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>
19 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
21 #define CREATE_TRACE_POINTS
22 #include <trace/events/compaction.h>
24 static unsigned long release_freepages(struct list_head *freelist)
26 struct page *page, *next;
27 unsigned long count = 0;
29 list_for_each_entry_safe(page, next, freelist, lru) {
38 static void map_pages(struct list_head *list)
42 list_for_each_entry(page, list, lru) {
43 arch_alloc_page(page, 0);
44 kernel_map_pages(page, 1, 1);
48 static inline bool migrate_async_suitable(int migratetype)
50 return is_migrate_cma(migratetype) || migratetype == MIGRATE_MOVABLE;
53 #ifdef CONFIG_COMPACTION
54 /* Returns true if the pageblock should be scanned for pages to isolate. */
55 static inline bool isolation_suitable(struct compact_control *cc,
58 if (cc->ignore_skip_hint)
61 return !get_pageblock_skip(page);
65 * This function is called to clear all cached information on pageblocks that
66 * should be skipped for page isolation when the migrate and free page scanner
69 static void __reset_isolation_suitable(struct zone *zone)
71 unsigned long start_pfn = zone->zone_start_pfn;
72 unsigned long end_pfn = zone->zone_start_pfn + zone->spanned_pages;
75 zone->compact_cached_migrate_pfn = start_pfn;
76 zone->compact_cached_free_pfn = end_pfn;
77 zone->compact_blockskip_flush = false;
79 /* Walk the zone and mark every pageblock as suitable for isolation */
80 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
88 page = pfn_to_page(pfn);
89 if (zone != page_zone(page))
92 clear_pageblock_skip(page);
96 void reset_isolation_suitable(pg_data_t *pgdat)
100 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
101 struct zone *zone = &pgdat->node_zones[zoneid];
102 if (!populated_zone(zone))
105 /* Only flush if a full compaction finished recently */
106 if (zone->compact_blockskip_flush)
107 __reset_isolation_suitable(zone);
112 * If no pages were isolated then mark this pageblock to be skipped in the
113 * future. The information is later cleared by __reset_isolation_suitable().
115 static void update_pageblock_skip(struct compact_control *cc,
116 struct page *page, unsigned long nr_isolated,
117 bool migrate_scanner)
119 struct zone *zone = cc->zone;
124 unsigned long pfn = page_to_pfn(page);
125 set_pageblock_skip(page);
127 /* Update where compaction should restart */
128 if (migrate_scanner) {
129 if (!cc->finished_update_migrate &&
130 pfn > zone->compact_cached_migrate_pfn)
131 zone->compact_cached_migrate_pfn = pfn;
133 if (!cc->finished_update_free &&
134 pfn < zone->compact_cached_free_pfn)
135 zone->compact_cached_free_pfn = pfn;
140 static inline bool isolation_suitable(struct compact_control *cc,
146 static void update_pageblock_skip(struct compact_control *cc,
147 struct page *page, unsigned long nr_isolated,
148 bool migrate_scanner)
151 #endif /* CONFIG_COMPACTION */
153 static inline bool should_release_lock(spinlock_t *lock)
155 return need_resched() || spin_is_contended(lock);
159 * Compaction requires the taking of some coarse locks that are potentially
160 * very heavily contended. Check if the process needs to be scheduled or
161 * if the lock is contended. For async compaction, back out in the event
162 * if contention is severe. For sync compaction, schedule.
164 * Returns true if the lock is held.
