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);
225 /* Returns true if the page is within a block suitable for migration to */
226 static bool suitable_migration_target(struct page *page)
228 /* If the page is a large free page, then disallow migration */
229 if (PageBuddy(page) && page_order(page) >= pageblock_order)
232 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
233 if (migrate_async_suitable(get_pageblock_migratetype(page)))
236 /* Otherwise skip the block */
241 * Isolate free pages onto a private freelist. If @strict is true, will abort
242 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
243 * (even though it may still end up isolating some pages).
245 static unsigned long isolate_freepages_block(struct compact_control *cc,
246 unsigned long blockpfn,
247 unsigned long end_pfn,
248 struct list_head *freelist,
251 int nr_scanned = 0, total_isolated = 0;
252 struct page *cursor, *valid_page = NULL;
255 bool checked_pageblock = false;
257 cursor = pfn_to_page(blockpfn);
259 /* Isolate free pages. */
260 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
262 struct page *page = cursor;
265 if (!pfn_valid_within(blockpfn))
270 if (!PageBuddy(page))
274 * The zone lock must be held to isolate freepages.
275 * Unfortunately this is a very coarse lock and can be
276 * heavily contended if there are parallel allocations
277 * or parallel compactions. For async compaction do not
278 * spin on the lock and we acquire the lock as late as
281 locked = compact_checklock_irqsave(&cc->zone->lock, &flags,
286 /* Recheck this is a suitable migration target under lock */
287 if (!strict && !checked_pageblock) {
289 * We need to check suitability of pageblock only once
290 * and this isolate_freepages_block() is called with
291 * pageblock range, so just check once is sufficient.
293 checked_pageblock = true;
294 if (!suitable_migration_target(page))
298 /* Recheck this is a buddy page under lock */
299 if (!PageBuddy(page))
302 /* Found a free page, break it into order-0 pages */
303 isolated = split_free_page(page);
304 total_isolated += isolated;
305 for (i = 0; i < isolated; i++) {
306 list_add(&page->lru, freelist);
310 /* If a page was split, advance to the end of it */
312 blockpfn += isolated - 1;
313 cursor += isolated - 1;
325 trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
328 * If strict isolation is requested by CMA then check that all the
329 * pages requested were isolated. If there were any failures, 0 is
330 * returned and CMA will fail.
332 if (strict && blockpfn < end_pfn)
336 spin_unlock_irqrestore(&cc->zone->lock, flags);
338 /* Update the pageblock-skip if the whole pageblock was scanned */
339 if (blockpfn == end_pfn)
340 update_pageblock_skip(cc, valid_page, total_isolated, true,
343 count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
345 count_compact_events(COMPACTISOLATED, total_isolated);
346 return total_isolated;
350 * isolate_freepages_range() - isolate free pages.
351 * @start_pfn: The first PFN to start isolating.
352 * @end_pfn: The one-past-last PFN.
354 * Non-free pages, invalid PFNs, or zone boundaries within the
355 * [start_pfn, end_pfn) range are considered errors, cause function to
356 * undo its actions and return zero.
358 * Otherwise, function returns one-past-the-last PFN of isolated page
359 * (which may be greater then end_pfn if end fell in a middle of
363 isolate_freepages_range(struct compact_control *cc,
364 unsigned long start_pfn, unsigned long end_pfn)
366 unsigned long isolated, pfn, block_end_pfn;
369 for (pfn = start_pfn; pfn < end_pfn; pfn += isolated) {
370 if (!pfn_valid(pfn) || cc->zone != page_zone(pfn_to_page(pfn)))
374 * On subsequent iterations ALIGN() is actually not needed,
375 * but we keep it that we not to complicate the code.
377 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
378 block_end_pfn = min(block_end_pfn, end_pfn);
380 isolated = isolate_freepages_block(cc, pfn, block_end_pfn,
384 * In strict mode, isolate_freepages_block() returns 0 if
385 * there are any holes in the block (ie. invalid PFNs or
392 * If we managed to isolate pages, it is always (1 << n) *
393 * pageblock_nr_pages for some non-negative n. (Max order
394 * page may span two pageblocks).
