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;
71 * Check that the whole (or subset of) a pageblock given by the interval of
72 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
73 * with the migration of free compaction scanner. The scanners then need to
74 * use only pfn_valid_within() check for arches that allow holes within
77 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
79 * It's possible on some configurations to have a setup like node0 node1 node0
80 * i.e. it's possible that all pages within a zones range of pages do not
81 * belong to a single zone. We assume that a border between node0 and node1
82 * can occur within a single pageblock, but not a node0 node1 node0
83 * interleaving within a single pageblock. It is therefore sufficient to check
84 * the first and last page of a pageblock and avoid checking each individual
85 * page in a pageblock.
87 static struct page *pageblock_pfn_to_page(unsigned long start_pfn,
88 unsigned long end_pfn, struct zone *zone)
90 struct page *start_page;
91 struct page *end_page;
93 /* end_pfn is one past the range we are checking */
96 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
99 start_page = pfn_to_page(start_pfn);
101 if (page_zone(start_page) != zone)
104 end_page = pfn_to_page(end_pfn);
106 /* This gives a shorter code than deriving page_zone(end_page) */
107 if (page_zone_id(start_page) != page_zone_id(end_page))
113 #ifdef CONFIG_COMPACTION
114 /* Returns true if the pageblock should be scanned for pages to isolate. */
115 static inline bool isolation_suitable(struct compact_control *cc,
118 if (cc->ignore_skip_hint)
121 return !get_pageblock_skip(page);
125 * This function is called to clear all cached information on pageblocks that
126 * should be skipped for page isolation when the migrate and free page scanner
129 static void __reset_isolation_suitable(struct zone *zone)
131 unsigned long start_pfn = zone->zone_start_pfn;
132 unsigned long end_pfn = zone_end_pfn(zone);
135 zone->compact_cached_migrate_pfn[0] = start_pfn;
136 zone->compact_cached_migrate_pfn[1] = start_pfn;
137 zone->compact_cached_free_pfn = end_pfn;
138 zone->compact_blockskip_flush = false;
140 /* Walk the zone and mark every pageblock as suitable for isolation */
141 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
149 page = pfn_to_page(pfn);
150 if (zone != page_zone(page))
153 clear_pageblock_skip(page);
157 void reset_isolation_suitable(pg_data_t *pgdat)
161 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
162 struct zone *zone = &pgdat->node_zones[zoneid];
163 if (!populated_zone(zone))
166 /* Only flush if a full compaction finished recently */
167 if (zone->compact_blockskip_flush)
168 __reset_isolation_suitable(zone);
173 * If no pages were isolated then mark this pageblock to be skipped in the
174 * future. The information is later cleared by __reset_isolation_suitable().
176 static void update_pageblock_skip(struct compact_control *cc,
177 struct page *page, unsigned long nr_isolated,
178 bool migrate_scanner)
180 struct zone *zone = cc->zone;
183 if (cc->ignore_skip_hint)
192 set_pageblock_skip(page);
194 pfn = page_to_pfn(page);
196 /* Update where async and sync compaction should restart */
197 if (migrate_scanner) {
198 if (cc->finished_update_migrate)
200 if (pfn > zone->compact_cached_migrate_pfn[0])
201 zone->compact_cached_migrate_pfn[0] = pfn;
202 if (cc->mode != MIGRATE_ASYNC &&
203 pfn > zone->compact_cached_migrate_pfn[1])
204 zone->compact_cached_migrate_pfn[1] = pfn;
206 if (cc->finished_update_free)
208 if (pfn < zone->compact_cached_free_pfn)
209 zone->compact_cached_free_pfn = pfn;
213 static inline bool isolation_suitable(struct compact_control *cc,
219 static void update_pageblock_skip(struct compact_control *cc,
220 struct page *page, unsigned long nr_isolated,
221 bool migrate_scanner)
224 #endif /* CONFIG_COMPACTION */
227 * Compaction requires the taking of some coarse locks that are potentially
228 * very heavily contended. For async compaction, back out if the lock cannot
229 * be taken immediately. For sync compaction, spin on the lock if needed.