165 * Returns false if the lock is released and compaction should abort
167 static bool compact_checklock_irqsave(spinlock_t *lock, unsigned long *flags,
168 bool locked, struct compact_control *cc)
170 if (should_release_lock(lock)) {
172 spin_unlock_irqrestore(lock, *flags);
176 /* async aborts if taking too long or contended */
178 cc->contended = true;
186 spin_lock_irqsave(lock, *flags);
190 static inline bool compact_trylock_irqsave(spinlock_t *lock,
191 unsigned long *flags, struct compact_control *cc)
193 return compact_checklock_irqsave(lock, flags, false, cc);
196 /* Returns true if the page is within a block suitable for migration to */
197 static bool suitable_migration_target(struct page *page)
199 int migratetype = get_pageblock_migratetype(page);
201 /* Don't interfere with memory hot-remove or the min_free_kbytes blocks */
202 if (migratetype == MIGRATE_ISOLATE || migratetype == MIGRATE_RESERVE)
205 /* If the page is a large free page, then allow migration */
206 if (PageBuddy(page) && page_order(page) >= pageblock_order)
209 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
210 if (migrate_async_suitable(migratetype))
213 /* Otherwise skip the block */
217 static void compact_capture_page(struct compact_control *cc)
220 int mtype, mtype_low, mtype_high;
222 if (!cc->page || *cc->page)
226 * For MIGRATE_MOVABLE allocations we capture a suitable page ASAP
227 * regardless of the migratetype of the freelist is is captured from.
228 * This is fine because the order for a high-order MIGRATE_MOVABLE
229 * allocation is typically at least a pageblock size and overall
230 * fragmentation is not impaired. Other allocation types must
231 * capture pages from their own migratelist because otherwise they
232 * could pollute other pageblocks like MIGRATE_MOVABLE with
233 * difficult to move pages and making fragmentation worse overall.
235 if (cc->migratetype == MIGRATE_MOVABLE) {
237 mtype_high = MIGRATE_PCPTYPES;
239 mtype_low = cc->migratetype;
240 mtype_high = cc->migratetype + 1;
243 /* Speculatively examine the free lists without zone lock */
244 for (mtype = mtype_low; mtype < mtype_high; mtype++) {
246 for (order = cc->order; order < MAX_ORDER; order++) {
248 struct free_area *area;
249 area = &(cc->zone->free_area[order]);
250 if (list_empty(&area->free_list[mtype]))
253 /* Take the lock and attempt capture of the page */
254 if (!compact_trylock_irqsave(&cc->zone->lock, &flags, cc))
256 if (!list_empty(&area->free_list[mtype])) {
257 page = list_entry(area->free_list[mtype].next,
259 if (capture_free_page(page, cc->order, mtype)) {
260 spin_unlock_irqrestore(&cc->zone->lock,
266 spin_unlock_irqrestore(&cc->zone->lock, flags);
272 * Isolate free pages onto a private freelist. Caller must hold zone->lock.
273 * If @strict is true, will abort returning 0 on any invalid PFNs or non-free
274 * pages inside of the pageblock (even though it may still end up isolating
277 static unsigned long isolate_freepages_block(struct compact_control *cc,
278 unsigned long blockpfn,
279 unsigned long end_pfn,
280 struct list_head *freelist,
283 int nr_scanned = 0, total_isolated = 0;
284 struct page *cursor, *valid_page = NULL;
285 unsigned long nr_strict_required = end_pfn - blockpfn;
289 cursor = pfn_to_page(blockpfn);
291 /* Isolate free pages. */
292 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
294 struct page *page = cursor;
297 if (!pfn_valid_within(blockpfn))
301 if (!PageBuddy(page))
305 * The zone lock must be held to isolate freepages.
306 * Unfortunately this is a very coarse lock and can be
307 * heavily contended if there are parallel allocations
308 * or parallel compactions. For async compaction do not
309 * spin on the lock and we acquire the lock as late as
312 locked = compact_checklock_irqsave(&cc->zone->lock, &flags,
317 /* Recheck this is a suitable migration target under lock */
318 if (!strict && !suitable_migration_target(page))
321 /* Recheck this is a buddy page under lock */
322 if (!PageBuddy(page))
325 /* Found a free page, break it into order-0 pages */
326 isolated = split_free_page(page);
327 if (!isolated && strict)
329 total_isolated += isolated;
330 for (i = 0; i < isolated; i++) {
331 list_add(&page->lru, freelist);
335 /* If a page was split, advance to the end of it */
337 blockpfn += isolated - 1;
338 cursor += isolated - 1;
342 trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
345 * If strict isolation is requested by CMA then check that all the
346 * pages requested were isolated. If there were any failures, 0 is
347 * returned and CMA will fail.
