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;
76 * Do not reset more than once every five seconds. If allocations are
77 * failing sufficiently quickly to allow this to happen then continually
78 * scanning for compaction is not going to help. The choice of five
79 * seconds is arbitrary but will mitigate excessive scanning.
81 if (time_before(jiffies, zone->compact_blockskip_expire))
84 zone->compact_cached_migrate_pfn = start_pfn;
85 zone->compact_cached_free_pfn = end_pfn;
86 zone->compact_blockskip_expire = jiffies + (HZ * 5);
88 /* Walk the zone and mark every pageblock as suitable for isolation */
89 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
97 page = pfn_to_page(pfn);
98 if (zone != page_zone(page))
101 clear_pageblock_skip(page);
106 * If no pages were isolated then mark this pageblock to be skipped in the
107 * future. The information is later cleared by reset_isolation_suitable().
109 static void update_pageblock_skip(struct compact_control *cc,
110 struct page *page, unsigned long nr_isolated,
111 bool migrate_scanner)
113 struct zone *zone = cc->zone;
118 unsigned long pfn = page_to_pfn(page);
119 set_pageblock_skip(page);
121 /* Update where compaction should restart */
122 if (migrate_scanner) {
123 if (!cc->finished_update_migrate &&
124 pfn > zone->compact_cached_migrate_pfn)
125 zone->compact_cached_migrate_pfn = pfn;
127 if (!cc->finished_update_free &&
128 pfn < zone->compact_cached_free_pfn)
129 zone->compact_cached_free_pfn = pfn;
134 static inline bool isolation_suitable(struct compact_control *cc,
140 static void update_pageblock_skip(struct compact_control *cc,
141 struct page *page, unsigned long nr_isolated,
142 bool migrate_scanner)
145 #endif /* CONFIG_COMPACTION */
147 static inline bool should_release_lock(spinlock_t *lock)
149 return need_resched() || spin_is_contended(lock);
153 * Compaction requires the taking of some coarse locks that are potentially
154 * very heavily contended. Check if the process needs to be scheduled or
155 * if the lock is contended. For async compaction, back out in the event
156 * if contention is severe. For sync compaction, schedule.
158 * Returns true if the lock is held.
159 * Returns false if the lock is released and compaction should abort
161 static bool compact_checklock_irqsave(spinlock_t *lock, unsigned long *flags,
162 bool locked, struct compact_control *cc)
164 if (should_release_lock(lock)) {
166 spin_unlock_irqrestore(lock, *flags);
170 /* async aborts if taking too long or contended */
172 cc->contended = true;
180 spin_lock_irqsave(lock, *flags);
184 static inline bool compact_trylock_irqsave(spinlock_t *lock,
185 unsigned long *flags, struct compact_control *cc)
187 return compact_checklock_irqsave(lock, flags, false, cc);
190 /* Returns true if the page is within a block suitable for migration to */
191 static bool suitable_migration_target(struct page *page)
193 int migratetype = get_pageblock_migratetype(page);
195 /* Don't interfere with memory hot-remove or the min_free_kbytes blocks */
196 if (migratetype == MIGRATE_ISOLATE || migratetype == MIGRATE_RESERVE)
199 /* If the page is a large free page, then allow migration */
200 if (PageBuddy(page) && page_order(page) >= pageblock_order)
203 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
204 if (migrate_async_suitable(migratetype))
207 /* Otherwise skip the block */
211 static void compact_capture_page(struct compact_control *cc)
214 int mtype, mtype_low, mtype_high;
216 if (!cc->page || *cc->page)
220 * For MIGRATE_MOVABLE allocations we capture a suitable page ASAP
221 * regardless of the migratetype of the freelist is is captured from.
222 * This is fine because the order for a high-order MIGRATE_MOVABLE
223 * allocation is typically at least a pageblock size and overall
224 * fragmentation is not impaired. Other allocation types must
225 * capture pages from their own migratelist because otherwise they
226 * could pollute other pageblocks like MIGRATE_MOVABLE with
227 * difficult to move pages and making fragmentation worse overall.
