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 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 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
;
186 if (zone
->compact_considered
>= defer_limit
)
189 trace_mm_compaction_deferred(zone
, order
);
195 * Update defer tracking counters after successful compaction of given order,
196 * which means an allocation either succeeded (alloc_success == true) or is
197 * expected to succeed.
199 void compaction_defer_reset(struct zone
*zone
, int order
,
203 zone
->compact_considered
= 0;
204 zone
->compact_defer_shift
= 0;
206 if (order
>= zone
->compact_order_failed
)
207 zone
->compact_order_failed
= order
+ 1;
209 trace_mm_compaction_defer_reset(zone
, order
);
212 /* Returns true if restarting compaction after many failures */
213 bool compaction_restarting(struct zone
*zone
, int order
)
215 if (order
< zone
->compact_order_failed
)
218 return zone
->compact_defer_shift
== COMPACT_MAX_DEFER_SHIFT
&&
219 zone
->compact_considered
>= 1UL << zone
->compact_defer_shift
;
222 /* Returns true if the pageblock should be scanned for pages to isolate. */
223 static inline bool isolation_suitable(struct compact_control
*cc
,
226 if (cc
->ignore_skip_hint
)
229 return !get_pageblock_skip(page
);
232 static void reset_cached_positions(struct zone
*zone
)
234 zone
->compact_cached_migrate_pfn
[0] = zone
->zone_start_pfn
;
235 zone
->compact_cached_migrate_pfn
[1] = zone
->zone_start_pfn
;
236 zone
->compact_cached_free_pfn
=
237 pageblock_start_pfn(zone_end_pfn(zone
) - 1);
241 * Compound pages of >= pageblock_order should consistenly be skipped until
242 * released. It is always pointless to compact pages of such order (if they are
243 * migratable), and the pageblocks they occupy cannot contain any free pages.
245 static bool pageblock_skip_persistent(struct page
*page
)
247 if (!PageCompound(page
))
250 page
= compound_head(page
);
252 if (compound_order(page
) >= pageblock_order
)
259 __reset_isolation_pfn(struct zone
*zone
, unsigned long pfn
, bool check_source
,
262 struct page
*page
= pfn_to_online_page(pfn
);
263 struct page
*block_page
;
264 struct page
*end_page
;
265 unsigned long block_pfn
;
269 if (zone
!= page_zone(page
))
271 if (pageblock_skip_persistent(page
))
275 * If skip is already cleared do no further checking once the
276 * restart points have been set.
278 if (check_source
&& check_target
&& !get_pageblock_skip(page
))
282 * If clearing skip for the target scanner, do not select a
283 * non-movable pageblock as the starting point.
285 if (!check_source
&& check_target
&&
286 get_pageblock_migratetype(page
) != MIGRATE_MOVABLE
)
289 /* Ensure the start of the pageblock or zone is online and valid */
290 block_pfn
= pageblock_start_pfn(pfn
);
291 block_pfn
= max(block_pfn
, zone
->zone_start_pfn
);
292 block_page
= pfn_to_online_page(block_pfn
);
298 /* Ensure the end of the pageblock or zone is online and valid */
299 block_pfn
= pageblock_end_pfn(pfn
) - 1;
300 block_pfn
= min(block_pfn
, zone_end_pfn(zone
) - 1);
301 end_page
= pfn_to_online_page(block_pfn
);
306 * Only clear the hint if a sample indicates there is either a
307 * free page or an LRU page in the block. One or other condition
308 * is necessary for the block to be a migration source/target.
311 if (pfn_valid_within(pfn
)) {
312 if (check_source
&& PageLRU(page
)) {
313 clear_pageblock_skip(page
);
317 if (check_target
&& PageBuddy(page
)) {
318 clear_pageblock_skip(page
);
323 page
+= (1 << PAGE_ALLOC_COSTLY_ORDER
);
324 pfn
+= (1 << PAGE_ALLOC_COSTLY_ORDER
);
325 } while (page
<= end_page
);
331 * This function is called to clear all cached information on pageblocks that
332 * should be skipped for page isolation when the migrate and free page scanner
335 static void __reset_isolation_suitable(struct zone
*zone
)
337 unsigned long migrate_pfn
= zone
->zone_start_pfn
;
338 unsigned long free_pfn
= zone_end_pfn(zone
) - 1;
339 unsigned long reset_migrate
= free_pfn
;
340 unsigned long reset_free
= migrate_pfn
;
341 bool source_set
= false;
342 bool free_set
= false;
344 if (!zone
->compact_blockskip_flush
)
347 zone
->compact_blockskip_flush
= false;
350 * Walk the zone and update pageblock skip information. Source looks
351 * for PageLRU while target looks for PageBuddy. When the scanner
352 * is found, both PageBuddy and PageLRU are checked as the pageblock
353 * is suitable as both source and target.
355 for (; migrate_pfn
< free_pfn
; migrate_pfn
+= pageblock_nr_pages
,
356 free_pfn
-= pageblock_nr_pages
) {
359 /* Update the migrate PFN */
360 if (__reset_isolation_pfn(zone
, migrate_pfn
, true, source_set
) &&
361 migrate_pfn
< reset_migrate
) {
363 reset_migrate
= migrate_pfn
;
364 zone
->compact_init_migrate_pfn
= reset_migrate
;
365 zone
->compact_cached_migrate_pfn
[0] = reset_migrate
;
366 zone
->compact_cached_migrate_pfn
[1] = reset_migrate
;
369 /* Update the free PFN */
370 if (__reset_isolation_pfn(zone
, free_pfn
, free_set
, true) &&
371 free_pfn
> reset_free
) {
373 reset_free
= free_pfn
;
374 zone
->compact_init_free_pfn
= reset_free
;
375 zone
->compact_cached_free_pfn
= reset_free
;
379 /* Leave no distance if no suitable block was reset */
380 if (reset_migrate
>= reset_free
) {
381 zone
->compact_cached_migrate_pfn
[0] = migrate_pfn
;
382 zone
->compact_cached_migrate_pfn
[1] = migrate_pfn
;
383 zone
->compact_cached_free_pfn
= free_pfn
;
387 void reset_isolation_suitable(pg_data_t
*pgdat
)
391 for (zoneid
= 0; zoneid
< MAX_NR_ZONES
; zoneid
++) {
392 struct zone
*zone
= &pgdat
->node_zones
[zoneid
];
393 if (!populated_zone(zone
))
396 /* Only flush if a full compaction finished recently */
397 if (zone
->compact_blockskip_flush
)
398 __reset_isolation_suitable(zone
);
403 * Sets the pageblock skip bit if it was clear. Note that this is a hint as
404 * locks are not required for read/writers. Returns true if it was already set.
406 static bool test_and_set_skip(struct compact_control
*cc
, struct page
*page
,
411 /* Do no update if skip hint is being ignored */
412 if (cc
->ignore_skip_hint
)
415 if (!IS_ALIGNED(pfn
, pageblock_nr_pages
))
418 skip
= get_pageblock_skip(page
);
419 if (!skip
&& !cc
->no_set_skip_hint
)
420 set_pageblock_skip(page
);
425 static void update_cached_migrate(struct compact_control
*cc
, unsigned long pfn
)
427 struct zone
*zone
= cc
->zone
;
429 pfn
= pageblock_end_pfn(pfn
);
431 /* Set for isolation rather than compaction */
432 if (cc
->no_set_skip_hint
)
435 if (pfn
> zone
->compact_cached_migrate_pfn
[0])
436 zone
->compact_cached_migrate_pfn
[0] = pfn
;
437 if (cc
->mode
!= MIGRATE_ASYNC
&&
438 pfn
> zone
->compact_cached_migrate_pfn
[1])
439 zone
->compact_cached_migrate_pfn
[1] = pfn
;
443 * If no pages were isolated then mark this pageblock to be skipped in the
444 * future. The information is later cleared by __reset_isolation_suitable().
446 static void update_pageblock_skip(struct compact_control
*cc
,
447 struct page
*page
, unsigned long pfn
)
449 struct zone
*zone
= cc
->zone
;
451 if (cc
->no_set_skip_hint
)
457 set_pageblock_skip(page
);
459 /* Update where async and sync compaction should restart */
460 if (pfn
< zone
->compact_cached_free_pfn
)
461 zone
->compact_cached_free_pfn
= pfn
;
464 static inline bool isolation_suitable(struct compact_control
*cc
,
470 static inline bool pageblock_skip_persistent(struct page
*page
)
475 static inline void update_pageblock_skip(struct compact_control
*cc
,
476 struct page
*page
, unsigned long pfn
)
480 static void update_cached_migrate(struct compact_control
*cc
, unsigned long pfn
)
484 static bool test_and_set_skip(struct compact_control
*cc
, struct page
*page
,
489 #endif /* CONFIG_COMPACTION */
492 * Compaction requires the taking of some coarse locks that are potentially
493 * very heavily contended. For async compaction, trylock and record if the
494 * lock is contended. The lock will still be acquired but compaction will
495 * abort when the current block is finished regardless of success rate.
496 * Sync compaction acquires the lock.
498 * Always returns true which makes it easier to track lock state in callers.
500 static bool compact_lock_irqsave(spinlock_t
*lock
, unsigned long *flags
,
501 struct compact_control
*cc
)
504 /* Track if the lock is contended in async mode */
505 if (cc
->mode
== MIGRATE_ASYNC
&& !cc
->contended
) {
506 if (spin_trylock_irqsave(lock
, *flags
))
509 cc
->contended
= true;
512 spin_lock_irqsave(lock
, *flags
);
517 * Compaction requires the taking of some coarse locks that are potentially
518 * very heavily contended. The lock should be periodically unlocked to avoid
519 * having disabled IRQs for a long time, even when there is nobody waiting on
520 * the lock. It might also be that allowing the IRQs will result in
521 * need_resched() becoming true. If scheduling is needed, async compaction
522 * aborts. Sync compaction schedules.
523 * Either compaction type will also abort if a fatal signal is pending.
524 * In either case if the lock was locked, it is dropped and not regained.
526 * Returns true if compaction should abort due to fatal signal pending, or
527 * async compaction due to need_resched()
528 * Returns false when compaction can continue (sync compaction might have
531 static bool compact_unlock_should_abort(spinlock_t
*lock
,
532 unsigned long flags
, bool *locked
, struct compact_control
*cc
)
535 spin_unlock_irqrestore(lock
, flags
);
539 if (fatal_signal_pending(current
)) {
540 cc
->contended
= true;
550 * Isolate free pages onto a private freelist. If @strict is true, will abort
551 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
552 * (even though it may still end up isolating some pages).
