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)
54 * Fragmentation score check interval for proactive compaction purposes.
56 static const unsigned int HPAGE_FRAG_CHECK_INTERVAL_MSEC
= 500;
59 * Page order with-respect-to which proactive compaction
60 * calculates external fragmentation, which is used as
61 * the "fragmentation score" of a node/zone.
63 #if defined CONFIG_TRANSPARENT_HUGEPAGE
64 #define COMPACTION_HPAGE_ORDER HPAGE_PMD_ORDER
65 #elif defined CONFIG_HUGETLBFS
66 #define COMPACTION_HPAGE_ORDER HUGETLB_PAGE_ORDER
68 #define COMPACTION_HPAGE_ORDER (PMD_SHIFT - PAGE_SHIFT)
71 static unsigned long release_freepages(struct list_head
*freelist
)
73 struct page
*page
, *next
;
74 unsigned long high_pfn
= 0;
76 list_for_each_entry_safe(page
, next
, freelist
, lru
) {
77 unsigned long pfn
= page_to_pfn(page
);
87 static void split_map_pages(struct list_head
*list
)
89 unsigned int i
, order
, nr_pages
;
90 struct page
*page
, *next
;
93 list_for_each_entry_safe(page
, next
, list
, lru
) {
96 order
= page_private(page
);
97 nr_pages
= 1 << order
;
99 post_alloc_hook(page
, order
, __GFP_MOVABLE
);
101 split_page(page
, order
);
103 for (i
= 0; i
< nr_pages
; i
++) {
104 list_add(&page
->lru
, &tmp_list
);
109 list_splice(&tmp_list
, list
);
112 #ifdef CONFIG_COMPACTION
114 int PageMovable(struct page
*page
)
116 struct address_space
*mapping
;
118 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
119 if (!__PageMovable(page
))
122 mapping
= page_mapping(page
);
123 if (mapping
&& mapping
->a_ops
&& mapping
->a_ops
->isolate_page
)
128 EXPORT_SYMBOL(PageMovable
);
130 void __SetPageMovable(struct page
*page
, struct address_space
*mapping
)
132 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
133 VM_BUG_ON_PAGE((unsigned long)mapping
& PAGE_MAPPING_MOVABLE
, page
);
134 page
->mapping
= (void *)((unsigned long)mapping
| PAGE_MAPPING_MOVABLE
);
136 EXPORT_SYMBOL(__SetPageMovable
);
138 void __ClearPageMovable(struct page
*page
)
140 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
141 VM_BUG_ON_PAGE(!PageMovable(page
), page
);
143 * Clear registered address_space val with keeping PAGE_MAPPING_MOVABLE
144 * flag so that VM can catch up released page by driver after isolation.
145 * With it, VM migration doesn't try to put it back.
147 page
->mapping
= (void *)((unsigned long)page
->mapping
&
148 PAGE_MAPPING_MOVABLE
);
150 EXPORT_SYMBOL(__ClearPageMovable
);
152 /* Do not skip compaction more than 64 times */
153 #define COMPACT_MAX_DEFER_SHIFT 6
156 * Compaction is deferred when compaction fails to result in a page
157 * allocation success. 1 << compact_defer_shift, compactions are skipped up
158 * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
160 static void defer_compaction(struct zone
*zone
, int order
)
162 zone
->compact_considered
= 0;
163 zone
->compact_defer_shift
++;
165 if (order
< zone
->compact_order_failed
)
166 zone
->compact_order_failed
= order
;
168 if (zone
->compact_defer_shift
> COMPACT_MAX_DEFER_SHIFT
)
169 zone
->compact_defer_shift
= COMPACT_MAX_DEFER_SHIFT
;
171 trace_mm_compaction_defer_compaction(zone
, order
);
174 /* Returns true if compaction should be skipped this time */
175 static bool compaction_deferred(struct zone
*zone
, int order
)
177 unsigned long defer_limit
= 1UL << zone
->compact_defer_shift
;
179 if (order
< zone
->compact_order_failed
)
182 /* Avoid possible overflow */
183 if (++zone
->compact_considered
>= defer_limit
) {
184 zone
->compact_considered
= defer_limit
;
188 trace_mm_compaction_deferred(zone
, order
);
194 * Update defer tracking counters after successful compaction of given order,
195 * which means an allocation either succeeded (alloc_success == true) or is
196 * expected to succeed.
198 void compaction_defer_reset(struct zone
*zone
, int order
,
202 zone
->compact_considered
= 0;
203 zone
->compact_defer_shift
= 0;
205 if (order
>= zone
->compact_order_failed
)
206 zone
->compact_order_failed
= order
+ 1;
208 trace_mm_compaction_defer_reset(zone
, order
);
211 /* Returns true if restarting compaction after many failures */
212 static bool compaction_restarting(struct zone
*zone
, int order
)
214 if (order
< zone
->compact_order_failed
)
217 return zone
->compact_defer_shift
== COMPACT_MAX_DEFER_SHIFT
&&
218 zone
->compact_considered
>= 1UL << zone
->compact_defer_shift
;
221 /* Returns true if the pageblock should be scanned for pages to isolate. */
222 static inline bool isolation_suitable(struct compact_control
*cc
,
225 if (cc
->ignore_skip_hint
)
228 return !get_pageblock_skip(page
);
231 static void reset_cached_positions(struct zone
*zone
)
233 zone
->compact_cached_migrate_pfn
[0] = zone
->zone_start_pfn
;
234 zone
->compact_cached_migrate_pfn
[1] = zone
->zone_start_pfn
;
235 zone
->compact_cached_free_pfn
=
236 pageblock_start_pfn(zone_end_pfn(zone
) - 1);
240 * Compound pages of >= pageblock_order should consistently be skipped until
241 * released. It is always pointless to compact pages of such order (if they are
242 * migratable), and the pageblocks they occupy cannot contain any free pages.
244 static bool pageblock_skip_persistent(struct page
*page
)
246 if (!PageCompound(page
))
249 page
= compound_head(page
);
251 if (compound_order(page
) >= pageblock_order
)
258 __reset_isolation_pfn(struct zone
*zone
, unsigned long pfn
, bool check_source
,
261 struct page
*page
= pfn_to_online_page(pfn
);
262 struct page
*block_page
;
263 struct page
*end_page
;
264 unsigned long block_pfn
;
268 if (zone
!= page_zone(page
))
270 if (pageblock_skip_persistent(page
))
274 * If skip is already cleared do no further checking once the
275 * restart points have been set.
277 if (check_source
&& check_target
&& !get_pageblock_skip(page
))
281 * If clearing skip for the target scanner, do not select a
282 * non-movable pageblock as the starting point.
284 if (!check_source
&& check_target
&&
285 get_pageblock_migratetype(page
) != MIGRATE_MOVABLE
)
288 /* Ensure the start of the pageblock or zone is online and valid */
289 block_pfn
= pageblock_start_pfn(pfn
);
290 block_pfn
= max(block_pfn
, zone
->zone_start_pfn
);
291 block_page
= pfn_to_online_page(block_pfn
);
297 /* Ensure the end of the pageblock or zone is online and valid */
298 block_pfn
= pageblock_end_pfn(pfn
) - 1;
299 block_pfn
= min(block_pfn
, zone_end_pfn(zone
) - 1);
300 end_page
= pfn_to_online_page(block_pfn
);
305 * Only clear the hint if a sample indicates there is either a
306 * free page or an LRU page in the block. One or other condition
307 * is necessary for the block to be a migration source/target.
310 if (pfn_valid_within(pfn
)) {
311 if (check_source
&& PageLRU(page
)) {
312 clear_pageblock_skip(page
);
316 if (check_target
&& PageBuddy(page
)) {
317 clear_pageblock_skip(page
);
322 page
+= (1 << PAGE_ALLOC_COSTLY_ORDER
);
323 pfn
+= (1 << PAGE_ALLOC_COSTLY_ORDER
);
324 } while (page
<= end_page
);
330 * This function is called to clear all cached information on pageblocks that
331 * should be skipped for page isolation when the migrate and free page scanner
334 static void __reset_isolation_suitable(struct zone
*zone
)
336 unsigned long migrate_pfn
= zone
->zone_start_pfn
;
337 unsigned long free_pfn
= zone_end_pfn(zone
) - 1;
338 unsigned long reset_migrate
= free_pfn
;
339 unsigned long reset_free
= migrate_pfn
;
340 bool source_set
= false;
341 bool free_set
= false;
343 if (!zone
->compact_blockskip_flush
)
346 zone
->compact_blockskip_flush
= false;
349 * Walk the zone and update pageblock skip information. Source looks
350 * for PageLRU while target looks for PageBuddy. When the scanner
351 * is found, both PageBuddy and PageLRU are checked as the pageblock
352 * is suitable as both source and target.
354 for (; migrate_pfn
< free_pfn
; migrate_pfn
+= pageblock_nr_pages
,
355 free_pfn
-= pageblock_nr_pages
) {
358 /* Update the migrate PFN */
359 if (__reset_isolation_pfn(zone
, migrate_pfn
, true, source_set
) &&
360 migrate_pfn
< reset_migrate
) {
362 reset_migrate
= migrate_pfn
;
363 zone
->compact_init_migrate_pfn
= reset_migrate
;
364 zone
->compact_cached_migrate_pfn
[0] = reset_migrate
;
365 zone
->compact_cached_migrate_pfn
[1] = reset_migrate
;
368 /* Update the free PFN */
369 if (__reset_isolation_pfn(zone
, free_pfn
, free_set
, true) &&
370 free_pfn
> reset_free
) {
372 reset_free
= free_pfn
;
373 zone
->compact_init_free_pfn
= reset_free
;
374 zone
->compact_cached_free_pfn
= reset_free
;
378 /* Leave no distance if no suitable block was reset */
379 if (reset_migrate
>= reset_free
) {
380 zone
->compact_cached_migrate_pfn
[0] = migrate_pfn
;
381 zone
->compact_cached_migrate_pfn
[1] = migrate_pfn
;
382 zone
->compact_cached_free_pfn
= free_pfn
;
386 void reset_isolation_suitable(pg_data_t
*pgdat
)
390 for (zoneid
= 0; zoneid
< MAX_NR_ZONES
; zoneid
++) {
391 struct zone
*zone
= &pgdat
->node_zones
[zoneid
];
392 if (!populated_zone(zone
))
395 /* Only flush if a full compaction finished recently */
396 if (zone
->compact_blockskip_flush
)
397 __reset_isolation_suitable(zone
);
402 * Sets the pageblock skip bit if it was clear. Note that this is a hint as
403 * locks are not required for read/writers. Returns true if it was already set.
405 static bool test_and_set_skip(struct compact_control
*cc
, struct page
*page
,
410 /* Do no update if skip hint is being ignored */
411 if (cc
->ignore_skip_hint
)
414 if (!IS_ALIGNED(pfn
, pageblock_nr_pages
))
417 skip
= get_pageblock_skip(page
);
418 if (!skip
&& !cc
->no_set_skip_hint
)
419 set_pageblock_skip(page
);
424 static void update_cached_migrate(struct compact_control
*cc
, unsigned long pfn
)
426 struct zone
*zone
= cc
->zone
;
428 pfn
= pageblock_end_pfn(pfn
);
430 /* Set for isolation rather than compaction */
431 if (cc
->no_set_skip_hint
)
434 if (pfn
> zone
->compact_cached_migrate_pfn
[0])
435 zone
->compact_cached_migrate_pfn
[0] = pfn
;
436 if (cc
->mode
!= MIGRATE_ASYNC
&&
437 pfn
> zone
->compact_cached_migrate_pfn
[1])
438 zone
->compact_cached_migrate_pfn
[1] = pfn
;
442 * If no pages were isolated then mark this pageblock to be skipped in the
443 * future. The information is later cleared by __reset_isolation_suitable().
