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
30 * Fragmentation score check interval for proactive compaction purposes.
32 #define HPAGE_FRAG_CHECK_INTERVAL_MSEC (500)
34 static inline void count_compact_event(enum vm_event_item item
)
39 static inline void count_compact_events(enum vm_event_item item
, long delta
)
41 count_vm_events(item
, delta
);
44 #define count_compact_event(item) do { } while (0)
45 #define count_compact_events(item, delta) do { } while (0)
48 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
50 #define CREATE_TRACE_POINTS
51 #include <trace/events/compaction.h>
53 #define block_start_pfn(pfn, order) round_down(pfn, 1UL << (order))
54 #define block_end_pfn(pfn, order) ALIGN((pfn) + 1, 1UL << (order))
57 * Page order with-respect-to which proactive compaction
58 * calculates external fragmentation, which is used as
59 * the "fragmentation score" of a node/zone.
61 #if defined CONFIG_TRANSPARENT_HUGEPAGE
62 #define COMPACTION_HPAGE_ORDER HPAGE_PMD_ORDER
63 #elif defined CONFIG_HUGETLBFS
64 #define COMPACTION_HPAGE_ORDER HUGETLB_PAGE_ORDER
66 #define COMPACTION_HPAGE_ORDER (PMD_SHIFT - PAGE_SHIFT)
69 static unsigned long release_freepages(struct list_head
*freelist
)
71 struct page
*page
, *next
;
72 unsigned long high_pfn
= 0;
74 list_for_each_entry_safe(page
, next
, freelist
, lru
) {
75 unsigned long pfn
= page_to_pfn(page
);
85 static void split_map_pages(struct list_head
*list
)
87 unsigned int i
, order
, nr_pages
;
88 struct page
*page
, *next
;
91 list_for_each_entry_safe(page
, next
, list
, lru
) {
94 order
= page_private(page
);
95 nr_pages
= 1 << order
;
97 post_alloc_hook(page
, order
, __GFP_MOVABLE
);
99 split_page(page
, order
);
101 for (i
= 0; i
< nr_pages
; i
++) {
102 list_add(&page
->lru
, &tmp_list
);
107 list_splice(&tmp_list
, list
);
110 #ifdef CONFIG_COMPACTION
111 bool PageMovable(struct page
*page
)
113 const struct movable_operations
*mops
;
115 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
116 if (!__PageMovable(page
))
119 mops
= page_movable_ops(page
);
126 void __SetPageMovable(struct page
*page
, const struct movable_operations
*mops
)
128 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
129 VM_BUG_ON_PAGE((unsigned long)mops
& PAGE_MAPPING_MOVABLE
, page
);
130 page
->mapping
= (void *)((unsigned long)mops
| PAGE_MAPPING_MOVABLE
);
132 EXPORT_SYMBOL(__SetPageMovable
);
134 void __ClearPageMovable(struct page
*page
)
136 VM_BUG_ON_PAGE(!PageMovable(page
), page
);
138 * This page still has the type of a movable page, but it's
139 * actually not movable any more.
141 page
->mapping
= (void *)PAGE_MAPPING_MOVABLE
;
143 EXPORT_SYMBOL(__ClearPageMovable
);
145 /* Do not skip compaction more than 64 times */
146 #define COMPACT_MAX_DEFER_SHIFT 6
149 * Compaction is deferred when compaction fails to result in a page
150 * allocation success. 1 << compact_defer_shift, compactions are skipped up
151 * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
153 static void defer_compaction(struct zone
*zone
, int order
)
155 zone
->compact_considered
= 0;
156 zone
->compact_defer_shift
++;
158 if (order
< zone
->compact_order_failed
)
159 zone
->compact_order_failed
= order
;
161 if (zone
->compact_defer_shift
> COMPACT_MAX_DEFER_SHIFT
)
162 zone
->compact_defer_shift
= COMPACT_MAX_DEFER_SHIFT
;
164 trace_mm_compaction_defer_compaction(zone
, order
);
167 /* Returns true if compaction should be skipped this time */
168 static bool compaction_deferred(struct zone
*zone
, int order
)
170 unsigned long defer_limit
= 1UL << zone
->compact_defer_shift
;
172 if (order
< zone
->compact_order_failed
)
175 /* Avoid possible overflow */
176 if (++zone
->compact_considered
>= defer_limit
) {
177 zone
->compact_considered
= defer_limit
;
181 trace_mm_compaction_deferred(zone
, order
);
187 * Update defer tracking counters after successful compaction of given order,
188 * which means an allocation either succeeded (alloc_success == true) or is
189 * expected to succeed.
