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;
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);
482 if (!pageblock_pfn_to_page(pfn, block_end_pfn, cc->zone))
485 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
486 block_end_pfn, &freelist, true);
489 * In strict mode, isolate_freepages_block() returns 0 if
490 * there are any holes in the block (ie. invalid PFNs or
497 * If we managed to isolate pages, it is always (1 << n) *
498 * pageblock_nr_pages for some non-negative n. (Max order
499 * page may span two pageblocks).
503 /* split_free_page does not map the pages */
504 map_pages(&freelist);
507 /* Loop terminated early, cleanup. */
508 release_freepages(&freelist);
512 /* We don't use freelists for anything. */
516 /* Update the number of anon and file isolated pages in the zone */
517 static void acct_isolated(struct zone *zone, struct compact_control *cc)
520 unsigned int count[2] = { 0, };
522 if (list_empty(&cc->migratepages))
525 list_for_each_entry(page, &cc->migratepages, lru)
526 count[!!page_is_file_cache(page)]++;
528 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
529 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
532 /* Similar to reclaim, but different enough that they don't share logic */
533 static bool too_many_isolated(struct zone *zone)
535 unsigned long active, inactive, isolated;
537 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
538 zone_page_state(zone, NR_INACTIVE_ANON);
539 active = zone_page_state(zone, NR_ACTIVE_FILE) +
540 zone_page_state(zone, NR_ACTIVE_ANON);
541 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
542 zone_page_state(zone, NR_ISOLATED_ANON);
544 return isolated > (inactive + active) / 2;
548 * isolate_migratepages_block() - isolate all migrate-able pages within
550 * @cc: Compaction control structure.
551 * @low_pfn: The first PFN to isolate
552 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
553 * @isolate_mode: Isolation mode to be used.
555 * Isolate all pages that can be migrated from the range specified by
556 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
557 * Returns zero if there is a fatal signal pending, otherwise PFN of the
558 * first page that was not scanned (which may be both less, equal to or more
561 * The pages are isolated on cc->migratepages list (not required to be empty),
562 * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
563 * is neither read nor updated.
566 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
567 unsigned long end_pfn, isolate_mode_t isolate_mode)
569 struct zone *zone = cc->zone;
570 unsigned long nr_scanned = 0, nr_isolated = 0;
571 struct list_head *migratelist = &cc->migratepages;
572 struct lruvec *lruvec;
575 struct page *page = NULL, *valid_page = NULL;
578 * Ensure that there are not too many pages isolated from the LRU
579 * list by either parallel reclaimers or compaction. If there are,
580 * delay for some time until fewer pages are isolated
582 while (unlikely(too_many_isolated(zone))) {
583 /* async migration should just abort */
584 if (cc->mode == MIGRATE_ASYNC)
587 congestion_wait(BLK_RW_ASYNC, HZ/10);
589 if (fatal_signal_pending(current))
593 if (compact_should_abort(cc))
596 /* Time to isolate some pages for migration */
597 for (; low_pfn < end_pfn; low_pfn++) {
599 * Periodically drop the lock (if held) regardless of its
600 * contention, to give chance to IRQs. Abort async compaction
603 if (!(low_pfn % SWAP_CLUSTER_MAX)
604 && compact_unlock_should_abort(&zone->lru_lock, flags,
608 if (!pfn_valid_within(low_pfn))
612 page = pfn_to_page(low_pfn);
618 * Skip if free. We read page order here without zone lock
619 * which is generally unsafe, but the race window is small and
620 * the worst thing that can happen is that we skip some
621 * potential isolation targets.
623 if (PageBuddy(page)) {
624 unsigned long freepage_order = page_order_unsafe(page);
627 * Without lock, we cannot be sure that what we got is
628 * a valid page order. Consider only values in the
629 * valid order range to prevent low_pfn overflow.
631 if (freepage_order > 0 && freepage_order < MAX_ORDER)
632 low_pfn += (1UL << freepage_order) - 1;
637 * Check may be lockless but that's ok as we recheck later.
638 * It's possible to migrate LRU pages and balloon pages
639 * Skip any other type of page
641 if (!PageLRU(page)) {
642 if (unlikely(balloon_page_movable(page))) {
643 if (locked && balloon_page_isolate(page)) {
644 /* Successfully isolated */
645 goto isolate_success;
652 * PageLRU is set. lru_lock normally excludes isolation
653 * splitting and collapsing (collapsing has already happened
654 * if PageLRU is set) but the lock is not necessarily taken
655 * here and it is wasteful to take it just to check transhuge.