554 static unsigned long isolate_freepages_block(struct compact_control
*cc
,
555 unsigned long *start_pfn
,
556 unsigned long end_pfn
,
557 struct list_head
*freelist
,
561 int nr_scanned
= 0, total_isolated
= 0;
563 unsigned long flags
= 0;
565 unsigned long blockpfn
= *start_pfn
;
568 /* Strict mode is for isolation, speed is secondary */
572 cursor
= pfn_to_page(blockpfn
);
574 /* Isolate free pages. */
575 for (; blockpfn
< end_pfn
; blockpfn
+= stride
, cursor
+= stride
) {
577 struct page
*page
= cursor
;
580 * Periodically drop the lock (if held) regardless of its
581 * contention, to give chance to IRQs. Abort if fatal signal
582 * pending or async compaction detects need_resched()
584 if (!(blockpfn
% SWAP_CLUSTER_MAX
)
585 && compact_unlock_should_abort(&cc
->zone
->lock
, flags
,
590 if (!pfn_valid_within(blockpfn
))
594 * For compound pages such as THP and hugetlbfs, we can save
595 * potentially a lot of iterations if we skip them at once.
596 * The check is racy, but we can consider only valid values
597 * and the only danger is skipping too much.
599 if (PageCompound(page
)) {
600 const unsigned int order
= compound_order(page
);
602 if (likely(order
< MAX_ORDER
)) {
603 blockpfn
+= (1UL << order
) - 1;
604 cursor
+= (1UL << order
) - 1;
609 if (!PageBuddy(page
))
613 * If we already hold the lock, we can skip some rechecking.
614 * Note that if we hold the lock now, checked_pageblock was
615 * already set in some previous iteration (or strict is true),
616 * so it is correct to skip the suitable migration target
620 locked
= compact_lock_irqsave(&cc
->zone
->lock
,
623 /* Recheck this is a buddy page under lock */
624 if (!PageBuddy(page
))
628 /* Found a free page, will break it into order-0 pages */
629 order
= page_order(page
);
630 isolated
= __isolate_free_page(page
, order
);
633 set_page_private(page
, order
);
635 total_isolated
+= isolated
;
636 cc
->nr_freepages
+= isolated
;
637 list_add_tail(&page
->lru
, freelist
);
639 if (!strict
&& cc
->nr_migratepages
<= cc
->nr_freepages
) {
640 blockpfn
+= isolated
;
643 /* Advance to the end of split page */
644 blockpfn
+= isolated
- 1;
645 cursor
+= isolated
- 1;
657 spin_unlock_irqrestore(&cc
->zone
->lock
, flags
);
660 * There is a tiny chance that we have read bogus compound_order(),
661 * so be careful to not go outside of the pageblock.
663 if (unlikely(blockpfn
> end_pfn
))
666 trace_mm_compaction_isolate_freepages(*start_pfn
, blockpfn
,
667 nr_scanned
, total_isolated
);
669 /* Record how far we have got within the block */
670 *start_pfn
= blockpfn
;
673 * If strict isolation is requested by CMA then check that all the
674 * pages requested were isolated. If there were any failures, 0 is
675 * returned and CMA will fail.
677 if (strict
&& blockpfn
< end_pfn
)
680 cc
->total_free_scanned
+= nr_scanned
;
682 count_compact_events(COMPACTISOLATED
, total_isolated
);
683 return total_isolated
;
687 * isolate_freepages_range() - isolate free pages.
688 * @cc: Compaction control structure.
689 * @start_pfn: The first PFN to start isolating.
690 * @end_pfn: The one-past-last PFN.
692 * Non-free pages, invalid PFNs, or zone boundaries within the
693 * [start_pfn, end_pfn) range are considered errors, cause function to
694 * undo its actions and return zero.
696 * Otherwise, function returns one-past-the-last PFN of isolated page
697 * (which may be greater then end_pfn if end fell in a middle of
701 isolate_freepages_range(struct compact_control
*cc
,
702 unsigned long start_pfn
, unsigned long end_pfn
)
704 unsigned long isolated
, pfn
, block_start_pfn
, block_end_pfn
;
708 block_start_pfn
= pageblock_start_pfn(pfn
);
709 if (block_start_pfn
< cc
->zone
->zone_start_pfn
)
710 block_start_pfn
= cc
->zone
->zone_start_pfn
;
711 block_end_pfn
= pageblock_end_pfn(pfn
);
713 for (; pfn
< end_pfn
; pfn
+= isolated
,
714 block_start_pfn
= block_end_pfn
,
715 block_end_pfn
+= pageblock_nr_pages
) {
716 /* Protect pfn from changing by isolate_freepages_block */
717 unsigned long isolate_start_pfn
= pfn
;
719 block_end_pfn
= min(block_end_pfn
, end_pfn
);
722 * pfn could pass the block_end_pfn if isolated freepage
723 * is more than pageblock order. In this case, we adjust
724 * scanning range to right one.
726 if (pfn
>= block_end_pfn
) {
727 block_start_pfn
= pageblock_start_pfn(pfn
);
728 block_end_pfn
= pageblock_end_pfn(pfn
);
729 block_end_pfn
= min(block_end_pfn
, end_pfn
);
732 if (!pageblock_pfn_to_page(block_start_pfn
,
733 block_end_pfn
, cc
->zone
))
736 isolated
= isolate_freepages_block(cc
, &isolate_start_pfn
,
737 block_end_pfn
, &freelist
, 0, true);
740 * In strict mode, isolate_freepages_block() returns 0 if
741 * there are any holes in the block (ie. invalid PFNs or
748 * If we managed to isolate pages, it is always (1 << n) *
749 * pageblock_nr_pages for some non-negative n. (Max order
750 * page may span two pageblocks).
754 /* __isolate_free_page() does not map the pages */
755 split_map_pages(&freelist
);
758 /* Loop terminated early, cleanup. */
759 release_freepages(&freelist
);
763 /* We don't use freelists for anything. */
767 /* Similar to reclaim, but different enough that they don't share logic */
768 static bool too_many_isolated(pg_data_t
*pgdat
)
770 unsigned long active
, inactive
, isolated
;
772 inactive
= node_page_state(pgdat
, NR_INACTIVE_FILE
) +
773 node_page_state(pgdat
, NR_INACTIVE_ANON
);
774 active
= node_page_state(pgdat
, NR_ACTIVE_FILE
) +
775 node_page_state(pgdat
, NR_ACTIVE_ANON
);
776 isolated
= node_page_state(pgdat
, NR_ISOLATED_FILE
) +
777 node_page_state(pgdat
, NR_ISOLATED_ANON
);
779 return isolated
> (inactive
+ active
) / 2;
783 * isolate_migratepages_block() - isolate all migrate-able pages within
785 * @cc: Compaction control structure.
786 * @low_pfn: The first PFN to isolate
787 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
788 * @isolate_mode: Isolation mode to be used.
790 * Isolate all pages that can be migrated from the range specified by
791 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
792 * Returns zero if there is a fatal signal pending, otherwise PFN of the
793 * first page that was not scanned (which may be both less, equal to or more
796 * The pages are isolated on cc->migratepages list (not required to be empty),
797 * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
798 * is neither read nor updated.
801 isolate_migratepages_block(struct compact_control
*cc
, unsigned long low_pfn
,
802 unsigned long end_pfn
, isolate_mode_t isolate_mode
)
804 pg_data_t
*pgdat
= cc
->zone
->zone_pgdat
;
805 unsigned long nr_scanned
= 0, nr_isolated
= 0;
806 struct lruvec
*lruvec
;
807 unsigned long flags
= 0;
809 struct page
*page
= NULL
, *valid_page
= NULL
;
810 unsigned long start_pfn
= low_pfn
;
811 bool skip_on_failure
= false;
812 unsigned long next_skip_pfn
= 0;
813 bool skip_updated
= false;
816 * Ensure that there are not too many pages isolated from the LRU
817 * list by either parallel reclaimers or compaction. If there are,
818 * delay for some time until fewer pages are isolated
820 while (unlikely(too_many_isolated(pgdat
))) {
821 /* async migration should just abort */
822 if (cc
->mode
== MIGRATE_ASYNC
)
825 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
827 if (fatal_signal_pending(current
))
833 if (cc
->direct_compaction
&& (cc
->mode
== MIGRATE_ASYNC
)) {
834 skip_on_failure
= true;
835 next_skip_pfn
= block_end_pfn(low_pfn
, cc
->order
);
838 /* Time to isolate some pages for migration */
839 for (; low_pfn
< end_pfn
; low_pfn
++) {
841 if (skip_on_failure
&& low_pfn
>= next_skip_pfn
) {
843 * We have isolated all migration candidates in the
844 * previous order-aligned block, and did not skip it due
845 * to failure. We should migrate the pages now and
846 * hopefully succeed compaction.
852 * We failed to isolate in the previous order-aligned
853 * block. Set the new boundary to the end of the
854 * current block. Note we can't simply increase
855 * next_skip_pfn by 1 << order, as low_pfn might have
856 * been incremented by a higher number due to skipping
857 * a compound or a high-order buddy page in the
858 * previous loop iteration.
860 next_skip_pfn
= block_end_pfn(low_pfn
, cc
->order
);
864 * Periodically drop the lock (if held) regardless of its
865 * contention, to give chance to IRQs. Abort completely if
866 * a fatal signal is pending.
868 if (!(low_pfn
% SWAP_CLUSTER_MAX
)
869 && compact_unlock_should_abort(&pgdat
->lru_lock
,
870 flags
, &locked
, cc
)) {
875 if (!pfn_valid_within(low_pfn
))
879 page
= pfn_to_page(low_pfn
);
882 * Check if the pageblock has already been marked skipped.
883 * Only the aligned PFN is checked as the caller isolates
884 * COMPACT_CLUSTER_MAX at a time so the second call must
885 * not falsely conclude that the block should be skipped.
887 if (!valid_page
&& IS_ALIGNED(low_pfn
, pageblock_nr_pages
)) {
888 if (!cc
->ignore_skip_hint
&& get_pageblock_skip(page
)) {
896 * Skip if free. We read page order here without zone lock
897 * which is generally unsafe, but the race window is small and
898 * the worst thing that can happen is that we skip some
899 * potential isolation targets.
901 if (PageBuddy(page
)) {
902 unsigned long freepage_order
= page_order_unsafe(page
);
905 * Without lock, we cannot be sure that what we got is
906 * a valid page order. Consider only values in the
907 * valid order range to prevent low_pfn overflow.
909 if (freepage_order
> 0 && freepage_order
< MAX_ORDER
)
910 low_pfn
+= (1UL << freepage_order
) - 1;
915 * Regardless of being on LRU, compound pages such as THP and
916 * hugetlbfs are not to be compacted unless we are attempting
917 * an allocation much larger than the huge page size (eg CMA).
918 * We can potentially save a lot of iterations if we skip them
919 * at once. The check is racy, but we can consider only valid
920 * values and the only danger is skipping too much.
