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;
141 unsigned long pfn = page_to_pfn(page);
142 set_pageblock_skip(page);
144 /* Update where compaction should restart */
145 if (migrate_scanner) {
146 if (!cc->finished_update_migrate &&
147 pfn > zone->compact_cached_migrate_pfn)
148 zone->compact_cached_migrate_pfn = pfn;
150 if (!cc->finished_update_free &&
151 pfn < zone->compact_cached_free_pfn)
152 zone->compact_cached_free_pfn = pfn;
157 static inline bool isolation_suitable(struct compact_control *cc,
163 static void update_pageblock_skip(struct compact_control *cc,
164 struct page *page, unsigned long nr_isolated,
165 bool migrate_scanner)
168 #endif /* CONFIG_COMPACTION */
170 static inline bool should_release_lock(spinlock_t *lock)
172 return need_resched() || spin_is_contended(lock);
176 * Compaction requires the taking of some coarse locks that are potentially
177 * very heavily contended. Check if the process needs to be scheduled or
178 * if the lock is contended. For async compaction, back out in the event
179 * if contention is severe. For sync compaction, schedule.
181 * Returns true if the lock is held.
182 * Returns false if the lock is released and compaction should abort
184 static bool compact_checklock_irqsave(spinlock_t *lock, unsigned long *flags,
185 bool locked, struct compact_control *cc)
187 if (should_release_lock(lock)) {
189 spin_unlock_irqrestore(lock, *flags);
193 /* async aborts if taking too long or contended */
195 cc->contended = true;
203 spin_lock_irqsave(lock, *flags);
207 static inline bool compact_trylock_irqsave(spinlock_t *lock,
208 unsigned long *flags, struct compact_control *cc)
210 return compact_checklock_irqsave(lock, flags, false, cc);
213 /* Returns true if the page is within a block suitable for migration to */
214 static bool suitable_migration_target(struct page *page)
216 int migratetype = get_pageblock_migratetype(page);
218 /* Don't interfere with memory hot-remove or the min_free_kbytes blocks */
219 if (migratetype == MIGRATE_RESERVE)
222 if (is_migrate_isolate(migratetype))
225 /* If the page is a large free page, then allow migration */
226 if (PageBuddy(page) && page_order(page) >= pageblock_order)
229 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
230 if (migrate_async_suitable(migratetype))
233 /* Otherwise skip the block */
238 * Isolate free pages onto a private freelist. If @strict is true, will abort
239 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
240 * (even though it may still end up isolating some pages).
242 static unsigned long isolate_freepages_block(struct compact_control *cc,
243 unsigned long blockpfn,
244 unsigned long end_pfn,
245 struct list_head *freelist,
248 int nr_scanned = 0, total_isolated = 0;
249 struct page *cursor, *valid_page = NULL;
250 unsigned long nr_strict_required = end_pfn - blockpfn;
254 cursor = pfn_to_page(blockpfn);
256 /* Isolate free pages. */
257 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
259 struct page *page = cursor;
262 if (!pfn_valid_within(blockpfn))
266 if (!PageBuddy(page))
270 * The zone lock must be held to isolate freepages.
271 * Unfortunately this is a very coarse lock and can be
272 * heavily contended if there are parallel allocations
273 * or parallel compactions. For async compaction do not
274 * spin on the lock and we acquire the lock as late as
277 locked = compact_checklock_irqsave(&cc->zone->lock, &flags,
282 /* Recheck this is a suitable migration target under lock */
283 if (!strict && !suitable_migration_target(page))
286 /* Recheck this is a buddy page under lock */
287 if (!PageBuddy(page))
290 /* Found a free page, break it into order-0 pages */
291 isolated = split_free_page(page);
292 if (!isolated && strict)
294 total_isolated += isolated;
295 for (i = 0; i < isolated; i++) {
296 list_add(&page->lru, freelist);
300 /* If a page was split, advance to the end of it */
302 blockpfn += isolated - 1;
303 cursor += isolated - 1;
307 trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
310 * If strict isolation is requested by CMA then check that all the
311 * pages requested were isolated. If there were any failures, 0 is
312 * returned and CMA will fail.
