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 return total_isolated;
310 * isolate_freepages_range() - isolate free pages.
311 * @start_pfn: The first PFN to start isolating.
312 * @end_pfn: The one-past-last PFN.
314 * Non-free pages, invalid PFNs, or zone boundaries within the
315 * [start_pfn, end_pfn) range are considered errors, cause function to
316 * undo its actions and return zero.
318 * Otherwise, function returns one-past-the-last PFN of isolated page
319 * (which may be greater then end_pfn if end fell in a middle of
323 isolate_freepages_range(struct compact_control *cc,
324 unsigned long start_pfn, unsigned long end_pfn)
326 unsigned long isolated, pfn, block_end_pfn;
329 for (pfn = start_pfn; pfn < end_pfn; pfn += isolated) {
330 if (!pfn_valid(pfn) || cc->zone != page_zone(pfn_to_page(pfn)))
334 * On subsequent iterations ALIGN() is actually not needed,
335 * but we keep it that we not to complicate the code.
337 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
338 block_end_pfn = min(block_end_pfn, end_pfn);
340 isolated = isolate_freepages_block(cc, pfn, block_end_pfn,
344 * In strict mode, isolate_freepages_block() returns 0 if
345 * there are any holes in the block (ie. invalid PFNs or
352 * If we managed to isolate pages, it is always (1 << n) *
353 * pageblock_nr_pages for some non-negative n. (Max order
354 * page may span two pageblocks).
358 /* split_free_page does not map the pages */
359 map_pages(&freelist);
362 /* Loop terminated early, cleanup. */
363 release_freepages(&freelist);
367 /* We don't use freelists for anything. */
371 /* Update the number of anon and file isolated pages in the zone */
372 static void acct_isolated(struct zone *zone, bool locked, struct compact_control *cc)
375 unsigned int count[2] = { 0, };
377 list_for_each_entry(page, &cc->migratepages, lru)
378 count[!!page_is_file_cache(page)]++;
380 /* If locked we can use the interrupt unsafe versions */
382 __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
383 __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
385 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
386 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
390 /* Similar to reclaim, but different enough that they don't share logic */
391 static bool too_many_isolated(struct zone *zone)
393 unsigned long active, inactive, isolated;
395 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
396 zone_page_state(zone, NR_INACTIVE_ANON);
397 active = zone_page_state(zone, NR_ACTIVE_FILE) +
398 zone_page_state(zone, NR_ACTIVE_ANON);
399 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
400 zone_page_state(zone, NR_ISOLATED_ANON);
402 return isolated > (inactive + active) / 2;
406 * isolate_migratepages_range() - isolate all migrate-able pages in range.
407 * @zone: Zone pages are in.
408 * @cc: Compaction control structure.
409 * @low_pfn: The first PFN of the range.
410 * @end_pfn: The one-past-the-last PFN of the range.
411 * @unevictable: true if it allows to isolate unevictable pages
413 * Isolate all pages that can be migrated from the range specified by
414 * [low_pfn, end_pfn). Returns zero if there is a fatal signal
415 * pending), otherwise PFN of the first page that was not scanned
416 * (which may be both less, equal to or more then end_pfn).
418 * Assumes that cc->migratepages is empty and cc->nr_migratepages is
421 * Apart from cc->migratepages and cc->nr_migratetypes this function
422 * does not modify any cc's fields, in particular it does not modify
423 * (or read for that matter) cc->migrate_pfn.
426 isolate_migratepages_range(struct zone *zone, struct compact_control *cc,
427 unsigned long low_pfn, unsigned long end_pfn, bool unevictable)
429 unsigned long last_pageblock_nr = 0, pageblock_nr;
430 unsigned long nr_scanned = 0, nr_isolated = 0;
431 struct list_head *migratelist = &cc->migratepages;
432 isolate_mode_t mode = 0;
433 struct lruvec *lruvec;
436 struct page *page = NULL, *valid_page = NULL;
439 * Ensure that there are not too many pages isolated from the LRU
440 * list by either parallel reclaimers or compaction. If there are,
441 * delay for some time until fewer pages are isolated
443 while (unlikely(too_many_isolated(zone))) {
444 /* async migration should just abort */
448 congestion_wait(BLK_RW_ASYNC, HZ/10);
450 if (fatal_signal_pending(current))
454 /* Time to isolate some pages for migration */
456 for (; low_pfn < end_pfn; low_pfn++) {
457 /* give a chance to irqs before checking need_resched() */
458 if (locked && !((low_pfn+1) % SWAP_CLUSTER_MAX)) {
459 if (should_release_lock(&zone->lru_lock)) {
460 spin_unlock_irqrestore(&zone->lru_lock, flags);
466 * migrate_pfn does not necessarily start aligned to a
467 * pageblock. Ensure that pfn_valid is called when moving
468 * into a new MAX_ORDER_NR_PAGES range in case of large
469 * memory holes within the zone
471 if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
472 if (!pfn_valid(low_pfn)) {
473 low_pfn += MAX_ORDER_NR_PAGES - 1;
478 if (!pfn_valid_within(low_pfn))
483 * Get the page and ensure the page is within the same zone.