445 static void update_pageblock_skip(struct compact_control
*cc
,
446 struct page
*page
, unsigned long pfn
)
448 struct zone
*zone
= cc
->zone
;
450 if (cc
->no_set_skip_hint
)
456 set_pageblock_skip(page
);
458 /* Update where async and sync compaction should restart */
459 if (pfn
< zone
->compact_cached_free_pfn
)
460 zone
->compact_cached_free_pfn
= pfn
;
463 static inline bool isolation_suitable(struct compact_control
*cc
,
469 static inline bool pageblock_skip_persistent(struct page
*page
)
474 static inline void update_pageblock_skip(struct compact_control
*cc
,
475 struct page
*page
, unsigned long pfn
)
479 static void update_cached_migrate(struct compact_control
*cc
, unsigned long pfn
)
483 static bool test_and_set_skip(struct compact_control
*cc
, struct page
*page
,
488 #endif /* CONFIG_COMPACTION */
491 * Compaction requires the taking of some coarse locks that are potentially
492 * very heavily contended. For async compaction, trylock and record if the
493 * lock is contended. The lock will still be acquired but compaction will
494 * abort when the current block is finished regardless of success rate.
495 * Sync compaction acquires the lock.
497 * Always returns true which makes it easier to track lock state in callers.
499 static bool compact_lock_irqsave(spinlock_t
*lock
, unsigned long *flags
,
500 struct compact_control
*cc
)
503 /* Track if the lock is contended in async mode */
504 if (cc
->mode
== MIGRATE_ASYNC
&& !cc
->contended
) {
505 if (spin_trylock_irqsave(lock
, *flags
))
508 cc
->contended
= true;
511 spin_lock_irqsave(lock
, *flags
);
516 * Compaction requires the taking of some coarse locks that are potentially
517 * very heavily contended. The lock should be periodically unlocked to avoid
518 * having disabled IRQs for a long time, even when there is nobody waiting on
519 * the lock. It might also be that allowing the IRQs will result in
520 * need_resched() becoming true. If scheduling is needed, async compaction
521 * aborts. Sync compaction schedules.
522 * Either compaction type will also abort if a fatal signal is pending.
523 * In either case if the lock was locked, it is dropped and not regained.
525 * Returns true if compaction should abort due to fatal signal pending, or
526 * async compaction due to need_resched()
527 * Returns false when compaction can continue (sync compaction might have
530 static bool compact_unlock_should_abort(spinlock_t
*lock
,
531 unsigned long flags
, bool *locked
, struct compact_control
*cc
)
534 spin_unlock_irqrestore(lock
, flags
);
538 if (fatal_signal_pending(current
)) {
539 cc
->contended
= true;
549 * Isolate free pages onto a private freelist. If @strict is true, will abort
550 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
551 * (even though it may still end up isolating some pages).
553 static unsigned long isolate_freepages_block(struct compact_control
*cc
,
554 unsigned long *start_pfn
,
555 unsigned long end_pfn
,
556 struct list_head
*freelist
,
560 int nr_scanned
= 0, total_isolated
= 0;
562 unsigned long flags
= 0;
564 unsigned long blockpfn
= *start_pfn
;
567 /* Strict mode is for isolation, speed is secondary */
571 cursor
= pfn_to_page(blockpfn
);
573 /* Isolate free pages. */
574 for (; blockpfn
< end_pfn
; blockpfn
+= stride
, cursor
+= stride
) {
576 struct page
*page
= cursor
;
579 * Periodically drop the lock (if held) regardless of its
580 * contention, to give chance to IRQs. Abort if fatal signal
581 * pending or async compaction detects need_resched()
583 if (!(blockpfn
% SWAP_CLUSTER_MAX
)
584 && compact_unlock_should_abort(&cc
->zone
->lock
, flags
,
589 if (!pfn_valid_within(blockpfn
))
593 * For compound pages such as THP and hugetlbfs, we can save
594 * potentially a lot of iterations if we skip them at once.
595 * The check is racy, but we can consider only valid values
596 * and the only danger is skipping too much.
598 if (PageCompound(page
)) {
599 const unsigned int order
= compound_order(page
);
601 if (likely(order
< MAX_ORDER
)) {
602 blockpfn
+= (1UL << order
) - 1;
603 cursor
+= (1UL << order
) - 1;
608 if (!PageBuddy(page
))
612 * If we already hold the lock, we can skip some rechecking.
613 * Note that if we hold the lock now, checked_pageblock was
614 * already set in some previous iteration (or strict is true),
615 * so it is correct to skip the suitable migration target
619 locked
= compact_lock_irqsave(&cc
->zone
->lock
,
622 /* Recheck this is a buddy page under lock */
623 if (!PageBuddy(page
))
627 /* Found a free page, will break it into order-0 pages */
628 order
= buddy_order(page
);
629 isolated
= __isolate_free_page(page
, order
);
632 set_page_private(page
, order
);
634 total_isolated
+= isolated
;
635 cc
->nr_freepages
+= isolated
;
636 list_add_tail(&page
->lru
, freelist
);
638 if (!strict
&& cc
->nr_migratepages
<= cc
->nr_freepages
) {
639 blockpfn
+= isolated
;
642 /* Advance to the end of split page */
643 blockpfn
+= isolated
- 1;
644 cursor
+= isolated
- 1;
656 spin_unlock_irqrestore(&cc
->zone
->lock
, flags
);
659 * There is a tiny chance that we have read bogus compound_order(),
660 * so be careful to not go outside of the pageblock.
662 if (unlikely(blockpfn
> end_pfn
))
665 trace_mm_compaction_isolate_freepages(*start_pfn
, blockpfn
,
666 nr_scanned
, total_isolated
);
668 /* Record how far we have got within the block */
669 *start_pfn
= blockpfn
;
672 * If strict isolation is requested by CMA then check that all the
673 * pages requested were isolated. If there were any failures, 0 is
674 * returned and CMA will fail.
676 if (strict
&& blockpfn
< end_pfn
)
679 cc
->total_free_scanned
+= nr_scanned
;
681 count_compact_events(COMPACTISOLATED
, total_isolated
);
682 return total_isolated
;
686 * isolate_freepages_range() - isolate free pages.
687 * @cc: Compaction control structure.
688 * @start_pfn: The first PFN to start isolating.
689 * @end_pfn: The one-past-last PFN.
691 * Non-free pages, invalid PFNs, or zone boundaries within the
692 * [start_pfn, end_pfn) range are considered errors, cause function to
693 * undo its actions and return zero.
695 * Otherwise, function returns one-past-the-last PFN of isolated page
696 * (which may be greater then end_pfn if end fell in a middle of
700 isolate_freepages_range(struct compact_control
*cc
,
701 unsigned long start_pfn
, unsigned long end_pfn
)
703 unsigned long isolated
, pfn
, block_start_pfn
, block_end_pfn
;
707 block_start_pfn
= pageblock_start_pfn(pfn
);
708 if (block_start_pfn
< cc
->zone
->zone_start_pfn
)
709 block_start_pfn
= cc
->zone
->zone_start_pfn
;
710 block_end_pfn
= pageblock_end_pfn(pfn
);
712 for (; pfn
< end_pfn
; pfn
+= isolated
,
713 block_start_pfn
= block_end_pfn
,
714 block_end_pfn
+= pageblock_nr_pages
) {
715 /* Protect pfn from changing by isolate_freepages_block */
716 unsigned long isolate_start_pfn
= pfn
;
718 block_end_pfn
= min(block_end_pfn
, end_pfn
);
721 * pfn could pass the block_end_pfn if isolated freepage
722 * is more than pageblock order. In this case, we adjust
723 * scanning range to right one.
725 if (pfn
>= block_end_pfn
) {
726 block_start_pfn
= pageblock_start_pfn(pfn
);
727 block_end_pfn
= pageblock_end_pfn(pfn
);
728 block_end_pfn
= min(block_end_pfn
, end_pfn
);
731 if (!pageblock_pfn_to_page(block_start_pfn
,
732 block_end_pfn
, cc
->zone
))
735 isolated
= isolate_freepages_block(cc
, &isolate_start_pfn
,
736 block_end_pfn
, &freelist
, 0, true);
739 * In strict mode, isolate_freepages_block() returns 0 if
740 * there are any holes in the block (ie. invalid PFNs or
747 * If we managed to isolate pages, it is always (1 << n) *
748 * pageblock_nr_pages for some non-negative n. (Max order
749 * page may span two pageblocks).
753 /* __isolate_free_page() does not map the pages */
754 split_map_pages(&freelist
);
757 /* Loop terminated early, cleanup. */
758 release_freepages(&freelist
);
762 /* We don't use freelists for anything. */
766 /* Similar to reclaim, but different enough that they don't share logic */
767 static bool too_many_isolated(pg_data_t
*pgdat
)
769 unsigned long active
, inactive
, isolated
;
771 inactive
= node_page_state(pgdat
, NR_INACTIVE_FILE
) +
772 node_page_state(pgdat
, NR_INACTIVE_ANON
);
773 active
= node_page_state(pgdat
, NR_ACTIVE_FILE
) +
774 node_page_state(pgdat
, NR_ACTIVE_ANON
);
775 isolated
= node_page_state(pgdat
, NR_ISOLATED_FILE
) +
776 node_page_state(pgdat
, NR_ISOLATED_ANON
);
778 return isolated
> (inactive
+ active
) / 2;
782 * isolate_migratepages_block() - isolate all migrate-able pages within
784 * @cc: Compaction control structure.
785 * @low_pfn: The first PFN to isolate
786 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
787 * @isolate_mode: Isolation mode to be used.
789 * Isolate all pages that can be migrated from the range specified by
790 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
791 * Returns zero if there is a fatal signal pending, otherwise PFN of the
792 * first page that was not scanned (which may be both less, equal to or more
795 * The pages are isolated on cc->migratepages list (not required to be empty),
796 * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
797 * is neither read nor updated.
800 isolate_migratepages_block(struct compact_control
*cc
, unsigned long low_pfn
,
801 unsigned long end_pfn
, isolate_mode_t isolate_mode
)
803 pg_data_t
*pgdat
= cc
->zone
->zone_pgdat
;
804 unsigned long nr_scanned
= 0, nr_isolated
= 0;
805 struct lruvec
*lruvec
;
806 unsigned long flags
= 0;
807 struct lruvec
*locked
= NULL
;
808 struct page
*page
= NULL
, *valid_page
= NULL
;
809 unsigned long start_pfn
= low_pfn
;
810 bool skip_on_failure
= false;
811 unsigned long next_skip_pfn
= 0;
812 bool skip_updated
= false;
815 * Ensure that there are not too many pages isolated from the LRU
816 * list by either parallel reclaimers or compaction. If there are,
817 * delay for some time until fewer pages are isolated
819 while (unlikely(too_many_isolated(pgdat
))) {
820 /* stop isolation if there are still pages not migrated */
821 if (cc
->nr_migratepages
)
824 /* async migration should just abort */
825 if (cc
->mode
== MIGRATE_ASYNC
)
828 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
830 if (fatal_signal_pending(current
))
836 if (cc
->direct_compaction
&& (cc
->mode
== MIGRATE_ASYNC
)) {
837 skip_on_failure
= true;
838 next_skip_pfn
= block_end_pfn(low_pfn
, cc
->order
);
841 /* Time to isolate some pages for migration */
842 for (; low_pfn
< end_pfn
; low_pfn
++) {
844 if (skip_on_failure
&& low_pfn
>= next_skip_pfn
) {
846 * We have isolated all migration candidates in the
847 * previous order-aligned block, and did not skip it due
848 * to failure. We should migrate the pages now and
849 * hopefully succeed compaction.
