2 * linux/mm/compaction.c
4 * Memory compaction for the reduction of external fragmentation. Note that
5 * this heavily depends upon page migration to do all the real heavy
8 * Copyright IBM Corp. 2007-2010 Mel Gorman <mel@csn.ul.ie>
10 #include <linux/swap.h>
11 #include <linux/migrate.h>
12 #include <linux/compaction.h>
13 #include <linux/mm_inline.h>
14 #include <linux/backing-dev.h>
15 #include <linux/sysctl.h>
16 #include <linux/sysfs.h>
17 #include <linux/balloon_compaction.h>
18 #include <linux/page-isolation.h>
21 #ifdef CONFIG_COMPACTION
22 static inline void count_compact_event(enum vm_event_item item)
27 static inline void count_compact_events(enum vm_event_item item, long delta)
29 count_vm_events(item, delta);
32 #define count_compact_event(item) do { } while (0)
33 #define count_compact_events(item, delta) do { } while (0)
36 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
38 #define CREATE_TRACE_POINTS
39 #include <trace/events/compaction.h>
41 static unsigned long release_freepages(struct list_head *freelist)
43 struct page *page, *next;
44 unsigned long count = 0;
46 list_for_each_entry_safe(page, next, freelist, lru) {
55 static void map_pages(struct list_head *list)
59 list_for_each_entry(page, list, lru) {
60 arch_alloc_page(page, 0);
61 kernel_map_pages(page, 1, 1);
65 static inline bool migrate_async_suitable(int migratetype)
67 return is_migrate_cma(migratetype) || migratetype == MIGRATE_MOVABLE;
70 #ifdef CONFIG_COMPACTION
71 /* Returns true if the pageblock should be scanned for pages to isolate. */
72 static inline bool isolation_suitable(struct compact_control *cc,
75 if (cc->ignore_skip_hint)
78 return !get_pageblock_skip(page);
82 * This function is called to clear all cached information on pageblocks that
83 * should be skipped for page isolation when the migrate and free page scanner
86 static void __reset_isolation_suitable(struct zone *zone)
88 unsigned long start_pfn = zone->zone_start_pfn;
89 unsigned long end_pfn = zone_end_pfn(zone);
92 zone->compact_cached_migrate_pfn[0] = start_pfn;
93 zone->compact_cached_migrate_pfn[1] = start_pfn;
94 zone->compact_cached_free_pfn = end_pfn;
95 zone->compact_blockskip_flush = false;
97 /* Walk the zone and mark every pageblock as suitable for isolation */
98 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
106 page = pfn_to_page(pfn);
107 if (zone != page_zone(page))
110 clear_pageblock_skip(page);
114 void reset_isolation_suitable(pg_data_t *pgdat)
118 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
119 struct zone *zone = &pgdat->node_zones[zoneid];
120 if (!populated_zone(zone))
123 /* Only flush if a full compaction finished recently */
124 if (zone->compact_blockskip_flush)
125 __reset_isolation_suitable(zone);
130 * If no pages were isolated then mark this pageblock to be skipped in the
131 * future. The information is later cleared by __reset_isolation_suitable().
133 static void update_pageblock_skip(struct compact_control *cc,
134 struct page *page, unsigned long nr_isolated,
135 bool set_unsuitable, bool migrate_scanner)
137 struct zone *zone = cc->zone;
140 if (cc->ignore_skip_hint)
150 * Only skip pageblocks when all forms of compaction will be known to
151 * fail in the near future.
154 set_pageblock_skip(page);
156 pfn = page_to_pfn(page);
158 /* Update where async and sync compaction should restart */
159 if (migrate_scanner) {
160 if (cc->finished_update_migrate)
162 if (pfn > zone->compact_cached_migrate_pfn[0])
163 zone->compact_cached_migrate_pfn[0] = pfn;
164 if (cc->mode != MIGRATE_ASYNC &&
165 pfn > zone->compact_cached_migrate_pfn[1])
166 zone->compact_cached_migrate_pfn[1] = pfn;
168 if (cc->finished_update_free)
170 if (pfn < zone->compact_cached_free_pfn)
171 zone->compact_cached_free_pfn = pfn;
175 static inline bool isolation_suitable(struct compact_control *cc,
181 static void update_pageblock_skip(struct compact_control *cc,
182 struct page *page, unsigned long nr_isolated,
183 bool set_unsuitable, bool migrate_scanner)
186 #endif /* CONFIG_COMPACTION */
188 static inline bool should_release_lock(spinlock_t *lock)
190 return need_resched() || spin_is_contended(lock);
194 * Compaction requires the taking of some coarse locks that are potentially
195 * very heavily contended. Check if the process needs to be scheduled or
196 * if the lock is contended. For async compaction, back out in the event
197 * if contention is severe. For sync compaction, schedule.