314 if (strict && nr_strict_required > total_isolated)
318 spin_unlock_irqrestore(&cc->zone->lock, flags);
320 /* Update the pageblock-skip if the whole pageblock was scanned */
321 if (blockpfn == end_pfn)
322 update_pageblock_skip(cc, valid_page, total_isolated, false);
324 count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
326 count_compact_events(COMPACTISOLATED, total_isolated);
327 return total_isolated;
331 * isolate_freepages_range() - isolate free pages.
332 * @start_pfn: The first PFN to start isolating.
333 * @end_pfn: The one-past-last PFN.
335 * Non-free pages, invalid PFNs, or zone boundaries within the
336 * [start_pfn, end_pfn) range are considered errors, cause function to
337 * undo its actions and return zero.
339 * Otherwise, function returns one-past-the-last PFN of isolated page
340 * (which may be greater then end_pfn if end fell in a middle of
344 isolate_freepages_range(struct compact_control *cc,
345 unsigned long start_pfn, unsigned long end_pfn)
347 unsigned long isolated, pfn, block_end_pfn;
350 for (pfn = start_pfn; pfn < end_pfn; pfn += isolated) {
351 if (!pfn_valid(pfn) || cc->zone != page_zone(pfn_to_page(pfn)))
355 * On subsequent iterations ALIGN() is actually not needed,
356 * but we keep it that we not to complicate the code.
358 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
359 block_end_pfn = min(block_end_pfn, end_pfn);
361 isolated = isolate_freepages_block(cc, pfn, block_end_pfn,
365 * In strict mode, isolate_freepages_block() returns 0 if
366 * there are any holes in the block (ie. invalid PFNs or
373 * If we managed to isolate pages, it is always (1 << n) *
374 * pageblock_nr_pages for some non-negative n. (Max order
375 * page may span two pageblocks).
379 /* split_free_page does not map the pages */
380 map_pages(&freelist);
383 /* Loop terminated early, cleanup. */
384 release_freepages(&freelist);
388 /* We don't use freelists for anything. */
392 /* Update the number of anon and file isolated pages in the zone */
393 static void acct_isolated(struct zone *zone, bool locked, struct compact_control *cc)
396 unsigned int count[2] = { 0, };
398 list_for_each_entry(page, &cc->migratepages, lru)
399 count[!!page_is_file_cache(page)]++;
401 /* If locked we can use the interrupt unsafe versions */
403 __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
404 __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
406 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
407 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
411 /* Similar to reclaim, but different enough that they don't share logic */
412 static bool too_many_isolated(struct zone *zone)
414 unsigned long active, inactive, isolated;
416 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
417 zone_page_state(zone, NR_INACTIVE_ANON);
418 active = zone_page_state(zone, NR_ACTIVE_FILE) +
419 zone_page_state(zone, NR_ACTIVE_ANON);
420 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
421 zone_page_state(zone, NR_ISOLATED_ANON);
423 return isolated > (inactive + active) / 2;
427 * isolate_migratepages_range() - isolate all migrate-able pages in range.
428 * @zone: Zone pages are in.
429 * @cc: Compaction control structure.
430 * @low_pfn: The first PFN of the range.
431 * @end_pfn: The one-past-the-last PFN of the range.
432 * @unevictable: true if it allows to isolate unevictable pages
434 * Isolate all pages that can be migrated from the range specified by
435 * [low_pfn, end_pfn). Returns zero if there is a fatal signal
436 * pending), otherwise PFN of the first page that was not scanned
437 * (which may be both less, equal to or more then end_pfn).
439 * Assumes that cc->migratepages is empty and cc->nr_migratepages is
442 * Apart from cc->migratepages and cc->nr_migratetypes this function
443 * does not modify any cc's fields, in particular it does not modify
444 * (or read for that matter) cc->migrate_pfn.
