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
37 #ifdef CONFIG_TRACEPOINTS
38 static const char *const compaction_status_string[] = {
47 #define CREATE_TRACE_POINTS
48 #include <trace/events/compaction.h>
50 static unsigned long release_freepages(struct list_head *freelist)
52 struct page *page, *next;
53 unsigned long high_pfn = 0;
55 list_for_each_entry_safe(page, next, freelist, lru) {
56 unsigned long pfn = page_to_pfn(page);
66 static void map_pages(struct list_head *list)
70 list_for_each_entry(page, list, lru) {
71 arch_alloc_page(page, 0);
72 kernel_map_pages(page, 1, 1);
76 static inline bool migrate_async_suitable(int migratetype)
78 return is_migrate_cma(migratetype) || migratetype == MIGRATE_MOVABLE;
82 * Check that the whole (or subset of) a pageblock given by the interval of
83 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
84 * with the migration of free compaction scanner. The scanners then need to
85 * use only pfn_valid_within() check for arches that allow holes within
88 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
90 * It's possible on some configurations to have a setup like node0 node1 node0
91 * i.e. it's possible that all pages within a zones range of pages do not
92 * belong to a single zone. We assume that a border between node0 and node1
93 * can occur within a single pageblock, but not a node0 node1 node0
94 * interleaving within a single pageblock. It is therefore sufficient to check
95 * the first and last page of a pageblock and avoid checking each individual
96 * page in a pageblock.
98 static struct page *pageblock_pfn_to_page(unsigned long start_pfn,
99 unsigned long end_pfn, struct zone *zone)
101 struct page *start_page;
102 struct page *end_page;
104 /* end_pfn is one past the range we are checking */
107 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
110 start_page = pfn_to_page(start_pfn);
112 if (page_zone(start_page) != zone)
115 end_page = pfn_to_page(end_pfn);
117 /* This gives a shorter code than deriving page_zone(end_page) */
118 if (page_zone_id(start_page) != page_zone_id(end_page))
124 #ifdef CONFIG_COMPACTION
125 /* Returns true if the pageblock should be scanned for pages to isolate. */
126 static inline bool isolation_suitable(struct compact_control *cc,
129 if (cc->ignore_skip_hint)
132 return !get_pageblock_skip(page);
136 * This function is called to clear all cached information on pageblocks that
137 * should be skipped for page isolation when the migrate and free page scanner
140 static void __reset_isolation_suitable(struct zone *zone)
142 unsigned long start_pfn = zone->zone_start_pfn;
143 unsigned long end_pfn = zone_end_pfn(zone);
146 zone->compact_cached_migrate_pfn[0] = start_pfn;
147 zone->compact_cached_migrate_pfn[1] = start_pfn;
148 zone->compact_cached_free_pfn = end_pfn;
149 zone->compact_blockskip_flush = false;
151 /* Walk the zone and mark every pageblock as suitable for isolation */
152 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
160 page = pfn_to_page(pfn);
161 if (zone != page_zone(page))
164 clear_pageblock_skip(page);
168 void reset_isolation_suitable(pg_data_t *pgdat)
172 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
173 struct zone *zone = &pgdat->node_zones[zoneid];
174 if (!populated_zone(zone))
177 /* Only flush if a full compaction finished recently */
178 if (zone->compact_blockskip_flush)
179 __reset_isolation_suitable(zone);
184 * If no pages were isolated then mark this pageblock to be skipped in the
185 * future. The information is later cleared by __reset_isolation_suitable().
187 static void update_pageblock_skip(struct compact_control *cc,
188 struct page *page, unsigned long nr_isolated,
189 bool migrate_scanner)
191 struct zone *zone = cc->zone;
194 if (cc->ignore_skip_hint)
203 set_pageblock_skip(page);
205 pfn = page_to_pfn(page);
207 /* Update where async and sync compaction should restart */
208 if (migrate_scanner) {
209 if (pfn > zone->compact_cached_migrate_pfn[0])
210 zone->compact_cached_migrate_pfn[0] = pfn;
211 if (cc->mode != MIGRATE_ASYNC &&
212 pfn > zone->compact_cached_migrate_pfn[1])
213 zone->compact_cached_migrate_pfn[1] = pfn;
215 if (pfn < zone->compact_cached_free_pfn)
216 zone->compact_cached_free_pfn = pfn;
220 static inline bool isolation_suitable(struct compact_control *cc,
226 static void update_pageblock_skip(struct compact_control *cc,
227 struct page *page, unsigned long nr_isolated,
228 bool migrate_scanner)
231 #endif /* CONFIG_COMPACTION */
234 * Compaction requires the taking of some coarse locks that are potentially
235 * very heavily contended. For async compaction, back out if the lock cannot
236 * be taken immediately. For sync compaction, spin on the lock if needed.
