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 static inline bool should_release_lock(spinlock_t *lock)
55 return need_resched() || spin_is_contended(lock);
59 * Compaction requires the taking of some coarse locks that are potentially
60 * very heavily contended. Check if the process needs to be scheduled or
61 * if the lock is contended. For async compaction, back out in the event
62 * if contention is severe. For sync compaction, schedule.
64 * Returns true if the lock is held.
65 * Returns false if the lock is released and compaction should abort
67 static bool compact_checklock_irqsave(spinlock_t *lock, unsigned long *flags,
68 bool locked, struct compact_control *cc)
70 if (should_release_lock(lock)) {
72 spin_unlock_irqrestore(lock, *flags);
76 /* async aborts if taking too long or contended */
86 spin_lock_irqsave(lock, *flags);
90 static inline bool compact_trylock_irqsave(spinlock_t *lock,
91 unsigned long *flags, struct compact_control *cc)
93 return compact_checklock_irqsave(lock, flags, false, cc);
96 static void compact_capture_page(struct compact_control *cc)
99 int mtype, mtype_low, mtype_high;
101 if (!cc->page || *cc->page)
105 * For MIGRATE_MOVABLE allocations we capture a suitable page ASAP
106 * regardless of the migratetype of the freelist is is captured from.
107 * This is fine because the order for a high-order MIGRATE_MOVABLE
108 * allocation is typically at least a pageblock size and overall
109 * fragmentation is not impaired. Other allocation types must
110 * capture pages from their own migratelist because otherwise they
111 * could pollute other pageblocks like MIGRATE_MOVABLE with
112 * difficult to move pages and making fragmentation worse overall.
114 if (cc->migratetype == MIGRATE_MOVABLE) {
116 mtype_high = MIGRATE_PCPTYPES;
118 mtype_low = cc->migratetype;
119 mtype_high = cc->migratetype + 1;
122 /* Speculatively examine the free lists without zone lock */
123 for (mtype = mtype_low; mtype < mtype_high; mtype++) {
125 for (order = cc->order; order < MAX_ORDER; order++) {
127 struct free_area *area;
128 area = &(cc->zone->free_area[order]);
129 if (list_empty(&area->free_list[mtype]))
132 /* Take the lock and attempt capture of the page */
133 if (!compact_trylock_irqsave(&cc->zone->lock, &flags, cc))
135 if (!list_empty(&area->free_list[mtype])) {
136 page = list_entry(area->free_list[mtype].next,
138 if (capture_free_page(page, cc->order, mtype)) {
139 spin_unlock_irqrestore(&cc->zone->lock,
145 spin_unlock_irqrestore(&cc->zone->lock, flags);
151 * Isolate free pages onto a private freelist. Caller must hold zone->lock.
152 * If @strict is true, will abort returning 0 on any invalid PFNs or non-free
153 * pages inside of the pageblock (even though it may still end up isolating
156 static unsigned long isolate_freepages_block(unsigned long blockpfn,
157 unsigned long end_pfn,
158 struct list_head *freelist,
161 int nr_scanned = 0, total_isolated = 0;
164 cursor = pfn_to_page(blockpfn);
166 /* Isolate free pages. This assumes the block is valid */
167 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
169 struct page *page = cursor;
171 if (!pfn_valid_within(blockpfn)) {
178 if (!PageBuddy(page)) {
184 /* Found a free page, break it into order-0 pages */
185 isolated = split_free_page(page);
186 if (!isolated && strict)
188 total_isolated += isolated;
189 for (i = 0; i < isolated; i++) {
190 list_add(&page->lru, freelist);
194 /* If a page was split, advance to the end of it */
196 blockpfn += isolated - 1;
197 cursor += isolated - 1;
201 trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
202 return total_isolated;
206 * isolate_freepages_range() - isolate free pages.
207 * @start_pfn: The first PFN to start isolating.
208 * @end_pfn: The one-past-last PFN.
210 * Non-free pages, invalid PFNs, or zone boundaries within the
211 * [start_pfn, end_pfn) range are considered errors, cause function to
212 * undo its actions and return zero.