447 isolate_migratepages_range(struct zone *zone, struct compact_control *cc,
448 unsigned long low_pfn, unsigned long end_pfn, bool unevictable)
450 unsigned long last_pageblock_nr = 0, pageblock_nr;
451 unsigned long nr_scanned = 0, nr_isolated = 0;
452 struct list_head *migratelist = &cc->migratepages;
453 isolate_mode_t mode = 0;
454 struct lruvec *lruvec;
457 struct page *page = NULL, *valid_page = NULL;
460 * Ensure that there are not too many pages isolated from the LRU
461 * list by either parallel reclaimers or compaction. If there are,
462 * delay for some time until fewer pages are isolated
464 while (unlikely(too_many_isolated(zone))) {
465 /* async migration should just abort */
469 congestion_wait(BLK_RW_ASYNC, HZ/10);
471 if (fatal_signal_pending(current))
475 /* Time to isolate some pages for migration */
477 for (; low_pfn < end_pfn; low_pfn++) {
478 /* give a chance to irqs before checking need_resched() */
479 if (locked && !((low_pfn+1) % SWAP_CLUSTER_MAX)) {
480 if (should_release_lock(&zone->lru_lock)) {
481 spin_unlock_irqrestore(&zone->lru_lock, flags);
487 * migrate_pfn does not necessarily start aligned to a
488 * pageblock. Ensure that pfn_valid is called when moving
489 * into a new MAX_ORDER_NR_PAGES range in case of large
490 * memory holes within the zone
492 if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
493 if (!pfn_valid(low_pfn)) {
494 low_pfn += MAX_ORDER_NR_PAGES - 1;
499 if (!pfn_valid_within(low_pfn))
504 * Get the page and ensure the page is within the same zone.
505 * See the comment in isolate_freepages about overlapping
506 * nodes. It is deliberate that the new zone lock is not taken
507 * as memory compaction should not move pages between nodes.
509 page = pfn_to_page(low_pfn);
510 if (page_zone(page) != zone)
516 /* If isolation recently failed, do not retry */
517 pageblock_nr = low_pfn >> pageblock_order;
518 if (!isolation_suitable(cc, page))
526 * For async migration, also only scan in MOVABLE blocks. Async
527 * migration is optimistic to see if the minimum amount of work
528 * satisfies the allocation
530 if (!cc->sync && last_pageblock_nr != pageblock_nr &&
531 !migrate_async_suitable(get_pageblock_migratetype(page))) {
532 cc->finished_update_migrate = true;
537 * Check may be lockless but that's ok as we recheck later.
538 * It's possible to migrate LRU pages and balloon pages
539 * Skip any other type of page
541 if (!PageLRU(page)) {
542 if (unlikely(balloon_page_movable(page))) {
543 if (locked && balloon_page_isolate(page)) {
544 /* Successfully isolated */
545 cc->finished_update_migrate = true;
546 list_add(&page->lru, migratelist);
547 cc->nr_migratepages++;
549 goto check_compact_cluster;
556 * PageLRU is set. lru_lock normally excludes isolation
557 * splitting and collapsing (collapsing has already happened
558 * if PageLRU is set) but the lock is not necessarily taken
559 * here and it is wasteful to take it just to check transhuge.
560 * Check TransHuge without lock and skip the whole pageblock if
561 * it's either a transhuge or hugetlbfs page, as calling
562 * compound_order() without preventing THP from splitting the
563 * page underneath us may return surprising results.