238 * Returns true if the lock is held
239 * Returns false if the lock is not held and compaction should abort
241 static bool compact_trylock_irqsave(spinlock_t *lock, unsigned long *flags,
242 struct compact_control *cc)
244 if (cc->mode == MIGRATE_ASYNC) {
245 if (!spin_trylock_irqsave(lock, *flags)) {
246 cc->contended = COMPACT_CONTENDED_LOCK;
250 spin_lock_irqsave(lock, *flags);
257 * Compaction requires the taking of some coarse locks that are potentially
258 * very heavily contended. The lock should be periodically unlocked to avoid
259 * having disabled IRQs for a long time, even when there is nobody waiting on
260 * the lock. It might also be that allowing the IRQs will result in
261 * need_resched() becoming true. If scheduling is needed, async compaction
262 * aborts. Sync compaction schedules.
263 * Either compaction type will also abort if a fatal signal is pending.
264 * In either case if the lock was locked, it is dropped and not regained.
266 * Returns true if compaction should abort due to fatal signal pending, or
267 * async compaction due to need_resched()
268 * Returns false when compaction can continue (sync compaction might have
271 static bool compact_unlock_should_abort(spinlock_t *lock,
272 unsigned long flags, bool *locked, struct compact_control *cc)
275 spin_unlock_irqrestore(lock, flags);
279 if (fatal_signal_pending(current)) {
280 cc->contended = COMPACT_CONTENDED_SCHED;
284 if (need_resched()) {
285 if (cc->mode == MIGRATE_ASYNC) {
286 cc->contended = COMPACT_CONTENDED_SCHED;
296 * Aside from avoiding lock contention, compaction also periodically checks
297 * need_resched() and either schedules in sync compaction or aborts async
298 * compaction. This is similar to what compact_unlock_should_abort() does, but
299 * is used where no lock is concerned.
301 * Returns false when no scheduling was needed, or sync compaction scheduled.
302 * Returns true when async compaction should abort.
304 static inline bool compact_should_abort(struct compact_control *cc)
306 /* async compaction aborts if contended */
307 if (need_resched()) {
308 if (cc->mode == MIGRATE_ASYNC) {
309 cc->contended = COMPACT_CONTENDED_SCHED;
319 /* Returns true if the page is within a block suitable for migration to */
320 static bool suitable_migration_target(struct page *page)
322 /* If the page is a large free page, then disallow migration */
323 if (PageBuddy(page)) {
325 * We are checking page_order without zone->lock taken. But
326 * the only small danger is that we skip a potentially suitable
327 * pageblock, so it's not worth to check order for valid range.
329 if (page_order_unsafe(page) >= pageblock_order)
333 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
334 if (migrate_async_suitable(get_pageblock_migratetype(page)))
337 /* Otherwise skip the block */
342 * Isolate free pages onto a private freelist. If @strict is true, will abort
343 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
344 * (even though it may still end up isolating some pages).
346 static unsigned long isolate_freepages_block(struct compact_control *cc,
347 unsigned long *start_pfn,
348 unsigned long end_pfn,
349 struct list_head *freelist,
352 int nr_scanned = 0, total_isolated = 0;
353 struct page *cursor, *valid_page = NULL;
354 unsigned long flags = 0;
356 unsigned long blockpfn = *start_pfn;
358 cursor = pfn_to_page(blockpfn);
360 /* Isolate free pages. */
361 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
363 struct page *page = cursor;
366 * Periodically drop the lock (if held) regardless of its
367 * contention, to give chance to IRQs. Abort if fatal signal
368 * pending or async compaction detects need_resched()
370 if (!(blockpfn % SWAP_CLUSTER_MAX)
371 && compact_unlock_should_abort(&cc->zone->lock, flags,
376 if (!pfn_valid_within(blockpfn))
381 if (!PageBuddy(page))
385 * If we already hold the lock, we can skip some rechecking.
386 * Note that if we hold the lock now, checked_pageblock was
387 * already set in some previous iteration (or strict is true),
388 * so it is correct to skip the suitable migration target
393 * The zone lock must be held to isolate freepages.
394 * Unfortunately this is a very coarse lock and can be
395 * heavily contended if there are parallel allocations
396 * or parallel compactions. For async compaction do not
397 * spin on the lock and we acquire the lock as late as
400 locked = compact_trylock_irqsave(&cc->zone->lock,
405 /* Recheck this is a buddy page under lock */
406 if (!PageBuddy(page))
410 /* Found a free page, break it into order-0 pages */
411 isolated = split_free_page(page);
412 total_isolated += isolated;
413 for (i = 0; i < isolated; i++) {
414 list_add(&page->lru, freelist);
418 /* If a page was split, advance to the end of it */
420 blockpfn += isolated - 1;
421 cursor += isolated - 1;
433 trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
434 nr_scanned, total_isolated);
436 /* Record how far we have got within the block */
437 *start_pfn = blockpfn;
440 * If strict isolation is requested by CMA then check that all the
441 * pages requested were isolated. If there were any failures, 0 is
442 * returned and CMA will fail.
444 if (strict && blockpfn < end_pfn)
448 spin_unlock_irqrestore(&cc->zone->lock, flags);
450 /* Update the pageblock-skip if the whole pageblock was scanned */
451 if (blockpfn == end_pfn)
452 update_pageblock_skip(cc, valid_page, total_isolated, false);
454 count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
456 count_compact_events(COMPACTISOLATED, total_isolated);
457 return total_isolated;
461 * isolate_freepages_range() - isolate free pages.