855 * We failed to isolate in the previous order-aligned
856 * block. Set the new boundary to the end of the
857 * current block. Note we can't simply increase
858 * next_skip_pfn by 1 << order, as low_pfn might have
859 * been incremented by a higher number due to skipping
860 * a compound or a high-order buddy page in the
861 * previous loop iteration.
863 next_skip_pfn
= block_end_pfn(low_pfn
, cc
->order
);
867 * Periodically drop the lock (if held) regardless of its
868 * contention, to give chance to IRQs. Abort completely if
869 * a fatal signal is pending.
871 if (!(low_pfn
% SWAP_CLUSTER_MAX
)) {
873 unlock_page_lruvec_irqrestore(locked
, flags
);
877 if (fatal_signal_pending(current
)) {
878 cc
->contended
= true;
887 if (!pfn_valid_within(low_pfn
))
891 page
= pfn_to_page(low_pfn
);
894 * Check if the pageblock has already been marked skipped.
895 * Only the aligned PFN is checked as the caller isolates
896 * COMPACT_CLUSTER_MAX at a time so the second call must
897 * not falsely conclude that the block should be skipped.
899 if (!valid_page
&& IS_ALIGNED(low_pfn
, pageblock_nr_pages
)) {
900 if (!cc
->ignore_skip_hint
&& get_pageblock_skip(page
)) {
909 * Skip if free. We read page order here without zone lock
910 * which is generally unsafe, but the race window is small and
911 * the worst thing that can happen is that we skip some
912 * potential isolation targets.
914 if (PageBuddy(page
)) {
915 unsigned long freepage_order
= buddy_order_unsafe(page
);
918 * Without lock, we cannot be sure that what we got is
919 * a valid page order. Consider only values in the
920 * valid order range to prevent low_pfn overflow.
922 if (freepage_order
> 0 && freepage_order
< MAX_ORDER
)
923 low_pfn
+= (1UL << freepage_order
) - 1;
928 * Regardless of being on LRU, compound pages such as THP and
929 * hugetlbfs are not to be compacted unless we are attempting
930 * an allocation much larger than the huge page size (eg CMA).
931 * We can potentially save a lot of iterations if we skip them
932 * at once. The check is racy, but we can consider only valid
933 * values and the only danger is skipping too much.
935 if (PageCompound(page
) && !cc
->alloc_contig
) {
936 const unsigned int order
= compound_order(page
);
938 if (likely(order
< MAX_ORDER
))
939 low_pfn
+= (1UL << order
) - 1;
944 * Check may be lockless but that's ok as we recheck later.
945 * It's possible to migrate LRU and non-lru movable pages.
946 * Skip any other type of page
948 if (!PageLRU(page
)) {
950 * __PageMovable can return false positive so we need
951 * to verify it under page_lock.
953 if (unlikely(__PageMovable(page
)) &&
954 !PageIsolated(page
)) {
956 unlock_page_lruvec_irqrestore(locked
, flags
);
960 if (!isolate_movable_page(page
, isolate_mode
))
961 goto isolate_success
;
968 * Migration will fail if an anonymous page is pinned in memory,
969 * so avoid taking lru_lock and isolating it unnecessarily in an
970 * admittedly racy check.
972 if (!page_mapping(page
) &&
973 page_count(page
) > page_mapcount(page
))
977 * Only allow to migrate anonymous pages in GFP_NOFS context
978 * because those do not depend on fs locks.
980 if (!(cc
->gfp_mask
& __GFP_FS
) && page_mapping(page
))
984 * Be careful not to clear PageLRU until after we're
985 * sure the page is not being freed elsewhere -- the
986 * page release code relies on it.
988 if (unlikely(!get_page_unless_zero(page
)))
991 if (__isolate_lru_page_prepare(page
, isolate_mode
) != 0)
992 goto isolate_fail_put
;
994 /* Try isolate the page */
995 if (!TestClearPageLRU(page
))
996 goto isolate_fail_put
;
999 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1001 /* If we already hold the lock, we can skip some rechecking */
1002 if (lruvec
!= locked
) {
1004 unlock_page_lruvec_irqrestore(locked
, flags
);
1006 compact_lock_irqsave(&lruvec
->lru_lock
, &flags
, cc
);
1010 lruvec_memcg_debug(lruvec
, page
);
1012 /* Try get exclusive access under lock */
1013 if (!skip_updated
) {
1014 skip_updated
= true;
1015 if (test_and_set_skip(cc
, page
, low_pfn
))
1020 * Page become compound since the non-locked check,
1021 * and it's on LRU. It can only be a THP so the order
1022 * is safe to read and it's 0 for tail pages.
1024 if (unlikely(PageCompound(page
) && !cc
->alloc_contig
)) {
1025 low_pfn
+= compound_nr(page
) - 1;
1027 goto isolate_fail_put
;
1032 /* The whole page is taken off the LRU; skip the tail pages. */
1033 if (PageCompound(page
))
1034 low_pfn
+= compound_nr(page
) - 1;
1036 /* Successfully isolated */
1037 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
1038 mod_node_page_state(page_pgdat(page
),
1039 NR_ISOLATED_ANON
+ page_is_file_lru(page
),
1040 thp_nr_pages(page
));
1043 list_add(&page
->lru
, &cc
->migratepages
);
1044 cc
->nr_migratepages
+= compound_nr(page
);
1045 nr_isolated
+= compound_nr(page
);
1048 * Avoid isolating too much unless this block is being
1049 * rescanned (e.g. dirty/writeback pages, parallel allocation)
1050 * or a lock is contended. For contention, isolate quickly to
1051 * potentially remove one source of contention.
1053 if (cc
->nr_migratepages
>= COMPACT_CLUSTER_MAX
&&
1054 !cc
->rescan
&& !cc
->contended
) {
1062 /* Avoid potential deadlock in freeing page under lru_lock */
1064 unlock_page_lruvec_irqrestore(locked
, flags
);
1070 if (!skip_on_failure
)
1074 * We have isolated some pages, but then failed. Release them
1075 * instead of migrating, as we cannot form the cc->order buddy
1080 unlock_page_lruvec_irqrestore(locked
, flags
);
1083 putback_movable_pages(&cc
->migratepages
);
1084 cc
->nr_migratepages
= 0;
1088 if (low_pfn
< next_skip_pfn
) {
1089 low_pfn
= next_skip_pfn
- 1;
1091 * The check near the loop beginning would have updated
1092 * next_skip_pfn too, but this is a bit simpler.
1094 next_skip_pfn
+= 1UL << cc
->order
;
1099 * The PageBuddy() check could have potentially brought us outside
1100 * the range to be scanned.
1102 if (unlikely(low_pfn
> end_pfn
))
1109 unlock_page_lruvec_irqrestore(locked
, flags
);
1116 * Updated the cached scanner pfn once the pageblock has been scanned
1117 * Pages will either be migrated in which case there is no point
1118 * scanning in the near future or migration failed in which case the
1119 * failure reason may persist. The block is marked for skipping if
1120 * there were no pages isolated in the block or if the block is
1121 * rescanned twice in a row.
1123 if (low_pfn
== end_pfn
&& (!nr_isolated
|| cc
->rescan
)) {
1124 if (valid_page
&& !skip_updated
)
1125 set_pageblock_skip(valid_page
);
1126 update_cached_migrate(cc
, low_pfn
);
1129 trace_mm_compaction_isolate_migratepages(start_pfn
, low_pfn
,
1130 nr_scanned
, nr_isolated
);
1133 cc
->total_migrate_scanned
+= nr_scanned
;
1135 count_compact_events(COMPACTISOLATED
, nr_isolated
);
1141 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
1142 * @cc: Compaction control structure.
1143 * @start_pfn: The first PFN to start isolating.
1144 * @end_pfn: The one-past-last PFN.
1146 * Returns zero if isolation fails fatally due to e.g. pending signal.
1147 * Otherwise, function returns one-past-the-last PFN of isolated page
1148 * (which may be greater than end_pfn if end fell in a middle of a THP page).
1151 isolate_migratepages_range(struct compact_control
*cc
, unsigned long start_pfn
,
1152 unsigned long end_pfn
)
1154 unsigned long pfn
, block_start_pfn
, block_end_pfn
;
1156 /* Scan block by block. First and last block may be incomplete */
1158 block_start_pfn
= pageblock_start_pfn(pfn
);
1159 if (block_start_pfn
< cc
->zone
->zone_start_pfn
)
1160 block_start_pfn
= cc
->zone
->zone_start_pfn
;
1161 block_end_pfn
= pageblock_end_pfn(pfn
);
1163 for (; pfn
< end_pfn
; pfn
= block_end_pfn
,
1164 block_start_pfn
= block_end_pfn
,
1165 block_end_pfn
+= pageblock_nr_pages
) {
1167 block_end_pfn
= min(block_end_pfn
, end_pfn
);
1169 if (!pageblock_pfn_to_page(block_start_pfn
,
1170 block_end_pfn
, cc
->zone
))
1173 pfn
= isolate_migratepages_block(cc
, pfn
, block_end_pfn
,
1174 ISOLATE_UNEVICTABLE
);
1179 if (cc
->nr_migratepages
>= COMPACT_CLUSTER_MAX
)
1186 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
1187 #ifdef CONFIG_COMPACTION
1189 static bool suitable_migration_source(struct compact_control
*cc
,
1194 if (pageblock_skip_persistent(page
))
1197 if ((cc
->mode
!= MIGRATE_ASYNC
) || !cc
->direct_compaction
)
1200 block_mt
= get_pageblock_migratetype(page
);
1202 if (cc
->migratetype
== MIGRATE_MOVABLE
)
1203 return is_migrate_movable(block_mt
);
1205 return block_mt
== cc
->migratetype
;
1208 /* Returns true if the page is within a block suitable for migration to */
1209 static bool suitable_migration_target(struct compact_control
*cc
,
1212 /* If the page is a large free page, then disallow migration */
1213 if (PageBuddy(page
)) {
1215 * We are checking page_order without zone->lock taken. But
1216 * the only small danger is that we skip a potentially suitable
1217 * pageblock, so it's not worth to check order for valid range.
1219 if (buddy_order_unsafe(page
) >= pageblock_order
)
1223 if (cc
->ignore_block_suitable
)
1226 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1227 if (is_migrate_movable(get_pageblock_migratetype(page
)))
1230 /* Otherwise skip the block */
1234 static inline unsigned int
1235 freelist_scan_limit(struct compact_control
*cc
)
1237 unsigned short shift
= BITS_PER_LONG
- 1;
1239 return (COMPACT_CLUSTER_MAX
>> min(shift
, cc
->fast_search_fail
)) + 1;
1243 * Test whether the free scanner has reached the same or lower pageblock than
1244 * the migration scanner, and compaction should thus terminate.
