [low_pfn, end_pfn). Returns zero if there is a fatal signal
454 * pending), otherwise PFN of the first page that was not scanned
455 * (which may be both less, equal to or more then end_pfn).
457 * Assumes that cc->migratepages is empty and cc->nr_migratepages is
460 * Apart from cc->migratepages and cc->nr_migratetypes this function
461 * does not modify any cc's fields, in particular it does not modify
462 * (or read for that matter) cc->migrate_pfn.
465 isolate_migratepages_range(struct zone *zone, struct compact_control *cc,
466 unsigned long low_pfn, unsigned long end_pfn, bool unevictable)
468 unsigned long last_pageblock_nr = 0, pageblock_nr;
469 unsigned long nr_scanned = 0, nr_isolated = 0;
470 struct list_head *migratelist = &cc->migratepages;
471 struct lruvec *lruvec;
474 struct page *page = NULL, *valid_page = NULL;
475 bool set_unsuitable = true;
476 const isolate_mode_t mode = (!cc->sync ? ISOLATE_ASYNC_MIGRATE : 0) |
477 (unevictable ? ISOLATE_UNEVICTABLE : 0);
480 * Ensure that there are not too many pages isolated from the LRU
481 * list by either parallel reclaimers or compaction. If there are,
482 * delay for some time until fewer pages are isolated
484 while (unlikely(too_many_isolated(zone))) {
485 /* async migration should just abort */
489 congestion_wait(BLK_RW_ASYNC, HZ/10);
491 if (fatal_signal_pending(current))
495 /* Time to isolate some pages for migration */
497 for (; low_pfn < end_pfn; low_pfn++) {
498 /* give a chance to irqs before checking need_resched() */
499 if (locked && !(low_pfn % SWAP_CLUSTER_MAX)) {
500 if (should_release_lock(&zone->lru_lock)) {
501 spin_unlock_irqrestore(&zone->lru_lock, flags);
507 * migrate_pfn does not necessarily start aligned to a
508 * pageblock. Ensure that pfn_valid is called when moving
509 * into a new MAX_ORDER_NR_PAGES range in case of large
510 * memory holes within the zone
512 if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
513 if (!pfn_valid(low_pfn)) {
514 low_pfn += MAX_ORDER_NR_PAGES - 1;
519 if (!pfn_valid_within(low_pfn))
524 * Get the page and ensure the page is within the same zone.
525 * See the comment in isolate_freepages about overlapping
526 * nodes. It is deliberate that the new zone lock is not taken
527 * as memory compaction should not move pages between nodes.
529 page = pfn_to_page(low_pfn);
530 if (page_zone(page) != zone)
536 /* If isolation recently failed, do not retry */
537 pageblock_nr = low_pfn >> pageblock_order;
538 if (last_pageblock_nr != pageblock_nr) {
541 last_pageblock_nr = pageblock_nr;
542 if (!isolation_suitable(cc, page))
546 * For async migration, also only scan in MOVABLE
547 * blocks. Async migration is optimistic to see if
548 * the minimum amount of work satisfies the allocation
550 mt = get_pageblock_migratetype(page);
551 if (!cc->sync && !migrate_async_suitable(mt)) {
552 set_unsuitable = false;
558 * Skip if free. page_order cannot be used without zone->lock
559 * as nothing prevents parallel allocations or buddy merging.
565 * Check may be lockless but that's ok as we recheck later.
566 * It's possible to migrate LRU pages and balloon pages
567 * Skip any other type of page
569 if (!PageLRU(page)) {
570 if (unlikely(balloon_page_movable(page))) {
571 if (locked && balloon_page_isolate(page)) {
572 /* Successfully isolated */
573 goto isolate_success;
580 * PageLRU is set. lru_lock normally excludes isolation
581 * splitting and collapsing (collapsing has already happened
582 * if PageLRU is set) but the lock is not necessarily taken
583 * here and it is wasteful to take it just to check transhuge.
584 * Check TransHuge without lock and skip the whole pageblock if
585 * it's either a transhuge or hugetlbfs page, as calling
586 * compound_order() without preventing THP from splitting the
587 * page underneath us may return surprising results.
589 if (PageTransHuge(page)) {
592 low_pfn += (1 << compound_order(page)) - 1;
597 * Migration will fail if an anonymous page is pinned in memory,
598 * so avoid taking lru_lock and isolating it unnecessarily in an
599 * admittedly racy check.
