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>
20 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
22 #define CREATE_TRACE_POINTS
23 #include <trace/events/compaction.h>
25 static unsigned long release_freepages(struct list_head *freelist)
27 struct page *page, *next;
28 unsigned long count = 0;
30 list_for_each_entry_safe(page, next, freelist, lru) {
39 static void map_pages(struct list_head *list)
43 list_for_each_entry(page, list, lru) {
44 arch_alloc_page(page, 0);
45 kernel_map_pages(page, 1, 1);
49 static inline bool migrate_async_suitable(int migratetype)
51 return is_migrate_cma(migratetype) || migratetype == MIGRATE_MOVABLE;
54 #ifdef CONFIG_COMPACTION
55 /* Returns true if the pageblock should be scanned for pages to isolate. */
56 static inline bool isolation_suitable(struct compact_control *cc,
59 if (cc->ignore_skip_hint)
62 return !get_pageblock_skip(page);
66 * This function is called to clear all cached information on pageblocks that
67 * should be skipped for page isolation when the migrate and free page scanner
70 static void __reset_isolation_suitable(struct zone *zone)
72 unsigned long start_pfn = zone->zone_start_pfn;
73 unsigned long end_pfn = zone->zone_start_pfn + zone->spanned_pages;
76 zone->compact_cached_migrate_pfn = start_pfn;
77 zone->compact_cached_free_pfn = end_pfn;
78 zone->compact_blockskip_flush = false;
80 /* Walk the zone and mark every pageblock as suitable for isolation */
81 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
89 page = pfn_to_page(pfn);
90 if (zone != page_zone(page))
93 clear_pageblock_skip(page);
97 void reset_isolation_suitable(pg_data_t *pgdat)
101 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
102 struct zone *zone = &pgdat->node_zones[zoneid];
103 if (!populated_zone(zone))
106 /* Only flush if a full compaction finished recently */
107 if (zone->compact_blockskip_flush)
108 __reset_isolation_suitable(zone);
113 * If no pages were isolated then mark this pageblock to be skipped in the
114 * future. The information is later cleared by __reset_isolation_suitable().
116 static void update_pageblock_skip(struct compact_control *cc,
117 struct page *page, unsigned long nr_isolated,
118 bool migrate_scanner)
120 struct zone *zone = cc->zone;
125 unsigned long pfn = page_to_pfn(page);
126 set_pageblock_skip(page);
128 /* Update where compaction should restart */
129 if (migrate_scanner) {
130 if (!cc->finished_update_migrate &&
131 pfn > zone->compact_cached_migrate_pfn)
132 zone->compact_cached_migrate_pfn = pfn;
134 if (!cc->finished_update_free &&
135 pfn < zone->compact_cached_free_pfn)
136 zone->compact_cached_free_pfn = pfn;
141 static inline bool isolation_suitable(struct compact_control *cc,
147 static void update_pageblock_skip(struct compact_control *cc,
148 struct page *page, unsigned long nr_isolated,
149 bool migrate_scanner)
152 #endif /* CONFIG_COMPACTION */
154 static inline bool should_release_lock(spinlock_t *lock)
156 return need_resched() || spin_is_contended(lock);
160 * Compaction requires the taking of some coarse locks that are potentially
161 * very heavily contended. Check if the process needs to be scheduled or
162 * if the lock is contended. For async compaction, back out in the event
163 * if contention is severe. For sync compaction, schedule.
165 * Returns true if the lock is held.
166 * Returns false if the lock is released and compaction should abort
168 static bool compact_checklock_irqsave(spinlock_t *lock, unsigned long *flags,
169 bool locked, struct compact_control *cc)
171 if (should_release_lock(lock)) {
173 spin_unlock_irqrestore(lock, *flags);
177 /* async aborts if taking too long or contended */
179 cc->contended = true;
187 spin_lock_irqsave(lock, *flags);
191 static inline bool compact_trylock_irqsave(spinlock_t *lock,
192 unsigned long *flags, struct compact_control *cc)
194 return compact_checklock_irqsave(lock, flags, false, cc);
197 /* Returns true if the page is within a block suitable for migration to */
198 static bool suitable_migration_target(struct page *page)
200 int migratetype = get_pageblock_migratetype(page);
202 /* Don't interfere with memory hot-remove or the min_free_kbytes blocks */
203 if (migratetype == MIGRATE_ISOLATE || migratetype == MIGRATE_RESERVE)
206 /* If the page is a large free page, then allow migration */
207 if (PageBuddy(page) && page_order(page) >= pageblock_order)
210 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
211 if (migrate_async_suitable(migratetype))
214 /* Otherwise skip the block */
219 * Isolate free pages onto a private freelist. Caller must hold zone->lock.
