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;
254 unsigned long nr_strict_required = end_pfn - blockpfn;
258 cursor = pfn_to_page(blockpfn);
260 /* Isolate free pages. */
261 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
263 struct page *page = cursor;
266 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 if (!isolated && strict)
298 total_isolated += isolated;
299 for (i = 0; i < isolated; i++) {
300 list_add(&page->lru, freelist);
304 /* If a page was split, advance to the end of it */
306 blockpfn += isolated - 1;
307 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 && nr_strict_required > total_isolated)
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 isolate_mode_t mode = 0;
458 struct lruvec *lruvec;
461 struct page *page = NULL, *valid_page = NULL;
462 bool skipped_async_unsuitable = false;
465 * Ensure that there are not too many pages isolated from the LRU
466 * list by either parallel reclaimers or compaction. If there are,
467 * delay for some time until fewer pages are isolated
469 while (unlikely(too_many_isolated(zone))) {
470 /* async migration should just abort */
474 congestion_wait(BLK_RW_ASYNC, HZ/10);
476 if (fatal_signal_pending(current))
480 /* Time to isolate some pages for migration */
482 for (; low_pfn < end_pfn; low_pfn++) {
483 /* give a chance to irqs before checking need_resched() */
484 if (locked && !((low_pfn+1) % SWAP_CLUSTER_MAX)) {
485 if (should_release_lock(&zone->lru_lock)) {
486 spin_unlock_irqrestore(&zone->lru_lock, flags);
492 * migrate_pfn does not necessarily start aligned to a
493 * pageblock. Ensure that pfn_valid is called when moving
494 * into a new MAX_ORDER_NR_PAGES range in case of large
495 * memory holes within the zone
497 if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
498 if (!pfn_valid(low_pfn)) {
499 low_pfn += MAX_ORDER_NR_PAGES - 1;
504 if (!pfn_valid_within(low_pfn))
509 * Get the page and ensure the page is within the same zone.
510 * See the comment in isolate_freepages about overlapping
511 * nodes. It is deliberate that the new zone lock is not taken
512 * as memory compaction should not move pages between nodes.
514 page = pfn_to_page(low_pfn);
515 if (page_zone(page) != zone)
521 /* If isolation recently failed, do not retry */
522 pageblock_nr = low_pfn >> pageblock_order;
523 if (!isolation_suitable(cc, page))
531 * For async migration, also only scan in MOVABLE blocks. Async
532 * migration is optimistic to see if the minimum amount of work
533 * satisfies the allocation
535 if (!cc->sync && last_pageblock_nr != pageblock_nr &&
536 !migrate_async_suitable(get_pageblock_migratetype(page))) {
537 cc->finished_update_migrate = true;
538 skipped_async_unsuitable = true;
543 * Check may be lockless but that's ok as we recheck later.
544 * It's possible to migrate LRU pages and balloon pages
545 * Skip any other type of page
547 if (!PageLRU(page)) {
548 if (unlikely(balloon_page_movable(page))) {
549 if (locked && balloon_page_isolate(page)) {
550 /* Successfully isolated */
551 cc->finished_update_migrate = true;
552 list_add(&page->lru, migratelist);
553 cc->nr_migratepages++;
555 goto check_compact_cluster;
562 * PageLRU is set. lru_lock normally excludes isolation
563 * splitting and collapsing (collapsing has already happened
564 * if PageLRU is set) but the lock is not necessarily taken
565 * here and it is wasteful to take it just to check transhuge.
566 * Check TransHuge without lock and skip the whole pageblock if
567 * it's either a transhuge or hugetlbfs page, as calling
568 * compound_order() without preventing THP from splitting the
569 * page underneath us may return surprising results.
