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)) {
318 * We are checking page_order without zone->lock taken. But
319 * the only small danger is that we skip a potentially suitable
320 * pageblock, so it's not worth to check order for valid range.
322 if (page_order_unsafe(page) >= pageblock_order)
326 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
327 if (migrate_async_suitable(get_pageblock_migratetype(page)))
330 /* Otherwise skip the block */
335 * Isolate free pages onto a private freelist. If @strict is true, will abort
336 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
337 * (even though it may still end up isolating some pages).
339 static unsigned long isolate_freepages_block(struct compact_control *cc,
340 unsigned long *start_pfn,
341 unsigned long end_pfn,
342 struct list_head *freelist,
345 int nr_scanned = 0, total_isolated = 0;
346 struct page *cursor, *valid_page = NULL;
347 unsigned long flags = 0;
349 unsigned long blockpfn = *start_pfn;
351 cursor = pfn_to_page(blockpfn);
353 /* Isolate free pages. */
354 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
356 struct page *page = cursor;
359 * Periodically drop the lock (if held) regardless of its
360 * contention, to give chance to IRQs. Abort if fatal signal
361 * pending or async compaction detects need_resched()
363 if (!(blockpfn % SWAP_CLUSTER_MAX)
364 && compact_unlock_should_abort(&cc->zone->lock, flags,
369 if (!pfn_valid_within(blockpfn))
374 if (!PageBuddy(page))
378 * If we already hold the lock, we can skip some rechecking.
379 * Note that if we hold the lock now, checked_pageblock was
380 * already set in some previous iteration (or strict is true),
381 * so it is correct to skip the suitable migration target
386 * The zone lock must be held to isolate freepages.
387 * Unfortunately this is a very coarse lock and can be
388 * heavily contended if there are parallel allocations
389 * or parallel compactions. For async compaction do not
390 * spin on the lock and we acquire the lock as late as
393 locked = compact_trylock_irqsave(&cc->zone->lock,
398 /* Recheck this is a buddy page under lock */
399 if (!PageBuddy(page))
403 /* Found a free page, break it into order-0 pages */
404 isolated = split_free_page(page);
405 total_isolated += isolated;
406 for (i = 0; i < isolated; i++) {
407 list_add(&page->lru, freelist);
411 /* If a page was split, advance to the end of it */
413 blockpfn += isolated - 1;
414 cursor += isolated - 1;
426 /* Record how far we have got within the block */
427 *start_pfn = blockpfn;
429 trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
432 * If strict isolation is requested by CMA then check that all the
433 * pages requested were isolated. If there were any failures, 0 is
434 * returned and CMA will fail.
436 if (strict && blockpfn < end_pfn)
440 spin_unlock_irqrestore(&cc->zone->lock, flags);
442 /* Update the pageblock-skip if the whole pageblock was scanned */
443 if (blockpfn == end_pfn)
444 update_pageblock_skip(cc, valid_page, total_isolated, false);
446 count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
448 count_compact_events(COMPACTISOLATED, total_isolated);
449 return total_isolated;
453 * isolate_freepages_range() - isolate free pages.
454 * @start_pfn: The first PFN to start isolating.
455 * @end_pfn: The one-past-last PFN.
457 * Non-free pages, invalid PFNs, or zone boundaries within the
458 * [start_pfn, end_pfn) range are considered errors, cause function to
459 * undo its actions and return zero.
461 * Otherwise, function returns one-past-the-last PFN of isolated page
462 * (which may be greater then end_pfn if end fell in a middle of
466 isolate_freepages_range(struct compact_control *cc,
467 unsigned long start_pfn, unsigned long end_pfn)
469 unsigned long isolated, pfn, block_end_pfn;
473 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
475 for (; pfn < end_pfn; pfn += isolated,
476 block_end_pfn += pageblock_nr_pages) {
477 /* Protect pfn from changing by isolate_freepages_block */
478 unsigned long isolate_start_pfn = pfn;
480 block_end_pfn = min(block_end_pfn, end_pfn);
483 * pfn could pass the block_end_pfn if isolated freepage
484 * is more than pageblock order. In this case, we adjust
485 * scanning range to right one.
