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/cpu.h>
11 #include <linux/swap.h>
12 #include <linux/migrate.h>
13 #include <linux/compaction.h>
14 #include <linux/mm_inline.h>
15 #include <linux/backing-dev.h>
16 #include <linux/sysctl.h>
17 #include <linux/sysfs.h>
18 #include <linux/page-isolation.h>
19 #include <linux/kasan.h>
20 #include <linux/kthread.h>
21 #include <linux/freezer.h>
24 #ifdef CONFIG_COMPACTION
25 static inline void count_compact_event(enum vm_event_item item
)
30 static inline void count_compact_events(enum vm_event_item item
, long delta
)
32 count_vm_events(item
, delta
);
35 #define count_compact_event(item) do { } while (0)
36 #define count_compact_events(item, delta) do { } while (0)
39 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
41 #define CREATE_TRACE_POINTS
42 #include <trace/events/compaction.h>
44 #define block_start_pfn(pfn, order) round_down(pfn, 1UL << (order))
45 #define block_end_pfn(pfn, order) ALIGN((pfn) + 1, 1UL << (order))
46 #define pageblock_start_pfn(pfn) block_start_pfn(pfn, pageblock_order)
47 #define pageblock_end_pfn(pfn) block_end_pfn(pfn, pageblock_order)
49 static unsigned long release_freepages(struct list_head
*freelist
)
51 struct page
*page
, *next
;
52 unsigned long high_pfn
= 0;
54 list_for_each_entry_safe(page
, next
, freelist
, lru
) {
55 unsigned long pfn
= page_to_pfn(page
);
65 static void map_pages(struct list_head
*list
)
69 list_for_each_entry(page
, list
, lru
) {
70 arch_alloc_page(page
, 0);
71 kernel_map_pages(page
, 1, 1);
72 kasan_alloc_pages(page
, 0);
76 static inline bool migrate_async_suitable(int migratetype
)
78 return is_migrate_cma(migratetype
) || migratetype
== MIGRATE_MOVABLE
;
81 #ifdef CONFIG_COMPACTION
83 int PageMovable(struct page
*page
)
85 struct address_space
*mapping
;
87 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
88 if (!__PageMovable(page
))
91 mapping
= page_mapping(page
);
92 if (mapping
&& mapping
->a_ops
&& mapping
->a_ops
->isolate_page
)
97 EXPORT_SYMBOL(PageMovable
);
99 void __SetPageMovable(struct page
*page
, struct address_space
*mapping
)
101 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
102 VM_BUG_ON_PAGE((unsigned long)mapping
& PAGE_MAPPING_MOVABLE
, page
);
103 page
->mapping
= (void *)((unsigned long)mapping
| PAGE_MAPPING_MOVABLE
);
105 EXPORT_SYMBOL(__SetPageMovable
);
107 void __ClearPageMovable(struct page
*page
)
109 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
110 VM_BUG_ON_PAGE(!PageMovable(page
), page
);
112 * Clear registered address_space val with keeping PAGE_MAPPING_MOVABLE
113 * flag so that VM can catch up released page by driver after isolation.
114 * With it, VM migration doesn't try to put it back.
116 page
->mapping
= (void *)((unsigned long)page
->mapping
&
117 PAGE_MAPPING_MOVABLE
);
119 EXPORT_SYMBOL(__ClearPageMovable
);
121 /* Do not skip compaction more than 64 times */
122 #define COMPACT_MAX_DEFER_SHIFT 6
125 * Compaction is deferred when compaction fails to result in a page
126 * allocation success. 1 << compact_defer_limit compactions are skipped up
127 * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
129 void defer_compaction(struct zone
*zone
, int order
)
131 zone
->compact_considered
= 0;
132 zone
->compact_defer_shift
++;
134 if (order
< zone
->compact_order_failed
)
135 zone
->compact_order_failed
= order
;
137 if (zone
->compact_defer_shift
> COMPACT_MAX_DEFER_SHIFT
)
138 zone
->compact_defer_shift
= COMPACT_MAX_DEFER_SHIFT
;
140 trace_mm_compaction_defer_compaction(zone
, order
);
143 /* Returns true if compaction should be skipped this time */
144 bool compaction_deferred(struct zone
*zone
, int order
)
146 unsigned long defer_limit
= 1UL << zone
->compact_defer_shift
;
148 if (order
< zone
->compact_order_failed
)
151 /* Avoid possible overflow */
152 if (++zone
->compact_considered
> defer_limit
)
153 zone
->compact_considered
= defer_limit
;
155 if (zone
->compact_considered
>= defer_limit
)
158 trace_mm_compaction_deferred(zone
, order
);
164 * Update defer tracking counters after successful compaction of given order,
165 * which means an allocation either succeeded (alloc_success == true) or is
166 * expected to succeed.
168 void compaction_defer_reset(struct zone
*zone
, int order
,
172 zone
->compact_considered
= 0;
173 zone
->compact_defer_shift
= 0;
175 if (order
>= zone
->compact_order_failed
)
176 zone
->compact_order_failed
= order
+ 1;
178 trace_mm_compaction_defer_reset(zone
, order
);
181 /* Returns true if restarting compaction after many failures */
182 bool compaction_restarting(struct zone
*zone
, int order
)
184 if (order
< zone
->compact_order_failed
)
187 return zone
->compact_defer_shift
== COMPACT_MAX_DEFER_SHIFT
&&
188 zone
->compact_considered
>= 1UL << zone
->compact_defer_shift
;
191 /* Returns true if the pageblock should be scanned for pages to isolate. */
192 static inline bool isolation_suitable(struct compact_control
*cc
,
195 if (cc
->ignore_skip_hint
)
198 return !get_pageblock_skip(page
);
201 static void reset_cached_positions(struct zone
*zone
)
203 zone
->compact_cached_migrate_pfn
[0] = zone
->zone_start_pfn
;
204 zone
->compact_cached_migrate_pfn
[1] = zone
->zone_start_pfn
;
205 zone
->compact_cached_free_pfn
=
206 pageblock_start_pfn(zone_end_pfn(zone
) - 1);
210 * This function is called to clear all cached information on pageblocks that
211 * should be skipped for page isolation when the migrate and free page scanner
214 static void __reset_isolation_suitable(struct zone
*zone
)
216 unsigned long start_pfn
= zone
->zone_start_pfn
;
217 unsigned long end_pfn
= zone_end_pfn(zone
);
220 zone
->compact_blockskip_flush
= false;
222 /* Walk the zone and mark every pageblock as suitable for isolation */
223 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
231 page
= pfn_to_page(pfn
);
232 if (zone
!= page_zone(page
))
235 clear_pageblock_skip(page
);
238 reset_cached_positions(zone
);
241 void reset_isolation_suitable(pg_data_t
*pgdat
)
245 for (zoneid
= 0; zoneid
< MAX_NR_ZONES
; zoneid
++) {
246 struct zone
*zone
= &pgdat
->node_zones
[zoneid
];
247 if (!populated_zone(zone
))
250 /* Only flush if a full compaction finished recently */
251 if (zone
->compact_blockskip_flush
)
252 __reset_isolation_suitable(zone
);
257 * If no pages were isolated then mark this pageblock to be skipped in the
258 * future. The information is later cleared by __reset_isolation_suitable().