191 void compaction_defer_reset(struct zone
*zone
, int order
,
195 zone
->compact_considered
= 0;
196 zone
->compact_defer_shift
= 0;
198 if (order
>= zone
->compact_order_failed
)
199 zone
->compact_order_failed
= order
+ 1;
201 trace_mm_compaction_defer_reset(zone
, order
);
204 /* Returns true if restarting compaction after many failures */
205 static bool compaction_restarting(struct zone
*zone
, int order
)
207 if (order
< zone
->compact_order_failed
)
210 return zone
->compact_defer_shift
== COMPACT_MAX_DEFER_SHIFT
&&
211 zone
->compact_considered
>= 1UL << zone
->compact_defer_shift
;
214 /* Returns true if the pageblock should be scanned for pages to isolate. */
215 static inline bool isolation_suitable(struct compact_control
*cc
,
218 if (cc
->ignore_skip_hint
)
221 return !get_pageblock_skip(page
);
224 static void reset_cached_positions(struct zone
*zone
)
226 zone
->compact_cached_migrate_pfn
[0] = zone
->zone_start_pfn
;
227 zone
->compact_cached_migrate_pfn
[1] = zone
->zone_start_pfn
;
228 zone
->compact_cached_free_pfn
=
229 pageblock_start_pfn(zone_end_pfn(zone
) - 1);
233 * Compound pages of >= pageblock_order should consistently be skipped until
234 * released. It is always pointless to compact pages of such order (if they are
235 * migratable), and the pageblocks they occupy cannot contain any free pages.
237 static bool pageblock_skip_persistent(struct page
*page
)
239 if (!PageCompound(page
))
242 page
= compound_head(page
);
244 if (compound_order(page
) >= pageblock_order
)
251 __reset_isolation_pfn(struct zone
*zone
, unsigned long pfn
, bool check_source
,
254 struct page
*page
= pfn_to_online_page(pfn
);
255 struct page
*block_page
;
256 struct page
*end_page
;
257 unsigned long block_pfn
;
261 if (zone
!= page_zone(page
))
263 if (pageblock_skip_persistent(page
))
267 * If skip is already cleared do no further checking once the
268 * restart points have been set.
270 if (check_source
&& check_target
&& !get_pageblock_skip(page
))
274 * If clearing skip for the target scanner, do not select a
275 * non-movable pageblock as the starting point.
277 if (!check_source
&& check_target
&&
278 get_pageblock_migratetype(page
) != MIGRATE_MOVABLE
)
281 /* Ensure the start of the pageblock or zone is online and valid */
282 block_pfn
= pageblock_start_pfn(pfn
);
283 block_pfn
= max(block_pfn
, zone
->zone_start_pfn
);
284 block_page
= pfn_to_online_page(block_pfn
);
290 /* Ensure the end of the pageblock or zone is online and valid */
291 block_pfn
= pageblock_end_pfn(pfn
) - 1;
292 block_pfn
= min(block_pfn
, zone_end_pfn(zone
) - 1);
293 end_page
= pfn_to_online_page(block_pfn
);
298 * Only clear the hint if a sample indicates there is either a
299 * free page or an LRU page in the block. One or other condition
300 * is necessary for the block to be a migration source/target.
303 if (check_source
&& PageLRU(page
)) {
304 clear_pageblock_skip(page
);
308 if (check_target
&& PageBuddy(page
)) {
309 clear_pageblock_skip(page
);
313 page
+= (1 << PAGE_ALLOC_COSTLY_ORDER
);
314 } while (page
<= end_page
);
320 * This function is called to clear all cached information on pageblocks that
321 * should be skipped for page isolation when the migrate and free page scanner
324 static void __reset_isolation_suitable(struct zone
*zone
)
326 unsigned long migrate_pfn
= zone
->zone_start_pfn
;
327 unsigned long free_pfn
= zone_end_pfn(zone
) - 1;
328 unsigned long reset_migrate
= free_pfn
;
329 unsigned long reset_free
= migrate_pfn
;
330 bool source_set
= false;
331 bool free_set
= false;
333 if (!zone
->compact_blockskip_flush
)
336 zone
->compact_blockskip_flush
= false;
339 * Walk the zone and update pageblock skip information. Source looks
340 * for PageLRU while target looks for PageBuddy. When the scanner
341 * is found, both PageBuddy and PageLRU are checked as the pageblock
342 * is suitable as both source and target.