214 * Otherwise, function returns one-past-the-last PFN of isolated page
215 * (which may be greater then end_pfn if end fell in a middle of
219 isolate_freepages_range(unsigned long start_pfn, unsigned long end_pfn)
221 unsigned long isolated, pfn, block_end_pfn, flags;
222 struct zone *zone = NULL;
225 if (pfn_valid(start_pfn))
226 zone = page_zone(pfn_to_page(start_pfn));
228 for (pfn = start_pfn; pfn < end_pfn; pfn += isolated) {
229 if (!pfn_valid(pfn) || zone != page_zone(pfn_to_page(pfn)))
233 * On subsequent iterations ALIGN() is actually not needed,
234 * but we keep it that we not to complicate the code.
236 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
237 block_end_pfn = min(block_end_pfn, end_pfn);
239 spin_lock_irqsave(&zone->lock, flags);
240 isolated = isolate_freepages_block(pfn, block_end_pfn,
242 spin_unlock_irqrestore(&zone->lock, flags);
245 * In strict mode, isolate_freepages_block() returns 0 if
246 * there are any holes in the block (ie. invalid PFNs or
253 * If we managed to isolate pages, it is always (1 << n) *
254 * pageblock_nr_pages for some non-negative n. (Max order
255 * page may span two pageblocks).
259 /* split_free_page does not map the pages */
260 map_pages(&freelist);
263 /* Loop terminated early, cleanup. */
264 release_freepages(&freelist);
268 /* We don't use freelists for anything. */
272 /* Update the number of anon and file isolated pages in the zone */
273 static void acct_isolated(struct zone *zone, bool locked, struct compact_control *cc)
276 unsigned int count[2] = { 0, };
278 list_for_each_entry(page, &cc->migratepages, lru)
279 count[!!page_is_file_cache(page)]++;
281 /* If locked we can use the interrupt unsafe versions */
283 __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
284 __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
286 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
287 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
291 /* Similar to reclaim, but different enough that they don't share logic */
292 static bool too_many_isolated(struct zone *zone)
294 unsigned long active, inactive, isolated;
296 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
297 zone_page_state(zone, NR_INACTIVE_ANON);
298 active = zone_page_state(zone, NR_ACTIVE_FILE) +
299 zone_page_state(zone, NR_ACTIVE_ANON);
300 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
301 zone_page_state(zone, NR_ISOLATED_ANON);
303 return isolated > (inactive + active) / 2;
307 * isolate_migratepages_range() - isolate all migrate-able pages in range.
308 * @zone: Zone pages are in.
309 * @cc: Compaction control structure.
310 * @low_pfn: The first PFN of the range.
311 * @end_pfn: The one-past-the-last PFN of the range.
313 * Isolate all pages that can be migrated from the range specified by
314 * [low_pfn, end_pfn). Returns zero if there is a fatal signal
315 * pending), otherwise PFN of the first page that was not scanned
316 * (which may be both less, equal to or more then end_pfn).
318 * Assumes that cc->migratepages is empty and cc->nr_migratepages is
321 * Apart from cc->migratepages and cc->nr_migratetypes this function
322 * does not modify any cc's fields, in particular it does not modify
323 * (or read for that matter) cc->migrate_pfn.
