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 = start_pfn;
93 zone->compact_cached_free_pfn = end_pfn;
94 zone->compact_blockskip_flush = false;
96 /* Walk the zone and mark every pageblock as suitable for isolation */
97 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
105 page = pfn_to_page(pfn);
106 if (zone != page_zone(page))
109 clear_pageblock_skip(page);
113 void reset_isolation_suitable(pg_data_t *pgdat)
117 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
118 struct zone *zone = &pgdat->node_zones[zoneid];
119 if (!populated_zone(zone))
122 /* Only flush if a full compaction finished recently */
123 if (zone->compact_blockskip_flush)
124 __reset_isolation_suitable(zone);
129 * If no pages were isolated then mark this pageblock to be skipped in the
130 * future. The information is later cleared by __reset_isolation_suitable().
132 static void update_pageblock_skip(struct compact_control *cc,
133 struct page *page, unsigned long nr_isolated,
134 bool migrate_scanner)
136 struct zone *zone = cc->zone;
138 if (cc->ignore_skip_hint)
145 unsigned long pfn = page_to_pfn(page);
146 set_pageblock_skip(page);
148 /* Update where compaction should restart */
149 if (migrate_scanner) {
150 if (!cc->finished_update_migrate &&
151 pfn > zone->compact_cached_migrate_pfn)
152 zone->compact_cached_migrate_pfn = pfn;
154 if (!cc->finished_update_free &&
155 pfn < zone->compact_cached_free_pfn)
156 zone->compact_cached_free_pfn = pfn;
161 static inline bool isolation_suitable(struct compact_control *cc,
167 static void update_pageblock_skip(struct compact_control *cc,
168 struct page *page, unsigned long nr_isolated,
169 bool migrate_scanner)
172 #endif /* CONFIG_COMPACTION */
174 static inline bool should_release_lock(spinlock_t *lock)
176 return need_resched() || spin_is_contended(lock);
180 * Compaction requires the taking of some coarse locks that are potentially
181 * very heavily contended. Check if the process needs to be scheduled or
182 * if the lock is contended. For async compaction, back out in the event
183 * if contention is severe. For sync compaction, schedule.
185 * Returns true if the lock is held.
186 * Returns false if the lock is released and compaction should abort
188 static bool compact_checklock_irqsave(spinlock_t *lock, unsigned long *flags,
189 bool locked, struct compact_control *cc)
191 if (should_release_lock(lock)) {
193 spin_unlock_irqrestore(lock, *flags);
197 /* async aborts if taking too long or contended */
199 cc->contended = true;
207 spin_lock_irqsave(lock, *flags);
211 static inline bool compact_trylock_irqsave(spinlock_t *lock,
212 unsigned long *flags, struct compact_control *cc)
214 return compact_checklock_irqsave(lock, flags, false, cc);
217 /* Returns true if the page is within a block suitable for migration to */
218 static bool suitable_migration_target(struct page *page)
220 int migratetype = get_pageblock_migratetype(page);
222 /* Don't interfere with memory hot-remove or the min_free_kbytes blocks */
223 if (migratetype == MIGRATE_RESERVE)
226 if (is_migrate_isolate(migratetype))
229 /* If the page is a large free page, then allow migration */
230 if (PageBuddy(page) && page_order(page) >= pageblock_order)
233 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
234 if (migrate_async_suitable(migratetype))
237 /* Otherwise skip the block */
242 * Isolate free pages onto a private freelist. If @strict is true, will abort
243 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
244 * (even though it may still end up isolating some pages).
246 static unsigned long isolate_freepages_block(struct compact_control *cc,
247 unsigned long blockpfn,
248 unsigned long end_pfn,
249 struct list_head *freelist,
252 int nr_scanned = 0, total_isolated = 0;
253 struct page *cursor, *valid_page = NULL;
257 cursor = pfn_to_page(blockpfn);
259 /* Isolate free pages. */
260 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
262 struct page *page = cursor;
265 if (!pfn_valid_within(blockpfn))
270 if (!PageBuddy(page))
274 * The zone lock must be held to isolate freepages.
