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>
25 #include <linux/psi.h>
28 #ifdef CONFIG_COMPACTION
29 static inline void count_compact_event(enum vm_event_item item
)
34 static inline void count_compact_events(enum vm_event_item item
, long delta
)
36 count_vm_events(item
, delta
);
39 #define count_compact_event(item) do { } while (0)
40 #define count_compact_events(item, delta) do { } while (0)
43 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
45 #define CREATE_TRACE_POINTS
46 #include <trace/events/compaction.h>
48 #define block_start_pfn(pfn, order) round_down(pfn, 1UL << (order))
49 #define block_end_pfn(pfn, order) ALIGN((pfn) + 1, 1UL << (order))
50 #define pageblock_start_pfn(pfn) block_start_pfn(pfn, pageblock_order)
51 #define pageblock_end_pfn(pfn) block_end_pfn(pfn, pageblock_order)
53 static unsigned long release_freepages(struct list_head
*freelist
)
55 struct page
*page
, *next
;
56 unsigned long high_pfn
= 0;
58 list_for_each_entry_safe(page
, next
, freelist
, lru
) {
59 unsigned long pfn
= page_to_pfn(page
);
69 static void split_map_pages(struct list_head
*list
)
71 unsigned int i
, order
, nr_pages
;
72 struct page
*page
, *next
;
75 list_for_each_entry_safe(page
, next
, list
, lru
) {
78 order
= page_private(page
);
79 nr_pages
= 1 << order
;
81 post_alloc_hook(page
, order
, __GFP_MOVABLE
);
83 split_page(page
, order
);
85 for (i
= 0; i
< nr_pages
; i
++) {
86 list_add(&page
->lru
, &tmp_list
);
91 list_splice(&tmp_list
, list
);
94 #ifdef CONFIG_COMPACTION
96 int PageMovable(struct page
*page
)
98 struct address_space
*mapping
;
100 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
101 if (!__PageMovable(page
))
104 mapping
= page_mapping(page
);
105 if (mapping
&& mapping
->a_ops
&& mapping
->a_ops
->isolate_page
)
110 EXPORT_SYMBOL(PageMovable
);
112 void __SetPageMovable(struct page
*page
, struct address_space
*mapping
)
114 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
115 VM_BUG_ON_PAGE((unsigned long)mapping
& PAGE_MAPPING_MOVABLE
, page
);
116 page
->mapping
= (void *)((unsigned long)mapping
| PAGE_MAPPING_MOVABLE
);
118 EXPORT_SYMBOL(__SetPageMovable
);
120 void __ClearPageMovable(struct page
*page
)
122 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
123 VM_BUG_ON_PAGE(!PageMovable(page
), page
);
125 * Clear registered address_space val with keeping PAGE_MAPPING_MOVABLE
126 * flag so that VM can catch up released page by driver after isolation.
127 * With it, VM migration doesn't try to put it back.
129 page
->mapping
= (void *)((unsigned long)page
->mapping
&
130 PAGE_MAPPING_MOVABLE
);
132 EXPORT_SYMBOL(__ClearPageMovable
);
134 /* Do not skip compaction more than 64 times */
135 #define COMPACT_MAX_DEFER_SHIFT 6
138 * Compaction is deferred when compaction fails to result in a page
139 * allocation success. 1 << compact_defer_limit compactions are skipped up
140 * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
142 void defer_compaction(struct zone
*zone
, int order
)
144 zone
->compact_considered
= 0;
145 zone
->compact_defer_shift
++;
147 if (order
< zone
->compact_order_failed
)
148 zone
->compact_order_failed
= order
;
150 if (zone
->compact_defer_shift
> COMPACT_MAX_DEFER_SHIFT
)
151 zone
->compact_defer_shift
= COMPACT_MAX_DEFER_SHIFT
;
153 trace_mm_compaction_defer_compaction(zone
, order
);
156 /* Returns true if compaction should be skipped this time */
157 bool compaction_deferred(struct zone
*zone
, int order
)
159 unsigned long defer_limit
= 1UL << zone
->compact_defer_shift
;
161 if (order
< zone
->compact_order_failed
)
164 /* Avoid possible overflow */
165 if (++zone
->compact_considered
> defer_limit
)
166 zone
->compact_considered
= defer_limit
;
168 if (zone
->compact_considered
>= defer_limit
)
171 trace_mm_compaction_deferred(zone
, order
);
177 * Update defer tracking counters after successful compaction of given order,
178 * which means an allocation either succeeded (alloc_success == true) or is
179 * expected to succeed.
181 void compaction_defer_reset(struct zone
*zone
, int order
,
185 zone
->compact_considered
= 0;
186 zone
->compact_defer_shift
= 0;
188 if (order
>= zone
->compact_order_failed
)
189 zone
->compact_order_failed
= order
+ 1;
191 trace_mm_compaction_defer_reset(zone
, order
);
194 /* Returns true if restarting compaction after many failures */
195 bool compaction_restarting(struct zone
*zone
, int order
)
197 if (order
< zone
->compact_order_failed
)
200 return zone
->compact_defer_shift
== COMPACT_MAX_DEFER_SHIFT
&&
201 zone
->compact_considered
>= 1UL << zone
->compact_defer_shift
;
204 /* Returns true if the pageblock should be scanned for pages to isolate. */
205 static inline bool isolation_suitable(struct compact_control
*cc
,
208 if (cc
->ignore_skip_hint
)
211 return !get_pageblock_skip(page
);
214 static void reset_cached_positions(struct zone
*zone
)
216 zone
->compact_cached_migrate_pfn
[0] = zone
->zone_start_pfn
;
217 zone
->compact_cached_migrate_pfn
[1] = zone
->zone_start_pfn
;
218 zone
->compact_cached_free_pfn
=
219 pageblock_start_pfn(zone_end_pfn(zone
) - 1);
223 * Compound pages of >= pageblock_order should consistenly be skipped until
224 * released. It is always pointless to compact pages of such order (if they are
225 * migratable), and the pageblocks they occupy cannot contain any free pages.
227 static bool pageblock_skip_persistent(struct page
*page
)
229 if (!PageCompound(page
))
232 page
= compound_head(page
);
234 if (compound_order(page
) >= pageblock_order
)
241 __reset_isolation_pfn(struct zone
*zone
, unsigned long pfn
, bool check_source
,
244 struct page
*page
= pfn_to_online_page(pfn
);
245 struct page
*block_page
;
246 struct page
*end_page
;
247 unsigned long block_pfn
;
251 if (zone
!= page_zone(page
))
253 if (pageblock_skip_persistent(page
))
257 * If skip is already cleared do no further checking once the
258 * restart points have been set.
260 if (check_source
&& check_target
&& !get_pageblock_skip(page
))
264 * If clearing skip for the target scanner, do not select a
265 * non-movable pageblock as the starting point.
267 if (!check_source
&& check_target
&&
268 get_pageblock_migratetype(page
) != MIGRATE_MOVABLE
)
271 /* Ensure the start of the pageblock or zone is online and valid */
272 block_pfn
= pageblock_start_pfn(pfn
);
273 block_pfn
= max(block_pfn
, zone
->zone_start_pfn
);
274 block_page
= pfn_to_online_page(block_pfn
);
280 /* Ensure the end of the pageblock or zone is online and valid */
281 block_pfn
= pageblock_end_pfn(pfn
) - 1;
282 block_pfn
= min(block_pfn
, zone_end_pfn(zone
) - 1);
283 end_page
= pfn_to_online_page(block_pfn
);
288 * Only clear the hint if a sample indicates there is either a
289 * free page or an LRU page in the block. One or other condition
290 * is necessary for the block to be a migration source/target.
293 if (pfn_valid_within(pfn
)) {
294 if (check_source
&& PageLRU(page
)) {
295 clear_pageblock_skip(page
);
299 if (check_target
&& PageBuddy(page
)) {
300 clear_pageblock_skip(page
);
305 page
+= (1 << PAGE_ALLOC_COSTLY_ORDER
);
306 pfn
+= (1 << PAGE_ALLOC_COSTLY_ORDER
);
307 } while (page
<= end_page
);
313 * This function is called to clear all cached information on pageblocks that
314 * should be skipped for page isolation when the migrate and free page scanner
317 static void __reset_isolation_suitable(struct zone
*zone
)
319 unsigned long migrate_pfn
= zone
->zone_start_pfn
;
320 unsigned long free_pfn
= zone_end_pfn(zone
) - 1;
321 unsigned long reset_migrate
= free_pfn
;
322 unsigned long reset_free
= migrate_pfn
;
323 bool source_set
= false;
324 bool free_set
= false;
326 if (!zone
->compact_blockskip_flush
)
329 zone
->compact_blockskip_flush
= false;
332 * Walk the zone and update pageblock skip information. Source looks
333 * for PageLRU while target looks for PageBuddy. When the scanner
334 * is found, both PageBuddy and PageLRU are checked as the pageblock
335 * is suitable as both source and target.
337 for (; migrate_pfn
< free_pfn
; migrate_pfn
+= pageblock_nr_pages
,
338 free_pfn
-= pageblock_nr_pages
) {
341 /* Update the migrate PFN */
342 if (__reset_isolation_pfn(zone
, migrate_pfn
, true, source_set
) &&
343 migrate_pfn
< reset_migrate
) {
345 reset_migrate
= migrate_pfn
;
346 zone
->compact_init_migrate_pfn
= reset_migrate
;
347 zone
->compact_cached_migrate_pfn
[0] = reset_migrate
;
348 zone
->compact_cached_migrate_pfn
[1] = reset_migrate
;
351 /* Update the free PFN */
352 if (__reset_isolation_pfn(zone
, free_pfn
, free_set
, true) &&
353 free_pfn
> reset_free
) {
355 reset_free
= free_pfn
;
356 zone
->compact_init_free_pfn
= reset_free
;
357 zone
->compact_cached_free_pfn
= reset_free
;
361 /* Leave no distance if no suitable block was reset */
362 if (reset_migrate
>= reset_free
) {
363 zone
->compact_cached_migrate_pfn
[0] = migrate_pfn
;
364 zone
->compact_cached_migrate_pfn
[1] = migrate_pfn
;
365 zone
->compact_cached_free_pfn
= free_pfn
;
369 void reset_isolation_suitable(pg_data_t
*pgdat
)
373 for (zoneid
= 0; zoneid
< MAX_NR_ZONES
; zoneid
++) {
374 struct zone
*zone
= &pgdat
->node_zones
[zoneid
];
375 if (!populated_zone(zone
))
378 /* Only flush if a full compaction finished recently */
379 if (zone
->compact_blockskip_flush
)
380 __reset_isolation_suitable(zone
);
385 * Sets the pageblock skip bit if it was clear. Note that this is a hint as
386 * locks are not required for read/writers. Returns true if it was already set.
