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 /* Returns true if the page is within a block suitable for migration to */
212 static bool suitable_migration_target(struct page *page)
214 /* If the page is a large free page, then disallow migration */
215 if (PageBuddy(page) && page_order(page) >= pageblock_order)
218 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
219 if (migrate_async_suitable(get_pageblock_migratetype(page)))
222 /* Otherwise skip the block */
227 * Isolate free pages onto a private freelist. If @strict is true, will abort
228 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
229 * (even though it may still end up isolating some pages).
231 static unsigned long isolate_freepages_block(struct compact_control *cc,
232 unsigned long blockpfn,
233 unsigned long end_pfn,
234 struct list_head *freelist,
237 int nr_scanned = 0, total_isolated = 0;
238 struct page *cursor, *valid_page = NULL;
241 bool checked_pageblock = false;
243 cursor = pfn_to_page(blockpfn);
245 /* Isolate free pages. */
246 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
248 struct page *page = cursor;
251 if (!pfn_valid_within(blockpfn))
256 if (!PageBuddy(page))
260 * The zone lock must be held to isolate freepages.
261 * Unfortunately this is a very coarse lock and can be
262 * heavily contended if there are parallel allocations
263 * or parallel compactions. For async compaction do not
264 * spin on the lock and we acquire the lock as late as
267 locked = compact_checklock_irqsave(&cc->zone->lock, &flags,
272 /* Recheck this is a suitable migration target under lock */
273 if (!strict && !checked_pageblock) {
275 * We need to check suitability of pageblock only once
276 * and this isolate_freepages_block() is called with
277 * pageblock range, so just check once is sufficient.
279 checked_pageblock = true;
280 if (!suitable_migration_target(page))
284 /* Recheck this is a buddy page under lock */
285 if (!PageBuddy(page))
288 /* Found a free page, break it into order-0 pages */
289 isolated = split_free_page(page);
290 total_isolated += isolated;
291 for (i = 0; i < isolated; i++) {
292 list_add(&page->lru, freelist);
296 /* If a page was split, advance to the end of it */
298 blockpfn += isolated - 1;
299 cursor += isolated - 1;
311 trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
314 * If strict isolation is requested by CMA then check that all the
315 * pages requested were isolated. If there were any failures, 0 is
316 * returned and CMA will fail.
318 if (strict && blockpfn < end_pfn)
322 spin_unlock_irqrestore(&cc->zone->lock, flags);
324 /* Update the pageblock-skip if the whole pageblock was scanned */
325 if (blockpfn == end_pfn)
326 update_pageblock_skip(cc, valid_page, total_isolated, false);
328 count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
330 count_compact_events(COMPACTISOLATED, total_isolated);
331 return total_isolated;
335 * isolate_freepages_range() - isolate free pages.
336 * @start_pfn: The first PFN to start isolating.
337 * @end_pfn: The one-past-last PFN.
339 * Non-free pages, invalid PFNs, or zone boundaries within the
340 * [start_pfn, end_pfn) range are considered errors, cause function to
341 * undo its actions and return zero.
343 * Otherwise, function returns one-past-the-last PFN of isolated page
344 * (which may be greater then end_pfn if end fell in a middle of
348 isolate_freepages_range(struct compact_control *cc,
349 unsigned long start_pfn, unsigned long end_pfn)
351 unsigned long isolated, pfn, block_end_pfn;
354 for (pfn = start_pfn; pfn < end_pfn; pfn += isolated) {
355 if (!pfn_valid(pfn) || cc->zone != page_zone(pfn_to_page(pfn)))
359 * On subsequent iterations ALIGN() is actually not needed,
360 * but we keep it that we not to complicate the code.