275 * Unfortunately this is a very coarse lock and can be
276 * heavily contended if there are parallel allocations
277 * or parallel compactions. For async compaction do not
278 * spin on the lock and we acquire the lock as late as
281 locked = compact_checklock_irqsave(&cc->zone->lock, &flags,
286 /* Recheck this is a suitable migration target under lock */
287 if (!strict && !suitable_migration_target(page))
290 /* Recheck this is a buddy page under lock */
291 if (!PageBuddy(page))
294 /* Found a free page, break it into order-0 pages */
295 isolated = split_free_page(page);
296 total_isolated += isolated;
297 for (i = 0; i < isolated; i++) {
298 list_add(&page->lru, freelist);
302 /* If a page was split, advance to the end of it */
304 blockpfn += isolated - 1;
305 cursor += isolated - 1;
317 trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
320 * If strict isolation is requested by CMA then check that all the
321 * pages requested were isolated. If there were any failures, 0 is
322 * returned and CMA will fail.
324 if (strict && blockpfn < end_pfn)
328 spin_unlock_irqrestore(&cc->zone->lock, flags);
330 /* Update the pageblock-skip if the whole pageblock was scanned */
331 if (blockpfn == end_pfn)
332 update_pageblock_skip(cc, valid_page, total_isolated, false);
334 count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
336 count_compact_events(COMPACTISOLATED, total_isolated);
337 return total_isolated;
341 * isolate_freepages_range() - isolate free pages.
342 * @start_pfn: The first PFN to start isolating.
343 * @end_pfn: The one-past-last PFN.
345 * Non-free pages, invalid PFNs, or zone boundaries within the
346 * [start_pfn, end_pfn) range are considered errors, cause function to
347 * undo its actions and return zero.
349 * Otherwise, function returns one-past-the-last PFN of isolated page
350 * (which may be greater then end_pfn if end fell in a middle of
354 isolate_freepages_range(struct compact_control *cc,
355 unsigned long start_pfn, unsigned long end_pfn)
357 unsigned long isolated, pfn, block_end_pfn;
360 for (pfn = start_pfn; pfn < end_pfn; pfn += isolated) {
361 if (!pfn_valid(pfn) || cc->zone != page_zone(pfn_to_page(pfn)))
365 * On subsequent iterations ALIGN() is actually not needed,
366 * but we keep it that we not to complicate the code.
368 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
369 block_end_pfn = min(block_end_pfn, end_pfn);
371 isolated = isolate_freepages_block(cc, pfn, block_end_pfn,
375 * In strict mode, isolate_freepages_block() returns 0 if
376 * there are any holes in the block (ie. invalid PFNs or
383 * If we managed to isolate pages, it is always (1 << n) *
384 * pageblock_nr_pages for some non-negative n. (Max order
385 * page may span two pageblocks).
389 /* split_free_page does not map the pages */
390 map_pages(&freelist);
393 /* Loop terminated early, cleanup. */
394 release_freepages(&freelist);
398 /* We don't use freelists for anything. */
402 /* Update the number of anon and file isolated pages in the zone */
403 static void acct_isolated(struct zone *zone, bool locked, struct compact_control *cc)
406 unsigned int count[2] = { 0, };
408 list_for_each_entry(page, &cc->migratepages, lru)
409 count[!!page_is_file_cache(page)]++;
411 /* If locked we can use the interrupt unsafe versions */
413 __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
414 __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
416 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
417 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
421 /* Similar to reclaim, but different enough that they don't share logic */
422 static bool too_many_isolated(struct zone *zone)
424 unsigned long active, inactive, isolated;
426 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
427 zone_page_state(zone, NR_INACTIVE_ANON);
428 active = zone_page_state(zone, NR_ACTIVE_FILE) +
429 zone_page_state(zone, NR_ACTIVE_ANON);
430 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
431 zone_page_state(zone, NR_ISOLATED_ANON);
433 return isolated > (inactive + active) / 2;
437 * isolate_migratepages_range() - isolate all migrate-able pages in range.
438 * @zone: Zone pages are in.
439 * @cc: Compaction control structure.
440 * @low_pfn: The first PFN of the range.
441 * @end_pfn: The one-past-the-last PFN of the range.
442 * @unevictable: true if it allows to isolate unevictable pages
444 * Isolate all pages that can be migrated from the range specified by
445 * [low_pfn, end_pfn). Returns zero if there is a fatal signal
446 * pending), otherwise PFN of the first page that was not scanned
447 * (which may be both less, equal to or more then end_pfn).