231 * Returns true if the lock is held
232 * Returns false if the lock is not held and compaction should abort
234 static bool compact_trylock_irqsave(spinlock_t *lock, unsigned long *flags,
235 struct compact_control *cc)
237 if (cc->mode == MIGRATE_ASYNC) {
238 if (!spin_trylock_irqsave(lock, *flags)) {
239 cc->contended = COMPACT_CONTENDED_LOCK;
243 spin_lock_irqsave(lock, *flags);
250 * Compaction requires the taking of some coarse locks that are potentially
251 * very heavily contended. The lock should be periodically unlocked to avoid
252 * having disabled IRQs for a long time, even when there is nobody waiting on
253 * the lock. It might also be that allowing the IRQs will result in
254 * need_resched() becoming true. If scheduling is needed, async compaction
255 * aborts. Sync compaction schedules.
256 * Either compaction type will also abort if a fatal signal is pending.
257 * In either case if the lock was locked, it is dropped and not regained.
259 * Returns true if compaction should abort due to fatal signal pending, or
260 * async compaction due to need_resched()
261 * Returns false when compaction can continue (sync compaction might have
264 static bool compact_unlock_should_abort(spinlock_t *lock,
265 unsigned long flags, bool *locked, struct compact_control *cc)
268 spin_unlock_irqrestore(lock, flags);
272 if (fatal_signal_pending(current)) {
273 cc->contended = COMPACT_CONTENDED_SCHED;
277 if (need_resched()) {
278 if (cc->mode == MIGRATE_ASYNC) {
279 cc->contended = COMPACT_CONTENDED_SCHED;
289 * Aside from avoiding lock contention, compaction also periodically checks
290 * need_resched() and either schedules in sync compaction or aborts async
291 * compaction. This is similar to what compact_unlock_should_abort() does, but
292 * is used where no lock is concerned.
294 * Returns false when no scheduling was needed, or sync compaction scheduled.
295 * Returns true when async compaction should abort.
297 static inline bool compact_should_abort(struct compact_control *cc)
299 /* async compaction aborts if contended */
300 if (need_resched()) {
301 if (cc->mode == MIGRATE_ASYNC) {
302 cc->contended = COMPACT_CONTENDED_SCHED;
312 /* Returns true if the page is within a block suitable for migration to */
313 static bool suitable_migration_target(struct page *page)
315 /* If the page is a large free page, then disallow migration */
316 if (PageBuddy(page) && page_order(page) >= pageblock_order)
319 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
320 if (migrate_async_suitable(get_pageblock_migratetype(page)))
323 /* Otherwise skip the block */
328 * Isolate free pages onto a private freelist. If @strict is true, will abort
329 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
330 * (even though it may still end up isolating some pages).
332 static unsigned long isolate_freepages_block(struct compact_control *cc,
333 unsigned long blockpfn,
334 unsigned long end_pfn,
335 struct list_head *freelist,
338 int nr_scanned = 0, total_isolated = 0;
339 struct page *cursor, *valid_page = NULL;
343 cursor = pfn_to_page(blockpfn);
345 /* Isolate free pages. */
346 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
348 struct page *page = cursor;
351 * Periodically drop the lock (if held) regardless of its
352 * contention, to give chance to IRQs. Abort if fatal signal
353 * pending or async compaction detects need_resched()
355 if (!(blockpfn % SWAP_CLUSTER_MAX)
356 && compact_unlock_should_abort(&cc->zone->lock, flags,
361 if (!pfn_valid_within(blockpfn))
366 if (!PageBuddy(page))
370 * The zone lock must be held to isolate freepages.
371 * Unfortunately this is a very coarse lock and can be
372 * heavily contended if there are parallel allocations
373 * or parallel compactions. For async compaction do not
374 * spin on the lock and we acquire the lock as late as
378 locked = compact_trylock_irqsave(&cc->zone->lock,
383 /* Recheck this is a buddy page under lock */
384 if (!PageBuddy(page))
387 /* Found a free page, break it into order-0 pages */
388 isolated = split_free_page(page);
389 total_isolated += isolated;
390 for (i = 0; i < isolated; i++) {
391 list_add(&page->lru, freelist);
395 /* If a page was split, advance to the end of it */
397 blockpfn += isolated - 1;
398 cursor += isolated - 1;
410 trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
413 * If strict isolation is requested by CMA then check that all the
414 * pages requested were isolated. If there were any failures, 0 is
415 * returned and CMA will fail.
417 if (strict && blockpfn < end_pfn)
421 spin_unlock_irqrestore(&cc->zone->lock, flags);
423 /* Update the pageblock-skip if the whole pageblock was scanned */
424 if (blockpfn == end_pfn)
425 update_pageblock_skip(cc, valid_page, total_isolated, false);
427 count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
429 count_compact_events(COMPACTISOLATED, total_isolated);
430 return total_isolated;
434 * isolate_freepages_range() - isolate free pages.