349 if (strict && nr_strict_required > total_isolated)
353 spin_unlock_irqrestore(&cc->zone->lock, flags);
355 /* Update the pageblock-skip if the whole pageblock was scanned */
356 if (blockpfn == end_pfn)
357 update_pageblock_skip(cc, valid_page, total_isolated, false);
359 return total_isolated;
363 * isolate_freepages_range() - isolate free pages.
364 * @start_pfn: The first PFN to start isolating.
365 * @end_pfn: The one-past-last PFN.
367 * Non-free pages, invalid PFNs, or zone boundaries within the
368 * [start_pfn, end_pfn) range are considered errors, cause function to
369 * undo its actions and return zero.
371 * Otherwise, function returns one-past-the-last PFN of isolated page
372 * (which may be greater then end_pfn if end fell in a middle of
376 isolate_freepages_range(struct compact_control *cc,
377 unsigned long start_pfn, unsigned long end_pfn)
379 unsigned long isolated, pfn, block_end_pfn;
382 for (pfn = start_pfn; pfn < end_pfn; pfn += isolated) {
383 if (!pfn_valid(pfn) || cc->zone != page_zone(pfn_to_page(pfn)))
387 * On subsequent iterations ALIGN() is actually not needed,
388 * but we keep it that we not to complicate the code.
390 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
391 block_end_pfn = min(block_end_pfn, end_pfn);
393 isolated = isolate_freepages_block(cc, pfn, block_end_pfn,
397 * In strict mode, isolate_freepages_block() returns 0 if
398 * there are any holes in the block (ie. invalid PFNs or
405 * If we managed to isolate pages, it is always (1 << n) *
406 * pageblock_nr_pages for some non-negative n. (Max order
407 * page may span two pageblocks).
411 /* split_free_page does not map the pages */
412 map_pages(&freelist);
415 /* Loop terminated early, cleanup. */
416 release_freepages(&freelist);
420 /* We don't use freelists for anything. */
424 /* Update the number of anon and file isolated pages in the zone */
425 static void acct_isolated(struct zone *zone, bool locked, struct compact_control *cc)
428 unsigned int count[2] = { 0, };
430 list_for_each_entry(page, &cc->migratepages, lru)
431 count[!!page_is_file_cache(page)]++;
433 /* If locked we can use the interrupt unsafe versions */
435 __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
436 __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
438 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
439 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
443 /* Similar to reclaim, but different enough that they don't share logic */
444 static bool too_many_isolated(struct zone *zone)
446 unsigned long active, inactive, isolated;
448 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
449 zone_page_state(zone, NR_INACTIVE_ANON);
450 active = zone_page_state(zone, NR_ACTIVE_FILE) +
451 zone_page_state(zone, NR_ACTIVE_ANON);
452 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
453 zone_page_state(zone, NR_ISOLATED_ANON);
455 return isolated > (inactive + active) / 2;
459 * isolate_migratepages_range() - isolate all migrate-able pages in range.
460 * @zone: Zone pages are in.
461 * @cc: Compaction control structure.
462 * @low_pfn: The first PFN of the range.
463 * @end_pfn: The one-past-the-last PFN of the range.
464 * @unevictable: true if it allows to isolate unevictable pages
466 * Isolate all pages that can be migrated from the range specified by
467 * [low_pfn, end_pfn). Returns zero if there is a fatal signal
468 * pending), otherwise PFN of the first page that was not scanned
469 * (which may be both less, equal to or more then end_pfn).
471 * Assumes that cc->migratepages is empty and cc->nr_migratepages is
474 * Apart from cc->migratepages and cc->nr_migratetypes this function
475 * does not modify any cc's fields, in particular it does not modify
476 * (or read for that matter) cc->migrate_pfn.