449 * Assumes that cc->migratepages is empty and cc->nr_migratepages is
452 * Apart from cc->migratepages and cc->nr_migratetypes this function
453 * does not modify any cc's fields, in particular it does not modify
454 * (or read for that matter) cc->migrate_pfn.
457 isolate_migratepages_range(struct zone *zone, struct compact_control *cc,
458 unsigned long low_pfn, unsigned long end_pfn, bool unevictable)
460 unsigned long last_pageblock_nr = 0, pageblock_nr;
461 unsigned long nr_scanned = 0, nr_isolated = 0;
462 struct list_head *migratelist = &cc->migratepages;
463 isolate_mode_t mode = 0;
464 struct lruvec *lruvec;
467 struct page *page = NULL, *valid_page = NULL;
468 bool skipped_async_unsuitable = false;
471 * Ensure that there are not too many pages isolated from the LRU
472 * list by either parallel reclaimers or compaction. If there are,
473 * delay for some time until fewer pages are isolated
475 while (unlikely(too_many_isolated(zone))) {
476 /* async migration should just abort */
480 congestion_wait(BLK_RW_ASYNC, HZ/10);
482 if (fatal_signal_pending(current))
486 /* Time to isolate some pages for migration */
488 for (; low_pfn < end_pfn; low_pfn++) {
489 /* give a chance to irqs before checking need_resched() */
490 if (locked && !((low_pfn+1) % SWAP_CLUSTER_MAX)) {
491 if (should_release_lock(&zone->lru_lock)) {
492 spin_unlock_irqrestore(&zone->lru_lock, flags);
498 * migrate_pfn does not necessarily start aligned to a
499 * pageblock. Ensure that pfn_valid is called when moving
500 * into a new MAX_ORDER_NR_PAGES range in case of large
501 * memory holes within the zone
503 if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
504 if (!pfn_valid(low_pfn)) {
505 low_pfn += MAX_ORDER_NR_PAGES - 1;
510 if (!pfn_valid_within(low_pfn))
515 * Get the page and ensure the page is within the same zone.
516 * See the comment in isolate_freepages about overlapping
517 * nodes. It is deliberate that the new zone lock is not taken
518 * as memory compaction should not move pages between nodes.
520 page = pfn_to_page(low_pfn);
521 if (page_zone(page) != zone)
527 /* If isolation recently failed, do not retry */
528 pageblock_nr = low_pfn >> pageblock_order;
529 if (!isolation_suitable(cc, page))
533 * Skip if free. page_order cannot be used without zone->lock
534 * as nothing prevents parallel allocations or buddy merging.
540 * For async migration, also only scan in MOVABLE blocks. Async
541 * migration is optimistic to see if the minimum amount of work
542 * satisfies the allocation
544 if (!cc->sync && last_pageblock_nr != pageblock_nr &&
545 !migrate_async_suitable(get_pageblock_migratetype(page))) {
546 cc->finished_update_migrate = true;
547 skipped_async_unsuitable = true;
552 * Check may be lockless but that's ok as we recheck later.
553 * It's possible to migrate LRU pages and balloon pages
554 * Skip any other type of page
556 if (!PageLRU(page)) {
557 if (unlikely(balloon_page_movable(page))) {
558 if (locked && balloon_page_isolate(page)) {
559 /* Successfully isolated */
560 cc->finished_update_migrate = true;
561 list_add(&page->lru, migratelist);
562 cc->nr_migratepages++;
564 goto check_compact_cluster;
571 * PageLRU is set. lru_lock normally excludes isolation
572 * splitting and collapsing (collapsing has already happened
573 * if PageLRU is set) but the lock is not necessarily taken
574 * here and it is wasteful to take it just to check transhuge.
575 * Check TransHuge without lock and skip the whole pageblock if
576 * it's either a transhuge or hugetlbfs page, as calling
577 * compound_order() without preventing THP from splitting the
578 * page underneath us may return surprising results.