922 if (PageCompound(page
) && !cc
->alloc_contig
) {
923 const unsigned int order
= compound_order(page
);
925 if (likely(order
< MAX_ORDER
))
926 low_pfn
+= (1UL << order
) - 1;
931 * Check may be lockless but that's ok as we recheck later.
932 * It's possible to migrate LRU and non-lru movable pages.
933 * Skip any other type of page
935 if (!PageLRU(page
)) {
937 * __PageMovable can return false positive so we need
938 * to verify it under page_lock.
940 if (unlikely(__PageMovable(page
)) &&
941 !PageIsolated(page
)) {
943 spin_unlock_irqrestore(&pgdat
->lru_lock
,
948 if (!isolate_movable_page(page
, isolate_mode
))
949 goto isolate_success
;
956 * Migration will fail if an anonymous page is pinned in memory,
957 * so avoid taking lru_lock and isolating it unnecessarily in an
958 * admittedly racy check.
960 if (!page_mapping(page
) &&
961 page_count(page
) > page_mapcount(page
))
965 * Only allow to migrate anonymous pages in GFP_NOFS context
966 * because those do not depend on fs locks.
968 if (!(cc
->gfp_mask
& __GFP_FS
) && page_mapping(page
))
971 /* If we already hold the lock, we can skip some rechecking */
973 locked
= compact_lock_irqsave(&pgdat
->lru_lock
,
976 /* Try get exclusive access under lock */
979 if (test_and_set_skip(cc
, page
, low_pfn
))
983 /* Recheck PageLRU and PageCompound under lock */
988 * Page become compound since the non-locked check,
989 * and it's on LRU. It can only be a THP so the order
990 * is safe to read and it's 0 for tail pages.
992 if (unlikely(PageCompound(page
) && !cc
->alloc_contig
)) {
993 low_pfn
+= compound_nr(page
) - 1;
998 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
1000 /* Try isolate the page */
1001 if (__isolate_lru_page(page
, isolate_mode
) != 0)
1004 /* The whole page is taken off the LRU; skip the tail pages. */
1005 if (PageCompound(page
))
1006 low_pfn
+= compound_nr(page
) - 1;
1008 /* Successfully isolated */
1009 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
1010 mod_node_page_state(page_pgdat(page
),
1011 NR_ISOLATED_ANON
+ page_is_file_lru(page
),
1012 hpage_nr_pages(page
));
1015 list_add(&page
->lru
, &cc
->migratepages
);
1016 cc
->nr_migratepages
++;
1020 * Avoid isolating too much unless this block is being
1021 * rescanned (e.g. dirty/writeback pages, parallel allocation)
1022 * or a lock is contended. For contention, isolate quickly to
1023 * potentially remove one source of contention.
1025 if (cc
->nr_migratepages
== COMPACT_CLUSTER_MAX
&&
1026 !cc
->rescan
&& !cc
->contended
) {
1033 if (!skip_on_failure
)
1037 * We have isolated some pages, but then failed. Release them
1038 * instead of migrating, as we cannot form the cc->order buddy
1043 spin_unlock_irqrestore(&pgdat
->lru_lock
, flags
);
1046 putback_movable_pages(&cc
->migratepages
);
1047 cc
->nr_migratepages
= 0;
1051 if (low_pfn
< next_skip_pfn
) {
1052 low_pfn
= next_skip_pfn
- 1;
1054 * The check near the loop beginning would have updated
1055 * next_skip_pfn too, but this is a bit simpler.
1057 next_skip_pfn
+= 1UL << cc
->order
;
1062 * The PageBuddy() check could have potentially brought us outside
1063 * the range to be scanned.
1065 if (unlikely(low_pfn
> end_pfn
))
1070 spin_unlock_irqrestore(&pgdat
->lru_lock
, flags
);
1073 * Updated the cached scanner pfn once the pageblock has been scanned
1074 * Pages will either be migrated in which case there is no point
1075 * scanning in the near future or migration failed in which case the
1076 * failure reason may persist. The block is marked for skipping if
1077 * there were no pages isolated in the block or if the block is
1078 * rescanned twice in a row.
1080 if (low_pfn
== end_pfn
&& (!nr_isolated
|| cc
->rescan
)) {
1081 if (valid_page
&& !skip_updated
)
1082 set_pageblock_skip(valid_page
);
1083 update_cached_migrate(cc
, low_pfn
);
1086 trace_mm_compaction_isolate_migratepages(start_pfn
, low_pfn
,
1087 nr_scanned
, nr_isolated
);
1090 cc
->total_migrate_scanned
+= nr_scanned
;
1092 count_compact_events(COMPACTISOLATED
, nr_isolated
);
1098 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
1099 * @cc: Compaction control structure.
1100 * @start_pfn: The first PFN to start isolating.
1101 * @end_pfn: The one-past-last PFN.
1103 * Returns zero if isolation fails fatally due to e.g. pending signal.
1104 * Otherwise, function returns one-past-the-last PFN of isolated page
1105 * (which may be greater than end_pfn if end fell in a middle of a THP page).
1108 isolate_migratepages_range(struct compact_control
*cc
, unsigned long start_pfn
,
1109 unsigned long end_pfn
)
1111 unsigned long pfn
, block_start_pfn
, block_end_pfn
;
1113 /* Scan block by block. First and last block may be incomplete */
1115 block_start_pfn
= pageblock_start_pfn(pfn
);
1116 if (block_start_pfn
< cc
->zone
->zone_start_pfn
)
1117 block_start_pfn
= cc
->zone
->zone_start_pfn
;
1118 block_end_pfn
= pageblock_end_pfn(pfn
);
1120 for (; pfn
< end_pfn
; pfn
= block_end_pfn
,
1121 block_start_pfn
= block_end_pfn
,
1122 block_end_pfn
+= pageblock_nr_pages
) {
1124 block_end_pfn
= min(block_end_pfn
, end_pfn
);
1126 if (!pageblock_pfn_to_page(block_start_pfn
,
1127 block_end_pfn
, cc
->zone
))
1130 pfn
= isolate_migratepages_block(cc
, pfn
, block_end_pfn
,
1131 ISOLATE_UNEVICTABLE
);
1136 if (cc
->nr_migratepages
== COMPACT_CLUSTER_MAX
)
1143 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
1144 #ifdef CONFIG_COMPACTION
1146 static bool suitable_migration_source(struct compact_control
*cc
,
1151 if (pageblock_skip_persistent(page
))
1154 if ((cc
->mode
!= MIGRATE_ASYNC
) || !cc
->direct_compaction
)
1157 block_mt
= get_pageblock_migratetype(page
);
1159 if (cc
->migratetype
== MIGRATE_MOVABLE
)
1160 return is_migrate_movable(block_mt
);
1162 return block_mt
== cc
->migratetype
;
1165 /* Returns true if the page is within a block suitable for migration to */
1166 static bool suitable_migration_target(struct compact_control
*cc
,
1169 /* If the page is a large free page, then disallow migration */
1170 if (PageBuddy(page
)) {
1172 * We are checking page_order without zone->lock taken. But
1173 * the only small danger is that we skip a potentially suitable
1174 * pageblock, so it's not worth to check order for valid range.
1176 if (page_order_unsafe(page
) >= pageblock_order
)
1180 if (cc
->ignore_block_suitable
)
1183 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1184 if (is_migrate_movable(get_pageblock_migratetype(page
)))
1187 /* Otherwise skip the block */
1191 static inline unsigned int
1192 freelist_scan_limit(struct compact_control
*cc
)
1194 unsigned short shift
= BITS_PER_LONG
- 1;
1196 return (COMPACT_CLUSTER_MAX
>> min(shift
, cc
->fast_search_fail
)) + 1;
1200 * Test whether the free scanner has reached the same or lower pageblock than
1201 * the migration scanner, and compaction should thus terminate.
1203 static inline bool compact_scanners_met(struct compact_control
*cc
)
1205 return (cc
->free_pfn
>> pageblock_order
)
1206 <= (cc
->migrate_pfn
>> pageblock_order
);
1210 * Used when scanning for a suitable migration target which scans freelists
1211 * in reverse. Reorders the list such as the unscanned pages are scanned
1212 * first on the next iteration of the free scanner
1215 move_freelist_head(struct list_head
*freelist
, struct page
*freepage
)
1219 if (!list_is_last(freelist
, &freepage
->lru
)) {
1220 list_cut_before(&sublist
, freelist
, &freepage
->lru
);
1221 if (!list_empty(&sublist
))
1222 list_splice_tail(&sublist
, freelist
);
1227 * Similar to move_freelist_head except used by the migration scanner
1228 * when scanning forward. It's possible for these list operations to
1229 * move against each other if they search the free list exactly in
1233 move_freelist_tail(struct list_head
*freelist
, struct page
*freepage
)
1237 if (!list_is_first(freelist
, &freepage
->lru
)) {
1238 list_cut_position(&sublist
, freelist
, &freepage
->lru
);
1239 if (!list_empty(&sublist
))
1240 list_splice_tail(&sublist
, freelist
);
1245 fast_isolate_around(struct compact_control
*cc
, unsigned long pfn
, unsigned long nr_isolated
)
1247 unsigned long start_pfn
, end_pfn
;
1248 struct page
*page
= pfn_to_page(pfn
);
1250 /* Do not search around if there are enough pages already */
1251 if (cc
->nr_freepages
>= cc
->nr_migratepages
)
1254 /* Minimise scanning during async compaction */
1255 if (cc
->direct_compaction
&& cc
->mode
== MIGRATE_ASYNC
)
1258 /* Pageblock boundaries */
1259 start_pfn
= pageblock_start_pfn(pfn
);
1260 end_pfn
= min(pageblock_end_pfn(pfn
), zone_end_pfn(cc
->zone
)) - 1;
1263 if (start_pfn
!= pfn
) {
1264 isolate_freepages_block(cc
, &start_pfn
, pfn
, &cc
->freepages
, 1, false);
1265 if (cc
->nr_freepages
>= cc
->nr_migratepages
)
1270 start_pfn
= pfn
+ nr_isolated
;
1271 if (start_pfn
< end_pfn
)
1272 isolate_freepages_block(cc
, &start_pfn
, end_pfn
, &cc
->freepages
, 1, false);
1274 /* Skip this pageblock in the future as it's full or nearly full */
1275 if (cc
->nr_freepages
< cc
->nr_migratepages
)
1276 set_pageblock_skip(page
);
1279 /* Search orders in round-robin fashion */
1280 static int next_search_order(struct compact_control
*cc
, int order
)
1284 order
= cc
->order
- 1;
1286 /* Search wrapped around? */
1287 if (order
== cc
->search_order
) {
1289 if (cc
->search_order
< 0)
1290 cc
->search_order
= cc
->order
- 1;
1297 static unsigned long
1298 fast_isolate_freepages(struct compact_control
*cc
)
1300 unsigned int limit
= min(1U, freelist_scan_limit(cc
) >> 1);
1301 unsigned int nr_scanned
= 0;
1302 unsigned long low_pfn
, min_pfn
, high_pfn
= 0, highest
= 0;
1303 unsigned long nr_isolated
= 0;
1304 unsigned long distance
;
1305 struct page
*page
= NULL
;
1306 bool scan_start
= false;
1309 /* Full compaction passes in a negative order */
1311 return cc
->free_pfn
;
1314 * If starting the scan, use a deeper search and use the highest
1315 * PFN found if a suitable one is not found.