388 static bool test_and_set_skip(struct compact_control
*cc
, struct page
*page
,
393 /* Do no update if skip hint is being ignored */
394 if (cc
->ignore_skip_hint
)
397 if (!IS_ALIGNED(pfn
, pageblock_nr_pages
))
400 skip
= get_pageblock_skip(page
);
401 if (!skip
&& !cc
->no_set_skip_hint
)
402 set_pageblock_skip(page
);
407 static void update_cached_migrate(struct compact_control
*cc
, unsigned long pfn
)
409 struct zone
*zone
= cc
->zone
;
411 pfn
= pageblock_end_pfn(pfn
);
413 /* Set for isolation rather than compaction */
414 if (cc
->no_set_skip_hint
)
417 if (pfn
> zone
->compact_cached_migrate_pfn
[0])
418 zone
->compact_cached_migrate_pfn
[0] = pfn
;
419 if (cc
->mode
!= MIGRATE_ASYNC
&&
420 pfn
> zone
->compact_cached_migrate_pfn
[1])
421 zone
->compact_cached_migrate_pfn
[1] = pfn
;
425 * If no pages were isolated then mark this pageblock to be skipped in the
426 * future. The information is later cleared by __reset_isolation_suitable().
428 static void update_pageblock_skip(struct compact_control
*cc
,
429 struct page
*page
, unsigned long pfn
)
431 struct zone
*zone
= cc
->zone
;
433 if (cc
->no_set_skip_hint
)
439 set_pageblock_skip(page
);
441 /* Update where async and sync compaction should restart */
442 if (pfn
< zone
->compact_cached_free_pfn
)
443 zone
->compact_cached_free_pfn
= pfn
;
446 static inline bool isolation_suitable(struct compact_control
*cc
,
452 static inline bool pageblock_skip_persistent(struct page
*page
)
457 static inline void update_pageblock_skip(struct compact_control
*cc
,
458 struct page
*page
, unsigned long pfn
)
462 static void update_cached_migrate(struct compact_control
*cc
, unsigned long pfn
)
466 static bool test_and_set_skip(struct compact_control
*cc
, struct page
*page
,
471 #endif /* CONFIG_COMPACTION */
474 * Compaction requires the taking of some coarse locks that are potentially
475 * very heavily contended. For async compaction, trylock and record if the
476 * lock is contended. The lock will still be acquired but compaction will
477 * abort when the current block is finished regardless of success rate.
478 * Sync compaction acquires the lock.
480 * Always returns true which makes it easier to track lock state in callers.
482 static bool compact_lock_irqsave(spinlock_t
*lock
, unsigned long *flags
,
483 struct compact_control
*cc
)
486 /* Track if the lock is contended in async mode */
487 if (cc
->mode
== MIGRATE_ASYNC
&& !cc
->contended
) {
488 if (spin_trylock_irqsave(lock
, *flags
))
491 cc
->contended
= true;
494 spin_lock_irqsave(lock
, *flags
);
499 * Compaction requires the taking of some coarse locks that are potentially
500 * very heavily contended. The lock should be periodically unlocked to avoid
501 * having disabled IRQs for a long time, even when there is nobody waiting on
502 * the lock. It might also be that allowing the IRQs will result in
503 * need_resched() becoming true. If scheduling is needed, async compaction
504 * aborts. Sync compaction schedules.
505 * Either compaction type will also abort if a fatal signal is pending.
506 * In either case if the lock was locked, it is dropped and not regained.
508 * Returns true if compaction should abort due to fatal signal pending, or
509 * async compaction due to need_resched()
510 * Returns false when compaction can continue (sync compaction might have
513 static bool compact_unlock_should_abort(spinlock_t
*lock
,
514 unsigned long flags
, bool *locked
, struct compact_control
*cc
)
517 spin_unlock_irqrestore(lock
, flags
);
521 if (fatal_signal_pending(current
)) {
522 cc
->contended
= true;
532 * Isolate free pages onto a private freelist. If @strict is true, will abort
533 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
534 * (even though it may still end up isolating some pages).
536 static unsigned long isolate_freepages_block(struct compact_control
*cc
,
537 unsigned long *start_pfn
,
538 unsigned long end_pfn
,
539 struct list_head
*freelist
,
543 int nr_scanned
= 0, total_isolated
= 0;
545 unsigned long flags
= 0;
547 unsigned long blockpfn
= *start_pfn
;
550 /* Strict mode is for isolation, speed is secondary */
554 cursor
= pfn_to_page(blockpfn
);
556 /* Isolate free pages. */
557 for (; blockpfn
< end_pfn
; blockpfn
+= stride
, cursor
+= stride
) {
559 struct page
*page
= cursor
;
562 * Periodically drop the lock (if held) regardless of its
563 * contention, to give chance to IRQs. Abort if fatal signal
564 * pending or async compaction detects need_resched()
566 if (!(blockpfn
% SWAP_CLUSTER_MAX
)
567 && compact_unlock_should_abort(&cc
->zone
->lock
, flags
,
572 if (!pfn_valid_within(blockpfn
))
576 * For compound pages such as THP and hugetlbfs, we can save
577 * potentially a lot of iterations if we skip them at once.
578 * The check is racy, but we can consider only valid values
579 * and the only danger is skipping too much.
581 if (PageCompound(page
)) {
582 const unsigned int order
= compound_order(page
);
584 if (likely(order
< MAX_ORDER
)) {
585 blockpfn
+= (1UL << order
) - 1;
586 cursor
+= (1UL << order
) - 1;
591 if (!PageBuddy(page
))
595 * If we already hold the lock, we can skip some rechecking.
596 * Note that if we hold the lock now, checked_pageblock was
597 * already set in some previous iteration (or strict is true),
598 * so it is correct to skip the suitable migration target
602 locked
= compact_lock_irqsave(&cc
->zone
->lock
,
605 /* Recheck this is a buddy page under lock */
606 if (!PageBuddy(page
))
610 /* Found a free page, will break it into order-0 pages */
611 order
= page_order(page
);
612 isolated
= __isolate_free_page(page
, order
);
615 set_page_private(page
, order
);
617 total_isolated
+= isolated
;
618 cc
->nr_freepages
+= isolated
;
619 list_add_tail(&page
->lru
, freelist
);
621 if (!strict
&& cc
->nr_migratepages
<= cc
->nr_freepages
) {
622 blockpfn
+= isolated
;
625 /* Advance to the end of split page */
626 blockpfn
+= isolated
- 1;
627 cursor
+= isolated
- 1;
639 spin_unlock_irqrestore(&cc
->zone
->lock
, flags
);
642 * There is a tiny chance that we have read bogus compound_order(),
643 * so be careful to not go outside of the pageblock.
645 if (unlikely(blockpfn
> end_pfn
))
648 trace_mm_compaction_isolate_freepages(*start_pfn
, blockpfn
,
649 nr_scanned
, total_isolated
);
651 /* Record how far we have got within the block */
652 *start_pfn
= blockpfn
;
655 * If strict isolation is requested by CMA then check that all the
656 * pages requested were isolated. If there were any failures, 0 is
657 * returned and CMA will fail.
659 if (strict
&& blockpfn
< end_pfn
)
662 cc
->total_free_scanned
+= nr_scanned
;
664 count_compact_events(COMPACTISOLATED
, total_isolated
);
665 return total_isolated
;
669 * isolate_freepages_range() - isolate free pages.
670 * @cc: Compaction control structure.
671 * @start_pfn: The first PFN to start isolating.
672 * @end_pfn: The one-past-last PFN.
674 * Non-free pages, invalid PFNs, or zone boundaries within the
675 * [start_pfn, end_pfn) range are considered errors, cause function to
676 * undo its actions and return zero.
678 * Otherwise, function returns one-past-the-last PFN of isolated page
679 * (which may be greater then end_pfn if end fell in a middle of
683 isolate_freepages_range(struct compact_control
*cc
,
684 unsigned long start_pfn
, unsigned long end_pfn
)
686 unsigned long isolated
, pfn
, block_start_pfn
, block_end_pfn
;
690 block_start_pfn
= pageblock_start_pfn(pfn
);
691 if (block_start_pfn
< cc
->zone
->zone_start_pfn
)
692 block_start_pfn
= cc
->zone
->zone_start_pfn
;
693 block_end_pfn
= pageblock_end_pfn(pfn
);
695 for (; pfn
< end_pfn
; pfn
+= isolated
,
696 block_start_pfn
= block_end_pfn
,
697 block_end_pfn
+= pageblock_nr_pages
) {
698 /* Protect pfn from changing by isolate_freepages_block */
699 unsigned long isolate_start_pfn
= pfn
;
701 block_end_pfn
= min(block_end_pfn
, end_pfn
);
704 * pfn could pass the block_end_pfn if isolated freepage
705 * is more than pageblock order. In this case, we adjust
706 * scanning range to right one.
708 if (pfn
>= block_end_pfn
) {
709 block_start_pfn
= pageblock_start_pfn(pfn
);
710 block_end_pfn
= pageblock_end_pfn(pfn
);
711 block_end_pfn
= min(block_end_pfn
, end_pfn
);
714 if (!pageblock_pfn_to_page(block_start_pfn
,
715 block_end_pfn
, cc
->zone
))
718 isolated
= isolate_freepages_block(cc
, &isolate_start_pfn
,
719 block_end_pfn
, &freelist
, 0, true);
722 * In strict mode, isolate_freepages_block() returns 0 if
723 * there are any holes in the block (ie. invalid PFNs or
730 * If we managed to isolate pages, it is always (1 << n) *
731 * pageblock_nr_pages for some non-negative n. (Max order
732 * page may span two pageblocks).
736 /* __isolate_free_page() does not map the pages */
737 split_map_pages(&freelist
);
740 /* Loop terminated early, cleanup. */
741 release_freepages(&freelist
);
745 /* We don't use freelists for anything. */
749 /* Similar to reclaim, but different enough that they don't share logic */
750 static bool too_many_isolated(pg_data_t
*pgdat
)
752 unsigned long active
, inactive
, isolated
;
754 inactive
= node_page_state(pgdat
, NR_INACTIVE_FILE
) +
755 node_page_state(pgdat
, NR_INACTIVE_ANON
);
756 active
= node_page_state(pgdat
, NR_ACTIVE_FILE
) +
757 node_page_state(pgdat
, NR_ACTIVE_ANON
);
758 isolated
= node_page_state(pgdat
, NR_ISOLATED_FILE
) +
759 node_page_state(pgdat
, NR_ISOLATED_ANON
);
761 return isolated
> (inactive
+ active
) / 2;
765 * isolate_migratepages_block() - isolate all migrate-able pages within
767 * @cc: Compaction control structure.