656 * Check TransHuge without lock and skip the whole pageblock if
657 * it's either a transhuge or hugetlbfs page, as calling
658 * compound_order() without preventing THP from splitting the
659 * page underneath us may return surprising results.
661 if (PageTransHuge(page)) {
663 low_pfn = ALIGN(low_pfn + 1,
664 pageblock_nr_pages) - 1;
666 low_pfn += (1 << compound_order(page)) - 1;
672 * Migration will fail if an anonymous page is pinned in memory,
673 * so avoid taking lru_lock and isolating it unnecessarily in an
674 * admittedly racy check.
676 if (!page_mapping(page) &&
677 page_count(page) > page_mapcount(page))
680 /* If we already hold the lock, we can skip some rechecking */
682 locked = compact_trylock_irqsave(&zone->lru_lock,
687 /* Recheck PageLRU and PageTransHuge under lock */
690 if (PageTransHuge(page)) {
691 low_pfn += (1 << compound_order(page)) - 1;
696 lruvec = mem_cgroup_page_lruvec(page, zone);
698 /* Try isolate the page */
699 if (__isolate_lru_page(page, isolate_mode) != 0)
702 VM_BUG_ON_PAGE(PageTransCompound(page), page);
704 /* Successfully isolated */
705 del_page_from_lru_list(page, lruvec, page_lru(page));
708 cc->finished_update_migrate = true;
709 list_add(&page->lru, migratelist);
710 cc->nr_migratepages++;
713 /* Avoid isolating too much */
714 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
721 * The PageBuddy() check could have potentially brought us outside
722 * the range to be scanned.
724 if (unlikely(low_pfn > end_pfn))
728 spin_unlock_irqrestore(&zone->lru_lock, flags);
731 * Update the pageblock-skip information and cached scanner pfn,
732 * if the whole pageblock was scanned without isolating any page.
734 if (low_pfn == end_pfn)
735 update_pageblock_skip(cc, valid_page, nr_isolated, true);
737 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
739 count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
741 count_compact_events(COMPACTISOLATED, nr_isolated);
747 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
748 * @cc: Compaction control structure.
749 * @start_pfn: The first PFN to start isolating.
750 * @end_pfn: The one-past-last PFN.
752 * Returns zero if isolation fails fatally due to e.g. pending signal.
753 * Otherwise, function returns one-past-the-last PFN of isolated page
754 * (which may be greater than end_pfn if end fell in a middle of a THP page).
757 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
758 unsigned long end_pfn)
760 unsigned long pfn, block_end_pfn;
762 /* Scan block by block. First and last block may be incomplete */
764 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
766 for (; pfn < end_pfn; pfn = block_end_pfn,
767 block_end_pfn += pageblock_nr_pages) {
769 block_end_pfn = min(block_end_pfn, end_pfn);
771 if (!pageblock_pfn_to_page(pfn, block_end_pfn, cc->zone))
774 pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
775 ISOLATE_UNEVICTABLE);
778 * In case of fatal failure, release everything that might
779 * have been isolated in the previous iteration, and signal
780 * the failure back to caller.
783 putback_movable_pages(&cc->migratepages);
784 cc->nr_migratepages = 0;
788 acct_isolated(cc->zone, cc);
793 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
794 #ifdef CONFIG_COMPACTION
796 * Based on information in the current compact_control, find blocks
797 * suitable for isolating free pages from and then isolate them.
799 static void isolate_freepages(struct compact_control *cc)
801 struct zone *zone = cc->zone;
803 unsigned long block_start_pfn; /* start of current pageblock */
804 unsigned long isolate_start_pfn; /* exact pfn we start at */
805 unsigned long block_end_pfn; /* end of current pageblock */
806 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
807 int nr_freepages = cc->nr_freepages;
808 struct list_head *freelist = &cc->freepages;
811 * Initialise the free scanner. The starting point is where we last
812 * successfully isolated from, zone-cached value, or the end of the
813 * zone when isolating for the first time. For looping we also need
814 * this pfn aligned down to the pageblock boundary, because we do
815 * block_start_pfn -= pageblock_nr_pages in the for loop.
816 * For ending point, take care when isolating in last pageblock of a
817 * a zone which ends in the middle of a pageblock.