1246 static inline bool compact_scanners_met(struct compact_control
*cc
)
1248 return (cc
->free_pfn
>> pageblock_order
)
1249 <= (cc
->migrate_pfn
>> pageblock_order
);
1253 * Used when scanning for a suitable migration target which scans freelists
1254 * in reverse. Reorders the list such as the unscanned pages are scanned
1255 * first on the next iteration of the free scanner
1258 move_freelist_head(struct list_head
*freelist
, struct page
*freepage
)
1262 if (!list_is_last(freelist
, &freepage
->lru
)) {
1263 list_cut_before(&sublist
, freelist
, &freepage
->lru
);
1264 if (!list_empty(&sublist
))
1265 list_splice_tail(&sublist
, freelist
);
1270 * Similar to move_freelist_head except used by the migration scanner
1271 * when scanning forward. It's possible for these list operations to
1272 * move against each other if they search the free list exactly in
1276 move_freelist_tail(struct list_head
*freelist
, struct page
*freepage
)
1280 if (!list_is_first(freelist
, &freepage
->lru
)) {
1281 list_cut_position(&sublist
, freelist
, &freepage
->lru
);
1282 if (!list_empty(&sublist
))
1283 list_splice_tail(&sublist
, freelist
);
1288 fast_isolate_around(struct compact_control
*cc
, unsigned long pfn
, unsigned long nr_isolated
)
1290 unsigned long start_pfn
, end_pfn
;
1293 /* Do not search around if there are enough pages already */
1294 if (cc
->nr_freepages
>= cc
->nr_migratepages
)
1297 /* Minimise scanning during async compaction */
1298 if (cc
->direct_compaction
&& cc
->mode
== MIGRATE_ASYNC
)
1301 /* Pageblock boundaries */
1302 start_pfn
= max(pageblock_start_pfn(pfn
), cc
->zone
->zone_start_pfn
);
1303 end_pfn
= min(pageblock_end_pfn(pfn
), zone_end_pfn(cc
->zone
));
1305 page
= pageblock_pfn_to_page(start_pfn
, end_pfn
, cc
->zone
);
1310 if (start_pfn
!= pfn
) {
1311 isolate_freepages_block(cc
, &start_pfn
, pfn
, &cc
->freepages
, 1, false);
1312 if (cc
->nr_freepages
>= cc
->nr_migratepages
)
1317 start_pfn
= pfn
+ nr_isolated
;
1318 if (start_pfn
< end_pfn
)
1319 isolate_freepages_block(cc
, &start_pfn
, end_pfn
, &cc
->freepages
, 1, false);
1321 /* Skip this pageblock in the future as it's full or nearly full */
1322 if (cc
->nr_freepages
< cc
->nr_migratepages
)
1323 set_pageblock_skip(page
);
1326 /* Search orders in round-robin fashion */
1327 static int next_search_order(struct compact_control
*cc
, int order
)
1331 order
= cc
->order
- 1;
1333 /* Search wrapped around? */
1334 if (order
== cc
->search_order
) {
1336 if (cc
->search_order
< 0)
1337 cc
->search_order
= cc
->order
- 1;
1344 static unsigned long
1345 fast_isolate_freepages(struct compact_control
*cc
)
1347 unsigned int limit
= min(1U, freelist_scan_limit(cc
) >> 1);
1348 unsigned int nr_scanned
= 0;
1349 unsigned long low_pfn
, min_pfn
, highest
= 0;
1350 unsigned long nr_isolated
= 0;
1351 unsigned long distance
;
1352 struct page
*page
= NULL
;
1353 bool scan_start
= false;
1356 /* Full compaction passes in a negative order */
1358 return cc
->free_pfn
;
1361 * If starting the scan, use a deeper search and use the highest
1362 * PFN found if a suitable one is not found.
1364 if (cc
->free_pfn
>= cc
->zone
->compact_init_free_pfn
) {
1365 limit
= pageblock_nr_pages
>> 1;
1370 * Preferred point is in the top quarter of the scan space but take
1371 * a pfn from the top half if the search is problematic.
1373 distance
= (cc
->free_pfn
- cc
->migrate_pfn
);
1374 low_pfn
= pageblock_start_pfn(cc
->free_pfn
- (distance
>> 2));
1375 min_pfn
= pageblock_start_pfn(cc
->free_pfn
- (distance
>> 1));
1377 if (WARN_ON_ONCE(min_pfn
> low_pfn
))
1381 * Search starts from the last successful isolation order or the next
1382 * order to search after a previous failure
1384 cc
->search_order
= min_t(unsigned int, cc
->order
- 1, cc
->search_order
);
1386 for (order
= cc
->search_order
;
1387 !page
&& order
>= 0;
1388 order
= next_search_order(cc
, order
)) {
1389 struct free_area
*area
= &cc
->zone
->free_area
[order
];
1390 struct list_head
*freelist
;
1391 struct page
*freepage
;
1392 unsigned long flags
;
1393 unsigned int order_scanned
= 0;
1394 unsigned long high_pfn
= 0;
1399 spin_lock_irqsave(&cc
->zone
->lock
, flags
);
1400 freelist
= &area
->free_list
[MIGRATE_MOVABLE
];
1401 list_for_each_entry_reverse(freepage
, freelist
, lru
) {
1406 pfn
= page_to_pfn(freepage
);
1409 highest
= max(pageblock_start_pfn(pfn
),
1410 cc
->zone
->zone_start_pfn
);
1412 if (pfn
>= low_pfn
) {
1413 cc
->fast_search_fail
= 0;
1414 cc
->search_order
= order
;
1419 if (pfn
>= min_pfn
&& pfn
> high_pfn
) {
1422 /* Shorten the scan if a candidate is found */
1426 if (order_scanned
>= limit
)
1430 /* Use a minimum pfn if a preferred one was not found */
1431 if (!page
&& high_pfn
) {
1432 page
= pfn_to_page(high_pfn
);
1434 /* Update freepage for the list reorder below */
1438 /* Reorder to so a future search skips recent pages */
1439 move_freelist_head(freelist
, freepage
);
1441 /* Isolate the page if available */
1443 if (__isolate_free_page(page
, order
)) {
1444 set_page_private(page
, order
);
1445 nr_isolated
= 1 << order
;
1446 cc
->nr_freepages
+= nr_isolated
;
1447 list_add_tail(&page
->lru
, &cc
->freepages
);
1448 count_compact_events(COMPACTISOLATED
, nr_isolated
);
1450 /* If isolation fails, abort the search */
1451 order
= cc
->search_order
+ 1;
1456 spin_unlock_irqrestore(&cc
->zone
->lock
, flags
);
1459 * Smaller scan on next order so the total scan ig related
1460 * to freelist_scan_limit.
1462 if (order_scanned
>= limit
)
1463 limit
= min(1U, limit
>> 1);
1467 cc
->fast_search_fail
++;
1470 * Use the highest PFN found above min. If one was
1471 * not found, be pessimistic for direct compaction
1472 * and use the min mark.
1475 page
= pfn_to_page(highest
);
1476 cc
->free_pfn
= highest
;
1478 if (cc
->direct_compaction
&& pfn_valid(min_pfn
)) {
1479 page
= pageblock_pfn_to_page(min_pfn
,
1480 min(pageblock_end_pfn(min_pfn
),
1481 zone_end_pfn(cc
->zone
)),
1483 cc
->free_pfn
= min_pfn
;
1489 if (highest
&& highest
>= cc
->zone
->compact_cached_free_pfn
) {
1490 highest
-= pageblock_nr_pages
;
1491 cc
->zone
->compact_cached_free_pfn
= highest
;
1494 cc
->total_free_scanned
+= nr_scanned
;
1496 return cc
->free_pfn
;
1498 low_pfn
= page_to_pfn(page
);
1499 fast_isolate_around(cc
, low_pfn
, nr_isolated
);
1504 * Based on information in the current compact_control, find blocks
1505 * suitable for isolating free pages from and then isolate them.
1507 static void isolate_freepages(struct compact_control
*cc
)
1509 struct zone
*zone
= cc
->zone
;
1511 unsigned long block_start_pfn
; /* start of current pageblock */
1512 unsigned long isolate_start_pfn
; /* exact pfn we start at */
1513 unsigned long block_end_pfn
; /* end of current pageblock */
1514 unsigned long low_pfn
; /* lowest pfn scanner is able to scan */
1515 struct list_head
*freelist
= &cc
->freepages
;
1516 unsigned int stride
;
1518 /* Try a small search of the free lists for a candidate */
1519 isolate_start_pfn
= fast_isolate_freepages(cc
);
1520 if (cc
->nr_freepages
)
1524 * Initialise the free scanner. The starting point is where we last
1525 * successfully isolated from, zone-cached value, or the end of the
1526 * zone when isolating for the first time. For looping we also need
1527 * this pfn aligned down to the pageblock boundary, because we do
1528 * block_start_pfn -= pageblock_nr_pages in the for loop.
1529 * For ending point, take care when isolating in last pageblock of a
1530 * zone which ends in the middle of a pageblock.
1531 * The low boundary is the end of the pageblock the migration scanner
1534 isolate_start_pfn
= cc
->free_pfn
;
1535 block_start_pfn
= pageblock_start_pfn(isolate_start_pfn
);
1536 block_end_pfn
= min(block_start_pfn
+ pageblock_nr_pages
,
1537 zone_end_pfn(zone
));
1538 low_pfn
= pageblock_end_pfn(cc
->migrate_pfn
);
1539 stride
= cc
->mode
== MIGRATE_ASYNC
? COMPACT_CLUSTER_MAX
: 1;
1542 * Isolate free pages until enough are available to migrate the
1543 * pages on cc->migratepages. We stop searching if the migrate
1544 * and free page scanners meet or enough free pages are isolated.
1546 for (; block_start_pfn
>= low_pfn
;
1547 block_end_pfn
= block_start_pfn
,
1548 block_start_pfn
-= pageblock_nr_pages
,
1549 isolate_start_pfn
= block_start_pfn
) {
1550 unsigned long nr_isolated
;
1553 * This can iterate a massively long zone without finding any
1554 * suitable migration targets, so periodically check resched.
1556 if (!(block_start_pfn
% (SWAP_CLUSTER_MAX
* pageblock_nr_pages
)))
1559 page
= pageblock_pfn_to_page(block_start_pfn
, block_end_pfn
,
1564 /* Check the block is suitable for migration */
1565 if (!suitable_migration_target(cc
, page
))
1568 /* If isolation recently failed, do not retry */
1569 if (!isolation_suitable(cc
, page
))
1572 /* Found a block suitable for isolating free pages from. */
1573 nr_isolated
= isolate_freepages_block(cc
, &isolate_start_pfn
,
1574 block_end_pfn
, freelist
, stride
, false);
1576 /* Update the skip hint if the full pageblock was scanned */
1577 if (isolate_start_pfn
== block_end_pfn
)
1578 update_pageblock_skip(cc
, page
, block_start_pfn
);
1580 /* Are enough freepages isolated? */
1581 if (cc
->nr_freepages
>= cc
->nr_migratepages
) {
1582 if (isolate_start_pfn
>= block_end_pfn
) {
1584 * Restart at previous pageblock if more
1585 * freepages can be isolated next time.