571 if (PageTransHuge(page)) {
574 low_pfn += (1 << compound_order(page)) - 1;
578 /* Check if it is ok to still hold the lock */
579 locked = compact_checklock_irqsave(&zone->lru_lock, &flags,
581 if (!locked || fatal_signal_pending(current))
584 /* Recheck PageLRU and PageTransHuge under lock */
587 if (PageTransHuge(page)) {
588 low_pfn += (1 << compound_order(page)) - 1;
593 mode |= ISOLATE_ASYNC_MIGRATE;
596 mode |= ISOLATE_UNEVICTABLE;
598 lruvec = mem_cgroup_page_lruvec(page, zone);
600 /* Try isolate the page */
601 if (__isolate_lru_page(page, mode) != 0)
604 VM_BUG_ON(PageTransCompound(page));
606 /* Successfully isolated */
607 cc->finished_update_migrate = true;
608 del_page_from_lru_list(page, lruvec, page_lru(page));
609 list_add(&page->lru, migratelist);
610 cc->nr_migratepages++;
613 check_compact_cluster:
614 /* Avoid isolating too much */
615 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
623 low_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages) - 1;
624 last_pageblock_nr = pageblock_nr;
627 acct_isolated(zone, locked, cc);
630 spin_unlock_irqrestore(&zone->lru_lock, flags);
633 * Update the pageblock-skip information and cached scanner pfn,
634 * if the whole pageblock was scanned without isolating any page.
635 * This is not done when pageblock was skipped due to being unsuitable
636 * for async compaction, so that eventual sync compaction can try.
638 if (low_pfn == end_pfn && !skipped_async_unsuitable)
639 update_pageblock_skip(cc, valid_page, nr_isolated, true);
641 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
643 count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
645 count_compact_events(COMPACTISOLATED, nr_isolated);
650 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
651 #ifdef CONFIG_COMPACTION
653 * Based on information in the current compact_control, find blocks
654 * suitable for isolating free pages from and then isolate them.
656 static void isolate_freepages(struct zone *zone,
657 struct compact_control *cc)
660 unsigned long high_pfn, low_pfn, pfn, z_end_pfn, end_pfn;
661 int nr_freepages = cc->nr_freepages;
662 struct list_head *freelist = &cc->freepages;
665 * Initialise the free scanner. The starting point is where we last
666 * scanned from (or the end of the zone if starting). The low point
667 * is the end of the pageblock the migration scanner is using.
670 low_pfn = ALIGN(cc->migrate_pfn + 1, pageblock_nr_pages);
673 * Take care that if the migration scanner is at the end of the zone
674 * that the free scanner does not accidentally move to the next zone
675 * in the next isolation cycle.
677 high_pfn = min(low_pfn, pfn);
679 z_end_pfn = zone_end_pfn(zone);
682 * Isolate free pages until enough are available to migrate the
683 * pages on cc->migratepages. We stop searching if the migrate
684 * and free page scanners meet or enough free pages are isolated.
686 for (; pfn >= low_pfn && cc->nr_migratepages > nr_freepages;
687 pfn -= pageblock_nr_pages) {
688 unsigned long isolated;
691 * This can iterate a massively long zone without finding any
692 * suitable migration targets, so periodically check if we need
701 * Check for overlapping nodes/zones. It's possible on some
702 * configurations to have a setup like
704 * i.e. it's possible that all pages within a zones range of
705 * pages do not belong to a single zone.
707 page = pfn_to_page(pfn);
708 if (page_zone(page) != zone)
711 /* Check the block is suitable for migration */
712 if (!suitable_migration_target(page))
715 /* If isolation recently failed, do not retry */
716 if (!isolation_suitable(cc, page))
719 /* Found a block suitable for isolating free pages from */
723 * As pfn may not start aligned, pfn+pageblock_nr_page
724 * may cross a MAX_ORDER_NR_PAGES boundary and miss
725 * a pfn_valid check. Ensure isolate_freepages_block()
726 * only scans within a pageblock
728 end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
729 end_pfn = min(end_pfn, z_end_pfn);
730 isolated = isolate_freepages_block(cc, pfn, end_pfn,
732 nr_freepages += isolated;
735 * Record the highest PFN we isolated pages from. When next
736 * looking for free pages, the search will restart here as
737 * page migration may have returned some pages to the allocator
740 cc->finished_update_free = true;
741 high_pfn = max(high_pfn, pfn);
745 /* split_free_page does not map the pages */
749 * If we crossed the migrate scanner, we want to keep it that way
750 * so that compact_finished() may detect this
753 cc->free_pfn = max(pfn, zone->zone_start_pfn);
755 cc->free_pfn = high_pfn;
756 cc->nr_freepages = nr_freepages;
760 * This is a migrate-callback that "allocates" freepages by taking pages
761 * from the isolated freelists in the block we are migrating to.