229 if (cc->migratetype == MIGRATE_MOVABLE) {
231 mtype_high = MIGRATE_PCPTYPES;
233 mtype_low = cc->migratetype;
234 mtype_high = cc->migratetype + 1;
237 /* Speculatively examine the free lists without zone lock */
238 for (mtype = mtype_low; mtype < mtype_high; mtype++) {
240 for (order = cc->order; order < MAX_ORDER; order++) {
242 struct free_area *area;
243 area = &(cc->zone->free_area[order]);
244 if (list_empty(&area->free_list[mtype]))
247 /* Take the lock and attempt capture of the page */
248 if (!compact_trylock_irqsave(&cc->zone->lock, &flags, cc))
250 if (!list_empty(&area->free_list[mtype])) {
251 page = list_entry(area->free_list[mtype].next,
253 if (capture_free_page(page, cc->order, mtype)) {
254 spin_unlock_irqrestore(&cc->zone->lock,
260 spin_unlock_irqrestore(&cc->zone->lock, flags);
266 * Isolate free pages onto a private freelist. Caller must hold zone->lock.
267 * If @strict is true, will abort returning 0 on any invalid PFNs or non-free
268 * pages inside of the pageblock (even though it may still end up isolating
271 static unsigned long isolate_freepages_block(struct compact_control *cc,
272 unsigned long blockpfn,
273 unsigned long end_pfn,
274 struct list_head *freelist,
277 int nr_scanned = 0, total_isolated = 0;
278 struct page *cursor, *valid_page = NULL;
279 unsigned long nr_strict_required = end_pfn - blockpfn;
283 cursor = pfn_to_page(blockpfn);
285 /* Isolate free pages. */
286 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
288 struct page *page = cursor;
291 if (!pfn_valid_within(blockpfn))
295 if (!PageBuddy(page))
299 * The zone lock must be held to isolate freepages.
300 * Unfortunately this is a very coarse lock and can be
301 * heavily contended if there are parallel allocations
302 * or parallel compactions. For async compaction do not
303 * spin on the lock and we acquire the lock as late as
306 locked = compact_checklock_irqsave(&cc->zone->lock, &flags,
311 /* Recheck this is a suitable migration target under lock */
312 if (!strict && !suitable_migration_target(page))
315 /* Recheck this is a buddy page under lock */
316 if (!PageBuddy(page))
319 /* Found a free page, break it into order-0 pages */
320 isolated = split_free_page(page);
321 if (!isolated && strict)
323 total_isolated += isolated;
324 for (i = 0; i < isolated; i++) {
325 list_add(&page->lru, freelist);
329 /* If a page was split, advance to the end of it */
331 blockpfn += isolated - 1;
332 cursor += isolated - 1;
336 trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
339 * If strict isolation is requested by CMA then check that all the
340 * pages requested were isolated. If there were any failures, 0 is
341 * returned and CMA will fail.
343 if (strict && nr_strict_required != total_isolated)
347 spin_unlock_irqrestore(&cc->zone->lock, flags);
349 /* Update the pageblock-skip if the whole pageblock was scanned */
350 if (blockpfn == end_pfn)
351 update_pageblock_skip(cc, valid_page, total_isolated, false);
353 return total_isolated;
357 * isolate_freepages_range() - isolate free pages.
358 * @start_pfn: The first PFN to start isolating.
359 * @end_pfn: The one-past-last PFN.
361 * Non-free pages, invalid PFNs, or zone boundaries within the
362 * [start_pfn, end_pfn) range are considered errors, cause function to
363 * undo its actions and return zero.
365 * Otherwise, function returns one-past-the-last PFN of isolated page
366 * (which may be greater then end_pfn if end fell in a middle of
370 isolate_freepages_range(struct compact_control *cc,
371 unsigned long start_pfn, unsigned long end_pfn)
373 unsigned long isolated, pfn, block_end_pfn;
376 for (pfn = start_pfn; pfn < end_pfn; pfn += isolated) {
377 if (!pfn_valid(pfn) || cc->zone != page_zone(pfn_to_page(pfn)))
381 * On subsequent iterations ALIGN() is actually not needed,
382 * but we keep it that we not to complicate the code.
384 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
385 block_end_pfn = min(block_end_pfn, end_pfn);
387 isolated = isolate_freepages_block(cc, pfn, block_end_pfn,
391 * In strict mode, isolate_freepages_block() returns 0 if
392 * there are any holes in the block (ie. invalid PFNs or
399 * If we managed to isolate pages, it is always (1 << n) *
400 * pageblock_nr_pages for some non-negative n. (Max order
401 * page may span two pageblocks).