220 * If @strict is true, will abort returning 0 on any invalid PFNs or non-free
221 * pages inside of the pageblock (even though it may still end up isolating
224 static unsigned long isolate_freepages_block(struct compact_control *cc,
225 unsigned long blockpfn,
226 unsigned long end_pfn,
227 struct list_head *freelist,
230 int nr_scanned = 0, total_isolated = 0;
231 struct page *cursor, *valid_page = NULL;
232 unsigned long nr_strict_required = end_pfn - blockpfn;
236 cursor = pfn_to_page(blockpfn);
238 /* Isolate free pages. */
239 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
241 struct page *page = cursor;
244 if (!pfn_valid_within(blockpfn))
248 if (!PageBuddy(page))
252 * The zone lock must be held to isolate freepages.
253 * Unfortunately this is a very coarse lock and can be
254 * heavily contended if there are parallel allocations
255 * or parallel compactions. For async compaction do not
256 * spin on the lock and we acquire the lock as late as
259 locked = compact_checklock_irqsave(&cc->zone->lock, &flags,
264 /* Recheck this is a suitable migration target under lock */
265 if (!strict && !suitable_migration_target(page))
268 /* Recheck this is a buddy page under lock */
269 if (!PageBuddy(page))
272 /* Found a free page, break it into order-0 pages */
273 isolated = split_free_page(page);
274 if (!isolated && strict)
276 total_isolated += isolated;
277 for (i = 0; i < isolated; i++) {
278 list_add(&page->lru, freelist);
282 /* If a page was split, advance to the end of it */
284 blockpfn += isolated - 1;
285 cursor += isolated - 1;
289 trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
292 * If strict isolation is requested by CMA then check that all the
293 * pages requested were isolated. If there were any failures, 0 is
294 * returned and CMA will fail.
296 if (strict && nr_strict_required > total_isolated)
300 spin_unlock_irqrestore(&cc->zone->lock, flags);
302 /* Update the pageblock-skip if the whole pageblock was scanned */
303 if (blockpfn == end_pfn)
304 update_pageblock_skip(cc, valid_page, total_isolated, false);
306 count_vm_events(COMPACTFREE_SCANNED, nr_scanned);
308 count_vm_events(COMPACTISOLATED, total_isolated);
310 return total_isolated;
314 * isolate_freepages_range() - isolate free pages.
315 * @start_pfn: The first PFN to start isolating.
316 * @end_pfn: The one-past-last PFN.
318 * Non-free pages, invalid PFNs, or zone boundaries within the
319 * [start_pfn, end_pfn) range are considered errors, cause function to
320 * undo its actions and return zero.
322 * Otherwise, function returns one-past-the-last PFN of isolated page
323 * (which may be greater then end_pfn if end fell in a middle of
327 isolate_freepages_range(struct compact_control *cc,
328 unsigned long start_pfn, unsigned long end_pfn)
330 unsigned long isolated, pfn, block_end_pfn;
333 for (pfn = start_pfn; pfn < end_pfn; pfn += isolated) {
334 if (!pfn_valid(pfn) || cc->zone != page_zone(pfn_to_page(pfn)))
338 * On subsequent iterations ALIGN() is actually not needed,
339 * but we keep it that we not to complicate the code.