768 * @low_pfn: The first PFN to isolate
769 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
770 * @isolate_mode: Isolation mode to be used.
772 * Isolate all pages that can be migrated from the range specified by
773 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
774 * Returns zero if there is a fatal signal pending, otherwise PFN of the
775 * first page that was not scanned (which may be both less, equal to or more
778 * The pages are isolated on cc->migratepages list (not required to be empty),
779 * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
780 * is neither read nor updated.
783 isolate_migratepages_block(struct compact_control
*cc
, unsigned long low_pfn
,
784 unsigned long end_pfn
, isolate_mode_t isolate_mode
)
786 pg_data_t
*pgdat
= cc
->zone
->zone_pgdat
;
787 unsigned long nr_scanned
= 0, nr_isolated
= 0;
788 struct lruvec
*lruvec
;
789 unsigned long flags
= 0;
791 struct page
*page
= NULL
, *valid_page
= NULL
;
792 unsigned long start_pfn
= low_pfn
;
793 bool skip_on_failure
= false;
794 unsigned long next_skip_pfn
= 0;
795 bool skip_updated
= false;
798 * Ensure that there are not too many pages isolated from the LRU
799 * list by either parallel reclaimers or compaction. If there are,
800 * delay for some time until fewer pages are isolated
802 while (unlikely(too_many_isolated(pgdat
))) {
803 /* async migration should just abort */
804 if (cc
->mode
== MIGRATE_ASYNC
)
807 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
809 if (fatal_signal_pending(current
))
815 if (cc
->direct_compaction
&& (cc
->mode
== MIGRATE_ASYNC
)) {
816 skip_on_failure
= true;
817 next_skip_pfn
= block_end_pfn(low_pfn
, cc
->order
);
820 /* Time to isolate some pages for migration */
821 for (; low_pfn
< end_pfn
; low_pfn
++) {
823 if (skip_on_failure
&& low_pfn
>= next_skip_pfn
) {
825 * We have isolated all migration candidates in the
826 * previous order-aligned block, and did not skip it due
827 * to failure. We should migrate the pages now and
828 * hopefully succeed compaction.
834 * We failed to isolate in the previous order-aligned
835 * block. Set the new boundary to the end of the
836 * current block. Note we can't simply increase
837 * next_skip_pfn by 1 << order, as low_pfn might have
838 * been incremented by a higher number due to skipping
839 * a compound or a high-order buddy page in the
840 * previous loop iteration.
842 next_skip_pfn
= block_end_pfn(low_pfn
, cc
->order
);
846 * Periodically drop the lock (if held) regardless of its
847 * contention, to give chance to IRQs. Abort completely if
848 * a fatal signal is pending.
850 if (!(low_pfn
% SWAP_CLUSTER_MAX
)
851 && compact_unlock_should_abort(&pgdat
->lru_lock
,
852 flags
, &locked
, cc
)) {
857 if (!pfn_valid_within(low_pfn
))
861 page
= pfn_to_page(low_pfn
);
864 * Check if the pageblock has already been marked skipped.
865 * Only the aligned PFN is checked as the caller isolates
866 * COMPACT_CLUSTER_MAX at a time so the second call must
867 * not falsely conclude that the block should be skipped.
869 if (!valid_page
&& IS_ALIGNED(low_pfn
, pageblock_nr_pages
)) {
870 if (!cc
->ignore_skip_hint
&& get_pageblock_skip(page
)) {
878 * Skip if free. We read page order here without zone lock
879 * which is generally unsafe, but the race window is small and
880 * the worst thing that can happen is that we skip some
881 * potential isolation targets.
883 if (PageBuddy(page
)) {
884 unsigned long freepage_order
= page_order_unsafe(page
);
887 * Without lock, we cannot be sure that what we got is
888 * a valid page order. Consider only values in the
889 * valid order range to prevent low_pfn overflow.
891 if (freepage_order
> 0 && freepage_order
< MAX_ORDER
)
892 low_pfn
+= (1UL << freepage_order
) - 1;
897 * Regardless of being on LRU, compound pages such as THP and
898 * hugetlbfs are not to be compacted unless we are attempting
899 * an allocation much larger than the huge page size (eg CMA).
900 * We can potentially save a lot of iterations if we skip them
901 * at once. The check is racy, but we can consider only valid
902 * values and the only danger is skipping too much.
904 if (PageCompound(page
) && !cc
->alloc_contig
) {
905 const unsigned int order
= compound_order(page
);
907 if (likely(order
< MAX_ORDER
))
908 low_pfn
+= (1UL << order
) - 1;
913 * Check may be lockless but that's ok as we recheck later.
914 * It's possible to migrate LRU and non-lru movable pages.
915 * Skip any other type of page
917 if (!PageLRU(page
)) {
919 * __PageMovable can return false positive so we need
920 * to verify it under page_lock.
922 if (unlikely(__PageMovable(page
)) &&
923 !PageIsolated(page
)) {
925 spin_unlock_irqrestore(&pgdat
->lru_lock
,
930 if (!isolate_movable_page(page
, isolate_mode
))
931 goto isolate_success
;
938 * Migration will fail if an anonymous page is pinned in memory,
939 * so avoid taking lru_lock and isolating it unnecessarily in an
940 * admittedly racy check.
942 if (!page_mapping(page
) &&
943 page_count(page
) > page_mapcount(page
))
947 * Only allow to migrate anonymous pages in GFP_NOFS context
948 * because those do not depend on fs locks.
950 if (!(cc
->gfp_mask
& __GFP_FS
) && page_mapping(page
))
953 /* If we already hold the lock, we can skip some rechecking */
955 locked
= compact_lock_irqsave(&pgdat
->lru_lock
,
958 /* Try get exclusive access under lock */
961 if (test_and_set_skip(cc
, page
, low_pfn
))
965 /* Recheck PageLRU and PageCompound under lock */
970 * Page become compound since the non-locked check,
971 * and it's on LRU. It can only be a THP so the order
972 * is safe to read and it's 0 for tail pages.
974 if (unlikely(PageCompound(page
) && !cc
->alloc_contig
)) {
975 low_pfn
+= compound_nr(page
) - 1;
980 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
982 /* Try isolate the page */
983 if (__isolate_lru_page(page
, isolate_mode
) != 0)
986 /* The whole page is taken off the LRU; skip the tail pages. */
987 if (PageCompound(page
))
988 low_pfn
+= compound_nr(page
) - 1;
990 /* Successfully isolated */
991 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
992 mod_node_page_state(page_pgdat(page
),
993 NR_ISOLATED_ANON
+ page_is_file_lru(page
),
994 hpage_nr_pages(page
));
997 list_add(&page
->lru
, &cc
->migratepages
);
998 cc
->nr_migratepages
++;
1002 * Avoid isolating too much unless this block is being
1003 * rescanned (e.g. dirty/writeback pages, parallel allocation)
1004 * or a lock is contended. For contention, isolate quickly to
1005 * potentially remove one source of contention.
1007 if (cc
->nr_migratepages
== COMPACT_CLUSTER_MAX
&&
1008 !cc
->rescan
&& !cc
->contended
) {
1015 if (!skip_on_failure
)
1019 * We have isolated some pages, but then failed. Release them
1020 * instead of migrating, as we cannot form the cc->order buddy
1025 spin_unlock_irqrestore(&pgdat
->lru_lock
, flags
);
1028 putback_movable_pages(&cc
->migratepages
);
1029 cc
->nr_migratepages
= 0;
1033 if (low_pfn
< next_skip_pfn
) {
1034 low_pfn
= next_skip_pfn
- 1;
1036 * The check near the loop beginning would have updated
1037 * next_skip_pfn too, but this is a bit simpler.
1039 next_skip_pfn
+= 1UL << cc
->order
;
1044 * The PageBuddy() check could have potentially brought us outside
1045 * the range to be scanned.
1047 if (unlikely(low_pfn
> end_pfn
))
1052 spin_unlock_irqrestore(&pgdat
->lru_lock
, flags
);
1055 * Updated the cached scanner pfn once the pageblock has been scanned
1056 * Pages will either be migrated in which case there is no point
1057 * scanning in the near future or migration failed in which case the
1058 * failure reason may persist. The block is marked for skipping if
1059 * there were no pages isolated in the block or if the block is
1060 * rescanned twice in a row.
1062 if (low_pfn
== end_pfn
&& (!nr_isolated
|| cc
->rescan
)) {
1063 if (valid_page
&& !skip_updated
)
1064 set_pageblock_skip(valid_page
);
1065 update_cached_migrate(cc
, low_pfn
);
1068 trace_mm_compaction_isolate_migratepages(start_pfn
, low_pfn
,
1069 nr_scanned
, nr_isolated
);
1072 cc
->total_migrate_scanned
+= nr_scanned
;
1074 count_compact_events(COMPACTISOLATED
, nr_isolated
);
1080 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
1081 * @cc: Compaction control structure.
1082 * @start_pfn: The first PFN to start isolating.
1083 * @end_pfn: The one-past-last PFN.
1085 * Returns zero if isolation fails fatally due to e.g. pending signal.
1086 * Otherwise, function returns one-past-the-last PFN of isolated page
1087 * (which may be greater than end_pfn if end fell in a middle of a THP page).
