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>
18 #include <linux/page-isolation.h>
19 #include <linux/kasan.h>
22 #ifdef CONFIG_COMPACTION
23 static inline void count_compact_event(enum vm_event_item item)
28 static inline void count_compact_events(enum vm_event_item item, long delta)
30 count_vm_events(item, delta);
33 #define count_compact_event(item) do { } while (0)
34 #define count_compact_events(item, delta) do { } while (0)
37 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
39 #define CREATE_TRACE_POINTS
40 #include <trace/events/compaction.h>
42 static unsigned long release_freepages(struct list_head *freelist)
44 struct page *page, *next;
45 unsigned long high_pfn = 0;
47 list_for_each_entry_safe(page, next, freelist, lru) {
48 unsigned long pfn = page_to_pfn(page);
58 static void map_pages(struct list_head *list)
62 list_for_each_entry(page, list, lru) {
63 arch_alloc_page(page, 0);
64 kernel_map_pages(page, 1, 1);
65 kasan_alloc_pages(page, 0);
69 static inline bool migrate_async_suitable(int migratetype)
71 return is_migrate_cma(migratetype) || migratetype == MIGRATE_MOVABLE;
75 * Check that the whole (or subset of) a pageblock given by the interval of
76 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
77 * with the migration of free compaction scanner. The scanners then need to
78 * use only pfn_valid_within() check for arches that allow holes within
81 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
83 * It's possible on some configurations to have a setup like node0 node1 node0
84 * i.e. it's possible that all pages within a zones range of pages do not
85 * belong to a single zone. We assume that a border between node0 and node1
86 * can occur within a single pageblock, but not a node0 node1 node0
87 * interleaving within a single pageblock. It is therefore sufficient to check
88 * the first and last page of a pageblock and avoid checking each individual
89 * page in a pageblock.
91 static struct page *pageblock_pfn_to_page(unsigned long start_pfn,
92 unsigned long end_pfn, struct zone *zone)
94 struct page *start_page;
95 struct page *end_page;
97 /* end_pfn is one past the range we are checking */
100 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
103 start_page = pfn_to_page(start_pfn);
105 if (page_zone(start_page) != zone)
108 end_page = pfn_to_page(end_pfn);
110 /* This gives a shorter code than deriving page_zone(end_page) */
111 if (page_zone_id(start_page) != page_zone_id(end_page))
117 #ifdef CONFIG_COMPACTION
119 /* Do not skip compaction more than 64 times */
120 #define COMPACT_MAX_DEFER_SHIFT 6
123 * Compaction is deferred when compaction fails to result in a page
124 * allocation success. 1 << compact_defer_limit compactions are skipped up
125 * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
127 void defer_compaction(struct zone *zone, int order)
129 zone->compact_considered = 0;
130 zone->compact_defer_shift++;
132 if (order < zone->compact_order_failed)
133 zone->compact_order_failed = order;
135 if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
136 zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
138 trace_mm_compaction_defer_compaction(zone, order);
141 /* Returns true if compaction should be skipped this time */
142 bool compaction_deferred(struct zone *zone, int order)
144 unsigned long defer_limit = 1UL << zone->compact_defer_shift;
146 if (order < zone->compact_order_failed)
149 /* Avoid possible overflow */
150 if (++zone->compact_considered > defer_limit)
151 zone->compact_considered = defer_limit;
153 if (zone->compact_considered >= defer_limit)
156 trace_mm_compaction_deferred(zone, order);
162 * Update defer tracking counters after successful compaction of given order,
163 * which means an allocation either succeeded (alloc_success == true) or is
164 * expected to succeed.
166 void compaction_defer_reset(struct zone *zone, int order,
170 zone->compact_considered = 0;
171 zone->compact_defer_shift = 0;
173 if (order >= zone->compact_order_failed)
174 zone->compact_order_failed = order + 1;
176 trace_mm_compaction_defer_reset(zone, order);
179 /* Returns true if restarting compaction after many failures */
180 bool compaction_restarting(struct zone *zone, int order)
182 if (order < zone->compact_order_failed)
185 return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
186 zone->compact_considered >= 1UL << zone->compact_defer_shift;
189 /* Returns true if the pageblock should be scanned for pages to isolate. */
190 static inline bool isolation_suitable(struct compact_control *cc,
193 if (cc->ignore_skip_hint)
196 return !get_pageblock_skip(page);
199 static void reset_cached_positions(struct zone *zone)
201 zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
202 zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
203 zone->compact_cached_free_pfn = zone_end_pfn(zone);
207 * This function is called to clear all cached information on pageblocks that
208 * should be skipped for page isolation when the migrate and free page scanner
211 static void __reset_isolation_suitable(struct zone *zone)
213 unsigned long start_pfn = zone->zone_start_pfn;
214 unsigned long end_pfn = zone_end_pfn(zone);
217 zone->compact_blockskip_flush = false;
219 /* Walk the zone and mark every pageblock as suitable for isolation */
220 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
228 page = pfn_to_page(pfn);
229 if (zone != page_zone(page))
232 clear_pageblock_skip(page);
235 reset_cached_positions(zone);
238 void reset_isolation_suitable(pg_data_t *pgdat)
242 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
243 struct zone *zone = &pgdat->node_zones[zoneid];
244 if (!populated_zone(zone))
247 /* Only flush if a full compaction finished recently */
248 if (zone->compact_blockskip_flush)
249 __reset_isolation_suitable(zone);
254 * If no pages were isolated then mark this pageblock to be skipped in the
255 * future. The information is later cleared by __reset_isolation_suitable().