341 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
342 block_end_pfn = min(block_end_pfn, end_pfn);
344 isolated = isolate_freepages_block(cc, pfn, block_end_pfn,
348 * In strict mode, isolate_freepages_block() returns 0 if
349 * there are any holes in the block (ie. invalid PFNs or
356 * If we managed to isolate pages, it is always (1 << n) *
357 * pageblock_nr_pages for some non-negative n. (Max order
358 * page may span two pageblocks).
362 /* split_free_page does not map the pages */
363 map_pages(&freelist);
366 /* Loop terminated early, cleanup. */
367 release_freepages(&freelist);
371 /* We don't use freelists for anything. */
375 /* Update the number of anon and file isolated pages in the zone */
376 static void acct_isolated(struct zone *zone, bool locked, struct compact_control *cc)
379 unsigned int count[2] = { 0, };
381 list_for_each_entry(page, &cc->migratepages, lru)
382 count[!!page_is_file_cache(page)]++;
384 /* If locked we can use the interrupt unsafe versions */
386 __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
387 __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
389 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
390 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
394 /* Similar to reclaim, but different enough that they don't share logic */
395 static bool too_many_isolated(struct zone *zone)
397 unsigned long active, inactive, isolated;
399 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
400 zone_page_state(zone, NR_INACTIVE_ANON);
401 active = zone_page_state(zone, NR_ACTIVE_FILE) +
402 zone_page_state(zone, NR_ACTIVE_ANON);
403 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
404 zone_page_state(zone, NR_ISOLATED_ANON);
406 return isolated > (inactive + active) / 2;
410 * isolate_migratepages_range() - isolate all migrate-able pages in range.
411 * @zone: Zone pages are in.
412 * @cc: Compaction control structure.
413 * @low_pfn: The first PFN of the range.
414 * @end_pfn: The one-past-the-last PFN of the range.
415 * @unevictable: true if it allows to isolate unevictable pages
417 * Isolate all pages that can be migrated from the range specified by
418 * [low_pfn, end_pfn). Returns zero if there is a fatal signal
419 * pending), otherwise PFN of the first page that was not scanned
420 * (which may be both less, equal to or more then end_pfn).
422 * Assumes that cc->migratepages is empty and cc->nr_migratepages is
425 * Apart from cc->migratepages and cc->nr_migratetypes this function
426 * does not modify any cc's fields, in particular it does not modify
427 * (or read for that matter) cc->migrate_pfn.
430 isolate_migratepages_range(struct zone *zone, struct compact_control *cc,
431 unsigned long low_pfn, unsigned long end_pfn, bool unevictable)
433 unsigned long last_pageblock_nr = 0, pageblock_nr;
434 unsigned long nr_scanned = 0, nr_isolated = 0;
435 struct list_head *migratelist = &cc->migratepages;
436 isolate_mode_t mode = 0;
437 struct lruvec *lruvec;
440 struct page *page = NULL, *valid_page = NULL;
443 * Ensure that there are not too many pages isolated from the LRU
444 * list by either parallel reclaimers or compaction. If there are,
445 * delay for some time until fewer pages are isolated
447 while (unlikely(too_many_isolated(zone))) {
448 /* async migration should just abort */
452 congestion_wait(BLK_RW_ASYNC, HZ/10);
454 if (fatal_signal_pending(current))
458 /* Time to isolate some pages for migration */
460 for (; low_pfn < end_pfn; low_pfn++) {
461 /* give a chance to irqs before checking need_resched() */
462 if (locked && !((low_pfn+1) % SWAP_CLUSTER_MAX)) {
463 if (should_release_lock(&zone->lru_lock)) {
464 spin_unlock_irqrestore(&zone->lru_lock, flags);
470 * migrate_pfn does not necessarily start aligned to a
471 * pageblock. Ensure that pfn_valid is called when moving
472 * into a new MAX_ORDER_NR_PAGES range in case of large
473 * memory holes within the zone
475 if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
476 if (!pfn_valid(low_pfn)) {
477 low_pfn += MAX_ORDER_NR_PAGES - 1;
482 if (!pfn_valid_within(low_pfn))
487 * Get the page and ensure the page is within the same zone.
488 * See the comment in isolate_freepages about overlapping
489 * nodes. It is deliberate that the new zone lock is not taken
490 * as memory compaction should not move pages between nodes.
