1 // SPDX-License-Identifier: GPL-2.0
3 * linux/mm/compaction.c
5 * Memory compaction for the reduction of external fragmentation. Note that
6 * this heavily depends upon page migration to do all the real heavy
9 * Copyright IBM Corp. 2007-2010 Mel Gorman <mel@csn.ul.ie>
11 #include <linux/cpu.h>
12 #include <linux/swap.h>
13 #include <linux/migrate.h>
14 #include <linux/compaction.h>
15 #include <linux/mm_inline.h>
16 #include <linux/sched/signal.h>
17 #include <linux/backing-dev.h>
18 #include <linux/sysctl.h>
19 #include <linux/sysfs.h>
20 #include <linux/page-isolation.h>
21 #include <linux/kasan.h>
22 #include <linux/kthread.h>
23 #include <linux/freezer.h>
24 #include <linux/page_owner.h>
27 #ifdef CONFIG_COMPACTION
28 static inline void count_compact_event(enum vm_event_item item
)
33 static inline void count_compact_events(enum vm_event_item item
, long delta
)
35 count_vm_events(item
, delta
);
38 #define count_compact_event(item) do { } while (0)
39 #define count_compact_events(item, delta) do { } while (0)
42 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
44 #define CREATE_TRACE_POINTS
45 #include <trace/events/compaction.h>
47 #define block_start_pfn(pfn, order) round_down(pfn, 1UL << (order))
48 #define block_end_pfn(pfn, order) ALIGN((pfn) + 1, 1UL << (order))
49 #define pageblock_start_pfn(pfn) block_start_pfn(pfn, pageblock_order)
50 #define pageblock_end_pfn(pfn) block_end_pfn(pfn, pageblock_order)
52 static unsigned long release_freepages(struct list_head
*freelist
)
54 struct page
*page
, *next
;
55 unsigned long high_pfn
= 0;
57 list_for_each_entry_safe(page
, next
, freelist
, lru
) {
58 unsigned long pfn
= page_to_pfn(page
);
68 static void map_pages(struct list_head
*list
)
70 unsigned int i
, order
, nr_pages
;
71 struct page
*page
, *next
;
74 list_for_each_entry_safe(page
, next
, list
, lru
) {
77 order
= page_private(page
);
78 nr_pages
= 1 << order
;
80 post_alloc_hook(page
, order
, __GFP_MOVABLE
);
82 split_page(page
, order
);
84 for (i
= 0; i
< nr_pages
; i
++) {
85 list_add(&page
->lru
, &tmp_list
);
90 list_splice(&tmp_list
, list
);
93 #ifdef CONFIG_COMPACTION
95 int PageMovable(struct page
*page
)
97 struct address_space
*mapping
;
99 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
100 if (!__PageMovable(page
))
103 mapping
= page_mapping(page
);
104 if (mapping
&& mapping
->a_ops
&& mapping
->a_ops
->isolate_page
)
109 EXPORT_SYMBOL(PageMovable
);
111 void __SetPageMovable(struct page
*page
, struct address_space
*mapping
)
113 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
114 VM_BUG_ON_PAGE((unsigned long)mapping
& PAGE_MAPPING_MOVABLE
, page
);
115 page
->mapping
= (void *)((unsigned long)mapping
| PAGE_MAPPING_MOVABLE
);
117 EXPORT_SYMBOL(__SetPageMovable
);
119 void __ClearPageMovable(struct page
*page
)
121 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
122 VM_BUG_ON_PAGE(!PageMovable(page
), page
);
124 * Clear registered address_space val with keeping PAGE_MAPPING_MOVABLE
125 * flag so that VM can catch up released page by driver after isolation.
126 * With it, VM migration doesn't try to put it back.
128 page
->mapping
= (void *)((unsigned long)page
->mapping
&
129 PAGE_MAPPING_MOVABLE
);
131 EXPORT_SYMBOL(__ClearPageMovable
);
133 /* Do not skip compaction more than 64 times */
134 #define COMPACT_MAX_DEFER_SHIFT 6
137 * Compaction is deferred when compaction fails to result in a page
138 * allocation success. 1 << compact_defer_limit compactions are skipped up
139 * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
141 void defer_compaction(struct zone
*zone
, int order
)
143 zone
->compact_considered
= 0;
144 zone
->compact_defer_shift
++;
146 if (order
< zone
->compact_order_failed
)
147 zone
->compact_order_failed
= order
;
149 if (zone
->compact_defer_shift
> COMPACT_MAX_DEFER_SHIFT
)
150 zone
->compact_defer_shift
= COMPACT_MAX_DEFER_SHIFT
;
152 trace_mm_compaction_defer_compaction(zone
, order
);
155 /* Returns true if compaction should be skipped this time */
156 bool compaction_deferred(struct zone
*zone
, int order
)
158 unsigned long defer_limit
= 1UL << zone
->compact_defer_shift
;
160 if (order
< zone
->compact_order_failed
)
163 /* Avoid possible overflow */
164 if (++zone
->compact_considered
> defer_limit
)
165 zone
->compact_considered
= defer_limit
;
167 if (zone
->compact_considered
>= defer_limit
)
170 trace_mm_compaction_deferred(zone
, order
);
176 * Update defer tracking counters after successful compaction of given order,
177 * which means an allocation either succeeded (alloc_success == true) or is
178 * expected to succeed.
180 void compaction_defer_reset(struct zone
*zone
, int order
,
184 zone
->compact_considered
= 0;
185 zone
->compact_defer_shift
= 0;
187 if (order
>= zone
->compact_order_failed
)
188 zone
->compact_order_failed
= order
+ 1;
190 trace_mm_compaction_defer_reset(zone
, order
);
193 /* Returns true if restarting compaction after many failures */
194 bool compaction_restarting(struct zone
*zone
, int order
)
196 if (order
< zone
->compact_order_failed
)
199 return zone
->compact_defer_shift
== COMPACT_MAX_DEFER_SHIFT
&&
200 zone
->compact_considered
>= 1UL << zone
->compact_defer_shift
;
203 /* Returns true if the pageblock should be scanned for pages to isolate. */
204 static inline bool isolation_suitable(struct compact_control
*cc
,
207 if (cc
->ignore_skip_hint
)
210 return !get_pageblock_skip(page
);
213 static void reset_cached_positions(struct zone
*zone
)
215 zone
->compact_cached_migrate_pfn
[0] = zone
->zone_start_pfn
;
216 zone
->compact_cached_migrate_pfn
[1] = zone
->zone_start_pfn
;
217 zone
->compact_cached_free_pfn
=
218 pageblock_start_pfn(zone_end_pfn(zone
) - 1);
222 * Compound pages of >= pageblock_order should consistenly be skipped until
223 * released. It is always pointless to compact pages of such order (if they are
224 * migratable), and the pageblocks they occupy cannot contain any free pages.
226 static bool pageblock_skip_persistent(struct page
*page
)
228 if (!PageCompound(page
))
231 page
= compound_head(page
);
233 if (compound_order(page
) >= pageblock_order
)
240 * This function is called to clear all cached information on pageblocks that
241 * should be skipped for page isolation when the migrate and free page scanner
244 static void __reset_isolation_suitable(struct zone
*zone
)
246 unsigned long start_pfn
= zone
->zone_start_pfn
;
247 unsigned long end_pfn
= zone_end_pfn(zone
);
250 zone
->compact_blockskip_flush
= false;
252 /* Walk the zone and mark every pageblock as suitable for isolation */
253 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
258 page
= pfn_to_online_page(pfn
);
261 if (zone
!= page_zone(page
))
263 if (pageblock_skip_persistent(page
))
266 clear_pageblock_skip(page
);
269 reset_cached_positions(zone
);
272 void reset_isolation_suitable(pg_data_t
*pgdat
)
276 for (zoneid
= 0; zoneid
< MAX_NR_ZONES
; zoneid
++) {
277 struct zone
*zone
= &pgdat
->node_zones
[zoneid
];
278 if (!populated_zone(zone
))
281 /* Only flush if a full compaction finished recently */
282 if (zone
->compact_blockskip_flush
)
283 __reset_isolation_suitable(zone
);
288 * If no pages were isolated then mark this pageblock to be skipped in the
289 * future. The information is later cleared by __reset_isolation_suitable().