199 * Returns true if the lock is held.
200 * Returns false if the lock is released and compaction should abort
202 static bool compact_checklock_irqsave(spinlock_t *lock, unsigned long *flags,
203 bool locked, struct compact_control *cc)
205 if (should_release_lock(lock)) {
207 spin_unlock_irqrestore(lock, *flags);
211 /* async aborts if taking too long or contended */
212 if (cc->mode == MIGRATE_ASYNC) {
213 cc->contended = true;
221 spin_lock_irqsave(lock, *flags);
226 * Aside from avoiding lock contention, compaction also periodically checks
227 * need_resched() and either schedules in sync compaction or aborts async
228 * compaction. This is similar to what compact_checklock_irqsave() does, but
229 * is used where no lock is concerned.
231 * Returns false when no scheduling was needed, or sync compaction scheduled.
232 * Returns true when async compaction should abort.
234 static inline bool compact_should_abort(struct compact_control *cc)
236 /* async compaction aborts if contended */
237 if (need_resched()) {
238 if (cc->mode == MIGRATE_ASYNC) {
239 cc->contended = true;
249 /* Returns true if the page is within a block suitable for migration to */
250 static bool suitable_migration_target(struct page *page)
252 /* If the page is a large free page, then disallow migration */
253 if (PageBuddy(page) && page_order(page) >= pageblock_order)
256 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
257 if (migrate_async_suitable(get_pageblock_migratetype(page)))
260 /* Otherwise skip the block */
265 * Isolate free pages onto a private freelist. If @strict is true, will abort
266 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
267 * (even though it may still end up isolating some pages).
269 static unsigned long isolate_freepages_block(struct compact_control *cc,
270 unsigned long blockpfn,
271 unsigned long end_pfn,
272 struct list_head *freelist,
275 int nr_scanned = 0, total_isolated = 0;
276 struct page *cursor, *valid_page = NULL;
279 bool checked_pageblock = false;
281 cursor = pfn_to_page(blockpfn);
283 /* Isolate free pages. */
284 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
286 struct page *page = cursor;
289 if (!pfn_valid_within(blockpfn))
294 if (!PageBuddy(page))
298 * The zone lock must be held to isolate freepages.
299 * Unfortunately this is a very coarse lock and can be
300 * heavily contended if there are parallel allocations
301 * or parallel compactions. For async compaction do not
302 * spin on the lock and we acquire the lock as late as
305 locked = compact_checklock_irqsave(&cc->zone->lock, &flags,
310 /* Recheck this is a suitable migration target under lock */
311 if (!strict && !checked_pageblock) {
313 * We need to check suitability of pageblock only once
314 * and this isolate_freepages_block() is called with
315 * pageblock range, so just check once is sufficient.
317 checked_pageblock = true;
318 if (!suitable_migration_target(page))
322 /* Recheck this is a buddy page under lock */
323 if (!PageBuddy(page))
326 /* Found a free page, break it into order-0 pages */
327 isolated = split_free_page(page);
328 total_isolated += isolated;
329 for (i = 0; i < isolated; i++) {
330 list_add(&page->lru, freelist);
334 /* If a page was split, advance to the end of it */
336 blockpfn += isolated - 1;
337 cursor += isolated - 1;
349 trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
352 * If strict isolation is requested by CMA then check that all the
353 * pages requested were isolated. If there were any failures, 0 is
354 * returned and CMA will fail.