435 * @start_pfn: The first PFN to start isolating.
436 * @end_pfn: The one-past-last PFN.
438 * Non-free pages, invalid PFNs, or zone boundaries within the
439 * [start_pfn, end_pfn) range are considered errors, cause function to
440 * undo its actions and return zero.
442 * Otherwise, function returns one-past-the-last PFN of isolated page
443 * (which may be greater then end_pfn if end fell in a middle of
447 isolate_freepages_range(struct compact_control *cc,
448 unsigned long start_pfn, unsigned long end_pfn)
450 unsigned long isolated, pfn, block_end_pfn;
454 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
456 for (; pfn < end_pfn; pfn += isolated,
457 block_end_pfn += pageblock_nr_pages) {
459 block_end_pfn = min(block_end_pfn, end_pfn);
461 if (!pageblock_pfn_to_page(pfn, block_end_pfn, cc->zone))
464 isolated = isolate_freepages_block(cc, pfn, block_end_pfn,
468 * In strict mode, isolate_freepages_block() returns 0 if
469 * there are any holes in the block (ie. invalid PFNs or
476 * If we managed to isolate pages, it is always (1 << n) *
477 * pageblock_nr_pages for some non-negative n. (Max order
478 * page may span two pageblocks).
482 /* split_free_page does not map the pages */
483 map_pages(&freelist);
486 /* Loop terminated early, cleanup. */
487 release_freepages(&freelist);
491 /* We don't use freelists for anything. */
495 /* Update the number of anon and file isolated pages in the zone */
496 static void acct_isolated(struct zone *zone, struct compact_control *cc)
499 unsigned int count[2] = { 0, };
501 if (list_empty(&cc->migratepages))
504 list_for_each_entry(page, &cc->migratepages, lru)
505 count[!!page_is_file_cache(page)]++;
507 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
508 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
511 /* Similar to reclaim, but different enough that they don't share logic */
512 static bool too_many_isolated(struct zone *zone)
514 unsigned long active, inactive, isolated;
516 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
517 zone_page_state(zone, NR_INACTIVE_ANON);
518 active = zone_page_state(zone, NR_ACTIVE_FILE) +
519 zone_page_state(zone, NR_ACTIVE_ANON);
520 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
521 zone_page_state(zone, NR_ISOLATED_ANON);
523 return isolated > (inactive + active) / 2;
527 * isolate_migratepages_block() - isolate all migrate-able pages within
529 * @cc: Compaction control structure.
530 * @low_pfn: The first PFN to isolate
531 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
532 * @isolate_mode: Isolation mode to be used.
534 * Isolate all pages that can be migrated from the range specified by
535 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
536 * Returns zero if there is a fatal signal pending, otherwise PFN of the
537 * first page that was not scanned (which may be both less, equal to or more
540 * The pages are isolated on cc->migratepages list (not required to be empty),
541 * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
542 * is neither read nor updated.
545 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
546 unsigned long end_pfn, isolate_mode_t isolate_mode)
548 struct zone *zone = cc->zone;
549 unsigned long nr_scanned = 0, nr_isolated = 0;
550 struct list_head *migratelist = &cc->migratepages;
551 struct lruvec *lruvec;
554 struct page *page = NULL, *valid_page = NULL;
557 * Ensure that there are not too many pages isolated from the LRU
558 * list by either parallel reclaimers or compaction. If there are,
559 * delay for some time until fewer pages are isolated
561 while (unlikely(too_many_isolated(zone))) {
562 /* async migration should just abort */
563 if (cc->mode == MIGRATE_ASYNC)
566 congestion_wait(BLK_RW_ASYNC, HZ/10);
568 if (fatal_signal_pending(current))
572 if (compact_should_abort(cc))
575 /* Time to isolate some pages for migration */
576 for (; low_pfn < end_pfn; low_pfn++) {
578 * Periodically drop the lock (if held) regardless of its
579 * contention, to give chance to IRQs. Abort async compaction
582 if (!(low_pfn % SWAP_CLUSTER_MAX)
583 && compact_unlock_should_abort(&zone->lru_lock, flags,
587 if (!pfn_valid_within(low_pfn))
591 page = pfn_to_page(low_pfn);
597 * Skip if free. page_order cannot be used without zone->lock
598 * as nothing prevents parallel allocations or buddy merging.
604 * Check may be lockless but that's ok as we recheck later.