484 * See the comment in isolate_freepages about overlapping
485 * nodes. It is deliberate that the new zone lock is not taken
486 * as memory compaction should not move pages between nodes.
488 page = pfn_to_page(low_pfn);
489 if (page_zone(page) != zone)
495 /* If isolation recently failed, do not retry */
496 pageblock_nr = low_pfn >> pageblock_order;
497 if (!isolation_suitable(cc, page))
505 * For async migration, also only scan in MOVABLE blocks. Async
506 * migration is optimistic to see if the minimum amount of work
507 * satisfies the allocation
509 if (!cc->sync && last_pageblock_nr != pageblock_nr &&
510 !migrate_async_suitable(get_pageblock_migratetype(page))) {
511 cc->finished_update_migrate = true;
516 * Check may be lockless but that's ok as we recheck later.
517 * It's possible to migrate LRU pages and balloon pages
518 * Skip any other type of page
520 if (!PageLRU(page)) {
521 if (unlikely(balloon_page_movable(page))) {
522 if (locked && balloon_page_isolate(page)) {
523 /* Successfully isolated */
524 cc->finished_update_migrate = true;
525 list_add(&page->lru, migratelist);
526 cc->nr_migratepages++;
528 goto check_compact_cluster;
535 * PageLRU is set. lru_lock normally excludes isolation
536 * splitting and collapsing (collapsing has already happened
537 * if PageLRU is set) but the lock is not necessarily taken
538 * here and it is wasteful to take it just to check transhuge.
539 * Check TransHuge without lock and skip the whole pageblock if
540 * it's either a transhuge or hugetlbfs page, as calling
541 * compound_order() without preventing THP from splitting the
542 * page underneath us may return surprising results.
544 if (PageTransHuge(page)) {
547 low_pfn += (1 << compound_order(page)) - 1;
551 /* Check if it is ok to still hold the lock */
552 locked = compact_checklock_irqsave(&zone->lru_lock, &flags,
554 if (!locked || fatal_signal_pending(current))
557 /* Recheck PageLRU and PageTransHuge under lock */
560 if (PageTransHuge(page)) {
561 low_pfn += (1 << compound_order(page)) - 1;
566 mode |= ISOLATE_ASYNC_MIGRATE;
569 mode |= ISOLATE_UNEVICTABLE;
571 lruvec = mem_cgroup_page_lruvec(page, zone);
573 /* Try isolate the page */
574 if (__isolate_lru_page(page, mode) != 0)
577 VM_BUG_ON(PageTransCompound(page));
579 /* Successfully isolated */
580 cc->finished_update_migrate = true;
581 del_page_from_lru_list(page, lruvec, page_lru(page));
582 list_add(&page->lru, migratelist);
583 cc->nr_migratepages++;
586 check_compact_cluster:
587 /* Avoid isolating too much */
588 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
596 low_pfn += pageblock_nr_pages;
597 low_pfn = ALIGN(low_pfn, pageblock_nr_pages) - 1;
598 last_pageblock_nr = pageblock_nr;
601 acct_isolated(zone, locked, cc);
604 spin_unlock_irqrestore(&zone->lru_lock, flags);
606 /* Update the pageblock-skip if the whole pageblock was scanned */
607 if (low_pfn == end_pfn)
608 update_pageblock_skip(cc, valid_page, nr_isolated, true);
610 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
615 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
616 #ifdef CONFIG_COMPACTION
618 * Based on information in the current compact_control, find blocks
619 * suitable for isolating free pages from and then isolate them.