487 if (pfn >= block_end_pfn) {
488 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
489 block_end_pfn = min(block_end_pfn, end_pfn);
492 if (!pageblock_pfn_to_page(pfn, block_end_pfn, cc->zone))
495 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
496 block_end_pfn, &freelist, true);
499 * In strict mode, isolate_freepages_block() returns 0 if
500 * there are any holes in the block (ie. invalid PFNs or
507 * If we managed to isolate pages, it is always (1 << n) *
508 * pageblock_nr_pages for some non-negative n. (Max order
509 * page may span two pageblocks).
513 /* split_free_page does not map the pages */
514 map_pages(&freelist);
517 /* Loop terminated early, cleanup. */
518 release_freepages(&freelist);
522 /* We don't use freelists for anything. */
526 /* Update the number of anon and file isolated pages in the zone */
527 static void acct_isolated(struct zone *zone, struct compact_control *cc)
530 unsigned int count[2] = { 0, };
532 if (list_empty(&cc->migratepages))
535 list_for_each_entry(page, &cc->migratepages, lru)
536 count[!!page_is_file_cache(page)]++;
538 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
539 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
542 /* Similar to reclaim, but different enough that they don't share logic */
543 static bool too_many_isolated(struct zone *zone)
545 unsigned long active, inactive, isolated;
547 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
548 zone_page_state(zone, NR_INACTIVE_ANON);
549 active = zone_page_state(zone, NR_ACTIVE_FILE) +
550 zone_page_state(zone, NR_ACTIVE_ANON);
551 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
552 zone_page_state(zone, NR_ISOLATED_ANON);
554 return isolated > (inactive + active) / 2;
558 * isolate_migratepages_block() - isolate all migrate-able pages within
560 * @cc: Compaction control structure.
561 * @low_pfn: The first PFN to isolate
562 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
563 * @isolate_mode: Isolation mode to be used.
565 * Isolate all pages that can be migrated from the range specified by
566 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
567 * Returns zero if there is a fatal signal pending, otherwise PFN of the
568 * first page that was not scanned (which may be both less, equal to or more
571 * The pages are isolated on cc->migratepages list (not required to be empty),
572 * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
573 * is neither read nor updated.
576 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
577 unsigned long end_pfn, isolate_mode_t isolate_mode)
579 struct zone *zone = cc->zone;
580 unsigned long nr_scanned = 0, nr_isolated = 0;
581 struct list_head *migratelist = &cc->migratepages;
582 struct lruvec *lruvec;
583 unsigned long flags = 0;
585 struct page *page = NULL, *valid_page = NULL;
588 * Ensure that there are not too many pages isolated from the LRU
589 * list by either parallel reclaimers or compaction. If there are,
590 * delay for some time until fewer pages are isolated
592 while (unlikely(too_many_isolated(zone))) {
593 /* async migration should just abort */
594 if (cc->mode == MIGRATE_ASYNC)
597 congestion_wait(BLK_RW_ASYNC, HZ/10);
599 if (fatal_signal_pending(current))
603 if (compact_should_abort(cc))
606 /* Time to isolate some pages for migration */
607 for (; low_pfn < end_pfn; low_pfn++) {
609 * Periodically drop the lock (if held) regardless of its
610 * contention, to give chance to IRQs. Abort async compaction
613 if (!(low_pfn % SWAP_CLUSTER_MAX)
614 && compact_unlock_should_abort(&zone->lru_lock, flags,
618 if (!pfn_valid_within(low_pfn))
622 page = pfn_to_page(low_pfn);
628 * Skip if free. We read page order here without zone lock
629 * which is generally unsafe, but the race window is small and
630 * the worst thing that can happen is that we skip some
631 * potential isolation targets.
633 if (PageBuddy(page)) {
634 unsigned long freepage_order = page_order_unsafe(page);
637 * Without lock, we cannot be sure that what we got is
638 * a valid page order. Consider only values in the
639 * valid order range to prevent low_pfn overflow.
641 if (freepage_order > 0 && freepage_order < MAX_ORDER)
642 low_pfn += (1UL << freepage_order) - 1;
647 * Check may be lockless but that's ok as we recheck later.