291 static void update_pageblock_skip(struct compact_control
*cc
,
292 struct page
*page
, unsigned long nr_isolated
,
293 bool migrate_scanner
)
295 struct zone
*zone
= cc
->zone
;
298 if (cc
->ignore_skip_hint
)
307 set_pageblock_skip(page
);
309 pfn
= page_to_pfn(page
);
311 /* Update where async and sync compaction should restart */
312 if (migrate_scanner
) {
313 if (pfn
> zone
->compact_cached_migrate_pfn
[0])
314 zone
->compact_cached_migrate_pfn
[0] = pfn
;
315 if (cc
->mode
!= MIGRATE_ASYNC
&&
316 pfn
> zone
->compact_cached_migrate_pfn
[1])
317 zone
->compact_cached_migrate_pfn
[1] = pfn
;
319 if (pfn
< zone
->compact_cached_free_pfn
)
320 zone
->compact_cached_free_pfn
= pfn
;
324 static inline bool isolation_suitable(struct compact_control
*cc
,
330 static inline bool pageblock_skip_persistent(struct page
*page
)
335 static inline void update_pageblock_skip(struct compact_control
*cc
,
336 struct page
*page
, unsigned long nr_isolated
,
337 bool migrate_scanner
)
340 #endif /* CONFIG_COMPACTION */
343 * Compaction requires the taking of some coarse locks that are potentially
344 * very heavily contended. For async compaction, back out if the lock cannot
345 * be taken immediately. For sync compaction, spin on the lock if needed.
347 * Returns true if the lock is held
348 * Returns false if the lock is not held and compaction should abort
350 static bool compact_trylock_irqsave(spinlock_t
*lock
, unsigned long *flags
,
351 struct compact_control
*cc
)
353 if (cc
->mode
== MIGRATE_ASYNC
) {
354 if (!spin_trylock_irqsave(lock
, *flags
)) {
355 cc
->contended
= true;
359 spin_lock_irqsave(lock
, *flags
);
366 * Compaction requires the taking of some coarse locks that are potentially
367 * very heavily contended. The lock should be periodically unlocked to avoid
368 * having disabled IRQs for a long time, even when there is nobody waiting on
369 * the lock. It might also be that allowing the IRQs will result in
370 * need_resched() becoming true. If scheduling is needed, async compaction
371 * aborts. Sync compaction schedules.
372 * Either compaction type will also abort if a fatal signal is pending.
373 * In either case if the lock was locked, it is dropped and not regained.
375 * Returns true if compaction should abort due to fatal signal pending, or
376 * async compaction due to need_resched()
377 * Returns false when compaction can continue (sync compaction might have
380 static bool compact_unlock_should_abort(spinlock_t
*lock
,
381 unsigned long flags
, bool *locked
, struct compact_control
*cc
)
384 spin_unlock_irqrestore(lock
, flags
);
388 if (fatal_signal_pending(current
)) {
389 cc
->contended
= true;
393 if (need_resched()) {
394 if (cc
->mode
== MIGRATE_ASYNC
) {
395 cc
->contended
= true;
405 * Aside from avoiding lock contention, compaction also periodically checks
406 * need_resched() and either schedules in sync compaction or aborts async
407 * compaction. This is similar to what compact_unlock_should_abort() does, but
408 * is used where no lock is concerned.
410 * Returns false when no scheduling was needed, or sync compaction scheduled.
411 * Returns true when async compaction should abort.
413 static inline bool compact_should_abort(struct compact_control
*cc
)
415 /* async compaction aborts if contended */
416 if (need_resched()) {
417 if (cc
->mode
== MIGRATE_ASYNC
) {
418 cc
->contended
= true;
429 * Isolate free pages onto a private freelist. If @strict is true, will abort
430 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
431 * (even though it may still end up isolating some pages).
433 static unsigned long isolate_freepages_block(struct compact_control
*cc
,
434 unsigned long *start_pfn
,
435 unsigned long end_pfn
,
436 struct list_head
*freelist
,
439 int nr_scanned
= 0, total_isolated
= 0;
440 struct page
*cursor
, *valid_page
= NULL
;
441 unsigned long flags
= 0;
443 unsigned long blockpfn
= *start_pfn
;
446 cursor
= pfn_to_page(blockpfn
);
448 /* Isolate free pages. */
449 for (; blockpfn
< end_pfn
; blockpfn
++, cursor
++) {
451 struct page
*page
= cursor
;
454 * Periodically drop the lock (if held) regardless of its
455 * contention, to give chance to IRQs. Abort if fatal signal
456 * pending or async compaction detects need_resched()
458 if (!(blockpfn
% SWAP_CLUSTER_MAX
)
459 && compact_unlock_should_abort(&cc
->zone
->lock
, flags
,
464 if (!pfn_valid_within(blockpfn
))
471 * For compound pages such as THP and hugetlbfs, we can save
472 * potentially a lot of iterations if we skip them at once.
473 * The check is racy, but we can consider only valid values
474 * and the only danger is skipping too much.
476 if (PageCompound(page
)) {
477 const unsigned int order
= compound_order(page
);
479 if (pageblock_skip_persistent(page
, order
)) {
480 set_pageblock_skip(page
);
482 } else if (likely(order
< MAX_ORDER
)) {
483 blockpfn
+= (1UL << order
) - 1;
484 cursor
+= (1UL << order
) - 1;
489 if (!PageBuddy(page
))
493 * If we already hold the lock, we can skip some rechecking.
494 * Note that if we hold the lock now, checked_pageblock was
495 * already set in some previous iteration (or strict is true),
496 * so it is correct to skip the suitable migration target
501 * The zone lock must be held to isolate freepages.
502 * Unfortunately this is a very coarse lock and can be
503 * heavily contended if there are parallel allocations
504 * or parallel compactions. For async compaction do not
505 * spin on the lock and we acquire the lock as late as
508 locked
= compact_trylock_irqsave(&cc
->zone
->lock
,
513 /* Recheck this is a buddy page under lock */
514 if (!PageBuddy(page
))
518 /* Found a free page, will break it into order-0 pages */
519 order
= page_order(page
);
520 isolated
= __isolate_free_page(page
, order
);
523 set_page_private(page
, order
);
525 total_isolated
+= isolated
;
526 cc
->nr_freepages
+= isolated
;
527 list_add_tail(&page
->lru
, freelist
);
529 if (!strict
&& cc
->nr_migratepages
<= cc
->nr_freepages
) {
530 blockpfn
+= isolated
;
533 /* Advance to the end of split page */
534 blockpfn
+= isolated
- 1;
535 cursor
+= isolated
- 1;
547 spin_unlock_irqrestore(&cc
->zone
->lock
, flags
);
550 * There is a tiny chance that we have read bogus compound_order(),
551 * so be careful to not go outside of the pageblock.
553 if (unlikely(blockpfn
> end_pfn
))
556 trace_mm_compaction_isolate_freepages(*start_pfn
, blockpfn
,
557 nr_scanned
, total_isolated
);
559 /* Record how far we have got within the block */
560 *start_pfn
= blockpfn
;
563 * If strict isolation is requested by CMA then check that all the
564 * pages requested were isolated. If there were any failures, 0 is
565 * returned and CMA will fail.
567 if (strict
&& blockpfn
< end_pfn
)
570 /* Update the pageblock-skip if the whole pageblock was scanned */
571 if (blockpfn
== end_pfn
)
572 update_pageblock_skip(cc
, valid_page
, total_isolated
, false);
574 cc
->total_free_scanned
+= nr_scanned
;
576 count_compact_events(COMPACTISOLATED
, total_isolated
);
577 return total_isolated
;
581 * isolate_freepages_range() - isolate free pages.
582 * @start_pfn: The first PFN to start isolating.
583 * @end_pfn: The one-past-last PFN.
585 * Non-free pages, invalid PFNs, or zone boundaries within the
586 * [start_pfn, end_pfn) range are considered errors, cause function to
587 * undo its actions and return zero.