621 static void isolate_freepages(struct zone *zone,
622 struct compact_control *cc)
625 unsigned long high_pfn, low_pfn, pfn, zone_end_pfn, end_pfn;
626 int nr_freepages = cc->nr_freepages;
627 struct list_head *freelist = &cc->freepages;
630 * Initialise the free scanner. The starting point is where we last
631 * scanned from (or the end of the zone if starting). The low point
632 * is the end of the pageblock the migration scanner is using.
635 low_pfn = cc->migrate_pfn + pageblock_nr_pages;
638 * Take care that if the migration scanner is at the end of the zone
639 * that the free scanner does not accidentally move to the next zone
640 * in the next isolation cycle.
642 high_pfn = min(low_pfn, pfn);
644 zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
647 * Isolate free pages until enough are available to migrate the
648 * pages on cc->migratepages. We stop searching if the migrate
649 * and free page scanners meet or enough free pages are isolated.
651 for (; pfn > low_pfn && cc->nr_migratepages > nr_freepages;
652 pfn -= pageblock_nr_pages) {
653 unsigned long isolated;
659 * Check for overlapping nodes/zones. It's possible on some
660 * configurations to have a setup like
662 * i.e. it's possible that all pages within a zones range of
663 * pages do not belong to a single zone.
665 page = pfn_to_page(pfn);
666 if (page_zone(page) != zone)
669 /* Check the block is suitable for migration */
670 if (!suitable_migration_target(page))
673 /* If isolation recently failed, do not retry */
674 if (!isolation_suitable(cc, page))
677 /* Found a block suitable for isolating free pages from */
681 * As pfn may not start aligned, pfn+pageblock_nr_page
682 * may cross a MAX_ORDER_NR_PAGES boundary and miss
683 * a pfn_valid check. Ensure isolate_freepages_block()
684 * only scans within a pageblock
686 end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
687 end_pfn = min(end_pfn, zone_end_pfn);
688 isolated = isolate_freepages_block(cc, pfn, end_pfn,
690 nr_freepages += isolated;
693 * Record the highest PFN we isolated pages from. When next
694 * looking for free pages, the search will restart here as
695 * page migration may have returned some pages to the allocator
698 cc->finished_update_free = true;
699 high_pfn = max(high_pfn, pfn);
703 /* split_free_page does not map the pages */
706 cc->free_pfn = high_pfn;
707 cc->nr_freepages = nr_freepages;
711 * This is a migrate-callback that "allocates" freepages by taking pages
712 * from the isolated freelists in the block we are migrating to.
714 static struct page *compaction_alloc(struct page *migratepage,
718 struct compact_control *cc = (struct compact_control *)data;
719 struct page *freepage;
721 /* Isolate free pages if necessary */
722 if (list_empty(&cc->freepages)) {
723 isolate_freepages(cc->zone, cc);
725 if (list_empty(&cc->freepages))
729 freepage = list_entry(cc->freepages.next, struct page, lru);
730 list_del(&freepage->lru);
737 * We cannot control nr_migratepages and nr_freepages fully when migration is
738 * running as migrate_pages() has no knowledge of compact_control. When
739 * migration is complete, we count the number of pages on the lists by hand.
741 static void update_nr_listpages(struct compact_control *cc)
743 int nr_migratepages = 0;
744 int nr_freepages = 0;
747 list_for_each_entry(page, &cc->migratepages, lru)
749 list_for_each_entry(page, &cc->freepages, lru)
752 cc->nr_migratepages = nr_migratepages;
753 cc->nr_freepages = nr_freepages;
756 /* possible outcome of isolate_migratepages */
758 ISOLATE_ABORT, /* Abort compaction now */
759 ISOLATE_NONE, /* No pages isolated, continue scanning */
760 ISOLATE_SUCCESS, /* Pages isolated, migrate */
764 * Isolate all pages that can be migrated from the block pointed to by
765 * the migrate scanner within compact_control.