818 * The low boundary is the end of the pageblock the migration scanner
821 isolate_start_pfn = cc->free_pfn;
822 block_start_pfn = cc->free_pfn & ~(pageblock_nr_pages-1);
823 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
825 low_pfn = ALIGN(cc->migrate_pfn + 1, pageblock_nr_pages);
828 * Isolate free pages until enough are available to migrate the
829 * pages on cc->migratepages. We stop searching if the migrate
830 * and free page scanners meet or enough free pages are isolated.
832 for (; block_start_pfn >= low_pfn && cc->nr_migratepages > nr_freepages;
833 block_end_pfn = block_start_pfn,
834 block_start_pfn -= pageblock_nr_pages,
835 isolate_start_pfn = block_start_pfn) {
836 unsigned long isolated;
839 * This can iterate a massively long zone without finding any
840 * suitable migration targets, so periodically check if we need
841 * to schedule, or even abort async compaction.
843 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
844 && compact_should_abort(cc))
847 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
852 /* Check the block is suitable for migration */
853 if (!suitable_migration_target(page))
856 /* If isolation recently failed, do not retry */
857 if (!isolation_suitable(cc, page))
860 /* Found a block suitable for isolating free pages from. */
861 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
862 block_end_pfn, freelist, false);
863 nr_freepages += isolated;
866 * Remember where the free scanner should restart next time,
867 * which is where isolate_freepages_block() left off.
868 * But if it scanned the whole pageblock, isolate_start_pfn
869 * now points at block_end_pfn, which is the start of the next
871 * In that case we will however want to restart at the start
872 * of the previous pageblock.
874 cc->free_pfn = (isolate_start_pfn < block_end_pfn) ?
876 block_start_pfn - pageblock_nr_pages;
879 * Set a flag that we successfully isolated in this pageblock.
880 * In the next loop iteration, zone->compact_cached_free_pfn
881 * will not be updated and thus it will effectively contain the
882 * highest pageblock we isolated pages from.
885 cc->finished_update_free = true;
888 * isolate_freepages_block() might have aborted due to async
889 * compaction being contended
895 /* split_free_page does not map the pages */
899 * If we crossed the migrate scanner, we want to keep it that way
900 * so that compact_finished() may detect this
902 if (block_start_pfn < low_pfn)
903 cc->free_pfn = cc->migrate_pfn;
905 cc->nr_freepages = nr_freepages;
909 * This is a migrate-callback that "allocates" freepages by taking pages
910 * from the isolated freelists in the block we are migrating to.
912 static struct page *compaction_alloc(struct page *migratepage,
916 struct compact_control *cc = (struct compact_control *)data;
917 struct page *freepage;
920 * Isolate free pages if necessary, and if we are not aborting due to
923 if (list_empty(&cc->freepages)) {
925 isolate_freepages(cc);
927 if (list_empty(&cc->freepages))
931 freepage = list_entry(cc->freepages.next, struct page, lru);
932 list_del(&freepage->lru);
939 * This is a migrate-callback that "frees" freepages back to the isolated
940 * freelist. All pages on the freelist are from the same zone, so there is no
941 * special handling needed for NUMA.
943 static void compaction_free(struct page *page, unsigned long data)
945 struct compact_control *cc = (struct compact_control *)data;
947 list_add(&page->lru, &cc->freepages);
951 /* possible outcome of isolate_migratepages */
953 ISOLATE_ABORT, /* Abort compaction now */
954 ISOLATE_NONE, /* No pages isolated, continue scanning */
955 ISOLATE_SUCCESS, /* Pages isolated, migrate */
959 * Isolate all pages that can be migrated from the first suitable block,
960 * starting at the block pointed to by the migrate scanner pfn within
963 static isolate_migrate_t isolate_migratepages(struct zone *zone,
964 struct compact_control *cc)
966 unsigned long low_pfn, end_pfn;
968 const isolate_mode_t isolate_mode =
969 (cc->mode == MIGRATE_ASYNC ? ISOLATE_ASYNC_MIGRATE : 0);
972 * Start at where we last stopped, or beginning of the zone as
973 * initialized by compact_zone()
975 low_pfn = cc->migrate_pfn;
977 /* Only scan within a pageblock boundary */
978 end_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages);
981 * Iterate over whole pageblocks until we find the first suitable.
982 * Do not cross the free scanner.
984 for (; end_pfn <= cc->free_pfn;
985 low_pfn = end_pfn, end_pfn += pageblock_nr_pages) {
988 * This can potentially iterate a massively long zone with
989 * many pageblocks unsuitable, so periodically check if we
990 * need to schedule, or even abort async compaction.