398 /* split_free_page does not map the pages */
399 map_pages(&freelist);
402 /* Loop terminated early, cleanup. */
403 release_freepages(&freelist);
407 /* We don't use freelists for anything. */
411 /* Update the number of anon and file isolated pages in the zone */
412 static void acct_isolated(struct zone *zone, bool locked, struct compact_control *cc)
415 unsigned int count[2] = { 0, };
417 list_for_each_entry(page, &cc->migratepages, lru)
418 count[!!page_is_file_cache(page)]++;
420 /* If locked we can use the interrupt unsafe versions */
422 __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
423 __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
425 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
426 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
430 /* Similar to reclaim, but different enough that they don't share logic */
431 static bool too_many_isolated(struct zone *zone)
433 unsigned long active, inactive, isolated;
435 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
436 zone_page_state(zone, NR_INACTIVE_ANON);
437 active = zone_page_state(zone, NR_ACTIVE_FILE) +
438 zone_page_state(zone, NR_ACTIVE_ANON);
439 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
440 zone_page_state(zone, NR_ISOLATED_ANON);
442 return isolated > (inactive + active) / 2;
446 * isolate_migratepages_range() - isolate all migrate-able pages in range.
447 * @zone: Zone pages are in.
448 * @cc: Compaction control structure.
449 * @low_pfn: The first PFN of the range.
450 * @end_pfn: The one-past-the-last PFN of the range.
451 * @unevictable: true if it allows to isolate unevictable pages
453 * Isolate all pages that can be migrated from the range specified by
454 * [low_pfn, end_pfn). Returns zero if there is a fatal signal
455 * pending), otherwise PFN of the first page that was not scanned
456 * (which may be both less, equal to or more then end_pfn).
458 * Assumes that cc->migratepages is empty and cc->nr_migratepages is
461 * Apart from cc->migratepages and cc->nr_migratetypes this function
462 * does not modify any cc's fields, in particular it does not modify
463 * (or read for that matter) cc->migrate_pfn.
466 isolate_migratepages_range(struct zone *zone, struct compact_control *cc,
467 unsigned long low_pfn, unsigned long end_pfn, bool unevictable)
469 unsigned long last_pageblock_nr = 0, pageblock_nr;
470 unsigned long nr_scanned = 0, nr_isolated = 0;
471 struct list_head *migratelist = &cc->migratepages;
472 struct lruvec *lruvec;
475 struct page *page = NULL, *valid_page = NULL;
476 bool set_unsuitable = true;
477 const isolate_mode_t mode = (cc->mode == MIGRATE_ASYNC ?
478 ISOLATE_ASYNC_MIGRATE : 0) |
479 (unevictable ? ISOLATE_UNEVICTABLE : 0);
482 * Ensure that there are not too many pages isolated from the LRU
483 * list by either parallel reclaimers or compaction. If there are,
484 * delay for some time until fewer pages are isolated
486 while (unlikely(too_many_isolated(zone))) {
487 /* async migration should just abort */
488 if (cc->mode == MIGRATE_ASYNC)
491 congestion_wait(BLK_RW_ASYNC, HZ/10);
493 if (fatal_signal_pending(current))
497 if (cond_resched()) {
498 /* Async terminates prematurely on need_resched() */
499 if (cc->mode == MIGRATE_ASYNC)
503 /* Time to isolate some pages for migration */
504 for (; low_pfn < end_pfn; low_pfn++) {
505 /* give a chance to irqs before checking need_resched() */
506 if (locked && !(low_pfn % SWAP_CLUSTER_MAX)) {
507 if (should_release_lock(&zone->lru_lock)) {
508 spin_unlock_irqrestore(&zone->lru_lock, flags);
514 * migrate_pfn does not necessarily start aligned to a
515 * pageblock. Ensure that pfn_valid is called when moving
516 * into a new MAX_ORDER_NR_PAGES range in case of large
517 * memory holes within the zone
519 if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
520 if (!pfn_valid(low_pfn)) {
521 low_pfn += MAX_ORDER_NR_PAGES - 1;
526 if (!pfn_valid_within(low_pfn))
531 * Get the page and ensure the page is within the same zone.
