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 /* Returns true if the page is within a block suitable for migration to */
97 static bool suitable_migration_target(struct page *page)
99 int migratetype = get_pageblock_migratetype(page);
101 /* Don't interfere with memory hot-remove or the min_free_kbytes blocks */
102 if (migratetype == MIGRATE_ISOLATE || migratetype == MIGRATE_RESERVE)
105 /* If the page is a large free page, then allow migration */
106 if (PageBuddy(page) && page_order(page) >= pageblock_order)
109 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
110 if (migrate_async_suitable(migratetype))
113 /* Otherwise skip the block */
117 static void compact_capture_page(struct compact_control *cc)
120 int mtype, mtype_low, mtype_high;
122 if (!cc->page || *cc->page)
126 * For MIGRATE_MOVABLE allocations we capture a suitable page ASAP
127 * regardless of the migratetype of the freelist is is captured from.
128 * This is fine because the order for a high-order MIGRATE_MOVABLE
129 * allocation is typically at least a pageblock size and overall
130 * fragmentation is not impaired. Other allocation types must
131 * capture pages from their own migratelist because otherwise they
132 * could pollute other pageblocks like MIGRATE_MOVABLE with
133 * difficult to move pages and making fragmentation worse overall.
135 if (cc->migratetype == MIGRATE_MOVABLE) {
137 mtype_high = MIGRATE_PCPTYPES;
139 mtype_low = cc->migratetype;
140 mtype_high = cc->migratetype + 1;
143 /* Speculatively examine the free lists without zone lock */
144 for (mtype = mtype_low; mtype < mtype_high; mtype++) {
146 for (order = cc->order; order < MAX_ORDER; order++) {
148 struct free_area *area;
149 area = &(cc->zone->free_area[order]);
150 if (list_empty(&area->free_list[mtype]))
153 /* Take the lock and attempt capture of the page */
154 if (!compact_trylock_irqsave(&cc->zone->lock, &flags, cc))
156 if (!list_empty(&area->free_list[mtype])) {
157 page = list_entry(area->free_list[mtype].next,
159 if (capture_free_page(page, cc->order, mtype)) {
160 spin_unlock_irqrestore(&cc->zone->lock,
166 spin_unlock_irqrestore(&cc->zone->lock, flags);
172 * Isolate free pages onto a private freelist. Caller must hold zone->lock.
173 * If @strict is true, will abort returning 0 on any invalid PFNs or non-free
174 * pages inside of the pageblock (even though it may still end up isolating
177 static unsigned long isolate_freepages_block(struct compact_control *cc,
178 unsigned long blockpfn,
179 unsigned long end_pfn,
180 struct list_head *freelist,
183 int nr_scanned = 0, total_isolated = 0;
185 unsigned long nr_strict_required = end_pfn - blockpfn;
189 cursor = pfn_to_page(blockpfn);
191 /* Isolate free pages. */
192 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
194 struct page *page = cursor;
197 if (!pfn_valid_within(blockpfn))
199 if (!PageBuddy(page))
203 * The zone lock must be held to isolate freepages.
204 * Unfortunately this is a very coarse lock and can be
205 * heavily contended if there are parallel allocations
206 * or parallel compactions. For async compaction do not
207 * spin on the lock and we acquire the lock as late as
210 locked = compact_checklock_irqsave(&cc->zone->lock, &flags,
215 /* Recheck this is a suitable migration target under lock */
216 if (!strict && !suitable_migration_target(page))
219 /* Recheck this is a buddy page under lock */
220 if (!PageBuddy(page))
223 /* Found a free page, break it into order-0 pages */
224 isolated = split_free_page(page);
225 if (!isolated && strict)
227 total_isolated += isolated;
228 for (i = 0; i < isolated; i++) {
229 list_add(&page->lru, freelist);
233 /* If a page was split, advance to the end of it */
235 blockpfn += isolated - 1;
236 cursor += isolated - 1;
240 trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
243 * If strict isolation is requested by CMA then check that all the
244 * pages requested were isolated. If there were any failures, 0 is
245 * returned and CMA will fail.
247 if (strict && nr_strict_required != total_isolated)
251 spin_unlock_irqrestore(&cc->zone->lock, flags);
253 return total_isolated;
257 * isolate_freepages_range() - isolate free pages.