1090 isolate_migratepages_range(struct compact_control
*cc
, unsigned long start_pfn
,
1091 unsigned long end_pfn
)
1093 unsigned long pfn
, block_start_pfn
, block_end_pfn
;
1095 /* Scan block by block. First and last block may be incomplete */
1097 block_start_pfn
= pageblock_start_pfn(pfn
);
1098 if (block_start_pfn
< cc
->zone
->zone_start_pfn
)
1099 block_start_pfn
= cc
->zone
->zone_start_pfn
;
1100 block_end_pfn
= pageblock_end_pfn(pfn
);
1102 for (; pfn
< end_pfn
; pfn
= block_end_pfn
,
1103 block_start_pfn
= block_end_pfn
,
1104 block_end_pfn
+= pageblock_nr_pages
) {
1106 block_end_pfn
= min(block_end_pfn
, end_pfn
);
1108 if (!pageblock_pfn_to_page(block_start_pfn
,
1109 block_end_pfn
, cc
->zone
))
1112 pfn
= isolate_migratepages_block(cc
, pfn
, block_end_pfn
,
1113 ISOLATE_UNEVICTABLE
);
1118 if (cc
->nr_migratepages
== COMPACT_CLUSTER_MAX
)
1125 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
1126 #ifdef CONFIG_COMPACTION
1128 static bool suitable_migration_source(struct compact_control
*cc
,
1133 if (pageblock_skip_persistent(page
))
1136 if ((cc
->mode
!= MIGRATE_ASYNC
) || !cc
->direct_compaction
)
1139 block_mt
= get_pageblock_migratetype(page
);
1141 if (cc
->migratetype
== MIGRATE_MOVABLE
)
1142 return is_migrate_movable(block_mt
);
1144 return block_mt
== cc
->migratetype
;
1147 /* Returns true if the page is within a block suitable for migration to */
1148 static bool suitable_migration_target(struct compact_control
*cc
,
1151 /* If the page is a large free page, then disallow migration */
1152 if (PageBuddy(page
)) {
1154 * We are checking page_order without zone->lock taken. But
1155 * the only small danger is that we skip a potentially suitable
1156 * pageblock, so it's not worth to check order for valid range.
1158 if (page_order_unsafe(page
) >= pageblock_order
)
1162 if (cc
->ignore_block_suitable
)
1165 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1166 if (is_migrate_movable(get_pageblock_migratetype(page
)))
1169 /* Otherwise skip the block */
1173 static inline unsigned int
1174 freelist_scan_limit(struct compact_control
*cc
)
1176 unsigned short shift
= BITS_PER_LONG
- 1;
1178 return (COMPACT_CLUSTER_MAX
>> min(shift
, cc
->fast_search_fail
)) + 1;
1182 * Test whether the free scanner has reached the same or lower pageblock than
1183 * the migration scanner, and compaction should thus terminate.
1185 static inline bool compact_scanners_met(struct compact_control
*cc
)
1187 return (cc
->free_pfn
>> pageblock_order
)
1188 <= (cc
->migrate_pfn
>> pageblock_order
);
1192 * Used when scanning for a suitable migration target which scans freelists
1193 * in reverse. Reorders the list such as the unscanned pages are scanned
1194 * first on the next iteration of the free scanner
1197 move_freelist_head(struct list_head
*freelist
, struct page
*freepage
)
1201 if (!list_is_last(freelist
, &freepage
->lru
)) {
1202 list_cut_before(&sublist
, freelist
, &freepage
->lru
);
1203 if (!list_empty(&sublist
))
1204 list_splice_tail(&sublist
, freelist
);
1209 * Similar to move_freelist_head except used by the migration scanner
1210 * when scanning forward. It's possible for these list operations to
1211 * move against each other if they search the free list exactly in
1215 move_freelist_tail(struct list_head
*freelist
, struct page
*freepage
)
1219 if (!list_is_first(freelist
, &freepage
->lru
)) {
1220 list_cut_position(&sublist
, freelist
, &freepage
->lru
);
1221 if (!list_empty(&sublist
))
1222 list_splice_tail(&sublist
, freelist
);
1227 fast_isolate_around(struct compact_control
*cc
, unsigned long pfn
, unsigned long nr_isolated
)
1229 unsigned long start_pfn
, end_pfn
;
1230 struct page
*page
= pfn_to_page(pfn
);
1232 /* Do not search around if there are enough pages already */
1233 if (cc
->nr_freepages
>= cc
->nr_migratepages
)
1236 /* Minimise scanning during async compaction */
1237 if (cc
->direct_compaction
&& cc
->mode
== MIGRATE_ASYNC
)
1240 /* Pageblock boundaries */
1241 start_pfn
= pageblock_start_pfn(pfn
);
1242 end_pfn
= min(pageblock_end_pfn(pfn
), zone_end_pfn(cc
->zone
)) - 1;
1245 if (start_pfn
!= pfn
) {
1246 isolate_freepages_block(cc
, &start_pfn
, pfn
, &cc
->freepages
, 1, false);
1247 if (cc
->nr_freepages
>= cc
->nr_migratepages
)
1252 start_pfn
= pfn
+ nr_isolated
;
1253 if (start_pfn
< end_pfn
)
1254 isolate_freepages_block(cc
, &start_pfn
, end_pfn
, &cc
->freepages
, 1, false);
1256 /* Skip this pageblock in the future as it's full or nearly full */
1257 if (cc
->nr_freepages
< cc
->nr_migratepages
)
1258 set_pageblock_skip(page
);
1261 /* Search orders in round-robin fashion */
1262 static int next_search_order(struct compact_control
*cc
, int order
)
1266 order
= cc
->order
- 1;
1268 /* Search wrapped around? */
1269 if (order
== cc
->search_order
) {
1271 if (cc
->search_order
< 0)
1272 cc
->search_order
= cc
->order
- 1;
1279 static unsigned long
1280 fast_isolate_freepages(struct compact_control
*cc
)
1282 unsigned int limit
= min(1U, freelist_scan_limit(cc
) >> 1);
1283 unsigned int nr_scanned
= 0;
1284 unsigned long low_pfn
, min_pfn
, high_pfn
= 0, highest
= 0;
1285 unsigned long nr_isolated
= 0;
1286 unsigned long distance
;
1287 struct page
*page
= NULL
;
1288 bool scan_start
= false;
1291 /* Full compaction passes in a negative order */
1293 return cc
->free_pfn
;
1296 * If starting the scan, use a deeper search and use the highest
1297 * PFN found if a suitable one is not found.
1299 if (cc
->free_pfn
>= cc
->zone
->compact_init_free_pfn
) {
1300 limit
= pageblock_nr_pages
>> 1;
1305 * Preferred point is in the top quarter of the scan space but take
1306 * a pfn from the top half if the search is problematic.
1308 distance
= (cc
->free_pfn
- cc
->migrate_pfn
);
1309 low_pfn
= pageblock_start_pfn(cc
->free_pfn
- (distance
>> 2));
1310 min_pfn
= pageblock_start_pfn(cc
->free_pfn
- (distance
>> 1));
1312 if (WARN_ON_ONCE(min_pfn
> low_pfn
))
1316 * Search starts from the last successful isolation order or the next
1317 * order to search after a previous failure
1319 cc
->search_order
= min_t(unsigned int, cc
->order
- 1, cc
->search_order
);
1321 for (order
= cc
->search_order
;
1322 !page
&& order
>= 0;
1323 order
= next_search_order(cc
, order
)) {
1324 struct free_area
*area
= &cc
->zone
->free_area
[order
];
1325 struct list_head
*freelist
;
1326 struct page
*freepage
;
1327 unsigned long flags
;
1328 unsigned int order_scanned
= 0;
1333 spin_lock_irqsave(&cc
->zone
->lock
, flags
);
1334 freelist
= &area
->free_list
[MIGRATE_MOVABLE
];
1335 list_for_each_entry_reverse(freepage
, freelist
, lru
) {
1340 pfn
= page_to_pfn(freepage
);
1343 highest
= pageblock_start_pfn(pfn
);
1345 if (pfn
>= low_pfn
) {
1346 cc
->fast_search_fail
= 0;
1347 cc
->search_order
= order
;
1352 if (pfn
>= min_pfn
&& pfn
> high_pfn
) {
1355 /* Shorten the scan if a candidate is found */
1359 if (order_scanned
>= limit
)
1363 /* Use a minimum pfn if a preferred one was not found */
1364 if (!page
&& high_pfn
) {
1365 page
= pfn_to_page(high_pfn
);
1367 /* Update freepage for the list reorder below */
1371 /* Reorder to so a future search skips recent pages */
1372 move_freelist_head(freelist
, freepage
);
1374 /* Isolate the page if available */
1376 if (__isolate_free_page(page
, order
)) {
1377 set_page_private(page
, order
);
1378 nr_isolated
= 1 << order
;
1379 cc
->nr_freepages
+= nr_isolated
;
1380 list_add_tail(&page
->lru
, &cc
->freepages
);
1381 count_compact_events(COMPACTISOLATED
, nr_isolated
);
1383 /* If isolation fails, abort the search */
1384 order
= cc
->search_order
+ 1;
1389 spin_unlock_irqrestore(&cc
->zone
->lock
, flags
);
1392 * Smaller scan on next order so the total scan ig related
1393 * to freelist_scan_limit.
1395 if (order_scanned
>= limit
)
1396 limit
= min(1U, limit
>> 1);
1400 cc
->fast_search_fail
++;
1403 * Use the highest PFN found above min. If one was
1404 * not found, be pessemistic for direct compaction
1405 * and use the min mark.
1408 page
= pfn_to_page(highest
);
1409 cc
->free_pfn
= highest
;
1411 if (cc
->direct_compaction
&& pfn_valid(min_pfn
)) {
1412 page
= pfn_to_page(min_pfn
);
1413 cc
->free_pfn
= min_pfn
;
1419 if (highest
&& highest
>= cc
->zone
->compact_cached_free_pfn
) {
1420 highest
-= pageblock_nr_pages
;
1421 cc
->zone
->compact_cached_free_pfn
= highest
;
1424 cc
->total_free_scanned
+= nr_scanned
;
1426 return cc
->free_pfn
;
1428 low_pfn
= page_to_pfn(page
);
1429 fast_isolate_around(cc
, low_pfn
, nr_isolated
);
1434 * Based on information in the current compact_control, find blocks
1435 * suitable for isolating free pages from and then isolate them.
1437 static void isolate_freepages(struct compact_control
*cc
)
1439 struct zone
*zone
= cc
->zone
;
1441 unsigned long block_start_pfn
; /* start of current pageblock */
1442 unsigned long isolate_start_pfn
; /* exact pfn we start at */
1443 unsigned long block_end_pfn
; /* end of current pageblock */
1444 unsigned long low_pfn
; /* lowest pfn scanner is able to scan */
1445 struct list_head
*freelist
= &cc
->freepages
;
1446 unsigned int stride
;
1448 /* Try a small search of the free lists for a candidate */
1449 isolate_start_pfn
= fast_isolate_freepages(cc
);
1450 if (cc
->nr_freepages
)
1454 * Initialise the free scanner. The starting point is where we last
1455 * successfully isolated from, zone-cached value, or the end of the
1456 * zone when isolating for the first time. For looping we also need
1457 * this pfn aligned down to the pageblock boundary, because we do
1458 * block_start_pfn -= pageblock_nr_pages in the for loop.
1459 * For ending point, take care when isolating in last pageblock of a
1460 * a zone which ends in the middle of a pageblock.