257 static void update_pageblock_skip(struct compact_control *cc,
258 struct page *page, unsigned long nr_isolated,
259 bool migrate_scanner)
261 struct zone *zone = cc->zone;
264 if (cc->ignore_skip_hint)
273 set_pageblock_skip(page);
275 pfn = page_to_pfn(page);
277 /* Update where async and sync compaction should restart */
278 if (migrate_scanner) {
279 if (pfn > zone->compact_cached_migrate_pfn[0])
280 zone->compact_cached_migrate_pfn[0] = pfn;
281 if (cc->mode != MIGRATE_ASYNC &&
282 pfn > zone->compact_cached_migrate_pfn[1])
283 zone->compact_cached_migrate_pfn[1] = pfn;
285 if (pfn < zone->compact_cached_free_pfn)
286 zone->compact_cached_free_pfn = pfn;
290 static inline bool isolation_suitable(struct compact_control *cc,
296 static void update_pageblock_skip(struct compact_control *cc,
297 struct page *page, unsigned long nr_isolated,
298 bool migrate_scanner)
301 #endif /* CONFIG_COMPACTION */
304 * Compaction requires the taking of some coarse locks that are potentially
305 * very heavily contended. For async compaction, back out if the lock cannot
306 * be taken immediately. For sync compaction, spin on the lock if needed.
308 * Returns true if the lock is held
309 * Returns false if the lock is not held and compaction should abort
311 static bool compact_trylock_irqsave(spinlock_t *lock, unsigned long *flags,
312 struct compact_control *cc)
314 if (cc->mode == MIGRATE_ASYNC) {
315 if (!spin_trylock_irqsave(lock, *flags)) {
316 cc->contended = COMPACT_CONTENDED_LOCK;
320 spin_lock_irqsave(lock, *flags);
327 * Compaction requires the taking of some coarse locks that are potentially
328 * very heavily contended. The lock should be periodically unlocked to avoid
329 * having disabled IRQs for a long time, even when there is nobody waiting on
330 * the lock. It might also be that allowing the IRQs will result in
331 * need_resched() becoming true. If scheduling is needed, async compaction
332 * aborts. Sync compaction schedules.
333 * Either compaction type will also abort if a fatal signal is pending.
334 * In either case if the lock was locked, it is dropped and not regained.
336 * Returns true if compaction should abort due to fatal signal pending, or
337 * async compaction due to need_resched()
338 * Returns false when compaction can continue (sync compaction might have
341 static bool compact_unlock_should_abort(spinlock_t *lock,
342 unsigned long flags, bool *locked, struct compact_control *cc)
345 spin_unlock_irqrestore(lock, flags);
349 if (fatal_signal_pending(current)) {
350 cc->contended = COMPACT_CONTENDED_SCHED;
354 if (need_resched()) {
355 if (cc->mode == MIGRATE_ASYNC) {
356 cc->contended = COMPACT_CONTENDED_SCHED;
366 * Aside from avoiding lock contention, compaction also periodically checks
367 * need_resched() and either schedules in sync compaction or aborts async
368 * compaction. This is similar to what compact_unlock_should_abort() does, but
369 * is used where no lock is concerned.
371 * Returns false when no scheduling was needed, or sync compaction scheduled.
372 * Returns true when async compaction should abort.
374 static inline bool compact_should_abort(struct compact_control *cc)
376 /* async compaction aborts if contended */
377 if (need_resched()) {
378 if (cc->mode == MIGRATE_ASYNC) {
379 cc->contended = COMPACT_CONTENDED_SCHED;
390 * Isolate free pages onto a private freelist. If @strict is true, will abort
391 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
392 * (even though it may still end up isolating some pages).
394 static unsigned long isolate_freepages_block(struct compact_control *cc,
395 unsigned long *start_pfn,
396 unsigned long end_pfn,
397 struct list_head *freelist,
400 int nr_scanned = 0, total_isolated = 0;
401 struct page *cursor, *valid_page = NULL;
402 unsigned long flags = 0;
404 unsigned long blockpfn = *start_pfn;
406 cursor = pfn_to_page(blockpfn);
408 /* Isolate free pages. */
409 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
411 struct page *page = cursor;
414 * Periodically drop the lock (if held) regardless of its
415 * contention, to give chance to IRQs. Abort if fatal signal
416 * pending or async compaction detects need_resched()
418 if (!(blockpfn % SWAP_CLUSTER_MAX)
419 && compact_unlock_should_abort(&cc->zone->lock, flags,
424 if (!pfn_valid_within(blockpfn))
431 * For compound pages such as THP and hugetlbfs, we can save
432 * potentially a lot of iterations if we skip them at once.
433 * The check is racy, but we can consider only valid values
434 * and the only danger is skipping too much.
436 if (PageCompound(page)) {
437 unsigned int comp_order = compound_order(page);
439 if (likely(comp_order < MAX_ORDER)) {
440 blockpfn += (1UL << comp_order) - 1;
441 cursor += (1UL << comp_order) - 1;
447 if (!PageBuddy(page))
451 * If we already hold the lock, we can skip some rechecking.
452 * Note that if we hold the lock now, checked_pageblock was
453 * already set in some previous iteration (or strict is true),
454 * so it is correct to skip the suitable migration target
459 * The zone lock must be held to isolate freepages.
460 * Unfortunately this is a very coarse lock and can be
461 * heavily contended if there are parallel allocations
462 * or parallel compactions. For async compaction do not
463 * spin on the lock and we acquire the lock as late as
466 locked = compact_trylock_irqsave(&cc->zone->lock,
471 /* Recheck this is a buddy page under lock */
472 if (!PageBuddy(page))
476 /* Found a free page, break it into order-0 pages */
477 isolated = split_free_page(page);
478 total_isolated += isolated;
479 for (i = 0; i < isolated; i++) {
480 list_add(&page->lru, freelist);
484 /* If a page was split, advance to the end of it */
486 cc->nr_freepages += isolated;
488 cc->nr_migratepages <= cc->nr_freepages) {
489 blockpfn += isolated;
493 blockpfn += isolated - 1;
494 cursor += isolated - 1;
507 * There is a tiny chance that we have read bogus compound_order(),
508 * so be careful to not go outside of the pageblock.