258 * @start_pfn: The first PFN to start isolating.
259 * @end_pfn: The one-past-last PFN.
261 * Non-free pages, invalid PFNs, or zone boundaries within the
262 * [start_pfn, end_pfn) range are considered errors, cause function to
263 * undo its actions and return zero.
265 * Otherwise, function returns one-past-the-last PFN of isolated page
266 * (which may be greater then end_pfn if end fell in a middle of
270 isolate_freepages_range(unsigned long start_pfn, unsigned long end_pfn)
272 unsigned long isolated, pfn, block_end_pfn;
273 struct zone *zone = NULL;
276 /* cc needed for isolate_freepages_block to acquire zone->lock */
277 struct compact_control cc = {
281 if (pfn_valid(start_pfn))
282 cc.zone = zone = page_zone(pfn_to_page(start_pfn));
284 for (pfn = start_pfn; pfn < end_pfn; pfn += isolated) {
285 if (!pfn_valid(pfn) || zone != page_zone(pfn_to_page(pfn)))
289 * On subsequent iterations ALIGN() is actually not needed,
290 * but we keep it that we not to complicate the code.
292 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
293 block_end_pfn = min(block_end_pfn, end_pfn);
295 isolated = isolate_freepages_block(&cc, pfn, block_end_pfn,
299 * In strict mode, isolate_freepages_block() returns 0 if
300 * there are any holes in the block (ie. invalid PFNs or
307 * If we managed to isolate pages, it is always (1 << n) *
308 * pageblock_nr_pages for some non-negative n. (Max order
309 * page may span two pageblocks).
313 /* split_free_page does not map the pages */
314 map_pages(&freelist);
317 /* Loop terminated early, cleanup. */
318 release_freepages(&freelist);
322 /* We don't use freelists for anything. */
326 /* Update the number of anon and file isolated pages in the zone */
327 static void acct_isolated(struct zone *zone, bool locked, struct compact_control *cc)
330 unsigned int count[2] = { 0, };
332 list_for_each_entry(page, &cc->migratepages, lru)
333 count[!!page_is_file_cache(page)]++;
335 /* If locked we can use the interrupt unsafe versions */
337 __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
338 __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
340 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
341 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
345 /* Similar to reclaim, but different enough that they don't share logic */
346 static bool too_many_isolated(struct zone *zone)
348 unsigned long active, inactive, isolated;
350 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
351 zone_page_state(zone, NR_INACTIVE_ANON);
352 active = zone_page_state(zone, NR_ACTIVE_FILE) +
353 zone_page_state(zone, NR_ACTIVE_ANON);
354 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
355 zone_page_state(zone, NR_ISOLATED_ANON);
357 return isolated > (inactive + active) / 2;
361 * isolate_migratepages_range() - isolate all migrate-able pages in range.
362 * @zone: Zone pages are in.
363 * @cc: Compaction control structure.
364 * @low_pfn: The first PFN of the range.
365 * @end_pfn: The one-past-the-last PFN of the range.
367 * Isolate all pages that can be migrated from the range specified by
368 * [low_pfn, end_pfn). Returns zero if there is a fatal signal
369 * pending), otherwise PFN of the first page that was not scanned
370 * (which may be both less, equal to or more then end_pfn).
372 * Assumes that cc->migratepages is empty and cc->nr_migratepages is
375 * Apart from cc->migratepages and cc->nr_migratetypes this function
376 * does not modify any cc's fields, in particular it does not modify
377 * (or read for that matter) cc->migrate_pfn.