1461 * The low boundary is the end of the pageblock the migration scanner
1464 isolate_start_pfn
= cc
->free_pfn
;
1465 block_start_pfn
= pageblock_start_pfn(isolate_start_pfn
);
1466 block_end_pfn
= min(block_start_pfn
+ pageblock_nr_pages
,
1467 zone_end_pfn(zone
));
1468 low_pfn
= pageblock_end_pfn(cc
->migrate_pfn
);
1469 stride
= cc
->mode
== MIGRATE_ASYNC
? COMPACT_CLUSTER_MAX
: 1;
1472 * Isolate free pages until enough are available to migrate the
1473 * pages on cc->migratepages. We stop searching if the migrate
1474 * and free page scanners meet or enough free pages are isolated.
1476 for (; block_start_pfn
>= low_pfn
;
1477 block_end_pfn
= block_start_pfn
,
1478 block_start_pfn
-= pageblock_nr_pages
,
1479 isolate_start_pfn
= block_start_pfn
) {
1480 unsigned long nr_isolated
;
1483 * This can iterate a massively long zone without finding any
1484 * suitable migration targets, so periodically check resched.
1486 if (!(block_start_pfn
% (SWAP_CLUSTER_MAX
* pageblock_nr_pages
)))
1489 page
= pageblock_pfn_to_page(block_start_pfn
, block_end_pfn
,
1494 /* Check the block is suitable for migration */
1495 if (!suitable_migration_target(cc
, page
))
1498 /* If isolation recently failed, do not retry */
1499 if (!isolation_suitable(cc
, page
))
1502 /* Found a block suitable for isolating free pages from. */
1503 nr_isolated
= isolate_freepages_block(cc
, &isolate_start_pfn
,
1504 block_end_pfn
, freelist
, stride
, false);
1506 /* Update the skip hint if the full pageblock was scanned */
1507 if (isolate_start_pfn
== block_end_pfn
)
1508 update_pageblock_skip(cc
, page
, block_start_pfn
);
1510 /* Are enough freepages isolated? */
1511 if (cc
->nr_freepages
>= cc
->nr_migratepages
) {
1512 if (isolate_start_pfn
>= block_end_pfn
) {
1514 * Restart at previous pageblock if more
1515 * freepages can be isolated next time.
1518 block_start_pfn
- pageblock_nr_pages
;
1521 } else if (isolate_start_pfn
< block_end_pfn
) {
1523 * If isolation failed early, do not continue
1529 /* Adjust stride depending on isolation */
1534 stride
= min_t(unsigned int, COMPACT_CLUSTER_MAX
, stride
<< 1);
1538 * Record where the free scanner will restart next time. Either we
1539 * broke from the loop and set isolate_start_pfn based on the last
1540 * call to isolate_freepages_block(), or we met the migration scanner
1541 * and the loop terminated due to isolate_start_pfn < low_pfn
1543 cc
->free_pfn
= isolate_start_pfn
;
1546 /* __isolate_free_page() does not map the pages */
1547 split_map_pages(freelist
);
1551 * This is a migrate-callback that "allocates" freepages by taking pages
1552 * from the isolated freelists in the block we are migrating to.
1554 static struct page
*compaction_alloc(struct page
*migratepage
,
1557 struct compact_control
*cc
= (struct compact_control
*)data
;
1558 struct page
*freepage
;
1560 if (list_empty(&cc
->freepages
)) {
1561 isolate_freepages(cc
);
1563 if (list_empty(&cc
->freepages
))
1567 freepage
= list_entry(cc
->freepages
.next
, struct page
, lru
);
1568 list_del(&freepage
->lru
);
1575 * This is a migrate-callback that "frees" freepages back to the isolated
1576 * freelist. All pages on the freelist are from the same zone, so there is no
1577 * special handling needed for NUMA.
1579 static void compaction_free(struct page
*page
, unsigned long data
)
1581 struct compact_control
*cc
= (struct compact_control
*)data
;
1583 list_add(&page
->lru
, &cc
->freepages
);
1587 /* possible outcome of isolate_migratepages */
1589 ISOLATE_ABORT
, /* Abort compaction now */
1590 ISOLATE_NONE
, /* No pages isolated, continue scanning */
1591 ISOLATE_SUCCESS
, /* Pages isolated, migrate */
1592 } isolate_migrate_t
;
1595 * Allow userspace to control policy on scanning the unevictable LRU for
1596 * compactable pages.
1598 #ifdef CONFIG_PREEMPT_RT
1599 int sysctl_compact_unevictable_allowed __read_mostly
= 0;
1601 int sysctl_compact_unevictable_allowed __read_mostly
= 1;
1605 update_fast_start_pfn(struct compact_control
*cc
, unsigned long pfn
)
1607 if (cc
->fast_start_pfn
== ULONG_MAX
)
1610 if (!cc
->fast_start_pfn
)
1611 cc
->fast_start_pfn
= pfn
;
1613 cc
->fast_start_pfn
= min(cc
->fast_start_pfn
, pfn
);
1616 static inline unsigned long
1617 reinit_migrate_pfn(struct compact_control
*cc
)
1619 if (!cc
->fast_start_pfn
|| cc
->fast_start_pfn
== ULONG_MAX
)
1620 return cc
->migrate_pfn
;
1622 cc
->migrate_pfn
= cc
->fast_start_pfn
;
1623 cc
->fast_start_pfn
= ULONG_MAX
;
1625 return cc
->migrate_pfn
;
1629 * Briefly search the free lists for a migration source that already has
1630 * some free pages to reduce the number of pages that need migration
1631 * before a pageblock is free.
1633 static unsigned long fast_find_migrateblock(struct compact_control
*cc
)
1635 unsigned int limit
= freelist_scan_limit(cc
);
1636 unsigned int nr_scanned
= 0;
1637 unsigned long distance
;
1638 unsigned long pfn
= cc
->migrate_pfn
;
1639 unsigned long high_pfn
;
1642 /* Skip hints are relied on to avoid repeats on the fast search */
1643 if (cc
->ignore_skip_hint
)
1647 * If the migrate_pfn is not at the start of a zone or the start
1648 * of a pageblock then assume this is a continuation of a previous
1649 * scan restarted due to COMPACT_CLUSTER_MAX.
1651 if (pfn
!= cc
->zone
->zone_start_pfn
&& pfn
!= pageblock_start_pfn(pfn
))
1655 * For smaller orders, just linearly scan as the number of pages
1656 * to migrate should be relatively small and does not necessarily
1657 * justify freeing up a large block for a small allocation.
1659 if (cc
->order
<= PAGE_ALLOC_COSTLY_ORDER
)
1663 * Only allow kcompactd and direct requests for movable pages to
1664 * quickly clear out a MOVABLE pageblock for allocation. This
1665 * reduces the risk that a large movable pageblock is freed for
1666 * an unmovable/reclaimable small allocation.
1668 if (cc
->direct_compaction
&& cc
->migratetype
!= MIGRATE_MOVABLE
)
1672 * When starting the migration scanner, pick any pageblock within the
1673 * first half of the search space. Otherwise try and pick a pageblock
1674 * within the first eighth to reduce the chances that a migration
1675 * target later becomes a source.
1677 distance
= (cc
->free_pfn
- cc
->migrate_pfn
) >> 1;
1678 if (cc
->migrate_pfn
!= cc
->zone
->zone_start_pfn
)
1680 high_pfn
= pageblock_start_pfn(cc
->migrate_pfn
+ distance
);
1682 for (order
= cc
->order
- 1;
1683 order
>= PAGE_ALLOC_COSTLY_ORDER
&& pfn
== cc
->migrate_pfn
&& nr_scanned
< limit
;
1685 struct free_area
*area
= &cc
->zone
->free_area
[order
];
1686 struct list_head
*freelist
;
1687 unsigned long flags
;
1688 struct page
*freepage
;
1693 spin_lock_irqsave(&cc
->zone
->lock
, flags
);
1694 freelist
= &area
->free_list
[MIGRATE_MOVABLE
];
1695 list_for_each_entry(freepage
, freelist
, lru
) {
1696 unsigned long free_pfn
;
1699 free_pfn
= page_to_pfn(freepage
);
1700 if (free_pfn
< high_pfn
) {
1702 * Avoid if skipped recently. Ideally it would
1703 * move to the tail but even safe iteration of
1704 * the list assumes an entry is deleted, not
1707 if (get_pageblock_skip(freepage
)) {
1708 if (list_is_last(freelist
, &freepage
->lru
))
1714 /* Reorder to so a future search skips recent pages */
1715 move_freelist_tail(freelist
, freepage
);
1717 update_fast_start_pfn(cc
, free_pfn
);
1718 pfn
= pageblock_start_pfn(free_pfn
);
1719 cc
->fast_search_fail
= 0;
1720 set_pageblock_skip(freepage
);
1724 if (nr_scanned
>= limit
) {
1725 cc
->fast_search_fail
++;
1726 move_freelist_tail(freelist
, freepage
);
1730 spin_unlock_irqrestore(&cc
->zone
->lock
, flags
);
1733 cc
->total_migrate_scanned
+= nr_scanned
;
1736 * If fast scanning failed then use a cached entry for a page block
1737 * that had free pages as the basis for starting a linear scan.
1739 if (pfn
== cc
->migrate_pfn
)
1740 pfn
= reinit_migrate_pfn(cc
);
1746 * Isolate all pages that can be migrated from the first suitable block,
1747 * starting at the block pointed to by the migrate scanner pfn within
1750 static isolate_migrate_t
isolate_migratepages(struct compact_control
*cc
)
1752 unsigned long block_start_pfn
;
1753 unsigned long block_end_pfn
;
1754 unsigned long low_pfn
;
1756 const isolate_mode_t isolate_mode
=
1757 (sysctl_compact_unevictable_allowed
? ISOLATE_UNEVICTABLE
: 0) |
1758 (cc
->mode
!= MIGRATE_SYNC
? ISOLATE_ASYNC_MIGRATE
: 0);
1759 bool fast_find_block
;
1762 * Start at where we last stopped, or beginning of the zone as
1763 * initialized by compact_zone(). The first failure will use
1764 * the lowest PFN as the starting point for linear scanning.
1766 low_pfn
= fast_find_migrateblock(cc
);
1767 block_start_pfn
= pageblock_start_pfn(low_pfn
);
1768 if (block_start_pfn
< cc
->zone
->zone_start_pfn
)
1769 block_start_pfn
= cc
->zone
->zone_start_pfn
;
1772 * fast_find_migrateblock marks a pageblock skipped so to avoid
1773 * the isolation_suitable check below, check whether the fast
1774 * search was successful.