260 static void update_pageblock_skip(struct compact_control
*cc
,
261 struct page
*page
, unsigned long nr_isolated
,
262 bool migrate_scanner
)
264 struct zone
*zone
= cc
->zone
;
267 if (cc
->ignore_skip_hint
)
276 set_pageblock_skip(page
);
278 pfn
= page_to_pfn(page
);
280 /* Update where async and sync compaction should restart */
281 if (migrate_scanner
) {
282 if (pfn
> zone
->compact_cached_migrate_pfn
[0])
283 zone
->compact_cached_migrate_pfn
[0] = pfn
;
284 if (cc
->mode
!= MIGRATE_ASYNC
&&
285 pfn
> zone
->compact_cached_migrate_pfn
[1])
286 zone
->compact_cached_migrate_pfn
[1] = pfn
;
288 if (pfn
< zone
->compact_cached_free_pfn
)
289 zone
->compact_cached_free_pfn
= pfn
;
293 static inline bool isolation_suitable(struct compact_control
*cc
,
299 static void update_pageblock_skip(struct compact_control
*cc
,
300 struct page
*page
, unsigned long nr_isolated
,
301 bool migrate_scanner
)
304 #endif /* CONFIG_COMPACTION */
307 * Compaction requires the taking of some coarse locks that are potentially
308 * very heavily contended. For async compaction, back out if the lock cannot
309 * be taken immediately. For sync compaction, spin on the lock if needed.
311 * Returns true if the lock is held
312 * Returns false if the lock is not held and compaction should abort
314 static bool compact_trylock_irqsave(spinlock_t
*lock
, unsigned long *flags
,
315 struct compact_control
*cc
)
317 if (cc
->mode
== MIGRATE_ASYNC
) {
318 if (!spin_trylock_irqsave(lock
, *flags
)) {
319 cc
->contended
= COMPACT_CONTENDED_LOCK
;
323 spin_lock_irqsave(lock
, *flags
);
330 * Compaction requires the taking of some coarse locks that are potentially
331 * very heavily contended. The lock should be periodically unlocked to avoid
332 * having disabled IRQs for a long time, even when there is nobody waiting on
333 * the lock. It might also be that allowing the IRQs will result in
334 * need_resched() becoming true. If scheduling is needed, async compaction
335 * aborts. Sync compaction schedules.
336 * Either compaction type will also abort if a fatal signal is pending.
337 * In either case if the lock was locked, it is dropped and not regained.
339 * Returns true if compaction should abort due to fatal signal pending, or
340 * async compaction due to need_resched()
341 * Returns false when compaction can continue (sync compaction might have
344 static bool compact_unlock_should_abort(spinlock_t
*lock
,
345 unsigned long flags
, bool *locked
, struct compact_control
*cc
)
348 spin_unlock_irqrestore(lock
, flags
);
352 if (fatal_signal_pending(current
)) {
353 cc
->contended
= COMPACT_CONTENDED_SCHED
;
357 if (need_resched()) {
358 if (cc
->mode
== MIGRATE_ASYNC
) {
359 cc
->contended
= COMPACT_CONTENDED_SCHED
;
369 * Aside from avoiding lock contention, compaction also periodically checks
370 * need_resched() and either schedules in sync compaction or aborts async
371 * compaction. This is similar to what compact_unlock_should_abort() does, but
372 * is used where no lock is concerned.
374 * Returns false when no scheduling was needed, or sync compaction scheduled.
375 * Returns true when async compaction should abort.
377 static inline bool compact_should_abort(struct compact_control
*cc
)
379 /* async compaction aborts if contended */
380 if (need_resched()) {
381 if (cc
->mode
== MIGRATE_ASYNC
) {
382 cc
->contended
= COMPACT_CONTENDED_SCHED
;
393 * Isolate free pages onto a private freelist. If @strict is true, will abort
394 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
395 * (even though it may still end up isolating some pages).
397 static unsigned long isolate_freepages_block(struct compact_control
*cc
,
398 unsigned long *start_pfn
,
399 unsigned long end_pfn
,
400 struct list_head
*freelist
,
403 int nr_scanned
= 0, total_isolated
= 0;
404 struct page
*cursor
, *valid_page
= NULL
;
405 unsigned long flags
= 0;
407 unsigned long blockpfn
= *start_pfn
;
409 cursor
= pfn_to_page(blockpfn
);
411 /* Isolate free pages. */
412 for (; blockpfn
< end_pfn
; blockpfn
++, cursor
++) {
414 struct page
*page
= cursor
;
417 * Periodically drop the lock (if held) regardless of its
418 * contention, to give chance to IRQs. Abort if fatal signal
419 * pending or async compaction detects need_resched()
421 if (!(blockpfn
% SWAP_CLUSTER_MAX
)
422 && compact_unlock_should_abort(&cc
->zone
->lock
, flags
,
427 if (!pfn_valid_within(blockpfn
))
434 * For compound pages such as THP and hugetlbfs, we can save
435 * potentially a lot of iterations if we skip them at once.
436 * The check is racy, but we can consider only valid values
437 * and the only danger is skipping too much.
439 if (PageCompound(page
)) {
440 unsigned int comp_order
= compound_order(page
);
442 if (likely(comp_order
< MAX_ORDER
)) {
443 blockpfn
+= (1UL << comp_order
) - 1;
444 cursor
+= (1UL << comp_order
) - 1;
450 if (!PageBuddy(page
))
454 * If we already hold the lock, we can skip some rechecking.
455 * Note that if we hold the lock now, checked_pageblock was
456 * already set in some previous iteration (or strict is true),
457 * so it is correct to skip the suitable migration target
462 * The zone lock must be held to isolate freepages.
463 * Unfortunately this is a very coarse lock and can be
464 * heavily contended if there are parallel allocations
465 * or parallel compactions. For async compaction do not
466 * spin on the lock and we acquire the lock as late as
469 locked
= compact_trylock_irqsave(&cc
->zone
->lock
,
474 /* Recheck this is a buddy page under lock */
475 if (!PageBuddy(page
))
479 /* Found a free page, break it into order-0 pages */
480 isolated
= split_free_page(page
);
484 total_isolated
+= isolated
;
485 cc
->nr_freepages
+= isolated
;
486 for (i
= 0; i
< isolated
; i
++) {
487 list_add(&page
->lru
, freelist
);
490 if (!strict
&& cc
->nr_migratepages
<= cc
->nr_freepages
) {
491 blockpfn
+= isolated
;
494 /* Advance to the end of split page */
495 blockpfn
+= isolated
- 1;
496 cursor
+= isolated
- 1;
508 spin_unlock_irqrestore(&cc
->zone
->lock
, flags
);
511 * There is a tiny chance that we have read bogus compound_order(),
512 * so be careful to not go outside of the pageblock.
514 if (unlikely(blockpfn
> end_pfn
))
517 trace_mm_compaction_isolate_freepages(*start_pfn
, blockpfn
,
518 nr_scanned
, total_isolated
);
520 /* Record how far we have got within the block */
521 *start_pfn
= blockpfn
;
524 * If strict isolation is requested by CMA then check that all the
525 * pages requested were isolated. If there were any failures, 0 is
526 * returned and CMA will fail.
528 if (strict
&& blockpfn
< end_pfn
)
531 /* Update the pageblock-skip if the whole pageblock was scanned */
532 if (blockpfn
== end_pfn
)
533 update_pageblock_skip(cc
, valid_page
, total_isolated
, false);
535 count_compact_events(COMPACTFREE_SCANNED
, nr_scanned
);
537 count_compact_events(COMPACTISOLATED
, total_isolated
);
538 return total_isolated
;
542 * isolate_freepages_range() - isolate free pages.
543 * @start_pfn: The first PFN to start isolating.
544 * @end_pfn: The one-past-last PFN.
546 * Non-free pages, invalid PFNs, or zone boundaries within the
547 * [start_pfn, end_pfn) range are considered errors, cause function to
548 * undo its actions and return zero.
550 * Otherwise, function returns one-past-the-last PFN of isolated page
551 * (which may be greater then end_pfn if end fell in a middle of
555 isolate_freepages_range(struct compact_control
*cc
,
556 unsigned long start_pfn
, unsigned long end_pfn
)
558 unsigned long isolated
, pfn
, block_start_pfn
, block_end_pfn
;
562 block_start_pfn
= pageblock_start_pfn(pfn
);
563 if (block_start_pfn
< cc
->zone
->zone_start_pfn
)
564 block_start_pfn
= cc
->zone
->zone_start_pfn
;
565 block_end_pfn
= pageblock_end_pfn(pfn
);
567 for (; pfn
< end_pfn
; pfn
+= isolated
,
568 block_start_pfn
= block_end_pfn
,
569 block_end_pfn
+= pageblock_nr_pages
) {
570 /* Protect pfn from changing by isolate_freepages_block */
571 unsigned long isolate_start_pfn
= pfn
;
573 block_end_pfn
= min(block_end_pfn
, end_pfn
);
576 * pfn could pass the block_end_pfn if isolated freepage
577 * is more than pageblock order. In this case, we adjust
578 * scanning range to right one.