380 isolate_migratepages_range(struct zone *zone, struct compact_control *cc,
381 unsigned long low_pfn, unsigned long end_pfn)
383 unsigned long last_pageblock_nr = 0, pageblock_nr;
384 unsigned long nr_scanned = 0, nr_isolated = 0;
385 struct list_head *migratelist = &cc->migratepages;
386 isolate_mode_t mode = 0;
387 struct lruvec *lruvec;
392 * Ensure that there are not too many pages isolated from the LRU
393 * list by either parallel reclaimers or compaction. If there are,
394 * delay for some time until fewer pages are isolated
396 while (unlikely(too_many_isolated(zone))) {
397 /* async migration should just abort */
401 congestion_wait(BLK_RW_ASYNC, HZ/10);
403 if (fatal_signal_pending(current))
407 /* Time to isolate some pages for migration */
409 for (; low_pfn < end_pfn; low_pfn++) {
412 /* give a chance to irqs before checking need_resched() */
413 if (locked && !((low_pfn+1) % SWAP_CLUSTER_MAX)) {
414 if (should_release_lock(&zone->lru_lock)) {
415 spin_unlock_irqrestore(&zone->lru_lock, flags);
421 * migrate_pfn does not necessarily start aligned to a
422 * pageblock. Ensure that pfn_valid is called when moving
423 * into a new MAX_ORDER_NR_PAGES range in case of large
424 * memory holes within the zone
426 if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
427 if (!pfn_valid(low_pfn)) {
428 low_pfn += MAX_ORDER_NR_PAGES - 1;
433 if (!pfn_valid_within(low_pfn))
438 * Get the page and ensure the page is within the same zone.
439 * See the comment in isolate_freepages about overlapping
440 * nodes. It is deliberate that the new zone lock is not taken
441 * as memory compaction should not move pages between nodes.
443 page = pfn_to_page(low_pfn);
444 if (page_zone(page) != zone)
452 * For async migration, also only scan in MOVABLE blocks. Async
453 * migration is optimistic to see if the minimum amount of work
454 * satisfies the allocation
456 pageblock_nr = low_pfn >> pageblock_order;
457 if (!cc->sync && last_pageblock_nr != pageblock_nr &&
458 !migrate_async_suitable(get_pageblock_migratetype(page))) {
462 /* Check may be lockless but that's ok as we recheck later */
467 * PageLRU is set. lru_lock normally excludes isolation
468 * splitting and collapsing (collapsing has already happened
469 * if PageLRU is set) but the lock is not necessarily taken
470 * here and it is wasteful to take it just to check transhuge.
471 * Check TransHuge without lock and skip the whole pageblock if
472 * it's either a transhuge or hugetlbfs page, as calling
473 * compound_order() without preventing THP from splitting the
474 * page underneath us may return surprising results.
476 if (PageTransHuge(page)) {
479 low_pfn += (1 << compound_order(page)) - 1;
483 /* Check if it is ok to still hold the lock */
484 locked = compact_checklock_irqsave(&zone->lru_lock, &flags,
486 if (!locked || fatal_signal_pending(current))
489 /* Recheck PageLRU and PageTransHuge under lock */
492 if (PageTransHuge(page)) {
493 low_pfn += (1 << compound_order(page)) - 1;
498 mode |= ISOLATE_ASYNC_MIGRATE;
500 lruvec = mem_cgroup_page_lruvec(page, zone);
502 /* Try isolate the page */
503 if (__isolate_lru_page(page, mode) != 0)
506 VM_BUG_ON(PageTransCompound(page));
508 /* Successfully isolated */
509 del_page_from_lru_list(page, lruvec, page_lru(page));
510 list_add(&page->lru, migratelist);
511 cc->nr_migratepages++;
514 /* Avoid isolating too much */
515 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
523 low_pfn += pageblock_nr_pages;
524 low_pfn = ALIGN(low_pfn, pageblock_nr_pages) - 1;
525 last_pageblock_nr = pageblock_nr;
528 acct_isolated(zone, locked, cc);
531 spin_unlock_irqrestore(&zone->lru_lock, flags);
533 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
538 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
539 #ifdef CONFIG_COMPACTION
541 * Based on information in the current compact_control, find blocks
542 * suitable for isolating free pages from and then isolate them.
544 static void isolate_freepages(struct zone *zone,
545 struct compact_control *cc)
548 unsigned long high_pfn, low_pfn, pfn, zone_end_pfn, end_pfn;
549 int nr_freepages = cc->nr_freepages;
550 struct list_head *freelist = &cc->freepages;
553 * Initialise the free scanner. The starting point is where we last
554 * scanned from (or the end of the zone if starting). The low point
555 * is the end of the pageblock the migration scanner is using.
558 low_pfn = cc->migrate_pfn + pageblock_nr_pages;
561 * Take care that if the migration scanner is at the end of the zone
562 * that the free scanner does not accidentally move to the next zone
563 * in the next isolation cycle.