479 isolate_migratepages_range(struct zone *zone, struct compact_control *cc,
480 unsigned long low_pfn, unsigned long end_pfn, bool unevictable)
482 unsigned long last_pageblock_nr = 0, pageblock_nr;
483 unsigned long nr_scanned = 0, nr_isolated = 0;
484 struct list_head *migratelist = &cc->migratepages;
485 isolate_mode_t mode = 0;
486 struct lruvec *lruvec;
489 struct page *page = NULL, *valid_page = NULL;
492 * Ensure that there are not too many pages isolated from the LRU
493 * list by either parallel reclaimers or compaction. If there are,
494 * delay for some time until fewer pages are isolated
496 while (unlikely(too_many_isolated(zone))) {
497 /* async migration should just abort */
501 congestion_wait(BLK_RW_ASYNC, HZ/10);
503 if (fatal_signal_pending(current))
507 /* Time to isolate some pages for migration */
509 for (; low_pfn < end_pfn; low_pfn++) {
510 /* give a chance to irqs before checking need_resched() */
511 if (locked && !((low_pfn+1) % SWAP_CLUSTER_MAX)) {
512 if (should_release_lock(&zone->lru_lock)) {
513 spin_unlock_irqrestore(&zone->lru_lock, flags);
519 * migrate_pfn does not necessarily start aligned to a
520 * pageblock. Ensure that pfn_valid is called when moving
521 * into a new MAX_ORDER_NR_PAGES range in case of large
522 * memory holes within the zone
524 if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
525 if (!pfn_valid(low_pfn)) {
526 low_pfn += MAX_ORDER_NR_PAGES - 1;
531 if (!pfn_valid_within(low_pfn))
536 * Get the page and ensure the page is within the same zone.
537 * See the comment in isolate_freepages about overlapping
538 * nodes. It is deliberate that the new zone lock is not taken
539 * as memory compaction should not move pages between nodes.
541 page = pfn_to_page(low_pfn);
542 if (page_zone(page) != zone)
548 /* If isolation recently failed, do not retry */
549 pageblock_nr = low_pfn >> pageblock_order;
550 if (!isolation_suitable(cc, page))
558 * For async migration, also only scan in MOVABLE blocks. Async
559 * migration is optimistic to see if the minimum amount of work
560 * satisfies the allocation
562 if (!cc->sync && last_pageblock_nr != pageblock_nr &&
563 !migrate_async_suitable(get_pageblock_migratetype(page))) {
564 cc->finished_update_migrate = true;
568 /* Check may be lockless but that's ok as we recheck later */
573 * PageLRU is set. lru_lock normally excludes isolation
574 * splitting and collapsing (collapsing has already happened
575 * if PageLRU is set) but the lock is not necessarily taken
576 * here and it is wasteful to take it just to check transhuge.
577 * Check TransHuge without lock and skip the whole pageblock if
578 * it's either a transhuge or hugetlbfs page, as calling
579 * compound_order() without preventing THP from splitting the
580 * page underneath us may return surprising results.