648 * It's possible to migrate LRU pages and balloon pages
649 * Skip any other type of page
651 if (!PageLRU(page)) {
652 if (unlikely(balloon_page_movable(page))) {
653 if (balloon_page_isolate(page)) {
654 /* Successfully isolated */
655 goto isolate_success;
662 * PageLRU is set. lru_lock normally excludes isolation
663 * splitting and collapsing (collapsing has already happened
664 * if PageLRU is set) but the lock is not necessarily taken
665 * here and it is wasteful to take it just to check transhuge.
666 * Check TransHuge without lock and skip the whole pageblock if
667 * it's either a transhuge or hugetlbfs page, as calling
668 * compound_order() without preventing THP from splitting the
669 * page underneath us may return surprising results.
671 if (PageTransHuge(page)) {
673 low_pfn = ALIGN(low_pfn + 1,
674 pageblock_nr_pages) - 1;
676 low_pfn += (1 << compound_order(page)) - 1;
682 * Migration will fail if an anonymous page is pinned in memory,
683 * so avoid taking lru_lock and isolating it unnecessarily in an
684 * admittedly racy check.
686 if (!page_mapping(page) &&
687 page_count(page) > page_mapcount(page))
690 /* If we already hold the lock, we can skip some rechecking */
692 locked = compact_trylock_irqsave(&zone->lru_lock,
697 /* Recheck PageLRU and PageTransHuge under lock */
700 if (PageTransHuge(page)) {
701 low_pfn += (1 << compound_order(page)) - 1;
706 lruvec = mem_cgroup_page_lruvec(page, zone);
708 /* Try isolate the page */
709 if (__isolate_lru_page(page, isolate_mode) != 0)
712 VM_BUG_ON_PAGE(PageTransCompound(page), page);
714 /* Successfully isolated */
715 del_page_from_lru_list(page, lruvec, page_lru(page));
718 cc->finished_update_migrate = true;
719 list_add(&page->lru, migratelist);
720 cc->nr_migratepages++;
723 /* Avoid isolating too much */
724 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
731 * The PageBuddy() check could have potentially brought us outside
732 * the range to be scanned.
734 if (unlikely(low_pfn > end_pfn))
738 spin_unlock_irqrestore(&zone->lru_lock, flags);
741 * Update the pageblock-skip information and cached scanner pfn,
742 * if the whole pageblock was scanned without isolating any page.
744 if (low_pfn == end_pfn)
745 update_pageblock_skip(cc, valid_page, nr_isolated, true);
747 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
749 count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
751 count_compact_events(COMPACTISOLATED, nr_isolated);
757 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
758 * @cc: Compaction control structure.
759 * @start_pfn: The first PFN to start isolating.
760 * @end_pfn: The one-past-last PFN.
762 * Returns zero if isolation fails fatally due to e.g. pending signal.
763 * Otherwise, function returns one-past-the-last PFN of isolated page
764 * (which may be greater than end_pfn if end fell in a middle of a THP page).
767 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
768 unsigned long end_pfn)
770 unsigned long pfn, block_end_pfn;
772 /* Scan block by block. First and last block may be incomplete */
774 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
776 for (; pfn < end_pfn; pfn = block_end_pfn,
777 block_end_pfn += pageblock_nr_pages) {
779 block_end_pfn = min(block_end_pfn, end_pfn);
781 if (!pageblock_pfn_to_page(pfn, block_end_pfn, cc->zone))
784 pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
785 ISOLATE_UNEVICTABLE);
788 * In case of fatal failure, release everything that might
789 * have been isolated in the previous iteration, and signal
790 * the failure back to caller.
793 putback_movable_pages(&cc->migratepages);
794 cc->nr_migratepages = 0;
798 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
801 acct_isolated(cc->zone, cc);
806 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
807 #ifdef CONFIG_COMPACTION
809 * Based on information in the current compact_control, find blocks
810 * suitable for isolating free pages from and then isolate them.
812 static void isolate_freepages(struct compact_control *cc)
814 struct zone *zone = cc->zone;
816 unsigned long block_start_pfn; /* start of current pageblock */
817 unsigned long isolate_start_pfn; /* exact pfn we start at */
818 unsigned long block_end_pfn; /* end of current pageblock */
819 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
820 int nr_freepages = cc->nr_freepages;
821 struct list_head *freelist = &cc->freepages;
824 * Initialise the free scanner. The starting point is where we last
825 * successfully isolated from, zone-cached value, or the end of the
826 * zone when isolating for the first time. For looping we also need
827 * this pfn aligned down to the pageblock boundary, because we do
828 * block_start_pfn -= pageblock_nr_pages in the for loop.