362 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
363 block_end_pfn = min(block_end_pfn, end_pfn);
365 isolated = isolate_freepages_block(cc, pfn, block_end_pfn,
369 * In strict mode, isolate_freepages_block() returns 0 if
370 * there are any holes in the block (ie. invalid PFNs or
377 * If we managed to isolate pages, it is always (1 << n) *
378 * pageblock_nr_pages for some non-negative n. (Max order
379 * page may span two pageblocks).
383 /* split_free_page does not map the pages */
384 map_pages(&freelist);
387 /* Loop terminated early, cleanup. */
388 release_freepages(&freelist);
392 /* We don't use freelists for anything. */
396 /* Update the number of anon and file isolated pages in the zone */
397 static void acct_isolated(struct zone *zone, bool locked, struct compact_control *cc)
400 unsigned int count[2] = { 0, };
402 list_for_each_entry(page, &cc->migratepages, lru)
403 count[!!page_is_file_cache(page)]++;
405 /* If locked we can use the interrupt unsafe versions */
407 __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
408 __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
410 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
411 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
415 /* Similar to reclaim, but different enough that they don't share logic */
416 static bool too_many_isolated(struct zone *zone)
418 unsigned long active, inactive, isolated;
420 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
421 zone_page_state(zone, NR_INACTIVE_ANON);
422 active = zone_page_state(zone, NR_ACTIVE_FILE) +
423 zone_page_state(zone, NR_ACTIVE_ANON);
424 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
425 zone_page_state(zone, NR_ISOLATED_ANON);
427 return isolated > (inactive + active) / 2;
431 * isolate_migratepages_range() - isolate all migrate-able pages in range.
432 * @zone: Zone pages are in.
433 * @cc: Compaction control structure.
434 * @low_pfn: The first PFN of the range.
435 * @end_pfn: The one-past-the-last PFN of the range.
436 * @unevictable: true if it allows to isolate unevictable pages
438 * Isolate all pages that can be migrated from the range specified by
439 * [low_pfn, end_pfn). Returns zero if there is a fatal signal
440 * pending), otherwise PFN of the first page that was not scanned
441 * (which may be both less, equal to or more then end_pfn).
443 * Assumes that cc->migratepages is empty and cc->nr_migratepages is
446 * Apart from cc->migratepages and cc->nr_migratetypes this function
447 * does not modify any cc's fields, in particular it does not modify
448 * (or read for that matter) cc->migrate_pfn.
451 isolate_migratepages_range(struct zone *zone, struct compact_control *cc,
452 unsigned long low_pfn, unsigned long end_pfn, bool unevictable)
454 unsigned long last_pageblock_nr = 0, pageblock_nr;
455 unsigned long nr_scanned = 0, nr_isolated = 0;
456 struct list_head *migratelist = &cc->migratepages;
457 struct lruvec *lruvec;
460 struct page *page = NULL, *valid_page = NULL;
461 bool skipped_async_unsuitable = false;
462 const isolate_mode_t mode = (!cc->sync ? ISOLATE_ASYNC_MIGRATE : 0) |
463 (unevictable ? ISOLATE_UNEVICTABLE : 0);
466 * Ensure that there are not too many pages isolated from the LRU
467 * list by either parallel reclaimers or compaction. If there are,
468 * delay for some time until fewer pages are isolated
470 while (unlikely(too_many_isolated(zone))) {
471 /* async migration should just abort */
475 congestion_wait(BLK_RW_ASYNC, HZ/10);
477 if (fatal_signal_pending(current))
481 /* Time to isolate some pages for migration */
483 for (; low_pfn < end_pfn; low_pfn++) {
484 /* give a chance to irqs before checking need_resched() */
485 if (locked && !(low_pfn % SWAP_CLUSTER_MAX)) {
486 if (should_release_lock(&zone->lru_lock)) {
487 spin_unlock_irqrestore(&zone->lru_lock, flags);
493 * migrate_pfn does not necessarily start aligned to a
494 * pageblock. Ensure that pfn_valid is called when moving
495 * into a new MAX_ORDER_NR_PAGES range in case of large
496 * memory holes within the zone
498 if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
499 if (!pfn_valid(low_pfn)) {
500 low_pfn += MAX_ORDER_NR_PAGES - 1;
505 if (!pfn_valid_within(low_pfn))
510 * Get the page and ensure the page is within the same zone.