510 if (unlikely(blockpfn > end_pfn))
513 trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
514 nr_scanned, total_isolated);
516 /* Record how far we have got within the block */
517 *start_pfn = blockpfn;
520 * If strict isolation is requested by CMA then check that all the
521 * pages requested were isolated. If there were any failures, 0 is
522 * returned and CMA will fail.
524 if (strict && blockpfn < end_pfn)
528 spin_unlock_irqrestore(&cc->zone->lock, flags);
530 /* Update the pageblock-skip if the whole pageblock was scanned */
531 if (blockpfn == end_pfn)
532 update_pageblock_skip(cc, valid_page, total_isolated, false);
534 count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
536 count_compact_events(COMPACTISOLATED, total_isolated);
537 return total_isolated;
541 * isolate_freepages_range() - isolate free pages.
542 * @start_pfn: The first PFN to start isolating.
543 * @end_pfn: The one-past-last PFN.
545 * Non-free pages, invalid PFNs, or zone boundaries within the
546 * [start_pfn, end_pfn) range are considered errors, cause function to
547 * undo its actions and return zero.
549 * Otherwise, function returns one-past-the-last PFN of isolated page
550 * (which may be greater then end_pfn if end fell in a middle of
554 isolate_freepages_range(struct compact_control *cc,
555 unsigned long start_pfn, unsigned long end_pfn)
557 unsigned long isolated, pfn, block_end_pfn;
561 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
563 for (; pfn < end_pfn; pfn += isolated,
564 block_end_pfn += pageblock_nr_pages) {
565 /* Protect pfn from changing by isolate_freepages_block */
566 unsigned long isolate_start_pfn = pfn;
568 block_end_pfn = min(block_end_pfn, end_pfn);
571 * pfn could pass the block_end_pfn if isolated freepage
572 * is more than pageblock order. In this case, we adjust
573 * scanning range to right one.
575 if (pfn >= block_end_pfn) {
576 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
577 block_end_pfn = min(block_end_pfn, end_pfn);
580 if (!pageblock_pfn_to_page(pfn, block_end_pfn, cc->zone))
583 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
584 block_end_pfn, &freelist, true);
587 * In strict mode, isolate_freepages_block() returns 0 if
588 * there are any holes in the block (ie. invalid PFNs or
595 * If we managed to isolate pages, it is always (1 << n) *
596 * pageblock_nr_pages for some non-negative n. (Max order
597 * page may span two pageblocks).
601 /* split_free_page does not map the pages */
602 map_pages(&freelist);
605 /* Loop terminated early, cleanup. */
606 release_freepages(&freelist);
610 /* We don't use freelists for anything. */
614 /* Update the number of anon and file isolated pages in the zone */
615 static void acct_isolated(struct zone *zone, struct compact_control *cc)
618 unsigned int count[2] = { 0, };
620 if (list_empty(&cc->migratepages))
623 list_for_each_entry(page, &cc->migratepages, lru)
624 count[!!page_is_file_cache(page)]++;
626 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
627 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
630 /* Similar to reclaim, but different enough that they don't share logic */
631 static bool too_many_isolated(struct zone *zone)
633 unsigned long active, inactive, isolated;
635 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
636 zone_page_state(zone, NR_INACTIVE_ANON);
637 active = zone_page_state(zone, NR_ACTIVE_FILE) +
638 zone_page_state(zone, NR_ACTIVE_ANON);
639 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
640 zone_page_state(zone, NR_ISOLATED_ANON);
642 return isolated > (inactive + active) / 2;
646 * isolate_migratepages_block() - isolate all migrate-able pages within
648 * @cc: Compaction control structure.
649 * @low_pfn: The first PFN to isolate
650 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
651 * @isolate_mode: Isolation mode to be used.
653 * Isolate all pages that can be migrated from the range specified by
654 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
655 * Returns zero if there is a fatal signal pending, otherwise PFN of the
656 * first page that was not scanned (which may be both less, equal to or more
659 * The pages are isolated on cc->migratepages list (not required to be empty),
660 * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
661 * is neither read nor updated.
664 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
665 unsigned long end_pfn, isolate_mode_t isolate_mode)
667 struct zone *zone = cc->zone;
668 unsigned long nr_scanned = 0, nr_isolated = 0;
669 struct list_head *migratelist = &cc->migratepages;
670 struct lruvec *lruvec;
671 unsigned long flags = 0;
673 struct page *page = NULL, *valid_page = NULL;
674 unsigned long start_pfn = low_pfn;
677 * Ensure that there are not too many pages isolated from the LRU
678 * list by either parallel reclaimers or compaction. If there are,
679 * delay for some time until fewer pages are isolated
681 while (unlikely(too_many_isolated(zone))) {
682 /* async migration should just abort */
683 if (cc->mode == MIGRATE_ASYNC)
686 congestion_wait(BLK_RW_ASYNC, HZ/10);
688 if (fatal_signal_pending(current))
692 if (compact_should_abort(cc))
695 /* Time to isolate some pages for migration */
696 for (; low_pfn < end_pfn; low_pfn++) {
700 * Periodically drop the lock (if held) regardless of its
701 * contention, to give chance to IRQs. Abort async compaction
704 if (!(low_pfn % SWAP_CLUSTER_MAX)
705 && compact_unlock_should_abort(&zone->lru_lock, flags,
709 if (!pfn_valid_within(low_pfn))
713 page = pfn_to_page(low_pfn);
719 * Skip if free. We read page order here without zone lock
720 * which is generally unsafe, but the race window is small and
721 * the worst thing that can happen is that we skip some
722 * potential isolation targets.