1588 block_start_pfn
- pageblock_nr_pages
;
1591 } else if (isolate_start_pfn
< block_end_pfn
) {
1593 * If isolation failed early, do not continue
1599 /* Adjust stride depending on isolation */
1604 stride
= min_t(unsigned int, COMPACT_CLUSTER_MAX
, stride
<< 1);
1608 * Record where the free scanner will restart next time. Either we
1609 * broke from the loop and set isolate_start_pfn based on the last
1610 * call to isolate_freepages_block(), or we met the migration scanner
1611 * and the loop terminated due to isolate_start_pfn < low_pfn
1613 cc
->free_pfn
= isolate_start_pfn
;
1616 /* __isolate_free_page() does not map the pages */
1617 split_map_pages(freelist
);
1621 * This is a migrate-callback that "allocates" freepages by taking pages
1622 * from the isolated freelists in the block we are migrating to.
1624 static struct page
*compaction_alloc(struct page
*migratepage
,
1627 struct compact_control
*cc
= (struct compact_control
*)data
;
1628 struct page
*freepage
;
1630 if (list_empty(&cc
->freepages
)) {
1631 isolate_freepages(cc
);
1633 if (list_empty(&cc
->freepages
))
1637 freepage
= list_entry(cc
->freepages
.next
, struct page
, lru
);
1638 list_del(&freepage
->lru
);
1645 * This is a migrate-callback that "frees" freepages back to the isolated
1646 * freelist. All pages on the freelist are from the same zone, so there is no
1647 * special handling needed for NUMA.
1649 static void compaction_free(struct page
*page
, unsigned long data
)
1651 struct compact_control
*cc
= (struct compact_control
*)data
;
1653 list_add(&page
->lru
, &cc
->freepages
);
1657 /* possible outcome of isolate_migratepages */
1659 ISOLATE_ABORT
, /* Abort compaction now */
1660 ISOLATE_NONE
, /* No pages isolated, continue scanning */
1661 ISOLATE_SUCCESS
, /* Pages isolated, migrate */
1662 } isolate_migrate_t
;
1665 * Allow userspace to control policy on scanning the unevictable LRU for
1666 * compactable pages.
1668 #ifdef CONFIG_PREEMPT_RT
1669 int sysctl_compact_unevictable_allowed __read_mostly
= 0;
1671 int sysctl_compact_unevictable_allowed __read_mostly
= 1;
1675 update_fast_start_pfn(struct compact_control
*cc
, unsigned long pfn
)
1677 if (cc
->fast_start_pfn
== ULONG_MAX
)
1680 if (!cc
->fast_start_pfn
)
1681 cc
->fast_start_pfn
= pfn
;
1683 cc
->fast_start_pfn
= min(cc
->fast_start_pfn
, pfn
);
1686 static inline unsigned long
1687 reinit_migrate_pfn(struct compact_control
*cc
)
1689 if (!cc
->fast_start_pfn
|| cc
->fast_start_pfn
== ULONG_MAX
)
1690 return cc
->migrate_pfn
;
1692 cc
->migrate_pfn
= cc
->fast_start_pfn
;
1693 cc
->fast_start_pfn
= ULONG_MAX
;
1695 return cc
->migrate_pfn
;
1699 * Briefly search the free lists for a migration source that already has
1700 * some free pages to reduce the number of pages that need migration
1701 * before a pageblock is free.
1703 static unsigned long fast_find_migrateblock(struct compact_control
*cc
)
1705 unsigned int limit
= freelist_scan_limit(cc
);
1706 unsigned int nr_scanned
= 0;
1707 unsigned long distance
;
1708 unsigned long pfn
= cc
->migrate_pfn
;
1709 unsigned long high_pfn
;
1711 bool found_block
= false;
1713 /* Skip hints are relied on to avoid repeats on the fast search */
1714 if (cc
->ignore_skip_hint
)
1718 * If the migrate_pfn is not at the start of a zone or the start
1719 * of a pageblock then assume this is a continuation of a previous
1720 * scan restarted due to COMPACT_CLUSTER_MAX.
1722 if (pfn
!= cc
->zone
->zone_start_pfn
&& pfn
!= pageblock_start_pfn(pfn
))
1726 * For smaller orders, just linearly scan as the number of pages
1727 * to migrate should be relatively small and does not necessarily
1728 * justify freeing up a large block for a small allocation.
1730 if (cc
->order
<= PAGE_ALLOC_COSTLY_ORDER
)
1734 * Only allow kcompactd and direct requests for movable pages to
1735 * quickly clear out a MOVABLE pageblock for allocation. This
1736 * reduces the risk that a large movable pageblock is freed for
1737 * an unmovable/reclaimable small allocation.
1739 if (cc
->direct_compaction
&& cc
->migratetype
!= MIGRATE_MOVABLE
)
1743 * When starting the migration scanner, pick any pageblock within the
1744 * first half of the search space. Otherwise try and pick a pageblock
1745 * within the first eighth to reduce the chances that a migration
1746 * target later becomes a source.
1748 distance
= (cc
->free_pfn
- cc
->migrate_pfn
) >> 1;
1749 if (cc
->migrate_pfn
!= cc
->zone
->zone_start_pfn
)
1751 high_pfn
= pageblock_start_pfn(cc
->migrate_pfn
+ distance
);
1753 for (order
= cc
->order
- 1;
1754 order
>= PAGE_ALLOC_COSTLY_ORDER
&& !found_block
&& nr_scanned
< limit
;
1756 struct free_area
*area
= &cc
->zone
->free_area
[order
];
1757 struct list_head
*freelist
;
1758 unsigned long flags
;
1759 struct page
*freepage
;
1764 spin_lock_irqsave(&cc
->zone
->lock
, flags
);
1765 freelist
= &area
->free_list
[MIGRATE_MOVABLE
];
1766 list_for_each_entry(freepage
, freelist
, lru
) {
1767 unsigned long free_pfn
;
1769 if (nr_scanned
++ >= limit
) {
1770 move_freelist_tail(freelist
, freepage
);
1774 free_pfn
= page_to_pfn(freepage
);
1775 if (free_pfn
< high_pfn
) {
1777 * Avoid if skipped recently. Ideally it would
1778 * move to the tail but even safe iteration of
1779 * the list assumes an entry is deleted, not
1782 if (get_pageblock_skip(freepage
))
1785 /* Reorder to so a future search skips recent pages */
1786 move_freelist_tail(freelist
, freepage
);
1788 update_fast_start_pfn(cc
, free_pfn
);
1789 pfn
= pageblock_start_pfn(free_pfn
);
1790 cc
->fast_search_fail
= 0;
1792 set_pageblock_skip(freepage
);
1796 spin_unlock_irqrestore(&cc
->zone
->lock
, flags
);
1799 cc
->total_migrate_scanned
+= nr_scanned
;
1802 * If fast scanning failed then use a cached entry for a page block
1803 * that had free pages as the basis for starting a linear scan.
1806 cc
->fast_search_fail
++;
1807 pfn
= reinit_migrate_pfn(cc
);
1813 * Isolate all pages that can be migrated from the first suitable block,
1814 * starting at the block pointed to by the migrate scanner pfn within
1817 static isolate_migrate_t
isolate_migratepages(struct compact_control
*cc
)
1819 unsigned long block_start_pfn
;
1820 unsigned long block_end_pfn
;
1821 unsigned long low_pfn
;
1823 const isolate_mode_t isolate_mode
=
1824 (sysctl_compact_unevictable_allowed
? ISOLATE_UNEVICTABLE
: 0) |
1825 (cc
->mode
!= MIGRATE_SYNC
? ISOLATE_ASYNC_MIGRATE
: 0);
1826 bool fast_find_block
;
1829 * Start at where we last stopped, or beginning of the zone as
1830 * initialized by compact_zone(). The first failure will use
1831 * the lowest PFN as the starting point for linear scanning.
1833 low_pfn
= fast_find_migrateblock(cc
);
1834 block_start_pfn
= pageblock_start_pfn(low_pfn
);
1835 if (block_start_pfn
< cc
->zone
->zone_start_pfn
)
1836 block_start_pfn
= cc
->zone
->zone_start_pfn
;
1839 * fast_find_migrateblock marks a pageblock skipped so to avoid
1840 * the isolation_suitable check below, check whether the fast
1841 * search was successful.
1843 fast_find_block
= low_pfn
!= cc
->migrate_pfn
&& !cc
->fast_search_fail
;
1845 /* Only scan within a pageblock boundary */
1846 block_end_pfn
= pageblock_end_pfn(low_pfn
);
1849 * Iterate over whole pageblocks until we find the first suitable.
1850 * Do not cross the free scanner.
1852 for (; block_end_pfn
<= cc
->free_pfn
;
1853 fast_find_block
= false,
1854 low_pfn
= block_end_pfn
,
1855 block_start_pfn
= block_end_pfn
,
1856 block_end_pfn
+= pageblock_nr_pages
) {
1859 * This can potentially iterate a massively long zone with
1860 * many pageblocks unsuitable, so periodically check if we
1863 if (!(low_pfn
% (SWAP_CLUSTER_MAX
* pageblock_nr_pages
)))
1866 page
= pageblock_pfn_to_page(block_start_pfn
,
1867 block_end_pfn
, cc
->zone
);
1872 * If isolation recently failed, do not retry. Only check the
1873 * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock
1874 * to be visited multiple times. Assume skip was checked
1875 * before making it "skip" so other compaction instances do
1876 * not scan the same block.
1878 if (IS_ALIGNED(low_pfn
, pageblock_nr_pages
) &&
1879 !fast_find_block
&& !isolation_suitable(cc
, page
))
1883 * For async compaction, also only scan in MOVABLE blocks
1884 * without huge pages. Async compaction is optimistic to see
1885 * if the minimum amount of work satisfies the allocation.
1886 * The cached PFN is updated as it's possible that all
1887 * remaining blocks between source and target are unsuitable
1888 * and the compaction scanners fail to meet.
1890 if (!suitable_migration_source(cc
, page
)) {
1891 update_cached_migrate(cc
, block_end_pfn
);
1895 /* Perform the isolation */
1896 low_pfn
= isolate_migratepages_block(cc
, low_pfn
,
1897 block_end_pfn
, isolate_mode
);
1900 return ISOLATE_ABORT
;
1903 * Either we isolated something and proceed with migration. Or
1904 * we failed and compact_zone should decide if we should
1910 /* Record where migration scanner will be restarted. */
1911 cc
->migrate_pfn
= low_pfn
;
1913 return cc
->nr_migratepages
? ISOLATE_SUCCESS
: ISOLATE_NONE
;
1917 * order == -1 is expected when compacting via
1918 * /proc/sys/vm/compact_memory
1920 static inline bool is_via_compact_memory(int order
)
1925 static bool kswapd_is_running(pg_data_t
*pgdat
)
1927 return pgdat
->kswapd
&& (pgdat
->kswapd
->state
== TASK_RUNNING
);
1931 * A zone's fragmentation score is the external fragmentation wrt to the
1932 * COMPACTION_HPAGE_ORDER scaled by the zone's size. It returns a value
1933 * in the range [0, 100].
1935 * The scaling factor ensures that proactive compaction focuses on larger
1936 * zones like ZONE_NORMAL, rather than smaller, specialized zones like
1937 * ZONE_DMA32. For smaller zones, the score value remains close to zero,
1938 * and thus never exceeds the high threshold for proactive compaction.