511 * See the comment in isolate_freepages about overlapping
512 * nodes. It is deliberate that the new zone lock is not taken
513 * as memory compaction should not move pages between nodes.
515 page = pfn_to_page(low_pfn);
516 if (page_zone(page) != zone)
522 /* If isolation recently failed, do not retry */
523 pageblock_nr = low_pfn >> pageblock_order;
524 if (last_pageblock_nr != pageblock_nr) {
527 last_pageblock_nr = pageblock_nr;
528 if (!isolation_suitable(cc, page))
532 * For async migration, also only scan in MOVABLE
533 * blocks. Async migration is optimistic to see if
534 * the minimum amount of work satisfies the allocation
536 mt = get_pageblock_migratetype(page);
537 if (!cc->sync && !migrate_async_suitable(mt)) {
538 cc->finished_update_migrate = true;
539 skipped_async_unsuitable = true;
545 * Skip if free. page_order cannot be used without zone->lock
546 * as nothing prevents parallel allocations or buddy merging.
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 goto isolate_success;
567 * PageLRU is set. lru_lock normally excludes isolation
568 * splitting and collapsing (collapsing has already happened
569 * if PageLRU is set) but the lock is not necessarily taken
570 * here and it is wasteful to take it just to check transhuge.
571 * Check TransHuge without lock and skip the whole pageblock if
572 * it's either a transhuge or hugetlbfs page, as calling
573 * compound_order() without preventing THP from splitting the
574 * page underneath us may return surprising results.
576 if (PageTransHuge(page)) {
579 low_pfn += (1 << compound_order(page)) - 1;
584 * Migration will fail if an anonymous page is pinned in memory,
585 * so avoid taking lru_lock and isolating it unnecessarily in an
586 * admittedly racy check.
588 if (!page_mapping(page) &&
589 page_count(page) > page_mapcount(page))
592 /* Check if it is ok to still hold the lock */
593 locked = compact_checklock_irqsave(&zone->lru_lock, &flags,
595 if (!locked || fatal_signal_pending(current))
598 /* Recheck PageLRU and PageTransHuge under lock */
601 if (PageTransHuge(page)) {
602 low_pfn += (1 << compound_order(page)) - 1;
606 lruvec = mem_cgroup_page_lruvec(page, zone);
608 /* Try isolate the page */
609 if (__isolate_lru_page(page, mode) != 0)
612 VM_BUG_ON_PAGE(PageTransCompound(page), page);
614 /* Successfully isolated */
615 del_page_from_lru_list(page, lruvec, page_lru(page));
618 cc->finished_update_migrate = true;
619 list_add(&page->lru, migratelist);
620 cc->nr_migratepages++;
623 /* Avoid isolating too much */
624 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
632 low_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages) - 1;
635 acct_isolated(zone, locked, cc);
638 spin_unlock_irqrestore(&zone->lru_lock, flags);
641 * Update the pageblock-skip information and cached scanner pfn,
642 * if the whole pageblock was scanned without isolating any page.
643 * This is not done when pageblock was skipped due to being unsuitable
644 * for async compaction, so that eventual sync compaction can try.
646 if (low_pfn == end_pfn && !skipped_async_unsuitable)
647 update_pageblock_skip(cc, valid_page, nr_isolated, true);
649 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
651 count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
653 count_compact_events(COMPACTISOLATED, nr_isolated);
658 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
659 #ifdef CONFIG_COMPACTION
661 * Based on information in the current compact_control, find blocks
662 * suitable for isolating free pages from and then isolate them.