1317 if (cc
->free_pfn
>= cc
->zone
->compact_init_free_pfn
) {
1318 limit
= pageblock_nr_pages
>> 1;
1323 * Preferred point is in the top quarter of the scan space but take
1324 * a pfn from the top half if the search is problematic.
1326 distance
= (cc
->free_pfn
- cc
->migrate_pfn
);
1327 low_pfn
= pageblock_start_pfn(cc
->free_pfn
- (distance
>> 2));
1328 min_pfn
= pageblock_start_pfn(cc
->free_pfn
- (distance
>> 1));
1330 if (WARN_ON_ONCE(min_pfn
> low_pfn
))
1334 * Search starts from the last successful isolation order or the next
1335 * order to search after a previous failure
1337 cc
->search_order
= min_t(unsigned int, cc
->order
- 1, cc
->search_order
);
1339 for (order
= cc
->search_order
;
1340 !page
&& order
>= 0;
1341 order
= next_search_order(cc
, order
)) {
1342 struct free_area
*area
= &cc
->zone
->free_area
[order
];
1343 struct list_head
*freelist
;
1344 struct page
*freepage
;
1345 unsigned long flags
;
1346 unsigned int order_scanned
= 0;
1351 spin_lock_irqsave(&cc
->zone
->lock
, flags
);
1352 freelist
= &area
->free_list
[MIGRATE_MOVABLE
];
1353 list_for_each_entry_reverse(freepage
, freelist
, lru
) {
1358 pfn
= page_to_pfn(freepage
);
1361 highest
= pageblock_start_pfn(pfn
);
1363 if (pfn
>= low_pfn
) {
1364 cc
->fast_search_fail
= 0;
1365 cc
->search_order
= order
;
1370 if (pfn
>= min_pfn
&& pfn
> high_pfn
) {
1373 /* Shorten the scan if a candidate is found */
1377 if (order_scanned
>= limit
)
1381 /* Use a minimum pfn if a preferred one was not found */
1382 if (!page
&& high_pfn
) {
1383 page
= pfn_to_page(high_pfn
);
1385 /* Update freepage for the list reorder below */
1389 /* Reorder to so a future search skips recent pages */
1390 move_freelist_head(freelist
, freepage
);
1392 /* Isolate the page if available */
1394 if (__isolate_free_page(page
, order
)) {
1395 set_page_private(page
, order
);
1396 nr_isolated
= 1 << order
;
1397 cc
->nr_freepages
+= nr_isolated
;
1398 list_add_tail(&page
->lru
, &cc
->freepages
);
1399 count_compact_events(COMPACTISOLATED
, nr_isolated
);
1401 /* If isolation fails, abort the search */
1402 order
= cc
->search_order
+ 1;
1407 spin_unlock_irqrestore(&cc
->zone
->lock
, flags
);
1410 * Smaller scan on next order so the total scan ig related
1411 * to freelist_scan_limit.
1413 if (order_scanned
>= limit
)
1414 limit
= min(1U, limit
>> 1);
1418 cc
->fast_search_fail
++;
1421 * Use the highest PFN found above min. If one was
1422 * not found, be pessimistic for direct compaction
1423 * and use the min mark.
1426 page
= pfn_to_page(highest
);
1427 cc
->free_pfn
= highest
;
1429 if (cc
->direct_compaction
&& pfn_valid(min_pfn
)) {
1430 page
= pageblock_pfn_to_page(min_pfn
,
1431 pageblock_end_pfn(min_pfn
),
1433 cc
->free_pfn
= min_pfn
;
1439 if (highest
&& highest
>= cc
->zone
->compact_cached_free_pfn
) {
1440 highest
-= pageblock_nr_pages
;
1441 cc
->zone
->compact_cached_free_pfn
= highest
;
1444 cc
->total_free_scanned
+= nr_scanned
;
1446 return cc
->free_pfn
;
1448 low_pfn
= page_to_pfn(page
);
1449 fast_isolate_around(cc
, low_pfn
, nr_isolated
);
1454 * Based on information in the current compact_control, find blocks
1455 * suitable for isolating free pages from and then isolate them.
1457 static void isolate_freepages(struct compact_control
*cc
)
1459 struct zone
*zone
= cc
->zone
;
1461 unsigned long block_start_pfn
; /* start of current pageblock */
1462 unsigned long isolate_start_pfn
; /* exact pfn we start at */
1463 unsigned long block_end_pfn
; /* end of current pageblock */
1464 unsigned long low_pfn
; /* lowest pfn scanner is able to scan */
1465 struct list_head
*freelist
= &cc
->freepages
;
1466 unsigned int stride
;
1468 /* Try a small search of the free lists for a candidate */
1469 isolate_start_pfn
= fast_isolate_freepages(cc
);
1470 if (cc
->nr_freepages
)
1474 * Initialise the free scanner. The starting point is where we last
1475 * successfully isolated from, zone-cached value, or the end of the
1476 * zone when isolating for the first time. For looping we also need
1477 * this pfn aligned down to the pageblock boundary, because we do
1478 * block_start_pfn -= pageblock_nr_pages in the for loop.
1479 * For ending point, take care when isolating in last pageblock of a
1480 * zone which ends in the middle of a pageblock.
1481 * The low boundary is the end of the pageblock the migration scanner
1484 isolate_start_pfn
= cc
->free_pfn
;
1485 block_start_pfn
= pageblock_start_pfn(isolate_start_pfn
);
1486 block_end_pfn
= min(block_start_pfn
+ pageblock_nr_pages
,
1487 zone_end_pfn(zone
));
1488 low_pfn
= pageblock_end_pfn(cc
->migrate_pfn
);
1489 stride
= cc
->mode
== MIGRATE_ASYNC
? COMPACT_CLUSTER_MAX
: 1;
1492 * Isolate free pages until enough are available to migrate the
1493 * pages on cc->migratepages. We stop searching if the migrate
1494 * and free page scanners meet or enough free pages are isolated.
1496 for (; block_start_pfn
>= low_pfn
;
1497 block_end_pfn
= block_start_pfn
,
1498 block_start_pfn
-= pageblock_nr_pages
,
1499 isolate_start_pfn
= block_start_pfn
) {
1500 unsigned long nr_isolated
;
1503 * This can iterate a massively long zone without finding any
1504 * suitable migration targets, so periodically check resched.
1506 if (!(block_start_pfn
% (SWAP_CLUSTER_MAX
* pageblock_nr_pages
)))
1509 page
= pageblock_pfn_to_page(block_start_pfn
, block_end_pfn
,
1514 /* Check the block is suitable for migration */
1515 if (!suitable_migration_target(cc
, page
))
1518 /* If isolation recently failed, do not retry */
1519 if (!isolation_suitable(cc
, page
))
1522 /* Found a block suitable for isolating free pages from. */
1523 nr_isolated
= isolate_freepages_block(cc
, &isolate_start_pfn
,
1524 block_end_pfn
, freelist
, stride
, false);
1526 /* Update the skip hint if the full pageblock was scanned */
1527 if (isolate_start_pfn
== block_end_pfn
)
1528 update_pageblock_skip(cc
, page
, block_start_pfn
);
1530 /* Are enough freepages isolated? */
1531 if (cc
->nr_freepages
>= cc
->nr_migratepages
) {
1532 if (isolate_start_pfn
>= block_end_pfn
) {
1534 * Restart at previous pageblock if more
1535 * freepages can be isolated next time.
1538 block_start_pfn
- pageblock_nr_pages
;
1541 } else if (isolate_start_pfn
< block_end_pfn
) {
1543 * If isolation failed early, do not continue
1549 /* Adjust stride depending on isolation */
1554 stride
= min_t(unsigned int, COMPACT_CLUSTER_MAX
, stride
<< 1);
1558 * Record where the free scanner will restart next time. Either we
1559 * broke from the loop and set isolate_start_pfn based on the last
1560 * call to isolate_freepages_block(), or we met the migration scanner
1561 * and the loop terminated due to isolate_start_pfn < low_pfn
1563 cc
->free_pfn
= isolate_start_pfn
;
1566 /* __isolate_free_page() does not map the pages */
1567 split_map_pages(freelist
);
1571 * This is a migrate-callback that "allocates" freepages by taking pages
1572 * from the isolated freelists in the block we are migrating to.
1574 static struct page
*compaction_alloc(struct page
*migratepage
,
1577 struct compact_control
*cc
= (struct compact_control
*)data
;
1578 struct page
*freepage
;
1580 if (list_empty(&cc
->freepages
)) {
1581 isolate_freepages(cc
);
1583 if (list_empty(&cc
->freepages
))
1587 freepage
= list_entry(cc
->freepages
.next
, struct page
, lru
);
1588 list_del(&freepage
->lru
);
1595 * This is a migrate-callback that "frees" freepages back to the isolated
1596 * freelist. All pages on the freelist are from the same zone, so there is no
1597 * special handling needed for NUMA.
1599 static void compaction_free(struct page
*page
, unsigned long data
)
1601 struct compact_control
*cc
= (struct compact_control
*)data
;
1603 list_add(&page
->lru
, &cc
->freepages
);
1607 /* possible outcome of isolate_migratepages */
1609 ISOLATE_ABORT
, /* Abort compaction now */
1610 ISOLATE_NONE
, /* No pages isolated, continue scanning */
1611 ISOLATE_SUCCESS
, /* Pages isolated, migrate */
1612 } isolate_migrate_t
;
1615 * Allow userspace to control policy on scanning the unevictable LRU for
1616 * compactable pages.