1940 static unsigned int fragmentation_score_zone(struct zone
*zone
)
1942 unsigned long score
;
1944 score
= zone
->present_pages
*
1945 extfrag_for_order(zone
, COMPACTION_HPAGE_ORDER
);
1946 return div64_ul(score
, zone
->zone_pgdat
->node_present_pages
+ 1);
1950 * The per-node proactive (background) compaction process is started by its
1951 * corresponding kcompactd thread when the node's fragmentation score
1952 * exceeds the high threshold. The compaction process remains active till
1953 * the node's score falls below the low threshold, or one of the back-off
1954 * conditions is met.
1956 static unsigned int fragmentation_score_node(pg_data_t
*pgdat
)
1958 unsigned int score
= 0;
1961 for (zoneid
= 0; zoneid
< MAX_NR_ZONES
; zoneid
++) {
1964 zone
= &pgdat
->node_zones
[zoneid
];
1965 score
+= fragmentation_score_zone(zone
);
1971 static unsigned int fragmentation_score_wmark(pg_data_t
*pgdat
, bool low
)
1973 unsigned int wmark_low
;
1976 * Cap the low watermak to avoid excessive compaction
1977 * activity in case a user sets the proactivess tunable
1978 * close to 100 (maximum).
1980 wmark_low
= max(100U - sysctl_compaction_proactiveness
, 5U);
1981 return low
? wmark_low
: min(wmark_low
+ 10, 100U);
1984 static bool should_proactive_compact_node(pg_data_t
*pgdat
)
1988 if (!sysctl_compaction_proactiveness
|| kswapd_is_running(pgdat
))
1991 wmark_high
= fragmentation_score_wmark(pgdat
, false);
1992 return fragmentation_score_node(pgdat
) > wmark_high
;
1995 static enum compact_result
__compact_finished(struct compact_control
*cc
)
1998 const int migratetype
= cc
->migratetype
;
2001 /* Compaction run completes if the migrate and free scanner meet */
2002 if (compact_scanners_met(cc
)) {
2003 /* Let the next compaction start anew. */
2004 reset_cached_positions(cc
->zone
);
2007 * Mark that the PG_migrate_skip information should be cleared
2008 * by kswapd when it goes to sleep. kcompactd does not set the
2009 * flag itself as the decision to be clear should be directly
2010 * based on an allocation request.
2012 if (cc
->direct_compaction
)
2013 cc
->zone
->compact_blockskip_flush
= true;
2016 return COMPACT_COMPLETE
;
2018 return COMPACT_PARTIAL_SKIPPED
;
2021 if (cc
->proactive_compaction
) {
2022 int score
, wmark_low
;
2025 pgdat
= cc
->zone
->zone_pgdat
;
2026 if (kswapd_is_running(pgdat
))
2027 return COMPACT_PARTIAL_SKIPPED
;
2029 score
= fragmentation_score_zone(cc
->zone
);
2030 wmark_low
= fragmentation_score_wmark(pgdat
, true);
2032 if (score
> wmark_low
)
2033 ret
= COMPACT_CONTINUE
;
2035 ret
= COMPACT_SUCCESS
;
2040 if (is_via_compact_memory(cc
->order
))
2041 return COMPACT_CONTINUE
;
2044 * Always finish scanning a pageblock to reduce the possibility of
2045 * fallbacks in the future. This is particularly important when
2046 * migration source is unmovable/reclaimable but it's not worth
2049 if (!IS_ALIGNED(cc
->migrate_pfn
, pageblock_nr_pages
))
2050 return COMPACT_CONTINUE
;
2052 /* Direct compactor: Is a suitable page free? */
2053 ret
= COMPACT_NO_SUITABLE_PAGE
;
2054 for (order
= cc
->order
; order
< MAX_ORDER
; order
++) {
2055 struct free_area
*area
= &cc
->zone
->free_area
[order
];
2058 /* Job done if page is free of the right migratetype */
2059 if (!free_area_empty(area
, migratetype
))
2060 return COMPACT_SUCCESS
;
2063 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
2064 if (migratetype
== MIGRATE_MOVABLE
&&
2065 !free_area_empty(area
, MIGRATE_CMA
))
2066 return COMPACT_SUCCESS
;
2069 * Job done if allocation would steal freepages from
2070 * other migratetype buddy lists.
2072 if (find_suitable_fallback(area
, order
, migratetype
,
2073 true, &can_steal
) != -1) {
2075 /* movable pages are OK in any pageblock */
2076 if (migratetype
== MIGRATE_MOVABLE
)
2077 return COMPACT_SUCCESS
;
2080 * We are stealing for a non-movable allocation. Make
2081 * sure we finish compacting the current pageblock
2082 * first so it is as free as possible and we won't
2083 * have to steal another one soon. This only applies
2084 * to sync compaction, as async compaction operates
2085 * on pageblocks of the same migratetype.
2087 if (cc
->mode
== MIGRATE_ASYNC
||
2088 IS_ALIGNED(cc
->migrate_pfn
,
2089 pageblock_nr_pages
)) {
2090 return COMPACT_SUCCESS
;
2093 ret
= COMPACT_CONTINUE
;
2099 if (cc
->contended
|| fatal_signal_pending(current
))
2100 ret
= COMPACT_CONTENDED
;
2105 static enum compact_result
compact_finished(struct compact_control
*cc
)
2109 ret
= __compact_finished(cc
);
2110 trace_mm_compaction_finished(cc
->zone
, cc
->order
, ret
);
2111 if (ret
== COMPACT_NO_SUITABLE_PAGE
)
2112 ret
= COMPACT_CONTINUE
;
2117 static enum compact_result
__compaction_suitable(struct zone
*zone
, int order
,
2118 unsigned int alloc_flags
,
2119 int highest_zoneidx
,
2120 unsigned long wmark_target
)
2122 unsigned long watermark
;
2124 if (is_via_compact_memory(order
))
2125 return COMPACT_CONTINUE
;
2127 watermark
= wmark_pages(zone
, alloc_flags
& ALLOC_WMARK_MASK
);
2129 * If watermarks for high-order allocation are already met, there
2130 * should be no need for compaction at all.
2132 if (zone_watermark_ok(zone
, order
, watermark
, highest_zoneidx
,
2134 return COMPACT_SUCCESS
;
2137 * Watermarks for order-0 must be met for compaction to be able to
2138 * isolate free pages for migration targets. This means that the
2139 * watermark and alloc_flags have to match, or be more pessimistic than
2140 * the check in __isolate_free_page(). We don't use the direct
2141 * compactor's alloc_flags, as they are not relevant for freepage
2142 * isolation. We however do use the direct compactor's highest_zoneidx
2143 * to skip over zones where lowmem reserves would prevent allocation
2144 * even if compaction succeeds.
2145 * For costly orders, we require low watermark instead of min for
2146 * compaction to proceed to increase its chances.
2147 * ALLOC_CMA is used, as pages in CMA pageblocks are considered
2148 * suitable migration targets
2150 watermark
= (order
> PAGE_ALLOC_COSTLY_ORDER
) ?
2151 low_wmark_pages(zone
) : min_wmark_pages(zone
);
2152 watermark
+= compact_gap(order
);
2153 if (!__zone_watermark_ok(zone
, 0, watermark
, highest_zoneidx
,
2154 ALLOC_CMA
, wmark_target
))
2155 return COMPACT_SKIPPED
;
2157 return COMPACT_CONTINUE
;
2161 * compaction_suitable: Is this suitable to run compaction on this zone now?
2163 * COMPACT_SKIPPED - If there are too few free pages for compaction
2164 * COMPACT_SUCCESS - If the allocation would succeed without compaction
2165 * COMPACT_CONTINUE - If compaction should run now
2167 enum compact_result
compaction_suitable(struct zone
*zone
, int order
,
2168 unsigned int alloc_flags
,
2169 int highest_zoneidx
)
2171 enum compact_result ret
;
2174 ret
= __compaction_suitable(zone
, order
, alloc_flags
, highest_zoneidx
,
2175 zone_page_state(zone
, NR_FREE_PAGES
));
2177 * fragmentation index determines if allocation failures are due to
2178 * low memory or external fragmentation
2180 * index of -1000 would imply allocations might succeed depending on
2181 * watermarks, but we already failed the high-order watermark check
2182 * index towards 0 implies failure is due to lack of memory
2183 * index towards 1000 implies failure is due to fragmentation
2185 * Only compact if a failure would be due to fragmentation. Also
2186 * ignore fragindex for non-costly orders where the alternative to
2187 * a successful reclaim/compaction is OOM. Fragindex and the
2188 * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
2189 * excessive compaction for costly orders, but it should not be at the
2190 * expense of system stability.
2192 if (ret
== COMPACT_CONTINUE
&& (order
> PAGE_ALLOC_COSTLY_ORDER
)) {
2193 fragindex
= fragmentation_index(zone
, order
);
2194 if (fragindex
>= 0 && fragindex
<= sysctl_extfrag_threshold
)
2195 ret
= COMPACT_NOT_SUITABLE_ZONE
;
2198 trace_mm_compaction_suitable(zone
, order
, ret
);
2199 if (ret
== COMPACT_NOT_SUITABLE_ZONE
)
2200 ret
= COMPACT_SKIPPED
;
2205 bool compaction_zonelist_suitable(struct alloc_context
*ac
, int order
,
2212 * Make sure at least one zone would pass __compaction_suitable if we continue
2213 * retrying the reclaim.
2215 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
2216 ac
->highest_zoneidx
, ac
->nodemask
) {
2217 unsigned long available
;
2218 enum compact_result compact_result
;
2221 * Do not consider all the reclaimable memory because we do not
2222 * want to trash just for a single high order allocation which
2223 * is even not guaranteed to appear even if __compaction_suitable
2224 * is happy about the watermark check.
2226 available
= zone_reclaimable_pages(zone
) / order
;
2227 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
2228 compact_result
= __compaction_suitable(zone
, order
, alloc_flags
,
2229 ac
->highest_zoneidx
, available
);
2230 if (compact_result
!= COMPACT_SKIPPED
)
2237 static enum compact_result
2238 compact_zone(struct compact_control
*cc
, struct capture_control
*capc
)
2240 enum compact_result ret
;
2241 unsigned long start_pfn
= cc
->zone
->zone_start_pfn
;
2242 unsigned long end_pfn
= zone_end_pfn(cc
->zone
);
2243 unsigned long last_migrated_pfn
;
2244 const bool sync
= cc
->mode
!= MIGRATE_ASYNC
;
2248 * These counters track activities during zone compaction. Initialize
2249 * them before compacting a new zone.
2251 cc
->total_migrate_scanned
= 0;
2252 cc
->total_free_scanned
= 0;
2253 cc
->nr_migratepages
= 0;
2254 cc
->nr_freepages
= 0;
2255 INIT_LIST_HEAD(&cc
->freepages
);
2256 INIT_LIST_HEAD(&cc
->migratepages
);
2258 cc
->migratetype
= gfp_migratetype(cc
->gfp_mask
);
2259 ret
= compaction_suitable(cc
->zone
, cc
->order
, cc
->alloc_flags
,
2260 cc
->highest_zoneidx
);
2261 /* Compaction is likely to fail */
2262 if (ret
== COMPACT_SUCCESS
|| ret
== COMPACT_SKIPPED
)
2265 /* huh, compaction_suitable is returning something unexpected */
2266 VM_BUG_ON(ret
!= COMPACT_CONTINUE
);
2269 * Clear pageblock skip if there were failures recently and compaction
2270 * is about to be retried after being deferred.