1776 fast_find_block
= low_pfn
!= cc
->migrate_pfn
&& !cc
->fast_search_fail
;
1778 /* Only scan within a pageblock boundary */
1779 block_end_pfn
= pageblock_end_pfn(low_pfn
);
1782 * Iterate over whole pageblocks until we find the first suitable.
1783 * Do not cross the free scanner.
1785 for (; block_end_pfn
<= cc
->free_pfn
;
1786 fast_find_block
= false,
1787 low_pfn
= block_end_pfn
,
1788 block_start_pfn
= block_end_pfn
,
1789 block_end_pfn
+= pageblock_nr_pages
) {
1792 * This can potentially iterate a massively long zone with
1793 * many pageblocks unsuitable, so periodically check if we
1796 if (!(low_pfn
% (SWAP_CLUSTER_MAX
* pageblock_nr_pages
)))
1799 page
= pageblock_pfn_to_page(block_start_pfn
,
1800 block_end_pfn
, cc
->zone
);
1805 * If isolation recently failed, do not retry. Only check the
1806 * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock
1807 * to be visited multiple times. Assume skip was checked
1808 * before making it "skip" so other compaction instances do
1809 * not scan the same block.
1811 if (IS_ALIGNED(low_pfn
, pageblock_nr_pages
) &&
1812 !fast_find_block
&& !isolation_suitable(cc
, page
))
1816 * For async compaction, also only scan in MOVABLE blocks
1817 * without huge pages. Async compaction is optimistic to see
1818 * if the minimum amount of work satisfies the allocation.
1819 * The cached PFN is updated as it's possible that all
1820 * remaining blocks between source and target are unsuitable
1821 * and the compaction scanners fail to meet.
1823 if (!suitable_migration_source(cc
, page
)) {
1824 update_cached_migrate(cc
, block_end_pfn
);
1828 /* Perform the isolation */
1829 low_pfn
= isolate_migratepages_block(cc
, low_pfn
,
1830 block_end_pfn
, isolate_mode
);
1833 return ISOLATE_ABORT
;
1836 * Either we isolated something and proceed with migration. Or
1837 * we failed and compact_zone should decide if we should
1843 /* Record where migration scanner will be restarted. */
1844 cc
->migrate_pfn
= low_pfn
;
1846 return cc
->nr_migratepages
? ISOLATE_SUCCESS
: ISOLATE_NONE
;
1850 * order == -1 is expected when compacting via
1851 * /proc/sys/vm/compact_memory
1853 static inline bool is_via_compact_memory(int order
)
1858 static enum compact_result
__compact_finished(struct compact_control
*cc
)
1861 const int migratetype
= cc
->migratetype
;
1864 /* Compaction run completes if the migrate and free scanner meet */
1865 if (compact_scanners_met(cc
)) {
1866 /* Let the next compaction start anew. */
1867 reset_cached_positions(cc
->zone
);
1870 * Mark that the PG_migrate_skip information should be cleared
1871 * by kswapd when it goes to sleep. kcompactd does not set the
1872 * flag itself as the decision to be clear should be directly
1873 * based on an allocation request.
1875 if (cc
->direct_compaction
)
1876 cc
->zone
->compact_blockskip_flush
= true;
1879 return COMPACT_COMPLETE
;
1881 return COMPACT_PARTIAL_SKIPPED
;
1884 if (is_via_compact_memory(cc
->order
))
1885 return COMPACT_CONTINUE
;
1888 * Always finish scanning a pageblock to reduce the possibility of
1889 * fallbacks in the future. This is particularly important when
1890 * migration source is unmovable/reclaimable but it's not worth
1893 if (!IS_ALIGNED(cc
->migrate_pfn
, pageblock_nr_pages
))
1894 return COMPACT_CONTINUE
;
1896 /* Direct compactor: Is a suitable page free? */
1897 ret
= COMPACT_NO_SUITABLE_PAGE
;
1898 for (order
= cc
->order
; order
< MAX_ORDER
; order
++) {
1899 struct free_area
*area
= &cc
->zone
->free_area
[order
];
1902 /* Job done if page is free of the right migratetype */
1903 if (!free_area_empty(area
, migratetype
))
1904 return COMPACT_SUCCESS
;
1907 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
1908 if (migratetype
== MIGRATE_MOVABLE
&&
1909 !free_area_empty(area
, MIGRATE_CMA
))
1910 return COMPACT_SUCCESS
;
1913 * Job done if allocation would steal freepages from
1914 * other migratetype buddy lists.
1916 if (find_suitable_fallback(area
, order
, migratetype
,
1917 true, &can_steal
) != -1) {
1919 /* movable pages are OK in any pageblock */
1920 if (migratetype
== MIGRATE_MOVABLE
)
1921 return COMPACT_SUCCESS
;
1924 * We are stealing for a non-movable allocation. Make
1925 * sure we finish compacting the current pageblock
1926 * first so it is as free as possible and we won't
1927 * have to steal another one soon. This only applies
1928 * to sync compaction, as async compaction operates
1929 * on pageblocks of the same migratetype.
1931 if (cc
->mode
== MIGRATE_ASYNC
||
1932 IS_ALIGNED(cc
->migrate_pfn
,
1933 pageblock_nr_pages
)) {
1934 return COMPACT_SUCCESS
;
1937 ret
= COMPACT_CONTINUE
;
1942 if (cc
->contended
|| fatal_signal_pending(current
))
1943 ret
= COMPACT_CONTENDED
;
1948 static enum compact_result
compact_finished(struct compact_control
*cc
)
1952 ret
= __compact_finished(cc
);
1953 trace_mm_compaction_finished(cc
->zone
, cc
->order
, ret
);
1954 if (ret
== COMPACT_NO_SUITABLE_PAGE
)
1955 ret
= COMPACT_CONTINUE
;
1961 * compaction_suitable: Is this suitable to run compaction on this zone now?
1963 * COMPACT_SKIPPED - If there are too few free pages for compaction
1964 * COMPACT_SUCCESS - If the allocation would succeed without compaction
1965 * COMPACT_CONTINUE - If compaction should run now
1967 static enum compact_result
__compaction_suitable(struct zone
*zone
, int order
,
1968 unsigned int alloc_flags
,
1970 unsigned long wmark_target
)
1972 unsigned long watermark
;
1974 if (is_via_compact_memory(order
))
1975 return COMPACT_CONTINUE
;
1977 watermark
= wmark_pages(zone
, alloc_flags
& ALLOC_WMARK_MASK
);
1979 * If watermarks for high-order allocation are already met, there
1980 * should be no need for compaction at all.
1982 if (zone_watermark_ok(zone
, order
, watermark
, classzone_idx
,
1984 return COMPACT_SUCCESS
;
1987 * Watermarks for order-0 must be met for compaction to be able to
1988 * isolate free pages for migration targets. This means that the
1989 * watermark and alloc_flags have to match, or be more pessimistic than
1990 * the check in __isolate_free_page(). We don't use the direct
1991 * compactor's alloc_flags, as they are not relevant for freepage
1992 * isolation. We however do use the direct compactor's classzone_idx to
1993 * skip over zones where lowmem reserves would prevent allocation even
1994 * if compaction succeeds.
1995 * For costly orders, we require low watermark instead of min for
1996 * compaction to proceed to increase its chances.
1997 * ALLOC_CMA is used, as pages in CMA pageblocks are considered
1998 * suitable migration targets
2000 watermark
= (order
> PAGE_ALLOC_COSTLY_ORDER
) ?
2001 low_wmark_pages(zone
) : min_wmark_pages(zone
);
2002 watermark
+= compact_gap(order
);
2003 if (!__zone_watermark_ok(zone
, 0, watermark
, classzone_idx
,
2004 ALLOC_CMA
, wmark_target
))
2005 return COMPACT_SKIPPED
;
2007 return COMPACT_CONTINUE
;
2010 enum compact_result
compaction_suitable(struct zone
*zone
, int order
,
2011 unsigned int alloc_flags
,
2014 enum compact_result ret
;
2017 ret
= __compaction_suitable(zone
, order
, alloc_flags
, classzone_idx
,
2018 zone_page_state(zone
, NR_FREE_PAGES
));
2020 * fragmentation index determines if allocation failures are due to
2021 * low memory or external fragmentation
2023 * index of -1000 would imply allocations might succeed depending on
2024 * watermarks, but we already failed the high-order watermark check
2025 * index towards 0 implies failure is due to lack of memory
2026 * index towards 1000 implies failure is due to fragmentation
2028 * Only compact if a failure would be due to fragmentation. Also
2029 * ignore fragindex for non-costly orders where the alternative to
2030 * a successful reclaim/compaction is OOM. Fragindex and the
2031 * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
2032 * excessive compaction for costly orders, but it should not be at the
2033 * expense of system stability.
2035 if (ret
== COMPACT_CONTINUE
&& (order
> PAGE_ALLOC_COSTLY_ORDER
)) {
2036 fragindex
= fragmentation_index(zone
, order
);
2037 if (fragindex
>= 0 && fragindex
<= sysctl_extfrag_threshold
)
2038 ret
= COMPACT_NOT_SUITABLE_ZONE
;
2041 trace_mm_compaction_suitable(zone
, order
, ret
);
2042 if (ret
== COMPACT_NOT_SUITABLE_ZONE
)
2043 ret
= COMPACT_SKIPPED
;
2048 bool compaction_zonelist_suitable(struct alloc_context
*ac
, int order
,
2055 * Make sure at least one zone would pass __compaction_suitable if we continue
2056 * retrying the reclaim.
2058 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
2060 unsigned long available
;
2061 enum compact_result compact_result
;
2064 * Do not consider all the reclaimable memory because we do not
2065 * want to trash just for a single high order allocation which
2066 * is even not guaranteed to appear even if __compaction_suitable
2067 * is happy about the watermark check.
2069 available
= zone_reclaimable_pages(zone
) / order
;
2070 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
2071 compact_result
= __compaction_suitable(zone
, order
, alloc_flags
,
2072 ac_classzone_idx(ac
), available
);
2073 if (compact_result
!= COMPACT_SKIPPED
)
2080 static enum compact_result
2081 compact_zone(struct compact_control
*cc
, struct capture_control
*capc
)
2083 enum compact_result ret
;
2084 unsigned long start_pfn
= cc
->zone
->zone_start_pfn
;
2085 unsigned long end_pfn
= zone_end_pfn(cc
->zone
);
2086 unsigned long last_migrated_pfn
;
2087 const bool sync
= cc
->mode
!= MIGRATE_ASYNC
;
2091 * These counters track activities during zone compaction. Initialize
2092 * them before compacting a new zone.