326 isolate_migratepages_range(struct zone *zone, struct compact_control *cc,
327 unsigned long low_pfn, unsigned long end_pfn)
329 unsigned long last_pageblock_nr = 0, pageblock_nr;
330 unsigned long nr_scanned = 0, nr_isolated = 0;
331 struct list_head *migratelist = &cc->migratepages;
332 isolate_mode_t mode = 0;
333 struct lruvec *lruvec;
338 * Ensure that there are not too many pages isolated from the LRU
339 * list by either parallel reclaimers or compaction. If there are,
340 * delay for some time until fewer pages are isolated
342 while (unlikely(too_many_isolated(zone))) {
343 /* async migration should just abort */
347 congestion_wait(BLK_RW_ASYNC, HZ/10);
349 if (fatal_signal_pending(current))
353 /* Time to isolate some pages for migration */
355 for (; low_pfn < end_pfn; low_pfn++) {
358 /* give a chance to irqs before checking need_resched() */
359 if (locked && !((low_pfn+1) % SWAP_CLUSTER_MAX)) {
360 if (should_release_lock(&zone->lru_lock)) {
361 spin_unlock_irqrestore(&zone->lru_lock, flags);
367 * migrate_pfn does not necessarily start aligned to a
368 * pageblock. Ensure that pfn_valid is called when moving
369 * into a new MAX_ORDER_NR_PAGES range in case of large
370 * memory holes within the zone
372 if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
373 if (!pfn_valid(low_pfn)) {
374 low_pfn += MAX_ORDER_NR_PAGES - 1;
379 if (!pfn_valid_within(low_pfn))
384 * Get the page and ensure the page is within the same zone.
385 * See the comment in isolate_freepages about overlapping
386 * nodes. It is deliberate that the new zone lock is not taken
387 * as memory compaction should not move pages between nodes.
389 page = pfn_to_page(low_pfn);
390 if (page_zone(page) != zone)
398 * For async migration, also only scan in MOVABLE blocks. Async
399 * migration is optimistic to see if the minimum amount of work
400 * satisfies the allocation
402 pageblock_nr = low_pfn >> pageblock_order;
403 if (!cc->sync && last_pageblock_nr != pageblock_nr &&
404 !migrate_async_suitable(get_pageblock_migratetype(page))) {
408 /* Check may be lockless but that's ok as we recheck later */
413 * PageLRU is set. lru_lock normally excludes isolation
414 * splitting and collapsing (collapsing has already happened
415 * if PageLRU is set) but the lock is not necessarily taken
416 * here and it is wasteful to take it just to check transhuge.
417 * Check TransHuge without lock and skip the whole pageblock if
418 * it's either a transhuge or hugetlbfs page, as calling
419 * compound_order() without preventing THP from splitting the
420 * page underneath us may return surprising results.
422 if (PageTransHuge(page)) {
425 low_pfn += (1 << compound_order(page)) - 1;
429 /* Check if it is ok to still hold the lock */
430 locked = compact_checklock_irqsave(&zone->lru_lock, &flags,
432 if (!locked || fatal_signal_pending(current))
435 /* Recheck PageLRU and PageTransHuge under lock */
438 if (PageTransHuge(page)) {
439 low_pfn += (1 << compound_order(page)) - 1;
444 mode |= ISOLATE_ASYNC_MIGRATE;
446 lruvec = mem_cgroup_page_lruvec(page, zone);
448 /* Try isolate the page */
449 if (__isolate_lru_page(page, mode) != 0)
452 VM_BUG_ON(PageTransCompound(page));
454 /* Successfully isolated */
455 del_page_from_lru_list(page, lruvec, page_lru(page));
456 list_add(&page->lru, migratelist);
457 cc->nr_migratepages++;
460 /* Avoid isolating too much */
461 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
469 low_pfn += pageblock_nr_pages;
470 low_pfn = ALIGN(low_pfn, pageblock_nr_pages) - 1;
471 last_pageblock_nr = pageblock_nr;
474 acct_isolated(zone, locked, cc);
477 spin_unlock_irqrestore(&zone->lru_lock, flags);
479 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
484 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
485 #ifdef CONFIG_COMPACTION
487 /* Returns true if the page is within a block suitable for migration to */
488 static bool suitable_migration_target(struct page *page)
491 int migratetype = get_pageblock_migratetype(page);
493 /* Don't interfere with memory hot-remove or the min_free_kbytes blocks */
494 if (migratetype == MIGRATE_ISOLATE || migratetype == MIGRATE_RESERVE)
497 /* If the page is a large free page, then allow migration */
498 if (PageBuddy(page) && page_order(page) >= pageblock_order)
501 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
502 if (migrate_async_suitable(migratetype))
505 /* Otherwise skip the block */
510 * Returns the start pfn of the last page block in a zone. This is the starting
511 * point for full compaction of a zone. Compaction searches for free pages from
512 * the end of each zone, while isolate_freepages_block scans forward inside each
515 static unsigned long start_free_pfn(struct zone *zone)
517 unsigned long free_pfn;
518 free_pfn = zone->zone_start_pfn + zone->spanned_pages;
519 free_pfn &= ~(pageblock_nr_pages-1);
524 * Based on information in the current compact_control, find blocks
525 * suitable for isolating free pages from and then isolate them.