2272 if (compaction_restarting(cc
->zone
, cc
->order
))
2273 __reset_isolation_suitable(cc
->zone
);
2276 * Setup to move all movable pages to the end of the zone. Used cached
2277 * information on where the scanners should start (unless we explicitly
2278 * want to compact the whole zone), but check that it is initialised
2279 * by ensuring the values are within zone boundaries.
2281 cc
->fast_start_pfn
= 0;
2282 if (cc
->whole_zone
) {
2283 cc
->migrate_pfn
= start_pfn
;
2284 cc
->free_pfn
= pageblock_start_pfn(end_pfn
- 1);
2286 cc
->migrate_pfn
= cc
->zone
->compact_cached_migrate_pfn
[sync
];
2287 cc
->free_pfn
= cc
->zone
->compact_cached_free_pfn
;
2288 if (cc
->free_pfn
< start_pfn
|| cc
->free_pfn
>= end_pfn
) {
2289 cc
->free_pfn
= pageblock_start_pfn(end_pfn
- 1);
2290 cc
->zone
->compact_cached_free_pfn
= cc
->free_pfn
;
2292 if (cc
->migrate_pfn
< start_pfn
|| cc
->migrate_pfn
>= end_pfn
) {
2293 cc
->migrate_pfn
= start_pfn
;
2294 cc
->zone
->compact_cached_migrate_pfn
[0] = cc
->migrate_pfn
;
2295 cc
->zone
->compact_cached_migrate_pfn
[1] = cc
->migrate_pfn
;
2298 if (cc
->migrate_pfn
<= cc
->zone
->compact_init_migrate_pfn
)
2299 cc
->whole_zone
= true;
2302 last_migrated_pfn
= 0;
2305 * Migrate has separate cached PFNs for ASYNC and SYNC* migration on
2306 * the basis that some migrations will fail in ASYNC mode. However,
2307 * if the cached PFNs match and pageblocks are skipped due to having
2308 * no isolation candidates, then the sync state does not matter.
2309 * Until a pageblock with isolation candidates is found, keep the
2310 * cached PFNs in sync to avoid revisiting the same blocks.
2312 update_cached
= !sync
&&
2313 cc
->zone
->compact_cached_migrate_pfn
[0] == cc
->zone
->compact_cached_migrate_pfn
[1];
2315 trace_mm_compaction_begin(start_pfn
, cc
->migrate_pfn
,
2316 cc
->free_pfn
, end_pfn
, sync
);
2318 migrate_prep_local();
2320 while ((ret
= compact_finished(cc
)) == COMPACT_CONTINUE
) {
2322 unsigned long iteration_start_pfn
= cc
->migrate_pfn
;
2325 * Avoid multiple rescans which can happen if a page cannot be
2326 * isolated (dirty/writeback in async mode) or if the migrated
2327 * pages are being allocated before the pageblock is cleared.
2328 * The first rescan will capture the entire pageblock for
2329 * migration. If it fails, it'll be marked skip and scanning
2330 * will proceed as normal.
2333 if (pageblock_start_pfn(last_migrated_pfn
) ==
2334 pageblock_start_pfn(iteration_start_pfn
)) {
2338 switch (isolate_migratepages(cc
)) {
2340 ret
= COMPACT_CONTENDED
;
2341 putback_movable_pages(&cc
->migratepages
);
2342 cc
->nr_migratepages
= 0;
2345 if (update_cached
) {
2346 cc
->zone
->compact_cached_migrate_pfn
[1] =
2347 cc
->zone
->compact_cached_migrate_pfn
[0];
2351 * We haven't isolated and migrated anything, but
2352 * there might still be unflushed migrations from
2353 * previous cc->order aligned block.
2356 case ISOLATE_SUCCESS
:
2357 update_cached
= false;
2358 last_migrated_pfn
= iteration_start_pfn
;
2361 err
= migrate_pages(&cc
->migratepages
, compaction_alloc
,
2362 compaction_free
, (unsigned long)cc
, cc
->mode
,
2365 trace_mm_compaction_migratepages(cc
->nr_migratepages
, err
,
2368 /* All pages were either migrated or will be released */
2369 cc
->nr_migratepages
= 0;
2371 putback_movable_pages(&cc
->migratepages
);
2373 * migrate_pages() may return -ENOMEM when scanners meet
2374 * and we want compact_finished() to detect it
2376 if (err
== -ENOMEM
&& !compact_scanners_met(cc
)) {
2377 ret
= COMPACT_CONTENDED
;
2381 * We failed to migrate at least one page in the current
2382 * order-aligned block, so skip the rest of it.
2384 if (cc
->direct_compaction
&&
2385 (cc
->mode
== MIGRATE_ASYNC
)) {
2386 cc
->migrate_pfn
= block_end_pfn(
2387 cc
->migrate_pfn
- 1, cc
->order
);
2388 /* Draining pcplists is useless in this case */
2389 last_migrated_pfn
= 0;
2395 * Has the migration scanner moved away from the previous
2396 * cc->order aligned block where we migrated from? If yes,
2397 * flush the pages that were freed, so that they can merge and
2398 * compact_finished() can detect immediately if allocation
2401 if (cc
->order
> 0 && last_migrated_pfn
) {
2402 unsigned long current_block_start
=
2403 block_start_pfn(cc
->migrate_pfn
, cc
->order
);
2405 if (last_migrated_pfn
< current_block_start
) {
2406 lru_add_drain_cpu_zone(cc
->zone
);
2407 /* No more flushing until we migrate again */
2408 last_migrated_pfn
= 0;
2412 /* Stop if a page has been captured */
2413 if (capc
&& capc
->page
) {
2414 ret
= COMPACT_SUCCESS
;
2421 * Release free pages and update where the free scanner should restart,
2422 * so we don't leave any returned pages behind in the next attempt.
2424 if (cc
->nr_freepages
> 0) {
2425 unsigned long free_pfn
= release_freepages(&cc
->freepages
);
2427 cc
->nr_freepages
= 0;
2428 VM_BUG_ON(free_pfn
== 0);
2429 /* The cached pfn is always the first in a pageblock */
2430 free_pfn
= pageblock_start_pfn(free_pfn
);
2432 * Only go back, not forward. The cached pfn might have been
2433 * already reset to zone end in compact_finished()
2435 if (free_pfn
> cc
->zone
->compact_cached_free_pfn
)
2436 cc
->zone
->compact_cached_free_pfn
= free_pfn
;
2439 count_compact_events(COMPACTMIGRATE_SCANNED
, cc
->total_migrate_scanned
);
2440 count_compact_events(COMPACTFREE_SCANNED
, cc
->total_free_scanned
);
2442 trace_mm_compaction_end(start_pfn
, cc
->migrate_pfn
,
2443 cc
->free_pfn
, end_pfn
, sync
, ret
);
2448 static enum compact_result
compact_zone_order(struct zone
*zone
, int order
,
2449 gfp_t gfp_mask
, enum compact_priority prio
,
2450 unsigned int alloc_flags
, int highest_zoneidx
,
2451 struct page
**capture
)
2453 enum compact_result ret
;
2454 struct compact_control cc
= {
2456 .search_order
= order
,
2457 .gfp_mask
= gfp_mask
,
2459 .mode
= (prio
== COMPACT_PRIO_ASYNC
) ?
2460 MIGRATE_ASYNC
: MIGRATE_SYNC_LIGHT
,
2461 .alloc_flags
= alloc_flags
,
2462 .highest_zoneidx
= highest_zoneidx
,
2463 .direct_compaction
= true,
2464 .whole_zone
= (prio
== MIN_COMPACT_PRIORITY
),
2465 .ignore_skip_hint
= (prio
== MIN_COMPACT_PRIORITY
),
2466 .ignore_block_suitable
= (prio
== MIN_COMPACT_PRIORITY
)
2468 struct capture_control capc
= {
2474 * Make sure the structs are really initialized before we expose the
2475 * capture control, in case we are interrupted and the interrupt handler
2479 WRITE_ONCE(current
->capture_control
, &capc
);
2481 ret
= compact_zone(&cc
, &capc
);
2483 VM_BUG_ON(!list_empty(&cc
.freepages
));
2484 VM_BUG_ON(!list_empty(&cc
.migratepages
));
2487 * Make sure we hide capture control first before we read the captured
2488 * page pointer, otherwise an interrupt could free and capture a page
2489 * and we would leak it.
2491 WRITE_ONCE(current
->capture_control
, NULL
);
2492 *capture
= READ_ONCE(capc
.page
);
2497 int sysctl_extfrag_threshold
= 500;
2500 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
2501 * @gfp_mask: The GFP mask of the current allocation
2502 * @order: The order of the current allocation
2503 * @alloc_flags: The allocation flags of the current allocation
2504 * @ac: The context of current allocation
2505 * @prio: Determines how hard direct compaction should try to succeed
2506 * @capture: Pointer to free page created by compaction will be stored here
2508 * This is the main entry point for direct page compaction.
2510 enum compact_result
try_to_compact_pages(gfp_t gfp_mask
, unsigned int order
,
2511 unsigned int alloc_flags
, const struct alloc_context
*ac
,
2512 enum compact_priority prio
, struct page
**capture
)
2514 int may_perform_io
= gfp_mask
& __GFP_IO
;
2517 enum compact_result rc
= COMPACT_SKIPPED
;
2520 * Check if the GFP flags allow compaction - GFP_NOIO is really
2521 * tricky context because the migration might require IO
2523 if (!may_perform_io
)
2524 return COMPACT_SKIPPED
;
2526 trace_mm_compaction_try_to_compact_pages(order
, gfp_mask
, prio
);
2528 /* Compact each zone in the list */
2529 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
2530 ac
->highest_zoneidx
, ac
->nodemask
) {
2531 enum compact_result status
;
2533 if (prio
> MIN_COMPACT_PRIORITY
2534 && compaction_deferred(zone
, order
)) {
2535 rc
= max_t(enum compact_result
, COMPACT_DEFERRED
, rc
);
2539 status
= compact_zone_order(zone
, order
, gfp_mask
, prio
,
2540 alloc_flags
, ac
->highest_zoneidx
, capture
);
2541 rc
= max(status
, rc
);
2543 /* The allocation should succeed, stop compacting */
2544 if (status
== COMPACT_SUCCESS
) {
2546 * We think the allocation will succeed in this zone,
2547 * but it is not certain, hence the false. The caller
2548 * will repeat this with true if allocation indeed
2549 * succeeds in this zone.
2551 compaction_defer_reset(zone
, order
, false);
2556 if (prio
!= COMPACT_PRIO_ASYNC
&& (status
== COMPACT_COMPLETE
||
2557 status
== COMPACT_PARTIAL_SKIPPED
))
2559 * We think that allocation won't succeed in this zone
2560 * so we defer compaction there. If it ends up
2561 * succeeding after all, it will be reset.
2563 defer_compaction(zone
, order
);
2566 * We might have stopped compacting due to need_resched() in
2567 * async compaction, or due to a fatal signal detected. In that
2568 * case do not try further zones
2570 if ((prio
== COMPACT_PRIO_ASYNC
&& need_resched())
2571 || fatal_signal_pending(current
))
2579 * Compact all zones within a node till each zone's fragmentation score
2580 * reaches within proactive compaction thresholds (as determined by the
2581 * proactiveness tunable).