601 if (!page_mapping(page) &&
602 page_count(page) > page_mapcount(page))
605 /* Check if it is ok to still hold the lock */
606 locked = compact_checklock_irqsave(&zone->lru_lock, &flags,
608 if (!locked || fatal_signal_pending(current))
611 /* Recheck PageLRU and PageTransHuge under lock */
614 if (PageTransHuge(page)) {
615 low_pfn += (1 << compound_order(page)) - 1;
619 lruvec = mem_cgroup_page_lruvec(page, zone);
621 /* Try isolate the page */
622 if (__isolate_lru_page(page, mode) != 0)
625 VM_BUG_ON_PAGE(PageTransCompound(page), page);
627 /* Successfully isolated */
628 del_page_from_lru_list(page, lruvec, page_lru(page));
631 cc->finished_update_migrate = true;
632 list_add(&page->lru, migratelist);
633 cc->nr_migratepages++;
636 /* Avoid isolating too much */
637 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
645 low_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages) - 1;
648 acct_isolated(zone, locked, cc);
651 spin_unlock_irqrestore(&zone->lru_lock, flags);
654 * Update the pageblock-skip information and cached scanner pfn,
655 * if the whole pageblock was scanned without isolating any page.
657 if (low_pfn == end_pfn)
658 update_pageblock_skip(cc, valid_page, nr_isolated,
659 set_unsuitable, true);
661 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
663 count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
665 count_compact_events(COMPACTISOLATED, nr_isolated);
670 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
671 #ifdef CONFIG_COMPACTION
673 * Based on information in the current compact_control, find blocks
674 * suitable for isolating free pages from and then isolate them.
676 static void isolate_freepages(struct zone *zone,
677 struct compact_control *cc)
680 unsigned long block_start_pfn; /* start of current pageblock */
681 unsigned long block_end_pfn; /* end of current pageblock */
682 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
683 unsigned long next_free_pfn; /* start pfn for scaning at next round */
684 int nr_freepages = cc->nr_freepages;
685 struct list_head *freelist = &cc->freepages;
688 * Initialise the free scanner. The starting point is where we last
689 * successfully isolated from, zone-cached value, or the end of the
690 * zone when isolating for the first time. We need this aligned to
691 * the pageblock boundary, because we do
692 * block_start_pfn -= pageblock_nr_pages in the for loop.
693 * For ending point, take care when isolating in last pageblock of a
694 * a zone which ends in the middle of a pageblock.
695 * The low boundary is the end of the pageblock the migration scanner
698 block_start_pfn = cc->free_pfn & ~(pageblock_nr_pages-1);
699 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
701 low_pfn = ALIGN(cc->migrate_pfn + 1, pageblock_nr_pages);
704 * If no pages are isolated, the block_start_pfn < low_pfn check
710 * Isolate free pages until enough are available to migrate the
711 * pages on cc->migratepages. We stop searching if the migrate
712 * and free page scanners meet or enough free pages are isolated.
714 for (; block_start_pfn >= low_pfn && cc->nr_migratepages > nr_freepages;
715 block_end_pfn = block_start_pfn,
716 block_start_pfn -= pageblock_nr_pages) {
717 unsigned long isolated;
720 * This can iterate a massively long zone without finding any
721 * suitable migration targets, so periodically check if we need
726 if (!pfn_valid(block_start_pfn))
730 * Check for overlapping nodes/zones. It's possible on some
731 * configurations to have a setup like
733 * i.e. it's possible that all pages within a zones range of
734 * pages do not belong to a single zone.
736 page = pfn_to_page(block_start_pfn);
737 if (page_zone(page) != zone)
740 /* Check the block is suitable for migration */
741 if (!suitable_migration_target(page))
744 /* If isolation recently failed, do not retry */
745 if (!isolation_suitable(cc, page))
748 /* Found a block suitable for isolating free pages from */
749 isolated = isolate_freepages_block(cc, block_start_pfn,
750 block_end_pfn, freelist, false);
751 nr_freepages += isolated;
754 * Record the highest PFN we isolated pages from. When next
755 * looking for free pages, the search will restart here as
756 * page migration may have returned some pages to the allocator
758 if (isolated && next_free_pfn == 0) {
759 cc->finished_update_free = true;
760 next_free_pfn = block_start_pfn;
764 /* split_free_page does not map the pages */
768 * If we crossed the migrate scanner, we want to keep it that way
769 * so that compact_finished() may detect this
771 if (block_start_pfn < low_pfn)
772 next_free_pfn = cc->migrate_pfn;
774 cc->free_pfn = next_free_pfn;
775 cc->nr_freepages = nr_freepages;
779 * This is a migrate-callback that "allocates" freepages by taking pages
780 * from the isolated freelists in the block we are migrating to.