492 page = pfn_to_page(low_pfn);
493 if (page_zone(page) != zone)
499 /* If isolation recently failed, do not retry */
500 pageblock_nr = low_pfn >> pageblock_order;
501 if (!isolation_suitable(cc, page))
509 * For async migration, also only scan in MOVABLE blocks. Async
510 * migration is optimistic to see if the minimum amount of work
511 * satisfies the allocation
513 if (!cc->sync && last_pageblock_nr != pageblock_nr &&
514 !migrate_async_suitable(get_pageblock_migratetype(page))) {
515 cc->finished_update_migrate = true;
520 * Check may be lockless but that's ok as we recheck later.
521 * It's possible to migrate LRU pages and balloon pages
522 * Skip any other type of page
524 if (!PageLRU(page)) {
525 if (unlikely(balloon_page_movable(page))) {
526 if (locked && balloon_page_isolate(page)) {
527 /* Successfully isolated */
528 cc->finished_update_migrate = true;
529 list_add(&page->lru, migratelist);
530 cc->nr_migratepages++;
532 goto check_compact_cluster;
539 * PageLRU is set. lru_lock normally excludes isolation
540 * splitting and collapsing (collapsing has already happened
541 * if PageLRU is set) but the lock is not necessarily taken
542 * here and it is wasteful to take it just to check transhuge.
543 * Check TransHuge without lock and skip the whole pageblock if
544 * it's either a transhuge or hugetlbfs page, as calling
545 * compound_order() without preventing THP from splitting the
546 * page underneath us may return surprising results.
548 if (PageTransHuge(page)) {
551 low_pfn += (1 << compound_order(page)) - 1;
555 /* Check if it is ok to still hold the lock */
556 locked = compact_checklock_irqsave(&zone->lru_lock, &flags,
558 if (!locked || fatal_signal_pending(current))
561 /* Recheck PageLRU and PageTransHuge under lock */
564 if (PageTransHuge(page)) {
565 low_pfn += (1 << compound_order(page)) - 1;
570 mode |= ISOLATE_ASYNC_MIGRATE;
573 mode |= ISOLATE_UNEVICTABLE;
575 lruvec = mem_cgroup_page_lruvec(page, zone);
577 /* Try isolate the page */
578 if (__isolate_lru_page(page, mode) != 0)
581 VM_BUG_ON(PageTransCompound(page));
583 /* Successfully isolated */
584 cc->finished_update_migrate = true;
585 del_page_from_lru_list(page, lruvec, page_lru(page));
586 list_add(&page->lru, migratelist);
587 cc->nr_migratepages++;
590 check_compact_cluster:
591 /* Avoid isolating too much */
592 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
600 low_pfn += pageblock_nr_pages;
601 low_pfn = ALIGN(low_pfn, pageblock_nr_pages) - 1;
602 last_pageblock_nr = pageblock_nr;
605 acct_isolated(zone, locked, cc);
608 spin_unlock_irqrestore(&zone->lru_lock, flags);
610 /* Update the pageblock-skip if the whole pageblock was scanned */
611 if (low_pfn == end_pfn)
612 update_pageblock_skip(cc, valid_page, nr_isolated, true);
614 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
616 count_vm_events(COMPACTMIGRATE_SCANNED, nr_scanned);
618 count_vm_events(COMPACTISOLATED, nr_isolated);
623 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
624 #ifdef CONFIG_COMPACTION
626 * Based on information in the current compact_control, find blocks
627 * suitable for isolating free pages from and then isolate them.
629 static void isolate_freepages(struct zone *zone,
630 struct compact_control *cc)
633 unsigned long high_pfn, low_pfn, pfn, zone_end_pfn, end_pfn;
634 int nr_freepages = cc->nr_freepages;
635 struct list_head *freelist = &cc->freepages;
638 * Initialise the free scanner. The starting point is where we last
639 * scanned from (or the end of the zone if starting). The low point
640 * is the end of the pageblock the migration scanner is using.