763 static struct page *compaction_alloc(struct page *migratepage,
767 struct compact_control *cc = (struct compact_control *)data;
768 struct page *freepage;
770 /* Isolate free pages if necessary */
771 if (list_empty(&cc->freepages)) {
772 isolate_freepages(cc->zone, cc);
774 if (list_empty(&cc->freepages))
778 freepage = list_entry(cc->freepages.next, struct page, lru);
779 list_del(&freepage->lru);
786 * We cannot control nr_migratepages and nr_freepages fully when migration is
787 * running as migrate_pages() has no knowledge of compact_control. When
788 * migration is complete, we count the number of pages on the lists by hand.
790 static void update_nr_listpages(struct compact_control *cc)
792 int nr_migratepages = 0;
793 int nr_freepages = 0;
796 list_for_each_entry(page, &cc->migratepages, lru)
798 list_for_each_entry(page, &cc->freepages, lru)
801 cc->nr_migratepages = nr_migratepages;
802 cc->nr_freepages = nr_freepages;
805 /* possible outcome of isolate_migratepages */
807 ISOLATE_ABORT, /* Abort compaction now */
808 ISOLATE_NONE, /* No pages isolated, continue scanning */
809 ISOLATE_SUCCESS, /* Pages isolated, migrate */
813 * Isolate all pages that can be migrated from the block pointed to by
814 * the migrate scanner within compact_control.
816 static isolate_migrate_t isolate_migratepages(struct zone *zone,
817 struct compact_control *cc)
819 unsigned long low_pfn, end_pfn;
821 /* Do not scan outside zone boundaries */
822 low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
824 /* Only scan within a pageblock boundary */
825 end_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages);
827 /* Do not cross the free scanner or scan within a memory hole */
828 if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
829 cc->migrate_pfn = end_pfn;
833 /* Perform the isolation */
834 low_pfn = isolate_migratepages_range(zone, cc, low_pfn, end_pfn, false);
835 if (!low_pfn || cc->contended)
836 return ISOLATE_ABORT;
838 cc->migrate_pfn = low_pfn;
840 return ISOLATE_SUCCESS;
843 static int compact_finished(struct zone *zone,
844 struct compact_control *cc)
847 unsigned long watermark;
849 if (fatal_signal_pending(current))
850 return COMPACT_PARTIAL;
852 /* Compaction run completes if the migrate and free scanner meet */
853 if (cc->free_pfn <= cc->migrate_pfn) {
854 /* Let the next compaction start anew. */
855 zone->compact_cached_migrate_pfn = zone->zone_start_pfn;
856 zone->compact_cached_free_pfn = zone_end_pfn(zone);
859 * Mark that the PG_migrate_skip information should be cleared
860 * by kswapd when it goes to sleep. kswapd does not set the
861 * flag itself as the decision to be clear should be directly
862 * based on an allocation request.
864 if (!current_is_kswapd())
865 zone->compact_blockskip_flush = true;
867 return COMPACT_COMPLETE;
871 * order == -1 is expected when compacting via
872 * /proc/sys/vm/compact_memory
875 return COMPACT_CONTINUE;
877 /* Compaction run is not finished if the watermark is not met */
878 watermark = low_wmark_pages(zone);
879 watermark += (1 << cc->order);
881 if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
882 return COMPACT_CONTINUE;
884 /* Direct compactor: Is a suitable page free? */
885 for (order = cc->order; order < MAX_ORDER; order++) {
886 struct free_area *area = &zone->free_area[order];
888 /* Job done if page is free of the right migratetype */
889 if (!list_empty(&area->free_list[cc->migratetype]))
890 return COMPACT_PARTIAL;
892 /* Job done if allocation would set block type */
893 if (cc->order >= pageblock_order && area->nr_free)
894 return COMPACT_PARTIAL;
897 return COMPACT_CONTINUE;
901 * compaction_suitable: Is this suitable to run compaction on this zone now?