605 * It's possible to migrate LRU pages and balloon pages
606 * Skip any other type of page
608 if (!PageLRU(page)) {
609 if (unlikely(balloon_page_movable(page))) {
610 if (locked && balloon_page_isolate(page)) {
611 /* Successfully isolated */
612 goto isolate_success;
619 * PageLRU is set. lru_lock normally excludes isolation
620 * splitting and collapsing (collapsing has already happened
621 * if PageLRU is set) but the lock is not necessarily taken
622 * here and it is wasteful to take it just to check transhuge.
623 * Check TransHuge without lock and skip the whole pageblock if
624 * it's either a transhuge or hugetlbfs page, as calling
625 * compound_order() without preventing THP from splitting the
626 * page underneath us may return surprising results.
628 if (PageTransHuge(page)) {
630 low_pfn = ALIGN(low_pfn + 1,
631 pageblock_nr_pages) - 1;
633 low_pfn += (1 << compound_order(page)) - 1;
639 * Migration will fail if an anonymous page is pinned in memory,
640 * so avoid taking lru_lock and isolating it unnecessarily in an
641 * admittedly racy check.
643 if (!page_mapping(page) &&
644 page_count(page) > page_mapcount(page))
647 /* If the lock is not held, try to take it */
649 locked = compact_trylock_irqsave(&zone->lru_lock,
654 /* Recheck PageLRU and PageTransHuge under lock */
657 if (PageTransHuge(page)) {
658 low_pfn += (1 << compound_order(page)) - 1;
662 lruvec = mem_cgroup_page_lruvec(page, zone);
664 /* Try isolate the page */
665 if (__isolate_lru_page(page, isolate_mode) != 0)
668 VM_BUG_ON_PAGE(PageTransCompound(page), page);
670 /* Successfully isolated */
671 del_page_from_lru_list(page, lruvec, page_lru(page));
674 cc->finished_update_migrate = true;
675 list_add(&page->lru, migratelist);
676 cc->nr_migratepages++;
679 /* Avoid isolating too much */
680 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
687 spin_unlock_irqrestore(&zone->lru_lock, flags);
690 * Update the pageblock-skip information and cached scanner pfn,
691 * if the whole pageblock was scanned without isolating any page.
693 if (low_pfn == end_pfn)
694 update_pageblock_skip(cc, valid_page, nr_isolated, true);
696 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
698 count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
700 count_compact_events(COMPACTISOLATED, nr_isolated);
706 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
707 * @cc: Compaction control structure.
708 * @start_pfn: The first PFN to start isolating.
709 * @end_pfn: The one-past-last PFN.
711 * Returns zero if isolation fails fatally due to e.g. pending signal.
712 * Otherwise, function returns one-past-the-last PFN of isolated page
713 * (which may be greater than end_pfn if end fell in a middle of a THP page).
716 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
717 unsigned long end_pfn)
719 unsigned long pfn, block_end_pfn;
721 /* Scan block by block. First and last block may be incomplete */
723 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
725 for (; pfn < end_pfn; pfn = block_end_pfn,
726 block_end_pfn += pageblock_nr_pages) {
728 block_end_pfn = min(block_end_pfn, end_pfn);
730 if (!pageblock_pfn_to_page(pfn, block_end_pfn, cc->zone))
733 pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
734 ISOLATE_UNEVICTABLE);
737 * In case of fatal failure, release everything that might
738 * have been isolated in the previous iteration, and signal
739 * the failure back to caller.
742 putback_movable_pages(&cc->migratepages);
743 cc->nr_migratepages = 0;
747 acct_isolated(cc->zone, cc);
752 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
753 #ifdef CONFIG_COMPACTION
755 * Based on information in the current compact_control, find blocks
756 * suitable for isolating free pages from and then isolate them.
758 static void isolate_freepages(struct compact_control *cc)
760 struct zone *zone = cc->zone;
762 unsigned long block_start_pfn; /* start of current pageblock */
763 unsigned long block_end_pfn; /* end of current pageblock */
764 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
765 int nr_freepages = cc->nr_freepages;
766 struct list_head *freelist = &cc->freepages;
769 * Initialise the free scanner. The starting point is where we last
770 * successfully isolated from, zone-cached value, or the end of the
771 * zone when isolating for the first time. We need this aligned to
772 * the pageblock boundary, because we do
773 * block_start_pfn -= pageblock_nr_pages in the for loop.