643 low_pfn = cc->migrate_pfn + pageblock_nr_pages;
646 * Take care that if the migration scanner is at the end of the zone
647 * that the free scanner does not accidentally move to the next zone
648 * in the next isolation cycle.
650 high_pfn = min(low_pfn, pfn);
652 zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
655 * Isolate free pages until enough are available to migrate the
656 * pages on cc->migratepages. We stop searching if the migrate
657 * and free page scanners meet or enough free pages are isolated.
659 for (; pfn > low_pfn && cc->nr_migratepages > nr_freepages;
660 pfn -= pageblock_nr_pages) {
661 unsigned long isolated;
667 * Check for overlapping nodes/zones. It's possible on some
668 * configurations to have a setup like
670 * i.e. it's possible that all pages within a zones range of
671 * pages do not belong to a single zone.
673 page = pfn_to_page(pfn);
674 if (page_zone(page) != zone)
677 /* Check the block is suitable for migration */
678 if (!suitable_migration_target(page))
681 /* If isolation recently failed, do not retry */
682 if (!isolation_suitable(cc, page))
685 /* Found a block suitable for isolating free pages from */
689 * As pfn may not start aligned, pfn+pageblock_nr_page
690 * may cross a MAX_ORDER_NR_PAGES boundary and miss
691 * a pfn_valid check. Ensure isolate_freepages_block()
692 * only scans within a pageblock
694 end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
695 end_pfn = min(end_pfn, zone_end_pfn);
696 isolated = isolate_freepages_block(cc, pfn, end_pfn,
698 nr_freepages += isolated;
701 * Record the highest PFN we isolated pages from. When next
702 * looking for free pages, the search will restart here as
703 * page migration may have returned some pages to the allocator
706 cc->finished_update_free = true;
707 high_pfn = max(high_pfn, pfn);
711 /* split_free_page does not map the pages */
714 cc->free_pfn = high_pfn;
715 cc->nr_freepages = nr_freepages;
719 * This is a migrate-callback that "allocates" freepages by taking pages
720 * from the isolated freelists in the block we are migrating to.
722 static struct page *compaction_alloc(struct page *migratepage,
726 struct compact_control *cc = (struct compact_control *)data;
727 struct page *freepage;
729 /* Isolate free pages if necessary */
730 if (list_empty(&cc->freepages)) {
731 isolate_freepages(cc->zone, cc);
733 if (list_empty(&cc->freepages))
737 freepage = list_entry(cc->freepages.next, struct page, lru);
738 list_del(&freepage->lru);
745 * We cannot control nr_migratepages and nr_freepages fully when migration is
746 * running as migrate_pages() has no knowledge of compact_control. When
747 * migration is complete, we count the number of pages on the lists by hand.
749 static void update_nr_listpages(struct compact_control *cc)
751 int nr_migratepages = 0;
752 int nr_freepages = 0;
755 list_for_each_entry(page, &cc->migratepages, lru)
757 list_for_each_entry(page, &cc->freepages, lru)
760 cc->nr_migratepages = nr_migratepages;
761 cc->nr_freepages = nr_freepages;
764 /* possible outcome of isolate_migratepages */
766 ISOLATE_ABORT, /* Abort compaction now */
767 ISOLATE_NONE, /* No pages isolated, continue scanning */
768 ISOLATE_SUCCESS, /* Pages isolated, migrate */
772 * Isolate all pages that can be migrated from the block pointed to by
773 * the migrate scanner within compact_control.
775 static isolate_migrate_t isolate_migratepages(struct zone *zone,
776 struct compact_control *cc)
778 unsigned long low_pfn, end_pfn;
780 /* Do not scan outside zone boundaries */
781 low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
783 /* Only scan within a pageblock boundary */
784 end_pfn = ALIGN(low_pfn + pageblock_nr_pages, pageblock_nr_pages);
786 /* Do not cross the free scanner or scan within a memory hole */
787 if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
788 cc->migrate_pfn = end_pfn;
792 /* Perform the isolation */
793 low_pfn = isolate_migratepages_range(zone, cc, low_pfn, end_pfn, false);
794 if (!low_pfn || cc->contended)
795 return ISOLATE_ABORT;
797 cc->migrate_pfn = low_pfn;
799 return ISOLATE_SUCCESS;
802 static int compact_finished(struct zone *zone,
803 struct compact_control *cc)
805 unsigned long watermark;
807 if (fatal_signal_pending(current))
808 return COMPACT_PARTIAL;
810 /* Compaction run completes if the migrate and free scanner meet */
811 if (cc->free_pfn <= cc->migrate_pfn) {
813 * Mark that the PG_migrate_skip information should be cleared
814 * by kswapd when it goes to sleep. kswapd does not set the
815 * flag itself as the decision to be clear should be directly
816 * based on an allocation request.