356 if (strict && blockpfn < end_pfn)
360 spin_unlock_irqrestore(&cc->zone->lock, flags);
362 /* Update the pageblock-skip if the whole pageblock was scanned */
363 if (blockpfn == end_pfn)
364 update_pageblock_skip(cc, valid_page, total_isolated, true,
367 count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
369 count_compact_events(COMPACTISOLATED, total_isolated);
370 return total_isolated;
374 * isolate_freepages_range() - isolate free pages.
375 * @start_pfn: The first PFN to start isolating.
376 * @end_pfn: The one-past-last PFN.
378 * Non-free pages, invalid PFNs, or zone boundaries within the
379 * [start_pfn, end_pfn) range are considered errors, cause function to
380 * undo its actions and return zero.
382 * Otherwise, function returns one-past-the-last PFN of isolated page
383 * (which may be greater then end_pfn if end fell in a middle of
387 isolate_freepages_range(struct compact_control *cc,
388 unsigned long start_pfn, unsigned long end_pfn)
390 unsigned long isolated, pfn, block_end_pfn;
393 for (pfn = start_pfn; pfn < end_pfn; pfn += isolated) {
394 if (!pfn_valid(pfn) || cc->zone != page_zone(pfn_to_page(pfn)))
398 * On subsequent iterations ALIGN() is actually not needed,
399 * but we keep it that we not to complicate the code.
401 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
402 block_end_pfn = min(block_end_pfn, end_pfn);
404 isolated = isolate_freepages_block(cc, pfn, block_end_pfn,
408 * In strict mode, isolate_freepages_block() returns 0 if
409 * there are any holes in the block (ie. invalid PFNs or
416 * If we managed to isolate pages, it is always (1 << n) *
417 * pageblock_nr_pages for some non-negative n. (Max order
418 * page may span two pageblocks).
422 /* split_free_page does not map the pages */
423 map_pages(&freelist);
426 /* Loop terminated early, cleanup. */
427 release_freepages(&freelist);
431 /* We don't use freelists for anything. */
435 /* Update the number of anon and file isolated pages in the zone */
436 static void acct_isolated(struct zone *zone, bool locked, struct compact_control *cc)
439 unsigned int count[2] = { 0, };
441 list_for_each_entry(page, &cc->migratepages, lru)
442 count[!!page_is_file_cache(page)]++;
444 /* If locked we can use the interrupt unsafe versions */
446 __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
447 __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
449 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
450 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
454 /* Similar to reclaim, but different enough that they don't share logic */
455 static bool too_many_isolated(struct zone *zone)
457 unsigned long active, inactive, isolated;
459 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
460 zone_page_state(zone, NR_INACTIVE_ANON);
461 active = zone_page_state(zone, NR_ACTIVE_FILE) +
462 zone_page_state(zone, NR_ACTIVE_ANON);
463 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
464 zone_page_state(zone, NR_ISOLATED_ANON);
466 return isolated > (inactive + active) / 2;
470 * isolate_migratepages_range() - isolate all migrate-able pages in range.
471 * @zone: Zone pages are in.
472 * @cc: Compaction control structure.
473 * @low_pfn: The first PFN of the range.
474 * @end_pfn: The one-past-the-last PFN of the range.
475 * @unevictable: true if it allows to isolate unevictable pages
477 * Isolate all pages that can be migrated from the range specified by
478 * [low_pfn, end_pfn). Returns zero if there is a fatal signal
479 * pending), otherwise PFN of the first page that was not scanned
480 * (which may be both less, equal to or more then end_pfn).
482 * Assumes that cc->migratepages is empty and cc->nr_migratepages is
485 * Apart from cc->migratepages and cc->nr_migratetypes this function
486 * does not modify any cc's fields, in particular it does not modify
487 * (or read for that matter) cc->migrate_pfn.
490 isolate_migratepages_range(struct zone *zone, struct compact_control *cc,
491 unsigned long low_pfn, unsigned long end_pfn, bool unevictable)
493 unsigned long last_pageblock_nr = 0, pageblock_nr;
494 unsigned long nr_scanned = 0, nr_isolated = 0;
495 struct list_head *migratelist = &cc->migratepages;
496 struct lruvec *lruvec;
499 struct page *page = NULL, *valid_page = NULL;
500 bool set_unsuitable = true;
501 const isolate_mode_t mode = (cc->mode == MIGRATE_ASYNC ?