580 if (pfn
>= block_end_pfn
) {
581 block_start_pfn
= pageblock_start_pfn(pfn
);
582 block_end_pfn
= pageblock_end_pfn(pfn
);
583 block_end_pfn
= min(block_end_pfn
, end_pfn
);
586 if (!pageblock_pfn_to_page(block_start_pfn
,
587 block_end_pfn
, cc
->zone
))
590 isolated
= isolate_freepages_block(cc
, &isolate_start_pfn
,
591 block_end_pfn
, &freelist
, true);
594 * In strict mode, isolate_freepages_block() returns 0 if
595 * there are any holes in the block (ie. invalid PFNs or
602 * If we managed to isolate pages, it is always (1 << n) *
603 * pageblock_nr_pages for some non-negative n. (Max order
604 * page may span two pageblocks).
608 /* split_free_page does not map the pages */
609 map_pages(&freelist
);
612 /* Loop terminated early, cleanup. */
613 release_freepages(&freelist
);
617 /* We don't use freelists for anything. */
621 /* Update the number of anon and file isolated pages in the zone */
622 static void acct_isolated(struct zone
*zone
, struct compact_control
*cc
)
625 unsigned int count
[2] = { 0, };
627 if (list_empty(&cc
->migratepages
))
630 list_for_each_entry(page
, &cc
->migratepages
, lru
)
631 count
[!!page_is_file_cache(page
)]++;
633 mod_zone_page_state(zone
, NR_ISOLATED_ANON
, count
[0]);
634 mod_zone_page_state(zone
, NR_ISOLATED_FILE
, count
[1]);
637 /* Similar to reclaim, but different enough that they don't share logic */
638 static bool too_many_isolated(struct zone
*zone
)
640 unsigned long active
, inactive
, isolated
;
642 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
) +
643 zone_page_state(zone
, NR_INACTIVE_ANON
);
644 active
= zone_page_state(zone
, NR_ACTIVE_FILE
) +
645 zone_page_state(zone
, NR_ACTIVE_ANON
);
646 isolated
= zone_page_state(zone
, NR_ISOLATED_FILE
) +
647 zone_page_state(zone
, NR_ISOLATED_ANON
);
649 return isolated
> (inactive
+ active
) / 2;
653 * isolate_migratepages_block() - isolate all migrate-able pages within
655 * @cc: Compaction control structure.
656 * @low_pfn: The first PFN to isolate
657 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
658 * @isolate_mode: Isolation mode to be used.
660 * Isolate all pages that can be migrated from the range specified by
661 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
662 * Returns zero if there is a fatal signal pending, otherwise PFN of the
663 * first page that was not scanned (which may be both less, equal to or more
666 * The pages are isolated on cc->migratepages list (not required to be empty),
667 * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
668 * is neither read nor updated.
671 isolate_migratepages_block(struct compact_control
*cc
, unsigned long low_pfn
,
672 unsigned long end_pfn
, isolate_mode_t isolate_mode
)
674 struct zone
*zone
= cc
->zone
;
675 unsigned long nr_scanned
= 0, nr_isolated
= 0;
676 struct lruvec
*lruvec
;
677 unsigned long flags
= 0;
679 struct page
*page
= NULL
, *valid_page
= NULL
;
680 unsigned long start_pfn
= low_pfn
;
681 bool skip_on_failure
= false;
682 unsigned long next_skip_pfn
= 0;
685 * Ensure that there are not too many pages isolated from the LRU
686 * list by either parallel reclaimers or compaction. If there are,
687 * delay for some time until fewer pages are isolated
689 while (unlikely(too_many_isolated(zone
))) {
690 /* async migration should just abort */
691 if (cc
->mode
== MIGRATE_ASYNC
)
694 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
696 if (fatal_signal_pending(current
))
700 if (compact_should_abort(cc
))
703 if (cc
->direct_compaction
&& (cc
->mode
== MIGRATE_ASYNC
)) {
704 skip_on_failure
= true;
705 next_skip_pfn
= block_end_pfn(low_pfn
, cc
->order
);
708 /* Time to isolate some pages for migration */
709 for (; low_pfn
< end_pfn
; low_pfn
++) {
711 if (skip_on_failure
&& low_pfn
>= next_skip_pfn
) {
713 * We have isolated all migration candidates in the
714 * previous order-aligned block, and did not skip it due
715 * to failure. We should migrate the pages now and
716 * hopefully succeed compaction.
722 * We failed to isolate in the previous order-aligned
723 * block. Set the new boundary to the end of the
724 * current block. Note we can't simply increase
725 * next_skip_pfn by 1 << order, as low_pfn might have
726 * been incremented by a higher number due to skipping
727 * a compound or a high-order buddy page in the
728 * previous loop iteration.
730 next_skip_pfn
= block_end_pfn(low_pfn
, cc
->order
);
734 * Periodically drop the lock (if held) regardless of its
735 * contention, to give chance to IRQs. Abort async compaction
738 if (!(low_pfn
% SWAP_CLUSTER_MAX
)
739 && compact_unlock_should_abort(&zone
->lru_lock
, flags
,
743 if (!pfn_valid_within(low_pfn
))
747 page
= pfn_to_page(low_pfn
);
753 * Skip if free. We read page order here without zone lock
754 * which is generally unsafe, but the race window is small and
755 * the worst thing that can happen is that we skip some
756 * potential isolation targets.
758 if (PageBuddy(page
)) {
759 unsigned long freepage_order
= page_order_unsafe(page
);
762 * Without lock, we cannot be sure that what we got is
763 * a valid page order. Consider only values in the
764 * valid order range to prevent low_pfn overflow.
766 if (freepage_order
> 0 && freepage_order
< MAX_ORDER
)
767 low_pfn
+= (1UL << freepage_order
) - 1;
772 * Regardless of being on LRU, compound pages such as THP and
773 * hugetlbfs are not to be compacted. We can potentially save
774 * a lot of iterations if we skip them at once. The check is
775 * racy, but we can consider only valid values and the only
776 * danger is skipping too much.
778 if (PageCompound(page
)) {
779 unsigned int comp_order
= compound_order(page
);
781 if (likely(comp_order
< MAX_ORDER
))
782 low_pfn
+= (1UL << comp_order
) - 1;
788 * Check may be lockless but that's ok as we recheck later.
789 * It's possible to migrate LRU and non-lru movable pages.
790 * Skip any other type of page
792 if (!PageLRU(page
)) {
794 * __PageMovable can return false positive so we need
795 * to verify it under page_lock.
797 if (unlikely(__PageMovable(page
)) &&
798 !PageIsolated(page
)) {
800 spin_unlock_irqrestore(&zone
->lru_lock
,
805 if (isolate_movable_page(page
, isolate_mode
))
806 goto isolate_success
;
813 * Migration will fail if an anonymous page is pinned in memory,
814 * so avoid taking lru_lock and isolating it unnecessarily in an
815 * admittedly racy check.
817 if (!page_mapping(page
) &&
818 page_count(page
) > page_mapcount(page
))
821 /* If we already hold the lock, we can skip some rechecking */
823 locked
= compact_trylock_irqsave(&zone
->lru_lock
,
828 /* Recheck PageLRU and PageCompound under lock */
833 * Page become compound since the non-locked check,
834 * and it's on LRU. It can only be a THP so the order
835 * is safe to read and it's 0 for tail pages.