462 * @start_pfn: The first PFN to start isolating.
463 * @end_pfn: The one-past-last PFN.
465 * Non-free pages, invalid PFNs, or zone boundaries within the
466 * [start_pfn, end_pfn) range are considered errors, cause function to
467 * undo its actions and return zero.
469 * Otherwise, function returns one-past-the-last PFN of isolated page
470 * (which may be greater then end_pfn if end fell in a middle of
474 isolate_freepages_range(struct compact_control *cc,
475 unsigned long start_pfn, unsigned long end_pfn)
477 unsigned long isolated, pfn, block_end_pfn;
481 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
483 for (; pfn < end_pfn; pfn += isolated,
484 block_end_pfn += pageblock_nr_pages) {
485 /* Protect pfn from changing by isolate_freepages_block */
486 unsigned long isolate_start_pfn = pfn;
488 block_end_pfn = min(block_end_pfn, end_pfn);
491 * pfn could pass the block_end_pfn if isolated freepage
492 * is more than pageblock order. In this case, we adjust
493 * scanning range to right one.
495 if (pfn >= block_end_pfn) {
496 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
497 block_end_pfn = min(block_end_pfn, end_pfn);
500 if (!pageblock_pfn_to_page(pfn, block_end_pfn, cc->zone))
503 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
504 block_end_pfn, &freelist, true);
507 * In strict mode, isolate_freepages_block() returns 0 if
508 * there are any holes in the block (ie. invalid PFNs or
515 * If we managed to isolate pages, it is always (1 << n) *
516 * pageblock_nr_pages for some non-negative n. (Max order
517 * page may span two pageblocks).
521 /* split_free_page does not map the pages */
522 map_pages(&freelist);
525 /* Loop terminated early, cleanup. */
526 release_freepages(&freelist);
530 /* We don't use freelists for anything. */
534 /* Update the number of anon and file isolated pages in the zone */
535 static void acct_isolated(struct zone *zone, struct compact_control *cc)
538 unsigned int count[2] = { 0, };
540 if (list_empty(&cc->migratepages))
543 list_for_each_entry(page, &cc->migratepages, lru)
544 count[!!page_is_file_cache(page)]++;
546 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
547 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
550 /* Similar to reclaim, but different enough that they don't share logic */
551 static bool too_many_isolated(struct zone *zone)
553 unsigned long active, inactive, isolated;
555 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
556 zone_page_state(zone, NR_INACTIVE_ANON);
557 active = zone_page_state(zone, NR_ACTIVE_FILE) +
558 zone_page_state(zone, NR_ACTIVE_ANON);
559 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
560 zone_page_state(zone, NR_ISOLATED_ANON);
562 return isolated > (inactive + active) / 2;
566 * isolate_migratepages_block() - isolate all migrate-able pages within
568 * @cc: Compaction control structure.
569 * @low_pfn: The first PFN to isolate
570 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
571 * @isolate_mode: Isolation mode to be used.
573 * Isolate all pages that can be migrated from the range specified by
574 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
575 * Returns zero if there is a fatal signal pending, otherwise PFN of the
576 * first page that was not scanned (which may be both less, equal to or more
579 * The pages are isolated on cc->migratepages list (not required to be empty),
580 * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
581 * is neither read nor updated.
584 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
585 unsigned long end_pfn, isolate_mode_t isolate_mode)
587 struct zone *zone = cc->zone;
588 unsigned long nr_scanned = 0, nr_isolated = 0;
589 struct list_head *migratelist = &cc->migratepages;
590 struct lruvec *lruvec;
591 unsigned long flags = 0;
593 struct page *page = NULL, *valid_page = NULL;
594 unsigned long start_pfn = low_pfn;
597 * Ensure that there are not too many pages isolated from the LRU
598 * list by either parallel reclaimers or compaction. If there are,
599 * delay for some time until fewer pages are isolated
601 while (unlikely(too_many_isolated(zone))) {
602 /* async migration should just abort */
603 if (cc->mode == MIGRATE_ASYNC)
606 congestion_wait(BLK_RW_ASYNC, HZ/10);
608 if (fatal_signal_pending(current))
612 if (compact_should_abort(cc))
615 /* Time to isolate some pages for migration */
616 for (; low_pfn < end_pfn; low_pfn++) {
618 * Periodically drop the lock (if held) regardless of its
619 * contention, to give chance to IRQs. Abort async compaction
622 if (!(low_pfn % SWAP_CLUSTER_MAX)
623 && compact_unlock_should_abort(&zone->lru_lock, flags,
627 if (!pfn_valid_within(low_pfn))
631 page = pfn_to_page(low_pfn);
637 * Skip if free. We read page order here without zone lock
638 * which is generally unsafe, but the race window is small and
639 * the worst thing that can happen is that we skip some
640 * potential isolation targets.
642 if (PageBuddy(page)) {
643 unsigned long freepage_order = page_order_unsafe(page);
646 * Without lock, we cannot be sure that what we got is
647 * a valid page order. Consider only values in the
648 * valid order range to prevent low_pfn overflow.