992 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
993 && compact_should_abort(cc))
996 page = pageblock_pfn_to_page(low_pfn, end_pfn, zone);
1000 /* If isolation recently failed, do not retry */
1001 if (!isolation_suitable(cc, page))
1005 * For async compaction, also only scan in MOVABLE blocks.
1006 * Async compaction is optimistic to see if the minimum amount
1007 * of work satisfies the allocation.
1009 if (cc->mode == MIGRATE_ASYNC &&
1010 !migrate_async_suitable(get_pageblock_migratetype(page)))
1013 /* Perform the isolation */
1014 low_pfn = isolate_migratepages_block(cc, low_pfn, end_pfn,
1017 if (!low_pfn || cc->contended)
1018 return ISOLATE_ABORT;
1021 * Either we isolated something and proceed with migration. Or
1022 * we failed and compact_zone should decide if we should
1028 acct_isolated(zone, cc);
1029 /* Record where migration scanner will be restarted */
1030 cc->migrate_pfn = low_pfn;
1032 return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
1035 static int compact_finished(struct zone *zone,
1036 struct compact_control *cc)
1039 unsigned long watermark;
1041 if (cc->contended || fatal_signal_pending(current))
1042 return COMPACT_PARTIAL;
1044 /* Compaction run completes if the migrate and free scanner meet */
1045 if (cc->free_pfn <= cc->migrate_pfn) {
1046 /* Let the next compaction start anew. */
1047 zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
1048 zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
1049 zone->compact_cached_free_pfn = zone_end_pfn(zone);
1052 * Mark that the PG_migrate_skip information should be cleared
1053 * by kswapd when it goes to sleep. kswapd does not set the
1054 * flag itself as the decision to be clear should be directly
1055 * based on an allocation request.
1057 if (!current_is_kswapd())
1058 zone->compact_blockskip_flush = true;
1060 return COMPACT_COMPLETE;
1064 * order == -1 is expected when compacting via
1065 * /proc/sys/vm/compact_memory
1067 if (cc->order == -1)
1068 return COMPACT_CONTINUE;
1070 /* Compaction run is not finished if the watermark is not met */
1071 watermark = low_wmark_pages(zone);
1072 watermark += (1 << cc->order);
1074 if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
1075 return COMPACT_CONTINUE;
1077 /* Direct compactor: Is a suitable page free? */
1078 for (order = cc->order; order < MAX_ORDER; order++) {
1079 struct free_area *area = &zone->free_area[order];
1081 /* Job done if page is free of the right migratetype */
1082 if (!list_empty(&area->free_list[cc->migratetype]))
1083 return COMPACT_PARTIAL;
1085 /* Job done if allocation would set block type */
1086 if (cc->order >= pageblock_order && area->nr_free)
1087 return COMPACT_PARTIAL;
1090 return COMPACT_CONTINUE;
1094 * compaction_suitable: Is this suitable to run compaction on this zone now?
1096 * COMPACT_SKIPPED - If there are too few free pages for compaction
1097 * COMPACT_PARTIAL - If the allocation would succeed without compaction
1098 * COMPACT_CONTINUE - If compaction should run now
1100 unsigned long compaction_suitable(struct zone *zone, int order)
1103 unsigned long watermark;
1106 * order == -1 is expected when compacting via
1107 * /proc/sys/vm/compact_memory
1110 return COMPACT_CONTINUE;
1113 * Watermarks for order-0 must be met for compaction. Note the 2UL.
1114 * This is because during migration, copies of pages need to be
1115 * allocated and for a short time, the footprint is higher
1117 watermark = low_wmark_pages(zone) + (2UL << order);
1118 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1119 return COMPACT_SKIPPED;
1122 * fragmentation index determines if allocation failures are due to
1123 * low memory or external fragmentation
1125 * index of -1000 implies allocations might succeed depending on
1127 * index towards 0 implies failure is due to lack of memory
1128 * index towards 1000 implies failure is due to fragmentation
1130 * Only compact if a failure would be due to fragmentation.
1132 fragindex = fragmentation_index(zone, order);
1133 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
1134 return COMPACT_SKIPPED;
1136 if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
1138 return COMPACT_PARTIAL;
1140 return COMPACT_CONTINUE;
1143 static int compact_zone(struct zone *zone, struct compact_control *cc)
1146 unsigned long start_pfn = zone->zone_start_pfn;
1147 unsigned long end_pfn = zone_end_pfn(zone);
1148 const bool sync = cc->mode != MIGRATE_ASYNC;
1150 ret = compaction_suitable(zone, cc->order);
1152 case COMPACT_PARTIAL:
1153 case COMPACT_SKIPPED:
1154 /* Compaction is likely to fail */
1156 case COMPACT_CONTINUE:
1157 /* Fall through to compaction */
1162 * Clear pageblock skip if there were failures recently and compaction
1163 * is about to be retried after being deferred. kswapd does not do
1164 * this reset as it'll reset the cached information when going to sleep.