818 if (!current_is_kswapd())
819 zone->compact_blockskip_flush = true;
821 return COMPACT_COMPLETE;
825 * order == -1 is expected when compacting via
826 * /proc/sys/vm/compact_memory
829 return COMPACT_CONTINUE;
831 /* Compaction run is not finished if the watermark is not met */
832 watermark = low_wmark_pages(zone);
833 watermark += (1 << cc->order);
835 if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
836 return COMPACT_CONTINUE;
838 /* Direct compactor: Is a suitable page free? */
840 /* Was a suitable page captured? */
842 return COMPACT_PARTIAL;
845 for (order = cc->order; order < MAX_ORDER; order++) {
846 struct free_area *area = &zone->free_area[cc->order];
847 /* Job done if page is free of the right migratetype */
848 if (!list_empty(&area->free_list[cc->migratetype]))
849 return COMPACT_PARTIAL;
851 /* Job done if allocation would set block type */
852 if (cc->order >= pageblock_order && area->nr_free)
853 return COMPACT_PARTIAL;
857 return COMPACT_CONTINUE;
861 * compaction_suitable: Is this suitable to run compaction on this zone now?
863 * COMPACT_SKIPPED - If there are too few free pages for compaction
864 * COMPACT_PARTIAL - If the allocation would succeed without compaction
865 * COMPACT_CONTINUE - If compaction should run now
867 unsigned long compaction_suitable(struct zone *zone, int order)
870 unsigned long watermark;
873 * order == -1 is expected when compacting via
874 * /proc/sys/vm/compact_memory
877 return COMPACT_CONTINUE;
880 * Watermarks for order-0 must be met for compaction. Note the 2UL.
881 * This is because during migration, copies of pages need to be
882 * allocated and for a short time, the footprint is higher
884 watermark = low_wmark_pages(zone) + (2UL << order);
885 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
886 return COMPACT_SKIPPED;
889 * fragmentation index determines if allocation failures are due to
890 * low memory or external fragmentation
892 * index of -1000 implies allocations might succeed depending on
894 * index towards 0 implies failure is due to lack of memory
895 * index towards 1000 implies failure is due to fragmentation
897 * Only compact if a failure would be due to fragmentation.
899 fragindex = fragmentation_index(zone, order);
900 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
901 return COMPACT_SKIPPED;
903 if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
905 return COMPACT_PARTIAL;
907 return COMPACT_CONTINUE;
910 static void compact_capture_page(struct compact_control *cc)
913 int mtype, mtype_low, mtype_high;
915 if (!cc->page || *cc->page)
919 * For MIGRATE_MOVABLE allocations we capture a suitable page ASAP
920 * regardless of the migratetype of the freelist is is captured from.
921 * This is fine because the order for a high-order MIGRATE_MOVABLE
922 * allocation is typically at least a pageblock size and overall
923 * fragmentation is not impaired. Other allocation types must
924 * capture pages from their own migratelist because otherwise they
925 * could pollute other pageblocks like MIGRATE_MOVABLE with
926 * difficult to move pages and making fragmentation worse overall.