650 if (freepage_order > 0 && freepage_order < MAX_ORDER)
651 low_pfn += (1UL << freepage_order) - 1;
656 * Check may be lockless but that's ok as we recheck later.
657 * It's possible to migrate LRU pages and balloon pages
658 * Skip any other type of page
660 if (!PageLRU(page)) {
661 if (unlikely(balloon_page_movable(page))) {
662 if (balloon_page_isolate(page)) {
663 /* Successfully isolated */
664 goto isolate_success;
671 * PageLRU is set. lru_lock normally excludes isolation
672 * splitting and collapsing (collapsing has already happened
673 * if PageLRU is set) but the lock is not necessarily taken
674 * here and it is wasteful to take it just to check transhuge.
675 * Check TransHuge without lock and skip the whole pageblock if
676 * it's either a transhuge or hugetlbfs page, as calling
677 * compound_order() without preventing THP from splitting the
678 * page underneath us may return surprising results.
680 if (PageTransHuge(page)) {
682 low_pfn = ALIGN(low_pfn + 1,
683 pageblock_nr_pages) - 1;
685 low_pfn += (1 << compound_order(page)) - 1;
691 * Migration will fail if an anonymous page is pinned in memory,
692 * so avoid taking lru_lock and isolating it unnecessarily in an
693 * admittedly racy check.
695 if (!page_mapping(page) &&
696 page_count(page) > page_mapcount(page))
699 /* If we already hold the lock, we can skip some rechecking */
701 locked = compact_trylock_irqsave(&zone->lru_lock,
706 /* Recheck PageLRU and PageTransHuge under lock */
709 if (PageTransHuge(page)) {
710 low_pfn += (1 << compound_order(page)) - 1;
715 lruvec = mem_cgroup_page_lruvec(page, zone);
717 /* Try isolate the page */
718 if (__isolate_lru_page(page, isolate_mode) != 0)
721 VM_BUG_ON_PAGE(PageTransCompound(page), page);
723 /* Successfully isolated */
724 del_page_from_lru_list(page, lruvec, page_lru(page));
727 list_add(&page->lru, migratelist);
728 cc->nr_migratepages++;
731 /* Avoid isolating too much */
732 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
739 * The PageBuddy() check could have potentially brought us outside
740 * the range to be scanned.
742 if (unlikely(low_pfn > end_pfn))
746 spin_unlock_irqrestore(&zone->lru_lock, flags);
749 * Update the pageblock-skip information and cached scanner pfn,
750 * if the whole pageblock was scanned without isolating any page.
752 if (low_pfn == end_pfn)
753 update_pageblock_skip(cc, valid_page, nr_isolated, true);
755 trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
756 nr_scanned, nr_isolated);
758 count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
760 count_compact_events(COMPACTISOLATED, nr_isolated);
766 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
767 * @cc: Compaction control structure.
768 * @start_pfn: The first PFN to start isolating.
769 * @end_pfn: The one-past-last PFN.
771 * Returns zero if isolation fails fatally due to e.g. pending signal.
772 * Otherwise, function returns one-past-the-last PFN of isolated page
773 * (which may be greater than end_pfn if end fell in a middle of a THP page).
776 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
777 unsigned long end_pfn)
779 unsigned long pfn, block_end_pfn;
781 /* Scan block by block. First and last block may be incomplete */
783 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
785 for (; pfn < end_pfn; pfn = block_end_pfn,
786 block_end_pfn += pageblock_nr_pages) {
788 block_end_pfn = min(block_end_pfn, end_pfn);
790 if (!pageblock_pfn_to_page(pfn, block_end_pfn, cc->zone))
793 pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
794 ISOLATE_UNEVICTABLE);
797 * In case of fatal failure, release everything that might
798 * have been isolated in the previous iteration, and signal
799 * the failure back to caller.
802 putback_movable_pages(&cc->migratepages);
803 cc->nr_migratepages = 0;
807 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
810 acct_isolated(cc->zone, cc);
815 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
816 #ifdef CONFIG_COMPACTION
818 * Based on information in the current compact_control, find blocks
819 * suitable for isolating free pages from and then isolate them.
821 static void isolate_freepages(struct compact_control *cc)
823 struct zone *zone = cc->zone;
825 unsigned long block_start_pfn; /* start of current pageblock */
826 unsigned long isolate_start_pfn; /* exact pfn we start at */
827 unsigned long block_end_pfn; /* end of current pageblock */
828 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
829 int nr_freepages = cc->nr_freepages;
830 struct list_head *freelist = &cc->freepages;
833 * Initialise the free scanner. The starting point is where we last
834 * successfully isolated from, zone-cached value, or the end of the
835 * zone when isolating for the first time. For looping we also need
836 * this pfn aligned down to the pageblock boundary, because we do
837 * block_start_pfn -= pageblock_nr_pages in the for loop.
838 * For ending point, take care when isolating in last pageblock of a
839 * a zone which ends in the middle of a pageblock.