664 static void isolate_freepages(struct zone *zone,
665 struct compact_control *cc)
668 unsigned long block_start_pfn; /* start of current pageblock */
669 unsigned long block_end_pfn; /* end of current pageblock */
670 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
671 unsigned long next_free_pfn; /* start pfn for scaning at next round */
672 int nr_freepages = cc->nr_freepages;
673 struct list_head *freelist = &cc->freepages;
676 * Initialise the free scanner. The starting point is where we last
677 * successfully isolated from, zone-cached value, or the end of the
678 * zone when isolating for the first time. We need this aligned to
679 * the pageblock boundary, because we do
680 * block_start_pfn -= pageblock_nr_pages in the for loop.
681 * For ending point, take care when isolating in last pageblock of a
682 * a zone which ends in the middle of a pageblock.
683 * The low boundary is the end of the pageblock the migration scanner
686 block_start_pfn = cc->free_pfn & ~(pageblock_nr_pages-1);
687 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
689 low_pfn = ALIGN(cc->migrate_pfn + 1, pageblock_nr_pages);
692 * If no pages are isolated, the block_start_pfn < low_pfn check
698 * Isolate free pages until enough are available to migrate the
699 * pages on cc->migratepages. We stop searching if the migrate
700 * and free page scanners meet or enough free pages are isolated.
702 for (; block_start_pfn >= low_pfn && cc->nr_migratepages > nr_freepages;
703 block_end_pfn = block_start_pfn,
704 block_start_pfn -= pageblock_nr_pages) {
705 unsigned long isolated;
708 * This can iterate a massively long zone without finding any
709 * suitable migration targets, so periodically check if we need
714 if (!pfn_valid(block_start_pfn))
718 * Check for overlapping nodes/zones. It's possible on some
719 * configurations to have a setup like
721 * i.e. it's possible that all pages within a zones range of
722 * pages do not belong to a single zone.
724 page = pfn_to_page(block_start_pfn);
725 if (page_zone(page) != zone)
728 /* Check the block is suitable for migration */
729 if (!suitable_migration_target(page))
732 /* If isolation recently failed, do not retry */
733 if (!isolation_suitable(cc, page))
736 /* Found a block suitable for isolating free pages from */
737 isolated = isolate_freepages_block(cc, block_start_pfn,
738 block_end_pfn, freelist, false);
739 nr_freepages += isolated;
742 * Record the highest PFN we isolated pages from. When next
743 * looking for free pages, the search will restart here as
744 * page migration may have returned some pages to the allocator
746 if (isolated && next_free_pfn == 0) {
747 cc->finished_update_free = true;
748 next_free_pfn = block_start_pfn;
752 /* split_free_page does not map the pages */
756 * If we crossed the migrate scanner, we want to keep it that way
757 * so that compact_finished() may detect this
759 if (block_start_pfn < low_pfn)
760 next_free_pfn = cc->migrate_pfn;
762 cc->free_pfn = next_free_pfn;
763 cc->nr_freepages = nr_freepages;
767 * This is a migrate-callback that "allocates" freepages by taking pages
768 * from the isolated freelists in the block we are migrating to.
770 static struct page *compaction_alloc(struct page *migratepage,
774 struct compact_control *cc = (struct compact_control *)data;
775 struct page *freepage;
777 /* Isolate free pages if necessary */
778 if (list_empty(&cc->freepages)) {
779 isolate_freepages(cc->zone, cc);
781 if (list_empty(&cc->freepages))
785 freepage = list_entry(cc->freepages.next, struct page, lru);
786 list_del(&freepage->lru);
793 * We cannot control nr_migratepages and nr_freepages fully when migration is
794 * running as migrate_pages() has no knowledge of compact_control. When
795 * migration is complete, we count the number of pages on the lists by hand.