527 static void isolate_freepages(struct zone *zone,
528 struct compact_control *cc)
531 unsigned long high_pfn, low_pfn, pfn, zone_end_pfn, end_pfn;
533 int nr_freepages = cc->nr_freepages;
534 struct list_head *freelist = &cc->freepages;
537 * Initialise the free scanner. The starting point is where we last
538 * scanned from (or the end of the zone if starting). The low point
539 * is the end of the pageblock the migration scanner is using.
542 low_pfn = cc->migrate_pfn + pageblock_nr_pages;
545 * Take care that if the migration scanner is at the end of the zone
546 * that the free scanner does not accidentally move to the next zone
547 * in the next isolation cycle.
549 high_pfn = min(low_pfn, pfn);
551 zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
554 * Isolate free pages until enough are available to migrate the
555 * pages on cc->migratepages. We stop searching if the migrate
556 * and free page scanners meet or enough free pages are isolated.
558 for (; pfn > low_pfn && cc->nr_migratepages > nr_freepages;
559 pfn -= pageblock_nr_pages) {
560 unsigned long isolated;
566 * Check for overlapping nodes/zones. It's possible on some
567 * configurations to have a setup like
569 * i.e. it's possible that all pages within a zones range of
570 * pages do not belong to a single zone.
572 page = pfn_to_page(pfn);
573 if (page_zone(page) != zone)
576 /* Check the block is suitable for migration */
577 if (!suitable_migration_target(page))
581 * Found a block suitable for isolating free pages from. Now
582 * we disabled interrupts, double check things are ok and
583 * isolate the pages. This is to minimise the time IRQs
589 * The zone lock must be held to isolate freepages. This
590 * unfortunately this is a very coarse lock and can be
591 * heavily contended if there are parallel allocations
592 * or parallel compactions. For async compaction do not
595 if (!compact_trylock_irqsave(&zone->lock, &flags, cc))
597 if (suitable_migration_target(page)) {
598 end_pfn = min(pfn + pageblock_nr_pages, zone_end_pfn);
599 isolated = isolate_freepages_block(pfn, end_pfn,
601 nr_freepages += isolated;
603 spin_unlock_irqrestore(&zone->lock, flags);
606 * Record the highest PFN we isolated pages from. When next
607 * looking for free pages, the search will restart here as
608 * page migration may have returned some pages to the allocator
611 high_pfn = max(high_pfn, pfn);
614 * If the free scanner has wrapped, update
615 * compact_cached_free_pfn to point to the highest
616 * pageblock with free pages. This reduces excessive
617 * scanning of full pageblocks near the end of the
620 if (cc->order > 0 && cc->wrapped)
621 zone->compact_cached_free_pfn = high_pfn;
625 /* split_free_page does not map the pages */
628 cc->free_pfn = high_pfn;
629 cc->nr_freepages = nr_freepages;
631 /* If compact_cached_free_pfn is reset then set it now */
632 if (cc->order > 0 && !cc->wrapped &&
633 zone->compact_cached_free_pfn == start_free_pfn(zone))
634 zone->compact_cached_free_pfn = high_pfn;
638 * This is a migrate-callback that "allocates" freepages by taking pages
639 * from the isolated freelists in the block we are migrating to.