582 if (PageTransHuge(page)) {
585 low_pfn += (1 << compound_order(page)) - 1;
589 /* Check if it is ok to still hold the lock */
590 locked = compact_checklock_irqsave(&zone->lru_lock, &flags,
592 if (!locked || fatal_signal_pending(current))
595 /* Recheck PageLRU and PageTransHuge under lock */
598 if (PageTransHuge(page)) {
599 low_pfn += (1 << compound_order(page)) - 1;
604 mode |= ISOLATE_ASYNC_MIGRATE;
607 mode |= ISOLATE_UNEVICTABLE;
609 lruvec = mem_cgroup_page_lruvec(page, zone);
611 /* Try isolate the page */
612 if (__isolate_lru_page(page, mode) != 0)
615 VM_BUG_ON(PageTransCompound(page));
617 /* Successfully isolated */
618 cc->finished_update_migrate = true;
619 del_page_from_lru_list(page, lruvec, page_lru(page));
620 list_add(&page->lru, migratelist);
621 cc->nr_migratepages++;
624 /* Avoid isolating too much */
625 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
633 low_pfn += pageblock_nr_pages;
634 low_pfn = ALIGN(low_pfn, pageblock_nr_pages) - 1;
635 last_pageblock_nr = pageblock_nr;
638 acct_isolated(zone, locked, cc);
641 spin_unlock_irqrestore(&zone->lru_lock, flags);
643 /* Update the pageblock-skip if the whole pageblock was scanned */
644 if (low_pfn == end_pfn)
645 update_pageblock_skip(cc, valid_page, nr_isolated, true);
647 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
652 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
653 #ifdef CONFIG_COMPACTION
655 * Based on information in the current compact_control, find blocks
656 * suitable for isolating free pages from and then isolate them.
658 static void isolate_freepages(struct zone *zone,
659 struct compact_control *cc)
662 unsigned long high_pfn, low_pfn, pfn, zone_end_pfn, end_pfn;
663 int nr_freepages = cc->nr_freepages;
664 struct list_head *freelist = &cc->freepages;
667 * Initialise the free scanner. The starting point is where we last
668 * scanned from (or the end of the zone if starting). The low point
669 * is the end of the pageblock the migration scanner is using.
672 low_pfn = cc->migrate_pfn + pageblock_nr_pages;
675 * Take care that if the migration scanner is at the end of the zone
676 * that the free scanner does not accidentally move to the next zone
677 * in the next isolation cycle.
679 high_pfn = min(low_pfn, pfn);
681 zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
684 * Isolate free pages until enough are available to migrate the
685 * pages on cc->migratepages. We stop searching if the migrate
686 * and free page scanners meet or enough free pages are isolated.
688 for (; pfn > low_pfn && cc->nr_migratepages > nr_freepages;
689 pfn -= pageblock_nr_pages) {
690 unsigned long isolated;
696 * Check for overlapping nodes/zones. It's possible on some
697 * configurations to have a setup like
699 * i.e. it's possible that all pages within a zones range of
700 * pages do not belong to a single zone.
702 page = pfn_to_page(pfn);
703 if (page_zone(page) != zone)
706 /* Check the block is suitable for migration */
707 if (!suitable_migration_target(page))
710 /* If isolation recently failed, do not retry */
711 if (!isolation_suitable(cc, page))
714 /* Found a block suitable for isolating free pages from */
716 end_pfn = min(pfn + pageblock_nr_pages, zone_end_pfn);
717 isolated = isolate_freepages_block(cc, pfn, end_pfn,
719 nr_freepages += isolated;
722 * Record the highest PFN we isolated pages from. When next
723 * looking for free pages, the search will restart here as
724 * page migration may have returned some pages to the allocator
727 cc->finished_update_free = true;
728 high_pfn = max(high_pfn, pfn);
732 /* split_free_page does not map the pages */
735 cc->free_pfn = high_pfn;
736 cc->nr_freepages = nr_freepages;
740 * This is a migrate-callback that "allocates" freepages by taking pages
741 * from the isolated freelists in the block we are migrating to.
743 static struct page *compaction_alloc(struct page *migratepage,
747 struct compact_control *cc = (struct compact_control *)data;
748 struct page *freepage;
750 /* Isolate free pages if necessary */
751 if (list_empty(&cc->freepages)) {
752 isolate_freepages(cc->zone, cc);
754 if (list_empty(&cc->freepages))
758 freepage = list_entry(cc->freepages.next, struct page, lru);
759 list_del(&freepage->lru);
766 * We cannot control nr_migratepages and nr_freepages fully when migration is
767 * running as migrate_pages() has no knowledge of compact_control. When
768 * migration is complete, we count the number of pages on the lists by hand.