840 * The low boundary is the end of the pageblock the migration scanner
843 isolate_start_pfn = cc->free_pfn;
844 block_start_pfn = cc->free_pfn & ~(pageblock_nr_pages-1);
845 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
847 low_pfn = ALIGN(cc->migrate_pfn + 1, pageblock_nr_pages);
850 * Isolate free pages until enough are available to migrate the
851 * pages on cc->migratepages. We stop searching if the migrate
852 * and free page scanners meet or enough free pages are isolated.
854 for (; block_start_pfn >= low_pfn && cc->nr_migratepages > nr_freepages;
855 block_end_pfn = block_start_pfn,
856 block_start_pfn -= pageblock_nr_pages,
857 isolate_start_pfn = block_start_pfn) {
858 unsigned long isolated;
861 * This can iterate a massively long zone without finding any
862 * suitable migration targets, so periodically check if we need
863 * to schedule, or even abort async compaction.
865 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
866 && compact_should_abort(cc))
869 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
874 /* Check the block is suitable for migration */
875 if (!suitable_migration_target(page))
878 /* If isolation recently failed, do not retry */
879 if (!isolation_suitable(cc, page))
882 /* Found a block suitable for isolating free pages from. */
883 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
884 block_end_pfn, freelist, false);
885 nr_freepages += isolated;
888 * Remember where the free scanner should restart next time,
889 * which is where isolate_freepages_block() left off.
890 * But if it scanned the whole pageblock, isolate_start_pfn
891 * now points at block_end_pfn, which is the start of the next
893 * In that case we will however want to restart at the start
894 * of the previous pageblock.
896 cc->free_pfn = (isolate_start_pfn < block_end_pfn) ?
898 block_start_pfn - pageblock_nr_pages;
901 * isolate_freepages_block() might have aborted due to async
902 * compaction being contended
908 /* split_free_page does not map the pages */
912 * If we crossed the migrate scanner, we want to keep it that way
913 * so that compact_finished() may detect this
915 if (block_start_pfn < low_pfn)
916 cc->free_pfn = cc->migrate_pfn;
918 cc->nr_freepages = nr_freepages;
922 * This is a migrate-callback that "allocates" freepages by taking pages
923 * from the isolated freelists in the block we are migrating to.
925 static struct page *compaction_alloc(struct page *migratepage,
929 struct compact_control *cc = (struct compact_control *)data;
930 struct page *freepage;
933 * Isolate free pages if necessary, and if we are not aborting due to
936 if (list_empty(&cc->freepages)) {
938 isolate_freepages(cc);
940 if (list_empty(&cc->freepages))
944 freepage = list_entry(cc->freepages.next, struct page, lru);
945 list_del(&freepage->lru);
952 * This is a migrate-callback that "frees" freepages back to the isolated
953 * freelist. All pages on the freelist are from the same zone, so there is no
954 * special handling needed for NUMA.
956 static void compaction_free(struct page *page, unsigned long data)
958 struct compact_control *cc = (struct compact_control *)data;
960 list_add(&page->lru, &cc->freepages);
964 /* possible outcome of isolate_migratepages */
966 ISOLATE_ABORT, /* Abort compaction now */
967 ISOLATE_NONE, /* No pages isolated, continue scanning */
968 ISOLATE_SUCCESS, /* Pages isolated, migrate */
972 * Isolate all pages that can be migrated from the first suitable block,
973 * starting at the block pointed to by the migrate scanner pfn within
976 static isolate_migrate_t isolate_migratepages(struct zone *zone,
977 struct compact_control *cc)
979 unsigned long low_pfn, end_pfn;
981 const isolate_mode_t isolate_mode =
982 (cc->mode == MIGRATE_ASYNC ? ISOLATE_ASYNC_MIGRATE : 0);
985 * Start at where we last stopped, or beginning of the zone as
986 * initialized by compact_zone()
988 low_pfn = cc->migrate_pfn;
990 /* Only scan within a pageblock boundary */
991 end_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages);
994 * Iterate over whole pageblocks until we find the first suitable.
995 * Do not cross the free scanner.
997 for (; end_pfn <= cc->free_pfn;
998 low_pfn = end_pfn, end_pfn += pageblock_nr_pages) {
1001 * This can potentially iterate a massively long zone with
1002 * many pageblocks unsuitable, so periodically check if we
1003 * need to schedule, or even abort async compaction.
1005 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1006 && compact_should_abort(cc))
1009 page = pageblock_pfn_to_page(low_pfn, end_pfn, zone);
1013 /* If isolation recently failed, do not retry */
1014 if (!isolation_suitable(cc, page))
1018 * For async compaction, also only scan in MOVABLE blocks.
1019 * Async compaction is optimistic to see if the minimum amount
1020 * of work satisfies the allocation.
1022 if (cc->mode == MIGRATE_ASYNC &&
1023 !migrate_async_suitable(get_pageblock_migratetype(page)))
1026 /* Perform the isolation */
1027 low_pfn = isolate_migratepages_block(cc, low_pfn, end_pfn,
1030 if (!low_pfn || cc->contended)
1031 return ISOLATE_ABORT;
1034 * Either we isolated something and proceed with migration. Or
1035 * we failed and compact_zone should decide if we should
1041 acct_isolated(zone, cc);
1043 * Record where migration scanner will be restarted. If we end up in
1044 * the same pageblock as the free scanner, make the scanners fully
1045 * meet so that compact_finished() terminates compaction.