589 * Otherwise, function returns one-past-the-last PFN of isolated page
590 * (which may be greater then end_pfn if end fell in a middle of
594 isolate_freepages_range(struct compact_control
*cc
,
595 unsigned long start_pfn
, unsigned long end_pfn
)
597 unsigned long isolated
, pfn
, block_start_pfn
, block_end_pfn
;
601 block_start_pfn
= pageblock_start_pfn(pfn
);
602 if (block_start_pfn
< cc
->zone
->zone_start_pfn
)
603 block_start_pfn
= cc
->zone
->zone_start_pfn
;
604 block_end_pfn
= pageblock_end_pfn(pfn
);
606 for (; pfn
< end_pfn
; pfn
+= isolated
,
607 block_start_pfn
= block_end_pfn
,
608 block_end_pfn
+= pageblock_nr_pages
) {
609 /* Protect pfn from changing by isolate_freepages_block */
610 unsigned long isolate_start_pfn
= pfn
;
612 block_end_pfn
= min(block_end_pfn
, end_pfn
);
615 * pfn could pass the block_end_pfn if isolated freepage
616 * is more than pageblock order. In this case, we adjust
617 * scanning range to right one.
619 if (pfn
>= block_end_pfn
) {
620 block_start_pfn
= pageblock_start_pfn(pfn
);
621 block_end_pfn
= pageblock_end_pfn(pfn
);
622 block_end_pfn
= min(block_end_pfn
, end_pfn
);
625 if (!pageblock_pfn_to_page(block_start_pfn
,
626 block_end_pfn
, cc
->zone
))
629 isolated
= isolate_freepages_block(cc
, &isolate_start_pfn
,
630 block_end_pfn
, &freelist
, true);
633 * In strict mode, isolate_freepages_block() returns 0 if
634 * there are any holes in the block (ie. invalid PFNs or
641 * If we managed to isolate pages, it is always (1 << n) *
642 * pageblock_nr_pages for some non-negative n. (Max order
643 * page may span two pageblocks).
647 /* __isolate_free_page() does not map the pages */
648 map_pages(&freelist
);
651 /* Loop terminated early, cleanup. */
652 release_freepages(&freelist
);
656 /* We don't use freelists for anything. */
660 /* Similar to reclaim, but different enough that they don't share logic */
661 static bool too_many_isolated(struct zone
*zone
)
663 unsigned long active
, inactive
, isolated
;
665 inactive
= node_page_state(zone
->zone_pgdat
, NR_INACTIVE_FILE
) +
666 node_page_state(zone
->zone_pgdat
, NR_INACTIVE_ANON
);
667 active
= node_page_state(zone
->zone_pgdat
, NR_ACTIVE_FILE
) +
668 node_page_state(zone
->zone_pgdat
, NR_ACTIVE_ANON
);
669 isolated
= node_page_state(zone
->zone_pgdat
, NR_ISOLATED_FILE
) +
670 node_page_state(zone
->zone_pgdat
, NR_ISOLATED_ANON
);
672 return isolated
> (inactive
+ active
) / 2;
676 * isolate_migratepages_block() - isolate all migrate-able pages within
678 * @cc: Compaction control structure.
679 * @low_pfn: The first PFN to isolate
680 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
681 * @isolate_mode: Isolation mode to be used.
683 * Isolate all pages that can be migrated from the range specified by
684 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
685 * Returns zero if there is a fatal signal pending, otherwise PFN of the
686 * first page that was not scanned (which may be both less, equal to or more
689 * The pages are isolated on cc->migratepages list (not required to be empty),
690 * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
691 * is neither read nor updated.
694 isolate_migratepages_block(struct compact_control
*cc
, unsigned long low_pfn
,
695 unsigned long end_pfn
, isolate_mode_t isolate_mode
)
697 struct zone
*zone
= cc
->zone
;
698 unsigned long nr_scanned
= 0, nr_isolated
= 0;
699 struct lruvec
*lruvec
;
700 unsigned long flags
= 0;
702 struct page
*page
= NULL
, *valid_page
= NULL
;
703 unsigned long start_pfn
= low_pfn
;
704 bool skip_on_failure
= false;
705 unsigned long next_skip_pfn
= 0;
708 * Ensure that there are not too many pages isolated from the LRU
709 * list by either parallel reclaimers or compaction. If there are,
710 * delay for some time until fewer pages are isolated
712 while (unlikely(too_many_isolated(zone
))) {
713 /* async migration should just abort */
714 if (cc
->mode
== MIGRATE_ASYNC
)
717 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
719 if (fatal_signal_pending(current
))
723 if (compact_should_abort(cc
))
726 if (cc
->direct_compaction
&& (cc
->mode
== MIGRATE_ASYNC
)) {
727 skip_on_failure
= true;
728 next_skip_pfn
= block_end_pfn(low_pfn
, cc
->order
);
731 /* Time to isolate some pages for migration */
732 for (; low_pfn
< end_pfn
; low_pfn
++) {
734 if (skip_on_failure
&& low_pfn
>= next_skip_pfn
) {
736 * We have isolated all migration candidates in the
737 * previous order-aligned block, and did not skip it due
738 * to failure. We should migrate the pages now and
739 * hopefully succeed compaction.
745 * We failed to isolate in the previous order-aligned
746 * block. Set the new boundary to the end of the
747 * current block. Note we can't simply increase
748 * next_skip_pfn by 1 << order, as low_pfn might have
749 * been incremented by a higher number due to skipping
750 * a compound or a high-order buddy page in the
751 * previous loop iteration.
753 next_skip_pfn
= block_end_pfn(low_pfn
, cc
->order
);
757 * Periodically drop the lock (if held) regardless of its
758 * contention, to give chance to IRQs. Abort async compaction
761 if (!(low_pfn
% SWAP_CLUSTER_MAX
)
762 && compact_unlock_should_abort(zone_lru_lock(zone
), flags
,
766 if (!pfn_valid_within(low_pfn
))
770 page
= pfn_to_page(low_pfn
);
776 * Skip if free. We read page order here without zone lock
777 * which is generally unsafe, but the race window is small and
778 * the worst thing that can happen is that we skip some
779 * potential isolation targets.
781 if (PageBuddy(page
)) {
782 unsigned long freepage_order
= page_order_unsafe(page
);
785 * Without lock, we cannot be sure that what we got is
786 * a valid page order. Consider only values in the
787 * valid order range to prevent low_pfn overflow.
789 if (freepage_order
> 0 && freepage_order
< MAX_ORDER
)
790 low_pfn
+= (1UL << freepage_order
) - 1;
795 * Regardless of being on LRU, compound pages such as THP and
796 * hugetlbfs are not to be compacted. We can potentially save
797 * a lot of iterations if we skip them at once. The check is
798 * racy, but we can consider only valid values and the only
799 * danger is skipping too much.
801 if (PageCompound(page
)) {
802 const unsigned int order
= compound_order(page
);
804 if (pageblock_skip_persistent(page
, order
)) {
805 set_pageblock_skip(page
);
807 } else if (likely(order
< MAX_ORDER
))
808 low_pfn
+= (1UL << order
) - 1;
813 * Check may be lockless but that's ok as we recheck later.
814 * It's possible to migrate LRU and non-lru movable pages.
815 * Skip any other type of page
817 if (!PageLRU(page
)) {
819 * __PageMovable can return false positive so we need
820 * to verify it under page_lock.
822 if (unlikely(__PageMovable(page
)) &&
823 !PageIsolated(page
)) {
825 spin_unlock_irqrestore(zone_lru_lock(zone
),
830 if (!isolate_movable_page(page
, isolate_mode
))
831 goto isolate_success
;
838 * Migration will fail if an anonymous page is pinned in memory,
839 * so avoid taking lru_lock and isolating it unnecessarily in an
840 * admittedly racy check.
842 if (!page_mapping(page
) &&
843 page_count(page
) > page_mapcount(page
))
847 * Only allow to migrate anonymous pages in GFP_NOFS context
848 * because those do not depend on fs locks.
850 if (!(cc
->gfp_mask
& __GFP_FS
) && page_mapping(page
))
853 /* If we already hold the lock, we can skip some rechecking */
855 locked
= compact_trylock_irqsave(zone_lru_lock(zone
),
860 /* Recheck PageLRU and PageCompound under lock */
865 * Page become compound since the non-locked check,
866 * and it's on LRU. It can only be a THP so the order
867 * is safe to read and it's 0 for tail pages.