903 * COMPACT_SKIPPED - If there are too few free pages for compaction
904 * COMPACT_PARTIAL - If the allocation would succeed without compaction
905 * COMPACT_CONTINUE - If compaction should run now
907 unsigned long compaction_suitable(struct zone *zone, int order)
910 unsigned long watermark;
913 * order == -1 is expected when compacting via
914 * /proc/sys/vm/compact_memory
917 return COMPACT_CONTINUE;
920 * Watermarks for order-0 must be met for compaction. Note the 2UL.
921 * This is because during migration, copies of pages need to be
922 * allocated and for a short time, the footprint is higher
924 watermark = low_wmark_pages(zone) + (2UL << order);
925 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
926 return COMPACT_SKIPPED;
929 * fragmentation index determines if allocation failures are due to
930 * low memory or external fragmentation
932 * index of -1000 implies allocations might succeed depending on
934 * index towards 0 implies failure is due to lack of memory
935 * index towards 1000 implies failure is due to fragmentation
937 * Only compact if a failure would be due to fragmentation.
939 fragindex = fragmentation_index(zone, order);
940 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
941 return COMPACT_SKIPPED;
943 if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
945 return COMPACT_PARTIAL;
947 return COMPACT_CONTINUE;
950 static int compact_zone(struct zone *zone, struct compact_control *cc)
953 unsigned long start_pfn = zone->zone_start_pfn;
954 unsigned long end_pfn = zone_end_pfn(zone);
956 ret = compaction_suitable(zone, cc->order);
958 case COMPACT_PARTIAL:
959 case COMPACT_SKIPPED:
960 /* Compaction is likely to fail */
962 case COMPACT_CONTINUE:
963 /* Fall through to compaction */
968 * Clear pageblock skip if there were failures recently and compaction
969 * is about to be retried after being deferred. kswapd does not do
970 * this reset as it'll reset the cached information when going to sleep.
972 if (compaction_restarting(zone, cc->order) && !current_is_kswapd())
973 __reset_isolation_suitable(zone);
976 * Setup to move all movable pages to the end of the zone. Used cached
977 * information on where the scanners should start but check that it
978 * is initialised by ensuring the values are within zone boundaries.
980 cc->migrate_pfn = zone->compact_cached_migrate_pfn;
981 cc->free_pfn = zone->compact_cached_free_pfn;
982 if (cc->free_pfn < start_pfn || cc->free_pfn > end_pfn) {
983 cc->free_pfn = end_pfn & ~(pageblock_nr_pages-1);
984 zone->compact_cached_free_pfn = cc->free_pfn;
986 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn > end_pfn) {
987 cc->migrate_pfn = start_pfn;
988 zone->compact_cached_migrate_pfn = cc->migrate_pfn;
991 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn, cc->free_pfn, end_pfn);
993 migrate_prep_local();
995 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
996 unsigned long nr_migrate, nr_remaining;
999 switch (isolate_migratepages(zone, cc)) {
1001 ret = COMPACT_PARTIAL;
1002 putback_movable_pages(&cc->migratepages);
1003 cc->nr_migratepages = 0;
1007 case ISOLATE_SUCCESS:
1011 nr_migrate = cc->nr_migratepages;
1012 err = migrate_pages(&cc->migratepages, compaction_alloc,
1014 cc->sync ? MIGRATE_SYNC_LIGHT : MIGRATE_ASYNC,
1016 update_nr_listpages(cc);
1017 nr_remaining = cc->nr_migratepages;
1019 trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
1022 /* Release isolated pages not migrated */
1024 putback_movable_pages(&cc->migratepages);
1025 cc->nr_migratepages = 0;
1027 * migrate_pages() may return -ENOMEM when scanners meet
1028 * and we want compact_finished() to detect it
1030 if (err == -ENOMEM && cc->free_pfn > cc->migrate_pfn) {
1031 ret = COMPACT_PARTIAL;
1038 /* Release free pages and check accounting */
1039 cc->nr_freepages -= release_freepages(&cc->freepages);
1040 VM_BUG_ON(cc->nr_freepages != 0);
1042 trace_mm_compaction_end(ret);
1047 static unsigned long compact_zone_order(struct zone *zone,
1048 int order, gfp_t gfp_mask,
1049 bool sync, bool *contended)
1052 struct compact_control cc = {
1054 .nr_migratepages = 0,
1056 .migratetype = allocflags_to_migratetype(gfp_mask),
1060 INIT_LIST_HEAD(&cc.freepages);
1061 INIT_LIST_HEAD(&cc.migratepages);
1063 ret = compact_zone(zone, &cc);
1065 VM_BUG_ON(!list_empty(&cc.freepages));
1066 VM_BUG_ON(!list_empty(&cc.migratepages));
1068 *contended = cc.contended;
1072 int sysctl_extfrag_threshold = 500;
1075 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1076 * @zonelist: The zonelist used for the current allocation
1077 * @order: The order of the current allocation
1078 * @gfp_mask: The GFP mask of the current allocation
1079 * @nodemask: The allowed nodes to allocate from
1080 * @sync: Whether migration is synchronous or not
1081 * @contended: Return value that is true if compaction was aborted due to lock contention
1082 * @page: Optionally capture a free page of the requested order during compaction
1084 * This is the main entry point for direct page compaction.