782 static struct page *compaction_alloc(struct page *migratepage,
786 struct compact_control *cc = (struct compact_control *)data;
787 struct page *freepage;
789 /* Isolate free pages if necessary */
790 if (list_empty(&cc->freepages)) {
791 isolate_freepages(cc->zone, cc);
793 if (list_empty(&cc->freepages))
797 freepage = list_entry(cc->freepages.next, struct page, lru);
798 list_del(&freepage->lru);
805 * This is a migrate-callback that "frees" freepages back to the isolated
806 * freelist. All pages on the freelist are from the same zone, so there is no
807 * special handling needed for NUMA.
809 static void compaction_free(struct page *page, unsigned long data)
811 struct compact_control *cc = (struct compact_control *)data;
813 list_add(&page->lru, &cc->freepages);
818 * We cannot control nr_migratepages fully when migration is running as
819 * migrate_pages() has no knowledge of of compact_control. When migration is
820 * complete, we count the number of pages on the list by hand.
822 static void update_nr_listpages(struct compact_control *cc)
824 int nr_migratepages = 0;
827 list_for_each_entry(page, &cc->migratepages, lru)
830 cc->nr_migratepages = nr_migratepages;
833 /* possible outcome of isolate_migratepages */
835 ISOLATE_ABORT, /* Abort compaction now */
836 ISOLATE_NONE, /* No pages isolated, continue scanning */
837 ISOLATE_SUCCESS, /* Pages isolated, migrate */
841 * Isolate all pages that can be migrated from the block pointed to by
842 * the migrate scanner within compact_control.
844 static isolate_migrate_t isolate_migratepages(struct zone *zone,
845 struct compact_control *cc)
847 unsigned long low_pfn, end_pfn;
849 /* Do not scan outside zone boundaries */
850 low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
852 /* Only scan within a pageblock boundary */
853 end_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages);
855 /* Do not cross the free scanner or scan within a memory hole */
856 if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
857 cc->migrate_pfn = end_pfn;
861 /* Perform the isolation */
862 low_pfn = isolate_migratepages_range(zone, cc, low_pfn, end_pfn, false);
863 if (!low_pfn || cc->contended)
864 return ISOLATE_ABORT;
866 cc->migrate_pfn = low_pfn;
868 return ISOLATE_SUCCESS;
871 static int compact_finished(struct zone *zone,
872 struct compact_control *cc)
875 unsigned long watermark;
877 if (fatal_signal_pending(current))
878 return COMPACT_PARTIAL;
880 /* Compaction run completes if the migrate and free scanner meet */
881 if (cc->free_pfn <= cc->migrate_pfn) {
882 /* Let the next compaction start anew. */
883 zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
884 zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
885 zone->compact_cached_free_pfn = zone_end_pfn(zone);
888 * Mark that the PG_migrate_skip information should be cleared
889 * by kswapd when it goes to sleep. kswapd does not set the
890 * flag itself as the decision to be clear should be directly
891 * based on an allocation request.
893 if (!current_is_kswapd())
894 zone->compact_blockskip_flush = true;
896 return COMPACT_COMPLETE;
900 * order == -1 is expected when compacting via
901 * /proc/sys/vm/compact_memory
904 return COMPACT_CONTINUE;
906 /* Compaction run is not finished if the watermark is not met */
907 watermark = low_wmark_pages(zone);
908 watermark += (1 << cc->order);
910 if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
911 return COMPACT_CONTINUE;
913 /* Direct compactor: Is a suitable page free? */
914 for (order = cc->order; order < MAX_ORDER; order++) {
915 struct free_area *area = &zone->free_area[order];
917 /* Job done if page is free of the right migratetype */
918 if (!list_empty(&area->free_list[cc->migratetype]))
919 return COMPACT_PARTIAL;
921 /* Job done if allocation would set block type */
922 if (cc->order >= pageblock_order && area->nr_free)
923 return COMPACT_PARTIAL;
926 return COMPACT_CONTINUE;
930 * compaction_suitable: Is this suitable to run compaction on this zone now?