565 if (PageTransHuge(page)) {
568 low_pfn += (1 << compound_order(page)) - 1;
572 /* Check if it is ok to still hold the lock */
573 locked = compact_checklock_irqsave(&zone->lru_lock, &flags,
575 if (!locked || fatal_signal_pending(current))
578 /* Recheck PageLRU and PageTransHuge under lock */
581 if (PageTransHuge(page)) {
582 low_pfn += (1 << compound_order(page)) - 1;
587 mode |= ISOLATE_ASYNC_MIGRATE;
590 mode |= ISOLATE_UNEVICTABLE;
592 lruvec = mem_cgroup_page_lruvec(page, zone);
594 /* Try isolate the page */
595 if (__isolate_lru_page(page, mode) != 0)
598 VM_BUG_ON(PageTransCompound(page));
600 /* Successfully isolated */
601 cc->finished_update_migrate = true;
602 del_page_from_lru_list(page, lruvec, page_lru(page));
603 list_add(&page->lru, migratelist);
604 cc->nr_migratepages++;
607 check_compact_cluster:
608 /* Avoid isolating too much */
609 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
617 low_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages) - 1;
618 last_pageblock_nr = pageblock_nr;
621 acct_isolated(zone, locked, cc);
624 spin_unlock_irqrestore(&zone->lru_lock, flags);
626 /* Update the pageblock-skip if the whole pageblock was scanned */
627 if (low_pfn == end_pfn)
628 update_pageblock_skip(cc, valid_page, nr_isolated, true);
630 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
632 count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
634 count_compact_events(COMPACTISOLATED, nr_isolated);
639 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
640 #ifdef CONFIG_COMPACTION
642 * Based on information in the current compact_control, find blocks
643 * suitable for isolating free pages from and then isolate them.
645 static void isolate_freepages(struct zone *zone,
646 struct compact_control *cc)
649 unsigned long high_pfn, low_pfn, pfn, z_end_pfn, end_pfn;
650 int nr_freepages = cc->nr_freepages;
651 struct list_head *freelist = &cc->freepages;
654 * Initialise the free scanner. The starting point is where we last
655 * scanned from (or the end of the zone if starting). The low point
656 * is the end of the pageblock the migration scanner is using.
659 low_pfn = cc->migrate_pfn + pageblock_nr_pages;
662 * Take care that if the migration scanner is at the end of the zone
663 * that the free scanner does not accidentally move to the next zone
664 * in the next isolation cycle.
666 high_pfn = min(low_pfn, pfn);
668 z_end_pfn = zone_end_pfn(zone);
671 * Isolate free pages until enough are available to migrate the
672 * pages on cc->migratepages. We stop searching if the migrate
673 * and free page scanners meet or enough free pages are isolated.
675 for (; pfn > low_pfn && cc->nr_migratepages > nr_freepages;
676 pfn -= pageblock_nr_pages) {
677 unsigned long isolated;
680 * This can iterate a massively long zone without finding any
681 * suitable migration targets, so periodically check if we need
690 * Check for overlapping nodes/zones. It's possible on some
691 * configurations to have a setup like
693 * i.e. it's possible that all pages within a zones range of
694 * pages do not belong to a single zone.
696 page = pfn_to_page(pfn);
697 if (page_zone(page) != zone)
700 /* Check the block is suitable for migration */
701 if (!suitable_migration_target(page))
704 /* If isolation recently failed, do not retry */
705 if (!isolation_suitable(cc, page))
708 /* Found a block suitable for isolating free pages from */
712 * As pfn may not start aligned, pfn+pageblock_nr_page
713 * may cross a MAX_ORDER_NR_PAGES boundary and miss
714 * a pfn_valid check. Ensure isolate_freepages_block()
715 * only scans within a pageblock
717 end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
718 end_pfn = min(end_pfn, z_end_pfn);
719 isolated = isolate_freepages_block(cc, pfn, end_pfn,
721 nr_freepages += isolated;
724 * Record the highest PFN we isolated pages from. When next
725 * looking for free pages, the search will restart here as
726 * page migration may have returned some pages to the allocator
729 cc->finished_update_free = true;
730 high_pfn = max(high_pfn, pfn);
734 /* split_free_page does not map the pages */
737 cc->free_pfn = high_pfn;
738 cc->nr_freepages = nr_freepages;
742 * This is a migrate-callback that "allocates" freepages by taking pages
743 * from the isolated freelists in the block we are migrating to.