869 if (unlikely(PageCompound(page
))) {
870 const unsigned int order
= compound_order(page
);
872 if (pageblock_skip_persistent(page
, order
)) {
873 set_pageblock_skip(page
);
876 low_pfn
+= (1UL << order
) - 1;
881 lruvec
= mem_cgroup_page_lruvec(page
, zone
->zone_pgdat
);
883 /* Try isolate the page */
884 if (__isolate_lru_page(page
, isolate_mode
) != 0)
887 VM_BUG_ON_PAGE(PageCompound(page
), page
);
889 /* Successfully isolated */
890 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
891 inc_node_page_state(page
,
892 NR_ISOLATED_ANON
+ page_is_file_cache(page
));
895 list_add(&page
->lru
, &cc
->migratepages
);
896 cc
->nr_migratepages
++;
900 * Record where we could have freed pages by migration and not
901 * yet flushed them to buddy allocator.
902 * - this is the lowest page that was isolated and likely be
903 * then freed by migration.
905 if (!cc
->last_migrated_pfn
)
906 cc
->last_migrated_pfn
= low_pfn
;
908 /* Avoid isolating too much */
909 if (cc
->nr_migratepages
== COMPACT_CLUSTER_MAX
) {
916 if (!skip_on_failure
)
920 * We have isolated some pages, but then failed. Release them
921 * instead of migrating, as we cannot form the cc->order buddy
926 spin_unlock_irqrestore(zone_lru_lock(zone
), flags
);
929 putback_movable_pages(&cc
->migratepages
);
930 cc
->nr_migratepages
= 0;
931 cc
->last_migrated_pfn
= 0;
935 if (low_pfn
< next_skip_pfn
) {
936 low_pfn
= next_skip_pfn
- 1;
938 * The check near the loop beginning would have updated
939 * next_skip_pfn too, but this is a bit simpler.
941 next_skip_pfn
+= 1UL << cc
->order
;
946 * The PageBuddy() check could have potentially brought us outside
947 * the range to be scanned.
949 if (unlikely(low_pfn
> end_pfn
))
953 spin_unlock_irqrestore(zone_lru_lock(zone
), flags
);
956 * Update the pageblock-skip information and cached scanner pfn,
957 * if the whole pageblock was scanned without isolating any page.
959 if (low_pfn
== end_pfn
)
960 update_pageblock_skip(cc
, valid_page
, nr_isolated
, true);
962 trace_mm_compaction_isolate_migratepages(start_pfn
, low_pfn
,
963 nr_scanned
, nr_isolated
);
965 cc
->total_migrate_scanned
+= nr_scanned
;
967 count_compact_events(COMPACTISOLATED
, nr_isolated
);
973 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
974 * @cc: Compaction control structure.
975 * @start_pfn: The first PFN to start isolating.
976 * @end_pfn: The one-past-last PFN.
978 * Returns zero if isolation fails fatally due to e.g. pending signal.
979 * Otherwise, function returns one-past-the-last PFN of isolated page
980 * (which may be greater than end_pfn if end fell in a middle of a THP page).
983 isolate_migratepages_range(struct compact_control
*cc
, unsigned long start_pfn
,
984 unsigned long end_pfn
)
986 unsigned long pfn
, block_start_pfn
, block_end_pfn
;
988 /* Scan block by block. First and last block may be incomplete */
990 block_start_pfn
= pageblock_start_pfn(pfn
);
991 if (block_start_pfn
< cc
->zone
->zone_start_pfn
)
992 block_start_pfn
= cc
->zone
->zone_start_pfn
;
993 block_end_pfn
= pageblock_end_pfn(pfn
);
995 for (; pfn
< end_pfn
; pfn
= block_end_pfn
,
996 block_start_pfn
= block_end_pfn
,
997 block_end_pfn
+= pageblock_nr_pages
) {
999 block_end_pfn
= min(block_end_pfn
, end_pfn
);
1001 if (!pageblock_pfn_to_page(block_start_pfn
,
1002 block_end_pfn
, cc
->zone
))
1005 pfn
= isolate_migratepages_block(cc
, pfn
, block_end_pfn
,
1006 ISOLATE_UNEVICTABLE
);
1011 if (cc
->nr_migratepages
== COMPACT_CLUSTER_MAX
)
1018 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
1019 #ifdef CONFIG_COMPACTION
1021 static bool suitable_migration_source(struct compact_control
*cc
,
1026 if ((cc
->mode
!= MIGRATE_ASYNC
) || !cc
->direct_compaction
)
1029 block_mt
= get_pageblock_migratetype(page
);
1031 if (cc
->migratetype
== MIGRATE_MOVABLE
)
1032 return is_migrate_movable(block_mt
);
1034 return block_mt
== cc
->migratetype
;
1037 /* Returns true if the page is within a block suitable for migration to */
1038 static bool suitable_migration_target(struct compact_control
*cc
,
1041 /* If the page is a large free page, then disallow migration */
1042 if (PageBuddy(page
)) {
1044 * We are checking page_order without zone->lock taken. But
1045 * the only small danger is that we skip a potentially suitable
1046 * pageblock, so it's not worth to check order for valid range.
1048 if (page_order_unsafe(page
) >= pageblock_order
)
1052 if (cc
->ignore_block_suitable
)
1055 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1056 if (is_migrate_movable(get_pageblock_migratetype(page
)))
1059 /* Otherwise skip the block */
1064 * Test whether the free scanner has reached the same or lower pageblock than
1065 * the migration scanner, and compaction should thus terminate.
1067 static inline bool compact_scanners_met(struct compact_control
*cc
)
1069 return (cc
->free_pfn
>> pageblock_order
)
1070 <= (cc
->migrate_pfn
>> pageblock_order
);
1074 * Based on information in the current compact_control, find blocks
1075 * suitable for isolating free pages from and then isolate them.
1077 static void isolate_freepages(struct compact_control
*cc
)
1079 struct zone
*zone
= cc
->zone
;
1081 unsigned long block_start_pfn
; /* start of current pageblock */
1082 unsigned long isolate_start_pfn
; /* exact pfn we start at */
1083 unsigned long block_end_pfn
; /* end of current pageblock */
1084 unsigned long low_pfn
; /* lowest pfn scanner is able to scan */
1085 struct list_head
*freelist
= &cc
->freepages
;
1088 * Initialise the free scanner. The starting point is where we last
1089 * successfully isolated from, zone-cached value, or the end of the
1090 * zone when isolating for the first time. For looping we also need
1091 * this pfn aligned down to the pageblock boundary, because we do
1092 * block_start_pfn -= pageblock_nr_pages in the for loop.
1093 * For ending point, take care when isolating in last pageblock of a
1094 * a zone which ends in the middle of a pageblock.
1095 * The low boundary is the end of the pageblock the migration scanner
1098 isolate_start_pfn
= cc
->free_pfn
;
1099 block_start_pfn
= pageblock_start_pfn(cc
->free_pfn
);
1100 block_end_pfn
= min(block_start_pfn
+ pageblock_nr_pages
,
1101 zone_end_pfn(zone
));
1102 low_pfn
= pageblock_end_pfn(cc
->migrate_pfn
);
1105 * Isolate free pages until enough are available to migrate the
1106 * pages on cc->migratepages. We stop searching if the migrate
1107 * and free page scanners meet or enough free pages are isolated.
1109 for (; block_start_pfn
>= low_pfn
;
1110 block_end_pfn
= block_start_pfn
,
1111 block_start_pfn
-= pageblock_nr_pages
,
1112 isolate_start_pfn
= block_start_pfn
) {
1114 * This can iterate a massively long zone without finding any
1115 * suitable migration targets, so periodically check if we need
1116 * to schedule, or even abort async compaction.
1118 if (!(block_start_pfn
% (SWAP_CLUSTER_MAX
* pageblock_nr_pages
))
1119 && compact_should_abort(cc
))
1122 page
= pageblock_pfn_to_page(block_start_pfn
, block_end_pfn
,
1127 /* Check the block is suitable for migration */
1128 if (!suitable_migration_target(cc
, page
))
1131 /* If isolation recently failed, do not retry */
1132 if (!isolation_suitable(cc
, page
))
1135 /* Found a block suitable for isolating free pages from. */
1136 isolate_freepages_block(cc
, &isolate_start_pfn
, block_end_pfn
,
1140 * If we isolated enough freepages, or aborted due to lock
1141 * contention, terminate.