405 /* split_free_page does not map the pages */
406 map_pages(&freelist);
409 /* Loop terminated early, cleanup. */
410 release_freepages(&freelist);
414 /* We don't use freelists for anything. */
418 /* Update the number of anon and file isolated pages in the zone */
419 static void acct_isolated(struct zone *zone, bool locked, struct compact_control *cc)
422 unsigned int count[2] = { 0, };
424 list_for_each_entry(page, &cc->migratepages, lru)
425 count[!!page_is_file_cache(page)]++;
427 /* If locked we can use the interrupt unsafe versions */
429 __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
430 __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
432 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
433 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
437 /* Similar to reclaim, but different enough that they don't share logic */
438 static bool too_many_isolated(struct zone *zone)
440 unsigned long active, inactive, isolated;
442 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
443 zone_page_state(zone, NR_INACTIVE_ANON);
444 active = zone_page_state(zone, NR_ACTIVE_FILE) +
445 zone_page_state(zone, NR_ACTIVE_ANON);
446 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
447 zone_page_state(zone, NR_ISOLATED_ANON);
449 return isolated > (inactive + active) / 2;
453 * isolate_migratepages_range() - isolate all migrate-able pages in range.
454 * @zone: Zone pages are in.
455 * @cc: Compaction control structure.
456 * @low_pfn: The first PFN of the range.
457 * @end_pfn: The one-past-the-last PFN of the range.
459 * Isolate all pages that can be migrated from the range specified by
460 * [low_pfn, end_pfn). Returns zero if there is a fatal signal
461 * pending), otherwise PFN of the first page that was not scanned
462 * (which may be both less, equal to or more then end_pfn).
464 * Assumes that cc->migratepages is empty and cc->nr_migratepages is
467 * Apart from cc->migratepages and cc->nr_migratetypes this function
468 * does not modify any cc's fields, in particular it does not modify
469 * (or read for that matter) cc->migrate_pfn.
472 isolate_migratepages_range(struct zone *zone, struct compact_control *cc,
473 unsigned long low_pfn, unsigned long end_pfn)
475 unsigned long last_pageblock_nr = 0, pageblock_nr;
476 unsigned long nr_scanned = 0, nr_isolated = 0;
477 struct list_head *migratelist = &cc->migratepages;
478 isolate_mode_t mode = 0;
479 struct lruvec *lruvec;
482 struct page *page = NULL, *valid_page = NULL;
485 * Ensure that there are not too many pages isolated from the LRU
486 * list by either parallel reclaimers or compaction. If there are,
487 * delay for some time until fewer pages are isolated
489 while (unlikely(too_many_isolated(zone))) {
490 /* async migration should just abort */
494 congestion_wait(BLK_RW_ASYNC, HZ/10);
496 if (fatal_signal_pending(current))
500 /* Time to isolate some pages for migration */
502 for (; low_pfn < end_pfn; low_pfn++) {
503 /* give a chance to irqs before checking need_resched() */
504 if (locked && !((low_pfn+1) % SWAP_CLUSTER_MAX)) {
505 if (should_release_lock(&zone->lru_lock)) {
506 spin_unlock_irqrestore(&zone->lru_lock, flags);
512 * migrate_pfn does not necessarily start aligned to a
513 * pageblock. Ensure that pfn_valid is called when moving
514 * into a new MAX_ORDER_NR_PAGES range in case of large
515 * memory holes within the zone
517 if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
518 if (!pfn_valid(low_pfn)) {
519 low_pfn += MAX_ORDER_NR_PAGES - 1;
524 if (!pfn_valid_within(low_pfn))
529 * Get the page and ensure the page is within the same zone.
530 * See the comment in isolate_freepages about overlapping
531 * nodes. It is deliberate that the new zone lock is not taken
532 * as memory compaction should not move pages between nodes.