502 ISOLATE_ASYNC_MIGRATE : 0) |
503 (unevictable ? ISOLATE_UNEVICTABLE : 0);
506 * Ensure that there are not too many pages isolated from the LRU
507 * list by either parallel reclaimers or compaction. If there are,
508 * delay for some time until fewer pages are isolated
510 while (unlikely(too_many_isolated(zone))) {
511 /* async migration should just abort */
512 if (cc->mode == MIGRATE_ASYNC)
515 congestion_wait(BLK_RW_ASYNC, HZ/10);
517 if (fatal_signal_pending(current))
521 if (compact_should_abort(cc))
524 /* Time to isolate some pages for migration */
525 for (; low_pfn < end_pfn; low_pfn++) {
526 /* give a chance to irqs before checking need_resched() */
527 if (locked && !(low_pfn % SWAP_CLUSTER_MAX)) {
528 if (should_release_lock(&zone->lru_lock)) {
529 spin_unlock_irqrestore(&zone->lru_lock, flags);
535 * migrate_pfn does not necessarily start aligned to a
536 * pageblock. Ensure that pfn_valid is called when moving
537 * into a new MAX_ORDER_NR_PAGES range in case of large
538 * memory holes within the zone
540 if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
541 if (!pfn_valid(low_pfn)) {
542 low_pfn += MAX_ORDER_NR_PAGES - 1;
547 if (!pfn_valid_within(low_pfn))
552 * Get the page and ensure the page is within the same zone.
553 * See the comment in isolate_freepages about overlapping
554 * nodes. It is deliberate that the new zone lock is not taken
555 * as memory compaction should not move pages between nodes.
557 page = pfn_to_page(low_pfn);
558 if (page_zone(page) != zone)
564 /* If isolation recently failed, do not retry */
565 pageblock_nr = low_pfn >> pageblock_order;
566 if (last_pageblock_nr != pageblock_nr) {
569 last_pageblock_nr = pageblock_nr;
570 if (!isolation_suitable(cc, page))
574 * For async migration, also only scan in MOVABLE
575 * blocks. Async migration is optimistic to see if
576 * the minimum amount of work satisfies the allocation
578 mt = get_pageblock_migratetype(page);
579 if (cc->mode == MIGRATE_ASYNC &&
580 !migrate_async_suitable(mt)) {
581 set_unsuitable = false;
587 * Skip if free. page_order cannot be used without zone->lock
588 * as nothing prevents parallel allocations or buddy merging.
594 * Check may be lockless but that's ok as we recheck later.
595 * It's possible to migrate LRU pages and balloon pages
596 * Skip any other type of page
598 if (!PageLRU(page)) {
599 if (unlikely(balloon_page_movable(page))) {
600 if (locked && balloon_page_isolate(page)) {
601 /* Successfully isolated */
602 goto isolate_success;
609 * PageLRU is set. lru_lock normally excludes isolation
610 * splitting and collapsing (collapsing has already happened
611 * if PageLRU is set) but the lock is not necessarily taken
612 * here and it is wasteful to take it just to check transhuge.
613 * Check TransHuge without lock and skip the whole pageblock if
614 * it's either a transhuge or hugetlbfs page, as calling
615 * compound_order() without preventing THP from splitting the
616 * page underneath us may return surprising results.
618 if (PageTransHuge(page)) {
621 low_pfn += (1 << compound_order(page)) - 1;
626 * Migration will fail if an anonymous page is pinned in memory,
627 * so avoid taking lru_lock and isolating it unnecessarily in an
628 * admittedly racy check.