1143 if ((cc
->nr_freepages
>= cc
->nr_migratepages
)
1145 if (isolate_start_pfn
>= block_end_pfn
) {
1147 * Restart at previous pageblock if more
1148 * freepages can be isolated next time.
1151 block_start_pfn
- pageblock_nr_pages
;
1154 } else if (isolate_start_pfn
< block_end_pfn
) {
1156 * If isolation failed early, do not continue
1163 /* __isolate_free_page() does not map the pages */
1164 map_pages(freelist
);
1167 * Record where the free scanner will restart next time. Either we
1168 * broke from the loop and set isolate_start_pfn based on the last
1169 * call to isolate_freepages_block(), or we met the migration scanner
1170 * and the loop terminated due to isolate_start_pfn < low_pfn
1172 cc
->free_pfn
= isolate_start_pfn
;
1176 * This is a migrate-callback that "allocates" freepages by taking pages
1177 * from the isolated freelists in the block we are migrating to.
1179 static struct page
*compaction_alloc(struct page
*migratepage
,
1183 struct compact_control
*cc
= (struct compact_control
*)data
;
1184 struct page
*freepage
;
1187 * Isolate free pages if necessary, and if we are not aborting due to
1190 if (list_empty(&cc
->freepages
)) {
1192 isolate_freepages(cc
);
1194 if (list_empty(&cc
->freepages
))
1198 freepage
= list_entry(cc
->freepages
.next
, struct page
, lru
);
1199 list_del(&freepage
->lru
);
1206 * This is a migrate-callback that "frees" freepages back to the isolated
1207 * freelist. All pages on the freelist are from the same zone, so there is no
1208 * special handling needed for NUMA.
1210 static void compaction_free(struct page
*page
, unsigned long data
)
1212 struct compact_control
*cc
= (struct compact_control
*)data
;
1214 list_add(&page
->lru
, &cc
->freepages
);
1218 /* possible outcome of isolate_migratepages */
1220 ISOLATE_ABORT
, /* Abort compaction now */
1221 ISOLATE_NONE
, /* No pages isolated, continue scanning */
1222 ISOLATE_SUCCESS
, /* Pages isolated, migrate */
1223 } isolate_migrate_t
;
1226 * Allow userspace to control policy on scanning the unevictable LRU for
1227 * compactable pages.
1229 int sysctl_compact_unevictable_allowed __read_mostly
= 1;
1232 * Isolate all pages that can be migrated from the first suitable block,
1233 * starting at the block pointed to by the migrate scanner pfn within
1236 static isolate_migrate_t
isolate_migratepages(struct zone
*zone
,
1237 struct compact_control
*cc
)
1239 unsigned long block_start_pfn
;
1240 unsigned long block_end_pfn
;
1241 unsigned long low_pfn
;
1243 const isolate_mode_t isolate_mode
=
1244 (sysctl_compact_unevictable_allowed
? ISOLATE_UNEVICTABLE
: 0) |
1245 (cc
->mode
!= MIGRATE_SYNC
? ISOLATE_ASYNC_MIGRATE
: 0);
1248 * Start at where we last stopped, or beginning of the zone as
1249 * initialized by compact_zone()
1251 low_pfn
= cc
->migrate_pfn
;
1252 block_start_pfn
= pageblock_start_pfn(low_pfn
);
1253 if (block_start_pfn
< zone
->zone_start_pfn
)
1254 block_start_pfn
= zone
->zone_start_pfn
;
1256 /* Only scan within a pageblock boundary */
1257 block_end_pfn
= pageblock_end_pfn(low_pfn
);
1260 * Iterate over whole pageblocks until we find the first suitable.
1261 * Do not cross the free scanner.
1263 for (; block_end_pfn
<= cc
->free_pfn
;
1264 low_pfn
= block_end_pfn
,
1265 block_start_pfn
= block_end_pfn
,
1266 block_end_pfn
+= pageblock_nr_pages
) {
1269 * This can potentially iterate a massively long zone with
1270 * many pageblocks unsuitable, so periodically check if we
1271 * need to schedule, or even abort async compaction.
1273 if (!(low_pfn
% (SWAP_CLUSTER_MAX
* pageblock_nr_pages
))
1274 && compact_should_abort(cc
))
1277 page
= pageblock_pfn_to_page(block_start_pfn
, block_end_pfn
,
1282 /* If isolation recently failed, do not retry */
1283 if (!isolation_suitable(cc
, page
))
1287 * For async compaction, also only scan in MOVABLE blocks.
1288 * Async compaction is optimistic to see if the minimum amount
1289 * of work satisfies the allocation.
1291 if (!suitable_migration_source(cc
, page
))
1294 /* Perform the isolation */
1295 low_pfn
= isolate_migratepages_block(cc
, low_pfn
,
1296 block_end_pfn
, isolate_mode
);
1298 if (!low_pfn
|| cc
->contended
)
1299 return ISOLATE_ABORT
;
1302 * Either we isolated something and proceed with migration. Or
1303 * we failed and compact_zone should decide if we should
1309 /* Record where migration scanner will be restarted. */
1310 cc
->migrate_pfn
= low_pfn
;
1312 return cc
->nr_migratepages
? ISOLATE_SUCCESS
: ISOLATE_NONE
;
1316 * order == -1 is expected when compacting via
1317 * /proc/sys/vm/compact_memory
1319 static inline bool is_via_compact_memory(int order
)
1324 static enum compact_result
__compact_finished(struct zone
*zone
,
1325 struct compact_control
*cc
)
1328 const int migratetype
= cc
->migratetype
;
1330 if (cc
->contended
|| fatal_signal_pending(current
))
1331 return COMPACT_CONTENDED
;
1333 /* Compaction run completes if the migrate and free scanner meet */
1334 if (compact_scanners_met(cc
)) {
1335 /* Let the next compaction start anew. */
1336 reset_cached_positions(zone
);
1339 * Mark that the PG_migrate_skip information should be cleared
1340 * by kswapd when it goes to sleep. kcompactd does not set the
1341 * flag itself as the decision to be clear should be directly
1342 * based on an allocation request.
1344 if (cc
->direct_compaction
)
1345 zone
->compact_blockskip_flush
= true;
1348 return COMPACT_COMPLETE
;
1350 return COMPACT_PARTIAL_SKIPPED
;
1353 if (is_via_compact_memory(cc
->order
))
1354 return COMPACT_CONTINUE
;
1356 if (cc
->finishing_block
) {
1358 * We have finished the pageblock, but better check again that
1359 * we really succeeded.
1361 if (IS_ALIGNED(cc
->migrate_pfn
, pageblock_nr_pages
))
1362 cc
->finishing_block
= false;
1364 return COMPACT_CONTINUE
;
1367 /* Direct compactor: Is a suitable page free? */
1368 for (order
= cc
->order
; order
< MAX_ORDER
; order
++) {
1369 struct free_area
*area
= &zone
->free_area
[order
];
1372 /* Job done if page is free of the right migratetype */
1373 if (!list_empty(&area
->free_list
[migratetype
]))
1374 return COMPACT_SUCCESS
;
1377 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
1378 if (migratetype
== MIGRATE_MOVABLE
&&
1379 !list_empty(&area
->free_list
[MIGRATE_CMA
]))
1380 return COMPACT_SUCCESS
;
1383 * Job done if allocation would steal freepages from
1384 * other migratetype buddy lists.
1386 if (find_suitable_fallback(area
, order
, migratetype
,
1387 true, &can_steal
) != -1) {
1389 /* movable pages are OK in any pageblock */
1390 if (migratetype
== MIGRATE_MOVABLE
)
1391 return COMPACT_SUCCESS
;
1394 * We are stealing for a non-movable allocation. Make
1395 * sure we finish compacting the current pageblock
1396 * first so it is as free as possible and we won't
1397 * have to steal another one soon. This only applies
1398 * to sync compaction, as async compaction operates
1399 * on pageblocks of the same migratetype.