641 static struct page *compaction_alloc(struct page *migratepage,
645 struct compact_control *cc = (struct compact_control *)data;
646 struct page *freepage;
648 /* Isolate free pages if necessary */
649 if (list_empty(&cc->freepages)) {
650 isolate_freepages(cc->zone, cc);
652 if (list_empty(&cc->freepages))
656 freepage = list_entry(cc->freepages.next, struct page, lru);
657 list_del(&freepage->lru);
664 * We cannot control nr_migratepages and nr_freepages fully when migration is
665 * running as migrate_pages() has no knowledge of compact_control. When
666 * migration is complete, we count the number of pages on the lists by hand.
668 static void update_nr_listpages(struct compact_control *cc)
670 int nr_migratepages = 0;
671 int nr_freepages = 0;
674 list_for_each_entry(page, &cc->migratepages, lru)
676 list_for_each_entry(page, &cc->freepages, lru)
679 cc->nr_migratepages = nr_migratepages;
680 cc->nr_freepages = nr_freepages;
683 /* possible outcome of isolate_migratepages */
685 ISOLATE_ABORT, /* Abort compaction now */
686 ISOLATE_NONE, /* No pages isolated, continue scanning */
687 ISOLATE_SUCCESS, /* Pages isolated, migrate */
691 * Isolate all pages that can be migrated from the block pointed to by
692 * the migrate scanner within compact_control.
694 static isolate_migrate_t isolate_migratepages(struct zone *zone,
695 struct compact_control *cc)
697 unsigned long low_pfn, end_pfn;
699 /* Do not scan outside zone boundaries */
700 low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
702 /* Only scan within a pageblock boundary */
703 end_pfn = ALIGN(low_pfn + pageblock_nr_pages, pageblock_nr_pages);
705 /* Do not cross the free scanner or scan within a memory hole */
706 if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
707 cc->migrate_pfn = end_pfn;
711 /* Perform the isolation */
712 low_pfn = isolate_migratepages_range(zone, cc, low_pfn, end_pfn);
713 if (!low_pfn || cc->contended)
714 return ISOLATE_ABORT;
716 cc->migrate_pfn = low_pfn;
718 return ISOLATE_SUCCESS;
721 static int compact_finished(struct zone *zone,
722 struct compact_control *cc)
724 unsigned long watermark;
726 if (fatal_signal_pending(current))
727 return COMPACT_PARTIAL;
730 * A full (order == -1) compaction run starts at the beginning and
731 * end of a zone; it completes when the migrate and free scanner meet.
732 * A partial (order > 0) compaction can start with the free scanner
733 * at a random point in the zone, and may have to restart.
735 if (cc->free_pfn <= cc->migrate_pfn) {
736 if (cc->order > 0 && !cc->wrapped) {
737 /* We started partway through; restart at the end. */
738 unsigned long free_pfn = start_free_pfn(zone);
739 zone->compact_cached_free_pfn = free_pfn;
740 cc->free_pfn = free_pfn;
742 return COMPACT_CONTINUE;
744 return COMPACT_COMPLETE;
747 /* We wrapped around and ended up where we started. */
748 if (cc->wrapped && cc->free_pfn <= cc->start_free_pfn)
749 return COMPACT_COMPLETE;
752 * order == -1 is expected when compacting via
753 * /proc/sys/vm/compact_memory
756 return COMPACT_CONTINUE;
758 /* Compaction run is not finished if the watermark is not met */
759 watermark = low_wmark_pages(zone);
760 watermark += (1 << cc->order);
762 if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
763 return COMPACT_CONTINUE;
765 /* Direct compactor: Is a suitable page free? */
767 /* Was a suitable page captured? */
769 return COMPACT_PARTIAL;
772 for (order = cc->order; order < MAX_ORDER; order++) {
773 struct free_area *area = &zone->free_area[cc->order];
774 /* Job done if page is free of the right migratetype */
775 if (!list_empty(&area->free_list[cc->migratetype]))
776 return COMPACT_PARTIAL;
778 /* Job done if allocation would set block type */
779 if (cc->order >= pageblock_order && area->nr_free)
780 return COMPACT_PARTIAL;
784 return COMPACT_CONTINUE;
788 * compaction_suitable: Is this suitable to run compaction on this zone now?