837 if (unlikely(PageCompound(page
))) {
838 low_pfn
+= (1UL << compound_order(page
)) - 1;
843 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
845 /* Try isolate the page */
846 if (__isolate_lru_page(page
, isolate_mode
) != 0)
849 VM_BUG_ON_PAGE(PageCompound(page
), page
);
851 /* Successfully isolated */
852 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
855 list_add(&page
->lru
, &cc
->migratepages
);
856 cc
->nr_migratepages
++;
860 * Record where we could have freed pages by migration and not
861 * yet flushed them to buddy allocator.
862 * - this is the lowest page that was isolated and likely be
863 * then freed by migration.
865 if (!cc
->last_migrated_pfn
)
866 cc
->last_migrated_pfn
= low_pfn
;
868 /* Avoid isolating too much */
869 if (cc
->nr_migratepages
== COMPACT_CLUSTER_MAX
) {
876 if (!skip_on_failure
)
880 * We have isolated some pages, but then failed. Release them
881 * instead of migrating, as we cannot form the cc->order buddy
886 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
889 acct_isolated(zone
, cc
);
890 putback_movable_pages(&cc
->migratepages
);
891 cc
->nr_migratepages
= 0;
892 cc
->last_migrated_pfn
= 0;
896 if (low_pfn
< next_skip_pfn
) {
897 low_pfn
= next_skip_pfn
- 1;
899 * The check near the loop beginning would have updated
900 * next_skip_pfn too, but this is a bit simpler.
902 next_skip_pfn
+= 1UL << cc
->order
;
907 * The PageBuddy() check could have potentially brought us outside
908 * the range to be scanned.
910 if (unlikely(low_pfn
> end_pfn
))
914 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
917 * Update the pageblock-skip information and cached scanner pfn,
918 * if the whole pageblock was scanned without isolating any page.
920 if (low_pfn
== end_pfn
)
921 update_pageblock_skip(cc
, valid_page
, nr_isolated
, true);
923 trace_mm_compaction_isolate_migratepages(start_pfn
, low_pfn
,
924 nr_scanned
, nr_isolated
);
926 count_compact_events(COMPACTMIGRATE_SCANNED
, nr_scanned
);
928 count_compact_events(COMPACTISOLATED
, nr_isolated
);
934 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
935 * @cc: Compaction control structure.
936 * @start_pfn: The first PFN to start isolating.
937 * @end_pfn: The one-past-last PFN.
939 * Returns zero if isolation fails fatally due to e.g. pending signal.
940 * Otherwise, function returns one-past-the-last PFN of isolated page
941 * (which may be greater than end_pfn if end fell in a middle of a THP page).
944 isolate_migratepages_range(struct compact_control
*cc
, unsigned long start_pfn
,
945 unsigned long end_pfn
)
947 unsigned long pfn
, block_start_pfn
, block_end_pfn
;
949 /* Scan block by block. First and last block may be incomplete */
951 block_start_pfn
= pageblock_start_pfn(pfn
);
952 if (block_start_pfn
< cc
->zone
->zone_start_pfn
)
953 block_start_pfn
= cc
->zone
->zone_start_pfn
;
954 block_end_pfn
= pageblock_end_pfn(pfn
);
956 for (; pfn
< end_pfn
; pfn
= block_end_pfn
,
957 block_start_pfn
= block_end_pfn
,
958 block_end_pfn
+= pageblock_nr_pages
) {
960 block_end_pfn
= min(block_end_pfn
, end_pfn
);
962 if (!pageblock_pfn_to_page(block_start_pfn
,
963 block_end_pfn
, cc
->zone
))
966 pfn
= isolate_migratepages_block(cc
, pfn
, block_end_pfn
,
967 ISOLATE_UNEVICTABLE
);
972 if (cc
->nr_migratepages
== COMPACT_CLUSTER_MAX
)
975 acct_isolated(cc
->zone
, cc
);
980 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
981 #ifdef CONFIG_COMPACTION
983 /* Returns true if the page is within a block suitable for migration to */
984 static bool suitable_migration_target(struct page
*page
)
986 /* If the page is a large free page, then disallow migration */
987 if (PageBuddy(page
)) {
989 * We are checking page_order without zone->lock taken. But
990 * the only small danger is that we skip a potentially suitable
991 * pageblock, so it's not worth to check order for valid range.
993 if (page_order_unsafe(page
) >= pageblock_order
)
997 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
998 if (migrate_async_suitable(get_pageblock_migratetype(page
)))
1001 /* Otherwise skip the block */
1006 * Test whether the free scanner has reached the same or lower pageblock than
1007 * the migration scanner, and compaction should thus terminate.
1009 static inline bool compact_scanners_met(struct compact_control
*cc
)
1011 return (cc
->free_pfn
>> pageblock_order
)
1012 <= (cc
->migrate_pfn
>> pageblock_order
);
1016 * Based on information in the current compact_control, find blocks
1017 * suitable for isolating free pages from and then isolate them.
1019 static void isolate_freepages(struct compact_control
*cc
)
1021 struct zone
*zone
= cc
->zone
;
1023 unsigned long block_start_pfn
; /* start of current pageblock */
1024 unsigned long isolate_start_pfn
; /* exact pfn we start at */
1025 unsigned long block_end_pfn
; /* end of current pageblock */
1026 unsigned long low_pfn
; /* lowest pfn scanner is able to scan */
1027 struct list_head
*freelist
= &cc
->freepages
;
1030 * Initialise the free scanner. The starting point is where we last
1031 * successfully isolated from, zone-cached value, or the end of the
1032 * zone when isolating for the first time. For looping we also need
1033 * this pfn aligned down to the pageblock boundary, because we do
1034 * block_start_pfn -= pageblock_nr_pages in the for loop.
1035 * For ending point, take care when isolating in last pageblock of a
1036 * a zone which ends in the middle of a pageblock.
1037 * The low boundary is the end of the pageblock the migration scanner
1040 isolate_start_pfn
= cc
->free_pfn
;
1041 block_start_pfn
= pageblock_start_pfn(cc
->free_pfn
);
1042 block_end_pfn
= min(block_start_pfn
+ pageblock_nr_pages
,
1043 zone_end_pfn(zone
));
1044 low_pfn
= pageblock_end_pfn(cc
->migrate_pfn
);
1047 * Isolate free pages until enough are available to migrate the
1048 * pages on cc->migratepages. We stop searching if the migrate
1049 * and free page scanners meet or enough free pages are isolated.
1051 for (; block_start_pfn
>= low_pfn
;
1052 block_end_pfn
= block_start_pfn
,
1053 block_start_pfn
-= pageblock_nr_pages
,
1054 isolate_start_pfn
= block_start_pfn
) {
1056 * This can iterate a massively long zone without finding any
1057 * suitable migration targets, so periodically check if we need
1058 * to schedule, or even abort async compaction.
1060 if (!(block_start_pfn
% (SWAP_CLUSTER_MAX
* pageblock_nr_pages
))
1061 && compact_should_abort(cc
))
1064 page
= pageblock_pfn_to_page(block_start_pfn
, block_end_pfn
,
1069 /* Check the block is suitable for migration */
1070 if (!suitable_migration_target(page
))
1073 /* If isolation recently failed, do not retry */
1074 if (!isolation_suitable(cc
, page
))
1077 /* Found a block suitable for isolating free pages from. */
1078 isolate_freepages_block(cc
, &isolate_start_pfn
, block_end_pfn
,
1082 * If we isolated enough freepages, or aborted due to lock
1083 * contention, terminate.
1085 if ((cc
->nr_freepages
>= cc
->nr_migratepages
)
1087 if (isolate_start_pfn
>= block_end_pfn
) {
1089 * Restart at previous pageblock if more
1090 * freepages can be isolated next time.