774 * For ending point, take care when isolating in last pageblock of a
775 * a zone which ends in the middle of a pageblock.
776 * The low boundary is the end of the pageblock the migration scanner
779 block_start_pfn = cc->free_pfn & ~(pageblock_nr_pages-1);
780 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
782 low_pfn = ALIGN(cc->migrate_pfn + 1, pageblock_nr_pages);
785 * Isolate free pages until enough are available to migrate the
786 * pages on cc->migratepages. We stop searching if the migrate
787 * and free page scanners meet or enough free pages are isolated.
789 for (; block_start_pfn >= low_pfn && cc->nr_migratepages > nr_freepages;
790 block_end_pfn = block_start_pfn,
791 block_start_pfn -= pageblock_nr_pages) {
792 unsigned long isolated;
795 * This can iterate a massively long zone without finding any
796 * suitable migration targets, so periodically check if we need
797 * to schedule, or even abort async compaction.
799 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
800 && compact_should_abort(cc))
803 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
808 /* Check the block is suitable for migration */
809 if (!suitable_migration_target(page))
812 /* If isolation recently failed, do not retry */
813 if (!isolation_suitable(cc, page))
816 /* Found a block suitable for isolating free pages from */
817 cc->free_pfn = block_start_pfn;
818 isolated = isolate_freepages_block(cc, block_start_pfn,
819 block_end_pfn, freelist, false);
820 nr_freepages += isolated;
823 * Set a flag that we successfully isolated in this pageblock.
824 * In the next loop iteration, zone->compact_cached_free_pfn
825 * will not be updated and thus it will effectively contain the
826 * highest pageblock we isolated pages from.
829 cc->finished_update_free = true;
832 * isolate_freepages_block() might have aborted due to async
833 * compaction being contended
839 /* split_free_page does not map the pages */
843 * If we crossed the migrate scanner, we want to keep it that way
844 * so that compact_finished() may detect this
846 if (block_start_pfn < low_pfn)
847 cc->free_pfn = cc->migrate_pfn;
849 cc->nr_freepages = nr_freepages;
853 * This is a migrate-callback that "allocates" freepages by taking pages
854 * from the isolated freelists in the block we are migrating to.
856 static struct page *compaction_alloc(struct page *migratepage,
860 struct compact_control *cc = (struct compact_control *)data;
861 struct page *freepage;
864 * Isolate free pages if necessary, and if we are not aborting due to
867 if (list_empty(&cc->freepages)) {
869 isolate_freepages(cc);
871 if (list_empty(&cc->freepages))
875 freepage = list_entry(cc->freepages.next, struct page, lru);
876 list_del(&freepage->lru);
883 * This is a migrate-callback that "frees" freepages back to the isolated
884 * freelist. All pages on the freelist are from the same zone, so there is no
885 * special handling needed for NUMA.
887 static void compaction_free(struct page *page, unsigned long data)
889 struct compact_control *cc = (struct compact_control *)data;
891 list_add(&page->lru, &cc->freepages);
895 /* possible outcome of isolate_migratepages */
897 ISOLATE_ABORT, /* Abort compaction now */
898 ISOLATE_NONE, /* No pages isolated, continue scanning */
899 ISOLATE_SUCCESS, /* Pages isolated, migrate */
903 * Isolate all pages that can be migrated from the first suitable block,
904 * starting at the block pointed to by the migrate scanner pfn within
907 static isolate_migrate_t isolate_migratepages(struct zone *zone,
908 struct compact_control *cc)
910 unsigned long low_pfn, end_pfn;
912 const isolate_mode_t isolate_mode =
913 (cc->mode == MIGRATE_ASYNC ? ISOLATE_ASYNC_MIGRATE : 0);
916 * Start at where we last stopped, or beginning of the zone as
917 * initialized by compact_zone()
919 low_pfn = cc->migrate_pfn;
921 /* Only scan within a pageblock boundary */
922 end_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages);
925 * Iterate over whole pageblocks until we find the first suitable.
926 * Do not cross the free scanner.
928 for (; end_pfn <= cc->free_pfn;
929 low_pfn = end_pfn, end_pfn += pageblock_nr_pages) {
932 * This can potentially iterate a massively long zone with
933 * many pageblocks unsuitable, so periodically check if we
934 * need to schedule, or even abort async compaction.
936 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
937 && compact_should_abort(cc))
940 page = pageblock_pfn_to_page(low_pfn, end_pfn, zone);
944 /* If isolation recently failed, do not retry */
945 if (!isolation_suitable(cc, page))
949 * For async compaction, also only scan in MOVABLE blocks.