790 * COMPACT_SKIPPED - If there are too few free pages for compaction
791 * COMPACT_PARTIAL - If the allocation would succeed without compaction
792 * COMPACT_CONTINUE - If compaction should run now
794 unsigned long compaction_suitable(struct zone *zone, int order)
797 unsigned long watermark;
800 * order == -1 is expected when compacting via
801 * /proc/sys/vm/compact_memory
804 return COMPACT_CONTINUE;
807 * Watermarks for order-0 must be met for compaction. Note the 2UL.
808 * This is because during migration, copies of pages need to be
809 * allocated and for a short time, the footprint is higher
811 watermark = low_wmark_pages(zone) + (2UL << order);
812 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
813 return COMPACT_SKIPPED;
816 * fragmentation index determines if allocation failures are due to
817 * low memory or external fragmentation
819 * index of -1000 implies allocations might succeed depending on
821 * index towards 0 implies failure is due to lack of memory
822 * index towards 1000 implies failure is due to fragmentation
824 * Only compact if a failure would be due to fragmentation.
826 fragindex = fragmentation_index(zone, order);
827 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
828 return COMPACT_SKIPPED;
830 if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
832 return COMPACT_PARTIAL;
834 return COMPACT_CONTINUE;
837 static int compact_zone(struct zone *zone, struct compact_control *cc)
841 ret = compaction_suitable(zone, cc->order);
843 case COMPACT_PARTIAL:
844 case COMPACT_SKIPPED:
845 /* Compaction is likely to fail */
847 case COMPACT_CONTINUE:
848 /* Fall through to compaction */
852 /* Setup to move all movable pages to the end of the zone */
853 cc->migrate_pfn = zone->zone_start_pfn;
856 /* Incremental compaction. Start where the last one stopped. */
857 cc->free_pfn = zone->compact_cached_free_pfn;
858 cc->start_free_pfn = cc->free_pfn;
860 /* Order == -1 starts at the end of the zone. */
861 cc->free_pfn = start_free_pfn(zone);
864 migrate_prep_local();
866 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
867 unsigned long nr_migrate, nr_remaining;
870 switch (isolate_migratepages(zone, cc)) {
872 ret = COMPACT_PARTIAL;
873 putback_lru_pages(&cc->migratepages);
874 cc->nr_migratepages = 0;
878 case ISOLATE_SUCCESS:
882 nr_migrate = cc->nr_migratepages;
883 err = migrate_pages(&cc->migratepages, compaction_alloc,
884 (unsigned long)cc, false,
885 cc->sync ? MIGRATE_SYNC_LIGHT : MIGRATE_ASYNC);
886 update_nr_listpages(cc);
887 nr_remaining = cc->nr_migratepages;
889 count_vm_event(COMPACTBLOCKS);
890 count_vm_events(COMPACTPAGES, nr_migrate - nr_remaining);
892 count_vm_events(COMPACTPAGEFAILED, nr_remaining);
893 trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
896 /* Release LRU pages not migrated */
898 putback_lru_pages(&cc->migratepages);
899 cc->nr_migratepages = 0;
900 if (err == -ENOMEM) {
901 ret = COMPACT_PARTIAL;
906 /* Capture a page now if it is a suitable size */
907 compact_capture_page(cc);
911 /* Release free pages and check accounting */
912 cc->nr_freepages -= release_freepages(&cc->freepages);
913 VM_BUG_ON(cc->nr_freepages != 0);
918 static unsigned long compact_zone_order(struct zone *zone,
919 int order, gfp_t gfp_mask,
920 bool sync, bool *contended,
924 struct compact_control cc = {
926 .nr_migratepages = 0,
928 .migratetype = allocflags_to_migratetype(gfp_mask),
933 INIT_LIST_HEAD(&cc.freepages);
934 INIT_LIST_HEAD(&cc.migratepages);
936 ret = compact_zone(zone, &cc);
938 VM_BUG_ON(!list_empty(&cc.freepages));
939 VM_BUG_ON(!list_empty(&cc.migratepages));
941 *contended = cc.contended;
945 int sysctl_extfrag_threshold = 500;
948 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
949 * @zonelist: The zonelist used for the current allocation
950 * @order: The order of the current allocation
951 * @gfp_mask: The GFP mask of the current allocation
952 * @nodemask: The allowed nodes to allocate from
953 * @sync: Whether migration is synchronous or not
954 * @contended: Return value that is true if compaction was aborted due to lock contention
955 * @page: Optionally capture a free page of the requested order during compaction
957 * This is the main entry point for direct page compaction.