1166 if (compaction_restarting(zone, cc->order) && !current_is_kswapd())
1167 __reset_isolation_suitable(zone);
1170 * Setup to move all movable pages to the end of the zone. Used cached
1171 * information on where the scanners should start but check that it
1172 * is initialised by ensuring the values are within zone boundaries.
1174 cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync];
1175 cc->free_pfn = zone->compact_cached_free_pfn;
1176 if (cc->free_pfn < start_pfn || cc->free_pfn > end_pfn) {
1177 cc->free_pfn = end_pfn & ~(pageblock_nr_pages-1);
1178 zone->compact_cached_free_pfn = cc->free_pfn;
1180 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn > end_pfn) {
1181 cc->migrate_pfn = start_pfn;
1182 zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
1183 zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
1186 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn, cc->free_pfn, end_pfn);
1188 migrate_prep_local();
1190 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
1193 switch (isolate_migratepages(zone, cc)) {
1195 ret = COMPACT_PARTIAL;
1196 putback_movable_pages(&cc->migratepages);
1197 cc->nr_migratepages = 0;
1201 case ISOLATE_SUCCESS:
1205 err = migrate_pages(&cc->migratepages, compaction_alloc,
1206 compaction_free, (unsigned long)cc, cc->mode,
1209 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
1212 /* All pages were either migrated or will be released */
1213 cc->nr_migratepages = 0;
1215 putback_movable_pages(&cc->migratepages);
1217 * migrate_pages() may return -ENOMEM when scanners meet
1218 * and we want compact_finished() to detect it
1220 if (err == -ENOMEM && cc->free_pfn > cc->migrate_pfn) {
1221 ret = COMPACT_PARTIAL;
1228 /* Release free pages and check accounting */
1229 cc->nr_freepages -= release_freepages(&cc->freepages);
1230 VM_BUG_ON(cc->nr_freepages != 0);
1232 trace_mm_compaction_end(ret);
1237 static unsigned long compact_zone_order(struct zone *zone, int order,
1238 gfp_t gfp_mask, enum migrate_mode mode, int *contended)
1241 struct compact_control cc = {
1243 .nr_migratepages = 0,
1245 .migratetype = allocflags_to_migratetype(gfp_mask),
1249 INIT_LIST_HEAD(&cc.freepages);
1250 INIT_LIST_HEAD(&cc.migratepages);
1252 ret = compact_zone(zone, &cc);
1254 VM_BUG_ON(!list_empty(&cc.freepages));
1255 VM_BUG_ON(!list_empty(&cc.migratepages));
1257 *contended = cc.contended;
1261 int sysctl_extfrag_threshold = 500;
1264 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1265 * @zonelist: The zonelist used for the current allocation
1266 * @order: The order of the current allocation
1267 * @gfp_mask: The GFP mask of the current allocation
1268 * @nodemask: The allowed nodes to allocate from
1269 * @mode: The migration mode for async, sync light, or sync migration
1270 * @contended: Return value that determines if compaction was aborted due to
1271 * need_resched() or lock contention
1272 * @candidate_zone: Return the zone where we think allocation should succeed
1274 * This is the main entry point for direct page compaction.
1276 unsigned long try_to_compact_pages(struct zonelist *zonelist,
1277 int order, gfp_t gfp_mask, nodemask_t *nodemask,
1278 enum migrate_mode mode, int *contended,
1279 struct zone **candidate_zone)
1281 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1282 int may_enter_fs = gfp_mask & __GFP_FS;
1283 int may_perform_io = gfp_mask & __GFP_IO;
1286 int rc = COMPACT_DEFERRED;
1287 int alloc_flags = 0;
1288 int all_zones_contended = COMPACT_CONTENDED_LOCK; /* init for &= op */
1290 *contended = COMPACT_CONTENDED_NONE;
1292 /* Check if the GFP flags allow compaction */
1293 if (!order || !may_enter_fs || !may_perform_io)
1294 return COMPACT_SKIPPED;
1297 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
1298 alloc_flags |= ALLOC_CMA;
1300 /* Compact each zone in the list */
1301 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1306 if (compaction_deferred(zone, order))
1309 status = compact_zone_order(zone, order, gfp_mask, mode,
1311 rc = max(status, rc);
1313 * It takes at least one zone that wasn't lock contended
1314 * to clear all_zones_contended.