565 high_pfn = min(low_pfn, pfn);
567 zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
570 * Isolate free pages until enough are available to migrate the
571 * pages on cc->migratepages. We stop searching if the migrate
572 * and free page scanners meet or enough free pages are isolated.
574 for (; pfn > low_pfn && cc->nr_migratepages > nr_freepages;
575 pfn -= pageblock_nr_pages) {
576 unsigned long isolated;
582 * Check for overlapping nodes/zones. It's possible on some
583 * configurations to have a setup like
585 * i.e. it's possible that all pages within a zones range of
586 * pages do not belong to a single zone.
588 page = pfn_to_page(pfn);
589 if (page_zone(page) != zone)
592 /* Check the block is suitable for migration */
593 if (!suitable_migration_target(page))
596 /* Found a block suitable for isolating free pages from */
598 end_pfn = min(pfn + pageblock_nr_pages, zone_end_pfn);
599 isolated = isolate_freepages_block(cc, pfn, end_pfn,
601 nr_freepages += isolated;
604 * Record the highest PFN we isolated pages from. When next
605 * looking for free pages, the search will restart here as
606 * page migration may have returned some pages to the allocator
609 high_pfn = max(high_pfn, pfn);
612 /* split_free_page does not map the pages */
615 cc->free_pfn = high_pfn;
616 cc->nr_freepages = nr_freepages;
620 * This is a migrate-callback that "allocates" freepages by taking pages
621 * from the isolated freelists in the block we are migrating to.
623 static struct page *compaction_alloc(struct page *migratepage,
627 struct compact_control *cc = (struct compact_control *)data;
628 struct page *freepage;
630 /* Isolate free pages if necessary */
631 if (list_empty(&cc->freepages)) {
632 isolate_freepages(cc->zone, cc);
634 if (list_empty(&cc->freepages))
638 freepage = list_entry(cc->freepages.next, struct page, lru);
639 list_del(&freepage->lru);
646 * We cannot control nr_migratepages and nr_freepages fully when migration is
647 * running as migrate_pages() has no knowledge of compact_control. When
648 * migration is complete, we count the number of pages on the lists by hand.
650 static void update_nr_listpages(struct compact_control *cc)
652 int nr_migratepages = 0;
653 int nr_freepages = 0;
656 list_for_each_entry(page, &cc->migratepages, lru)
658 list_for_each_entry(page, &cc->freepages, lru)
661 cc->nr_migratepages = nr_migratepages;
662 cc->nr_freepages = nr_freepages;
665 /* possible outcome of isolate_migratepages */
667 ISOLATE_ABORT, /* Abort compaction now */
668 ISOLATE_NONE, /* No pages isolated, continue scanning */
669 ISOLATE_SUCCESS, /* Pages isolated, migrate */
673 * Isolate all pages that can be migrated from the block pointed to by
674 * the migrate scanner within compact_control.
676 static isolate_migrate_t isolate_migratepages(struct zone *zone,
677 struct compact_control *cc)
679 unsigned long low_pfn, end_pfn;
681 /* Do not scan outside zone boundaries */
682 low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
684 /* Only scan within a pageblock boundary */
685 end_pfn = ALIGN(low_pfn + pageblock_nr_pages, pageblock_nr_pages);
687 /* Do not cross the free scanner or scan within a memory hole */
688 if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
689 cc->migrate_pfn = end_pfn;
693 /* Perform the isolation */
694 low_pfn = isolate_migratepages_range(zone, cc, low_pfn, end_pfn);
695 if (!low_pfn || cc->contended)
696 return ISOLATE_ABORT;
698 cc->migrate_pfn = low_pfn;
700 return ISOLATE_SUCCESS;
703 static int compact_finished(struct zone *zone,
704 struct compact_control *cc)
706 unsigned long watermark;
708 if (fatal_signal_pending(current))
709 return COMPACT_PARTIAL;
711 /* Compaction run completes if the migrate and free scanner meet */
712 if (cc->free_pfn <= cc->migrate_pfn)
713 return COMPACT_COMPLETE;
716 * order == -1 is expected when compacting via
717 * /proc/sys/vm/compact_memory
720 return COMPACT_CONTINUE;
722 /* Compaction run is not finished if the watermark is not met */
723 watermark = low_wmark_pages(zone);
724 watermark += (1 << cc->order);
726 if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
727 return COMPACT_CONTINUE;
729 /* Direct compactor: Is a suitable page free? */
731 /* Was a suitable page captured? */
733 return COMPACT_PARTIAL;
736 for (order = cc->order; order < MAX_ORDER; order++) {
737 struct free_area *area = &zone->free_area[cc->order];
738 /* Job done if page is free of the right migratetype */
739 if (!list_empty(&area->free_list[cc->migratetype]))
740 return COMPACT_PARTIAL;
742 /* Job done if allocation would set block type */
743 if (cc->order >= pageblock_order && area->nr_free)
744 return COMPACT_PARTIAL;
748 return COMPACT_CONTINUE;
752 * compaction_suitable: Is this suitable to run compaction on this zone now?