959 unsigned long try_to_compact_pages(struct zonelist *zonelist,
960 int order, gfp_t gfp_mask, nodemask_t *nodemask,
961 bool sync, bool *contended, struct page **page)
963 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
964 int may_enter_fs = gfp_mask & __GFP_FS;
965 int may_perform_io = gfp_mask & __GFP_IO;
968 int rc = COMPACT_SKIPPED;
971 /* Check if the GFP flags allow compaction */
972 if (!order || !may_enter_fs || !may_perform_io)
975 count_vm_event(COMPACTSTALL);
978 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
979 alloc_flags |= ALLOC_CMA;
981 /* Compact each zone in the list */
982 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
986 status = compact_zone_order(zone, order, gfp_mask, sync,
988 rc = max(status, rc);
990 /* If a normal allocation would succeed, stop compacting */
991 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0,
1000 /* Compact all zones within a node */
1001 static int __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
1006 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1008 zone = &pgdat->node_zones[zoneid];
1009 if (!populated_zone(zone))
1012 cc->nr_freepages = 0;
1013 cc->nr_migratepages = 0;
1015 INIT_LIST_HEAD(&cc->freepages);
1016 INIT_LIST_HEAD(&cc->migratepages);
1018 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
1019 compact_zone(zone, cc);
1021 if (cc->order > 0) {
1022 int ok = zone_watermark_ok(zone, cc->order,
1023 low_wmark_pages(zone), 0, 0);
1024 if (ok && cc->order >= zone->compact_order_failed)
1025 zone->compact_order_failed = cc->order + 1;
1026 /* Currently async compaction is never deferred. */
1027 else if (!ok && cc->sync)
1028 defer_compaction(zone, cc->order);
1031 VM_BUG_ON(!list_empty(&cc->freepages));
1032 VM_BUG_ON(!list_empty(&cc->migratepages));
1038 int compact_pgdat(pg_data_t *pgdat, int order)
1040 struct compact_control cc = {
1046 return __compact_pgdat(pgdat, &cc);
1049 static int compact_node(int nid)
1051 struct compact_control cc = {
1057 return __compact_pgdat(NODE_DATA(nid), &cc);
1060 /* Compact all nodes in the system */
1061 static int compact_nodes(void)
1065 /* Flush pending updates to the LRU lists */
1066 lru_add_drain_all();
1068 for_each_online_node(nid)
1071 return COMPACT_COMPLETE;
1074 /* The written value is actually unused, all memory is compacted */
1075 int sysctl_compact_memory;
1077 /* This is the entry point for compacting all nodes via /proc/sys/vm */
1078 int sysctl_compaction_handler(struct ctl_table *table, int write,
1079 void __user *buffer, size_t *length, loff_t *ppos)
1082 return compact_nodes();
1087 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1088 void __user *buffer, size_t *length, loff_t *ppos)
1090 proc_dointvec_minmax(table, write, buffer, length, ppos);
1095 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1096 ssize_t sysfs_compact_node(struct device *dev,
1097 struct device_attribute *attr,
1098 const char *buf, size_t count)
1102 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1103 /* Flush pending updates to the LRU lists */
1104 lru_add_drain_all();
1111 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1113 int compaction_register_node(struct node *node)
1115 return device_create_file(&node->dev, &dev_attr_compact);
1118 void compaction_unregister_node(struct node *node)
1120 return device_remove_file(&node->dev, &dev_attr_compact);
1122 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1124 #endif /* CONFIG_COMPACTION */