1618 #ifdef CONFIG_PREEMPT_RT
1619 int sysctl_compact_unevictable_allowed __read_mostly
= 0;
1621 int sysctl_compact_unevictable_allowed __read_mostly
= 1;
1625 update_fast_start_pfn(struct compact_control
*cc
, unsigned long pfn
)
1627 if (cc
->fast_start_pfn
== ULONG_MAX
)
1630 if (!cc
->fast_start_pfn
)
1631 cc
->fast_start_pfn
= pfn
;
1633 cc
->fast_start_pfn
= min(cc
->fast_start_pfn
, pfn
);
1636 static inline unsigned long
1637 reinit_migrate_pfn(struct compact_control
*cc
)
1639 if (!cc
->fast_start_pfn
|| cc
->fast_start_pfn
== ULONG_MAX
)
1640 return cc
->migrate_pfn
;
1642 cc
->migrate_pfn
= cc
->fast_start_pfn
;
1643 cc
->fast_start_pfn
= ULONG_MAX
;
1645 return cc
->migrate_pfn
;
1649 * Briefly search the free lists for a migration source that already has
1650 * some free pages to reduce the number of pages that need migration
1651 * before a pageblock is free.
1653 static unsigned long fast_find_migrateblock(struct compact_control
*cc
)
1655 unsigned int limit
= freelist_scan_limit(cc
);
1656 unsigned int nr_scanned
= 0;
1657 unsigned long distance
;
1658 unsigned long pfn
= cc
->migrate_pfn
;
1659 unsigned long high_pfn
;
1662 /* Skip hints are relied on to avoid repeats on the fast search */
1663 if (cc
->ignore_skip_hint
)
1667 * If the migrate_pfn is not at the start of a zone or the start
1668 * of a pageblock then assume this is a continuation of a previous
1669 * scan restarted due to COMPACT_CLUSTER_MAX.
1671 if (pfn
!= cc
->zone
->zone_start_pfn
&& pfn
!= pageblock_start_pfn(pfn
))
1675 * For smaller orders, just linearly scan as the number of pages
1676 * to migrate should be relatively small and does not necessarily
1677 * justify freeing up a large block for a small allocation.
1679 if (cc
->order
<= PAGE_ALLOC_COSTLY_ORDER
)
1683 * Only allow kcompactd and direct requests for movable pages to
1684 * quickly clear out a MOVABLE pageblock for allocation. This
1685 * reduces the risk that a large movable pageblock is freed for
1686 * an unmovable/reclaimable small allocation.
1688 if (cc
->direct_compaction
&& cc
->migratetype
!= MIGRATE_MOVABLE
)
1692 * When starting the migration scanner, pick any pageblock within the
1693 * first half of the search space. Otherwise try and pick a pageblock
1694 * within the first eighth to reduce the chances that a migration
1695 * target later becomes a source.
1697 distance
= (cc
->free_pfn
- cc
->migrate_pfn
) >> 1;
1698 if (cc
->migrate_pfn
!= cc
->zone
->zone_start_pfn
)
1700 high_pfn
= pageblock_start_pfn(cc
->migrate_pfn
+ distance
);
1702 for (order
= cc
->order
- 1;
1703 order
>= PAGE_ALLOC_COSTLY_ORDER
&& pfn
== cc
->migrate_pfn
&& nr_scanned
< limit
;
1705 struct free_area
*area
= &cc
->zone
->free_area
[order
];
1706 struct list_head
*freelist
;
1707 unsigned long flags
;
1708 struct page
*freepage
;
1713 spin_lock_irqsave(&cc
->zone
->lock
, flags
);
1714 freelist
= &area
->free_list
[MIGRATE_MOVABLE
];
1715 list_for_each_entry(freepage
, freelist
, lru
) {
1716 unsigned long free_pfn
;
1719 free_pfn
= page_to_pfn(freepage
);
1720 if (free_pfn
< high_pfn
) {
1722 * Avoid if skipped recently. Ideally it would
1723 * move to the tail but even safe iteration of
1724 * the list assumes an entry is deleted, not
1727 if (get_pageblock_skip(freepage
)) {
1728 if (list_is_last(freelist
, &freepage
->lru
))
1734 /* Reorder to so a future search skips recent pages */
1735 move_freelist_tail(freelist
, freepage
);
1737 update_fast_start_pfn(cc
, free_pfn
);
1738 pfn
= pageblock_start_pfn(free_pfn
);
1739 cc
->fast_search_fail
= 0;
1740 set_pageblock_skip(freepage
);
1744 if (nr_scanned
>= limit
) {
1745 cc
->fast_search_fail
++;
1746 move_freelist_tail(freelist
, freepage
);
1750 spin_unlock_irqrestore(&cc
->zone
->lock
, flags
);
1753 cc
->total_migrate_scanned
+= nr_scanned
;
1756 * If fast scanning failed then use a cached entry for a page block
1757 * that had free pages as the basis for starting a linear scan.
1759 if (pfn
== cc
->migrate_pfn
)
1760 pfn
= reinit_migrate_pfn(cc
);
1766 * Isolate all pages that can be migrated from the first suitable block,
1767 * starting at the block pointed to by the migrate scanner pfn within
1770 static isolate_migrate_t
isolate_migratepages(struct compact_control
*cc
)
1772 unsigned long block_start_pfn
;
1773 unsigned long block_end_pfn
;
1774 unsigned long low_pfn
;
1776 const isolate_mode_t isolate_mode
=
1777 (sysctl_compact_unevictable_allowed
? ISOLATE_UNEVICTABLE
: 0) |
1778 (cc
->mode
!= MIGRATE_SYNC
? ISOLATE_ASYNC_MIGRATE
: 0);
1779 bool fast_find_block
;
1782 * Start at where we last stopped, or beginning of the zone as
1783 * initialized by compact_zone(). The first failure will use
1784 * the lowest PFN as the starting point for linear scanning.
1786 low_pfn
= fast_find_migrateblock(cc
);
1787 block_start_pfn
= pageblock_start_pfn(low_pfn
);
1788 if (block_start_pfn
< cc
->zone
->zone_start_pfn
)
1789 block_start_pfn
= cc
->zone
->zone_start_pfn
;
1792 * fast_find_migrateblock marks a pageblock skipped so to avoid
1793 * the isolation_suitable check below, check whether the fast
1794 * search was successful.
1796 fast_find_block
= low_pfn
!= cc
->migrate_pfn
&& !cc
->fast_search_fail
;
1798 /* Only scan within a pageblock boundary */
1799 block_end_pfn
= pageblock_end_pfn(low_pfn
);
1802 * Iterate over whole pageblocks until we find the first suitable.
1803 * Do not cross the free scanner.
1805 for (; block_end_pfn
<= cc
->free_pfn
;
1806 fast_find_block
= false,
1807 low_pfn
= block_end_pfn
,
1808 block_start_pfn
= block_end_pfn
,
1809 block_end_pfn
+= pageblock_nr_pages
) {
1812 * This can potentially iterate a massively long zone with
1813 * many pageblocks unsuitable, so periodically check if we
1816 if (!(low_pfn
% (SWAP_CLUSTER_MAX
* pageblock_nr_pages
)))
1819 page
= pageblock_pfn_to_page(block_start_pfn
,
1820 block_end_pfn
, cc
->zone
);
1825 * If isolation recently failed, do not retry. Only check the
1826 * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock
1827 * to be visited multiple times. Assume skip was checked
1828 * before making it "skip" so other compaction instances do
1829 * not scan the same block.
1831 if (IS_ALIGNED(low_pfn
, pageblock_nr_pages
) &&
1832 !fast_find_block
&& !isolation_suitable(cc
, page
))
1836 * For async compaction, also only scan in MOVABLE blocks
1837 * without huge pages. Async compaction is optimistic to see
1838 * if the minimum amount of work satisfies the allocation.
1839 * The cached PFN is updated as it's possible that all
1840 * remaining blocks between source and target are unsuitable
1841 * and the compaction scanners fail to meet.
1843 if (!suitable_migration_source(cc
, page
)) {
1844 update_cached_migrate(cc
, block_end_pfn
);
1848 /* Perform the isolation */
1849 low_pfn
= isolate_migratepages_block(cc
, low_pfn
,
1850 block_end_pfn
, isolate_mode
);
1853 return ISOLATE_ABORT
;
1856 * Either we isolated something and proceed with migration. Or
1857 * we failed and compact_zone should decide if we should
1863 /* Record where migration scanner will be restarted. */
1864 cc
->migrate_pfn
= low_pfn
;
1866 return cc
->nr_migratepages
? ISOLATE_SUCCESS
: ISOLATE_NONE
;
1870 * order == -1 is expected when compacting via
1871 * /proc/sys/vm/compact_memory
1873 static inline bool is_via_compact_memory(int order
)
1878 static bool kswapd_is_running(pg_data_t
*pgdat
)
1880 return pgdat
->kswapd
&& (pgdat
->kswapd
->state
== TASK_RUNNING
);
1884 * A zone's fragmentation score is the external fragmentation wrt to the
1885 * COMPACTION_HPAGE_ORDER scaled by the zone's size. It returns a value
1886 * in the range [0, 100].
1888 * The scaling factor ensures that proactive compaction focuses on larger
1889 * zones like ZONE_NORMAL, rather than smaller, specialized zones like
1890 * ZONE_DMA32. For smaller zones, the score value remains close to zero,
1891 * and thus never exceeds the high threshold for proactive compaction.
1893 static unsigned int fragmentation_score_zone(struct zone
*zone
)
1895 unsigned long score
;
1897 score
= zone
->present_pages
*
1898 extfrag_for_order(zone
, COMPACTION_HPAGE_ORDER
);
1899 return div64_ul(score
, zone
->zone_pgdat
->node_present_pages
+ 1);
1903 * The per-node proactive (background) compaction process is started by its
1904 * corresponding kcompactd thread when the node's fragmentation score
1905 * exceeds the high threshold. The compaction process remains active till
1906 * the node's score falls below the low threshold, or one of the back-off
1907 * conditions is met.
1909 static unsigned int fragmentation_score_node(pg_data_t
*pgdat
)
1911 unsigned int score
= 0;
1914 for (zoneid
= 0; zoneid
< MAX_NR_ZONES
; zoneid
++) {
1917 zone
= &pgdat
->node_zones
[zoneid
];
1918 score
+= fragmentation_score_zone(zone
);
1924 static unsigned int fragmentation_score_wmark(pg_data_t
*pgdat
, bool low
)
1926 unsigned int wmark_low
;
1929 * Cap the low watermak to avoid excessive compaction
1930 * activity in case a user sets the proactivess tunable
1931 * close to 100 (maximum).
1933 wmark_low
= max(100U - sysctl_compaction_proactiveness
, 5U);
1934 return low
? wmark_low
: min(wmark_low
+ 10, 100U);
1937 static bool should_proactive_compact_node(pg_data_t
*pgdat
)
1941 if (!sysctl_compaction_proactiveness
|| kswapd_is_running(pgdat
))
1944 wmark_high
= fragmentation_score_wmark(pgdat
, false);
1945 return fragmentation_score_node(pgdat
) > wmark_high
;
1948 static enum compact_result
__compact_finished(struct compact_control
*cc
)
1951 const int migratetype
= cc
->migratetype
;
1954 /* Compaction run completes if the migrate and free scanner meet */
1955 if (compact_scanners_met(cc
)) {
1956 /* Let the next compaction start anew. */
1957 reset_cached_positions(cc
->zone
);
1960 * Mark that the PG_migrate_skip information should be cleared
1961 * by kswapd when it goes to sleep. kcompactd does not set the
1962 * flag itself as the decision to be clear should be directly
1963 * based on an allocation request.