928 if (cc->migratetype == MIGRATE_MOVABLE) {
930 mtype_high = MIGRATE_PCPTYPES;
932 mtype_low = cc->migratetype;
933 mtype_high = cc->migratetype + 1;
936 /* Speculatively examine the free lists without zone lock */
937 for (mtype = mtype_low; mtype < mtype_high; mtype++) {
939 for (order = cc->order; order < MAX_ORDER; order++) {
941 struct free_area *area;
942 area = &(cc->zone->free_area[order]);
943 if (list_empty(&area->free_list[mtype]))
946 /* Take the lock and attempt capture of the page */
947 if (!compact_trylock_irqsave(&cc->zone->lock, &flags, cc))
949 if (!list_empty(&area->free_list[mtype])) {
950 page = list_entry(area->free_list[mtype].next,
952 if (capture_free_page(page, cc->order, mtype)) {
953 spin_unlock_irqrestore(&cc->zone->lock,
959 spin_unlock_irqrestore(&cc->zone->lock, flags);
964 static int compact_zone(struct zone *zone, struct compact_control *cc)
967 unsigned long start_pfn = zone->zone_start_pfn;
968 unsigned long end_pfn = zone->zone_start_pfn + zone->spanned_pages;
970 ret = compaction_suitable(zone, cc->order);
972 case COMPACT_PARTIAL:
973 case COMPACT_SKIPPED:
974 /* Compaction is likely to fail */
976 case COMPACT_CONTINUE:
977 /* Fall through to compaction */
982 * Setup to move all movable pages to the end of the zone. Used cached
983 * information on where the scanners should start but check that it
984 * is initialised by ensuring the values are within zone boundaries.
986 cc->migrate_pfn = zone->compact_cached_migrate_pfn;
987 cc->free_pfn = zone->compact_cached_free_pfn;
988 if (cc->free_pfn < start_pfn || cc->free_pfn > end_pfn) {
989 cc->free_pfn = end_pfn & ~(pageblock_nr_pages-1);
990 zone->compact_cached_free_pfn = cc->free_pfn;
992 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn > end_pfn) {
993 cc->migrate_pfn = start_pfn;
994 zone->compact_cached_migrate_pfn = cc->migrate_pfn;
998 * Clear pageblock skip if there were failures recently and compaction
999 * is about to be retried after being deferred. kswapd does not do
1000 * this reset as it'll reset the cached information when going to sleep.
1002 if (compaction_restarting(zone, cc->order) && !current_is_kswapd())
1003 __reset_isolation_suitable(zone);
1005 migrate_prep_local();
1007 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
1008 unsigned long nr_migrate, nr_remaining;
1011 switch (isolate_migratepages(zone, cc)) {
1013 ret = COMPACT_PARTIAL;
1014 putback_movable_pages(&cc->migratepages);
1015 cc->nr_migratepages = 0;
1019 case ISOLATE_SUCCESS:
1023 nr_migrate = cc->nr_migratepages;
1024 err = migrate_pages(&cc->migratepages, compaction_alloc,
1025 (unsigned long)cc, false,
1026 cc->sync ? MIGRATE_SYNC_LIGHT : MIGRATE_ASYNC,
1028 update_nr_listpages(cc);
1029 nr_remaining = cc->nr_migratepages;
1031 trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
1034 /* Release isolated pages not migrated */
1036 putback_movable_pages(&cc->migratepages);
1037 cc->nr_migratepages = 0;
1038 if (err == -ENOMEM) {
1039 ret = COMPACT_PARTIAL;
1044 /* Capture a page now if it is a suitable size */
1045 compact_capture_page(cc);
1049 /* Release free pages and check accounting */
1050 cc->nr_freepages -= release_freepages(&cc->freepages);
1051 VM_BUG_ON(cc->nr_freepages != 0);
1056 static unsigned long compact_zone_order(struct zone *zone,
1057 int order, gfp_t gfp_mask,
1058 bool sync, bool *contended,
1062 struct compact_control cc = {
1064 .nr_migratepages = 0,
1066 .migratetype = allocflags_to_migratetype(gfp_mask),
1071 INIT_LIST_HEAD(&cc.freepages);
1072 INIT_LIST_HEAD(&cc.migratepages);
1074 ret = compact_zone(zone, &cc);
1076 VM_BUG_ON(!list_empty(&cc.freepages));
1077 VM_BUG_ON(!list_empty(&cc.migratepages));
1079 *contended = cc.contended;
1083 int sysctl_extfrag_threshold = 500;
1086 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1087 * @zonelist: The zonelist used for the current allocation
1088 * @order: The order of the current allocation
1089 * @gfp_mask: The GFP mask of the current allocation
1090 * @nodemask: The allowed nodes to allocate from
1091 * @sync: Whether migration is synchronous or not
1092 * @contended: Return value that is true if compaction was aborted due to lock contention
1093 * @page: Optionally capture a free page of the requested order during compaction
1095 * This is the main entry point for direct page compaction.