932 * COMPACT_SKIPPED - If there are too few free pages for compaction
933 * COMPACT_PARTIAL - If the allocation would succeed without compaction
934 * COMPACT_CONTINUE - If compaction should run now
936 unsigned long compaction_suitable(struct zone *zone, int order)
939 unsigned long watermark;
942 * order == -1 is expected when compacting via
943 * /proc/sys/vm/compact_memory
946 return COMPACT_CONTINUE;
949 * Watermarks for order-0 must be met for compaction. Note the 2UL.
950 * This is because during migration, copies of pages need to be
951 * allocated and for a short time, the footprint is higher
953 watermark = low_wmark_pages(zone) + (2UL << order);
954 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
955 return COMPACT_SKIPPED;
958 * fragmentation index determines if allocation failures are due to
959 * low memory or external fragmentation
961 * index of -1000 implies allocations might succeed depending on
963 * index towards 0 implies failure is due to lack of memory
964 * index towards 1000 implies failure is due to fragmentation
966 * Only compact if a failure would be due to fragmentation.
968 fragindex = fragmentation_index(zone, order);
969 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
970 return COMPACT_SKIPPED;
972 if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
974 return COMPACT_PARTIAL;
976 return COMPACT_CONTINUE;
979 static int compact_zone(struct zone *zone, struct compact_control *cc)
982 unsigned long start_pfn = zone->zone_start_pfn;
983 unsigned long end_pfn = zone_end_pfn(zone);
985 ret = compaction_suitable(zone, cc->order);
987 case COMPACT_PARTIAL:
988 case COMPACT_SKIPPED:
989 /* Compaction is likely to fail */
991 case COMPACT_CONTINUE:
992 /* Fall through to compaction */
997 * Clear pageblock skip if there were failures recently and compaction
998 * is about to be retried after being deferred. kswapd does not do
999 * this reset as it'll reset the cached information when going to sleep.
1001 if (compaction_restarting(zone, cc->order) && !current_is_kswapd())
1002 __reset_isolation_suitable(zone);
1005 * Setup to move all movable pages to the end of the zone. Used cached
1006 * information on where the scanners should start but check that it
1007 * is initialised by ensuring the values are within zone boundaries.
1009 cc->migrate_pfn = zone->compact_cached_migrate_pfn[cc->sync];
1010 cc->free_pfn = zone->compact_cached_free_pfn;
1011 if (cc->free_pfn < start_pfn || cc->free_pfn > end_pfn) {
1012 cc->free_pfn = end_pfn & ~(pageblock_nr_pages-1);
1013 zone->compact_cached_free_pfn = cc->free_pfn;
1015 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn > end_pfn) {
1016 cc->migrate_pfn = start_pfn;
1017 zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
1018 zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
1021 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn, cc->free_pfn, end_pfn);
1023 migrate_prep_local();
1025 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
1026 unsigned long nr_migrate, nr_remaining;
1029 switch (isolate_migratepages(zone, cc)) {
1031 ret = COMPACT_PARTIAL;
1032 putback_movable_pages(&cc->migratepages);
1033 cc->nr_migratepages = 0;
1037 case ISOLATE_SUCCESS:
1041 nr_migrate = cc->nr_migratepages;
1042 err = migrate_pages(&cc->migratepages, compaction_alloc,
1043 compaction_free, (unsigned long)cc,
1044 cc->sync ? MIGRATE_SYNC_LIGHT : MIGRATE_ASYNC,
1046 update_nr_listpages(cc);
1047 nr_remaining = cc->nr_migratepages;
1049 trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
1052 /* Release isolated pages not migrated */
1054 putback_movable_pages(&cc->migratepages);
1055 cc->nr_migratepages = 0;
1057 * migrate_pages() may return -ENOMEM when scanners meet
1058 * and we want compact_finished() to detect it
1060 if (err == -ENOMEM && cc->free_pfn > cc->migrate_pfn) {
1061 ret = COMPACT_PARTIAL;
1068 /* Release free pages and check accounting */
1069 cc->nr_freepages -= release_freepages(&cc->freepages);
1070 VM_BUG_ON(cc->nr_freepages != 0);
1072 trace_mm_compaction_end(ret);
1077 static unsigned long compact_zone_order(struct zone *zone,
1078 int order, gfp_t gfp_mask,
1079 bool sync, bool *contended)
1082 struct compact_control cc = {
1084 .nr_migratepages = 0,
1086 .migratetype = allocflags_to_migratetype(gfp_mask),
1090 INIT_LIST_HEAD(&cc.freepages);
1091 INIT_LIST_HEAD(&cc.migratepages);
1093 ret = compact_zone(zone, &cc);
1095 VM_BUG_ON(!list_empty(&cc.freepages));
1096 VM_BUG_ON(!list_empty(&cc.migratepages));
1098 *contended = cc.contended;
1102 int sysctl_extfrag_threshold = 500;
1105 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1106 * @zonelist: The zonelist used for the current allocation
1107 * @order: The order of the current allocation
1108 * @gfp_mask: The GFP mask of the current allocation
1109 * @nodemask: The allowed nodes to allocate from
1110 * @sync: Whether migration is synchronous or not
1111 * @contended: Return value that is true if compaction was aborted due to lock contention
1112 * @page: Optionally capture a free page of the requested order during compaction
1114 * This is the main entry point for direct page compaction.