829 * For ending point, take care when isolating in last pageblock of a
830 * a zone which ends in the middle of a pageblock.
831 * The low boundary is the end of the pageblock the migration scanner
834 isolate_start_pfn = cc->free_pfn;
835 block_start_pfn = cc->free_pfn & ~(pageblock_nr_pages-1);
836 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
838 low_pfn = ALIGN(cc->migrate_pfn + 1, pageblock_nr_pages);
841 * Isolate free pages until enough are available to migrate the
842 * pages on cc->migratepages. We stop searching if the migrate
843 * and free page scanners meet or enough free pages are isolated.
845 for (; block_start_pfn >= low_pfn && cc->nr_migratepages > nr_freepages;
846 block_end_pfn = block_start_pfn,
847 block_start_pfn -= pageblock_nr_pages,
848 isolate_start_pfn = block_start_pfn) {
849 unsigned long isolated;
852 * This can iterate a massively long zone without finding any
853 * suitable migration targets, so periodically check if we need
854 * to schedule, or even abort async compaction.
856 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
857 && compact_should_abort(cc))
860 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
865 /* Check the block is suitable for migration */
866 if (!suitable_migration_target(page))
869 /* If isolation recently failed, do not retry */
870 if (!isolation_suitable(cc, page))
873 /* Found a block suitable for isolating free pages from. */
874 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
875 block_end_pfn, freelist, false);
876 nr_freepages += isolated;
879 * Remember where the free scanner should restart next time,
880 * which is where isolate_freepages_block() left off.
881 * But if it scanned the whole pageblock, isolate_start_pfn
882 * now points at block_end_pfn, which is the start of the next
884 * In that case we will however want to restart at the start
885 * of the previous pageblock.
887 cc->free_pfn = (isolate_start_pfn < block_end_pfn) ?
889 block_start_pfn - pageblock_nr_pages;
892 * Set a flag that we successfully isolated in this pageblock.
893 * In the next loop iteration, zone->compact_cached_free_pfn
894 * will not be updated and thus it will effectively contain the
895 * highest pageblock we isolated pages from.
898 cc->finished_update_free = true;
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);
1089 watermark += (1 << cc->order);
1091 if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
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)
1120 unsigned long watermark;
1123 * order == -1 is expected when compacting via
1124 * /proc/sys/vm/compact_memory
1127 return COMPACT_CONTINUE;
1130 * Watermarks for order-0 must be met for compaction. Note the 2UL.
1131 * This is because during migration, copies of pages need to be
1132 * allocated and for a short time, the footprint is higher
1134 watermark = low_wmark_pages(zone) + (2UL << order);
1135 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1136 return COMPACT_SKIPPED;
1139 * fragmentation index determines if allocation failures are due to
1140 * low memory or external fragmentation
1142 * index of -1000 implies allocations might succeed depending on
1144 * index towards 0 implies failure is due to lack of memory
1145 * index towards 1000 implies failure is due to fragmentation
1147 * Only compact if a failure would be due to fragmentation.
1149 fragindex = fragmentation_index(zone, order);
1150 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
1151 return COMPACT_SKIPPED;
1153 if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
1155 return COMPACT_PARTIAL;
1157 return COMPACT_CONTINUE;
1160 static int compact_zone(struct zone *zone, struct compact_control *cc)
1163 unsigned long start_pfn = zone->zone_start_pfn;
1164 unsigned long end_pfn = zone_end_pfn(zone);
1165 const int migratetype = gfpflags_to_migratetype(cc->gfp_mask);
1166 const bool sync = cc->mode != MIGRATE_ASYNC;
1168 ret = compaction_suitable(zone, cc->order);
1170 case COMPACT_PARTIAL:
1171 case COMPACT_SKIPPED:
1172 /* Compaction is likely to fail */
1174 case COMPACT_CONTINUE:
1175 /* Fall through to compaction */
1180 * Clear pageblock skip if there were failures recently and compaction
1181 * is about to be retried after being deferred. kswapd does not do
1182 * this reset as it'll reset the cached information when going to sleep.