1965 if (cc
->direct_compaction
)
1966 cc
->zone
->compact_blockskip_flush
= true;
1969 return COMPACT_COMPLETE
;
1971 return COMPACT_PARTIAL_SKIPPED
;
1974 if (cc
->proactive_compaction
) {
1975 int score
, wmark_low
;
1978 pgdat
= cc
->zone
->zone_pgdat
;
1979 if (kswapd_is_running(pgdat
))
1980 return COMPACT_PARTIAL_SKIPPED
;
1982 score
= fragmentation_score_zone(cc
->zone
);
1983 wmark_low
= fragmentation_score_wmark(pgdat
, true);
1985 if (score
> wmark_low
)
1986 ret
= COMPACT_CONTINUE
;
1988 ret
= COMPACT_SUCCESS
;
1993 if (is_via_compact_memory(cc
->order
))
1994 return COMPACT_CONTINUE
;
1997 * Always finish scanning a pageblock to reduce the possibility of
1998 * fallbacks in the future. This is particularly important when
1999 * migration source is unmovable/reclaimable but it's not worth
2002 if (!IS_ALIGNED(cc
->migrate_pfn
, pageblock_nr_pages
))
2003 return COMPACT_CONTINUE
;
2005 /* Direct compactor: Is a suitable page free? */
2006 ret
= COMPACT_NO_SUITABLE_PAGE
;
2007 for (order
= cc
->order
; order
< MAX_ORDER
; order
++) {
2008 struct free_area
*area
= &cc
->zone
->free_area
[order
];
2011 /* Job done if page is free of the right migratetype */
2012 if (!free_area_empty(area
, migratetype
))
2013 return COMPACT_SUCCESS
;
2016 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
2017 if (migratetype
== MIGRATE_MOVABLE
&&
2018 !free_area_empty(area
, MIGRATE_CMA
))
2019 return COMPACT_SUCCESS
;
2022 * Job done if allocation would steal freepages from
2023 * other migratetype buddy lists.
2025 if (find_suitable_fallback(area
, order
, migratetype
,
2026 true, &can_steal
) != -1) {
2028 /* movable pages are OK in any pageblock */
2029 if (migratetype
== MIGRATE_MOVABLE
)
2030 return COMPACT_SUCCESS
;
2033 * We are stealing for a non-movable allocation. Make
2034 * sure we finish compacting the current pageblock
2035 * first so it is as free as possible and we won't
2036 * have to steal another one soon. This only applies
2037 * to sync compaction, as async compaction operates
2038 * on pageblocks of the same migratetype.
2040 if (cc
->mode
== MIGRATE_ASYNC
||
2041 IS_ALIGNED(cc
->migrate_pfn
,
2042 pageblock_nr_pages
)) {
2043 return COMPACT_SUCCESS
;
2046 ret
= COMPACT_CONTINUE
;
2052 if (cc
->contended
|| fatal_signal_pending(current
))
2053 ret
= COMPACT_CONTENDED
;
2058 static enum compact_result
compact_finished(struct compact_control
*cc
)
2062 ret
= __compact_finished(cc
);
2063 trace_mm_compaction_finished(cc
->zone
, cc
->order
, ret
);
2064 if (ret
== COMPACT_NO_SUITABLE_PAGE
)
2065 ret
= COMPACT_CONTINUE
;
2071 * compaction_suitable: Is this suitable to run compaction on this zone now?
2073 * COMPACT_SKIPPED - If there are too few free pages for compaction
2074 * COMPACT_SUCCESS - If the allocation would succeed without compaction
2075 * COMPACT_CONTINUE - If compaction should run now
2077 static enum compact_result
__compaction_suitable(struct zone
*zone
, int order
,
2078 unsigned int alloc_flags
,
2079 int highest_zoneidx
,
2080 unsigned long wmark_target
)
2082 unsigned long watermark
;
2084 if (is_via_compact_memory(order
))
2085 return COMPACT_CONTINUE
;
2087 watermark
= wmark_pages(zone
, alloc_flags
& ALLOC_WMARK_MASK
);
2089 * If watermarks for high-order allocation are already met, there
2090 * should be no need for compaction at all.
2092 if (zone_watermark_ok(zone
, order
, watermark
, highest_zoneidx
,
2094 return COMPACT_SUCCESS
;
2097 * Watermarks for order-0 must be met for compaction to be able to
2098 * isolate free pages for migration targets. This means that the
2099 * watermark and alloc_flags have to match, or be more pessimistic than
2100 * the check in __isolate_free_page(). We don't use the direct
2101 * compactor's alloc_flags, as they are not relevant for freepage
2102 * isolation. We however do use the direct compactor's highest_zoneidx
2103 * to skip over zones where lowmem reserves would prevent allocation
2104 * even if compaction succeeds.
2105 * For costly orders, we require low watermark instead of min for
2106 * compaction to proceed to increase its chances.
2107 * ALLOC_CMA is used, as pages in CMA pageblocks are considered
2108 * suitable migration targets
2110 watermark
= (order
> PAGE_ALLOC_COSTLY_ORDER
) ?
2111 low_wmark_pages(zone
) : min_wmark_pages(zone
);
2112 watermark
+= compact_gap(order
);
2113 if (!__zone_watermark_ok(zone
, 0, watermark
, highest_zoneidx
,
2114 ALLOC_CMA
, wmark_target
))
2115 return COMPACT_SKIPPED
;
2117 return COMPACT_CONTINUE
;
2120 enum compact_result
compaction_suitable(struct zone
*zone
, int order
,
2121 unsigned int alloc_flags
,
2122 int highest_zoneidx
)
2124 enum compact_result ret
;
2127 ret
= __compaction_suitable(zone
, order
, alloc_flags
, highest_zoneidx
,
2128 zone_page_state(zone
, NR_FREE_PAGES
));
2130 * fragmentation index determines if allocation failures are due to
2131 * low memory or external fragmentation
2133 * index of -1000 would imply allocations might succeed depending on
2134 * watermarks, but we already failed the high-order watermark check
2135 * index towards 0 implies failure is due to lack of memory
2136 * index towards 1000 implies failure is due to fragmentation
2138 * Only compact if a failure would be due to fragmentation. Also
2139 * ignore fragindex for non-costly orders where the alternative to
2140 * a successful reclaim/compaction is OOM. Fragindex and the
2141 * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
2142 * excessive compaction for costly orders, but it should not be at the
2143 * expense of system stability.
2145 if (ret
== COMPACT_CONTINUE
&& (order
> PAGE_ALLOC_COSTLY_ORDER
)) {
2146 fragindex
= fragmentation_index(zone
, order
);
2147 if (fragindex
>= 0 && fragindex
<= sysctl_extfrag_threshold
)
2148 ret
= COMPACT_NOT_SUITABLE_ZONE
;
2151 trace_mm_compaction_suitable(zone
, order
, ret
);
2152 if (ret
== COMPACT_NOT_SUITABLE_ZONE
)
2153 ret
= COMPACT_SKIPPED
;
2158 bool compaction_zonelist_suitable(struct alloc_context
*ac
, int order
,
2165 * Make sure at least one zone would pass __compaction_suitable if we continue
2166 * retrying the reclaim.
2168 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
2169 ac
->highest_zoneidx
, ac
->nodemask
) {
2170 unsigned long available
;
2171 enum compact_result compact_result
;
2174 * Do not consider all the reclaimable memory because we do not
2175 * want to trash just for a single high order allocation which
2176 * is even not guaranteed to appear even if __compaction_suitable
2177 * is happy about the watermark check.
2179 available
= zone_reclaimable_pages(zone
) / order
;
2180 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
2181 compact_result
= __compaction_suitable(zone
, order
, alloc_flags
,
2182 ac
->highest_zoneidx
, available
);
2183 if (compact_result
!= COMPACT_SKIPPED
)
2190 static enum compact_result
2191 compact_zone(struct compact_control
*cc
, struct capture_control
*capc
)
2193 enum compact_result ret
;
2194 unsigned long start_pfn
= cc
->zone
->zone_start_pfn
;
2195 unsigned long end_pfn
= zone_end_pfn(cc
->zone
);
2196 unsigned long last_migrated_pfn
;
2197 const bool sync
= cc
->mode
!= MIGRATE_ASYNC
;
2201 * These counters track activities during zone compaction. Initialize
2202 * them before compacting a new zone.
2204 cc
->total_migrate_scanned
= 0;
2205 cc
->total_free_scanned
= 0;
2206 cc
->nr_migratepages
= 0;
2207 cc
->nr_freepages
= 0;
2208 INIT_LIST_HEAD(&cc
->freepages
);
2209 INIT_LIST_HEAD(&cc
->migratepages
);
2211 cc
->migratetype
= gfp_migratetype(cc
->gfp_mask
);
2212 ret
= compaction_suitable(cc
->zone
, cc
->order
, cc
->alloc_flags
,
2213 cc
->highest_zoneidx
);
2214 /* Compaction is likely to fail */
2215 if (ret
== COMPACT_SUCCESS
|| ret
== COMPACT_SKIPPED
)
2218 /* huh, compaction_suitable is returning something unexpected */
2219 VM_BUG_ON(ret
!= COMPACT_CONTINUE
);
2222 * Clear pageblock skip if there were failures recently and compaction
2223 * is about to be retried after being deferred.
2225 if (compaction_restarting(cc
->zone
, cc
->order
))
2226 __reset_isolation_suitable(cc
->zone
);
2229 * Setup to move all movable pages to the end of the zone. Used cached
2230 * information on where the scanners should start (unless we explicitly
2231 * want to compact the whole zone), but check that it is initialised
2232 * by ensuring the values are within zone boundaries.