767 static isolate_migrate_t isolate_migratepages(struct zone *zone,
768 struct compact_control *cc)
770 unsigned long low_pfn, end_pfn;
772 /* Do not scan outside zone boundaries */
773 low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
775 /* Only scan within a pageblock boundary */
776 end_pfn = ALIGN(low_pfn + pageblock_nr_pages, pageblock_nr_pages);
778 /* Do not cross the free scanner or scan within a memory hole */
779 if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
780 cc->migrate_pfn = end_pfn;
784 /* Perform the isolation */
785 low_pfn = isolate_migratepages_range(zone, cc, low_pfn, end_pfn, false);
786 if (!low_pfn || cc->contended)
787 return ISOLATE_ABORT;
789 cc->migrate_pfn = low_pfn;
791 return ISOLATE_SUCCESS;
794 static int compact_finished(struct zone *zone,
795 struct compact_control *cc)
797 unsigned long watermark;
799 if (fatal_signal_pending(current))
800 return COMPACT_PARTIAL;
802 /* Compaction run completes if the migrate and free scanner meet */
803 if (cc->free_pfn <= cc->migrate_pfn) {
805 * Mark that the PG_migrate_skip information should be cleared
806 * by kswapd when it goes to sleep. kswapd does not set the
807 * flag itself as the decision to be clear should be directly
808 * based on an allocation request.
810 if (!current_is_kswapd())
811 zone->compact_blockskip_flush = true;
813 return COMPACT_COMPLETE;
817 * order == -1 is expected when compacting via
818 * /proc/sys/vm/compact_memory
821 return COMPACT_CONTINUE;
823 /* Compaction run is not finished if the watermark is not met */
824 watermark = low_wmark_pages(zone);
825 watermark += (1 << cc->order);
827 if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
828 return COMPACT_CONTINUE;
830 /* Direct compactor: Is a suitable page free? */
832 /* Was a suitable page captured? */
834 return COMPACT_PARTIAL;
837 for (order = cc->order; order < MAX_ORDER; order++) {
838 struct free_area *area = &zone->free_area[cc->order];
839 /* Job done if page is free of the right migratetype */
840 if (!list_empty(&area->free_list[cc->migratetype]))
841 return COMPACT_PARTIAL;
843 /* Job done if allocation would set block type */
844 if (cc->order >= pageblock_order && area->nr_free)
845 return COMPACT_PARTIAL;
849 return COMPACT_CONTINUE;
853 * compaction_suitable: Is this suitable to run compaction on this zone now?
855 * COMPACT_SKIPPED - If there are too few free pages for compaction
856 * COMPACT_PARTIAL - If the allocation would succeed without compaction
857 * COMPACT_CONTINUE - If compaction should run now
859 unsigned long compaction_suitable(struct zone *zone, int order)
862 unsigned long watermark;
865 * order == -1 is expected when compacting via
866 * /proc/sys/vm/compact_memory
869 return COMPACT_CONTINUE;
872 * Watermarks for order-0 must be met for compaction. Note the 2UL.
873 * This is because during migration, copies of pages need to be
874 * allocated and for a short time, the footprint is higher
876 watermark = low_wmark_pages(zone) + (2UL << order);
877 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
878 return COMPACT_SKIPPED;
881 * fragmentation index determines if allocation failures are due to
882 * low memory or external fragmentation
884 * index of -1000 implies allocations might succeed depending on
886 * index towards 0 implies failure is due to lack of memory
887 * index towards 1000 implies failure is due to fragmentation
889 * Only compact if a failure would be due to fragmentation.
891 fragindex = fragmentation_index(zone, order);
892 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
893 return COMPACT_SKIPPED;
895 if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
897 return COMPACT_PARTIAL;
899 return COMPACT_CONTINUE;
902 static void compact_capture_page(struct compact_control *cc)
905 int mtype, mtype_low, mtype_high;
907 if (!cc->page || *cc->page)
911 * For MIGRATE_MOVABLE allocations we capture a suitable page ASAP
912 * regardless of the migratetype of the freelist is is captured from.
913 * This is fine because the order for a high-order MIGRATE_MOVABLE
914 * allocation is typically at least a pageblock size and overall
915 * fragmentation is not impaired. Other allocation types must
916 * capture pages from their own migratelist because otherwise they
917 * could pollute other pageblocks like MIGRATE_MOVABLE with
918 * difficult to move pages and making fragmentation worse overall.