1184 if (compaction_restarting(zone, cc->order) && !current_is_kswapd())
1185 __reset_isolation_suitable(zone);
1188 * Setup to move all movable pages to the end of the zone. Used cached
1189 * information on where the scanners should start but check that it
1190 * is initialised by ensuring the values are within zone boundaries.
1192 cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync];
1193 cc->free_pfn = zone->compact_cached_free_pfn;
1194 if (cc->free_pfn < start_pfn || cc->free_pfn > end_pfn) {
1195 cc->free_pfn = end_pfn & ~(pageblock_nr_pages-1);
1196 zone->compact_cached_free_pfn = cc->free_pfn;
1198 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn > end_pfn) {
1199 cc->migrate_pfn = start_pfn;
1200 zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
1201 zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
1204 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn, cc->free_pfn, end_pfn);
1206 migrate_prep_local();
1208 while ((ret = compact_finished(zone, cc, migratetype)) ==
1212 switch (isolate_migratepages(zone, cc)) {
1214 ret = COMPACT_PARTIAL;
1215 putback_movable_pages(&cc->migratepages);
1216 cc->nr_migratepages = 0;
1220 case ISOLATE_SUCCESS:
1224 err = migrate_pages(&cc->migratepages, compaction_alloc,
1225 compaction_free, (unsigned long)cc, cc->mode,
1228 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
1231 /* All pages were either migrated or will be released */
1232 cc->nr_migratepages = 0;
1234 putback_movable_pages(&cc->migratepages);
1236 * migrate_pages() may return -ENOMEM when scanners meet
1237 * and we want compact_finished() to detect it
1239 if (err == -ENOMEM && cc->free_pfn > cc->migrate_pfn) {
1240 ret = COMPACT_PARTIAL;
1247 /* Release free pages and check accounting */
1248 cc->nr_freepages -= release_freepages(&cc->freepages);
1249 VM_BUG_ON(cc->nr_freepages != 0);
1251 trace_mm_compaction_end(ret);
1256 static unsigned long compact_zone_order(struct zone *zone, int order,
1257 gfp_t gfp_mask, enum migrate_mode mode, int *contended)
1260 struct compact_control cc = {
1262 .nr_migratepages = 0,
1264 .gfp_mask = gfp_mask,
1268 INIT_LIST_HEAD(&cc.freepages);
1269 INIT_LIST_HEAD(&cc.migratepages);
1271 ret = compact_zone(zone, &cc);
1273 VM_BUG_ON(!list_empty(&cc.freepages));
1274 VM_BUG_ON(!list_empty(&cc.migratepages));
1276 *contended = cc.contended;
1280 int sysctl_extfrag_threshold = 500;
1283 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1284 * @zonelist: The zonelist used for the current allocation
1285 * @order: The order of the current allocation
1286 * @gfp_mask: The GFP mask of the current allocation
1287 * @nodemask: The allowed nodes to allocate from
1288 * @mode: The migration mode for async, sync light, or sync migration
1289 * @contended: Return value that determines if compaction was aborted due to
1290 * need_resched() or lock contention
1291 * @candidate_zone: Return the zone where we think allocation should succeed
1293 * This is the main entry point for direct page compaction.
1295 unsigned long try_to_compact_pages(struct zonelist *zonelist,
1296 int order, gfp_t gfp_mask, nodemask_t *nodemask,
1297 enum migrate_mode mode, int *contended,
1298 struct zone **candidate_zone)
1300 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1301 int may_enter_fs = gfp_mask & __GFP_FS;
1302 int may_perform_io = gfp_mask & __GFP_IO;
1305 int rc = COMPACT_DEFERRED;
1306 int alloc_flags = 0;
1307 int all_zones_contended = COMPACT_CONTENDED_LOCK; /* init for &= op */
1309 *contended = COMPACT_CONTENDED_NONE;
1311 /* Check if the GFP flags allow compaction */
1312 if (!order || !may_enter_fs || !may_perform_io)
1313 return COMPACT_SKIPPED;
1316 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
1317 alloc_flags |= ALLOC_CMA;
1319 /* Compact each zone in the list */
1320 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1325 if (compaction_deferred(zone, order))
1328 status = compact_zone_order(zone, order, gfp_mask, mode,
1330 rc = max(status, rc);
1332 * It takes at least one zone that wasn't lock contended
1333 * to clear all_zones_contended.