724 if (PageBuddy(page)) {
725 unsigned long freepage_order = page_order_unsafe(page);
728 * Without lock, we cannot be sure that what we got is
729 * a valid page order. Consider only values in the
730 * valid order range to prevent low_pfn overflow.
732 if (freepage_order > 0 && freepage_order < MAX_ORDER)
733 low_pfn += (1UL << freepage_order) - 1;
738 * Check may be lockless but that's ok as we recheck later.
739 * It's possible to migrate LRU pages and balloon pages
740 * Skip any other type of page
742 is_lru = PageLRU(page);
744 if (unlikely(balloon_page_movable(page))) {
745 if (balloon_page_isolate(page)) {
746 /* Successfully isolated */
747 goto isolate_success;
753 * Regardless of being on LRU, compound pages such as THP and
754 * hugetlbfs are not to be compacted. We can potentially save
755 * a lot of iterations if we skip them at once. The check is
756 * racy, but we can consider only valid values and the only
757 * danger is skipping too much.
759 if (PageCompound(page)) {
760 unsigned int comp_order = compound_order(page);
762 if (likely(comp_order < MAX_ORDER))
763 low_pfn += (1UL << comp_order) - 1;
772 * Migration will fail if an anonymous page is pinned in memory,
773 * so avoid taking lru_lock and isolating it unnecessarily in an
774 * admittedly racy check.
776 if (!page_mapping(page) &&
777 page_count(page) > page_mapcount(page))
780 /* If we already hold the lock, we can skip some rechecking */
782 locked = compact_trylock_irqsave(&zone->lru_lock,
787 /* Recheck PageLRU and PageCompound under lock */
792 * Page become compound since the non-locked check,
793 * and it's on LRU. It can only be a THP so the order
794 * is safe to read and it's 0 for tail pages.
796 if (unlikely(PageCompound(page))) {
797 low_pfn += (1UL << compound_order(page)) - 1;
802 lruvec = mem_cgroup_page_lruvec(page, zone);
804 /* Try isolate the page */
805 if (__isolate_lru_page(page, isolate_mode) != 0)
808 VM_BUG_ON_PAGE(PageCompound(page), page);
810 /* Successfully isolated */
811 del_page_from_lru_list(page, lruvec, page_lru(page));
814 list_add(&page->lru, migratelist);
815 cc->nr_migratepages++;
818 /* Avoid isolating too much */
819 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX &&
820 !is_cma_page(pfn_to_page(low_pfn))) {
827 * The PageBuddy() check could have potentially brought us outside
828 * the range to be scanned.
830 if (unlikely(low_pfn > end_pfn))
834 spin_unlock_irqrestore(&zone->lru_lock, flags);
837 * Update the pageblock-skip information and cached scanner pfn,
838 * if the whole pageblock was scanned without isolating any page.
840 if (low_pfn == end_pfn)
841 update_pageblock_skip(cc, valid_page, nr_isolated, true);
843 trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
844 nr_scanned, nr_isolated);
846 count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
848 count_compact_events(COMPACTISOLATED, nr_isolated);
854 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
855 * @cc: Compaction control structure.
856 * @start_pfn: The first PFN to start isolating.
857 * @end_pfn: The one-past-last PFN.
859 * Returns zero if isolation fails fatally due to e.g. pending signal.
860 * Otherwise, function returns one-past-the-last PFN of isolated page
861 * (which may be greater than end_pfn if end fell in a middle of a THP page).
864 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
865 unsigned long end_pfn)
867 unsigned long pfn, block_end_pfn;
869 /* Scan block by block. First and last block may be incomplete */
871 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
873 for (; pfn < end_pfn; pfn = block_end_pfn,
874 block_end_pfn += pageblock_nr_pages) {
876 block_end_pfn = min(block_end_pfn, end_pfn);
878 if (!pageblock_pfn_to_page(pfn, block_end_pfn, cc->zone))
881 pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
882 ISOLATE_UNEVICTABLE);
887 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
890 acct_isolated(cc->zone, cc);
895 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
896 #ifdef CONFIG_COMPACTION
898 /* Returns true if the page is within a block suitable for migration to */
899 static bool suitable_migration_target(struct page *page)
901 /* If the page is a large free page, then disallow migration */
902 if (PageBuddy(page)) {
904 * We are checking page_order without zone->lock taken. But
905 * the only small danger is that we skip a potentially suitable
906 * pageblock, so it's not worth to check order for valid range.
908 if (page_order_unsafe(page) >= pageblock_order)
912 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
913 if (migrate_async_suitable(get_pageblock_migratetype(page)))
916 /* Otherwise skip the block */
921 * Test whether the free scanner has reached the same or lower pageblock than
922 * the migration scanner, and compaction should thus terminate.
924 static inline bool compact_scanners_met(struct compact_control *cc)
926 return (cc->free_pfn >> pageblock_order)
927 <= (cc->migrate_pfn >> pageblock_order);
931 * Based on information in the current compact_control, find blocks
932 * suitable for isolating free pages from and then isolate them.
934 static void isolate_freepages(struct compact_control *cc)
936 struct zone *zone = cc->zone;
938 unsigned long block_start_pfn; /* start of current pageblock */
939 unsigned long isolate_start_pfn; /* exact pfn we start at */
940 unsigned long block_end_pfn; /* end of current pageblock */
941 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
942 struct list_head *freelist = &cc->freepages;
945 * Initialise the free scanner. The starting point is where we last
946 * successfully isolated from, zone-cached value, or the end of the
947 * zone when isolating for the first time. For looping we also need
948 * this pfn aligned down to the pageblock boundary, because we do
949 * block_start_pfn -= pageblock_nr_pages in the for loop.