1401 if (cc
->mode
== MIGRATE_ASYNC
||
1402 IS_ALIGNED(cc
->migrate_pfn
,
1403 pageblock_nr_pages
)) {
1404 return COMPACT_SUCCESS
;
1407 cc
->finishing_block
= true;
1408 return COMPACT_CONTINUE
;
1412 return COMPACT_NO_SUITABLE_PAGE
;
1415 static enum compact_result
compact_finished(struct zone
*zone
,
1416 struct compact_control
*cc
)
1420 ret
= __compact_finished(zone
, cc
);
1421 trace_mm_compaction_finished(zone
, cc
->order
, ret
);
1422 if (ret
== COMPACT_NO_SUITABLE_PAGE
)
1423 ret
= COMPACT_CONTINUE
;
1429 * compaction_suitable: Is this suitable to run compaction on this zone now?
1431 * COMPACT_SKIPPED - If there are too few free pages for compaction
1432 * COMPACT_SUCCESS - If the allocation would succeed without compaction
1433 * COMPACT_CONTINUE - If compaction should run now
1435 static enum compact_result
__compaction_suitable(struct zone
*zone
, int order
,
1436 unsigned int alloc_flags
,
1438 unsigned long wmark_target
)
1440 unsigned long watermark
;
1442 if (is_via_compact_memory(order
))
1443 return COMPACT_CONTINUE
;
1445 watermark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
1447 * If watermarks for high-order allocation are already met, there
1448 * should be no need for compaction at all.
1450 if (zone_watermark_ok(zone
, order
, watermark
, classzone_idx
,
1452 return COMPACT_SUCCESS
;
1455 * Watermarks for order-0 must be met for compaction to be able to
1456 * isolate free pages for migration targets. This means that the
1457 * watermark and alloc_flags have to match, or be more pessimistic than
1458 * the check in __isolate_free_page(). We don't use the direct
1459 * compactor's alloc_flags, as they are not relevant for freepage
1460 * isolation. We however do use the direct compactor's classzone_idx to
1461 * skip over zones where lowmem reserves would prevent allocation even
1462 * if compaction succeeds.
1463 * For costly orders, we require low watermark instead of min for
1464 * compaction to proceed to increase its chances.
1465 * ALLOC_CMA is used, as pages in CMA pageblocks are considered
1466 * suitable migration targets
1468 watermark
= (order
> PAGE_ALLOC_COSTLY_ORDER
) ?
1469 low_wmark_pages(zone
) : min_wmark_pages(zone
);
1470 watermark
+= compact_gap(order
);
1471 if (!__zone_watermark_ok(zone
, 0, watermark
, classzone_idx
,
1472 ALLOC_CMA
, wmark_target
))
1473 return COMPACT_SKIPPED
;
1475 return COMPACT_CONTINUE
;
1478 enum compact_result
compaction_suitable(struct zone
*zone
, int order
,
1479 unsigned int alloc_flags
,
1482 enum compact_result ret
;
1485 ret
= __compaction_suitable(zone
, order
, alloc_flags
, classzone_idx
,
1486 zone_page_state(zone
, NR_FREE_PAGES
));
1488 * fragmentation index determines if allocation failures are due to
1489 * low memory or external fragmentation
1491 * index of -1000 would imply allocations might succeed depending on
1492 * watermarks, but we already failed the high-order watermark check
1493 * index towards 0 implies failure is due to lack of memory
1494 * index towards 1000 implies failure is due to fragmentation
1496 * Only compact if a failure would be due to fragmentation. Also
1497 * ignore fragindex for non-costly orders where the alternative to
1498 * a successful reclaim/compaction is OOM. Fragindex and the
1499 * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
1500 * excessive compaction for costly orders, but it should not be at the
1501 * expense of system stability.
1503 if (ret
== COMPACT_CONTINUE
&& (order
> PAGE_ALLOC_COSTLY_ORDER
)) {
1504 fragindex
= fragmentation_index(zone
, order
);
1505 if (fragindex
>= 0 && fragindex
<= sysctl_extfrag_threshold
)
1506 ret
= COMPACT_NOT_SUITABLE_ZONE
;
1509 trace_mm_compaction_suitable(zone
, order
, ret
);
1510 if (ret
== COMPACT_NOT_SUITABLE_ZONE
)
1511 ret
= COMPACT_SKIPPED
;
1516 bool compaction_zonelist_suitable(struct alloc_context
*ac
, int order
,
1523 * Make sure at least one zone would pass __compaction_suitable if we continue
1524 * retrying the reclaim.
1526 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
1528 unsigned long available
;
1529 enum compact_result compact_result
;
1532 * Do not consider all the reclaimable memory because we do not
1533 * want to trash just for a single high order allocation which
1534 * is even not guaranteed to appear even if __compaction_suitable
1535 * is happy about the watermark check.
1537 available
= zone_reclaimable_pages(zone
) / order
;
1538 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
1539 compact_result
= __compaction_suitable(zone
, order
, alloc_flags
,
1540 ac_classzone_idx(ac
), available
);
1541 if (compact_result
!= COMPACT_SKIPPED
)
1548 static enum compact_result
compact_zone(struct zone
*zone
, struct compact_control
*cc
)
1550 enum compact_result ret
;
1551 unsigned long start_pfn
= zone
->zone_start_pfn
;
1552 unsigned long end_pfn
= zone_end_pfn(zone
);
1553 const bool sync
= cc
->mode
!= MIGRATE_ASYNC
;
1555 cc
->migratetype
= gfpflags_to_migratetype(cc
->gfp_mask
);
1556 ret
= compaction_suitable(zone
, cc
->order
, cc
->alloc_flags
,
1558 /* Compaction is likely to fail */
1559 if (ret
== COMPACT_SUCCESS
|| ret
== COMPACT_SKIPPED
)
1562 /* huh, compaction_suitable is returning something unexpected */
1563 VM_BUG_ON(ret
!= COMPACT_CONTINUE
);
1566 * Clear pageblock skip if there were failures recently and compaction
1567 * is about to be retried after being deferred.
1569 if (compaction_restarting(zone
, cc
->order
))
1570 __reset_isolation_suitable(zone
);
1573 * Setup to move all movable pages to the end of the zone. Used cached
1574 * information on where the scanners should start (unless we explicitly
1575 * want to compact the whole zone), but check that it is initialised
1576 * by ensuring the values are within zone boundaries.
1578 if (cc
->whole_zone
) {
1579 cc
->migrate_pfn
= start_pfn
;
1580 cc
->free_pfn
= pageblock_start_pfn(end_pfn
- 1);
1582 cc
->migrate_pfn
= zone
->compact_cached_migrate_pfn
[sync
];
1583 cc
->free_pfn
= zone
->compact_cached_free_pfn
;
1584 if (cc
->free_pfn
< start_pfn
|| cc
->free_pfn
>= end_pfn
) {
1585 cc
->free_pfn
= pageblock_start_pfn(end_pfn
- 1);
1586 zone
->compact_cached_free_pfn
= cc
->free_pfn
;
1588 if (cc
->migrate_pfn
< start_pfn
|| cc
->migrate_pfn
>= end_pfn
) {
1589 cc
->migrate_pfn
= start_pfn
;
1590 zone
->compact_cached_migrate_pfn
[0] = cc
->migrate_pfn
;
1591 zone
->compact_cached_migrate_pfn
[1] = cc
->migrate_pfn
;
1594 if (cc
->migrate_pfn
== start_pfn
)
1595 cc
->whole_zone
= true;
1598 cc
->last_migrated_pfn
= 0;
1600 trace_mm_compaction_begin(start_pfn
, cc
->migrate_pfn
,
1601 cc
->free_pfn
, end_pfn
, sync
);
1603 migrate_prep_local();
1605 while ((ret
= compact_finished(zone
, cc
)) == COMPACT_CONTINUE
) {
1608 switch (isolate_migratepages(zone
, cc
)) {
1610 ret
= COMPACT_CONTENDED
;
1611 putback_movable_pages(&cc
->migratepages
);
1612 cc
->nr_migratepages
= 0;
1616 * We haven't isolated and migrated anything, but
1617 * there might still be unflushed migrations from
1618 * previous cc->order aligned block.