532 * See the comment in isolate_freepages about overlapping
533 * nodes. It is deliberate that the new zone lock is not taken
534 * as memory compaction should not move pages between nodes.
536 page = pfn_to_page(low_pfn);
537 if (page_zone(page) != zone)
543 /* If isolation recently failed, do not retry */
544 pageblock_nr = low_pfn >> pageblock_order;
545 if (last_pageblock_nr != pageblock_nr) {
548 last_pageblock_nr = pageblock_nr;
549 if (!isolation_suitable(cc, page))
553 * For async migration, also only scan in MOVABLE
554 * blocks. Async migration is optimistic to see if
555 * the minimum amount of work satisfies the allocation
557 mt = get_pageblock_migratetype(page);
558 if (cc->mode == MIGRATE_ASYNC &&
559 !migrate_async_suitable(mt)) {
560 set_unsuitable = false;
566 * Skip if free. page_order cannot be used without zone->lock
567 * as nothing prevents parallel allocations or buddy merging.
573 * Check may be lockless but that's ok as we recheck later.
574 * It's possible to migrate LRU pages and balloon pages
575 * Skip any other type of page
577 if (!PageLRU(page)) {
578 if (unlikely(balloon_page_movable(page))) {
579 if (locked && balloon_page_isolate(page)) {
580 /* Successfully isolated */
581 goto isolate_success;
588 * PageLRU is set. lru_lock normally excludes isolation
589 * splitting and collapsing (collapsing has already happened
590 * if PageLRU is set) but the lock is not necessarily taken
591 * here and it is wasteful to take it just to check transhuge.
592 * Check TransHuge without lock and skip the whole pageblock if
593 * it's either a transhuge or hugetlbfs page, as calling
594 * compound_order() without preventing THP from splitting the
595 * page underneath us may return surprising results.
597 if (PageTransHuge(page)) {
600 low_pfn += (1 << compound_order(page)) - 1;
605 * Migration will fail if an anonymous page is pinned in memory,
606 * so avoid taking lru_lock and isolating it unnecessarily in an
607 * admittedly racy check.
609 if (!page_mapping(page) &&
610 page_count(page) > page_mapcount(page))
613 /* Check if it is ok to still hold the lock */
614 locked = compact_checklock_irqsave(&zone->lru_lock, &flags,
616 if (!locked || fatal_signal_pending(current))
619 /* Recheck PageLRU and PageTransHuge under lock */
622 if (PageTransHuge(page)) {
623 low_pfn += (1 << compound_order(page)) - 1;
627 lruvec = mem_cgroup_page_lruvec(page, zone);
629 /* Try isolate the page */
630 if (__isolate_lru_page(page, mode) != 0)
633 VM_BUG_ON_PAGE(PageTransCompound(page), page);
635 /* Successfully isolated */
636 del_page_from_lru_list(page, lruvec, page_lru(page));
639 cc->finished_update_migrate = true;
640 list_add(&page->lru, migratelist);
641 cc->nr_migratepages++;
644 /* Avoid isolating too much */
645 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
653 low_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages) - 1;
656 acct_isolated(zone, locked, cc);
659 spin_unlock_irqrestore(&zone->lru_lock, flags);
662 * Update the pageblock-skip information and cached scanner pfn,
663 * if the whole pageblock was scanned without isolating any page.
665 if (low_pfn == end_pfn)
666 update_pageblock_skip(cc, valid_page, nr_isolated,
667 set_unsuitable, true);
669 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
671 count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
673 count_compact_events(COMPACTISOLATED, nr_isolated);
678 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
679 #ifdef CONFIG_COMPACTION
681 * Based on information in the current compact_control, find blocks
682 * suitable for isolating free pages from and then isolate them.