950 * For ending point, take care when isolating in last pageblock of a
951 * a zone which ends in the middle of a pageblock.
952 * The low boundary is the end of the pageblock the migration scanner
955 isolate_start_pfn = cc->free_pfn;
956 block_start_pfn = cc->free_pfn & ~(pageblock_nr_pages-1);
957 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
959 low_pfn = ALIGN(cc->migrate_pfn + 1, pageblock_nr_pages);
962 * Isolate free pages until enough are available to migrate the
963 * pages on cc->migratepages. We stop searching if the migrate
964 * and free page scanners meet or enough free pages are isolated.
966 for (; block_start_pfn >= low_pfn;
967 block_end_pfn = block_start_pfn,
968 block_start_pfn -= pageblock_nr_pages,
969 isolate_start_pfn = block_start_pfn) {
972 * This can iterate a massively long zone without finding any
973 * suitable migration targets, so periodically check if we need
974 * to schedule, or even abort async compaction.
976 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
977 && compact_should_abort(cc))
980 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
985 /* Check the block is suitable for migration */
986 if (!suitable_migration_target(page))
989 /* If isolation recently failed, do not retry */
990 if (!isolation_suitable(cc, page))
993 /* Found a block suitable for isolating free pages from. */
994 isolate_freepages_block(cc, &isolate_start_pfn,
995 block_end_pfn, freelist, false);
998 * If we isolated enough freepages, or aborted due to async
999 * compaction being contended, terminate the loop.
1000 * Remember where the free scanner should restart next time,
1001 * which is where isolate_freepages_block() left off.
1002 * But if it scanned the whole pageblock, isolate_start_pfn
1003 * now points at block_end_pfn, which is the start of the next
1005 * In that case we will however want to restart at the start
1006 * of the previous pageblock.
1008 if ((cc->nr_freepages >= cc->nr_migratepages)
1010 if (isolate_start_pfn >= block_end_pfn)
1012 block_start_pfn - pageblock_nr_pages;
1016 * isolate_freepages_block() should not terminate
1017 * prematurely unless contended, or isolated enough
1019 VM_BUG_ON(isolate_start_pfn < block_end_pfn);
1023 /* split_free_page does not map the pages */
1024 map_pages(freelist);
1027 * Record where the free scanner will restart next time. Either we
1028 * broke from the loop and set isolate_start_pfn based on the last
1029 * call to isolate_freepages_block(), or we met the migration scanner
1030 * and the loop terminated due to isolate_start_pfn < low_pfn
1032 cc->free_pfn = isolate_start_pfn;
1036 * This is a migrate-callback that "allocates" freepages by taking pages
1037 * from the isolated freelists in the block we are migrating to.
1039 static struct page *compaction_alloc(struct page *migratepage,
1043 struct compact_control *cc = (struct compact_control *)data;
1044 struct page *freepage;
1047 * Isolate free pages if necessary, and if we are not aborting due to
1050 if (list_empty(&cc->freepages)) {
1052 isolate_freepages(cc);
1054 if (list_empty(&cc->freepages))
1058 freepage = list_entry(cc->freepages.next, struct page, lru);
1059 list_del(&freepage->lru);
1066 * This is a migrate-callback that "frees" freepages back to the isolated
1067 * freelist. All pages on the freelist are from the same zone, so there is no
1068 * special handling needed for NUMA.
1070 static void compaction_free(struct page *page, unsigned long data)
1072 struct compact_control *cc = (struct compact_control *)data;
1074 list_add(&page->lru, &cc->freepages);
1078 /* possible outcome of isolate_migratepages */
1080 ISOLATE_ABORT, /* Abort compaction now */
1081 ISOLATE_NONE, /* No pages isolated, continue scanning */
1082 ISOLATE_SUCCESS, /* Pages isolated, migrate */
1083 } isolate_migrate_t;
1086 * Allow userspace to control policy on scanning the unevictable LRU for
1087 * compactable pages.
1089 int sysctl_compact_unevictable_allowed __read_mostly = 1;
1092 * Isolate all pages that can be migrated from the first suitable block,
1093 * starting at the block pointed to by the migrate scanner pfn within
1096 static isolate_migrate_t isolate_migratepages(struct zone *zone,
1097 struct compact_control *cc)
1099 unsigned long low_pfn, end_pfn;
1100 unsigned long isolate_start_pfn;
1102 const isolate_mode_t isolate_mode =
1103 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1104 (cc->mode == MIGRATE_ASYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1107 * Start at where we last stopped, or beginning of the zone as
1108 * initialized by compact_zone()
1110 low_pfn = cc->migrate_pfn;
1112 /* Only scan within a pageblock boundary */
1113 end_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages);
1116 * Iterate over whole pageblocks until we find the first suitable.
1117 * Do not cross the free scanner.
1119 for (; end_pfn <= cc->free_pfn;
1120 low_pfn = end_pfn, end_pfn += pageblock_nr_pages) {
1123 * This can potentially iterate a massively long zone with
1124 * many pageblocks unsuitable, so periodically check if we
1125 * need to schedule, or even abort async compaction.
1127 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1128 && compact_should_abort(cc))
1131 page = pageblock_pfn_to_page(low_pfn, end_pfn, zone);
1135 /* If isolation recently failed, do not retry */
1136 if (!isolation_suitable(cc, page))
1140 * For async compaction, also only scan in MOVABLE blocks.
1141 * Async compaction is optimistic to see if the minimum amount
1142 * of work satisfies the allocation.