1093 block_start_pfn
- pageblock_nr_pages
;
1096 } else if (isolate_start_pfn
< block_end_pfn
) {
1098 * If isolation failed early, do not continue
1105 /* split_free_page does not map the pages */
1106 map_pages(freelist
);
1109 * Record where the free scanner will restart next time. Either we
1110 * broke from the loop and set isolate_start_pfn based on the last
1111 * call to isolate_freepages_block(), or we met the migration scanner
1112 * and the loop terminated due to isolate_start_pfn < low_pfn
1114 cc
->free_pfn
= isolate_start_pfn
;
1118 * This is a migrate-callback that "allocates" freepages by taking pages
1119 * from the isolated freelists in the block we are migrating to.
1121 static struct page
*compaction_alloc(struct page
*migratepage
,
1125 struct compact_control
*cc
= (struct compact_control
*)data
;
1126 struct page
*freepage
;
1129 * Isolate free pages if necessary, and if we are not aborting due to
1132 if (list_empty(&cc
->freepages
)) {
1134 isolate_freepages(cc
);
1136 if (list_empty(&cc
->freepages
))
1140 freepage
= list_entry(cc
->freepages
.next
, struct page
, lru
);
1141 list_del(&freepage
->lru
);
1148 * This is a migrate-callback that "frees" freepages back to the isolated
1149 * freelist. All pages on the freelist are from the same zone, so there is no
1150 * special handling needed for NUMA.
1152 static void compaction_free(struct page
*page
, unsigned long data
)
1154 struct compact_control
*cc
= (struct compact_control
*)data
;
1156 list_add(&page
->lru
, &cc
->freepages
);
1160 /* possible outcome of isolate_migratepages */
1162 ISOLATE_ABORT
, /* Abort compaction now */
1163 ISOLATE_NONE
, /* No pages isolated, continue scanning */
1164 ISOLATE_SUCCESS
, /* Pages isolated, migrate */
1165 } isolate_migrate_t
;
1168 * Allow userspace to control policy on scanning the unevictable LRU for
1169 * compactable pages.
1171 int sysctl_compact_unevictable_allowed __read_mostly
= 1;
1174 * Isolate all pages that can be migrated from the first suitable block,
1175 * starting at the block pointed to by the migrate scanner pfn within
1178 static isolate_migrate_t
isolate_migratepages(struct zone
*zone
,
1179 struct compact_control
*cc
)
1181 unsigned long block_start_pfn
;
1182 unsigned long block_end_pfn
;
1183 unsigned long low_pfn
;
1185 const isolate_mode_t isolate_mode
=
1186 (sysctl_compact_unevictable_allowed
? ISOLATE_UNEVICTABLE
: 0) |
1187 (cc
->mode
== MIGRATE_ASYNC
? ISOLATE_ASYNC_MIGRATE
: 0);
1190 * Start at where we last stopped, or beginning of the zone as
1191 * initialized by compact_zone()
1193 low_pfn
= cc
->migrate_pfn
;
1194 block_start_pfn
= pageblock_start_pfn(low_pfn
);
1195 if (block_start_pfn
< zone
->zone_start_pfn
)
1196 block_start_pfn
= zone
->zone_start_pfn
;
1198 /* Only scan within a pageblock boundary */
1199 block_end_pfn
= pageblock_end_pfn(low_pfn
);
1202 * Iterate over whole pageblocks until we find the first suitable.
1203 * Do not cross the free scanner.
1205 for (; block_end_pfn
<= cc
->free_pfn
;
1206 low_pfn
= block_end_pfn
,
1207 block_start_pfn
= block_end_pfn
,
1208 block_end_pfn
+= pageblock_nr_pages
) {
1211 * This can potentially iterate a massively long zone with
1212 * many pageblocks unsuitable, so periodically check if we
1213 * need to schedule, or even abort async compaction.
1215 if (!(low_pfn
% (SWAP_CLUSTER_MAX
* pageblock_nr_pages
))
1216 && compact_should_abort(cc
))
1219 page
= pageblock_pfn_to_page(block_start_pfn
, block_end_pfn
,
1224 /* If isolation recently failed, do not retry */
1225 if (!isolation_suitable(cc
, page
))
1229 * For async compaction, also only scan in MOVABLE blocks.
1230 * Async compaction is optimistic to see if the minimum amount
1231 * of work satisfies the allocation.
1233 if (cc
->mode
== MIGRATE_ASYNC
&&
1234 !migrate_async_suitable(get_pageblock_migratetype(page
)))
1237 /* Perform the isolation */
1238 low_pfn
= isolate_migratepages_block(cc
, low_pfn
,
1239 block_end_pfn
, isolate_mode
);
1241 if (!low_pfn
|| cc
->contended
) {
1242 acct_isolated(zone
, cc
);
1243 return ISOLATE_ABORT
;
1247 * Either we isolated something and proceed with migration. Or
1248 * we failed and compact_zone should decide if we should
1254 acct_isolated(zone
, cc
);
1255 /* Record where migration scanner will be restarted. */
1256 cc
->migrate_pfn
= low_pfn
;
1258 return cc
->nr_migratepages
? ISOLATE_SUCCESS
: ISOLATE_NONE
;
1262 * order == -1 is expected when compacting via
1263 * /proc/sys/vm/compact_memory
1265 static inline bool is_via_compact_memory(int order
)
1270 static enum compact_result
__compact_finished(struct zone
*zone
, struct compact_control
*cc
,
1271 const int migratetype
)
1274 unsigned long watermark
;
1276 if (cc
->contended
|| fatal_signal_pending(current
))
1277 return COMPACT_CONTENDED
;
1279 /* Compaction run completes if the migrate and free scanner meet */
1280 if (compact_scanners_met(cc
)) {
1281 /* Let the next compaction start anew. */
1282 reset_cached_positions(zone
);
1285 * Mark that the PG_migrate_skip information should be cleared
1286 * by kswapd when it goes to sleep. kcompactd does not set the
1287 * flag itself as the decision to be clear should be directly
1288 * based on an allocation request.
1290 if (cc
->direct_compaction
)
1291 zone
->compact_blockskip_flush
= true;
1294 return COMPACT_COMPLETE
;
1296 return COMPACT_PARTIAL_SKIPPED
;
1299 if (is_via_compact_memory(cc
->order
))
1300 return COMPACT_CONTINUE
;
1302 /* Compaction run is not finished if the watermark is not met */
1303 watermark
= low_wmark_pages(zone
);
1305 if (!zone_watermark_ok(zone
, cc
->order
, watermark
, cc
->classzone_idx
,
1307 return COMPACT_CONTINUE
;
1309 /* Direct compactor: Is a suitable page free? */
1310 for (order
= cc
->order
; order
< MAX_ORDER
; order
++) {
1311 struct free_area
*area
= &zone
->free_area
[order
];
1314 /* Job done if page is free of the right migratetype */
1315 if (!list_empty(&area
->free_list
[migratetype
]))
1316 return COMPACT_PARTIAL
;
1319 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
1320 if (migratetype
== MIGRATE_MOVABLE
&&
1321 !list_empty(&area
->free_list
[MIGRATE_CMA
]))
1322 return COMPACT_PARTIAL
;
1325 * Job done if allocation would steal freepages from
1326 * other migratetype buddy lists.
1328 if (find_suitable_fallback(area
, order
, migratetype
,
1329 true, &can_steal
) != -1)
1330 return COMPACT_PARTIAL
;
1333 return COMPACT_NO_SUITABLE_PAGE
;
1336 static enum compact_result
compact_finished(struct zone
*zone
,
1337 struct compact_control
*cc
,
1338 const int migratetype
)
1342 ret
= __compact_finished(zone
, cc
, migratetype
);
1343 trace_mm_compaction_finished(zone
, cc
->order
, ret
);
1344 if (ret
== COMPACT_NO_SUITABLE_PAGE
)
1345 ret
= COMPACT_CONTINUE
;
1351 * compaction_suitable: Is this suitable to run compaction on this zone now?