950 * Async compaction is optimistic to see if the minimum amount
951 * of work satisfies the allocation.
953 if (cc->mode == MIGRATE_ASYNC &&
954 !migrate_async_suitable(get_pageblock_migratetype(page)))
957 /* Perform the isolation */
958 low_pfn = isolate_migratepages_block(cc, low_pfn, end_pfn,
961 if (!low_pfn || cc->contended)
962 return ISOLATE_ABORT;
965 * Either we isolated something and proceed with migration. Or
966 * we failed and compact_zone should decide if we should
972 acct_isolated(zone, cc);
973 /* Record where migration scanner will be restarted */
974 cc->migrate_pfn = low_pfn;
976 return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
979 static int compact_finished(struct zone *zone,
980 struct compact_control *cc)
983 unsigned long watermark;
985 if (cc->contended || fatal_signal_pending(current))
986 return COMPACT_PARTIAL;
988 /* Compaction run completes if the migrate and free scanner meet */
989 if (cc->free_pfn <= cc->migrate_pfn) {
990 /* Let the next compaction start anew. */
991 zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
992 zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
993 zone->compact_cached_free_pfn = zone_end_pfn(zone);
996 * Mark that the PG_migrate_skip information should be cleared
997 * by kswapd when it goes to sleep. kswapd does not set the
998 * flag itself as the decision to be clear should be directly
999 * based on an allocation request.
1001 if (!current_is_kswapd())
1002 zone->compact_blockskip_flush = true;
1004 return COMPACT_COMPLETE;
1008 * order == -1 is expected when compacting via
1009 * /proc/sys/vm/compact_memory
1011 if (cc->order == -1)
1012 return COMPACT_CONTINUE;
1014 /* Compaction run is not finished if the watermark is not met */
1015 watermark = low_wmark_pages(zone);
1016 watermark += (1 << cc->order);
1018 if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
1019 return COMPACT_CONTINUE;
1021 /* Direct compactor: Is a suitable page free? */
1022 for (order = cc->order; order < MAX_ORDER; order++) {
1023 struct free_area *area = &zone->free_area[order];
1025 /* Job done if page is free of the right migratetype */
1026 if (!list_empty(&area->free_list[cc->migratetype]))
1027 return COMPACT_PARTIAL;
1029 /* Job done if allocation would set block type */
1030 if (cc->order >= pageblock_order && area->nr_free)
1031 return COMPACT_PARTIAL;
1034 return COMPACT_CONTINUE;
1038 * compaction_suitable: Is this suitable to run compaction on this zone now?
1040 * COMPACT_SKIPPED - If there are too few free pages for compaction
1041 * COMPACT_PARTIAL - If the allocation would succeed without compaction
1042 * COMPACT_CONTINUE - If compaction should run now
1044 unsigned long compaction_suitable(struct zone *zone, int order)
1047 unsigned long watermark;
1050 * order == -1 is expected when compacting via
1051 * /proc/sys/vm/compact_memory
1054 return COMPACT_CONTINUE;
1057 * Watermarks for order-0 must be met for compaction. Note the 2UL.
1058 * This is because during migration, copies of pages need to be
1059 * allocated and for a short time, the footprint is higher
1061 watermark = low_wmark_pages(zone) + (2UL << order);
1062 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1063 return COMPACT_SKIPPED;
1066 * fragmentation index determines if allocation failures are due to
1067 * low memory or external fragmentation
1069 * index of -1000 implies allocations might succeed depending on
1071 * index towards 0 implies failure is due to lack of memory
1072 * index towards 1000 implies failure is due to fragmentation
1074 * Only compact if a failure would be due to fragmentation.
1076 fragindex = fragmentation_index(zone, order);
1077 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
1078 return COMPACT_SKIPPED;
1080 if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
1082 return COMPACT_PARTIAL;
1084 return COMPACT_CONTINUE;
1087 static int compact_zone(struct zone *zone, struct compact_control *cc)
1090 unsigned long start_pfn = zone->zone_start_pfn;
1091 unsigned long end_pfn = zone_end_pfn(zone);
1092 const bool sync = cc->mode != MIGRATE_ASYNC;
1094 ret = compaction_suitable(zone, cc->order);
1096 case COMPACT_PARTIAL:
1097 case COMPACT_SKIPPED:
1098 /* Compaction is likely to fail */
1100 case COMPACT_CONTINUE:
1101 /* Fall through to compaction */
1106 * Clear pageblock skip if there were failures recently and compaction
1107 * is about to be retried after being deferred. kswapd does not do
1108 * this reset as it'll reset the cached information when going to sleep.