1047 cc->migrate_pfn = (end_pfn <= cc->free_pfn) ? low_pfn : cc->free_pfn;
1049 return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
1052 static int compact_finished(struct zone *zone, struct compact_control *cc,
1053 const int migratetype)
1056 unsigned long watermark;
1058 if (cc->contended || fatal_signal_pending(current))
1059 return COMPACT_PARTIAL;
1061 /* Compaction run completes if the migrate and free scanner meet */
1062 if (cc->free_pfn <= cc->migrate_pfn) {
1063 /* Let the next compaction start anew. */
1064 zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
1065 zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
1066 zone->compact_cached_free_pfn = zone_end_pfn(zone);
1069 * Mark that the PG_migrate_skip information should be cleared
1070 * by kswapd when it goes to sleep. kswapd does not set the
1071 * flag itself as the decision to be clear should be directly
1072 * based on an allocation request.
1074 if (!current_is_kswapd())
1075 zone->compact_blockskip_flush = true;
1077 return COMPACT_COMPLETE;
1081 * order == -1 is expected when compacting via
1082 * /proc/sys/vm/compact_memory
1084 if (cc->order == -1)
1085 return COMPACT_CONTINUE;
1087 /* Compaction run is not finished if the watermark is not met */
1088 watermark = low_wmark_pages(zone);
1090 if (!zone_watermark_ok(zone, cc->order, watermark, cc->classzone_idx,
1092 return COMPACT_CONTINUE;
1094 /* Direct compactor: Is a suitable page free? */
1095 for (order = cc->order; order < MAX_ORDER; order++) {
1096 struct free_area *area = &zone->free_area[order];
1098 /* Job done if page is free of the right migratetype */
1099 if (!list_empty(&area->free_list[migratetype]))
1100 return COMPACT_PARTIAL;
1102 /* Job done if allocation would set block type */
1103 if (cc->order >= pageblock_order && area->nr_free)
1104 return COMPACT_PARTIAL;
1107 return COMPACT_CONTINUE;
1111 * compaction_suitable: Is this suitable to run compaction on this zone now?
1113 * COMPACT_SKIPPED - If there are too few free pages for compaction
1114 * COMPACT_PARTIAL - If the allocation would succeed without compaction
1115 * COMPACT_CONTINUE - If compaction should run now
1117 unsigned long compaction_suitable(struct zone *zone, int order,
1118 int alloc_flags, int classzone_idx)
1121 unsigned long watermark;
1124 * order == -1 is expected when compacting via
1125 * /proc/sys/vm/compact_memory
1128 return COMPACT_CONTINUE;
1130 watermark = low_wmark_pages(zone);
1132 * If watermarks for high-order allocation are already met, there
1133 * should be no need for compaction at all.
1135 if (zone_watermark_ok(zone, order, watermark, classzone_idx,
1137 return COMPACT_PARTIAL;
1140 * Watermarks for order-0 must be met for compaction. Note the 2UL.
1141 * This is because during migration, copies of pages need to be
1142 * allocated and for a short time, the footprint is higher
1144 watermark += (2UL << order);
1145 if (!zone_watermark_ok(zone, 0, watermark, classzone_idx, alloc_flags))
1146 return COMPACT_SKIPPED;
1149 * fragmentation index determines if allocation failures are due to
1150 * low memory or external fragmentation
1152 * index of -1000 would imply allocations might succeed depending on
1153 * watermarks, but we already failed the high-order watermark check
1154 * index towards 0 implies failure is due to lack of memory
1155 * index towards 1000 implies failure is due to fragmentation
1157 * Only compact if a failure would be due to fragmentation.
1159 fragindex = fragmentation_index(zone, order);
1160 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
1161 return COMPACT_SKIPPED;
1163 return COMPACT_CONTINUE;
1166 static int compact_zone(struct zone *zone, struct compact_control *cc)
1169 unsigned long start_pfn = zone->zone_start_pfn;
1170 unsigned long end_pfn = zone_end_pfn(zone);
1171 const int migratetype = gfpflags_to_migratetype(cc->gfp_mask);
1172 const bool sync = cc->mode != MIGRATE_ASYNC;
1173 unsigned long last_migrated_pfn = 0;
1175 ret = compaction_suitable(zone, cc->order, cc->alloc_flags,
1178 case COMPACT_PARTIAL:
1179 case COMPACT_SKIPPED:
1180 /* Compaction is likely to fail */
1182 case COMPACT_CONTINUE:
1183 /* Fall through to compaction */
1188 * Clear pageblock skip if there were failures recently and compaction
1189 * is about to be retried after being deferred. kswapd does not do
1190 * this reset as it'll reset the cached information when going to sleep.
1192 if (compaction_restarting(zone, cc->order) && !current_is_kswapd())
1193 __reset_isolation_suitable(zone);
1196 * Setup to move all movable pages to the end of the zone. Used cached
1197 * information on where the scanners should start but check that it
1198 * is initialised by ensuring the values are within zone boundaries.