630 if (!page_mapping(page) &&
631 page_count(page) > page_mapcount(page))
634 /* Check if it is ok to still hold the lock */
635 locked = compact_checklock_irqsave(&zone->lru_lock, &flags,
637 if (!locked || fatal_signal_pending(current))
640 /* Recheck PageLRU and PageTransHuge under lock */
643 if (PageTransHuge(page)) {
644 low_pfn += (1 << compound_order(page)) - 1;
648 lruvec = mem_cgroup_page_lruvec(page, zone);
650 /* Try isolate the page */
651 if (__isolate_lru_page(page, mode) != 0)
654 VM_BUG_ON_PAGE(PageTransCompound(page), page);
656 /* Successfully isolated */
657 del_page_from_lru_list(page, lruvec, page_lru(page));
660 cc->finished_update_migrate = true;
661 list_add(&page->lru, migratelist);
662 cc->nr_migratepages++;
665 /* Avoid isolating too much */
666 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
674 low_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages) - 1;
677 acct_isolated(zone, locked, cc);
680 spin_unlock_irqrestore(&zone->lru_lock, flags);
683 * Update the pageblock-skip information and cached scanner pfn,
684 * if the whole pageblock was scanned without isolating any page.
686 if (low_pfn == end_pfn)
687 update_pageblock_skip(cc, valid_page, nr_isolated,
688 set_unsuitable, true);
690 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
692 count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
694 count_compact_events(COMPACTISOLATED, nr_isolated);
699 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
700 #ifdef CONFIG_COMPACTION
702 * Based on information in the current compact_control, find blocks
703 * suitable for isolating free pages from and then isolate them.
705 static void isolate_freepages(struct zone *zone,
706 struct compact_control *cc)
709 unsigned long block_start_pfn; /* start of current pageblock */
710 unsigned long block_end_pfn; /* end of current pageblock */
711 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
712 int nr_freepages = cc->nr_freepages;
713 struct list_head *freelist = &cc->freepages;
716 * Initialise the free scanner. The starting point is where we last
717 * successfully isolated from, zone-cached value, or the end of the
718 * zone when isolating for the first time. We need this aligned to
719 * the pageblock boundary, because we do
720 * block_start_pfn -= pageblock_nr_pages in the for loop.
721 * For ending point, take care when isolating in last pageblock of a
722 * a zone which ends in the middle of a pageblock.
723 * The low boundary is the end of the pageblock the migration scanner
726 block_start_pfn = cc->free_pfn & ~(pageblock_nr_pages-1);
727 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
729 low_pfn = ALIGN(cc->migrate_pfn + 1, pageblock_nr_pages);
732 * Isolate free pages until enough are available to migrate the
733 * pages on cc->migratepages. We stop searching if the migrate
734 * and free page scanners meet or enough free pages are isolated.
736 for (; block_start_pfn >= low_pfn && cc->nr_migratepages > nr_freepages;
737 block_end_pfn = block_start_pfn,
738 block_start_pfn -= pageblock_nr_pages) {
739 unsigned long isolated;
742 * This can iterate a massively long zone without finding any
743 * suitable migration targets, so periodically check if we need
744 * to schedule, or even abort async compaction.
746 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
747 && compact_should_abort(cc))
750 if (!pfn_valid(block_start_pfn))
754 * Check for overlapping nodes/zones. It's possible on some
755 * configurations to have a setup like
757 * i.e. it's possible that all pages within a zones range of
758 * pages do not belong to a single zone.
760 page = pfn_to_page(block_start_pfn);
761 if (page_zone(page) != zone)
764 /* Check the block is suitable for migration */
765 if (!suitable_migration_target(page))
768 /* If isolation recently failed, do not retry */
769 if (!isolation_suitable(cc, page))
772 /* Found a block suitable for isolating free pages from */
773 cc->free_pfn = block_start_pfn;
774 isolated = isolate_freepages_block(cc, block_start_pfn,
775 block_end_pfn, freelist, false);
776 nr_freepages += isolated;
779 * Set a flag that we successfully isolated in this pageblock.
780 * In the next loop iteration, zone->compact_cached_free_pfn
781 * will not be updated and thus it will effectively contain the
782 * highest pageblock we isolated pages from.
785 cc->finished_update_free = true;
788 * isolate_freepages_block() might have aborted due to async
789 * compaction being contended
795 /* split_free_page does not map the pages */
799 * If we crossed the migrate scanner, we want to keep it that way
800 * so that compact_finished() may detect this
802 if (block_start_pfn < low_pfn)
803 cc->free_pfn = cc->migrate_pfn;
805 cc->nr_freepages = nr_freepages;
809 * This is a migrate-callback that "allocates" freepages by taking pages
810 * from the isolated freelists in the block we are migrating to.