745 static struct page *compaction_alloc(struct page *migratepage,
749 struct compact_control *cc = (struct compact_control *)data;
750 struct page *freepage;
752 /* Isolate free pages if necessary */
753 if (list_empty(&cc->freepages)) {
754 isolate_freepages(cc->zone, cc);
756 if (list_empty(&cc->freepages))
760 freepage = list_entry(cc->freepages.next, struct page, lru);
761 list_del(&freepage->lru);
768 * We cannot control nr_migratepages and nr_freepages fully when migration is
769 * running as migrate_pages() has no knowledge of compact_control. When
770 * migration is complete, we count the number of pages on the lists by hand.
772 static void update_nr_listpages(struct compact_control *cc)
774 int nr_migratepages = 0;
775 int nr_freepages = 0;
778 list_for_each_entry(page, &cc->migratepages, lru)
780 list_for_each_entry(page, &cc->freepages, lru)
783 cc->nr_migratepages = nr_migratepages;
784 cc->nr_freepages = nr_freepages;
787 /* possible outcome of isolate_migratepages */
789 ISOLATE_ABORT, /* Abort compaction now */
790 ISOLATE_NONE, /* No pages isolated, continue scanning */
791 ISOLATE_SUCCESS, /* Pages isolated, migrate */
795 * Isolate all pages that can be migrated from the block pointed to by
796 * the migrate scanner within compact_control.
798 static isolate_migrate_t isolate_migratepages(struct zone *zone,
799 struct compact_control *cc)
801 unsigned long low_pfn, end_pfn;
803 /* Do not scan outside zone boundaries */
804 low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
806 /* Only scan within a pageblock boundary */
807 end_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages);
809 /* Do not cross the free scanner or scan within a memory hole */
810 if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
811 cc->migrate_pfn = end_pfn;
815 /* Perform the isolation */
816 low_pfn = isolate_migratepages_range(zone, cc, low_pfn, end_pfn, false);
817 if (!low_pfn || cc->contended)
818 return ISOLATE_ABORT;
820 cc->migrate_pfn = low_pfn;
822 return ISOLATE_SUCCESS;
825 static int compact_finished(struct zone *zone,
826 struct compact_control *cc)
829 unsigned long watermark;
831 if (fatal_signal_pending(current))
832 return COMPACT_PARTIAL;
834 /* Compaction run completes if the migrate and free scanner meet */
835 if (cc->free_pfn <= cc->migrate_pfn) {
837 * Mark that the PG_migrate_skip information should be cleared
838 * by kswapd when it goes to sleep. kswapd does not set the
839 * flag itself as the decision to be clear should be directly
840 * based on an allocation request.
842 if (!current_is_kswapd())
843 zone->compact_blockskip_flush = true;
845 return COMPACT_COMPLETE;
849 * order == -1 is expected when compacting via
850 * /proc/sys/vm/compact_memory
853 return COMPACT_CONTINUE;
855 /* Compaction run is not finished if the watermark is not met */
856 watermark = low_wmark_pages(zone);
857 watermark += (1 << cc->order);
859 if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
860 return COMPACT_CONTINUE;
862 /* Direct compactor: Is a suitable page free? */
863 for (order = cc->order; order < MAX_ORDER; order++) {
864 struct free_area *area = &zone->free_area[order];
866 /* Job done if page is free of the right migratetype */
867 if (!list_empty(&area->free_list[cc->migratetype]))
868 return COMPACT_PARTIAL;
870 /* Job done if allocation would set block type */
871 if (cc->order >= pageblock_order && area->nr_free)
872 return COMPACT_PARTIAL;
875 return COMPACT_CONTINUE;
879 * compaction_suitable: Is this suitable to run compaction on this zone now?
881 * COMPACT_SKIPPED - If there are too few free pages for compaction
882 * COMPACT_PARTIAL - If the allocation would succeed without compaction
883 * COMPACT_CONTINUE - If compaction should run now
885 unsigned long compaction_suitable(struct zone *zone, int order)
888 unsigned long watermark;
891 * order == -1 is expected when compacting via
892 * /proc/sys/vm/compact_memory
895 return COMPACT_CONTINUE;
898 * Watermarks for order-0 must be met for compaction. Note the 2UL.