1353 * COMPACT_SKIPPED - If there are too few free pages for compaction
1354 * COMPACT_PARTIAL - If the allocation would succeed without compaction
1355 * COMPACT_CONTINUE - If compaction should run now
1357 static enum compact_result
__compaction_suitable(struct zone
*zone
, int order
,
1358 unsigned int alloc_flags
,
1360 unsigned long wmark_target
)
1363 unsigned long watermark
;
1365 if (is_via_compact_memory(order
))
1366 return COMPACT_CONTINUE
;
1368 watermark
= low_wmark_pages(zone
);
1370 * If watermarks for high-order allocation are already met, there
1371 * should be no need for compaction at all.
1373 if (zone_watermark_ok(zone
, order
, watermark
, classzone_idx
,
1375 return COMPACT_PARTIAL
;
1378 * Watermarks for order-0 must be met for compaction. Note the 2UL.
1379 * This is because during migration, copies of pages need to be
1380 * allocated and for a short time, the footprint is higher
1382 watermark
+= (2UL << order
);
1383 if (!__zone_watermark_ok(zone
, 0, watermark
, classzone_idx
,
1384 alloc_flags
, wmark_target
))
1385 return COMPACT_SKIPPED
;
1388 * fragmentation index determines if allocation failures are due to
1389 * low memory or external fragmentation
1391 * index of -1000 would imply allocations might succeed depending on
1392 * watermarks, but we already failed the high-order watermark check
1393 * index towards 0 implies failure is due to lack of memory
1394 * index towards 1000 implies failure is due to fragmentation
1396 * Only compact if a failure would be due to fragmentation.
1398 fragindex
= fragmentation_index(zone
, order
);
1399 if (fragindex
>= 0 && fragindex
<= sysctl_extfrag_threshold
)
1400 return COMPACT_NOT_SUITABLE_ZONE
;
1402 return COMPACT_CONTINUE
;
1405 enum compact_result
compaction_suitable(struct zone
*zone
, int order
,
1406 unsigned int alloc_flags
,
1409 enum compact_result ret
;
1411 ret
= __compaction_suitable(zone
, order
, alloc_flags
, classzone_idx
,
1412 zone_page_state(zone
, NR_FREE_PAGES
));
1413 trace_mm_compaction_suitable(zone
, order
, ret
);
1414 if (ret
== COMPACT_NOT_SUITABLE_ZONE
)
1415 ret
= COMPACT_SKIPPED
;
1420 bool compaction_zonelist_suitable(struct alloc_context
*ac
, int order
,
1427 * Make sure at least one zone would pass __compaction_suitable if we continue
1428 * retrying the reclaim.
1430 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
1432 unsigned long available
;
1433 enum compact_result compact_result
;
1436 * Do not consider all the reclaimable memory because we do not
1437 * want to trash just for a single high order allocation which
1438 * is even not guaranteed to appear even if __compaction_suitable
1439 * is happy about the watermark check.
1441 available
= zone_reclaimable_pages(zone
) / order
;
1442 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
1443 compact_result
= __compaction_suitable(zone
, order
, alloc_flags
,
1444 ac_classzone_idx(ac
), available
);
1445 if (compact_result
!= COMPACT_SKIPPED
&&
1446 compact_result
!= COMPACT_NOT_SUITABLE_ZONE
)
1453 static enum compact_result
compact_zone(struct zone
*zone
, struct compact_control
*cc
)
1455 enum compact_result ret
;
1456 unsigned long start_pfn
= zone
->zone_start_pfn
;
1457 unsigned long end_pfn
= zone_end_pfn(zone
);
1458 const int migratetype
= gfpflags_to_migratetype(cc
->gfp_mask
);
1459 const bool sync
= cc
->mode
!= MIGRATE_ASYNC
;
1461 ret
= compaction_suitable(zone
, cc
->order
, cc
->alloc_flags
,
1463 /* Compaction is likely to fail */
1464 if (ret
== COMPACT_PARTIAL
|| ret
== COMPACT_SKIPPED
)
1467 /* huh, compaction_suitable is returning something unexpected */
1468 VM_BUG_ON(ret
!= COMPACT_CONTINUE
);
1471 * Clear pageblock skip if there were failures recently and compaction
1472 * is about to be retried after being deferred.
1474 if (compaction_restarting(zone
, cc
->order
))
1475 __reset_isolation_suitable(zone
);
1478 * Setup to move all movable pages to the end of the zone. Used cached
1479 * information on where the scanners should start but check that it
1480 * is initialised by ensuring the values are within zone boundaries.
1482 cc
->migrate_pfn
= zone
->compact_cached_migrate_pfn
[sync
];
1483 cc
->free_pfn
= zone
->compact_cached_free_pfn
;
1484 if (cc
->free_pfn
< start_pfn
|| cc
->free_pfn
>= end_pfn
) {
1485 cc
->free_pfn
= pageblock_start_pfn(end_pfn
- 1);
1486 zone
->compact_cached_free_pfn
= cc
->free_pfn
;
1488 if (cc
->migrate_pfn
< start_pfn
|| cc
->migrate_pfn
>= end_pfn
) {
1489 cc
->migrate_pfn
= start_pfn
;
1490 zone
->compact_cached_migrate_pfn
[0] = cc
->migrate_pfn
;
1491 zone
->compact_cached_migrate_pfn
[1] = cc
->migrate_pfn
;
1494 if (cc
->migrate_pfn
== start_pfn
)
1495 cc
->whole_zone
= true;
1497 cc
->last_migrated_pfn
= 0;
1499 trace_mm_compaction_begin(start_pfn
, cc
->migrate_pfn
,
1500 cc
->free_pfn
, end_pfn
, sync
);
1502 migrate_prep_local();
1504 while ((ret
= compact_finished(zone
, cc
, migratetype
)) ==
1508 switch (isolate_migratepages(zone
, cc
)) {
1510 ret
= COMPACT_CONTENDED
;
1511 putback_movable_pages(&cc
->migratepages
);
1512 cc
->nr_migratepages
= 0;
1516 * We haven't isolated and migrated anything, but
1517 * there might still be unflushed migrations from
1518 * previous cc->order aligned block.
1521 case ISOLATE_SUCCESS
:
1525 err
= migrate_pages(&cc
->migratepages
, compaction_alloc
,
1526 compaction_free
, (unsigned long)cc
, cc
->mode
,
1529 trace_mm_compaction_migratepages(cc
->nr_migratepages
, err
,
1532 /* All pages were either migrated or will be released */
1533 cc
->nr_migratepages
= 0;
1535 putback_movable_pages(&cc
->migratepages
);
1537 * migrate_pages() may return -ENOMEM when scanners meet
1538 * and we want compact_finished() to detect it
1540 if (err
== -ENOMEM
&& !compact_scanners_met(cc
)) {
1541 ret
= COMPACT_CONTENDED
;
1545 * We failed to migrate at least one page in the current
1546 * order-aligned block, so skip the rest of it.
1548 if (cc
->direct_compaction
&&
1549 (cc
->mode
== MIGRATE_ASYNC
)) {
1550 cc
->migrate_pfn
= block_end_pfn(
1551 cc
->migrate_pfn
- 1, cc
->order
);
1552 /* Draining pcplists is useless in this case */
1553 cc
->last_migrated_pfn
= 0;
1560 * Has the migration scanner moved away from the previous
1561 * cc->order aligned block where we migrated from? If yes,
1562 * flush the pages that were freed, so that they can merge and
1563 * compact_finished() can detect immediately if allocation
1566 if (cc
->order
> 0 && cc
->last_migrated_pfn
) {
1568 unsigned long current_block_start
=
1569 block_start_pfn(cc
->migrate_pfn
, cc
->order
);
1571 if (cc
->last_migrated_pfn
< current_block_start
) {
1573 lru_add_drain_cpu(cpu
);
1574 drain_local_pages(zone
);
1576 /* No more flushing until we migrate again */
1577 cc
->last_migrated_pfn
= 0;
1585 * Release free pages and update where the free scanner should restart,
1586 * so we don't leave any returned pages behind in the next attempt.