684 static void isolate_freepages(struct zone *zone,
685 struct compact_control *cc)
688 unsigned long block_start_pfn; /* start of current pageblock */
689 unsigned long block_end_pfn; /* end of current pageblock */
690 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
691 unsigned long next_free_pfn; /* start pfn for scaning at next round */
692 int nr_freepages = cc->nr_freepages;
693 struct list_head *freelist = &cc->freepages;
696 * Initialise the free scanner. The starting point is where we last
697 * successfully isolated from, zone-cached value, or the end of the
698 * zone when isolating for the first time. We need this aligned to
699 * the pageblock boundary, because we do
700 * block_start_pfn -= pageblock_nr_pages in the for loop.
701 * For ending point, take care when isolating in last pageblock of a
702 * a zone which ends in the middle of a pageblock.
703 * The low boundary is the end of the pageblock the migration scanner
706 block_start_pfn = cc->free_pfn & ~(pageblock_nr_pages-1);
707 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
709 low_pfn = ALIGN(cc->migrate_pfn + 1, pageblock_nr_pages);
712 * If no pages are isolated, the block_start_pfn < low_pfn check
718 * Isolate free pages until enough are available to migrate the
719 * pages on cc->migratepages. We stop searching if the migrate
720 * and free page scanners meet or enough free pages are isolated.
722 for (; block_start_pfn >= low_pfn && cc->nr_migratepages > nr_freepages;
723 block_end_pfn = block_start_pfn,
724 block_start_pfn -= pageblock_nr_pages) {
725 unsigned long isolated;
728 * This can iterate a massively long zone without finding any
729 * suitable migration targets, so periodically check if we need
734 if (!pfn_valid(block_start_pfn))
738 * Check for overlapping nodes/zones. It's possible on some
739 * configurations to have a setup like
741 * i.e. it's possible that all pages within a zones range of
742 * pages do not belong to a single zone.
744 page = pfn_to_page(block_start_pfn);
745 if (page_zone(page) != zone)
748 /* Check the block is suitable for migration */
749 if (!suitable_migration_target(page))
752 /* If isolation recently failed, do not retry */
753 if (!isolation_suitable(cc, page))
756 /* Found a block suitable for isolating free pages from */
757 isolated = isolate_freepages_block(cc, block_start_pfn,
758 block_end_pfn, freelist, false);
759 nr_freepages += isolated;
762 * Record the highest PFN we isolated pages from. When next
763 * looking for free pages, the search will restart here as
764 * page migration may have returned some pages to the allocator
766 if (isolated && next_free_pfn == 0) {
767 cc->finished_update_free = true;
768 next_free_pfn = block_start_pfn;
772 /* split_free_page does not map the pages */
776 * If we crossed the migrate scanner, we want to keep it that way
777 * so that compact_finished() may detect this
779 if (block_start_pfn < low_pfn)
780 next_free_pfn = cc->migrate_pfn;
782 cc->free_pfn = next_free_pfn;
783 cc->nr_freepages = nr_freepages;
787 * This is a migrate-callback that "allocates" freepages by taking pages
788 * from the isolated freelists in the block we are migrating to.
790 static struct page *compaction_alloc(struct page *migratepage,
794 struct compact_control *cc = (struct compact_control *)data;
795 struct page *freepage;
797 /* Isolate free pages if necessary */
798 if (list_empty(&cc->freepages)) {
799 isolate_freepages(cc->zone, cc);
801 if (list_empty(&cc->freepages))
805 freepage = list_entry(cc->freepages.next, struct page, lru);
806 list_del(&freepage->lru);
813 * This is a migrate-callback that "frees" freepages back to the isolated
814 * freelist. All pages on the freelist are from the same zone, so there is no
815 * special handling needed for NUMA.
817 static void compaction_free(struct page *page, unsigned long data)
819 struct compact_control *cc = (struct compact_control *)data;
821 list_add(&page->lru, &cc->freepages);
826 * We cannot control nr_migratepages fully when migration is running as
827 * migrate_pages() has no knowledge of of compact_control. When migration is
828 * complete, we count the number of pages on the list by hand.