812 static struct page *compaction_alloc(struct page *migratepage,
816 struct compact_control *cc = (struct compact_control *)data;
817 struct page *freepage;
820 * Isolate free pages if necessary, and if we are not aborting due to
823 if (list_empty(&cc->freepages)) {
825 isolate_freepages(cc->zone, cc);
827 if (list_empty(&cc->freepages))
831 freepage = list_entry(cc->freepages.next, struct page, lru);
832 list_del(&freepage->lru);
839 * This is a migrate-callback that "frees" freepages back to the isolated
840 * freelist. All pages on the freelist are from the same zone, so there is no
841 * special handling needed for NUMA.
843 static void compaction_free(struct page *page, unsigned long data)
845 struct compact_control *cc = (struct compact_control *)data;
847 list_add(&page->lru, &cc->freepages);
851 /* possible outcome of isolate_migratepages */
853 ISOLATE_ABORT, /* Abort compaction now */
854 ISOLATE_NONE, /* No pages isolated, continue scanning */
855 ISOLATE_SUCCESS, /* Pages isolated, migrate */
859 * Isolate all pages that can be migrated from the block pointed to by
860 * the migrate scanner within compact_control.
862 static isolate_migrate_t isolate_migratepages(struct zone *zone,
863 struct compact_control *cc)
865 unsigned long low_pfn, end_pfn;
867 /* Do not scan outside zone boundaries */
868 low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
870 /* Only scan within a pageblock boundary */
871 end_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages);
873 /* Do not cross the free scanner or scan within a memory hole */
874 if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
875 cc->migrate_pfn = end_pfn;
879 /* Perform the isolation */
880 low_pfn = isolate_migratepages_range(zone, cc, low_pfn, end_pfn, false);
881 if (!low_pfn || cc->contended)
882 return ISOLATE_ABORT;
884 cc->migrate_pfn = low_pfn;
886 return ISOLATE_SUCCESS;
889 static int compact_finished(struct zone *zone,
890 struct compact_control *cc)
893 unsigned long watermark;
895 if (cc->contended || fatal_signal_pending(current))
896 return COMPACT_PARTIAL;
898 /* Compaction run completes if the migrate and free scanner meet */
899 if (cc->free_pfn <= cc->migrate_pfn) {
900 /* Let the next compaction start anew. */
901 zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
902 zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
903 zone->compact_cached_free_pfn = zone_end_pfn(zone);
906 * Mark that the PG_migrate_skip information should be cleared
907 * by kswapd when it goes to sleep. kswapd does not set the
908 * flag itself as the decision to be clear should be directly
909 * based on an allocation request.
911 if (!current_is_kswapd())
912 zone->compact_blockskip_flush = true;
914 return COMPACT_COMPLETE;
918 * order == -1 is expected when compacting via
919 * /proc/sys/vm/compact_memory
922 return COMPACT_CONTINUE;
924 /* Compaction run is not finished if the watermark is not met */
925 watermark = low_wmark_pages(zone);
926 watermark += (1 << cc->order);
928 if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
929 return COMPACT_CONTINUE;
931 /* Direct compactor: Is a suitable page free? */
932 for (order = cc->order; order < MAX_ORDER; order++) {
933 struct free_area *area = &zone->free_area[order];
935 /* Job done if page is free of the right migratetype */
936 if (!list_empty(&area->free_list[cc->migratetype]))
937 return COMPACT_PARTIAL;
939 /* Job done if allocation would set block type */
940 if (cc->order >= pageblock_order && area->nr_free)
941 return COMPACT_PARTIAL;
944 return COMPACT_CONTINUE;
948 * compaction_suitable: Is this suitable to run compaction on this zone now?
950 * COMPACT_SKIPPED - If there are too few free pages for compaction
951 * COMPACT_PARTIAL - If the allocation would succeed without compaction
952 * COMPACT_CONTINUE - If compaction should run now
954 unsigned long compaction_suitable(struct zone *zone, int order)
957 unsigned long watermark;
960 * order == -1 is expected when compacting via
961 * /proc/sys/vm/compact_memory
964 return COMPACT_CONTINUE;
967 * Watermarks for order-0 must be met for compaction. Note the 2UL.