580 if (PageTransHuge(page)) {
583 low_pfn += (1 << compound_order(page)) - 1;
587 /* Check if it is ok to still hold the lock */
588 locked = compact_checklock_irqsave(&zone->lru_lock, &flags,
590 if (!locked || fatal_signal_pending(current))
593 /* Recheck PageLRU and PageTransHuge under lock */
596 if (PageTransHuge(page)) {
597 low_pfn += (1 << compound_order(page)) - 1;
602 mode |= ISOLATE_ASYNC_MIGRATE;
605 mode |= ISOLATE_UNEVICTABLE;
607 lruvec = mem_cgroup_page_lruvec(page, zone);
609 /* Try isolate the page */
610 if (__isolate_lru_page(page, mode) != 0)
613 VM_BUG_ON_PAGE(PageTransCompound(page), page);
615 /* Successfully isolated */
616 cc->finished_update_migrate = true;
617 del_page_from_lru_list(page, lruvec, page_lru(page));
618 list_add(&page->lru, migratelist);
619 cc->nr_migratepages++;
622 check_compact_cluster:
623 /* Avoid isolating too much */
624 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
632 low_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages) - 1;
633 last_pageblock_nr = pageblock_nr;
636 acct_isolated(zone, locked, cc);
639 spin_unlock_irqrestore(&zone->lru_lock, flags);
642 * Update the pageblock-skip information and cached scanner pfn,
643 * if the whole pageblock was scanned without isolating any page.
644 * This is not done when pageblock was skipped due to being unsuitable
645 * for async compaction, so that eventual sync compaction can try.
647 if (low_pfn == end_pfn && !skipped_async_unsuitable)
648 update_pageblock_skip(cc, valid_page, nr_isolated, true);
650 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
652 count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
654 count_compact_events(COMPACTISOLATED, nr_isolated);
659 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
660 #ifdef CONFIG_COMPACTION
662 * Based on information in the current compact_control, find blocks
663 * suitable for isolating free pages from and then isolate them.
665 static void isolate_freepages(struct zone *zone,
666 struct compact_control *cc)
669 unsigned long high_pfn, low_pfn, pfn, z_end_pfn;
670 int nr_freepages = cc->nr_freepages;
671 struct list_head *freelist = &cc->freepages;
674 * Initialise the free scanner. The starting point is where we last
675 * successfully isolated from, zone-cached value, or the end of the
676 * zone when isolating for the first time. We need this aligned to
677 * the pageblock boundary, because we do pfn -= pageblock_nr_pages
679 * The low boundary is the end of the pageblock the migration scanner
682 pfn = cc->free_pfn & ~(pageblock_nr_pages-1);
683 low_pfn = ALIGN(cc->migrate_pfn + 1, pageblock_nr_pages);
686 * Take care that if the migration scanner is at the end of the zone
687 * that the free scanner does not accidentally move to the next zone
688 * in the next isolation cycle.
690 high_pfn = min(low_pfn, pfn);
692 z_end_pfn = zone_end_pfn(zone);
695 * Isolate free pages until enough are available to migrate the
696 * pages on cc->migratepages. We stop searching if the migrate
697 * and free page scanners meet or enough free pages are isolated.
699 for (; pfn >= low_pfn && cc->nr_migratepages > nr_freepages;
700 pfn -= pageblock_nr_pages) {
701 unsigned long isolated;
702 unsigned long end_pfn;
705 * This can iterate a massively long zone without finding any
706 * suitable migration targets, so periodically check if we need
715 * Check for overlapping nodes/zones. It's possible on some
716 * configurations to have a setup like
718 * i.e. it's possible that all pages within a zones range of
719 * pages do not belong to a single zone.
721 page = pfn_to_page(pfn);
722 if (page_zone(page) != zone)
725 /* Check the block is suitable for migration */
726 if (!suitable_migration_target(page))
729 /* If isolation recently failed, do not retry */
730 if (!isolation_suitable(cc, page))
733 /* Found a block suitable for isolating free pages from */
737 * Take care when isolating in last pageblock of a zone which
738 * ends in the middle of a pageblock.
740 end_pfn = min(pfn + pageblock_nr_pages, z_end_pfn);
741 isolated = isolate_freepages_block(cc, pfn, end_pfn,
743 nr_freepages += isolated;
746 * Record the highest PFN we isolated pages from. When next
747 * looking for free pages, the search will restart here as
748 * page migration may have returned some pages to the allocator
751 cc->finished_update_free = true;
752 high_pfn = max(high_pfn, pfn);
756 /* split_free_page does not map the pages */
760 * If we crossed the migrate scanner, we want to keep it that way
761 * so that compact_finished() may detect this
764 cc->free_pfn = max(pfn, zone->zone_start_pfn);
766 cc->free_pfn = high_pfn;
767 cc->nr_freepages = nr_freepages;
771 * This is a migrate-callback that "allocates" freepages by taking pages
772 * from the isolated freelists in the block we are migrating to.