830 static void update_nr_listpages(struct compact_control *cc)
832 int nr_migratepages = 0;
835 list_for_each_entry(page, &cc->migratepages, lru)
838 cc->nr_migratepages = nr_migratepages;
841 /* possible outcome of isolate_migratepages */
843 ISOLATE_ABORT, /* Abort compaction now */
844 ISOLATE_NONE, /* No pages isolated, continue scanning */
845 ISOLATE_SUCCESS, /* Pages isolated, migrate */
849 * Isolate all pages that can be migrated from the block pointed to by
850 * the migrate scanner within compact_control.
852 static isolate_migrate_t isolate_migratepages(struct zone *zone,
853 struct compact_control *cc)
855 unsigned long low_pfn, end_pfn;
857 /* Do not scan outside zone boundaries */
858 low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
860 /* Only scan within a pageblock boundary */
861 end_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages);
863 /* Do not cross the free scanner or scan within a memory hole */
864 if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
865 cc->migrate_pfn = end_pfn;
869 /* Perform the isolation */
870 low_pfn = isolate_migratepages_range(zone, cc, low_pfn, end_pfn, false);
871 if (!low_pfn || cc->contended)
872 return ISOLATE_ABORT;
874 cc->migrate_pfn = low_pfn;
876 return ISOLATE_SUCCESS;
879 static int compact_finished(struct zone *zone,
880 struct compact_control *cc)
883 unsigned long watermark;
885 if (fatal_signal_pending(current))
886 return COMPACT_PARTIAL;
888 /* Compaction run completes if the migrate and free scanner meet */
889 if (cc->free_pfn <= cc->migrate_pfn) {
890 /* Let the next compaction start anew. */
891 zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
892 zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
893 zone->compact_cached_free_pfn = zone_end_pfn(zone);
896 * Mark that the PG_migrate_skip information should be cleared
897 * by kswapd when it goes to sleep. kswapd does not set the
898 * flag itself as the decision to be clear should be directly
899 * based on an allocation request.
901 if (!current_is_kswapd())
902 zone->compact_blockskip_flush = true;
904 return COMPACT_COMPLETE;
908 * order == -1 is expected when compacting via
909 * /proc/sys/vm/compact_memory
912 return COMPACT_CONTINUE;
914 /* Compaction run is not finished if the watermark is not met */
915 watermark = low_wmark_pages(zone);
916 watermark += (1 << cc->order);
918 if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
919 return COMPACT_CONTINUE;
921 /* Direct compactor: Is a suitable page free? */
922 for (order = cc->order; order < MAX_ORDER; order++) {
923 struct free_area *area = &zone->free_area[order];
925 /* Job done if page is free of the right migratetype */
926 if (!list_empty(&area->free_list[cc->migratetype]))
927 return COMPACT_PARTIAL;
929 /* Job done if allocation would set block type */
930 if (cc->order >= pageblock_order && area->nr_free)
931 return COMPACT_PARTIAL;
934 return COMPACT_CONTINUE;
938 * compaction_suitable: Is this suitable to run compaction on this zone now?
940 * COMPACT_SKIPPED - If there are too few free pages for compaction
941 * COMPACT_PARTIAL - If the allocation would succeed without compaction
942 * COMPACT_CONTINUE - If compaction should run now
944 unsigned long compaction_suitable(struct zone *zone, int order)
947 unsigned long watermark;
950 * order == -1 is expected when compacting via
951 * /proc/sys/vm/compact_memory
954 return COMPACT_CONTINUE;
957 * Watermarks for order-0 must be met for compaction. Note the 2UL.