920 if (cc->migratetype == MIGRATE_MOVABLE) {
922 mtype_high = MIGRATE_PCPTYPES;
924 mtype_low = cc->migratetype;
925 mtype_high = cc->migratetype + 1;
928 /* Speculatively examine the free lists without zone lock */
929 for (mtype = mtype_low; mtype < mtype_high; mtype++) {
931 for (order = cc->order; order < MAX_ORDER; order++) {
933 struct free_area *area;
934 area = &(cc->zone->free_area[order]);
935 if (list_empty(&area->free_list[mtype]))
938 /* Take the lock and attempt capture of the page */
939 if (!compact_trylock_irqsave(&cc->zone->lock, &flags, cc))
941 if (!list_empty(&area->free_list[mtype])) {
942 page = list_entry(area->free_list[mtype].next,
944 if (capture_free_page(page, cc->order, mtype)) {
945 spin_unlock_irqrestore(&cc->zone->lock,
951 spin_unlock_irqrestore(&cc->zone->lock, flags);
956 static int compact_zone(struct zone *zone, struct compact_control *cc)
959 unsigned long start_pfn = zone->zone_start_pfn;
960 unsigned long end_pfn = zone->zone_start_pfn + zone->spanned_pages;
962 ret = compaction_suitable(zone, cc->order);
964 case COMPACT_PARTIAL:
965 case COMPACT_SKIPPED:
966 /* Compaction is likely to fail */
968 case COMPACT_CONTINUE:
969 /* Fall through to compaction */
974 * Setup to move all movable pages to the end of the zone. Used cached
975 * information on where the scanners should start but check that it
976 * is initialised by ensuring the values are within zone boundaries.
978 cc->migrate_pfn = zone->compact_cached_migrate_pfn;
979 cc->free_pfn = zone->compact_cached_free_pfn;
980 if (cc->free_pfn < start_pfn || cc->free_pfn > end_pfn) {
981 cc->free_pfn = end_pfn & ~(pageblock_nr_pages-1);
982 zone->compact_cached_free_pfn = cc->free_pfn;
984 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn > end_pfn) {
985 cc->migrate_pfn = start_pfn;
986 zone->compact_cached_migrate_pfn = cc->migrate_pfn;
990 * Clear pageblock skip if there were failures recently and compaction
991 * is about to be retried after being deferred. kswapd does not do
992 * this reset as it'll reset the cached information when going to sleep.
994 if (compaction_restarting(zone, cc->order) && !current_is_kswapd())
995 __reset_isolation_suitable(zone);
997 migrate_prep_local();
999 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
1000 unsigned long nr_migrate, nr_remaining;
1003 switch (isolate_migratepages(zone, cc)) {
1005 ret = COMPACT_PARTIAL;
1006 putback_movable_pages(&cc->migratepages);
1007 cc->nr_migratepages = 0;
1011 case ISOLATE_SUCCESS:
1015 nr_migrate = cc->nr_migratepages;
1016 err = migrate_pages(&cc->migratepages, compaction_alloc,
1017 (unsigned long)cc, false,
1018 cc->sync ? MIGRATE_SYNC_LIGHT : MIGRATE_ASYNC);
1019 update_nr_listpages(cc);
1020 nr_remaining = cc->nr_migratepages;
1022 count_vm_event(COMPACTBLOCKS);
1023 count_vm_events(COMPACTPAGES, nr_migrate - nr_remaining);
1025 count_vm_events(COMPACTPAGEFAILED, nr_remaining);
1026 trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
1029 /* Release isolated pages not migrated */
1031 putback_movable_pages(&cc->migratepages);
1032 cc->nr_migratepages = 0;
1033 if (err == -ENOMEM) {
1034 ret = COMPACT_PARTIAL;
1039 /* Capture a page now if it is a suitable size */
1040 compact_capture_page(cc);
1044 /* Release free pages and check accounting */
1045 cc->nr_freepages -= release_freepages(&cc->freepages);
1046 VM_BUG_ON(cc->nr_freepages != 0);
1051 static unsigned long compact_zone_order(struct zone *zone,
1052 int order, gfp_t gfp_mask,
1053 bool sync, bool *contended,
1057 struct compact_control cc = {
1059 .nr_migratepages = 0,
1061 .migratetype = allocflags_to_migratetype(gfp_mask),
1066 INIT_LIST_HEAD(&cc.freepages);
1067 INIT_LIST_HEAD(&cc.migratepages);
1069 ret = compact_zone(zone, &cc);
1071 VM_BUG_ON(!list_empty(&cc.freepages));
1072 VM_BUG_ON(!list_empty(&cc.migratepages));
1074 *contended = cc.contended;
1078 int sysctl_extfrag_threshold = 500;
1081 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1082 * @zonelist: The zonelist used for the current allocation
1083 * @order: The order of the current allocation
1084 * @gfp_mask: The GFP mask of the current allocation
1085 * @nodemask: The allowed nodes to allocate from
1086 * @sync: Whether migration is synchronous or not
1087 * @contended: Return value that is true if compaction was aborted due to lock contention
1088 * @page: Optionally capture a free page of the requested order during compaction
1090 * This is the main entry point for direct page compaction.