1588 if (cc
->nr_freepages
> 0) {
1589 unsigned long free_pfn
= release_freepages(&cc
->freepages
);
1591 cc
->nr_freepages
= 0;
1592 VM_BUG_ON(free_pfn
== 0);
1593 /* The cached pfn is always the first in a pageblock */
1594 free_pfn
= pageblock_start_pfn(free_pfn
);
1596 * Only go back, not forward. The cached pfn might have been
1597 * already reset to zone end in compact_finished()
1599 if (free_pfn
> zone
->compact_cached_free_pfn
)
1600 zone
->compact_cached_free_pfn
= free_pfn
;
1603 trace_mm_compaction_end(start_pfn
, cc
->migrate_pfn
,
1604 cc
->free_pfn
, end_pfn
, sync
, ret
);
1606 if (ret
== COMPACT_CONTENDED
)
1607 ret
= COMPACT_PARTIAL
;
1612 static enum compact_result
compact_zone_order(struct zone
*zone
, int order
,
1613 gfp_t gfp_mask
, enum migrate_mode mode
, int *contended
,
1614 unsigned int alloc_flags
, int classzone_idx
)
1616 enum compact_result ret
;
1617 struct compact_control cc
= {
1619 .nr_migratepages
= 0,
1621 .gfp_mask
= gfp_mask
,
1624 .alloc_flags
= alloc_flags
,
1625 .classzone_idx
= classzone_idx
,
1626 .direct_compaction
= true,
1628 INIT_LIST_HEAD(&cc
.freepages
);
1629 INIT_LIST_HEAD(&cc
.migratepages
);
1631 ret
= compact_zone(zone
, &cc
);
1633 VM_BUG_ON(!list_empty(&cc
.freepages
));
1634 VM_BUG_ON(!list_empty(&cc
.migratepages
));
1636 *contended
= cc
.contended
;
1640 int sysctl_extfrag_threshold
= 500;
1643 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1644 * @gfp_mask: The GFP mask of the current allocation
1645 * @order: The order of the current allocation
1646 * @alloc_flags: The allocation flags of the current allocation
1647 * @ac: The context of current allocation
1648 * @mode: The migration mode for async, sync light, or sync migration
1649 * @contended: Return value that determines if compaction was aborted due to
1650 * need_resched() or lock contention
1652 * This is the main entry point for direct page compaction.
1654 enum compact_result
try_to_compact_pages(gfp_t gfp_mask
, unsigned int order
,
1655 unsigned int alloc_flags
, const struct alloc_context
*ac
,
1656 enum migrate_mode mode
, int *contended
)
1658 int may_enter_fs
= gfp_mask
& __GFP_FS
;
1659 int may_perform_io
= gfp_mask
& __GFP_IO
;
1662 enum compact_result rc
= COMPACT_SKIPPED
;
1663 int all_zones_contended
= COMPACT_CONTENDED_LOCK
; /* init for &= op */
1665 *contended
= COMPACT_CONTENDED_NONE
;
1667 /* Check if the GFP flags allow compaction */
1668 if (!order
|| !may_enter_fs
|| !may_perform_io
)
1669 return COMPACT_SKIPPED
;
1671 trace_mm_compaction_try_to_compact_pages(order
, gfp_mask
, mode
);
1673 /* Compact each zone in the list */
1674 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
1676 enum compact_result status
;
1679 if (compaction_deferred(zone
, order
)) {
1680 rc
= max_t(enum compact_result
, COMPACT_DEFERRED
, rc
);
1684 status
= compact_zone_order(zone
, order
, gfp_mask
, mode
,
1685 &zone_contended
, alloc_flags
,
1686 ac_classzone_idx(ac
));
1687 rc
= max(status
, rc
);
1689 * It takes at least one zone that wasn't lock contended
1690 * to clear all_zones_contended.
1692 all_zones_contended
&= zone_contended
;
1694 /* If a normal allocation would succeed, stop compacting */
1695 if (zone_watermark_ok(zone
, order
, low_wmark_pages(zone
),
1696 ac_classzone_idx(ac
), alloc_flags
)) {
1698 * We think the allocation will succeed in this zone,
1699 * but it is not certain, hence the false. The caller
1700 * will repeat this with true if allocation indeed
1701 * succeeds in this zone.
1703 compaction_defer_reset(zone
, order
, false);
1705 * It is possible that async compaction aborted due to
1706 * need_resched() and the watermarks were ok thanks to
1707 * somebody else freeing memory. The allocation can
1708 * however still fail so we better signal the
1709 * need_resched() contention anyway (this will not
1710 * prevent the allocation attempt).
1712 if (zone_contended
== COMPACT_CONTENDED_SCHED
)
1713 *contended
= COMPACT_CONTENDED_SCHED
;
1718 if (mode
!= MIGRATE_ASYNC
&& (status
== COMPACT_COMPLETE
||
1719 status
== COMPACT_PARTIAL_SKIPPED
)) {
1721 * We think that allocation won't succeed in this zone
1722 * so we defer compaction there. If it ends up
1723 * succeeding after all, it will be reset.
1725 defer_compaction(zone
, order
);
1729 * We might have stopped compacting due to need_resched() in
1730 * async compaction, or due to a fatal signal detected. In that
1731 * case do not try further zones and signal need_resched()
1734 if ((zone_contended
== COMPACT_CONTENDED_SCHED
)
1735 || fatal_signal_pending(current
)) {
1736 *contended
= COMPACT_CONTENDED_SCHED
;
1743 * We might not have tried all the zones, so be conservative
1744 * and assume they are not all lock contended.
1746 all_zones_contended
= 0;
1751 * If at least one zone wasn't deferred or skipped, we report if all
1752 * zones that were tried were lock contended.
1754 if (rc
> COMPACT_INACTIVE
&& all_zones_contended
)
1755 *contended
= COMPACT_CONTENDED_LOCK
;
1761 /* Compact all zones within a node */
1762 static void __compact_pgdat(pg_data_t
*pgdat
, struct compact_control
*cc
)
1767 for (zoneid
= 0; zoneid
< MAX_NR_ZONES
; zoneid
++) {
1769 zone
= &pgdat
->node_zones
[zoneid
];
1770 if (!populated_zone(zone
))
1773 cc
->nr_freepages
= 0;
1774 cc
->nr_migratepages
= 0;
1776 INIT_LIST_HEAD(&cc
->freepages
);
1777 INIT_LIST_HEAD(&cc
->migratepages
);
1780 * When called via /proc/sys/vm/compact_memory
1781 * this makes sure we compact the whole zone regardless of
1782 * cached scanner positions.