1200 cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync];
1201 cc->free_pfn = zone->compact_cached_free_pfn;
1202 if (cc->free_pfn < start_pfn || cc->free_pfn > end_pfn) {
1203 cc->free_pfn = end_pfn & ~(pageblock_nr_pages-1);
1204 zone->compact_cached_free_pfn = cc->free_pfn;
1206 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn > end_pfn) {
1207 cc->migrate_pfn = start_pfn;
1208 zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
1209 zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
1212 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
1213 cc->free_pfn, end_pfn, sync);
1215 migrate_prep_local();
1217 while ((ret = compact_finished(zone, cc, migratetype)) ==
1220 unsigned long isolate_start_pfn = cc->migrate_pfn;
1222 switch (isolate_migratepages(zone, cc)) {
1224 ret = COMPACT_PARTIAL;
1225 putback_movable_pages(&cc->migratepages);
1226 cc->nr_migratepages = 0;
1230 * We haven't isolated and migrated anything, but
1231 * there might still be unflushed migrations from
1232 * previous cc->order aligned block.
1235 case ISOLATE_SUCCESS:
1239 err = migrate_pages(&cc->migratepages, compaction_alloc,
1240 compaction_free, (unsigned long)cc, cc->mode,
1243 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
1246 /* All pages were either migrated or will be released */
1247 cc->nr_migratepages = 0;
1249 putback_movable_pages(&cc->migratepages);
1251 * migrate_pages() may return -ENOMEM when scanners meet
1252 * and we want compact_finished() to detect it
1254 if (err == -ENOMEM && cc->free_pfn > cc->migrate_pfn) {
1255 ret = COMPACT_PARTIAL;
1261 * Record where we could have freed pages by migration and not
1262 * yet flushed them to buddy allocator. We use the pfn that
1263 * isolate_migratepages() started from in this loop iteration
1264 * - this is the lowest page that could have been isolated and
1265 * then freed by migration.
1267 if (!last_migrated_pfn)
1268 last_migrated_pfn = isolate_start_pfn;
1272 * Has the migration scanner moved away from the previous
1273 * cc->order aligned block where we migrated from? If yes,
1274 * flush the pages that were freed, so that they can merge and
1275 * compact_finished() can detect immediately if allocation
1278 if (cc->order > 0 && last_migrated_pfn) {
1280 unsigned long current_block_start =
1281 cc->migrate_pfn & ~((1UL << cc->order) - 1);
1283 if (last_migrated_pfn < current_block_start) {
1285 lru_add_drain_cpu(cpu);
1286 drain_local_pages(zone);
1288 /* No more flushing until we migrate again */
1289 last_migrated_pfn = 0;
1297 * Release free pages and update where the free scanner should restart,
1298 * so we don't leave any returned pages behind in the next attempt.
1300 if (cc->nr_freepages > 0) {
1301 unsigned long free_pfn = release_freepages(&cc->freepages);
1303 cc->nr_freepages = 0;
1304 VM_BUG_ON(free_pfn == 0);
1305 /* The cached pfn is always the first in a pageblock */
1306 free_pfn &= ~(pageblock_nr_pages-1);
1308 * Only go back, not forward. The cached pfn might have been
1309 * already reset to zone end in compact_finished()
1311 if (free_pfn > zone->compact_cached_free_pfn)
1312 zone->compact_cached_free_pfn = free_pfn;
1315 trace_mm_compaction_end(start_pfn, cc->migrate_pfn,
1316 cc->free_pfn, end_pfn, sync, ret);
1321 static unsigned long compact_zone_order(struct zone *zone, int order,
1322 gfp_t gfp_mask, enum migrate_mode mode, int *contended,
1323 int alloc_flags, int classzone_idx)
1326 struct compact_control cc = {
1328 .nr_migratepages = 0,
1330 .gfp_mask = gfp_mask,
1333 .alloc_flags = alloc_flags,
1334 .classzone_idx = classzone_idx,
1336 INIT_LIST_HEAD(&cc.freepages);
1337 INIT_LIST_HEAD(&cc.migratepages);
1339 ret = compact_zone(zone, &cc);
1341 VM_BUG_ON(!list_empty(&cc.freepages));
1342 VM_BUG_ON(!list_empty(&cc.migratepages));
1344 *contended = cc.contended;
1348 int sysctl_extfrag_threshold = 500;
1351 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1352 * @gfp_mask: The GFP mask of the current allocation
1353 * @order: The order of the current allocation
1354 * @alloc_flags: The allocation flags of the current allocation
1355 * @ac: The context of current allocation
1356 * @mode: The migration mode for async, sync light, or sync migration
1357 * @contended: Return value that determines if compaction was aborted due to
1358 * need_resched() or lock contention
1360 * This is the main entry point for direct page compaction.