534 page = pfn_to_page(low_pfn);
535 if (page_zone(page) != zone)
541 /* If isolation recently failed, do not retry */
542 pageblock_nr = low_pfn >> pageblock_order;
543 if (!isolation_suitable(cc, page))
551 * For async migration, also only scan in MOVABLE blocks. Async
552 * migration is optimistic to see if the minimum amount of work
553 * satisfies the allocation
555 if (!cc->sync && last_pageblock_nr != pageblock_nr &&
556 !migrate_async_suitable(get_pageblock_migratetype(page))) {
557 cc->finished_update_migrate = true;
561 /* Check may be lockless but that's ok as we recheck later */
566 * PageLRU is set. lru_lock normally excludes isolation
567 * splitting and collapsing (collapsing has already happened
568 * if PageLRU is set) but the lock is not necessarily taken
569 * here and it is wasteful to take it just to check transhuge.
570 * Check TransHuge without lock and skip the whole pageblock if
571 * it's either a transhuge or hugetlbfs page, as calling
572 * compound_order() without preventing THP from splitting the
573 * page underneath us may return surprising results.
575 if (PageTransHuge(page)) {
578 low_pfn += (1 << compound_order(page)) - 1;
582 /* Check if it is ok to still hold the lock */
583 locked = compact_checklock_irqsave(&zone->lru_lock, &flags,
585 if (!locked || fatal_signal_pending(current))
588 /* Recheck PageLRU and PageTransHuge under lock */
591 if (PageTransHuge(page)) {
592 low_pfn += (1 << compound_order(page)) - 1;
597 mode |= ISOLATE_ASYNC_MIGRATE;
599 lruvec = mem_cgroup_page_lruvec(page, zone);
601 /* Try isolate the page */
602 if (__isolate_lru_page(page, mode) != 0)
605 VM_BUG_ON(PageTransCompound(page));
607 /* Successfully isolated */
608 cc->finished_update_migrate = true;
609 del_page_from_lru_list(page, lruvec, page_lru(page));
610 list_add(&page->lru, migratelist);
611 cc->nr_migratepages++;
614 /* Avoid isolating too much */
615 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
623 low_pfn += pageblock_nr_pages;
624 low_pfn = ALIGN(low_pfn, pageblock_nr_pages) - 1;
625 last_pageblock_nr = pageblock_nr;
628 acct_isolated(zone, locked, cc);
631 spin_unlock_irqrestore(&zone->lru_lock, flags);
633 /* Update the pageblock-skip if the whole pageblock was scanned */
634 if (low_pfn == end_pfn)
635 update_pageblock_skip(cc, valid_page, nr_isolated, true);
637 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
642 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
643 #ifdef CONFIG_COMPACTION
645 * Based on information in the current compact_control, find blocks
646 * suitable for isolating free pages from and then isolate them.
648 static void isolate_freepages(struct zone *zone,
649 struct compact_control *cc)
652 unsigned long high_pfn, low_pfn, pfn, zone_end_pfn, end_pfn;
653 int nr_freepages = cc->nr_freepages;
654 struct list_head *freelist = &cc->freepages;
657 * Initialise the free scanner. The starting point is where we last
658 * scanned from (or the end of the zone if starting). The low point
659 * is the end of the pageblock the migration scanner is using.
662 low_pfn = cc->migrate_pfn + pageblock_nr_pages;
665 * Take care that if the migration scanner is at the end of the zone
666 * that the free scanner does not accidentally move to the next zone
667 * in the next isolation cycle.
669 high_pfn = min(low_pfn, pfn);
671 zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
674 * Isolate free pages until enough are available to migrate the
675 * pages on cc->migratepages. We stop searching if the migrate
676 * and free page scanners meet or enough free pages are isolated.
678 for (; pfn > low_pfn && cc->nr_migratepages > nr_freepages;
679 pfn -= pageblock_nr_pages) {
680 unsigned long isolated;
686 * Check for overlapping nodes/zones. It's possible on some
687 * configurations to have a setup like
689 * i.e. it's possible that all pages within a zones range of
690 * pages do not belong to a single zone.
692 page = pfn_to_page(pfn);
693 if (page_zone(page) != zone)
696 /* Check the block is suitable for migration */
697 if (!suitable_migration_target(page))
700 /* If isolation recently failed, do not retry */
701 if (!isolation_suitable(cc, page))
704 /* Found a block suitable for isolating free pages from */
706 end_pfn = min(pfn + pageblock_nr_pages, zone_end_pfn);
707 isolated = isolate_freepages_block(cc, pfn, end_pfn,
709 nr_freepages += isolated;
712 * Record the highest PFN we isolated pages from. When next
713 * looking for free pages, the search will restart here as
714 * page migration may have returned some pages to the allocator
717 cc->finished_update_free = true;
718 high_pfn = max(high_pfn, pfn);
722 /* split_free_page does not map the pages */
725 cc->free_pfn = high_pfn;
726 cc->nr_freepages = nr_freepages;
730 * This is a migrate-callback that "allocates" freepages by taking pages
731 * from the isolated freelists in the block we are migrating to.