968 * This is because during migration, copies of pages need to be
969 * allocated and for a short time, the footprint is higher
971 watermark = low_wmark_pages(zone) + (2UL << order);
972 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
973 return COMPACT_SKIPPED;
976 * fragmentation index determines if allocation failures are due to
977 * low memory or external fragmentation
979 * index of -1000 implies allocations might succeed depending on
981 * index towards 0 implies failure is due to lack of memory
982 * index towards 1000 implies failure is due to fragmentation
984 * Only compact if a failure would be due to fragmentation.
986 fragindex = fragmentation_index(zone, order);
987 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
988 return COMPACT_SKIPPED;
990 if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
992 return COMPACT_PARTIAL;
994 return COMPACT_CONTINUE;
997 static int compact_zone(struct zone *zone, struct compact_control *cc)
1000 unsigned long start_pfn = zone->zone_start_pfn;
1001 unsigned long end_pfn = zone_end_pfn(zone);
1002 const bool sync = cc->mode != MIGRATE_ASYNC;
1004 ret = compaction_suitable(zone, cc->order);
1006 case COMPACT_PARTIAL:
1007 case COMPACT_SKIPPED:
1008 /* Compaction is likely to fail */
1010 case COMPACT_CONTINUE:
1011 /* Fall through to compaction */
1016 * Clear pageblock skip if there were failures recently and compaction
1017 * is about to be retried after being deferred. kswapd does not do
1018 * this reset as it'll reset the cached information when going to sleep.
1020 if (compaction_restarting(zone, cc->order) && !current_is_kswapd())
1021 __reset_isolation_suitable(zone);
1024 * Setup to move all movable pages to the end of the zone. Used cached
1025 * information on where the scanners should start but check that it
1026 * is initialised by ensuring the values are within zone boundaries.
1028 cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync];
1029 cc->free_pfn = zone->compact_cached_free_pfn;
1030 if (cc->free_pfn < start_pfn || cc->free_pfn > end_pfn) {
1031 cc->free_pfn = end_pfn & ~(pageblock_nr_pages-1);
1032 zone->compact_cached_free_pfn = cc->free_pfn;
1034 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn > end_pfn) {
1035 cc->migrate_pfn = start_pfn;
1036 zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
1037 zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
1040 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn, cc->free_pfn, end_pfn);
1042 migrate_prep_local();
1044 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
1047 switch (isolate_migratepages(zone, cc)) {
1049 ret = COMPACT_PARTIAL;
1050 putback_movable_pages(&cc->migratepages);
1051 cc->nr_migratepages = 0;
1055 case ISOLATE_SUCCESS:
1059 if (!cc->nr_migratepages)
1062 err = migrate_pages(&cc->migratepages, compaction_alloc,
1063 compaction_free, (unsigned long)cc, cc->mode,
1066 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
1069 /* All pages were either migrated or will be released */
1070 cc->nr_migratepages = 0;
1072 putback_movable_pages(&cc->migratepages);
1074 * migrate_pages() may return -ENOMEM when scanners meet
1075 * and we want compact_finished() to detect it
1077 if (err == -ENOMEM && cc->free_pfn > cc->migrate_pfn) {
1078 ret = COMPACT_PARTIAL;
1085 /* Release free pages and check accounting */
1086 cc->nr_freepages -= release_freepages(&cc->freepages);
1087 VM_BUG_ON(cc->nr_freepages != 0);
1089 trace_mm_compaction_end(ret);
1094 static unsigned long compact_zone_order(struct zone *zone, int order,
1095 gfp_t gfp_mask, enum migrate_mode mode, bool *contended)
1098 struct compact_control cc = {
1100 .nr_migratepages = 0,
1102 .migratetype = allocflags_to_migratetype(gfp_mask),
1106 INIT_LIST_HEAD(&cc.freepages);
1107 INIT_LIST_HEAD(&cc.migratepages);
1109 ret = compact_zone(zone, &cc);
1111 VM_BUG_ON(!list_empty(&cc.freepages));
1112 VM_BUG_ON(!list_empty(&cc.migratepages));
1114 *contended = cc.contended;
1118 int sysctl_extfrag_threshold = 500;
1121 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1122 * @zonelist: The zonelist used for the current allocation
1123 * @order: The order of the current allocation
1124 * @gfp_mask: The GFP mask of the current allocation
1125 * @nodemask: The allowed nodes to allocate from
1126 * @mode: The migration mode for async, sync light, or sync migration
1127 * @contended: Return value that is true if compaction was aborted due to lock contention
1128 * @page: Optionally capture a free page of the requested order during compaction
1130 * This is the main entry point for direct page compaction.