1110 if (compaction_restarting(zone, cc->order) && !current_is_kswapd())
1111 __reset_isolation_suitable(zone);
1114 * Setup to move all movable pages to the end of the zone. Used cached
1115 * information on where the scanners should start but check that it
1116 * is initialised by ensuring the values are within zone boundaries.
1118 cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync];
1119 cc->free_pfn = zone->compact_cached_free_pfn;
1120 if (cc->free_pfn < start_pfn || cc->free_pfn > end_pfn) {
1121 cc->free_pfn = end_pfn & ~(pageblock_nr_pages-1);
1122 zone->compact_cached_free_pfn = cc->free_pfn;
1124 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn > end_pfn) {
1125 cc->migrate_pfn = start_pfn;
1126 zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
1127 zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
1130 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn, cc->free_pfn, end_pfn);
1132 migrate_prep_local();
1134 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
1137 switch (isolate_migratepages(zone, cc)) {
1139 ret = COMPACT_PARTIAL;
1140 putback_movable_pages(&cc->migratepages);
1141 cc->nr_migratepages = 0;
1145 case ISOLATE_SUCCESS:
1149 err = migrate_pages(&cc->migratepages, compaction_alloc,
1150 compaction_free, (unsigned long)cc, cc->mode,
1153 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
1156 /* All pages were either migrated or will be released */
1157 cc->nr_migratepages = 0;
1159 putback_movable_pages(&cc->migratepages);
1161 * migrate_pages() may return -ENOMEM when scanners meet
1162 * and we want compact_finished() to detect it
1164 if (err == -ENOMEM && cc->free_pfn > cc->migrate_pfn) {
1165 ret = COMPACT_PARTIAL;
1172 /* Release free pages and check accounting */
1173 cc->nr_freepages -= release_freepages(&cc->freepages);
1174 VM_BUG_ON(cc->nr_freepages != 0);
1176 trace_mm_compaction_end(ret);
1181 static unsigned long compact_zone_order(struct zone *zone, int order,
1182 gfp_t gfp_mask, enum migrate_mode mode, int *contended)
1185 struct compact_control cc = {
1187 .nr_migratepages = 0,
1189 .migratetype = allocflags_to_migratetype(gfp_mask),
1193 INIT_LIST_HEAD(&cc.freepages);
1194 INIT_LIST_HEAD(&cc.migratepages);
1196 ret = compact_zone(zone, &cc);
1198 VM_BUG_ON(!list_empty(&cc.freepages));
1199 VM_BUG_ON(!list_empty(&cc.migratepages));
1201 *contended = cc.contended;
1205 int sysctl_extfrag_threshold = 500;
1208 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1209 * @zonelist: The zonelist used for the current allocation
1210 * @order: The order of the current allocation
1211 * @gfp_mask: The GFP mask of the current allocation
1212 * @nodemask: The allowed nodes to allocate from
1213 * @mode: The migration mode for async, sync light, or sync migration
1214 * @contended: Return value that determines if compaction was aborted due to
1215 * need_resched() or lock contention
1216 * @candidate_zone: Return the zone where we think allocation should succeed
1218 * This is the main entry point for direct page compaction.
1220 unsigned long try_to_compact_pages(struct zonelist *zonelist,
1221 int order, gfp_t gfp_mask, nodemask_t *nodemask,
1222 enum migrate_mode mode, int *contended,
1223 struct zone **candidate_zone)
1225 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1226 int may_enter_fs = gfp_mask & __GFP_FS;
1227 int may_perform_io = gfp_mask & __GFP_IO;
1230 int rc = COMPACT_DEFERRED;
1231 int alloc_flags = 0;
1232 int all_zones_contended = COMPACT_CONTENDED_LOCK; /* init for &= op */
1234 *contended = COMPACT_CONTENDED_NONE;
1236 /* Check if the GFP flags allow compaction */
1237 if (!order || !may_enter_fs || !may_perform_io)
1238 return COMPACT_SKIPPED;
1241 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
1242 alloc_flags |= ALLOC_CMA;
1244 /* Compact each zone in the list */
1245 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1250 if (compaction_deferred(zone, order))
1253 status = compact_zone_order(zone, order, gfp_mask, mode,
1255 rc = max(status, rc);
1257 * It takes at least one zone that wasn't lock contended
1258 * to clear all_zones_contended.