1086 unsigned long try_to_compact_pages(struct zonelist *zonelist,
1087 int order, gfp_t gfp_mask, nodemask_t *nodemask,
1088 bool sync, bool *contended)
1090 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1091 int may_enter_fs = gfp_mask & __GFP_FS;
1092 int may_perform_io = gfp_mask & __GFP_IO;
1095 int rc = COMPACT_SKIPPED;
1096 int alloc_flags = 0;
1098 /* Check if the GFP flags allow compaction */
1099 if (!order || !may_enter_fs || !may_perform_io)
1102 count_compact_event(COMPACTSTALL);
1105 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
1106 alloc_flags |= ALLOC_CMA;
1108 /* Compact each zone in the list */
1109 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1113 status = compact_zone_order(zone, order, gfp_mask, sync,
1115 rc = max(status, rc);
1117 /* If a normal allocation would succeed, stop compacting */
1118 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0,
1127 /* Compact all zones within a node */
1128 static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
1133 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1135 zone = &pgdat->node_zones[zoneid];
1136 if (!populated_zone(zone))
1139 cc->nr_freepages = 0;
1140 cc->nr_migratepages = 0;
1142 INIT_LIST_HEAD(&cc->freepages);
1143 INIT_LIST_HEAD(&cc->migratepages);
1145 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
1146 compact_zone(zone, cc);
1148 if (cc->order > 0) {
1149 if (zone_watermark_ok(zone, cc->order,
1150 low_wmark_pages(zone), 0, 0))
1151 compaction_defer_reset(zone, cc->order, false);
1152 /* Currently async compaction is never deferred. */
1154 defer_compaction(zone, cc->order);
1157 VM_BUG_ON(!list_empty(&cc->freepages));
1158 VM_BUG_ON(!list_empty(&cc->migratepages));
1162 void compact_pgdat(pg_data_t *pgdat, int order)
1164 struct compact_control cc = {
1172 __compact_pgdat(pgdat, &cc);
1175 static void compact_node(int nid)
1177 struct compact_control cc = {
1182 __compact_pgdat(NODE_DATA(nid), &cc);
1185 /* Compact all nodes in the system */
1186 static void compact_nodes(void)
1190 /* Flush pending updates to the LRU lists */
1191 lru_add_drain_all();
1193 for_each_online_node(nid)
1197 /* The written value is actually unused, all memory is compacted */
1198 int sysctl_compact_memory;
1200 /* This is the entry point for compacting all nodes via /proc/sys/vm */
1201 int sysctl_compaction_handler(struct ctl_table *table, int write,
1202 void __user *buffer, size_t *length, loff_t *ppos)
1210 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1211 void __user *buffer, size_t *length, loff_t *ppos)
1213 proc_dointvec_minmax(table, write, buffer, length, ppos);
1218 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1219 ssize_t sysfs_compact_node(struct device *dev,
1220 struct device_attribute *attr,
1221 const char *buf, size_t count)
1225 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1226 /* Flush pending updates to the LRU lists */
1227 lru_add_drain_all();
1234 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1236 int compaction_register_node(struct node *node)
1238 return device_create_file(&node->dev, &dev_attr_compact);
1241 void compaction_unregister_node(struct node *node)
1243 return device_remove_file(&node->dev, &dev_attr_compact);
1245 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1247 #endif /* CONFIG_COMPACTION */