797 static void update_nr_listpages(struct compact_control *cc)
799 int nr_migratepages = 0;
800 int nr_freepages = 0;
803 list_for_each_entry(page, &cc->migratepages, lru)
805 list_for_each_entry(page, &cc->freepages, lru)
808 cc->nr_migratepages = nr_migratepages;
809 cc->nr_freepages = nr_freepages;
812 /* possible outcome of isolate_migratepages */
814 ISOLATE_ABORT, /* Abort compaction now */
815 ISOLATE_NONE, /* No pages isolated, continue scanning */
816 ISOLATE_SUCCESS, /* Pages isolated, migrate */
820 * Isolate all pages that can be migrated from the block pointed to by
821 * the migrate scanner within compact_control.
823 static isolate_migrate_t isolate_migratepages(struct zone *zone,
824 struct compact_control *cc)
826 unsigned long low_pfn, end_pfn;
828 /* Do not scan outside zone boundaries */
829 low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
831 /* Only scan within a pageblock boundary */
832 end_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages);
834 /* Do not cross the free scanner or scan within a memory hole */
835 if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
836 cc->migrate_pfn = end_pfn;
840 /* Perform the isolation */
841 low_pfn = isolate_migratepages_range(zone, cc, low_pfn, end_pfn, false);
842 if (!low_pfn || cc->contended)
843 return ISOLATE_ABORT;
845 cc->migrate_pfn = low_pfn;
847 return ISOLATE_SUCCESS;
850 static int compact_finished(struct zone *zone,
851 struct compact_control *cc)
854 unsigned long watermark;
856 if (fatal_signal_pending(current))
857 return COMPACT_PARTIAL;
859 /* Compaction run completes if the migrate and free scanner meet */
860 if (cc->free_pfn <= cc->migrate_pfn) {
861 /* Let the next compaction start anew. */
862 zone->compact_cached_migrate_pfn = zone->zone_start_pfn;
863 zone->compact_cached_free_pfn = zone_end_pfn(zone);
866 * Mark that the PG_migrate_skip information should be cleared
867 * by kswapd when it goes to sleep. kswapd does not set the
868 * flag itself as the decision to be clear should be directly
869 * based on an allocation request.
871 if (!current_is_kswapd())
872 zone->compact_blockskip_flush = true;
874 return COMPACT_COMPLETE;
878 * order == -1 is expected when compacting via
879 * /proc/sys/vm/compact_memory
882 return COMPACT_CONTINUE;
884 /* Compaction run is not finished if the watermark is not met */
885 watermark = low_wmark_pages(zone);
886 watermark += (1 << cc->order);
888 if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
889 return COMPACT_CONTINUE;
891 /* Direct compactor: Is a suitable page free? */
892 for (order = cc->order; order < MAX_ORDER; order++) {
893 struct free_area *area = &zone->free_area[order];
895 /* Job done if page is free of the right migratetype */
896 if (!list_empty(&area->free_list[cc->migratetype]))
897 return COMPACT_PARTIAL;
899 /* Job done if allocation would set block type */
900 if (cc->order >= pageblock_order && area->nr_free)
901 return COMPACT_PARTIAL;
904 return COMPACT_CONTINUE;
908 * compaction_suitable: Is this suitable to run compaction on this zone now?
910 * COMPACT_SKIPPED - If there are too few free pages for compaction
911 * COMPACT_PARTIAL - If the allocation would succeed without compaction
912 * COMPACT_CONTINUE - If compaction should run now
914 unsigned long compaction_suitable(struct zone *zone, int order)
917 unsigned long watermark;
920 * order == -1 is expected when compacting via
921 * /proc/sys/vm/compact_memory
924 return COMPACT_CONTINUE;
927 * Watermarks for order-0 must be met for compaction. Note the 2UL.
928 * This is because during migration, copies of pages need to be
929 * allocated and for a short time, the footprint is higher
931 watermark = low_wmark_pages(zone) + (2UL << order);
932 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
933 return COMPACT_SKIPPED;
936 * fragmentation index determines if allocation failures are due to
937 * low memory or external fragmentation
939 * index of -1000 implies allocations might succeed depending on
941 * index towards 0 implies failure is due to lack of memory
942 * index towards 1000 implies failure is due to fragmentation
944 * Only compact if a failure would be due to fragmentation.