774 static struct page *compaction_alloc(struct page *migratepage,
778 struct compact_control *cc = (struct compact_control *)data;
779 struct page *freepage;
781 /* Isolate free pages if necessary */
782 if (list_empty(&cc->freepages)) {
783 isolate_freepages(cc->zone, cc);
785 if (list_empty(&cc->freepages))
789 freepage = list_entry(cc->freepages.next, struct page, lru);
790 list_del(&freepage->lru);
797 * We cannot control nr_migratepages and nr_freepages fully when migration is
798 * running as migrate_pages() has no knowledge of compact_control. When
799 * migration is complete, we count the number of pages on the lists by hand.
801 static void update_nr_listpages(struct compact_control *cc)
803 int nr_migratepages = 0;
804 int nr_freepages = 0;
807 list_for_each_entry(page, &cc->migratepages, lru)
809 list_for_each_entry(page, &cc->freepages, lru)
812 cc->nr_migratepages = nr_migratepages;
813 cc->nr_freepages = nr_freepages;
816 /* possible outcome of isolate_migratepages */
818 ISOLATE_ABORT, /* Abort compaction now */
819 ISOLATE_NONE, /* No pages isolated, continue scanning */
820 ISOLATE_SUCCESS, /* Pages isolated, migrate */
824 * Isolate all pages that can be migrated from the block pointed to by
825 * the migrate scanner within compact_control.
827 static isolate_migrate_t isolate_migratepages(struct zone *zone,
828 struct compact_control *cc)
830 unsigned long low_pfn, end_pfn;
832 /* Do not scan outside zone boundaries */
833 low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
835 /* Only scan within a pageblock boundary */
836 end_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages);
838 /* Do not cross the free scanner or scan within a memory hole */
839 if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
840 cc->migrate_pfn = end_pfn;
844 /* Perform the isolation */
845 low_pfn = isolate_migratepages_range(zone, cc, low_pfn, end_pfn, false);
846 if (!low_pfn || cc->contended)
847 return ISOLATE_ABORT;
849 cc->migrate_pfn = low_pfn;
851 return ISOLATE_SUCCESS;
854 static int compact_finished(struct zone *zone,
855 struct compact_control *cc)
858 unsigned long watermark;
860 if (fatal_signal_pending(current))
861 return COMPACT_PARTIAL;
863 /* Compaction run completes if the migrate and free scanner meet */
864 if (cc->free_pfn <= cc->migrate_pfn) {
865 /* Let the next compaction start anew. */
866 zone->compact_cached_migrate_pfn = zone->zone_start_pfn;
867 zone->compact_cached_free_pfn = zone_end_pfn(zone);
870 * Mark that the PG_migrate_skip information should be cleared
871 * by kswapd when it goes to sleep. kswapd does not set the
872 * flag itself as the decision to be clear should be directly
873 * based on an allocation request.
875 if (!current_is_kswapd())
876 zone->compact_blockskip_flush = true;
878 return COMPACT_COMPLETE;
882 * order == -1 is expected when compacting via
883 * /proc/sys/vm/compact_memory
886 return COMPACT_CONTINUE;
888 /* Compaction run is not finished if the watermark is not met */
889 watermark = low_wmark_pages(zone);
890 watermark += (1 << cc->order);
892 if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
893 return COMPACT_CONTINUE;
895 /* Direct compactor: Is a suitable page free? */
896 for (order = cc->order; order < MAX_ORDER; order++) {
897 struct free_area *area = &zone->free_area[order];
899 /* Job done if page is free of the right migratetype */
900 if (!list_empty(&area->free_list[cc->migratetype]))
901 return COMPACT_PARTIAL;
903 /* Job done if allocation would set block type */
904 if (cc->order >= pageblock_order && area->nr_free)
905 return COMPACT_PARTIAL;
908 return COMPACT_CONTINUE;
912 * compaction_suitable: Is this suitable to run compaction on this zone now?