1097 unsigned long try_to_compact_pages(struct zonelist *zonelist,
1098 int order, gfp_t gfp_mask, nodemask_t *nodemask,
1099 bool sync, bool *contended, struct page **page)
1101 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1102 int may_enter_fs = gfp_mask & __GFP_FS;
1103 int may_perform_io = gfp_mask & __GFP_IO;
1106 int rc = COMPACT_SKIPPED;
1107 int alloc_flags = 0;
1109 /* Check if the GFP flags allow compaction */
1110 if (!order || !may_enter_fs || !may_perform_io)
1113 count_vm_event(COMPACTSTALL);
1116 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
1117 alloc_flags |= ALLOC_CMA;
1119 /* Compact each zone in the list */
1120 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1124 status = compact_zone_order(zone, order, gfp_mask, sync,
1126 rc = max(status, rc);
1128 /* If a normal allocation would succeed, stop compacting */
1129 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0,
1138 /* Compact all zones within a node */
1139 static int __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
1144 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1146 zone = &pgdat->node_zones[zoneid];
1147 if (!populated_zone(zone))
1150 cc->nr_freepages = 0;
1151 cc->nr_migratepages = 0;
1153 INIT_LIST_HEAD(&cc->freepages);
1154 INIT_LIST_HEAD(&cc->migratepages);
1156 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
1157 compact_zone(zone, cc);
1159 if (cc->order > 0) {
1160 int ok = zone_watermark_ok(zone, cc->order,
1161 low_wmark_pages(zone), 0, 0);
1162 if (ok && cc->order >= zone->compact_order_failed)
1163 zone->compact_order_failed = cc->order + 1;
1164 /* Currently async compaction is never deferred. */
1165 else if (!ok && cc->sync)
1166 defer_compaction(zone, cc->order);
1169 VM_BUG_ON(!list_empty(&cc->freepages));
1170 VM_BUG_ON(!list_empty(&cc->migratepages));
1176 int compact_pgdat(pg_data_t *pgdat, int order)
1178 struct compact_control cc = {
1184 return __compact_pgdat(pgdat, &cc);
1187 static int compact_node(int nid)
1189 struct compact_control cc = {
1195 return __compact_pgdat(NODE_DATA(nid), &cc);
1198 /* Compact all nodes in the system */
1199 static int compact_nodes(void)
1203 /* Flush pending updates to the LRU lists */
1204 lru_add_drain_all();
1206 for_each_online_node(nid)
1209 return COMPACT_COMPLETE;
1212 /* The written value is actually unused, all memory is compacted */
1213 int sysctl_compact_memory;
1215 /* This is the entry point for compacting all nodes via /proc/sys/vm */
1216 int sysctl_compaction_handler(struct ctl_table *table, int write,
1217 void __user *buffer, size_t *length, loff_t *ppos)
1220 return compact_nodes();
1225 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1226 void __user *buffer, size_t *length, loff_t *ppos)
1228 proc_dointvec_minmax(table, write, buffer, length, ppos);
1233 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1234 ssize_t sysfs_compact_node(struct device *dev,
1235 struct device_attribute *attr,
1236 const char *buf, size_t count)
1240 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1241 /* Flush pending updates to the LRU lists */
1242 lru_add_drain_all();
1249 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1251 int compaction_register_node(struct node *node)
1253 return device_create_file(&node->dev, &dev_attr_compact);
1256 void compaction_unregister_node(struct node *node)
1258 return device_remove_file(&node->dev, &dev_attr_compact);
1260 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1262 #endif /* CONFIG_COMPACTION */