2234 cc
->fast_start_pfn
= 0;
2235 if (cc
->whole_zone
) {
2236 cc
->migrate_pfn
= start_pfn
;
2237 cc
->free_pfn
= pageblock_start_pfn(end_pfn
- 1);
2239 cc
->migrate_pfn
= cc
->zone
->compact_cached_migrate_pfn
[sync
];
2240 cc
->free_pfn
= cc
->zone
->compact_cached_free_pfn
;
2241 if (cc
->free_pfn
< start_pfn
|| cc
->free_pfn
>= end_pfn
) {
2242 cc
->free_pfn
= pageblock_start_pfn(end_pfn
- 1);
2243 cc
->zone
->compact_cached_free_pfn
= cc
->free_pfn
;
2245 if (cc
->migrate_pfn
< start_pfn
|| cc
->migrate_pfn
>= end_pfn
) {
2246 cc
->migrate_pfn
= start_pfn
;
2247 cc
->zone
->compact_cached_migrate_pfn
[0] = cc
->migrate_pfn
;
2248 cc
->zone
->compact_cached_migrate_pfn
[1] = cc
->migrate_pfn
;
2251 if (cc
->migrate_pfn
<= cc
->zone
->compact_init_migrate_pfn
)
2252 cc
->whole_zone
= true;
2255 last_migrated_pfn
= 0;
2258 * Migrate has separate cached PFNs for ASYNC and SYNC* migration on
2259 * the basis that some migrations will fail in ASYNC mode. However,
2260 * if the cached PFNs match and pageblocks are skipped due to having
2261 * no isolation candidates, then the sync state does not matter.
2262 * Until a pageblock with isolation candidates is found, keep the
2263 * cached PFNs in sync to avoid revisiting the same blocks.
2265 update_cached
= !sync
&&
2266 cc
->zone
->compact_cached_migrate_pfn
[0] == cc
->zone
->compact_cached_migrate_pfn
[1];
2268 trace_mm_compaction_begin(start_pfn
, cc
->migrate_pfn
,
2269 cc
->free_pfn
, end_pfn
, sync
);
2271 migrate_prep_local();
2273 while ((ret
= compact_finished(cc
)) == COMPACT_CONTINUE
) {
2275 unsigned long start_pfn
= cc
->migrate_pfn
;
2278 * Avoid multiple rescans which can happen if a page cannot be
2279 * isolated (dirty/writeback in async mode) or if the migrated
2280 * pages are being allocated before the pageblock is cleared.
2281 * The first rescan will capture the entire pageblock for
2282 * migration. If it fails, it'll be marked skip and scanning
2283 * will proceed as normal.
2286 if (pageblock_start_pfn(last_migrated_pfn
) ==
2287 pageblock_start_pfn(start_pfn
)) {
2291 switch (isolate_migratepages(cc
)) {
2293 ret
= COMPACT_CONTENDED
;
2294 putback_movable_pages(&cc
->migratepages
);
2295 cc
->nr_migratepages
= 0;
2298 if (update_cached
) {
2299 cc
->zone
->compact_cached_migrate_pfn
[1] =
2300 cc
->zone
->compact_cached_migrate_pfn
[0];
2304 * We haven't isolated and migrated anything, but
2305 * there might still be unflushed migrations from
2306 * previous cc->order aligned block.
2309 case ISOLATE_SUCCESS
:
2310 update_cached
= false;
2311 last_migrated_pfn
= start_pfn
;
2315 err
= migrate_pages(&cc
->migratepages
, compaction_alloc
,
2316 compaction_free
, (unsigned long)cc
, cc
->mode
,
2319 trace_mm_compaction_migratepages(cc
->nr_migratepages
, err
,
2322 /* All pages were either migrated or will be released */
2323 cc
->nr_migratepages
= 0;
2325 putback_movable_pages(&cc
->migratepages
);
2327 * migrate_pages() may return -ENOMEM when scanners meet
2328 * and we want compact_finished() to detect it
2330 if (err
== -ENOMEM
&& !compact_scanners_met(cc
)) {
2331 ret
= COMPACT_CONTENDED
;
2335 * We failed to migrate at least one page in the current
2336 * order-aligned block, so skip the rest of it.
2338 if (cc
->direct_compaction
&&
2339 (cc
->mode
== MIGRATE_ASYNC
)) {
2340 cc
->migrate_pfn
= block_end_pfn(
2341 cc
->migrate_pfn
- 1, cc
->order
);
2342 /* Draining pcplists is useless in this case */
2343 last_migrated_pfn
= 0;
2349 * Has the migration scanner moved away from the previous
2350 * cc->order aligned block where we migrated from? If yes,
2351 * flush the pages that were freed, so that they can merge and
2352 * compact_finished() can detect immediately if allocation
2355 if (cc
->order
> 0 && last_migrated_pfn
) {
2356 unsigned long current_block_start
=
2357 block_start_pfn(cc
->migrate_pfn
, cc
->order
);
2359 if (last_migrated_pfn
< current_block_start
) {
2360 lru_add_drain_cpu_zone(cc
->zone
);
2361 /* No more flushing until we migrate again */
2362 last_migrated_pfn
= 0;
2366 /* Stop if a page has been captured */
2367 if (capc
&& capc
->page
) {
2368 ret
= COMPACT_SUCCESS
;
2375 * Release free pages and update where the free scanner should restart,
2376 * so we don't leave any returned pages behind in the next attempt.
2378 if (cc
->nr_freepages
> 0) {
2379 unsigned long free_pfn
= release_freepages(&cc
->freepages
);
2381 cc
->nr_freepages
= 0;
2382 VM_BUG_ON(free_pfn
== 0);
2383 /* The cached pfn is always the first in a pageblock */
2384 free_pfn
= pageblock_start_pfn(free_pfn
);
2386 * Only go back, not forward. The cached pfn might have been
2387 * already reset to zone end in compact_finished()
2389 if (free_pfn
> cc
->zone
->compact_cached_free_pfn
)
2390 cc
->zone
->compact_cached_free_pfn
= free_pfn
;
2393 count_compact_events(COMPACTMIGRATE_SCANNED
, cc
->total_migrate_scanned
);
2394 count_compact_events(COMPACTFREE_SCANNED
, cc
->total_free_scanned
);
2396 trace_mm_compaction_end(start_pfn
, cc
->migrate_pfn
,
2397 cc
->free_pfn
, end_pfn
, sync
, ret
);
2402 static enum compact_result
compact_zone_order(struct zone
*zone
, int order
,
2403 gfp_t gfp_mask
, enum compact_priority prio
,
2404 unsigned int alloc_flags
, int highest_zoneidx
,
2405 struct page
**capture
)
2407 enum compact_result ret
;
2408 struct compact_control cc
= {
2410 .search_order
= order
,
2411 .gfp_mask
= gfp_mask
,
2413 .mode
= (prio
== COMPACT_PRIO_ASYNC
) ?
2414 MIGRATE_ASYNC
: MIGRATE_SYNC_LIGHT
,
2415 .alloc_flags
= alloc_flags
,
2416 .highest_zoneidx
= highest_zoneidx
,
2417 .direct_compaction
= true,
2418 .whole_zone
= (prio
== MIN_COMPACT_PRIORITY
),
2419 .ignore_skip_hint
= (prio
== MIN_COMPACT_PRIORITY
),
2420 .ignore_block_suitable
= (prio
== MIN_COMPACT_PRIORITY
)
2422 struct capture_control capc
= {
2428 * Make sure the structs are really initialized before we expose the
2429 * capture control, in case we are interrupted and the interrupt handler
2433 WRITE_ONCE(current
->capture_control
, &capc
);
2435 ret
= compact_zone(&cc
, &capc
);
2437 VM_BUG_ON(!list_empty(&cc
.freepages
));
2438 VM_BUG_ON(!list_empty(&cc
.migratepages
));
2441 * Make sure we hide capture control first before we read the captured
2442 * page pointer, otherwise an interrupt could free and capture a page
2443 * and we would leak it.
2445 WRITE_ONCE(current
->capture_control
, NULL
);
2446 *capture
= READ_ONCE(capc
.page
);
2451 int sysctl_extfrag_threshold
= 500;
2454 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
2455 * @gfp_mask: The GFP mask of the current allocation
2456 * @order: The order of the current allocation
2457 * @alloc_flags: The allocation flags of the current allocation
2458 * @ac: The context of current allocation
2459 * @prio: Determines how hard direct compaction should try to succeed
2460 * @capture: Pointer to free page created by compaction will be stored here
2462 * This is the main entry point for direct page compaction.
2464 enum compact_result
try_to_compact_pages(gfp_t gfp_mask
, unsigned int order
,
2465 unsigned int alloc_flags
, const struct alloc_context
*ac
,
2466 enum compact_priority prio
, struct page
**capture
)
2468 int may_perform_io
= gfp_mask
& __GFP_IO
;
2471 enum compact_result rc
= COMPACT_SKIPPED
;
2474 * Check if the GFP flags allow compaction - GFP_NOIO is really
2475 * tricky context because the migration might require IO
2477 if (!may_perform_io
)
2478 return COMPACT_SKIPPED
;
2480 trace_mm_compaction_try_to_compact_pages(order
, gfp_mask
, prio
);
2482 /* Compact each zone in the list */
2483 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
2484 ac
->highest_zoneidx
, ac
->nodemask
) {
2485 enum compact_result status
;
2487 if (prio
> MIN_COMPACT_PRIORITY
2488 && compaction_deferred(zone
, order
)) {
2489 rc
= max_t(enum compact_result
, COMPACT_DEFERRED
, rc
);
2493 status
= compact_zone_order(zone
, order
, gfp_mask
, prio
,
2494 alloc_flags
, ac
->highest_zoneidx
, capture
);
2495 rc
= max(status
, rc
);
2497 /* The allocation should succeed, stop compacting */
2498 if (status
== COMPACT_SUCCESS
) {
2500 * We think the allocation will succeed in this zone,
2501 * but it is not certain, hence the false. The caller
2502 * will repeat this with true if allocation indeed
2503 * succeeds in this zone.
2505 compaction_defer_reset(zone
, order
, false);
2510 if (prio
!= COMPACT_PRIO_ASYNC
&& (status
== COMPACT_COMPLETE
||
2511 status
== COMPACT_PARTIAL_SKIPPED
))
2513 * We think that allocation won't succeed in this zone
2514 * so we defer compaction there. If it ends up
2515 * succeeding after all, it will be reset.
2517 defer_compaction(zone
, order
);
2520 * We might have stopped compacting due to need_resched() in
2521 * async compaction, or due to a fatal signal detected. In that
2522 * case do not try further zones
2524 if ((prio
== COMPACT_PRIO_ASYNC
&& need_resched())
2525 || fatal_signal_pending(current
))
2533 * Compact all zones within a node till each zone's fragmentation score
2534 * reaches within proactive compaction thresholds (as determined by the
2535 * proactiveness tunable).