914 * COMPACT_SKIPPED - If there are too few free pages for compaction
915 * COMPACT_PARTIAL - If the allocation would succeed without compaction
916 * COMPACT_CONTINUE - If compaction should run now
918 unsigned long compaction_suitable(struct zone *zone, int order)
921 unsigned long watermark;
924 * order == -1 is expected when compacting via
925 * /proc/sys/vm/compact_memory
928 return COMPACT_CONTINUE;
931 * Watermarks for order-0 must be met for compaction. Note the 2UL.
932 * This is because during migration, copies of pages need to be
933 * allocated and for a short time, the footprint is higher
935 watermark = low_wmark_pages(zone) + (2UL << order);
936 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
937 return COMPACT_SKIPPED;
940 * fragmentation index determines if allocation failures are due to
941 * low memory or external fragmentation
943 * index of -1000 implies allocations might succeed depending on
945 * index towards 0 implies failure is due to lack of memory
946 * index towards 1000 implies failure is due to fragmentation
948 * Only compact if a failure would be due to fragmentation.
950 fragindex = fragmentation_index(zone, order);
951 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
952 return COMPACT_SKIPPED;
954 if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
956 return COMPACT_PARTIAL;
958 return COMPACT_CONTINUE;
961 static int compact_zone(struct zone *zone, struct compact_control *cc)
964 unsigned long start_pfn = zone->zone_start_pfn;
965 unsigned long end_pfn = zone_end_pfn(zone);
967 ret = compaction_suitable(zone, cc->order);
969 case COMPACT_PARTIAL:
970 case COMPACT_SKIPPED:
971 /* Compaction is likely to fail */
973 case COMPACT_CONTINUE:
974 /* Fall through to compaction */
979 * Clear pageblock skip if there were failures recently and compaction
980 * is about to be retried after being deferred. kswapd does not do
981 * this reset as it'll reset the cached information when going to sleep.
983 if (compaction_restarting(zone, cc->order) && !current_is_kswapd())
984 __reset_isolation_suitable(zone);
987 * Setup to move all movable pages to the end of the zone. Used cached
988 * information on where the scanners should start but check that it
989 * is initialised by ensuring the values are within zone boundaries.
991 cc->migrate_pfn = zone->compact_cached_migrate_pfn;
992 cc->free_pfn = zone->compact_cached_free_pfn;
993 if (cc->free_pfn < start_pfn || cc->free_pfn > end_pfn) {
994 cc->free_pfn = end_pfn & ~(pageblock_nr_pages-1);
995 zone->compact_cached_free_pfn = cc->free_pfn;
997 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn > end_pfn) {
998 cc->migrate_pfn = start_pfn;
999 zone->compact_cached_migrate_pfn = cc->migrate_pfn;
1002 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn, cc->free_pfn, end_pfn);
1004 migrate_prep_local();
1006 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
1007 unsigned long nr_migrate, nr_remaining;
1010 switch (isolate_migratepages(zone, cc)) {
1012 ret = COMPACT_PARTIAL;
1013 putback_movable_pages(&cc->migratepages);
1014 cc->nr_migratepages = 0;
1018 case ISOLATE_SUCCESS:
1022 nr_migrate = cc->nr_migratepages;
1023 err = migrate_pages(&cc->migratepages, compaction_alloc,
1025 cc->sync ? MIGRATE_SYNC_LIGHT : MIGRATE_ASYNC,
1027 update_nr_listpages(cc);
1028 nr_remaining = cc->nr_migratepages;
1030 trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
1033 /* Release isolated pages not migrated */
1035 putback_movable_pages(&cc->migratepages);
1036 cc->nr_migratepages = 0;
1038 * migrate_pages() may return -ENOMEM when scanners meet
1039 * and we want compact_finished() to detect it
1041 if (err == -ENOMEM && cc->free_pfn > cc->migrate_pfn) {
1042 ret = COMPACT_PARTIAL;
1049 /* Release free pages and check accounting */
1050 cc->nr_freepages -= release_freepages(&cc->freepages);
1051 VM_BUG_ON(cc->nr_freepages != 0);
1053 trace_mm_compaction_end(ret);
1058 static unsigned long compact_zone_order(struct zone *zone,
1059 int order, gfp_t gfp_mask,
1060 bool sync, bool *contended)
1063 struct compact_control cc = {
1065 .nr_migratepages = 0,
1067 .migratetype = allocflags_to_migratetype(gfp_mask),
1071 INIT_LIST_HEAD(&cc.freepages);
1072 INIT_LIST_HEAD(&cc.migratepages);
1074 ret = compact_zone(zone, &cc);
1076 VM_BUG_ON(!list_empty(&cc.freepages));
1077 VM_BUG_ON(!list_empty(&cc.migratepages));
1079 *contended = cc.contended;
1083 int sysctl_extfrag_threshold = 500;
1086 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1087 * @zonelist: The zonelist used for the current allocation
1088 * @order: The order of the current allocation
1089 * @gfp_mask: The GFP mask of the current allocation
1090 * @nodemask: The allowed nodes to allocate from
1091 * @sync: Whether migration is synchronous or not
1092 * @contended: Return value that is true if compaction was aborted due to lock contention
1093 * @page: Optionally capture a free page of the requested order during compaction
1095 * This is the main entry point for direct page compaction.