2583 * It is possible that the function returns before reaching score targets
2584 * due to various back-off conditions, such as, contention on per-node or
2587 static void proactive_compact_node(pg_data_t
*pgdat
)
2591 struct compact_control cc
= {
2593 .mode
= MIGRATE_SYNC_LIGHT
,
2594 .ignore_skip_hint
= true,
2596 .gfp_mask
= GFP_KERNEL
,
2597 .proactive_compaction
= true,
2600 for (zoneid
= 0; zoneid
< MAX_NR_ZONES
; zoneid
++) {
2601 zone
= &pgdat
->node_zones
[zoneid
];
2602 if (!populated_zone(zone
))
2607 compact_zone(&cc
, NULL
);
2609 VM_BUG_ON(!list_empty(&cc
.freepages
));
2610 VM_BUG_ON(!list_empty(&cc
.migratepages
));
2614 /* Compact all zones within a node */
2615 static void compact_node(int nid
)
2617 pg_data_t
*pgdat
= NODE_DATA(nid
);
2620 struct compact_control cc
= {
2622 .mode
= MIGRATE_SYNC
,
2623 .ignore_skip_hint
= true,
2625 .gfp_mask
= GFP_KERNEL
,
2629 for (zoneid
= 0; zoneid
< MAX_NR_ZONES
; zoneid
++) {
2631 zone
= &pgdat
->node_zones
[zoneid
];
2632 if (!populated_zone(zone
))
2637 compact_zone(&cc
, NULL
);
2639 VM_BUG_ON(!list_empty(&cc
.freepages
));
2640 VM_BUG_ON(!list_empty(&cc
.migratepages
));
2644 /* Compact all nodes in the system */
2645 static void compact_nodes(void)
2649 /* Flush pending updates to the LRU lists */
2650 lru_add_drain_all();
2652 for_each_online_node(nid
)
2656 /* The written value is actually unused, all memory is compacted */
2657 int sysctl_compact_memory
;
2660 * Tunable for proactive compaction. It determines how
2661 * aggressively the kernel should compact memory in the
2662 * background. It takes values in the range [0, 100].
2664 unsigned int __read_mostly sysctl_compaction_proactiveness
= 20;
2667 * This is the entry point for compacting all nodes via
2668 * /proc/sys/vm/compact_memory
2670 int sysctl_compaction_handler(struct ctl_table
*table
, int write
,
2671 void *buffer
, size_t *length
, loff_t
*ppos
)
2679 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
2680 static ssize_t
sysfs_compact_node(struct device
*dev
,
2681 struct device_attribute
*attr
,
2682 const char *buf
, size_t count
)
2686 if (nid
>= 0 && nid
< nr_node_ids
&& node_online(nid
)) {
2687 /* Flush pending updates to the LRU lists */
2688 lru_add_drain_all();
2695 static DEVICE_ATTR(compact
, 0200, NULL
, sysfs_compact_node
);
2697 int compaction_register_node(struct node
*node
)
2699 return device_create_file(&node
->dev
, &dev_attr_compact
);
2702 void compaction_unregister_node(struct node
*node
)
2704 return device_remove_file(&node
->dev
, &dev_attr_compact
);
2706 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
2708 static inline bool kcompactd_work_requested(pg_data_t
*pgdat
)
2710 return pgdat
->kcompactd_max_order
> 0 || kthread_should_stop();
2713 static bool kcompactd_node_suitable(pg_data_t
*pgdat
)
2717 enum zone_type highest_zoneidx
= pgdat
->kcompactd_highest_zoneidx
;
2719 for (zoneid
= 0; zoneid
<= highest_zoneidx
; zoneid
++) {
2720 zone
= &pgdat
->node_zones
[zoneid
];
2722 if (!populated_zone(zone
))
2725 if (compaction_suitable(zone
, pgdat
->kcompactd_max_order
, 0,
2726 highest_zoneidx
) == COMPACT_CONTINUE
)
2733 static void kcompactd_do_work(pg_data_t
*pgdat
)
2736 * With no special task, compact all zones so that a page of requested
2737 * order is allocatable.
2741 struct compact_control cc
= {
2742 .order
= pgdat
->kcompactd_max_order
,
2743 .search_order
= pgdat
->kcompactd_max_order
,
2744 .highest_zoneidx
= pgdat
->kcompactd_highest_zoneidx
,
2745 .mode
= MIGRATE_SYNC_LIGHT
,
2746 .ignore_skip_hint
= false,
2747 .gfp_mask
= GFP_KERNEL
,
2749 trace_mm_compaction_kcompactd_wake(pgdat
->node_id
, cc
.order
,
2750 cc
.highest_zoneidx
);
2751 count_compact_event(KCOMPACTD_WAKE
);
2753 for (zoneid
= 0; zoneid
<= cc
.highest_zoneidx
; zoneid
++) {
2756 zone
= &pgdat
->node_zones
[zoneid
];
2757 if (!populated_zone(zone
))
2760 if (compaction_deferred(zone
, cc
.order
))
2763 if (compaction_suitable(zone
, cc
.order
, 0, zoneid
) !=
2767 if (kthread_should_stop())
2771 status
= compact_zone(&cc
, NULL
);
2773 if (status
== COMPACT_SUCCESS
) {
2774 compaction_defer_reset(zone
, cc
.order
, false);
2775 } else if (status
== COMPACT_PARTIAL_SKIPPED
|| status
== COMPACT_COMPLETE
) {
2777 * Buddy pages may become stranded on pcps that could
2778 * otherwise coalesce on the zone's free area for
2779 * order >= cc.order. This is ratelimited by the
2780 * upcoming deferral.
2782 drain_all_pages(zone
);
2785 * We use sync migration mode here, so we defer like
2786 * sync direct compaction does.
2788 defer_compaction(zone
, cc
.order
);
2791 count_compact_events(KCOMPACTD_MIGRATE_SCANNED
,
2792 cc
.total_migrate_scanned
);
2793 count_compact_events(KCOMPACTD_FREE_SCANNED
,
2794 cc
.total_free_scanned
);
2796 VM_BUG_ON(!list_empty(&cc
.freepages
));
2797 VM_BUG_ON(!list_empty(&cc
.migratepages
));
2801 * Regardless of success, we are done until woken up next. But remember
2802 * the requested order/highest_zoneidx in case it was higher/tighter
2803 * than our current ones
2805 if (pgdat
->kcompactd_max_order
<= cc
.order
)
2806 pgdat
->kcompactd_max_order
= 0;
2807 if (pgdat
->kcompactd_highest_zoneidx
>= cc
.highest_zoneidx
)
2808 pgdat
->kcompactd_highest_zoneidx
= pgdat
->nr_zones
- 1;
2811 void wakeup_kcompactd(pg_data_t
*pgdat
, int order
, int highest_zoneidx
)
2816 if (pgdat
->kcompactd_max_order
< order
)
2817 pgdat
->kcompactd_max_order
= order
;
2819 if (pgdat
->kcompactd_highest_zoneidx
> highest_zoneidx
)
2820 pgdat
->kcompactd_highest_zoneidx
= highest_zoneidx
;
2823 * Pairs with implicit barrier in wait_event_freezable()
2824 * such that wakeups are not missed.
2826 if (!wq_has_sleeper(&pgdat
->kcompactd_wait
))
2829 if (!kcompactd_node_suitable(pgdat
))
2832 trace_mm_compaction_wakeup_kcompactd(pgdat
->node_id
, order
,
2834 wake_up_interruptible(&pgdat
->kcompactd_wait
);
2838 * The background compaction daemon, started as a kernel thread
2839 * from the init process.
2841 static int kcompactd(void *p
)
2843 pg_data_t
*pgdat
= (pg_data_t
*)p
;
2844 struct task_struct
*tsk
= current
;
2845 unsigned int proactive_defer
= 0;
2847 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
2849 if (!cpumask_empty(cpumask
))
2850 set_cpus_allowed_ptr(tsk
, cpumask
);
2854 pgdat
->kcompactd_max_order
= 0;
2855 pgdat
->kcompactd_highest_zoneidx
= pgdat
->nr_zones
- 1;
2857 while (!kthread_should_stop()) {
2858 unsigned long pflags
;
2860 trace_mm_compaction_kcompactd_sleep(pgdat
->node_id
);
2861 if (wait_event_freezable_timeout(pgdat
->kcompactd_wait
,
2862 kcompactd_work_requested(pgdat
),
2863 msecs_to_jiffies(HPAGE_FRAG_CHECK_INTERVAL_MSEC
))) {
2865 psi_memstall_enter(&pflags
);
2866 kcompactd_do_work(pgdat
);
2867 psi_memstall_leave(&pflags
);
2871 /* kcompactd wait timeout */
2872 if (should_proactive_compact_node(pgdat
)) {
2873 unsigned int prev_score
, score
;
2875 if (proactive_defer
) {
2879 prev_score
= fragmentation_score_node(pgdat
);
2880 proactive_compact_node(pgdat
);
2881 score
= fragmentation_score_node(pgdat
);
2883 * Defer proactive compaction if the fragmentation
2884 * score did not go down i.e. no progress made.
2886 proactive_defer
= score
< prev_score
?
2887 0 : 1 << COMPACT_MAX_DEFER_SHIFT
;
2895 * This kcompactd start function will be called by init and node-hot-add.
2896 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
2898 int kcompactd_run(int nid
)
2900 pg_data_t
*pgdat
= NODE_DATA(nid
);
2903 if (pgdat
->kcompactd
)
2906 pgdat
->kcompactd
= kthread_run(kcompactd
, pgdat
, "kcompactd%d", nid
);
2907 if (IS_ERR(pgdat
->kcompactd
)) {
2908 pr_err("Failed to start kcompactd on node %d\n", nid
);
2909 ret
= PTR_ERR(pgdat
->kcompactd
);
2910 pgdat
->kcompactd
= NULL
;
2916 * Called by memory hotplug when all memory in a node is offlined. Caller must
2917 * hold mem_hotplug_begin/end().
2919 void kcompactd_stop(int nid
)
2921 struct task_struct
*kcompactd
= NODE_DATA(nid
)->kcompactd
;
2924 kthread_stop(kcompactd
);
2925 NODE_DATA(nid
)->kcompactd
= NULL
;
2930 * It's optimal to keep kcompactd on the same CPUs as their memory, but
2931 * not required for correctness. So if the last cpu in a node goes
2932 * away, we get changed to run anywhere: as the first one comes back,
2933 * restore their cpu bindings.
2935 static int kcompactd_cpu_online(unsigned int cpu
)
2939 for_each_node_state(nid
, N_MEMORY
) {
2940 pg_data_t
*pgdat
= NODE_DATA(nid
);
2941 const struct cpumask
*mask
;
2943 mask
= cpumask_of_node(pgdat
->node_id
);
2945 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
2946 /* One of our CPUs online: restore mask */
2947 set_cpus_allowed_ptr(pgdat
->kcompactd
, mask
);
2952 static int __init
kcompactd_init(void)
2957 ret
= cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN
,
2958 "mm/compaction:online",
2959 kcompactd_cpu_online
, NULL
);
2961 pr_err("kcompactd: failed to register hotplug callbacks.\n");
2965 for_each_node_state(nid
, N_MEMORY
)
2969 subsys_initcall(kcompactd_init
)
2971 #endif /* CONFIG_COMPACTION */