1362 unsigned long try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
1363 int alloc_flags, const struct alloc_context *ac,
1364 enum migrate_mode mode, int *contended)
1366 int may_enter_fs = gfp_mask & __GFP_FS;
1367 int may_perform_io = gfp_mask & __GFP_IO;
1370 int rc = COMPACT_DEFERRED;
1371 int all_zones_contended = COMPACT_CONTENDED_LOCK; /* init for &= op */
1373 *contended = COMPACT_CONTENDED_NONE;
1375 /* Check if the GFP flags allow compaction */
1376 if (!order || !may_enter_fs || !may_perform_io)
1377 return COMPACT_SKIPPED;
1379 /* Compact each zone in the list */
1380 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1385 if (compaction_deferred(zone, order))
1388 status = compact_zone_order(zone, order, gfp_mask, mode,
1389 &zone_contended, alloc_flags,
1391 rc = max(status, rc);
1393 * It takes at least one zone that wasn't lock contended
1394 * to clear all_zones_contended.
1396 all_zones_contended &= zone_contended;
1398 /* If a normal allocation would succeed, stop compacting */
1399 if (zone_watermark_ok(zone, order, low_wmark_pages(zone),
1400 ac->classzone_idx, alloc_flags)) {
1402 * We think the allocation will succeed in this zone,
1403 * but it is not certain, hence the false. The caller
1404 * will repeat this with true if allocation indeed
1405 * succeeds in this zone.
1407 compaction_defer_reset(zone, order, false);
1409 * It is possible that async compaction aborted due to
1410 * need_resched() and the watermarks were ok thanks to
1411 * somebody else freeing memory. The allocation can
1412 * however still fail so we better signal the
1413 * need_resched() contention anyway (this will not
1414 * prevent the allocation attempt).
1416 if (zone_contended == COMPACT_CONTENDED_SCHED)
1417 *contended = COMPACT_CONTENDED_SCHED;
1422 if (mode != MIGRATE_ASYNC && status == COMPACT_COMPLETE) {
1424 * We think that allocation won't succeed in this zone
1425 * so we defer compaction there. If it ends up
1426 * succeeding after all, it will be reset.
1428 defer_compaction(zone, order);
1432 * We might have stopped compacting due to need_resched() in
1433 * async compaction, or due to a fatal signal detected. In that
1434 * case do not try further zones and signal need_resched()
1437 if ((zone_contended == COMPACT_CONTENDED_SCHED)
1438 || fatal_signal_pending(current)) {
1439 *contended = COMPACT_CONTENDED_SCHED;
1446 * We might not have tried all the zones, so be conservative
1447 * and assume they are not all lock contended.
1449 all_zones_contended = 0;
1454 * If at least one zone wasn't deferred or skipped, we report if all
1455 * zones that were tried were lock contended.
1457 if (rc > COMPACT_SKIPPED && all_zones_contended)
1458 *contended = COMPACT_CONTENDED_LOCK;
1464 /* Compact all zones within a node */
1465 static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
1470 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1472 zone = &pgdat->node_zones[zoneid];
1473 if (!populated_zone(zone))
1476 cc->nr_freepages = 0;
1477 cc->nr_migratepages = 0;
1479 INIT_LIST_HEAD(&cc->freepages);
1480 INIT_LIST_HEAD(&cc->migratepages);
1482 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
1483 compact_zone(zone, cc);
1485 if (cc->order > 0) {
1486 if (zone_watermark_ok(zone, cc->order,
1487 low_wmark_pages(zone), 0, 0))
1488 compaction_defer_reset(zone, cc->order, false);
1491 VM_BUG_ON(!list_empty(&cc->freepages));
1492 VM_BUG_ON(!list_empty(&cc->migratepages));
1496 void compact_pgdat(pg_data_t *pgdat, int order)
1498 struct compact_control cc = {
1500 .mode = MIGRATE_ASYNC,
1506 __compact_pgdat(pgdat, &cc);
1509 static void compact_node(int nid)
1511 struct compact_control cc = {
1513 .mode = MIGRATE_SYNC,
1514 .ignore_skip_hint = true,
1517 __compact_pgdat(NODE_DATA(nid), &cc);
1520 /* Compact all nodes in the system */
1521 static void compact_nodes(void)
1525 /* Flush pending updates to the LRU lists */
1526 lru_add_drain_all();
1528 for_each_online_node(nid)
1532 /* The written value is actually unused, all memory is compacted */
1533 int sysctl_compact_memory;
1535 /* This is the entry point for compacting all nodes via /proc/sys/vm */
1536 int sysctl_compaction_handler(struct ctl_table *table, int write,
1537 void __user *buffer, size_t *length, loff_t *ppos)
1545 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1546 void __user *buffer, size_t *length, loff_t *ppos)
1548 proc_dointvec_minmax(table, write, buffer, length, ppos);
1553 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1554 static ssize_t sysfs_compact_node(struct device *dev,
1555 struct device_attribute *attr,
1556 const char *buf, size_t count)
1560 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1561 /* Flush pending updates to the LRU lists */
1562 lru_add_drain_all();
1569 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1571 int compaction_register_node(struct node *node)
1573 return device_create_file(&node->dev, &dev_attr_compact);
1576 void compaction_unregister_node(struct node *node)
1578 return device_remove_file(&node->dev, &dev_attr_compact);
1580 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1582 #endif /* CONFIG_COMPACTION */