770 static void update_nr_listpages(struct compact_control *cc)
772 int nr_migratepages = 0;
773 int nr_freepages = 0;
776 list_for_each_entry(page, &cc->migratepages, lru)
778 list_for_each_entry(page, &cc->freepages, lru)
781 cc->nr_migratepages = nr_migratepages;
782 cc->nr_freepages = nr_freepages;
785 /* possible outcome of isolate_migratepages */
787 ISOLATE_ABORT, /* Abort compaction now */
788 ISOLATE_NONE, /* No pages isolated, continue scanning */
789 ISOLATE_SUCCESS, /* Pages isolated, migrate */
793 * Isolate all pages that can be migrated from the block pointed to by
794 * the migrate scanner within compact_control.
796 static isolate_migrate_t isolate_migratepages(struct zone *zone,
797 struct compact_control *cc)
799 unsigned long low_pfn, end_pfn;
801 /* Do not scan outside zone boundaries */
802 low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
804 /* Only scan within a pageblock boundary */
805 end_pfn = ALIGN(low_pfn + pageblock_nr_pages, pageblock_nr_pages);
807 /* Do not cross the free scanner or scan within a memory hole */
808 if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
809 cc->migrate_pfn = end_pfn;
813 /* Perform the isolation */
814 low_pfn = isolate_migratepages_range(zone, cc, low_pfn, end_pfn, false);
815 if (!low_pfn || cc->contended)
816 return ISOLATE_ABORT;
818 cc->migrate_pfn = low_pfn;
820 return ISOLATE_SUCCESS;
823 static int compact_finished(struct zone *zone,
824 struct compact_control *cc)
826 unsigned long watermark;
828 if (fatal_signal_pending(current))
829 return COMPACT_PARTIAL;
831 /* Compaction run completes if the migrate and free scanner meet */
832 if (cc->free_pfn <= cc->migrate_pfn) {
834 * Mark that the PG_migrate_skip information should be cleared
835 * by kswapd when it goes to sleep. kswapd does not set the
836 * flag itself as the decision to be clear should be directly
837 * based on an allocation request.
839 if (!current_is_kswapd())
840 zone->compact_blockskip_flush = true;
842 return COMPACT_COMPLETE;
846 * order == -1 is expected when compacting via
847 * /proc/sys/vm/compact_memory
850 return COMPACT_CONTINUE;
852 /* Compaction run is not finished if the watermark is not met */
853 watermark = low_wmark_pages(zone);
854 watermark += (1 << cc->order);
856 if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
857 return COMPACT_CONTINUE;
859 /* Direct compactor: Is a suitable page free? */
861 /* Was a suitable page captured? */
863 return COMPACT_PARTIAL;
866 for (order = cc->order; order < MAX_ORDER; order++) {
867 struct free_area *area = &zone->free_area[cc->order];
868 /* Job done if page is free of the right migratetype */
869 if (!list_empty(&area->free_list[cc->migratetype]))
870 return COMPACT_PARTIAL;
872 /* Job done if allocation would set block type */
873 if (cc->order >= pageblock_order && area->nr_free)
874 return COMPACT_PARTIAL;
878 return COMPACT_CONTINUE;
882 * compaction_suitable: Is this suitable to run compaction on this zone now?
884 * COMPACT_SKIPPED - If there are too few free pages for compaction
885 * COMPACT_PARTIAL - If the allocation would succeed without compaction
886 * COMPACT_CONTINUE - If compaction should run now
888 unsigned long compaction_suitable(struct zone *zone, int order)
891 unsigned long watermark;
894 * order == -1 is expected when compacting via
895 * /proc/sys/vm/compact_memory
898 return COMPACT_CONTINUE;
901 * Watermarks for order-0 must be met for compaction. Note the 2UL.
902 * This is because during migration, copies of pages need to be
903 * allocated and for a short time, the footprint is higher
905 watermark = low_wmark_pages(zone) + (2UL << order);
906 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
907 return COMPACT_SKIPPED;
910 * fragmentation index determines if allocation failures are due to
911 * low memory or external fragmentation
913 * index of -1000 implies allocations might succeed depending on
915 * index towards 0 implies failure is due to lack of memory
916 * index towards 1000 implies failure is due to fragmentation
918 * Only compact if a failure would be due to fragmentation.