1316 all_zones_contended &= zone_contended;
1318 /* If a normal allocation would succeed, stop compacting */
1319 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0,
1321 *candidate_zone = zone;
1323 * We think the allocation will succeed in this zone,
1324 * but it is not certain, hence the false. The caller
1325 * will repeat this with true if allocation indeed
1326 * succeeds in this zone.
1328 compaction_defer_reset(zone, order, false);
1330 * It is possible that async compaction aborted due to
1331 * need_resched() and the watermarks were ok thanks to
1332 * somebody else freeing memory. The allocation can
1333 * however still fail so we better signal the
1334 * need_resched() contention anyway (this will not
1335 * prevent the allocation attempt).
1337 if (zone_contended == COMPACT_CONTENDED_SCHED)
1338 *contended = COMPACT_CONTENDED_SCHED;
1343 if (mode != MIGRATE_ASYNC) {
1345 * We think that allocation won't succeed in this zone
1346 * so we defer compaction there. If it ends up
1347 * succeeding after all, it will be reset.
1349 defer_compaction(zone, order);
1353 * We might have stopped compacting due to need_resched() in
1354 * async compaction, or due to a fatal signal detected. In that
1355 * case do not try further zones and signal need_resched()
1358 if ((zone_contended == COMPACT_CONTENDED_SCHED)
1359 || fatal_signal_pending(current)) {
1360 *contended = COMPACT_CONTENDED_SCHED;
1367 * We might not have tried all the zones, so be conservative
1368 * and assume they are not all lock contended.
1370 all_zones_contended = 0;
1375 * If at least one zone wasn't deferred or skipped, we report if all
1376 * zones that were tried were lock contended.
1378 if (rc > COMPACT_SKIPPED && all_zones_contended)
1379 *contended = COMPACT_CONTENDED_LOCK;
1385 /* Compact all zones within a node */
1386 static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
1391 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1393 zone = &pgdat->node_zones[zoneid];
1394 if (!populated_zone(zone))
1397 cc->nr_freepages = 0;
1398 cc->nr_migratepages = 0;
1400 INIT_LIST_HEAD(&cc->freepages);
1401 INIT_LIST_HEAD(&cc->migratepages);
1403 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
1404 compact_zone(zone, cc);
1406 if (cc->order > 0) {
1407 if (zone_watermark_ok(zone, cc->order,
1408 low_wmark_pages(zone), 0, 0))
1409 compaction_defer_reset(zone, cc->order, false);
1412 VM_BUG_ON(!list_empty(&cc->freepages));
1413 VM_BUG_ON(!list_empty(&cc->migratepages));
1417 void compact_pgdat(pg_data_t *pgdat, int order)
1419 struct compact_control cc = {
1421 .mode = MIGRATE_ASYNC,
1427 __compact_pgdat(pgdat, &cc);
1430 static void compact_node(int nid)
1432 struct compact_control cc = {
1434 .mode = MIGRATE_SYNC,
1435 .ignore_skip_hint = true,
1438 __compact_pgdat(NODE_DATA(nid), &cc);
1441 /* Compact all nodes in the system */
1442 static void compact_nodes(void)
1446 /* Flush pending updates to the LRU lists */
1447 lru_add_drain_all();
1449 for_each_online_node(nid)
1453 /* The written value is actually unused, all memory is compacted */
1454 int sysctl_compact_memory;
1456 /* This is the entry point for compacting all nodes via /proc/sys/vm */
1457 int sysctl_compaction_handler(struct ctl_table *table, int write,
1458 void __user *buffer, size_t *length, loff_t *ppos)
1466 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1467 void __user *buffer, size_t *length, loff_t *ppos)
1469 proc_dointvec_minmax(table, write, buffer, length, ppos);
1474 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1475 static ssize_t sysfs_compact_node(struct device *dev,
1476 struct device_attribute *attr,
1477 const char *buf, size_t count)
1481 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1482 /* Flush pending updates to the LRU lists */
1483 lru_add_drain_all();
1490 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1492 int compaction_register_node(struct node *node)
1494 return device_create_file(&node->dev, &dev_attr_compact);
1497 void compaction_unregister_node(struct node *node)
1499 return device_remove_file(&node->dev, &dev_attr_compact);
1501 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1503 #endif /* CONFIG_COMPACTION */