1132 unsigned long try_to_compact_pages(struct zonelist *zonelist,
1133 int order, gfp_t gfp_mask, nodemask_t *nodemask,
1134 enum migrate_mode mode, bool *contended)
1136 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1137 int may_enter_fs = gfp_mask & __GFP_FS;
1138 int may_perform_io = gfp_mask & __GFP_IO;
1141 int rc = COMPACT_SKIPPED;
1142 int alloc_flags = 0;
1144 /* Check if the GFP flags allow compaction */
1145 if (!order || !may_enter_fs || !may_perform_io)
1148 count_compact_event(COMPACTSTALL);
1151 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
1152 alloc_flags |= ALLOC_CMA;
1154 /* Compact each zone in the list */
1155 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1159 status = compact_zone_order(zone, order, gfp_mask, mode,
1161 rc = max(status, rc);
1163 /* If a normal allocation would succeed, stop compacting */
1164 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0,
1173 /* Compact all zones within a node */
1174 static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
1179 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1181 zone = &pgdat->node_zones[zoneid];
1182 if (!populated_zone(zone))
1185 cc->nr_freepages = 0;
1186 cc->nr_migratepages = 0;
1188 INIT_LIST_HEAD(&cc->freepages);
1189 INIT_LIST_HEAD(&cc->migratepages);
1191 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
1192 compact_zone(zone, cc);
1194 if (cc->order > 0) {
1195 if (zone_watermark_ok(zone, cc->order,
1196 low_wmark_pages(zone), 0, 0))
1197 compaction_defer_reset(zone, cc->order, false);
1200 VM_BUG_ON(!list_empty(&cc->freepages));
1201 VM_BUG_ON(!list_empty(&cc->migratepages));
1205 void compact_pgdat(pg_data_t *pgdat, int order)
1207 struct compact_control cc = {
1209 .mode = MIGRATE_ASYNC,
1215 __compact_pgdat(pgdat, &cc);
1218 static void compact_node(int nid)
1220 struct compact_control cc = {
1222 .mode = MIGRATE_SYNC,
1223 .ignore_skip_hint = true,
1226 __compact_pgdat(NODE_DATA(nid), &cc);
1229 /* Compact all nodes in the system */
1230 static void compact_nodes(void)
1234 /* Flush pending updates to the LRU lists */
1235 lru_add_drain_all();
1237 for_each_online_node(nid)
1241 /* The written value is actually unused, all memory is compacted */
1242 int sysctl_compact_memory;
1244 /* This is the entry point for compacting all nodes via /proc/sys/vm */
1245 int sysctl_compaction_handler(struct ctl_table *table, int write,
1246 void __user *buffer, size_t *length, loff_t *ppos)
1254 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1255 void __user *buffer, size_t *length, loff_t *ppos)
1257 proc_dointvec_minmax(table, write, buffer, length, ppos);
1262 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1263 ssize_t sysfs_compact_node(struct device *dev,
1264 struct device_attribute *attr,
1265 const char *buf, size_t count)
1269 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1270 /* Flush pending updates to the LRU lists */
1271 lru_add_drain_all();
1278 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1280 int compaction_register_node(struct node *node)
1282 return device_create_file(&node->dev, &dev_attr_compact);
1285 void compaction_unregister_node(struct node *node)
1287 return device_remove_file(&node->dev, &dev_attr_compact);
1289 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1291 #endif /* CONFIG_COMPACTION */