1144 if (cc->mode == MIGRATE_ASYNC &&
1145 !migrate_async_suitable(get_pageblock_migratetype(page)))
1148 /* Perform the isolation */
1149 isolate_start_pfn = low_pfn;
1150 low_pfn = isolate_migratepages_block(cc, low_pfn, end_pfn,
1153 if (!low_pfn || cc->contended) {
1154 acct_isolated(zone, cc);
1155 return ISOLATE_ABORT;
1159 * Record where we could have freed pages by migration and not
1160 * yet flushed them to buddy allocator.
1161 * - this is the lowest page that could have been isolated and
1162 * then freed by migration.
1164 if (cc->nr_migratepages && !cc->last_migrated_pfn)
1165 cc->last_migrated_pfn = isolate_start_pfn;
1168 * Either we isolated something and proceed with migration. Or
1169 * we failed and compact_zone should decide if we should
1175 acct_isolated(zone, cc);
1176 /* Record where migration scanner will be restarted. */
1177 cc->migrate_pfn = low_pfn;
1179 return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
1183 * order == -1 is expected when compacting via
1184 * /proc/sys/vm/compact_memory
1186 static inline bool is_via_compact_memory(int order)
1191 static int __compact_finished(struct zone *zone, struct compact_control *cc,
1192 const int migratetype)
1195 unsigned long watermark;
1197 if (cc->contended || fatal_signal_pending(current))
1198 return COMPACT_CONTENDED;
1200 /* Compaction run completes if the migrate and free scanner meet */
1201 if (compact_scanners_met(cc)) {
1202 /* Let the next compaction start anew. */
1203 reset_cached_positions(zone);
1206 * Mark that the PG_migrate_skip information should be cleared
1207 * by kswapd when it goes to sleep. kswapd does not set the
1208 * flag itself as the decision to be clear should be directly
1209 * based on an allocation request.
1211 if (!current_is_kswapd())
1212 zone->compact_blockskip_flush = true;
1214 return COMPACT_COMPLETE;
1217 if (is_via_compact_memory(cc->order))
1218 return COMPACT_CONTINUE;
1220 /* Compaction run is not finished if the watermark is not met */
1221 watermark = low_wmark_pages(zone);
1223 if (!zone_watermark_ok(zone, cc->order, watermark, cc->classzone_idx,
1225 return COMPACT_CONTINUE;
1227 /* Direct compactor: Is a suitable page free? */
1228 for (order = cc->order; order < MAX_ORDER; order++) {
1229 struct free_area *area = &zone->free_area[order];
1232 /* Job done if page is free of the right migratetype */
1233 if (!list_empty(&area->free_list[migratetype]))
1234 return COMPACT_PARTIAL;
1237 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
1238 if (migratetype == MIGRATE_MOVABLE &&
1239 !list_empty(&area->free_list[MIGRATE_CMA]))
1240 return COMPACT_PARTIAL;
1243 * Job done if allocation would steal freepages from
1244 * other migratetype buddy lists.
1246 if (find_suitable_fallback(area, order, migratetype,
1247 true, &can_steal) != -1)
1248 return COMPACT_PARTIAL;
1251 return COMPACT_NO_SUITABLE_PAGE;
1254 static int compact_finished(struct zone *zone, struct compact_control *cc,
1255 const int migratetype)
1259 ret = __compact_finished(zone, cc, migratetype);
1260 trace_mm_compaction_finished(zone, cc->order, ret);
1261 if (ret == COMPACT_NO_SUITABLE_PAGE)
1262 ret = COMPACT_CONTINUE;
1268 * compaction_suitable: Is this suitable to run compaction on this zone now?
1270 * COMPACT_SKIPPED - If there are too few free pages for compaction
1271 * COMPACT_PARTIAL - If the allocation would succeed without compaction
1272 * COMPACT_CONTINUE - If compaction should run now
1274 static unsigned long __compaction_suitable(struct zone *zone, int order,
1275 int alloc_flags, int classzone_idx)
1278 unsigned long watermark;
1280 if (is_via_compact_memory(order))
1281 return COMPACT_CONTINUE;
1283 watermark = low_wmark_pages(zone);
1285 * If watermarks for high-order allocation are already met, there
1286 * should be no need for compaction at all.
1288 if (zone_watermark_ok(zone, order, watermark, classzone_idx,
1290 return COMPACT_PARTIAL;
1293 * Watermarks for order-0 must be met for compaction. Note the 2UL.
1294 * This is because during migration, copies of pages need to be
1295 * allocated and for a short time, the footprint is higher
1297 watermark += (2UL << order);
1298 if (!zone_watermark_ok(zone, 0, watermark, classzone_idx, alloc_flags))
1299 return COMPACT_SKIPPED;
1302 * fragmentation index determines if allocation failures are due to
1303 * low memory or external fragmentation
1305 * index of -1000 would imply allocations might succeed depending on
1306 * watermarks, but we already failed the high-order watermark check
1307 * index towards 0 implies failure is due to lack of memory
1308 * index towards 1000 implies failure is due to fragmentation
1310 * Only compact if a failure would be due to fragmentation.