2537 * It is possible that the function returns before reaching score targets
2538 * due to various back-off conditions, such as, contention on per-node or
2541 static void proactive_compact_node(pg_data_t
*pgdat
)
2545 struct compact_control cc
= {
2547 .mode
= MIGRATE_SYNC_LIGHT
,
2548 .ignore_skip_hint
= true,
2550 .gfp_mask
= GFP_KERNEL
,
2551 .proactive_compaction
= true,
2554 for (zoneid
= 0; zoneid
< MAX_NR_ZONES
; zoneid
++) {
2555 zone
= &pgdat
->node_zones
[zoneid
];
2556 if (!populated_zone(zone
))
2561 compact_zone(&cc
, NULL
);
2563 VM_BUG_ON(!list_empty(&cc
.freepages
));
2564 VM_BUG_ON(!list_empty(&cc
.migratepages
));
2568 /* Compact all zones within a node */
2569 static void compact_node(int nid
)
2571 pg_data_t
*pgdat
= NODE_DATA(nid
);
2574 struct compact_control cc
= {
2576 .mode
= MIGRATE_SYNC
,
2577 .ignore_skip_hint
= true,
2579 .gfp_mask
= GFP_KERNEL
,
2583 for (zoneid
= 0; zoneid
< MAX_NR_ZONES
; zoneid
++) {
2585 zone
= &pgdat
->node_zones
[zoneid
];
2586 if (!populated_zone(zone
))
2591 compact_zone(&cc
, NULL
);
2593 VM_BUG_ON(!list_empty(&cc
.freepages
));
2594 VM_BUG_ON(!list_empty(&cc
.migratepages
));
2598 /* Compact all nodes in the system */
2599 static void compact_nodes(void)
2603 /* Flush pending updates to the LRU lists */
2604 lru_add_drain_all();
2606 for_each_online_node(nid
)
2610 /* The written value is actually unused, all memory is compacted */
2611 int sysctl_compact_memory
;
2614 * Tunable for proactive compaction. It determines how
2615 * aggressively the kernel should compact memory in the
2616 * background. It takes values in the range [0, 100].
2618 unsigned int __read_mostly sysctl_compaction_proactiveness
= 20;
2621 * This is the entry point for compacting all nodes via
2622 * /proc/sys/vm/compact_memory
2624 int sysctl_compaction_handler(struct ctl_table
*table
, int write
,
2625 void *buffer
, size_t *length
, loff_t
*ppos
)
2633 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
2634 static ssize_t
sysfs_compact_node(struct device
*dev
,
2635 struct device_attribute
*attr
,
2636 const char *buf
, size_t count
)
2640 if (nid
>= 0 && nid
< nr_node_ids
&& node_online(nid
)) {
2641 /* Flush pending updates to the LRU lists */
2642 lru_add_drain_all();
2649 static DEVICE_ATTR(compact
, 0200, NULL
, sysfs_compact_node
);
2651 int compaction_register_node(struct node
*node
)
2653 return device_create_file(&node
->dev
, &dev_attr_compact
);
2656 void compaction_unregister_node(struct node
*node
)
2658 return device_remove_file(&node
->dev
, &dev_attr_compact
);
2660 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
2662 static inline bool kcompactd_work_requested(pg_data_t
*pgdat
)
2664 return pgdat
->kcompactd_max_order
> 0 || kthread_should_stop();
2667 static bool kcompactd_node_suitable(pg_data_t
*pgdat
)
2671 enum zone_type highest_zoneidx
= pgdat
->kcompactd_highest_zoneidx
;
2673 for (zoneid
= 0; zoneid
<= highest_zoneidx
; zoneid
++) {
2674 zone
= &pgdat
->node_zones
[zoneid
];
2676 if (!populated_zone(zone
))
2679 if (compaction_suitable(zone
, pgdat
->kcompactd_max_order
, 0,
2680 highest_zoneidx
) == COMPACT_CONTINUE
)
2687 static void kcompactd_do_work(pg_data_t
*pgdat
)
2690 * With no special task, compact all zones so that a page of requested
2691 * order is allocatable.
2695 struct compact_control cc
= {
2696 .order
= pgdat
->kcompactd_max_order
,
2697 .search_order
= pgdat
->kcompactd_max_order
,
2698 .highest_zoneidx
= pgdat
->kcompactd_highest_zoneidx
,
2699 .mode
= MIGRATE_SYNC_LIGHT
,
2700 .ignore_skip_hint
= false,
2701 .gfp_mask
= GFP_KERNEL
,
2703 trace_mm_compaction_kcompactd_wake(pgdat
->node_id
, cc
.order
,
2704 cc
.highest_zoneidx
);
2705 count_compact_event(KCOMPACTD_WAKE
);
2707 for (zoneid
= 0; zoneid
<= cc
.highest_zoneidx
; zoneid
++) {
2710 zone
= &pgdat
->node_zones
[zoneid
];
2711 if (!populated_zone(zone
))
2714 if (compaction_deferred(zone
, cc
.order
))
2717 if (compaction_suitable(zone
, cc
.order
, 0, zoneid
) !=
2721 if (kthread_should_stop())
2725 status
= compact_zone(&cc
, NULL
);
2727 if (status
== COMPACT_SUCCESS
) {
2728 compaction_defer_reset(zone
, cc
.order
, false);
2729 } else if (status
== COMPACT_PARTIAL_SKIPPED
|| status
== COMPACT_COMPLETE
) {
2731 * Buddy pages may become stranded on pcps that could
2732 * otherwise coalesce on the zone's free area for
2733 * order >= cc.order. This is ratelimited by the
2734 * upcoming deferral.
2736 drain_all_pages(zone
);
2739 * We use sync migration mode here, so we defer like
2740 * sync direct compaction does.
2742 defer_compaction(zone
, cc
.order
);
2745 count_compact_events(KCOMPACTD_MIGRATE_SCANNED
,
2746 cc
.total_migrate_scanned
);
2747 count_compact_events(KCOMPACTD_FREE_SCANNED
,
2748 cc
.total_free_scanned
);
2750 VM_BUG_ON(!list_empty(&cc
.freepages
));
2751 VM_BUG_ON(!list_empty(&cc
.migratepages
));
2755 * Regardless of success, we are done until woken up next. But remember
2756 * the requested order/highest_zoneidx in case it was higher/tighter
2757 * than our current ones
2759 if (pgdat
->kcompactd_max_order
<= cc
.order
)
2760 pgdat
->kcompactd_max_order
= 0;
2761 if (pgdat
->kcompactd_highest_zoneidx
>= cc
.highest_zoneidx
)
2762 pgdat
->kcompactd_highest_zoneidx
= pgdat
->nr_zones
- 1;
2765 void wakeup_kcompactd(pg_data_t
*pgdat
, int order
, int highest_zoneidx
)
2770 if (pgdat
->kcompactd_max_order
< order
)
2771 pgdat
->kcompactd_max_order
= order
;
2773 if (pgdat
->kcompactd_highest_zoneidx
> highest_zoneidx
)
2774 pgdat
->kcompactd_highest_zoneidx
= highest_zoneidx
;
2777 * Pairs with implicit barrier in wait_event_freezable()
2778 * such that wakeups are not missed.
2780 if (!wq_has_sleeper(&pgdat
->kcompactd_wait
))
2783 if (!kcompactd_node_suitable(pgdat
))
2786 trace_mm_compaction_wakeup_kcompactd(pgdat
->node_id
, order
,
2788 wake_up_interruptible(&pgdat
->kcompactd_wait
);
2792 * The background compaction daemon, started as a kernel thread
2793 * from the init process.
2795 static int kcompactd(void *p
)
2797 pg_data_t
*pgdat
= (pg_data_t
*)p
;
2798 struct task_struct
*tsk
= current
;
2799 unsigned int proactive_defer
= 0;
2801 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
2803 if (!cpumask_empty(cpumask
))
2804 set_cpus_allowed_ptr(tsk
, cpumask
);
2808 pgdat
->kcompactd_max_order
= 0;
2809 pgdat
->kcompactd_highest_zoneidx
= pgdat
->nr_zones
- 1;
2811 while (!kthread_should_stop()) {
2812 unsigned long pflags
;
2814 trace_mm_compaction_kcompactd_sleep(pgdat
->node_id
);
2815 if (wait_event_freezable_timeout(pgdat
->kcompactd_wait
,
2816 kcompactd_work_requested(pgdat
),
2817 msecs_to_jiffies(HPAGE_FRAG_CHECK_INTERVAL_MSEC
))) {
2819 psi_memstall_enter(&pflags
);
2820 kcompactd_do_work(pgdat
);
2821 psi_memstall_leave(&pflags
);
2825 /* kcompactd wait timeout */
2826 if (should_proactive_compact_node(pgdat
)) {
2827 unsigned int prev_score
, score
;
2829 if (proactive_defer
) {
2833 prev_score
= fragmentation_score_node(pgdat
);
2834 proactive_compact_node(pgdat
);
2835 score
= fragmentation_score_node(pgdat
);
2837 * Defer proactive compaction if the fragmentation
2838 * score did not go down i.e. no progress made.
2840 proactive_defer
= score
< prev_score
?
2841 0 : 1 << COMPACT_MAX_DEFER_SHIFT
;
2849 * This kcompactd start function will be called by init and node-hot-add.
2850 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
2852 int kcompactd_run(int nid
)
2854 pg_data_t
*pgdat
= NODE_DATA(nid
);
2857 if (pgdat
->kcompactd
)
2860 pgdat
->kcompactd
= kthread_run(kcompactd
, pgdat
, "kcompactd%d", nid
);
2861 if (IS_ERR(pgdat
->kcompactd
)) {
2862 pr_err("Failed to start kcompactd on node %d\n", nid
);
2863 ret
= PTR_ERR(pgdat
->kcompactd
);
2864 pgdat
->kcompactd
= NULL
;
2870 * Called by memory hotplug when all memory in a node is offlined. Caller must
2871 * hold mem_hotplug_begin/end().
2873 void kcompactd_stop(int nid
)
2875 struct task_struct
*kcompactd
= NODE_DATA(nid
)->kcompactd
;
2878 kthread_stop(kcompactd
);
2879 NODE_DATA(nid
)->kcompactd
= NULL
;
2884 * It's optimal to keep kcompactd on the same CPUs as their memory, but
2885 * not required for correctness. So if the last cpu in a node goes
2886 * away, we get changed to run anywhere: as the first one comes back,
2887 * restore their cpu bindings.
2889 static int kcompactd_cpu_online(unsigned int cpu
)
2893 for_each_node_state(nid
, N_MEMORY
) {
2894 pg_data_t
*pgdat
= NODE_DATA(nid
);
2895 const struct cpumask
*mask
;
2897 mask
= cpumask_of_node(pgdat
->node_id
);
2899 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
2900 /* One of our CPUs online: restore mask */
2901 set_cpus_allowed_ptr(pgdat
->kcompactd
, mask
);
2906 static int __init
kcompactd_init(void)
2911 ret
= cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN
,
2912 "mm/compaction:online",
2913 kcompactd_cpu_online
, NULL
);
2915 pr_err("kcompactd: failed to register hotplug callbacks.\n");
2919 for_each_node_state(nid
, N_MEMORY
)
2923 subsys_initcall(kcompactd_init
)
2925 #endif /* CONFIG_COMPACTION */