1621 case ISOLATE_SUCCESS
:
1625 err
= migrate_pages(&cc
->migratepages
, compaction_alloc
,
1626 compaction_free
, (unsigned long)cc
, cc
->mode
,
1629 trace_mm_compaction_migratepages(cc
->nr_migratepages
, err
,
1632 /* All pages were either migrated or will be released */
1633 cc
->nr_migratepages
= 0;
1635 putback_movable_pages(&cc
->migratepages
);
1637 * migrate_pages() may return -ENOMEM when scanners meet
1638 * and we want compact_finished() to detect it
1640 if (err
== -ENOMEM
&& !compact_scanners_met(cc
)) {
1641 ret
= COMPACT_CONTENDED
;
1645 * We failed to migrate at least one page in the current
1646 * order-aligned block, so skip the rest of it.
1648 if (cc
->direct_compaction
&&
1649 (cc
->mode
== MIGRATE_ASYNC
)) {
1650 cc
->migrate_pfn
= block_end_pfn(
1651 cc
->migrate_pfn
- 1, cc
->order
);
1652 /* Draining pcplists is useless in this case */
1653 cc
->last_migrated_pfn
= 0;
1660 * Has the migration scanner moved away from the previous
1661 * cc->order aligned block where we migrated from? If yes,
1662 * flush the pages that were freed, so that they can merge and
1663 * compact_finished() can detect immediately if allocation
1666 if (cc
->order
> 0 && cc
->last_migrated_pfn
) {
1668 unsigned long current_block_start
=
1669 block_start_pfn(cc
->migrate_pfn
, cc
->order
);
1671 if (cc
->last_migrated_pfn
< current_block_start
) {
1673 lru_add_drain_cpu(cpu
);
1674 drain_local_pages(zone
);
1676 /* No more flushing until we migrate again */
1677 cc
->last_migrated_pfn
= 0;
1685 * Release free pages and update where the free scanner should restart,
1686 * so we don't leave any returned pages behind in the next attempt.
1688 if (cc
->nr_freepages
> 0) {
1689 unsigned long free_pfn
= release_freepages(&cc
->freepages
);
1691 cc
->nr_freepages
= 0;
1692 VM_BUG_ON(free_pfn
== 0);
1693 /* The cached pfn is always the first in a pageblock */
1694 free_pfn
= pageblock_start_pfn(free_pfn
);
1696 * Only go back, not forward. The cached pfn might have been
1697 * already reset to zone end in compact_finished()
1699 if (free_pfn
> zone
->compact_cached_free_pfn
)
1700 zone
->compact_cached_free_pfn
= free_pfn
;
1703 count_compact_events(COMPACTMIGRATE_SCANNED
, cc
->total_migrate_scanned
);
1704 count_compact_events(COMPACTFREE_SCANNED
, cc
->total_free_scanned
);
1706 trace_mm_compaction_end(start_pfn
, cc
->migrate_pfn
,
1707 cc
->free_pfn
, end_pfn
, sync
, ret
);
1712 static enum compact_result
compact_zone_order(struct zone
*zone
, int order
,
1713 gfp_t gfp_mask
, enum compact_priority prio
,
1714 unsigned int alloc_flags
, int classzone_idx
)
1716 enum compact_result ret
;
1717 struct compact_control cc
= {
1719 .nr_migratepages
= 0,
1720 .total_migrate_scanned
= 0,
1721 .total_free_scanned
= 0,
1723 .gfp_mask
= gfp_mask
,
1725 .mode
= (prio
== COMPACT_PRIO_ASYNC
) ?
1726 MIGRATE_ASYNC
: MIGRATE_SYNC_LIGHT
,
1727 .alloc_flags
= alloc_flags
,
1728 .classzone_idx
= classzone_idx
,
1729 .direct_compaction
= true,
1730 .whole_zone
= (prio
== MIN_COMPACT_PRIORITY
),
1731 .ignore_skip_hint
= (prio
== MIN_COMPACT_PRIORITY
),
1732 .ignore_block_suitable
= (prio
== MIN_COMPACT_PRIORITY
)
1734 INIT_LIST_HEAD(&cc
.freepages
);
1735 INIT_LIST_HEAD(&cc
.migratepages
);
1737 ret
= compact_zone(zone
, &cc
);
1739 VM_BUG_ON(!list_empty(&cc
.freepages
));
1740 VM_BUG_ON(!list_empty(&cc
.migratepages
));
1745 int sysctl_extfrag_threshold
= 500;
1748 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1749 * @gfp_mask: The GFP mask of the current allocation
1750 * @order: The order of the current allocation
1751 * @alloc_flags: The allocation flags of the current allocation
1752 * @ac: The context of current allocation
1753 * @mode: The migration mode for async, sync light, or sync migration
1755 * This is the main entry point for direct page compaction.
1757 enum compact_result
try_to_compact_pages(gfp_t gfp_mask
, unsigned int order
,
1758 unsigned int alloc_flags
, const struct alloc_context
*ac
,
1759 enum compact_priority prio
)
1761 int may_perform_io
= gfp_mask
& __GFP_IO
;
1764 enum compact_result rc
= COMPACT_SKIPPED
;
1767 * Check if the GFP flags allow compaction - GFP_NOIO is really
1768 * tricky context because the migration might require IO
1770 if (!may_perform_io
)
1771 return COMPACT_SKIPPED
;
1773 trace_mm_compaction_try_to_compact_pages(order
, gfp_mask
, prio
);
1775 /* Compact each zone in the list */
1776 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
1778 enum compact_result status
;
1780 if (prio
> MIN_COMPACT_PRIORITY
1781 && compaction_deferred(zone
, order
)) {
1782 rc
= max_t(enum compact_result
, COMPACT_DEFERRED
, rc
);
1786 status
= compact_zone_order(zone
, order
, gfp_mask
, prio
,
1787 alloc_flags
, ac_classzone_idx(ac
));
1788 rc
= max(status
, rc
);
1790 /* The allocation should succeed, stop compacting */
1791 if (status
== COMPACT_SUCCESS
) {
1793 * We think the allocation will succeed in this zone,
1794 * but it is not certain, hence the false. The caller
1795 * will repeat this with true if allocation indeed
1796 * succeeds in this zone.
1798 compaction_defer_reset(zone
, order
, false);
1803 if (prio
!= COMPACT_PRIO_ASYNC
&& (status
== COMPACT_COMPLETE
||
1804 status
== COMPACT_PARTIAL_SKIPPED
))
1806 * We think that allocation won't succeed in this zone
1807 * so we defer compaction there. If it ends up
1808 * succeeding after all, it will be reset.
1810 defer_compaction(zone
, order
);
1813 * We might have stopped compacting due to need_resched() in
1814 * async compaction, or due to a fatal signal detected. In that
1815 * case do not try further zones
1817 if ((prio
== COMPACT_PRIO_ASYNC
&& need_resched())
1818 || fatal_signal_pending(current
))
1826 /* Compact all zones within a node */
1827 static void compact_node(int nid
)
1829 pg_data_t
*pgdat
= NODE_DATA(nid
);
1832 struct compact_control cc
= {
1834 .total_migrate_scanned
= 0,
1835 .total_free_scanned
= 0,
1836 .mode
= MIGRATE_SYNC
,
1837 .ignore_skip_hint
= true,
1839 .gfp_mask
= GFP_KERNEL
,
1843 for (zoneid
= 0; zoneid
< MAX_NR_ZONES
; zoneid
++) {
1845 zone
= &pgdat
->node_zones
[zoneid
];
1846 if (!populated_zone(zone
))
1849 cc
.nr_freepages
= 0;
1850 cc
.nr_migratepages
= 0;
1852 INIT_LIST_HEAD(&cc
.freepages
);
1853 INIT_LIST_HEAD(&cc
.migratepages
);
1855 compact_zone(zone
, &cc
);
1857 VM_BUG_ON(!list_empty(&cc
.freepages
));
1858 VM_BUG_ON(!list_empty(&cc
.migratepages
));
1862 /* Compact all nodes in the system */
1863 static void compact_nodes(void)
1867 /* Flush pending updates to the LRU lists */
1868 lru_add_drain_all();
1870 for_each_online_node(nid
)
1874 /* The written value is actually unused, all memory is compacted */
1875 int sysctl_compact_memory
;
1878 * This is the entry point for compacting all nodes via
1879 * /proc/sys/vm/compact_memory
1881 int sysctl_compaction_handler(struct ctl_table
*table
, int write
,
1882 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
1890 int sysctl_extfrag_handler(struct ctl_table
*table
, int write
,
1891 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
1893 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
1898 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1899 static ssize_t
sysfs_compact_node(struct device
*dev
,
1900 struct device_attribute
*attr
,
1901 const char *buf
, size_t count
)
1905 if (nid
>= 0 && nid
< nr_node_ids
&& node_online(nid
)) {
1906 /* Flush pending updates to the LRU lists */
1907 lru_add_drain_all();
1914 static DEVICE_ATTR(compact
, S_IWUSR
, NULL
, sysfs_compact_node
);
1916 int compaction_register_node(struct node
*node
)
1918 return device_create_file(&node
->dev
, &dev_attr_compact
);
1921 void compaction_unregister_node(struct node
*node
)
1923 return device_remove_file(&node
->dev
, &dev_attr_compact
);
1925 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1927 static inline bool kcompactd_work_requested(pg_data_t
*pgdat
)
1929 return pgdat
->kcompactd_max_order
> 0 || kthread_should_stop();
1932 static bool kcompactd_node_suitable(pg_data_t
*pgdat
)
1936 enum zone_type classzone_idx
= pgdat
->kcompactd_classzone_idx
;
1938 for (zoneid
= 0; zoneid
<= classzone_idx
; zoneid
++) {
1939 zone
= &pgdat
->node_zones
[zoneid
];
1941 if (!populated_zone(zone
))
1944 if (compaction_suitable(zone
, pgdat
->kcompactd_max_order
, 0,
1945 classzone_idx
) == COMPACT_CONTINUE
)
1952 static void kcompactd_do_work(pg_data_t
*pgdat
)
1955 * With no special task, compact all zones so that a page of requested
1956 * order is allocatable.