946 fragindex = fragmentation_index(zone, order);
947 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
948 return COMPACT_SKIPPED;
950 if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
952 return COMPACT_PARTIAL;
954 return COMPACT_CONTINUE;
957 static int compact_zone(struct zone *zone, struct compact_control *cc)
960 unsigned long start_pfn = zone->zone_start_pfn;
961 unsigned long end_pfn = zone_end_pfn(zone);
963 ret = compaction_suitable(zone, cc->order);
965 case COMPACT_PARTIAL:
966 case COMPACT_SKIPPED:
967 /* Compaction is likely to fail */
969 case COMPACT_CONTINUE:
970 /* Fall through to compaction */
975 * Clear pageblock skip if there were failures recently and compaction
976 * is about to be retried after being deferred. kswapd does not do
977 * this reset as it'll reset the cached information when going to sleep.
979 if (compaction_restarting(zone, cc->order) && !current_is_kswapd())
980 __reset_isolation_suitable(zone);
983 * Setup to move all movable pages to the end of the zone. Used cached
984 * information on where the scanners should start but check that it
985 * is initialised by ensuring the values are within zone boundaries.
987 cc->migrate_pfn = zone->compact_cached_migrate_pfn;
988 cc->free_pfn = zone->compact_cached_free_pfn;
989 if (cc->free_pfn < start_pfn || cc->free_pfn > end_pfn) {
990 cc->free_pfn = end_pfn & ~(pageblock_nr_pages-1);
991 zone->compact_cached_free_pfn = cc->free_pfn;
993 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn > end_pfn) {
994 cc->migrate_pfn = start_pfn;
995 zone->compact_cached_migrate_pfn = cc->migrate_pfn;
998 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn, cc->free_pfn, end_pfn);
1000 migrate_prep_local();
1002 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
1003 unsigned long nr_migrate, nr_remaining;
1006 switch (isolate_migratepages(zone, cc)) {
1008 ret = COMPACT_PARTIAL;
1009 putback_movable_pages(&cc->migratepages);
1010 cc->nr_migratepages = 0;
1014 case ISOLATE_SUCCESS:
1018 nr_migrate = cc->nr_migratepages;
1019 err = migrate_pages(&cc->migratepages, compaction_alloc,
1021 cc->sync ? MIGRATE_SYNC_LIGHT : MIGRATE_ASYNC,
1023 update_nr_listpages(cc);
1024 nr_remaining = cc->nr_migratepages;
1026 trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
1029 /* Release isolated pages not migrated */
1031 putback_movable_pages(&cc->migratepages);
1032 cc->nr_migratepages = 0;
1034 * migrate_pages() may return -ENOMEM when scanners meet
1035 * and we want compact_finished() to detect it
1037 if (err == -ENOMEM && cc->free_pfn > cc->migrate_pfn) {
1038 ret = COMPACT_PARTIAL;
1045 /* Release free pages and check accounting */
1046 cc->nr_freepages -= release_freepages(&cc->freepages);
1047 VM_BUG_ON(cc->nr_freepages != 0);
1049 trace_mm_compaction_end(ret);
1054 static unsigned long compact_zone_order(struct zone *zone,
1055 int order, gfp_t gfp_mask,
1056 bool sync, bool *contended)
1059 struct compact_control cc = {
1061 .nr_migratepages = 0,
1063 .migratetype = allocflags_to_migratetype(gfp_mask),
1067 INIT_LIST_HEAD(&cc.freepages);
1068 INIT_LIST_HEAD(&cc.migratepages);
1070 ret = compact_zone(zone, &cc);
1072 VM_BUG_ON(!list_empty(&cc.freepages));
1073 VM_BUG_ON(!list_empty(&cc.migratepages));
1075 *contended = cc.contended;
1079 int sysctl_extfrag_threshold = 500;
1082 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1083 * @zonelist: The zonelist used for the current allocation
1084 * @order: The order of the current allocation
1085 * @gfp_mask: The GFP mask of the current allocation
1086 * @nodemask: The allowed nodes to allocate from
1087 * @sync: Whether migration is synchronous or not
1088 * @contended: Return value that is true if compaction was aborted due to lock contention
1089 * @page: Optionally capture a free page of the requested order during compaction
1091 * This is the main entry point for direct page compaction.