1116 unsigned long try_to_compact_pages(struct zonelist *zonelist,
1117 int order, gfp_t gfp_mask, nodemask_t *nodemask,
1118 bool sync, bool *contended)
1120 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1121 int may_enter_fs = gfp_mask & __GFP_FS;
1122 int may_perform_io = gfp_mask & __GFP_IO;
1125 int rc = COMPACT_SKIPPED;
1126 int alloc_flags = 0;
1128 /* Check if the GFP flags allow compaction */
1129 if (!order || !may_enter_fs || !may_perform_io)
1132 count_compact_event(COMPACTSTALL);
1135 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
1136 alloc_flags |= ALLOC_CMA;
1138 /* Compact each zone in the list */
1139 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1143 status = compact_zone_order(zone, order, gfp_mask, sync,
1145 rc = max(status, rc);
1147 /* If a normal allocation would succeed, stop compacting */
1148 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0,
1157 /* Compact all zones within a node */
1158 static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
1163 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1165 zone = &pgdat->node_zones[zoneid];
1166 if (!populated_zone(zone))
1169 cc->nr_freepages = 0;
1170 cc->nr_migratepages = 0;
1172 INIT_LIST_HEAD(&cc->freepages);
1173 INIT_LIST_HEAD(&cc->migratepages);
1175 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
1176 compact_zone(zone, cc);
1178 if (cc->order > 0) {
1179 if (zone_watermark_ok(zone, cc->order,
1180 low_wmark_pages(zone), 0, 0))
1181 compaction_defer_reset(zone, cc->order, false);
1184 VM_BUG_ON(!list_empty(&cc->freepages));
1185 VM_BUG_ON(!list_empty(&cc->migratepages));
1189 void compact_pgdat(pg_data_t *pgdat, int order)
1191 struct compact_control cc = {
1199 __compact_pgdat(pgdat, &cc);
1202 static void compact_node(int nid)
1204 struct compact_control cc = {
1207 .ignore_skip_hint = true,
1210 __compact_pgdat(NODE_DATA(nid), &cc);
1213 /* Compact all nodes in the system */
1214 static void compact_nodes(void)
1218 /* Flush pending updates to the LRU lists */
1219 lru_add_drain_all();
1221 for_each_online_node(nid)
1225 /* The written value is actually unused, all memory is compacted */
1226 int sysctl_compact_memory;
1228 /* This is the entry point for compacting all nodes via /proc/sys/vm */
1229 int sysctl_compaction_handler(struct ctl_table *table, int write,
1230 void __user *buffer, size_t *length, loff_t *ppos)
1238 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1239 void __user *buffer, size_t *length, loff_t *ppos)
1241 proc_dointvec_minmax(table, write, buffer, length, ppos);
1246 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1247 ssize_t sysfs_compact_node(struct device *dev,
1248 struct device_attribute *attr,
1249 const char *buf, size_t count)
1253 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1254 /* Flush pending updates to the LRU lists */
1255 lru_add_drain_all();
1262 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1264 int compaction_register_node(struct node *node)
1266 return device_create_file(&node->dev, &dev_attr_compact);
1269 void compaction_unregister_node(struct node *node)
1271 return device_remove_file(&node->dev, &dev_attr_compact);
1273 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1275 #endif /* CONFIG_COMPACTION */