754 * COMPACT_SKIPPED - If there are too few free pages for compaction
755 * COMPACT_PARTIAL - If the allocation would succeed without compaction
756 * COMPACT_CONTINUE - If compaction should run now
758 unsigned long compaction_suitable(struct zone *zone, int order)
761 unsigned long watermark;
764 * order == -1 is expected when compacting via
765 * /proc/sys/vm/compact_memory
768 return COMPACT_CONTINUE;
771 * Watermarks for order-0 must be met for compaction. Note the 2UL.
772 * This is because during migration, copies of pages need to be
773 * allocated and for a short time, the footprint is higher
775 watermark = low_wmark_pages(zone) + (2UL << order);
776 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
777 return COMPACT_SKIPPED;
780 * fragmentation index determines if allocation failures are due to
781 * low memory or external fragmentation
783 * index of -1000 implies allocations might succeed depending on
785 * index towards 0 implies failure is due to lack of memory
786 * index towards 1000 implies failure is due to fragmentation
788 * Only compact if a failure would be due to fragmentation.
790 fragindex = fragmentation_index(zone, order);
791 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
792 return COMPACT_SKIPPED;
794 if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
796 return COMPACT_PARTIAL;
798 return COMPACT_CONTINUE;
801 static int compact_zone(struct zone *zone, struct compact_control *cc)
805 ret = compaction_suitable(zone, cc->order);
807 case COMPACT_PARTIAL:
808 case COMPACT_SKIPPED:
809 /* Compaction is likely to fail */
811 case COMPACT_CONTINUE:
812 /* Fall through to compaction */
816 /* Setup to move all movable pages to the end of the zone */
817 cc->migrate_pfn = zone->zone_start_pfn;
818 cc->free_pfn = cc->migrate_pfn + zone->spanned_pages;
819 cc->free_pfn &= ~(pageblock_nr_pages-1);
821 migrate_prep_local();
823 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
824 unsigned long nr_migrate, nr_remaining;
827 switch (isolate_migratepages(zone, cc)) {
829 ret = COMPACT_PARTIAL;
830 putback_lru_pages(&cc->migratepages);
831 cc->nr_migratepages = 0;
835 case ISOLATE_SUCCESS:
839 nr_migrate = cc->nr_migratepages;
840 err = migrate_pages(&cc->migratepages, compaction_alloc,
841 (unsigned long)cc, false,
842 cc->sync ? MIGRATE_SYNC_LIGHT : MIGRATE_ASYNC);
843 update_nr_listpages(cc);
844 nr_remaining = cc->nr_migratepages;
846 count_vm_event(COMPACTBLOCKS);
847 count_vm_events(COMPACTPAGES, nr_migrate - nr_remaining);
849 count_vm_events(COMPACTPAGEFAILED, nr_remaining);
850 trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
853 /* Release LRU pages not migrated */
855 putback_lru_pages(&cc->migratepages);
856 cc->nr_migratepages = 0;
857 if (err == -ENOMEM) {
858 ret = COMPACT_PARTIAL;
863 /* Capture a page now if it is a suitable size */
864 compact_capture_page(cc);
868 /* Release free pages and check accounting */
869 cc->nr_freepages -= release_freepages(&cc->freepages);
870 VM_BUG_ON(cc->nr_freepages != 0);
875 static unsigned long compact_zone_order(struct zone *zone,
876 int order, gfp_t gfp_mask,
877 bool sync, bool *contended,
881 struct compact_control cc = {
883 .nr_migratepages = 0,
885 .migratetype = allocflags_to_migratetype(gfp_mask),
890 INIT_LIST_HEAD(&cc.freepages);
891 INIT_LIST_HEAD(&cc.migratepages);
893 ret = compact_zone(zone, &cc);
895 VM_BUG_ON(!list_empty(&cc.freepages));
896 VM_BUG_ON(!list_empty(&cc.migratepages));
898 *contended = cc.contended;
902 int sysctl_extfrag_threshold = 500;
905 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
906 * @zonelist: The zonelist used for the current allocation
907 * @order: The order of the current allocation
908 * @gfp_mask: The GFP mask of the current allocation
909 * @nodemask: The allowed nodes to allocate from
910 * @sync: Whether migration is synchronous or not
911 * @contended: Return value that is true if compaction was aborted due to lock contention
912 * @page: Optionally capture a free page of the requested order during compaction
914 * This is the main entry point for direct page compaction.