920 fragindex = fragmentation_index(zone, order);
921 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
922 return COMPACT_SKIPPED;
924 if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
926 return COMPACT_PARTIAL;
928 return COMPACT_CONTINUE;
931 static int compact_zone(struct zone *zone, struct compact_control *cc)
934 unsigned long start_pfn = zone->zone_start_pfn;
935 unsigned long end_pfn = zone->zone_start_pfn + zone->spanned_pages;
937 ret = compaction_suitable(zone, cc->order);
939 case COMPACT_PARTIAL:
940 case COMPACT_SKIPPED:
941 /* Compaction is likely to fail */
943 case COMPACT_CONTINUE:
944 /* Fall through to compaction */
949 * Setup to move all movable pages to the end of the zone. Used cached
950 * information on where the scanners should start but check that it
951 * is initialised by ensuring the values are within zone boundaries.
953 cc->migrate_pfn = zone->compact_cached_migrate_pfn;
954 cc->free_pfn = zone->compact_cached_free_pfn;
955 if (cc->free_pfn < start_pfn || cc->free_pfn > end_pfn) {
956 cc->free_pfn = end_pfn & ~(pageblock_nr_pages-1);
957 zone->compact_cached_free_pfn = cc->free_pfn;
959 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn > end_pfn) {
960 cc->migrate_pfn = start_pfn;
961 zone->compact_cached_migrate_pfn = cc->migrate_pfn;
965 * Clear pageblock skip if there were failures recently and compaction
966 * is about to be retried after being deferred. kswapd does not do
967 * this reset as it'll reset the cached information when going to sleep.
969 if (compaction_restarting(zone, cc->order) && !current_is_kswapd())
970 __reset_isolation_suitable(zone);
972 migrate_prep_local();
974 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
975 unsigned long nr_migrate, nr_remaining;
978 switch (isolate_migratepages(zone, cc)) {
980 ret = COMPACT_PARTIAL;
981 putback_lru_pages(&cc->migratepages);
982 cc->nr_migratepages = 0;
986 case ISOLATE_SUCCESS:
990 nr_migrate = cc->nr_migratepages;
991 err = migrate_pages(&cc->migratepages, compaction_alloc,
992 (unsigned long)cc, false,
993 cc->sync ? MIGRATE_SYNC_LIGHT : MIGRATE_ASYNC,
995 update_nr_listpages(cc);
996 nr_remaining = cc->nr_migratepages;
998 trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
1001 /* Release LRU pages not migrated */
1003 putback_lru_pages(&cc->migratepages);
1004 cc->nr_migratepages = 0;
1005 if (err == -ENOMEM) {
1006 ret = COMPACT_PARTIAL;
1011 /* Capture a page now if it is a suitable size */
1012 compact_capture_page(cc);
1016 /* Release free pages and check accounting */
1017 cc->nr_freepages -= release_freepages(&cc->freepages);
1018 VM_BUG_ON(cc->nr_freepages != 0);
1023 static unsigned long compact_zone_order(struct zone *zone,
1024 int order, gfp_t gfp_mask,
1025 bool sync, bool *contended,
1029 struct compact_control cc = {
1031 .nr_migratepages = 0,
1033 .migratetype = allocflags_to_migratetype(gfp_mask),
1038 INIT_LIST_HEAD(&cc.freepages);
1039 INIT_LIST_HEAD(&cc.migratepages);
1041 ret = compact_zone(zone, &cc);
1043 VM_BUG_ON(!list_empty(&cc.freepages));
1044 VM_BUG_ON(!list_empty(&cc.migratepages));
1046 *contended = cc.contended;
1050 int sysctl_extfrag_threshold = 500;
1053 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1054 * @zonelist: The zonelist used for the current allocation
1055 * @order: The order of the current allocation
1056 * @gfp_mask: The GFP mask of the current allocation
1057 * @nodemask: The allowed nodes to allocate from
1058 * @sync: Whether migration is synchronous or not
1059 * @contended: Return value that is true if compaction was aborted due to lock contention
1060 * @page: Optionally capture a free page of the requested order during compaction
1062 * This is the main entry point for direct page compaction.