733 static struct page *compaction_alloc(struct page *migratepage,
737 struct compact_control *cc = (struct compact_control *)data;
738 struct page *freepage;
740 /* Isolate free pages if necessary */
741 if (list_empty(&cc->freepages)) {
742 isolate_freepages(cc->zone, cc);
744 if (list_empty(&cc->freepages))
748 freepage = list_entry(cc->freepages.next, struct page, lru);
749 list_del(&freepage->lru);
756 * We cannot control nr_migratepages and nr_freepages fully when migration is
757 * running as migrate_pages() has no knowledge of compact_control. When
758 * migration is complete, we count the number of pages on the lists by hand.
760 static void update_nr_listpages(struct compact_control *cc)
762 int nr_migratepages = 0;
763 int nr_freepages = 0;
766 list_for_each_entry(page, &cc->migratepages, lru)
768 list_for_each_entry(page, &cc->freepages, lru)
771 cc->nr_migratepages = nr_migratepages;
772 cc->nr_freepages = nr_freepages;
775 /* possible outcome of isolate_migratepages */
777 ISOLATE_ABORT, /* Abort compaction now */
778 ISOLATE_NONE, /* No pages isolated, continue scanning */
779 ISOLATE_SUCCESS, /* Pages isolated, migrate */
783 * Isolate all pages that can be migrated from the block pointed to by
784 * the migrate scanner within compact_control.
786 static isolate_migrate_t isolate_migratepages(struct zone *zone,
787 struct compact_control *cc)
789 unsigned long low_pfn, end_pfn;
791 /* Do not scan outside zone boundaries */
792 low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
794 /* Only scan within a pageblock boundary */
795 end_pfn = ALIGN(low_pfn + pageblock_nr_pages, pageblock_nr_pages);
797 /* Do not cross the free scanner or scan within a memory hole */
798 if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
799 cc->migrate_pfn = end_pfn;
803 /* Perform the isolation */
804 low_pfn = isolate_migratepages_range(zone, cc, low_pfn, end_pfn);
805 if (!low_pfn || cc->contended)
806 return ISOLATE_ABORT;
808 cc->migrate_pfn = low_pfn;
810 return ISOLATE_SUCCESS;
813 static int compact_finished(struct zone *zone,
814 struct compact_control *cc)
816 unsigned long watermark;
818 if (fatal_signal_pending(current))
819 return COMPACT_PARTIAL;
821 /* Compaction run completes if the migrate and free scanner meet */
822 if (cc->free_pfn <= cc->migrate_pfn) {
823 reset_isolation_suitable(cc->zone);
824 return COMPACT_COMPLETE;
828 * order == -1 is expected when compacting via
829 * /proc/sys/vm/compact_memory
832 return COMPACT_CONTINUE;
834 /* Compaction run is not finished if the watermark is not met */
835 watermark = low_wmark_pages(zone);
836 watermark += (1 << cc->order);
838 if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
839 return COMPACT_CONTINUE;
841 /* Direct compactor: Is a suitable page free? */
843 /* Was a suitable page captured? */
845 return COMPACT_PARTIAL;
848 for (order = cc->order; order < MAX_ORDER; order++) {
849 struct free_area *area = &zone->free_area[cc->order];
850 /* Job done if page is free of the right migratetype */
851 if (!list_empty(&area->free_list[cc->migratetype]))
852 return COMPACT_PARTIAL;
854 /* Job done if allocation would set block type */
855 if (cc->order >= pageblock_order && area->nr_free)
856 return COMPACT_PARTIAL;
860 return COMPACT_CONTINUE;
864 * compaction_suitable: Is this suitable to run compaction on this zone now?