1960 struct compact_control cc
= {
1961 .order
= pgdat
->kcompactd_max_order
,
1962 .total_migrate_scanned
= 0,
1963 .total_free_scanned
= 0,
1964 .classzone_idx
= pgdat
->kcompactd_classzone_idx
,
1965 .mode
= MIGRATE_SYNC_LIGHT
,
1966 .ignore_skip_hint
= false,
1967 .gfp_mask
= GFP_KERNEL
,
1969 trace_mm_compaction_kcompactd_wake(pgdat
->node_id
, cc
.order
,
1971 count_compact_event(KCOMPACTD_WAKE
);
1973 for (zoneid
= 0; zoneid
<= cc
.classzone_idx
; zoneid
++) {
1976 zone
= &pgdat
->node_zones
[zoneid
];
1977 if (!populated_zone(zone
))
1980 if (compaction_deferred(zone
, cc
.order
))
1983 if (compaction_suitable(zone
, cc
.order
, 0, zoneid
) !=
1987 cc
.nr_freepages
= 0;
1988 cc
.nr_migratepages
= 0;
1989 cc
.total_migrate_scanned
= 0;
1990 cc
.total_free_scanned
= 0;
1992 INIT_LIST_HEAD(&cc
.freepages
);
1993 INIT_LIST_HEAD(&cc
.migratepages
);
1995 if (kthread_should_stop())
1997 status
= compact_zone(zone
, &cc
);
1999 if (status
== COMPACT_SUCCESS
) {
2000 compaction_defer_reset(zone
, cc
.order
, false);
2001 } else if (status
== COMPACT_PARTIAL_SKIPPED
|| status
== COMPACT_COMPLETE
) {
2003 * We use sync migration mode here, so we defer like
2004 * sync direct compaction does.
2006 defer_compaction(zone
, cc
.order
);
2009 count_compact_events(KCOMPACTD_MIGRATE_SCANNED
,
2010 cc
.total_migrate_scanned
);
2011 count_compact_events(KCOMPACTD_FREE_SCANNED
,
2012 cc
.total_free_scanned
);
2014 VM_BUG_ON(!list_empty(&cc
.freepages
));
2015 VM_BUG_ON(!list_empty(&cc
.migratepages
));
2019 * Regardless of success, we are done until woken up next. But remember
2020 * the requested order/classzone_idx in case it was higher/tighter than
2023 if (pgdat
->kcompactd_max_order
<= cc
.order
)
2024 pgdat
->kcompactd_max_order
= 0;
2025 if (pgdat
->kcompactd_classzone_idx
>= cc
.classzone_idx
)
2026 pgdat
->kcompactd_classzone_idx
= pgdat
->nr_zones
- 1;
2029 void wakeup_kcompactd(pg_data_t
*pgdat
, int order
, int classzone_idx
)
2034 if (pgdat
->kcompactd_max_order
< order
)
2035 pgdat
->kcompactd_max_order
= order
;
2037 if (pgdat
->kcompactd_classzone_idx
> classzone_idx
)
2038 pgdat
->kcompactd_classzone_idx
= classzone_idx
;
2041 * Pairs with implicit barrier in wait_event_freezable()
2042 * such that wakeups are not missed.
2044 if (!wq_has_sleeper(&pgdat
->kcompactd_wait
))
2047 if (!kcompactd_node_suitable(pgdat
))
2050 trace_mm_compaction_wakeup_kcompactd(pgdat
->node_id
, order
,
2052 wake_up_interruptible(&pgdat
->kcompactd_wait
);
2056 * The background compaction daemon, started as a kernel thread
2057 * from the init process.
2059 static int kcompactd(void *p
)
2061 pg_data_t
*pgdat
= (pg_data_t
*)p
;
2062 struct task_struct
*tsk
= current
;
2064 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
2066 if (!cpumask_empty(cpumask
))
2067 set_cpus_allowed_ptr(tsk
, cpumask
);
2071 pgdat
->kcompactd_max_order
= 0;
2072 pgdat
->kcompactd_classzone_idx
= pgdat
->nr_zones
- 1;
2074 while (!kthread_should_stop()) {
2075 trace_mm_compaction_kcompactd_sleep(pgdat
->node_id
);
2076 wait_event_freezable(pgdat
->kcompactd_wait
,
2077 kcompactd_work_requested(pgdat
));
2079 kcompactd_do_work(pgdat
);
2086 * This kcompactd start function will be called by init and node-hot-add.
2087 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
2089 int kcompactd_run(int nid
)
2091 pg_data_t
*pgdat
= NODE_DATA(nid
);
2094 if (pgdat
->kcompactd
)
2097 pgdat
->kcompactd
= kthread_run(kcompactd
, pgdat
, "kcompactd%d", nid
);
2098 if (IS_ERR(pgdat
->kcompactd
)) {
2099 pr_err("Failed to start kcompactd on node %d\n", nid
);
2100 ret
= PTR_ERR(pgdat
->kcompactd
);
2101 pgdat
->kcompactd
= NULL
;
2107 * Called by memory hotplug when all memory in a node is offlined. Caller must
2108 * hold mem_hotplug_begin/end().
2110 void kcompactd_stop(int nid
)
2112 struct task_struct
*kcompactd
= NODE_DATA(nid
)->kcompactd
;
2115 kthread_stop(kcompactd
);
2116 NODE_DATA(nid
)->kcompactd
= NULL
;
2121 * It's optimal to keep kcompactd on the same CPUs as their memory, but
2122 * not required for correctness. So if the last cpu in a node goes
2123 * away, we get changed to run anywhere: as the first one comes back,
2124 * restore their cpu bindings.
2126 static int kcompactd_cpu_online(unsigned int cpu
)
2130 for_each_node_state(nid
, N_MEMORY
) {
2131 pg_data_t
*pgdat
= NODE_DATA(nid
);
2132 const struct cpumask
*mask
;
2134 mask
= cpumask_of_node(pgdat
->node_id
);
2136 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
2137 /* One of our CPUs online: restore mask */
2138 set_cpus_allowed_ptr(pgdat
->kcompactd
, mask
);
2143 static int __init
kcompactd_init(void)
2148 ret
= cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN
,
2149 "mm/compaction:online",
2150 kcompactd_cpu_online
, NULL
);
2152 pr_err("kcompactd: failed to register hotplug callbacks.\n");
2156 for_each_node_state(nid
, N_MEMORY
)
2160 subsys_initcall(kcompactd_init
)
2162 #endif /* CONFIG_COMPACTION */