2094 cc
->total_migrate_scanned
= 0;
2095 cc
->total_free_scanned
= 0;
2096 cc
->nr_migratepages
= 0;
2097 cc
->nr_freepages
= 0;
2098 INIT_LIST_HEAD(&cc
->freepages
);
2099 INIT_LIST_HEAD(&cc
->migratepages
);
2101 cc
->migratetype
= gfpflags_to_migratetype(cc
->gfp_mask
);
2102 ret
= compaction_suitable(cc
->zone
, cc
->order
, cc
->alloc_flags
,
2104 /* Compaction is likely to fail */
2105 if (ret
== COMPACT_SUCCESS
|| ret
== COMPACT_SKIPPED
)
2108 /* huh, compaction_suitable is returning something unexpected */
2109 VM_BUG_ON(ret
!= COMPACT_CONTINUE
);
2112 * Clear pageblock skip if there were failures recently and compaction
2113 * is about to be retried after being deferred.
2115 if (compaction_restarting(cc
->zone
, cc
->order
))
2116 __reset_isolation_suitable(cc
->zone
);
2119 * Setup to move all movable pages to the end of the zone. Used cached
2120 * information on where the scanners should start (unless we explicitly
2121 * want to compact the whole zone), but check that it is initialised
2122 * by ensuring the values are within zone boundaries.
2124 cc
->fast_start_pfn
= 0;
2125 if (cc
->whole_zone
) {
2126 cc
->migrate_pfn
= start_pfn
;
2127 cc
->free_pfn
= pageblock_start_pfn(end_pfn
- 1);
2129 cc
->migrate_pfn
= cc
->zone
->compact_cached_migrate_pfn
[sync
];
2130 cc
->free_pfn
= cc
->zone
->compact_cached_free_pfn
;
2131 if (cc
->free_pfn
< start_pfn
|| cc
->free_pfn
>= end_pfn
) {
2132 cc
->free_pfn
= pageblock_start_pfn(end_pfn
- 1);
2133 cc
->zone
->compact_cached_free_pfn
= cc
->free_pfn
;
2135 if (cc
->migrate_pfn
< start_pfn
|| cc
->migrate_pfn
>= end_pfn
) {
2136 cc
->migrate_pfn
= start_pfn
;
2137 cc
->zone
->compact_cached_migrate_pfn
[0] = cc
->migrate_pfn
;
2138 cc
->zone
->compact_cached_migrate_pfn
[1] = cc
->migrate_pfn
;
2141 if (cc
->migrate_pfn
<= cc
->zone
->compact_init_migrate_pfn
)
2142 cc
->whole_zone
= true;
2145 last_migrated_pfn
= 0;
2148 * Migrate has separate cached PFNs for ASYNC and SYNC* migration on
2149 * the basis that some migrations will fail in ASYNC mode. However,
2150 * if the cached PFNs match and pageblocks are skipped due to having
2151 * no isolation candidates, then the sync state does not matter.
2152 * Until a pageblock with isolation candidates is found, keep the
2153 * cached PFNs in sync to avoid revisiting the same blocks.
2155 update_cached
= !sync
&&
2156 cc
->zone
->compact_cached_migrate_pfn
[0] == cc
->zone
->compact_cached_migrate_pfn
[1];
2158 trace_mm_compaction_begin(start_pfn
, cc
->migrate_pfn
,
2159 cc
->free_pfn
, end_pfn
, sync
);
2161 migrate_prep_local();
2163 while ((ret
= compact_finished(cc
)) == COMPACT_CONTINUE
) {
2165 unsigned long start_pfn
= cc
->migrate_pfn
;
2168 * Avoid multiple rescans which can happen if a page cannot be
2169 * isolated (dirty/writeback in async mode) or if the migrated
2170 * pages are being allocated before the pageblock is cleared.
2171 * The first rescan will capture the entire pageblock for
2172 * migration. If it fails, it'll be marked skip and scanning
2173 * will proceed as normal.
2176 if (pageblock_start_pfn(last_migrated_pfn
) ==
2177 pageblock_start_pfn(start_pfn
)) {
2181 switch (isolate_migratepages(cc
)) {
2183 ret
= COMPACT_CONTENDED
;
2184 putback_movable_pages(&cc
->migratepages
);
2185 cc
->nr_migratepages
= 0;
2188 if (update_cached
) {
2189 cc
->zone
->compact_cached_migrate_pfn
[1] =
2190 cc
->zone
->compact_cached_migrate_pfn
[0];
2194 * We haven't isolated and migrated anything, but
2195 * there might still be unflushed migrations from
2196 * previous cc->order aligned block.
2199 case ISOLATE_SUCCESS
:
2200 update_cached
= false;
2201 last_migrated_pfn
= start_pfn
;
2205 err
= migrate_pages(&cc
->migratepages
, compaction_alloc
,
2206 compaction_free
, (unsigned long)cc
, cc
->mode
,
2209 trace_mm_compaction_migratepages(cc
->nr_migratepages
, err
,
2212 /* All pages were either migrated or will be released */
2213 cc
->nr_migratepages
= 0;
2215 putback_movable_pages(&cc
->migratepages
);
2217 * migrate_pages() may return -ENOMEM when scanners meet
2218 * and we want compact_finished() to detect it
2220 if (err
== -ENOMEM
&& !compact_scanners_met(cc
)) {
2221 ret
= COMPACT_CONTENDED
;
2225 * We failed to migrate at least one page in the current
2226 * order-aligned block, so skip the rest of it.
2228 if (cc
->direct_compaction
&&
2229 (cc
->mode
== MIGRATE_ASYNC
)) {
2230 cc
->migrate_pfn
= block_end_pfn(
2231 cc
->migrate_pfn
- 1, cc
->order
);
2232 /* Draining pcplists is useless in this case */
2233 last_migrated_pfn
= 0;
2239 * Has the migration scanner moved away from the previous
2240 * cc->order aligned block where we migrated from? If yes,
2241 * flush the pages that were freed, so that they can merge and
2242 * compact_finished() can detect immediately if allocation
2245 if (cc
->order
> 0 && last_migrated_pfn
) {
2246 unsigned long current_block_start
=
2247 block_start_pfn(cc
->migrate_pfn
, cc
->order
);
2249 if (last_migrated_pfn
< current_block_start
) {
2250 lru_add_drain_cpu_zone(cc
->zone
);
2251 /* No more flushing until we migrate again */
2252 last_migrated_pfn
= 0;
2256 /* Stop if a page has been captured */
2257 if (capc
&& capc
->page
) {
2258 ret
= COMPACT_SUCCESS
;
2265 * Release free pages and update where the free scanner should restart,
2266 * so we don't leave any returned pages behind in the next attempt.
2268 if (cc
->nr_freepages
> 0) {
2269 unsigned long free_pfn
= release_freepages(&cc
->freepages
);
2271 cc
->nr_freepages
= 0;
2272 VM_BUG_ON(free_pfn
== 0);
2273 /* The cached pfn is always the first in a pageblock */
2274 free_pfn
= pageblock_start_pfn(free_pfn
);
2276 * Only go back, not forward. The cached pfn might have been
2277 * already reset to zone end in compact_finished()
2279 if (free_pfn
> cc
->zone
->compact_cached_free_pfn
)
2280 cc
->zone
->compact_cached_free_pfn
= free_pfn
;
2283 count_compact_events(COMPACTMIGRATE_SCANNED
, cc
->total_migrate_scanned
);
2284 count_compact_events(COMPACTFREE_SCANNED
, cc
->total_free_scanned
);
2286 trace_mm_compaction_end(start_pfn
, cc
->migrate_pfn
,
2287 cc
->free_pfn
, end_pfn
, sync
, ret
);
2292 static enum compact_result
compact_zone_order(struct zone
*zone
, int order
,
2293 gfp_t gfp_mask
, enum compact_priority prio
,
2294 unsigned int alloc_flags
, int classzone_idx
,
2295 struct page
**capture
)
2297 enum compact_result ret
;
2298 struct compact_control cc
= {
2300 .search_order
= order
,
2301 .gfp_mask
= gfp_mask
,
2303 .mode
= (prio
== COMPACT_PRIO_ASYNC
) ?
2304 MIGRATE_ASYNC
: MIGRATE_SYNC_LIGHT
,
2305 .alloc_flags
= alloc_flags
,
2306 .classzone_idx
= classzone_idx
,
2307 .direct_compaction
= true,
2308 .whole_zone
= (prio
== MIN_COMPACT_PRIORITY
),
2309 .ignore_skip_hint
= (prio
== MIN_COMPACT_PRIORITY
),
2310 .ignore_block_suitable
= (prio
== MIN_COMPACT_PRIORITY
)
2312 struct capture_control capc
= {
2317 current
->capture_control
= &capc
;
2319 ret
= compact_zone(&cc
, &capc
);
2321 VM_BUG_ON(!list_empty(&cc
.freepages
));
2322 VM_BUG_ON(!list_empty(&cc
.migratepages
));
2324 *capture
= capc
.page
;
2325 current
->capture_control
= NULL
;
2330 int sysctl_extfrag_threshold
= 500;
2333 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
2334 * @gfp_mask: The GFP mask of the current allocation
2335 * @order: The order of the current allocation
2336 * @alloc_flags: The allocation flags of the current allocation
2337 * @ac: The context of current allocation
2338 * @prio: Determines how hard direct compaction should try to succeed
2339 * @capture: Pointer to free page created by compaction will be stored here
2341 * This is the main entry point for direct page compaction.
2343 enum compact_result
try_to_compact_pages(gfp_t gfp_mask
, unsigned int order
,
2344 unsigned int alloc_flags
, const struct alloc_context
*ac
,
2345 enum compact_priority prio
, struct page
**capture
)
2347 int may_perform_io
= gfp_mask
& __GFP_IO
;
2350 enum compact_result rc
= COMPACT_SKIPPED
;
2353 * Check if the GFP flags allow compaction - GFP_NOIO is really
2354 * tricky context because the migration might require IO
2356 if (!may_perform_io
)
2357 return COMPACT_SKIPPED
;
2359 trace_mm_compaction_try_to_compact_pages(order
, gfp_mask
, prio
);
2361 /* Compact each zone in the list */
2362 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
2364 enum compact_result status
;
2366 if (prio
> MIN_COMPACT_PRIORITY
2367 && compaction_deferred(zone
, order
)) {
2368 rc
= max_t(enum compact_result
, COMPACT_DEFERRED
, rc
);
2372 status
= compact_zone_order(zone
, order
, gfp_mask
, prio
,
2373 alloc_flags
, ac_classzone_idx(ac
), capture
);
2374 rc
= max(status
, rc
);
2376 /* The allocation should succeed, stop compacting */
2377 if (status
== COMPACT_SUCCESS
) {
2379 * We think the allocation will succeed in this zone,
2380 * but it is not certain, hence the false. The caller
2381 * will repeat this with true if allocation indeed
2382 * succeeds in this zone.