1312 fragindex = fragmentation_index(zone, order);
1313 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
1314 return COMPACT_NOT_SUITABLE_ZONE;
1316 return COMPACT_CONTINUE;
1319 unsigned long compaction_suitable(struct zone *zone, int order,
1320 int alloc_flags, int classzone_idx)
1324 ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx);
1325 trace_mm_compaction_suitable(zone, order, ret);
1326 if (ret == COMPACT_NOT_SUITABLE_ZONE)
1327 ret = COMPACT_SKIPPED;
1332 static int compact_zone(struct zone *zone, struct compact_control *cc)
1335 unsigned long start_pfn = zone->zone_start_pfn;
1336 unsigned long end_pfn = zone_end_pfn(zone);
1337 const int migratetype = gfpflags_to_migratetype(cc->gfp_mask);
1338 const bool sync = cc->mode != MIGRATE_ASYNC;
1340 ret = compaction_suitable(zone, cc->order, cc->alloc_flags,
1343 case COMPACT_PARTIAL:
1344 case COMPACT_SKIPPED:
1345 /* Compaction is likely to fail */
1347 case COMPACT_CONTINUE:
1348 /* Fall through to compaction */
1353 * Clear pageblock skip if there were failures recently and compaction
1354 * is about to be retried after being deferred. kswapd does not do
1355 * this reset as it'll reset the cached information when going to sleep.
1357 if (compaction_restarting(zone, cc->order) && !current_is_kswapd())
1358 __reset_isolation_suitable(zone);
1361 * Setup to move all movable pages to the end of the zone. Used cached
1362 * information on where the scanners should start but check that it
1363 * is initialised by ensuring the values are within zone boundaries.
1365 cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync];
1366 cc->free_pfn = zone->compact_cached_free_pfn;
1367 if (cc->free_pfn < start_pfn || cc->free_pfn > end_pfn) {
1368 cc->free_pfn = end_pfn & ~(pageblock_nr_pages-1);
1369 zone->compact_cached_free_pfn = cc->free_pfn;
1371 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn > end_pfn) {
1372 cc->migrate_pfn = start_pfn;
1373 zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
1374 zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
1376 cc->last_migrated_pfn = 0;
1378 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
1379 cc->free_pfn, end_pfn, sync);
1381 migrate_prep_local();
1383 while ((ret = compact_finished(zone, cc, migratetype)) ==
1387 switch (isolate_migratepages(zone, cc)) {
1389 ret = COMPACT_CONTENDED;
1390 putback_movable_pages(&cc->migratepages);
1391 cc->nr_migratepages = 0;
1395 * We haven't isolated and migrated anything, but
1396 * there might still be unflushed migrations from
1397 * previous cc->order aligned block.
1400 case ISOLATE_SUCCESS:
1404 err = migrate_pages(&cc->migratepages, compaction_alloc,
1405 compaction_free, (unsigned long)cc, cc->mode,
1408 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
1411 /* All pages were either migrated or will be released */
1412 cc->nr_migratepages = 0;
1414 putback_movable_pages(&cc->migratepages);
1416 * migrate_pages() may return -ENOMEM when scanners meet
1417 * and we want compact_finished() to detect it
1419 if (err == -ENOMEM && !compact_scanners_met(cc)) {
1420 ret = COMPACT_CONTENDED;
1427 * Has the migration scanner moved away from the previous
1428 * cc->order aligned block where we migrated from? If yes,
1429 * flush the pages that were freed, so that they can merge and
1430 * compact_finished() can detect immediately if allocation
1433 if (cc->order > 0 && cc->last_migrated_pfn) {
1435 unsigned long current_block_start =
1436 cc->migrate_pfn & ~((1UL << cc->order) - 1);
1438 if (cc->last_migrated_pfn < current_block_start) {
1440 lru_add_drain_cpu(cpu);
1441 drain_local_pages(zone);
1443 /* No more flushing until we migrate again */
1444 cc->last_migrated_pfn = 0;
1452 * Release free pages and update where the free scanner should restart,
1453 * so we don't leave any returned pages behind in the next attempt.
1455 if (cc->nr_freepages > 0) {
1456 unsigned long free_pfn = release_freepages(&cc->freepages);
1458 cc->nr_freepages = 0;
1459 VM_BUG_ON(free_pfn == 0);
1460 /* The cached pfn is always the first in a pageblock */
1461 free_pfn &= ~(pageblock_nr_pages-1);
1463 * Only go back, not forward. The cached pfn might have been
1464 * already reset to zone end in compact_finished()
1466 if (free_pfn > zone->compact_cached_free_pfn)
1467 zone->compact_cached_free_pfn = free_pfn;
1470 trace_mm_compaction_end(start_pfn, cc->migrate_pfn,
1471 cc->free_pfn, end_pfn, sync, ret);
1473 if (ret == COMPACT_CONTENDED)
1474 ret = COMPACT_PARTIAL;
1479 static unsigned long compact_zone_order(struct zone *zone, int order,
1480 gfp_t gfp_mask, enum migrate_mode mode, int *contended,
1481 int alloc_flags, int classzone_idx)
1484 struct compact_control cc = {
1486 .nr_migratepages = 0,
1488 .gfp_mask = gfp_mask,
1491 .alloc_flags = alloc_flags,
1492 .classzone_idx = classzone_idx,
1494 INIT_LIST_HEAD(&cc.freepages);
1495 INIT_LIST_HEAD(&cc.migratepages);
1497 ret = compact_zone(zone, &cc);
1499 VM_BUG_ON(!list_empty(&cc.freepages));
1500 VM_BUG_ON(!list_empty(&cc.migratepages));
1502 *contended = cc.contended;
1506 int sysctl_extfrag_threshold = 500;
1509 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1510 * @gfp_mask: The GFP mask of the current allocation
1511 * @order: The order of the current allocation
1512 * @alloc_flags: The allocation flags of the current allocation
1513 * @ac: The context of current allocation
1514 * @mode: The migration mode for async, sync light, or sync migration
1515 * @contended: Return value that determines if compaction was aborted due to
1516 * need_resched() or lock contention
1518 * This is the main entry point for direct page compaction.