958 * This is because during migration, copies of pages need to be
959 * allocated and for a short time, the footprint is higher
961 watermark = low_wmark_pages(zone) + (2UL << order);
962 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
963 return COMPACT_SKIPPED;
966 * fragmentation index determines if allocation failures are due to
967 * low memory or external fragmentation
969 * index of -1000 implies allocations might succeed depending on
971 * index towards 0 implies failure is due to lack of memory
972 * index towards 1000 implies failure is due to fragmentation
974 * Only compact if a failure would be due to fragmentation.
976 fragindex = fragmentation_index(zone, order);
977 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
978 return COMPACT_SKIPPED;
980 if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
982 return COMPACT_PARTIAL;
984 return COMPACT_CONTINUE;
987 static int compact_zone(struct zone *zone, struct compact_control *cc)
990 unsigned long start_pfn = zone->zone_start_pfn;
991 unsigned long end_pfn = zone_end_pfn(zone);
992 const bool sync = cc->mode != MIGRATE_ASYNC;
994 ret = compaction_suitable(zone, cc->order);
996 case COMPACT_PARTIAL:
997 case COMPACT_SKIPPED:
998 /* Compaction is likely to fail */
1000 case COMPACT_CONTINUE:
1001 /* Fall through to compaction */
1006 * Clear pageblock skip if there were failures recently and compaction
1007 * is about to be retried after being deferred. kswapd does not do
1008 * this reset as it'll reset the cached information when going to sleep.
1010 if (compaction_restarting(zone, cc->order) && !current_is_kswapd())
1011 __reset_isolation_suitable(zone);
1014 * Setup to move all movable pages to the end of the zone. Used cached
1015 * information on where the scanners should start but check that it
1016 * is initialised by ensuring the values are within zone boundaries.
1018 cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync];
1019 cc->free_pfn = zone->compact_cached_free_pfn;
1020 if (cc->free_pfn < start_pfn || cc->free_pfn > end_pfn) {
1021 cc->free_pfn = end_pfn & ~(pageblock_nr_pages-1);
1022 zone->compact_cached_free_pfn = cc->free_pfn;
1024 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn > end_pfn) {
1025 cc->migrate_pfn = start_pfn;
1026 zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
1027 zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
1030 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn, cc->free_pfn, end_pfn);
1032 migrate_prep_local();
1034 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
1035 unsigned long nr_migrate, nr_remaining;
1038 switch (isolate_migratepages(zone, cc)) {
1040 ret = COMPACT_PARTIAL;
1041 putback_movable_pages(&cc->migratepages);
1042 cc->nr_migratepages = 0;
1046 case ISOLATE_SUCCESS:
1050 nr_migrate = cc->nr_migratepages;
1051 err = migrate_pages(&cc->migratepages, compaction_alloc,
1052 compaction_free, (unsigned long)cc, cc->mode,
1054 update_nr_listpages(cc);
1055 nr_remaining = cc->nr_migratepages;
1057 trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
1060 /* Release isolated pages not migrated */
1062 putback_movable_pages(&cc->migratepages);
1063 cc->nr_migratepages = 0;
1065 * migrate_pages() may return -ENOMEM when scanners meet
1066 * and we want compact_finished() to detect it
1068 if (err == -ENOMEM && cc->free_pfn > cc->migrate_pfn) {
1069 ret = COMPACT_PARTIAL;
1076 /* Release free pages and check accounting */
1077 cc->nr_freepages -= release_freepages(&cc->freepages);
1078 VM_BUG_ON(cc->nr_freepages != 0);
1080 trace_mm_compaction_end(ret);
1085 static unsigned long compact_zone_order(struct zone *zone, int order,
1086 gfp_t gfp_mask, enum migrate_mode mode, bool *contended)
1089 struct compact_control cc = {
1091 .nr_migratepages = 0,
1093 .migratetype = allocflags_to_migratetype(gfp_mask),
1097 INIT_LIST_HEAD(&cc.freepages);
1098 INIT_LIST_HEAD(&cc.migratepages);
1100 ret = compact_zone(zone, &cc);
1102 VM_BUG_ON(!list_empty(&cc.freepages));
1103 VM_BUG_ON(!list_empty(&cc.migratepages));
1105 *contended = cc.contended;
1109 int sysctl_extfrag_threshold = 500;
1112 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1113 * @zonelist: The zonelist used for the current allocation
1114 * @order: The order of the current allocation
1115 * @gfp_mask: The GFP mask of the current allocation
1116 * @nodemask: The allowed nodes to allocate from
1117 * @mode: The migration mode for async, sync light, or sync migration
1118 * @contended: Return value that is true if compaction was aborted due to lock contention
1119 * @page: Optionally capture a free page of the requested order during compaction
1121 * This is the main entry point for direct page compaction.