1097 unsigned long try_to_compact_pages(struct zonelist *zonelist,
1098 int order, gfp_t gfp_mask, nodemask_t *nodemask,
1099 bool sync, bool *contended)
1101 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1102 int may_enter_fs = gfp_mask & __GFP_FS;
1103 int may_perform_io = gfp_mask & __GFP_IO;
1106 int rc = COMPACT_SKIPPED;
1107 int alloc_flags = 0;
1109 /* Check if the GFP flags allow compaction */
1110 if (!order || !may_enter_fs || !may_perform_io)
1113 count_compact_event(COMPACTSTALL);
1116 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
1117 alloc_flags |= ALLOC_CMA;
1119 /* Compact each zone in the list */
1120 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1124 status = compact_zone_order(zone, order, gfp_mask, sync,
1126 rc = max(status, rc);
1128 /* If a normal allocation would succeed, stop compacting */
1129 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0,
1138 /* Compact all zones within a node */
1139 static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
1144 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1146 zone = &pgdat->node_zones[zoneid];
1147 if (!populated_zone(zone))
1150 cc->nr_freepages = 0;
1151 cc->nr_migratepages = 0;
1153 INIT_LIST_HEAD(&cc->freepages);
1154 INIT_LIST_HEAD(&cc->migratepages);
1156 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
1157 compact_zone(zone, cc);
1159 if (cc->order > 0) {
1160 if (zone_watermark_ok(zone, cc->order,
1161 low_wmark_pages(zone), 0, 0))
1162 compaction_defer_reset(zone, cc->order, false);
1163 /* Currently async compaction is never deferred. */
1165 defer_compaction(zone, cc->order);
1168 VM_BUG_ON(!list_empty(&cc->freepages));
1169 VM_BUG_ON(!list_empty(&cc->migratepages));
1173 void compact_pgdat(pg_data_t *pgdat, int order)
1175 struct compact_control cc = {
1183 __compact_pgdat(pgdat, &cc);
1186 static void compact_node(int nid)
1188 struct compact_control cc = {
1193 __compact_pgdat(NODE_DATA(nid), &cc);
1196 /* Compact all nodes in the system */
1197 static void compact_nodes(void)
1201 /* Flush pending updates to the LRU lists */
1202 lru_add_drain_all();
1204 for_each_online_node(nid)
1208 /* The written value is actually unused, all memory is compacted */
1209 int sysctl_compact_memory;
1211 /* This is the entry point for compacting all nodes via /proc/sys/vm */
1212 int sysctl_compaction_handler(struct ctl_table *table, int write,
1213 void __user *buffer, size_t *length, loff_t *ppos)
1221 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1222 void __user *buffer, size_t *length, loff_t *ppos)
1224 proc_dointvec_minmax(table, write, buffer, length, ppos);
1229 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1230 ssize_t sysfs_compact_node(struct device *dev,
1231 struct device_attribute *attr,
1232 const char *buf, size_t count)
1236 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1237 /* Flush pending updates to the LRU lists */
1238 lru_add_drain_all();
1245 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1247 int compaction_register_node(struct node *node)
1249 return device_create_file(&node->dev, &dev_attr_compact);
1252 void compaction_unregister_node(struct node *node)
1254 return device_remove_file(&node->dev, &dev_attr_compact);
1256 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1258 #endif /* CONFIG_COMPACTION */