1093 unsigned long try_to_compact_pages(struct zonelist *zonelist,
1094 int order, gfp_t gfp_mask, nodemask_t *nodemask,
1095 bool sync, bool *contended)
1097 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1098 int may_enter_fs = gfp_mask & __GFP_FS;
1099 int may_perform_io = gfp_mask & __GFP_IO;
1102 int rc = COMPACT_SKIPPED;
1103 int alloc_flags = 0;
1105 /* Check if the GFP flags allow compaction */
1106 if (!order || !may_enter_fs || !may_perform_io)
1109 count_compact_event(COMPACTSTALL);
1112 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
1113 alloc_flags |= ALLOC_CMA;
1115 /* Compact each zone in the list */
1116 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1120 status = compact_zone_order(zone, order, gfp_mask, sync,
1122 rc = max(status, rc);
1124 /* If a normal allocation would succeed, stop compacting */
1125 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0,
1134 /* Compact all zones within a node */
1135 static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
1140 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1142 zone = &pgdat->node_zones[zoneid];
1143 if (!populated_zone(zone))
1146 cc->nr_freepages = 0;
1147 cc->nr_migratepages = 0;
1149 INIT_LIST_HEAD(&cc->freepages);
1150 INIT_LIST_HEAD(&cc->migratepages);
1152 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
1153 compact_zone(zone, cc);
1155 if (cc->order > 0) {
1156 if (zone_watermark_ok(zone, cc->order,
1157 low_wmark_pages(zone), 0, 0))
1158 compaction_defer_reset(zone, cc->order, false);
1161 VM_BUG_ON(!list_empty(&cc->freepages));
1162 VM_BUG_ON(!list_empty(&cc->migratepages));
1166 void compact_pgdat(pg_data_t *pgdat, int order)
1168 struct compact_control cc = {
1176 __compact_pgdat(pgdat, &cc);
1179 static void compact_node(int nid)
1181 struct compact_control cc = {
1184 .ignore_skip_hint = true,
1187 __compact_pgdat(NODE_DATA(nid), &cc);
1190 /* Compact all nodes in the system */
1191 static void compact_nodes(void)
1195 /* Flush pending updates to the LRU lists */
1196 lru_add_drain_all();
1198 for_each_online_node(nid)
1202 /* The written value is actually unused, all memory is compacted */
1203 int sysctl_compact_memory;
1205 /* This is the entry point for compacting all nodes via /proc/sys/vm */
1206 int sysctl_compaction_handler(struct ctl_table *table, int write,
1207 void __user *buffer, size_t *length, loff_t *ppos)
1215 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1216 void __user *buffer, size_t *length, loff_t *ppos)
1218 proc_dointvec_minmax(table, write, buffer, length, ppos);
1223 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1224 ssize_t sysfs_compact_node(struct device *dev,
1225 struct device_attribute *attr,
1226 const char *buf, size_t count)
1230 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1231 /* Flush pending updates to the LRU lists */
1232 lru_add_drain_all();
1239 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1241 int compaction_register_node(struct node *node)
1243 return device_create_file(&node->dev, &dev_attr_compact);
1246 void compaction_unregister_node(struct node *node)
1248 return device_remove_file(&node->dev, &dev_attr_compact);
1250 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1252 #endif /* CONFIG_COMPACTION */