1784 if (is_via_compact_memory(cc
->order
))
1785 __reset_isolation_suitable(zone
);
1787 if (is_via_compact_memory(cc
->order
) ||
1788 !compaction_deferred(zone
, cc
->order
))
1789 compact_zone(zone
, cc
);
1791 VM_BUG_ON(!list_empty(&cc
->freepages
));
1792 VM_BUG_ON(!list_empty(&cc
->migratepages
));
1794 if (is_via_compact_memory(cc
->order
))
1797 if (zone_watermark_ok(zone
, cc
->order
,
1798 low_wmark_pages(zone
), 0, 0))
1799 compaction_defer_reset(zone
, cc
->order
, false);
1803 void compact_pgdat(pg_data_t
*pgdat
, int order
)
1805 struct compact_control cc
= {
1807 .mode
= MIGRATE_ASYNC
,
1813 __compact_pgdat(pgdat
, &cc
);
1816 static void compact_node(int nid
)
1818 struct compact_control cc
= {
1820 .mode
= MIGRATE_SYNC
,
1821 .ignore_skip_hint
= true,
1824 __compact_pgdat(NODE_DATA(nid
), &cc
);
1827 /* Compact all nodes in the system */
1828 static void compact_nodes(void)
1832 /* Flush pending updates to the LRU lists */
1833 lru_add_drain_all();
1835 for_each_online_node(nid
)
1839 /* The written value is actually unused, all memory is compacted */
1840 int sysctl_compact_memory
;
1843 * This is the entry point for compacting all nodes via
1844 * /proc/sys/vm/compact_memory
1846 int sysctl_compaction_handler(struct ctl_table
*table
, int write
,
1847 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
1855 int sysctl_extfrag_handler(struct ctl_table
*table
, int write
,
1856 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
1858 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
1863 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1864 static ssize_t
sysfs_compact_node(struct device
*dev
,
1865 struct device_attribute
*attr
,
1866 const char *buf
, size_t count
)
1870 if (nid
>= 0 && nid
< nr_node_ids
&& node_online(nid
)) {
1871 /* Flush pending updates to the LRU lists */
1872 lru_add_drain_all();
1879 static DEVICE_ATTR(compact
, S_IWUSR
, NULL
, sysfs_compact_node
);
1881 int compaction_register_node(struct node
*node
)
1883 return device_create_file(&node
->dev
, &dev_attr_compact
);
1886 void compaction_unregister_node(struct node
*node
)
1888 return device_remove_file(&node
->dev
, &dev_attr_compact
);
1890 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1892 static inline bool kcompactd_work_requested(pg_data_t
*pgdat
)
1894 return pgdat
->kcompactd_max_order
> 0 || kthread_should_stop();
1897 static bool kcompactd_node_suitable(pg_data_t
*pgdat
)
1901 enum zone_type classzone_idx
= pgdat
->kcompactd_classzone_idx
;
1903 for (zoneid
= 0; zoneid
<= classzone_idx
; zoneid
++) {
1904 zone
= &pgdat
->node_zones
[zoneid
];
1906 if (!populated_zone(zone
))
1909 if (compaction_suitable(zone
, pgdat
->kcompactd_max_order
, 0,
1910 classzone_idx
) == COMPACT_CONTINUE
)
1917 static void kcompactd_do_work(pg_data_t
*pgdat
)
1920 * With no special task, compact all zones so that a page of requested
1921 * order is allocatable.
1925 struct compact_control cc
= {
1926 .order
= pgdat
->kcompactd_max_order
,
1927 .classzone_idx
= pgdat
->kcompactd_classzone_idx
,
1928 .mode
= MIGRATE_SYNC_LIGHT
,
1929 .ignore_skip_hint
= true,
1932 bool success
= false;
1934 trace_mm_compaction_kcompactd_wake(pgdat
->node_id
, cc
.order
,
1936 count_vm_event(KCOMPACTD_WAKE
);
1938 for (zoneid
= 0; zoneid
<= cc
.classzone_idx
; zoneid
++) {
1941 zone
= &pgdat
->node_zones
[zoneid
];
1942 if (!populated_zone(zone
))
1945 if (compaction_deferred(zone
, cc
.order
))
1948 if (compaction_suitable(zone
, cc
.order
, 0, zoneid
) !=
1952 cc
.nr_freepages
= 0;
1953 cc
.nr_migratepages
= 0;
1955 INIT_LIST_HEAD(&cc
.freepages
);
1956 INIT_LIST_HEAD(&cc
.migratepages
);
1958 if (kthread_should_stop())
1960 status
= compact_zone(zone
, &cc
);
1962 if (zone_watermark_ok(zone
, cc
.order
, low_wmark_pages(zone
),
1963 cc
.classzone_idx
, 0)) {
1965 compaction_defer_reset(zone
, cc
.order
, false);
1966 } else if (status
== COMPACT_PARTIAL_SKIPPED
|| status
== COMPACT_COMPLETE
) {
1968 * We use sync migration mode here, so we defer like
1969 * sync direct compaction does.
1971 defer_compaction(zone
, cc
.order
);
1974 VM_BUG_ON(!list_empty(&cc
.freepages
));
1975 VM_BUG_ON(!list_empty(&cc
.migratepages
));
1979 * Regardless of success, we are done until woken up next. But remember
1980 * the requested order/classzone_idx in case it was higher/tighter than
1983 if (pgdat
->kcompactd_max_order
<= cc
.order
)
1984 pgdat
->kcompactd_max_order
= 0;
1985 if (pgdat
->kcompactd_classzone_idx
>= cc
.classzone_idx
)
1986 pgdat
->kcompactd_classzone_idx
= pgdat
->nr_zones
- 1;
1989 void wakeup_kcompactd(pg_data_t
*pgdat
, int order
, int classzone_idx
)
1994 if (pgdat
->kcompactd_max_order
< order
)
1995 pgdat
->kcompactd_max_order
= order
;
1997 if (pgdat
->kcompactd_classzone_idx
> classzone_idx
)
1998 pgdat
->kcompactd_classzone_idx
= classzone_idx
;
2000 if (!waitqueue_active(&pgdat
->kcompactd_wait
))
2003 if (!kcompactd_node_suitable(pgdat
))
2006 trace_mm_compaction_wakeup_kcompactd(pgdat
->node_id
, order
,
2008 wake_up_interruptible(&pgdat
->kcompactd_wait
);
2012 * The background compaction daemon, started as a kernel thread
2013 * from the init process.
2015 static int kcompactd(void *p
)
2017 pg_data_t
*pgdat
= (pg_data_t
*)p
;
2018 struct task_struct
*tsk
= current
;
2020 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
2022 if (!cpumask_empty(cpumask
))
2023 set_cpus_allowed_ptr(tsk
, cpumask
);
2027 pgdat
->kcompactd_max_order
= 0;
2028 pgdat
->kcompactd_classzone_idx
= pgdat
->nr_zones
- 1;
2030 while (!kthread_should_stop()) {
2031 trace_mm_compaction_kcompactd_sleep(pgdat
->node_id
);
2032 wait_event_freezable(pgdat
->kcompactd_wait
,
2033 kcompactd_work_requested(pgdat
));
2035 kcompactd_do_work(pgdat
);
2042 * This kcompactd start function will be called by init and node-hot-add.
2043 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
2045 int kcompactd_run(int nid
)
2047 pg_data_t
*pgdat
= NODE_DATA(nid
);
2050 if (pgdat
->kcompactd
)
2053 pgdat
->kcompactd
= kthread_run(kcompactd
, pgdat
, "kcompactd%d", nid
);
2054 if (IS_ERR(pgdat
->kcompactd
)) {
2055 pr_err("Failed to start kcompactd on node %d\n", nid
);
2056 ret
= PTR_ERR(pgdat
->kcompactd
);
2057 pgdat
->kcompactd
= NULL
;
2063 * Called by memory hotplug when all memory in a node is offlined. Caller must
2064 * hold mem_hotplug_begin/end().
2066 void kcompactd_stop(int nid
)
2068 struct task_struct
*kcompactd
= NODE_DATA(nid
)->kcompactd
;
2071 kthread_stop(kcompactd
);
2072 NODE_DATA(nid
)->kcompactd
= NULL
;
2077 * It's optimal to keep kcompactd on the same CPUs as their memory, but
2078 * not required for correctness. So if the last cpu in a node goes
2079 * away, we get changed to run anywhere: as the first one comes back,
2080 * restore their cpu bindings.
2082 static int cpu_callback(struct notifier_block
*nfb
, unsigned long action
,
2087 if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
) {
2088 for_each_node_state(nid
, N_MEMORY
) {
2089 pg_data_t
*pgdat
= NODE_DATA(nid
);
2090 const struct cpumask
*mask
;
2092 mask
= cpumask_of_node(pgdat
->node_id
);
2094 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
2095 /* One of our CPUs online: restore mask */
2096 set_cpus_allowed_ptr(pgdat
->kcompactd
, mask
);
2102 static int __init
kcompactd_init(void)
2106 for_each_node_state(nid
, N_MEMORY
)
2108 hotcpu_notifier(cpu_callback
, 0);
2111 subsys_initcall(kcompactd_init
)
2113 #endif /* CONFIG_COMPACTION */