916 unsigned long try_to_compact_pages(struct zonelist *zonelist,
917 int order, gfp_t gfp_mask, nodemask_t *nodemask,
918 bool sync, bool *contended, struct page **page)
920 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
921 int may_enter_fs = gfp_mask & __GFP_FS;
922 int may_perform_io = gfp_mask & __GFP_IO;
925 int rc = COMPACT_SKIPPED;
928 /* Check if the GFP flags allow compaction */
929 if (!order || !may_enter_fs || !may_perform_io)
932 count_vm_event(COMPACTSTALL);
935 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
936 alloc_flags |= ALLOC_CMA;
938 /* Compact each zone in the list */
939 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
943 status = compact_zone_order(zone, order, gfp_mask, sync,
945 rc = max(status, rc);
947 /* If a normal allocation would succeed, stop compacting */
948 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0,
957 /* Compact all zones within a node */
958 static int __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
963 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
965 zone = &pgdat->node_zones[zoneid];
966 if (!populated_zone(zone))
969 cc->nr_freepages = 0;
970 cc->nr_migratepages = 0;
972 INIT_LIST_HEAD(&cc->freepages);
973 INIT_LIST_HEAD(&cc->migratepages);
975 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
976 compact_zone(zone, cc);
979 int ok = zone_watermark_ok(zone, cc->order,
980 low_wmark_pages(zone), 0, 0);
981 if (ok && cc->order >= zone->compact_order_failed)
982 zone->compact_order_failed = cc->order + 1;
983 /* Currently async compaction is never deferred. */
984 else if (!ok && cc->sync)
985 defer_compaction(zone, cc->order);
988 VM_BUG_ON(!list_empty(&cc->freepages));
989 VM_BUG_ON(!list_empty(&cc->migratepages));
995 int compact_pgdat(pg_data_t *pgdat, int order)
997 struct compact_control cc = {
1003 return __compact_pgdat(pgdat, &cc);
1006 static int compact_node(int nid)
1008 struct compact_control cc = {
1014 return __compact_pgdat(NODE_DATA(nid), &cc);
1017 /* Compact all nodes in the system */
1018 static int compact_nodes(void)
1022 /* Flush pending updates to the LRU lists */
1023 lru_add_drain_all();
1025 for_each_online_node(nid)
1028 return COMPACT_COMPLETE;
1031 /* The written value is actually unused, all memory is compacted */
1032 int sysctl_compact_memory;
1034 /* This is the entry point for compacting all nodes via /proc/sys/vm */
1035 int sysctl_compaction_handler(struct ctl_table *table, int write,
1036 void __user *buffer, size_t *length, loff_t *ppos)
1039 return compact_nodes();
1044 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1045 void __user *buffer, size_t *length, loff_t *ppos)
1047 proc_dointvec_minmax(table, write, buffer, length, ppos);
1052 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1053 ssize_t sysfs_compact_node(struct device *dev,
1054 struct device_attribute *attr,
1055 const char *buf, size_t count)
1059 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1060 /* Flush pending updates to the LRU lists */
1061 lru_add_drain_all();
1068 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1070 int compaction_register_node(struct node *node)
1072 return device_create_file(&node->dev, &dev_attr_compact);
1075 void compaction_unregister_node(struct node *node)
1077 return device_remove_file(&node->dev, &dev_attr_compact);
1079 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1081 #endif /* CONFIG_COMPACTION */