866 * COMPACT_SKIPPED - If there are too few free pages for compaction
867 * COMPACT_PARTIAL - If the allocation would succeed without compaction
868 * COMPACT_CONTINUE - If compaction should run now
870 unsigned long compaction_suitable(struct zone *zone, int order)
873 unsigned long watermark;
876 * order == -1 is expected when compacting via
877 * /proc/sys/vm/compact_memory
880 return COMPACT_CONTINUE;
883 * Watermarks for order-0 must be met for compaction. Note the 2UL.
884 * This is because during migration, copies of pages need to be
885 * allocated and for a short time, the footprint is higher
887 watermark = low_wmark_pages(zone) + (2UL << order);
888 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
889 return COMPACT_SKIPPED;
892 * fragmentation index determines if allocation failures are due to
893 * low memory or external fragmentation
895 * index of -1000 implies allocations might succeed depending on
897 * index towards 0 implies failure is due to lack of memory
898 * index towards 1000 implies failure is due to fragmentation
900 * Only compact if a failure would be due to fragmentation.
902 fragindex = fragmentation_index(zone, order);
903 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
904 return COMPACT_SKIPPED;
906 if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
908 return COMPACT_PARTIAL;
910 return COMPACT_CONTINUE;
913 static int compact_zone(struct zone *zone, struct compact_control *cc)
916 unsigned long start_pfn = zone->zone_start_pfn;
917 unsigned long end_pfn = zone->zone_start_pfn + zone->spanned_pages;
919 ret = compaction_suitable(zone, cc->order);
921 case COMPACT_PARTIAL:
922 case COMPACT_SKIPPED:
923 /* Compaction is likely to fail */
925 case COMPACT_CONTINUE:
926 /* Fall through to compaction */
931 * Setup to move all movable pages to the end of the zone. Used cached
932 * information on where the scanners should start but check that it
933 * is initialised by ensuring the values are within zone boundaries.
935 cc->migrate_pfn = zone->compact_cached_migrate_pfn;
936 cc->free_pfn = zone->compact_cached_free_pfn;
937 if (cc->free_pfn < start_pfn || cc->free_pfn > end_pfn) {
938 cc->free_pfn = end_pfn & ~(pageblock_nr_pages-1);
939 zone->compact_cached_free_pfn = cc->free_pfn;
941 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn > end_pfn) {
942 cc->migrate_pfn = start_pfn;
943 zone->compact_cached_migrate_pfn = cc->migrate_pfn;
946 /* Clear pageblock skip if there are numerous alloc failures */
947 if (zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT)
948 reset_isolation_suitable(zone);
950 migrate_prep_local();
952 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
953 unsigned long nr_migrate, nr_remaining;
956 switch (isolate_migratepages(zone, cc)) {
958 ret = COMPACT_PARTIAL;
959 putback_lru_pages(&cc->migratepages);
960 cc->nr_migratepages = 0;
964 case ISOLATE_SUCCESS:
968 nr_migrate = cc->nr_migratepages;
969 err = migrate_pages(&cc->migratepages, compaction_alloc,
970 (unsigned long)cc, false,
971 cc->sync ? MIGRATE_SYNC_LIGHT : MIGRATE_ASYNC);
972 update_nr_listpages(cc);
973 nr_remaining = cc->nr_migratepages;
975 count_vm_event(COMPACTBLOCKS);
976 count_vm_events(COMPACTPAGES, nr_migrate - nr_remaining);
978 count_vm_events(COMPACTPAGEFAILED, nr_remaining);
979 trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
982 /* Release LRU pages not migrated */
984 putback_lru_pages(&cc->migratepages);
985 cc->nr_migratepages = 0;
986 if (err == -ENOMEM) {
987 ret = COMPACT_PARTIAL;
992 /* Capture a page now if it is a suitable size */
993 compact_capture_page(cc);
997 /* Release free pages and check accounting */
998 cc->nr_freepages -= release_freepages(&cc->freepages);
999 VM_BUG_ON(cc->nr_freepages != 0);
1004 static unsigned long compact_zone_order(struct zone *zone,
1005 int order, gfp_t gfp_mask,
1006 bool sync, bool *contended,
1010 struct compact_control cc = {
1012 .nr_migratepages = 0,
1014 .migratetype = allocflags_to_migratetype(gfp_mask),
1019 INIT_LIST_HEAD(&cc.freepages);
1020 INIT_LIST_HEAD(&cc.migratepages);
1022 ret = compact_zone(zone, &cc);
1024 VM_BUG_ON(!list_empty(&cc.freepages));
1025 VM_BUG_ON(!list_empty(&cc.migratepages));
1027 *contended = cc.contended;
1031 int sysctl_extfrag_threshold = 500;
1034 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1035 * @zonelist: The zonelist used for the current allocation
1036 * @order: The order of the current allocation
1037 * @gfp_mask: The GFP mask of the current allocation
1038 * @nodemask: The allowed nodes to allocate from
1039 * @sync: Whether migration is synchronous or not
1040 * @contended: Return value that is true if compaction was aborted due to lock contention
1041 * @page: Optionally capture a free page of the requested order during compaction
1043 * This is the main entry point for direct page compaction.