2384 compaction_defer_reset(zone
, order
, false);
2389 if (prio
!= COMPACT_PRIO_ASYNC
&& (status
== COMPACT_COMPLETE
||
2390 status
== COMPACT_PARTIAL_SKIPPED
))
2392 * We think that allocation won't succeed in this zone
2393 * so we defer compaction there. If it ends up
2394 * succeeding after all, it will be reset.
2396 defer_compaction(zone
, order
);
2399 * We might have stopped compacting due to need_resched() in
2400 * async compaction, or due to a fatal signal detected. In that
2401 * case do not try further zones
2403 if ((prio
== COMPACT_PRIO_ASYNC
&& need_resched())
2404 || fatal_signal_pending(current
))
2412 /* Compact all zones within a node */
2413 static void compact_node(int nid
)
2415 pg_data_t
*pgdat
= NODE_DATA(nid
);
2418 struct compact_control cc
= {
2420 .mode
= MIGRATE_SYNC
,
2421 .ignore_skip_hint
= true,
2423 .gfp_mask
= GFP_KERNEL
,
2427 for (zoneid
= 0; zoneid
< MAX_NR_ZONES
; zoneid
++) {
2429 zone
= &pgdat
->node_zones
[zoneid
];
2430 if (!populated_zone(zone
))
2435 compact_zone(&cc
, NULL
);
2437 VM_BUG_ON(!list_empty(&cc
.freepages
));
2438 VM_BUG_ON(!list_empty(&cc
.migratepages
));
2442 /* Compact all nodes in the system */
2443 static void compact_nodes(void)
2447 /* Flush pending updates to the LRU lists */
2448 lru_add_drain_all();
2450 for_each_online_node(nid
)
2454 /* The written value is actually unused, all memory is compacted */
2455 int sysctl_compact_memory
;
2458 * This is the entry point for compacting all nodes via
2459 * /proc/sys/vm/compact_memory
2461 int sysctl_compaction_handler(struct ctl_table
*table
, int write
,
2462 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2470 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
2471 static ssize_t
sysfs_compact_node(struct device
*dev
,
2472 struct device_attribute
*attr
,
2473 const char *buf
, size_t count
)
2477 if (nid
>= 0 && nid
< nr_node_ids
&& node_online(nid
)) {
2478 /* Flush pending updates to the LRU lists */
2479 lru_add_drain_all();
2486 static DEVICE_ATTR(compact
, 0200, NULL
, sysfs_compact_node
);
2488 int compaction_register_node(struct node
*node
)
2490 return device_create_file(&node
->dev
, &dev_attr_compact
);
2493 void compaction_unregister_node(struct node
*node
)
2495 return device_remove_file(&node
->dev
, &dev_attr_compact
);
2497 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
2499 static inline bool kcompactd_work_requested(pg_data_t
*pgdat
)
2501 return pgdat
->kcompactd_max_order
> 0 || kthread_should_stop();
2504 static bool kcompactd_node_suitable(pg_data_t
*pgdat
)
2508 enum zone_type classzone_idx
= pgdat
->kcompactd_classzone_idx
;
2510 for (zoneid
= 0; zoneid
<= classzone_idx
; zoneid
++) {
2511 zone
= &pgdat
->node_zones
[zoneid
];
2513 if (!populated_zone(zone
))
2516 if (compaction_suitable(zone
, pgdat
->kcompactd_max_order
, 0,
2517 classzone_idx
) == COMPACT_CONTINUE
)
2524 static void kcompactd_do_work(pg_data_t
*pgdat
)
2527 * With no special task, compact all zones so that a page of requested
2528 * order is allocatable.
2532 struct compact_control cc
= {
2533 .order
= pgdat
->kcompactd_max_order
,
2534 .search_order
= pgdat
->kcompactd_max_order
,
2535 .classzone_idx
= pgdat
->kcompactd_classzone_idx
,
2536 .mode
= MIGRATE_SYNC_LIGHT
,
2537 .ignore_skip_hint
= false,
2538 .gfp_mask
= GFP_KERNEL
,
2540 trace_mm_compaction_kcompactd_wake(pgdat
->node_id
, cc
.order
,
2542 count_compact_event(KCOMPACTD_WAKE
);
2544 for (zoneid
= 0; zoneid
<= cc
.classzone_idx
; zoneid
++) {
2547 zone
= &pgdat
->node_zones
[zoneid
];
2548 if (!populated_zone(zone
))
2551 if (compaction_deferred(zone
, cc
.order
))
2554 if (compaction_suitable(zone
, cc
.order
, 0, zoneid
) !=
2558 if (kthread_should_stop())
2562 status
= compact_zone(&cc
, NULL
);
2564 if (status
== COMPACT_SUCCESS
) {
2565 compaction_defer_reset(zone
, cc
.order
, false);
2566 } else if (status
== COMPACT_PARTIAL_SKIPPED
|| status
== COMPACT_COMPLETE
) {
2568 * Buddy pages may become stranded on pcps that could
2569 * otherwise coalesce on the zone's free area for
2570 * order >= cc.order. This is ratelimited by the
2571 * upcoming deferral.
2573 drain_all_pages(zone
);
2576 * We use sync migration mode here, so we defer like
2577 * sync direct compaction does.
2579 defer_compaction(zone
, cc
.order
);
2582 count_compact_events(KCOMPACTD_MIGRATE_SCANNED
,
2583 cc
.total_migrate_scanned
);
2584 count_compact_events(KCOMPACTD_FREE_SCANNED
,
2585 cc
.total_free_scanned
);
2587 VM_BUG_ON(!list_empty(&cc
.freepages
));
2588 VM_BUG_ON(!list_empty(&cc
.migratepages
));
2592 * Regardless of success, we are done until woken up next. But remember
2593 * the requested order/classzone_idx in case it was higher/tighter than
2596 if (pgdat
->kcompactd_max_order
<= cc
.order
)
2597 pgdat
->kcompactd_max_order
= 0;
2598 if (pgdat
->kcompactd_classzone_idx
>= cc
.classzone_idx
)
2599 pgdat
->kcompactd_classzone_idx
= pgdat
->nr_zones
- 1;
2602 void wakeup_kcompactd(pg_data_t
*pgdat
, int order
, int classzone_idx
)
2607 if (pgdat
->kcompactd_max_order
< order
)
2608 pgdat
->kcompactd_max_order
= order
;
2610 if (pgdat
->kcompactd_classzone_idx
> classzone_idx
)
2611 pgdat
->kcompactd_classzone_idx
= classzone_idx
;
2614 * Pairs with implicit barrier in wait_event_freezable()
2615 * such that wakeups are not missed.
2617 if (!wq_has_sleeper(&pgdat
->kcompactd_wait
))
2620 if (!kcompactd_node_suitable(pgdat
))
2623 trace_mm_compaction_wakeup_kcompactd(pgdat
->node_id
, order
,
2625 wake_up_interruptible(&pgdat
->kcompactd_wait
);
2629 * The background compaction daemon, started as a kernel thread
2630 * from the init process.
2632 static int kcompactd(void *p
)
2634 pg_data_t
*pgdat
= (pg_data_t
*)p
;
2635 struct task_struct
*tsk
= current
;
2637 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
2639 if (!cpumask_empty(cpumask
))
2640 set_cpus_allowed_ptr(tsk
, cpumask
);
2644 pgdat
->kcompactd_max_order
= 0;
2645 pgdat
->kcompactd_classzone_idx
= pgdat
->nr_zones
- 1;
2647 while (!kthread_should_stop()) {
2648 unsigned long pflags
;
2650 trace_mm_compaction_kcompactd_sleep(pgdat
->node_id
);
2651 wait_event_freezable(pgdat
->kcompactd_wait
,
2652 kcompactd_work_requested(pgdat
));
2654 psi_memstall_enter(&pflags
);
2655 kcompactd_do_work(pgdat
);
2656 psi_memstall_leave(&pflags
);
2663 * This kcompactd start function will be called by init and node-hot-add.
2664 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
2666 int kcompactd_run(int nid
)
2668 pg_data_t
*pgdat
= NODE_DATA(nid
);
2671 if (pgdat
->kcompactd
)
2674 pgdat
->kcompactd
= kthread_run(kcompactd
, pgdat
, "kcompactd%d", nid
);
2675 if (IS_ERR(pgdat
->kcompactd
)) {
2676 pr_err("Failed to start kcompactd on node %d\n", nid
);
2677 ret
= PTR_ERR(pgdat
->kcompactd
);
2678 pgdat
->kcompactd
= NULL
;
2684 * Called by memory hotplug when all memory in a node is offlined. Caller must
2685 * hold mem_hotplug_begin/end().
2687 void kcompactd_stop(int nid
)
2689 struct task_struct
*kcompactd
= NODE_DATA(nid
)->kcompactd
;
2692 kthread_stop(kcompactd
);
2693 NODE_DATA(nid
)->kcompactd
= NULL
;
2698 * It's optimal to keep kcompactd on the same CPUs as their memory, but
2699 * not required for correctness. So if the last cpu in a node goes
2700 * away, we get changed to run anywhere: as the first one comes back,
2701 * restore their cpu bindings.
2703 static int kcompactd_cpu_online(unsigned int cpu
)
2707 for_each_node_state(nid
, N_MEMORY
) {
2708 pg_data_t
*pgdat
= NODE_DATA(nid
);
2709 const struct cpumask
*mask
;
2711 mask
= cpumask_of_node(pgdat
->node_id
);
2713 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
2714 /* One of our CPUs online: restore mask */
2715 set_cpus_allowed_ptr(pgdat
->kcompactd
, mask
);
2720 static int __init
kcompactd_init(void)
2725 ret
= cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN
,
2726 "mm/compaction:online",
2727 kcompactd_cpu_online
, NULL
);
2729 pr_err("kcompactd: failed to register hotplug callbacks.\n");
2733 for_each_node_state(nid
, N_MEMORY
)
2737 subsys_initcall(kcompactd_init
)
2739 #endif /* CONFIG_COMPACTION */