1064 unsigned long try_to_compact_pages(struct zonelist *zonelist,
1065 int order, gfp_t gfp_mask, nodemask_t *nodemask,
1066 bool sync, bool *contended, struct page **page)
1068 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1069 int may_enter_fs = gfp_mask & __GFP_FS;
1070 int may_perform_io = gfp_mask & __GFP_IO;
1073 int rc = COMPACT_SKIPPED;
1074 int alloc_flags = 0;
1076 /* Check if the GFP flags allow compaction */
1077 if (!order || !may_enter_fs || !may_perform_io)
1080 count_vm_event(COMPACTSTALL);
1083 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
1084 alloc_flags |= ALLOC_CMA;
1086 /* Compact each zone in the list */
1087 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1091 status = compact_zone_order(zone, order, gfp_mask, sync,
1093 rc = max(status, rc);
1095 /* If a normal allocation would succeed, stop compacting */
1096 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0,
1105 /* Compact all zones within a node */
1106 static int __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
1111 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1113 zone = &pgdat->node_zones[zoneid];
1114 if (!populated_zone(zone))
1117 cc->nr_freepages = 0;
1118 cc->nr_migratepages = 0;
1120 INIT_LIST_HEAD(&cc->freepages);
1121 INIT_LIST_HEAD(&cc->migratepages);
1123 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
1124 compact_zone(zone, cc);
1126 if (cc->order > 0) {
1127 int ok = zone_watermark_ok(zone, cc->order,
1128 low_wmark_pages(zone), 0, 0);
1129 if (ok && cc->order >= zone->compact_order_failed)
1130 zone->compact_order_failed = cc->order + 1;
1131 /* Currently async compaction is never deferred. */
1132 else if (!ok && cc->sync)
1133 defer_compaction(zone, cc->order);
1136 VM_BUG_ON(!list_empty(&cc->freepages));
1137 VM_BUG_ON(!list_empty(&cc->migratepages));
1143 int compact_pgdat(pg_data_t *pgdat, int order)
1145 struct compact_control cc = {
1151 return __compact_pgdat(pgdat, &cc);
1154 static int compact_node(int nid)
1156 struct compact_control cc = {
1162 return __compact_pgdat(NODE_DATA(nid), &cc);
1165 /* Compact all nodes in the system */
1166 static int compact_nodes(void)
1170 /* Flush pending updates to the LRU lists */
1171 lru_add_drain_all();
1173 for_each_online_node(nid)
1176 return COMPACT_COMPLETE;
1179 /* The written value is actually unused, all memory is compacted */
1180 int sysctl_compact_memory;
1182 /* This is the entry point for compacting all nodes via /proc/sys/vm */
1183 int sysctl_compaction_handler(struct ctl_table *table, int write,
1184 void __user *buffer, size_t *length, loff_t *ppos)
1187 return compact_nodes();
1192 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1193 void __user *buffer, size_t *length, loff_t *ppos)
1195 proc_dointvec_minmax(table, write, buffer, length, ppos);
1200 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1201 ssize_t sysfs_compact_node(struct device *dev,
1202 struct device_attribute *attr,
1203 const char *buf, size_t count)
1207 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1208 /* Flush pending updates to the LRU lists */
1209 lru_add_drain_all();
1216 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1218 int compaction_register_node(struct node *node)
1220 return device_create_file(&node->dev, &dev_attr_compact);
1223 void compaction_unregister_node(struct node *node)
1225 return device_remove_file(&node->dev, &dev_attr_compact);
1227 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1229 #endif /* CONFIG_COMPACTION */