1045 unsigned long try_to_compact_pages(struct zonelist *zonelist,
1046 int order, gfp_t gfp_mask, nodemask_t *nodemask,
1047 bool sync, bool *contended, struct page **page)
1049 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1050 int may_enter_fs = gfp_mask & __GFP_FS;
1051 int may_perform_io = gfp_mask & __GFP_IO;
1054 int rc = COMPACT_SKIPPED;
1055 int alloc_flags = 0;
1057 /* Check if the GFP flags allow compaction */
1058 if (!order || !may_enter_fs || !may_perform_io)
1061 count_vm_event(COMPACTSTALL);
1064 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
1065 alloc_flags |= ALLOC_CMA;
1067 /* Compact each zone in the list */
1068 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1072 status = compact_zone_order(zone, order, gfp_mask, sync,
1074 rc = max(status, rc);
1076 /* If a normal allocation would succeed, stop compacting */
1077 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0,
1086 /* Compact all zones within a node */
1087 static int __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
1092 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1094 zone = &pgdat->node_zones[zoneid];
1095 if (!populated_zone(zone))
1098 cc->nr_freepages = 0;
1099 cc->nr_migratepages = 0;
1101 INIT_LIST_HEAD(&cc->freepages);
1102 INIT_LIST_HEAD(&cc->migratepages);
1104 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
1105 compact_zone(zone, cc);
1107 if (cc->order > 0) {
1108 int ok = zone_watermark_ok(zone, cc->order,
1109 low_wmark_pages(zone), 0, 0);
1110 if (ok && cc->order >= zone->compact_order_failed)
1111 zone->compact_order_failed = cc->order + 1;
1112 /* Currently async compaction is never deferred. */
1113 else if (!ok && cc->sync)
1114 defer_compaction(zone, cc->order);
1117 VM_BUG_ON(!list_empty(&cc->freepages));
1118 VM_BUG_ON(!list_empty(&cc->migratepages));
1124 int compact_pgdat(pg_data_t *pgdat, int order)
1126 struct compact_control cc = {
1132 return __compact_pgdat(pgdat, &cc);
1135 static int compact_node(int nid)
1137 struct compact_control cc = {
1143 return __compact_pgdat(NODE_DATA(nid), &cc);
1146 /* Compact all nodes in the system */
1147 static int compact_nodes(void)
1151 /* Flush pending updates to the LRU lists */
1152 lru_add_drain_all();
1154 for_each_online_node(nid)
1157 return COMPACT_COMPLETE;
1160 /* The written value is actually unused, all memory is compacted */
1161 int sysctl_compact_memory;
1163 /* This is the entry point for compacting all nodes via /proc/sys/vm */
1164 int sysctl_compaction_handler(struct ctl_table *table, int write,
1165 void __user *buffer, size_t *length, loff_t *ppos)
1168 return compact_nodes();
1173 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1174 void __user *buffer, size_t *length, loff_t *ppos)
1176 proc_dointvec_minmax(table, write, buffer, length, ppos);
1181 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1182 ssize_t sysfs_compact_node(struct device *dev,
1183 struct device_attribute *attr,
1184 const char *buf, size_t count)
1188 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1189 /* Flush pending updates to the LRU lists */
1190 lru_add_drain_all();
1197 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1199 int compaction_register_node(struct node *node)
1201 return device_create_file(&node->dev, &dev_attr_compact);
1204 void compaction_unregister_node(struct node *node)
1206 return device_remove_file(&node->dev, &dev_attr_compact);
1208 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1210 #endif /* CONFIG_COMPACTION */