1520 unsigned long try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
1521 int alloc_flags, const struct alloc_context *ac,
1522 enum migrate_mode mode, int *contended)
1524 int may_enter_fs = gfp_mask & __GFP_FS;
1525 int may_perform_io = gfp_mask & __GFP_IO;
1528 int rc = COMPACT_DEFERRED;
1529 int all_zones_contended = COMPACT_CONTENDED_LOCK; /* init for &= op */
1531 *contended = COMPACT_CONTENDED_NONE;
1533 /* Check if the GFP flags allow compaction */
1534 if (!order || !may_enter_fs || !may_perform_io)
1535 return COMPACT_SKIPPED;
1537 trace_mm_compaction_try_to_compact_pages(order, gfp_mask, mode);
1539 /* Compact each zone in the list */
1540 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1545 if (compaction_deferred(zone, order))
1548 status = compact_zone_order(zone, order, gfp_mask, mode,
1549 &zone_contended, alloc_flags,
1551 rc = max(status, rc);
1553 * It takes at least one zone that wasn't lock contended
1554 * to clear all_zones_contended.
1556 all_zones_contended &= zone_contended;
1558 /* If a normal allocation would succeed, stop compacting */
1559 if (zone_watermark_ok(zone, order, low_wmark_pages(zone),
1560 ac->classzone_idx, alloc_flags)) {
1562 * We think the allocation will succeed in this zone,
1563 * but it is not certain, hence the false. The caller
1564 * will repeat this with true if allocation indeed
1565 * succeeds in this zone.
1567 compaction_defer_reset(zone, order, false);
1569 * It is possible that async compaction aborted due to
1570 * need_resched() and the watermarks were ok thanks to
1571 * somebody else freeing memory. The allocation can
1572 * however still fail so we better signal the
1573 * need_resched() contention anyway (this will not
1574 * prevent the allocation attempt).
1576 if (zone_contended == COMPACT_CONTENDED_SCHED)
1577 *contended = COMPACT_CONTENDED_SCHED;
1582 if (mode != MIGRATE_ASYNC && status == COMPACT_COMPLETE) {
1584 * We think that allocation won't succeed in this zone
1585 * so we defer compaction there. If it ends up
1586 * succeeding after all, it will be reset.
1588 defer_compaction(zone, order);
1592 * We might have stopped compacting due to need_resched() in
1593 * async compaction, or due to a fatal signal detected. In that
1594 * case do not try further zones and signal need_resched()
1597 if ((zone_contended == COMPACT_CONTENDED_SCHED)
1598 || fatal_signal_pending(current)) {
1599 *contended = COMPACT_CONTENDED_SCHED;
1606 * We might not have tried all the zones, so be conservative
1607 * and assume they are not all lock contended.
1609 all_zones_contended = 0;
1614 * If at least one zone wasn't deferred or skipped, we report if all
1615 * zones that were tried were lock contended.
1617 if (rc > COMPACT_SKIPPED && all_zones_contended)
1618 *contended = COMPACT_CONTENDED_LOCK;
1624 /* Compact all zones within a node */
1625 static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
1630 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1632 zone = &pgdat->node_zones[zoneid];
1633 if (!populated_zone(zone))
1636 cc->nr_freepages = 0;
1637 cc->nr_migratepages = 0;
1639 INIT_LIST_HEAD(&cc->freepages);
1640 INIT_LIST_HEAD(&cc->migratepages);
1643 * When called via /proc/sys/vm/compact_memory
1644 * this makes sure we compact the whole zone regardless of
1645 * cached scanner positions.
1647 if (is_via_compact_memory(cc->order))
1648 __reset_isolation_suitable(zone);
1650 if (is_via_compact_memory(cc->order) ||
1651 !compaction_deferred(zone, cc->order))
1652 compact_zone(zone, cc);
1654 if (cc->order > 0) {
1655 if (zone_watermark_ok(zone, cc->order,
1656 low_wmark_pages(zone), 0, 0))
1657 compaction_defer_reset(zone, cc->order, false);
1660 VM_BUG_ON(!list_empty(&cc->freepages));
1661 VM_BUG_ON(!list_empty(&cc->migratepages));
1665 void compact_pgdat(pg_data_t *pgdat, int order)
1667 struct compact_control cc = {
1669 .mode = MIGRATE_ASYNC,
1675 __compact_pgdat(pgdat, &cc);
1678 static void compact_node(int nid)
1680 struct compact_control cc = {
1682 .mode = MIGRATE_SYNC,
1683 .ignore_skip_hint = true,
1686 __compact_pgdat(NODE_DATA(nid), &cc);
1689 /* Compact all nodes in the system */
1690 static void compact_nodes(void)
1694 /* Flush pending updates to the LRU lists */
1695 lru_add_drain_all();
1697 for_each_online_node(nid)
1701 /* The written value is actually unused, all memory is compacted */
1702 int sysctl_compact_memory;
1704 /* This is the entry point for compacting all nodes via /proc/sys/vm */
1705 int sysctl_compaction_handler(struct ctl_table *table, int write,
1706 void __user *buffer, size_t *length, loff_t *ppos)
1714 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1715 void __user *buffer, size_t *length, loff_t *ppos)
1717 proc_dointvec_minmax(table, write, buffer, length, ppos);
1722 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1723 static ssize_t sysfs_compact_node(struct device *dev,
1724 struct device_attribute *attr,
1725 const char *buf, size_t count)
1729 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1730 /* Flush pending updates to the LRU lists */
1731 lru_add_drain_all();
1738 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1740 int compaction_register_node(struct node *node)
1742 return device_create_file(&node->dev, &dev_attr_compact);
1745 void compaction_unregister_node(struct node *node)
1747 return device_remove_file(&node->dev, &dev_attr_compact);
1749 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1751 #endif /* CONFIG_COMPACTION */