899 * This is because during migration, copies of pages need to be
900 * allocated and for a short time, the footprint is higher
902 watermark = low_wmark_pages(zone) + (2UL << order);
903 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
904 return COMPACT_SKIPPED;
907 * fragmentation index determines if allocation failures are due to
908 * low memory or external fragmentation
910 * index of -1000 implies allocations might succeed depending on
912 * index towards 0 implies failure is due to lack of memory
913 * index towards 1000 implies failure is due to fragmentation
915 * Only compact if a failure would be due to fragmentation.
917 fragindex = fragmentation_index(zone, order);
918 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
919 return COMPACT_SKIPPED;
921 if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
923 return COMPACT_PARTIAL;
925 return COMPACT_CONTINUE;
928 static int compact_zone(struct zone *zone, struct compact_control *cc)
931 unsigned long start_pfn = zone->zone_start_pfn;
932 unsigned long end_pfn = zone_end_pfn(zone);
934 ret = compaction_suitable(zone, cc->order);
936 case COMPACT_PARTIAL:
937 case COMPACT_SKIPPED:
938 /* Compaction is likely to fail */
940 case COMPACT_CONTINUE:
941 /* Fall through to compaction */
946 * Setup to move all movable pages to the end of the zone. Used cached
947 * information on where the scanners should start but check that it
948 * is initialised by ensuring the values are within zone boundaries.
950 cc->migrate_pfn = zone->compact_cached_migrate_pfn;
951 cc->free_pfn = zone->compact_cached_free_pfn;
952 if (cc->free_pfn < start_pfn || cc->free_pfn > end_pfn) {
953 cc->free_pfn = end_pfn & ~(pageblock_nr_pages-1);
954 zone->compact_cached_free_pfn = cc->free_pfn;
956 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn > end_pfn) {
957 cc->migrate_pfn = start_pfn;
958 zone->compact_cached_migrate_pfn = cc->migrate_pfn;
962 * Clear pageblock skip if there were failures recently and compaction
963 * is about to be retried after being deferred. kswapd does not do
964 * this reset as it'll reset the cached information when going to sleep.
966 if (compaction_restarting(zone, cc->order) && !current_is_kswapd())
967 __reset_isolation_suitable(zone);
969 migrate_prep_local();
971 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
972 unsigned long nr_migrate, nr_remaining;
975 switch (isolate_migratepages(zone, cc)) {
977 ret = COMPACT_PARTIAL;
978 putback_movable_pages(&cc->migratepages);
979 cc->nr_migratepages = 0;
983 case ISOLATE_SUCCESS:
987 nr_migrate = cc->nr_migratepages;
988 err = migrate_pages(&cc->migratepages, compaction_alloc,
990 cc->sync ? MIGRATE_SYNC_LIGHT : MIGRATE_ASYNC,
992 update_nr_listpages(cc);
993 nr_remaining = cc->nr_migratepages;
995 trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
998 /* Release isolated pages not migrated */
1000 putback_movable_pages(&cc->migratepages);
1001 cc->nr_migratepages = 0;
1002 if (err == -ENOMEM) {
1003 ret = COMPACT_PARTIAL;
1010 /* Release free pages and check accounting */
1011 cc->nr_freepages -= release_freepages(&cc->freepages);
1012 VM_BUG_ON(cc->nr_freepages != 0);
1017 static unsigned long compact_zone_order(struct zone *zone,
1018 int order, gfp_t gfp_mask,
1019 bool sync, bool *contended)
1022 struct compact_control cc = {
1024 .nr_migratepages = 0,
1026 .migratetype = allocflags_to_migratetype(gfp_mask),
1030 INIT_LIST_HEAD(&cc.freepages);
1031 INIT_LIST_HEAD(&cc.migratepages);
1033 ret = compact_zone(zone, &cc);
1035 VM_BUG_ON(!list_empty(&cc.freepages));
1036 VM_BUG_ON(!list_empty(&cc.migratepages));
1038 *contended = cc.contended;
1042 int sysctl_extfrag_threshold = 500;
1045 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1046 * @zonelist: The zonelist used for the current allocation
1047 * @order: The order of the current allocation
1048 * @gfp_mask: The GFP mask of the current allocation
1049 * @nodemask: The allowed nodes to allocate from
1050 * @sync: Whether migration is synchronous or not
1051 * @contended: Return value that is true if compaction was aborted due to lock contention
1052 * @page: Optionally capture a free page of the requested order during compaction
1054 * This is the main entry point for direct page compaction.