1123 unsigned long try_to_compact_pages(struct zonelist *zonelist,
1124 int order, gfp_t gfp_mask, nodemask_t *nodemask,
1125 enum migrate_mode mode, bool *contended)
1127 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1128 int may_enter_fs = gfp_mask & __GFP_FS;
1129 int may_perform_io = gfp_mask & __GFP_IO;
1132 int rc = COMPACT_SKIPPED;
1133 int alloc_flags = 0;
1135 /* Check if the GFP flags allow compaction */
1136 if (!order || !may_enter_fs || !may_perform_io)
1139 count_compact_event(COMPACTSTALL);
1142 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
1143 alloc_flags |= ALLOC_CMA;
1145 /* Compact each zone in the list */
1146 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1150 status = compact_zone_order(zone, order, gfp_mask, mode,
1152 rc = max(status, rc);
1154 /* If a normal allocation would succeed, stop compacting */
1155 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0,
1164 /* Compact all zones within a node */
1165 static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
1170 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1172 zone = &pgdat->node_zones[zoneid];
1173 if (!populated_zone(zone))
1176 cc->nr_freepages = 0;
1177 cc->nr_migratepages = 0;
1179 INIT_LIST_HEAD(&cc->freepages);
1180 INIT_LIST_HEAD(&cc->migratepages);
1182 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
1183 compact_zone(zone, cc);
1185 if (cc->order > 0) {
1186 if (zone_watermark_ok(zone, cc->order,
1187 low_wmark_pages(zone), 0, 0))
1188 compaction_defer_reset(zone, cc->order, false);
1191 VM_BUG_ON(!list_empty(&cc->freepages));
1192 VM_BUG_ON(!list_empty(&cc->migratepages));
1196 void compact_pgdat(pg_data_t *pgdat, int order)
1198 struct compact_control cc = {
1200 .mode = MIGRATE_ASYNC,
1206 __compact_pgdat(pgdat, &cc);
1209 static void compact_node(int nid)
1211 struct compact_control cc = {
1213 .mode = MIGRATE_SYNC,
1214 .ignore_skip_hint = true,
1217 __compact_pgdat(NODE_DATA(nid), &cc);
1220 /* Compact all nodes in the system */
1221 static void compact_nodes(void)
1225 /* Flush pending updates to the LRU lists */
1226 lru_add_drain_all();
1228 for_each_online_node(nid)
1232 /* The written value is actually unused, all memory is compacted */
1233 int sysctl_compact_memory;
1235 /* This is the entry point for compacting all nodes via /proc/sys/vm */
1236 int sysctl_compaction_handler(struct ctl_table *table, int write,
1237 void __user *buffer, size_t *length, loff_t *ppos)
1245 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1246 void __user *buffer, size_t *length, loff_t *ppos)
1248 proc_dointvec_minmax(table, write, buffer, length, ppos);
1253 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1254 static ssize_t sysfs_compact_node(struct device *dev,
1255 struct device_attribute *attr,
1256 const char *buf, size_t count)
1260 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1261 /* Flush pending updates to the LRU lists */
1262 lru_add_drain_all();
1269 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1271 int compaction_register_node(struct node *node)
1273 return device_create_file(&node->dev, &dev_attr_compact);
1276 void compaction_unregister_node(struct node *node)
1278 return device_remove_file(&node->dev, &dev_attr_compact);
1280 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1282 #endif /* CONFIG_COMPACTION */