1092 unsigned long try_to_compact_pages(struct zonelist *zonelist,
1093 int order, gfp_t gfp_mask, nodemask_t *nodemask,
1094 bool sync, bool *contended, struct page **page)
1096 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1097 int may_enter_fs = gfp_mask & __GFP_FS;
1098 int may_perform_io = gfp_mask & __GFP_IO;
1101 int rc = COMPACT_SKIPPED;
1102 int alloc_flags = 0;
1104 /* Check if the GFP flags allow compaction */
1105 if (!order || !may_enter_fs || !may_perform_io)
1108 count_vm_event(COMPACTSTALL);
1111 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
1112 alloc_flags |= ALLOC_CMA;
1114 /* Compact each zone in the list */
1115 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1119 status = compact_zone_order(zone, order, gfp_mask, sync,
1121 rc = max(status, rc);
1123 /* If a normal allocation would succeed, stop compacting */
1124 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0,
1133 /* Compact all zones within a node */
1134 static int __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
1139 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1141 zone = &pgdat->node_zones[zoneid];
1142 if (!populated_zone(zone))
1145 cc->nr_freepages = 0;
1146 cc->nr_migratepages = 0;
1148 INIT_LIST_HEAD(&cc->freepages);
1149 INIT_LIST_HEAD(&cc->migratepages);
1151 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
1152 compact_zone(zone, cc);
1154 if (cc->order > 0) {
1155 int ok = zone_watermark_ok(zone, cc->order,
1156 low_wmark_pages(zone), 0, 0);
1157 if (ok && cc->order >= zone->compact_order_failed)
1158 zone->compact_order_failed = cc->order + 1;
1159 /* Currently async compaction is never deferred. */
1160 else if (!ok && cc->sync)
1161 defer_compaction(zone, cc->order);
1164 VM_BUG_ON(!list_empty(&cc->freepages));
1165 VM_BUG_ON(!list_empty(&cc->migratepages));
1171 int compact_pgdat(pg_data_t *pgdat, int order)
1173 struct compact_control cc = {
1179 return __compact_pgdat(pgdat, &cc);
1182 static int compact_node(int nid)
1184 struct compact_control cc = {
1190 return __compact_pgdat(NODE_DATA(nid), &cc);
1193 /* Compact all nodes in the system */
1194 static int compact_nodes(void)
1198 /* Flush pending updates to the LRU lists */
1199 lru_add_drain_all();
1201 for_each_online_node(nid)
1204 return COMPACT_COMPLETE;
1207 /* The written value is actually unused, all memory is compacted */
1208 int sysctl_compact_memory;
1210 /* This is the entry point for compacting all nodes via /proc/sys/vm */
1211 int sysctl_compaction_handler(struct ctl_table *table, int write,
1212 void __user *buffer, size_t *length, loff_t *ppos)
1215 return compact_nodes();
1220 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1221 void __user *buffer, size_t *length, loff_t *ppos)
1223 proc_dointvec_minmax(table, write, buffer, length, ppos);
1228 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1229 ssize_t sysfs_compact_node(struct device *dev,
1230 struct device_attribute *attr,
1231 const char *buf, size_t count)
1235 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1236 /* Flush pending updates to the LRU lists */
1237 lru_add_drain_all();
1244 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1246 int compaction_register_node(struct node *node)
1248 return device_create_file(&node->dev, &dev_attr_compact);
1251 void compaction_unregister_node(struct node *node)
1253 return device_remove_file(&node->dev, &dev_attr_compact);
1255 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1257 #endif /* CONFIG_COMPACTION */