1056 unsigned long try_to_compact_pages(struct zonelist *zonelist,
1057 int order, gfp_t gfp_mask, nodemask_t *nodemask,
1058 bool sync, bool *contended)
1060 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1061 int may_enter_fs = gfp_mask & __GFP_FS;
1062 int may_perform_io = gfp_mask & __GFP_IO;
1065 int rc = COMPACT_SKIPPED;
1066 int alloc_flags = 0;
1068 /* Check if the GFP flags allow compaction */
1069 if (!order || !may_enter_fs || !may_perform_io)
1072 count_compact_event(COMPACTSTALL);
1075 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
1076 alloc_flags |= ALLOC_CMA;
1078 /* Compact each zone in the list */
1079 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1083 status = compact_zone_order(zone, order, gfp_mask, sync,
1085 rc = max(status, rc);
1087 /* If a normal allocation would succeed, stop compacting */
1088 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0,
1097 /* Compact all zones within a node */
1098 static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
1103 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1105 zone = &pgdat->node_zones[zoneid];
1106 if (!populated_zone(zone))
1109 cc->nr_freepages = 0;
1110 cc->nr_migratepages = 0;
1112 INIT_LIST_HEAD(&cc->freepages);
1113 INIT_LIST_HEAD(&cc->migratepages);
1115 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
1116 compact_zone(zone, cc);
1118 if (cc->order > 0) {
1119 int ok = zone_watermark_ok(zone, cc->order,
1120 low_wmark_pages(zone), 0, 0);
1121 if (ok && cc->order >= zone->compact_order_failed)
1122 zone->compact_order_failed = cc->order + 1;
1123 /* Currently async compaction is never deferred. */
1124 else if (!ok && cc->sync)
1125 defer_compaction(zone, cc->order);
1128 VM_BUG_ON(!list_empty(&cc->freepages));
1129 VM_BUG_ON(!list_empty(&cc->migratepages));
1133 void compact_pgdat(pg_data_t *pgdat, int order)
1135 struct compact_control cc = {
1143 __compact_pgdat(pgdat, &cc);
1146 static void compact_node(int nid)
1148 struct compact_control cc = {
1153 __compact_pgdat(NODE_DATA(nid), &cc);
1156 /* Compact all nodes in the system */
1157 static void compact_nodes(void)
1161 /* Flush pending updates to the LRU lists */
1162 lru_add_drain_all();
1164 for_each_online_node(nid)
1168 /* The written value is actually unused, all memory is compacted */
1169 int sysctl_compact_memory;
1171 /* This is the entry point for compacting all nodes via /proc/sys/vm */
1172 int sysctl_compaction_handler(struct ctl_table *table, int write,
1173 void __user *buffer, size_t *length, loff_t *ppos)
1181 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1182 void __user *buffer, size_t *length, loff_t *ppos)
1184 proc_dointvec_minmax(table, write, buffer, length, ppos);
1189 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1190 ssize_t sysfs_compact_node(struct device *dev,
1191 struct device_attribute *attr,
1192 const char *buf, size_t count)
1196 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1197 /* Flush pending updates to the LRU lists */
1198 lru_add_drain_all();
1205 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1207 int compaction_register_node(struct node *node)
1209 return device_create_file(&node->dev, &dev_attr_compact);
1212 void compaction_unregister_node(struct node *node)
1214 return device_remove_file(&node->dev, &dev_attr_compact);
1216 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1218 #endif /* CONFIG_COMPACTION */