1260 all_zones_contended &= zone_contended;
1262 /* If a normal allocation would succeed, stop compacting */
1263 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0,
1265 *candidate_zone = zone;
1267 * We think the allocation will succeed in this zone,
1268 * but it is not certain, hence the false. The caller
1269 * will repeat this with true if allocation indeed
1270 * succeeds in this zone.
1272 compaction_defer_reset(zone, order, false);
1274 * It is possible that async compaction aborted due to
1275 * need_resched() and the watermarks were ok thanks to
1276 * somebody else freeing memory. The allocation can
1277 * however still fail so we better signal the
1278 * need_resched() contention anyway (this will not
1279 * prevent the allocation attempt).
1281 if (zone_contended == COMPACT_CONTENDED_SCHED)
1282 *contended = COMPACT_CONTENDED_SCHED;
1287 if (mode != MIGRATE_ASYNC) {
1289 * We think that allocation won't succeed in this zone
1290 * so we defer compaction there. If it ends up
1291 * succeeding after all, it will be reset.
1293 defer_compaction(zone, order);
1297 * We might have stopped compacting due to need_resched() in
1298 * async compaction, or due to a fatal signal detected. In that
1299 * case do not try further zones and signal need_resched()
1302 if ((zone_contended == COMPACT_CONTENDED_SCHED)
1303 || fatal_signal_pending(current)) {
1304 *contended = COMPACT_CONTENDED_SCHED;
1311 * We might not have tried all the zones, so be conservative
1312 * and assume they are not all lock contended.
1314 all_zones_contended = 0;
1319 * If at least one zone wasn't deferred or skipped, we report if all
1320 * zones that were tried were lock contended.
1322 if (rc > COMPACT_SKIPPED && all_zones_contended)
1323 *contended = COMPACT_CONTENDED_LOCK;
1329 /* Compact all zones within a node */
1330 static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
1335 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1337 zone = &pgdat->node_zones[zoneid];
1338 if (!populated_zone(zone))
1341 cc->nr_freepages = 0;
1342 cc->nr_migratepages = 0;
1344 INIT_LIST_HEAD(&cc->freepages);
1345 INIT_LIST_HEAD(&cc->migratepages);
1347 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
1348 compact_zone(zone, cc);
1350 if (cc->order > 0) {
1351 if (zone_watermark_ok(zone, cc->order,
1352 low_wmark_pages(zone), 0, 0))
1353 compaction_defer_reset(zone, cc->order, false);
1356 VM_BUG_ON(!list_empty(&cc->freepages));
1357 VM_BUG_ON(!list_empty(&cc->migratepages));
1361 void compact_pgdat(pg_data_t *pgdat, int order)
1363 struct compact_control cc = {
1365 .mode = MIGRATE_ASYNC,
1371 __compact_pgdat(pgdat, &cc);
1374 static void compact_node(int nid)
1376 struct compact_control cc = {
1378 .mode = MIGRATE_SYNC,
1379 .ignore_skip_hint = true,
1382 __compact_pgdat(NODE_DATA(nid), &cc);
1385 /* Compact all nodes in the system */
1386 static void compact_nodes(void)
1390 /* Flush pending updates to the LRU lists */
1391 lru_add_drain_all();
1393 for_each_online_node(nid)
1397 /* The written value is actually unused, all memory is compacted */
1398 int sysctl_compact_memory;
1400 /* This is the entry point for compacting all nodes via /proc/sys/vm */
1401 int sysctl_compaction_handler(struct ctl_table *table, int write,
1402 void __user *buffer, size_t *length, loff_t *ppos)
1410 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1411 void __user *buffer, size_t *length, loff_t *ppos)
1413 proc_dointvec_minmax(table, write, buffer, length, ppos);
1418 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1419 static ssize_t sysfs_compact_node(struct device *dev,
1420 struct device_attribute *attr,
1421 const char *buf, size_t count)
1425 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1426 /* Flush pending updates to the LRU lists */
1427 lru_add_drain_all();
1434 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1436 int compaction_register_node(struct node *node)
1438 return device_create_file(&node->dev, &dev_attr_compact);
1441 void compaction_unregister_node(struct node *node)
1443 return device_remove_file(&node->dev, &dev_attr_compact);
1445 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1447 #endif /* CONFIG_COMPACTION */