1335 all_zones_contended &= zone_contended;
1337 /* If a normal allocation would succeed, stop compacting */
1338 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0,
1340 *candidate_zone = zone;
1342 * We think the allocation will succeed in this zone,
1343 * but it is not certain, hence the false. The caller
1344 * will repeat this with true if allocation indeed
1345 * succeeds in this zone.
1347 compaction_defer_reset(zone, order, false);
1349 * It is possible that async compaction aborted due to
1350 * need_resched() and the watermarks were ok thanks to
1351 * somebody else freeing memory. The allocation can
1352 * however still fail so we better signal the
1353 * need_resched() contention anyway (this will not
1354 * prevent the allocation attempt).
1356 if (zone_contended == COMPACT_CONTENDED_SCHED)
1357 *contended = COMPACT_CONTENDED_SCHED;
1362 if (mode != MIGRATE_ASYNC) {
1364 * We think that allocation won't succeed in this zone
1365 * so we defer compaction there. If it ends up
1366 * succeeding after all, it will be reset.
1368 defer_compaction(zone, order);
1372 * We might have stopped compacting due to need_resched() in
1373 * async compaction, or due to a fatal signal detected. In that
1374 * case do not try further zones and signal need_resched()
1377 if ((zone_contended == COMPACT_CONTENDED_SCHED)
1378 || fatal_signal_pending(current)) {
1379 *contended = COMPACT_CONTENDED_SCHED;
1386 * We might not have tried all the zones, so be conservative
1387 * and assume they are not all lock contended.
1389 all_zones_contended = 0;
1394 * If at least one zone wasn't deferred or skipped, we report if all
1395 * zones that were tried were lock contended.
1397 if (rc > COMPACT_SKIPPED && all_zones_contended)
1398 *contended = COMPACT_CONTENDED_LOCK;
1404 /* Compact all zones within a node */
1405 static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
1410 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1412 zone = &pgdat->node_zones[zoneid];
1413 if (!populated_zone(zone))
1416 cc->nr_freepages = 0;
1417 cc->nr_migratepages = 0;
1419 INIT_LIST_HEAD(&cc->freepages);
1420 INIT_LIST_HEAD(&cc->migratepages);
1422 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
1423 compact_zone(zone, cc);
1425 if (cc->order > 0) {
1426 if (zone_watermark_ok(zone, cc->order,
1427 low_wmark_pages(zone), 0, 0))
1428 compaction_defer_reset(zone, cc->order, false);
1431 VM_BUG_ON(!list_empty(&cc->freepages));
1432 VM_BUG_ON(!list_empty(&cc->migratepages));
1436 void compact_pgdat(pg_data_t *pgdat, int order)
1438 struct compact_control cc = {
1440 .mode = MIGRATE_ASYNC,
1446 __compact_pgdat(pgdat, &cc);
1449 static void compact_node(int nid)
1451 struct compact_control cc = {
1453 .mode = MIGRATE_SYNC,
1454 .ignore_skip_hint = true,
1457 __compact_pgdat(NODE_DATA(nid), &cc);
1460 /* Compact all nodes in the system */
1461 static void compact_nodes(void)
1465 /* Flush pending updates to the LRU lists */
1466 lru_add_drain_all();
1468 for_each_online_node(nid)
1472 /* The written value is actually unused, all memory is compacted */
1473 int sysctl_compact_memory;
1475 /* This is the entry point for compacting all nodes via /proc/sys/vm */
1476 int sysctl_compaction_handler(struct ctl_table *table, int write,
1477 void __user *buffer, size_t *length, loff_t *ppos)
1485 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1486 void __user *buffer, size_t *length, loff_t *ppos)
1488 proc_dointvec_minmax(table, write, buffer, length, ppos);
1493 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1494 static ssize_t sysfs_compact_node(struct device *dev,
1495 struct device_attribute *attr,
1496 const char *buf, size_t count)
1500 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1501 /* Flush pending updates to the LRU lists */
1502 lru_add_drain_all();
1509 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1511 int compaction_register_node(struct node *node)
1513 return device_create_file(&node->dev, &dev_attr_compact);
1516 void compaction_unregister_node(struct node *node)
1518 return device_remove_file(&node->dev, &dev_attr_compact);
1520 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1522 #endif /* CONFIG_COMPACTION */