344 for (; migrate_pfn
< free_pfn
; migrate_pfn
+= pageblock_nr_pages
,
345 free_pfn
-= pageblock_nr_pages
) {
348 /* Update the migrate PFN */
349 if (__reset_isolation_pfn(zone
, migrate_pfn
, true, source_set
) &&
350 migrate_pfn
< reset_migrate
) {
352 reset_migrate
= migrate_pfn
;
353 zone
->compact_init_migrate_pfn
= reset_migrate
;
354 zone
->compact_cached_migrate_pfn
[0] = reset_migrate
;
355 zone
->compact_cached_migrate_pfn
[1] = reset_migrate
;
358 /* Update the free PFN */
359 if (__reset_isolation_pfn(zone
, free_pfn
, free_set
, true) &&
360 free_pfn
> reset_free
) {
362 reset_free
= free_pfn
;
363 zone
->compact_init_free_pfn
= reset_free
;
364 zone
->compact_cached_free_pfn
= reset_free
;
368 /* Leave no distance if no suitable block was reset */
369 if (reset_migrate
>= reset_free
) {
370 zone
->compact_cached_migrate_pfn
[0] = migrate_pfn
;
371 zone
->compact_cached_migrate_pfn
[1] = migrate_pfn
;
372 zone
->compact_cached_free_pfn
= free_pfn
;
376 void reset_isolation_suitable(pg_data_t
*pgdat
)
380 for (zoneid
= 0; zoneid
< MAX_NR_ZONES
; zoneid
++) {
381 struct zone
*zone
= &pgdat
->node_zones
[zoneid
];
382 if (!populated_zone(zone
))
385 /* Only flush if a full compaction finished recently */
386 if (zone
->compact_blockskip_flush
)
387 __reset_isolation_suitable(zone
);
392 * Sets the pageblock skip bit if it was clear. Note that this is a hint as
393 * locks are not required for read/writers. Returns true if it was already set.
395 static bool test_and_set_skip(struct compact_control
*cc
, struct page
*page
,
400 /* Do no update if skip hint is being ignored */
401 if (cc
->ignore_skip_hint
)
404 if (!pageblock_aligned(pfn
))
407 skip
= get_pageblock_skip(page
);
408 if (!skip
&& !cc
->no_set_skip_hint
)
409 set_pageblock_skip(page
);
414 static void update_cached_migrate(struct compact_control
*cc
, unsigned long pfn
)
416 struct zone
*zone
= cc
->zone
;
418 pfn
= pageblock_end_pfn(pfn
);
420 /* Set for isolation rather than compaction */
421 if (cc
->no_set_skip_hint
)
424 if (pfn
> zone
->compact_cached_migrate_pfn
[0])
425 zone
->compact_cached_migrate_pfn
[0] = pfn
;
426 if (cc
->mode
!= MIGRATE_ASYNC
&&
427 pfn
> zone
->compact_cached_migrate_pfn
[1])
428 zone
->compact_cached_migrate_pfn
[1] = pfn
;
432 * If no pages were isolated then mark this pageblock to be skipped in the
433 * future. The information is later cleared by __reset_isolation_suitable().
435 static void update_pageblock_skip(struct compact_control
*cc
,
436 struct page
*page
, unsigned long pfn
)
438 struct zone
*zone
= cc
->zone
;
440 if (cc
->no_set_skip_hint
)
446 set_pageblock_skip(page
);
448 /* Update where async and sync compaction should restart */
449 if (pfn
< zone
->compact_cached_free_pfn
)
450 zone
->compact_cached_free_pfn
= pfn
;
453 static inline bool isolation_suitable(struct compact_control
*cc
,
459 static inline bool pageblock_skip_persistent(struct page
*page
)
464 static inline void update_pageblock_skip(struct compact_control
*cc
,
465 struct page
*page
, unsigned long pfn
)
469 static void update_cached_migrate(struct compact_control
*cc
, unsigned long pfn
)
473 static bool test_and_set_skip(struct compact_control
*cc
, struct page
*page
,
478 #endif /* CONFIG_COMPACTION */
481 * Compaction requires the taking of some coarse locks that are potentially
482 * very heavily contended. For async compaction, trylock and record if the
483 * lock is contended. The lock will still be acquired but compaction will
484 * abort when the current block is finished regardless of success rate.
485 * Sync compaction acquires the lock.
487 * Always returns true which makes it easier to track lock state in callers.
489 static bool compact_lock_irqsave(spinlock_t
*lock
, unsigned long *flags
,
490 struct compact_control
*cc
)
493 /* Track if the lock is contended in async mode */
494 if (cc
->mode
== MIGRATE_ASYNC
&& !cc
->contended
) {
495 if (spin_trylock_irqsave(lock
, *flags
))
498 cc
->contended
= true;
501 spin_lock_irqsave(lock
, *flags
);
506 * Compaction requires the taking of some coarse locks that are potentially
507 * very heavily contended. The lock should be periodically unlocked to avoid
508 * having disabled IRQs for a long time, even when there is nobody waiting on
509 * the lock. It might also be that allowing the IRQs will result in
510 * need_resched() becoming true. If scheduling is needed, compaction schedules.
511 * Either compaction type will also abort if a fatal signal is pending.
512 * In either case if the lock was locked, it is dropped and not regained.
514 * Returns true if compaction should abort due to fatal signal pending.
515 * Returns false when compaction can continue.
517 static bool compact_unlock_should_abort(spinlock_t
*lock
,
518 unsigned long flags
, bool *locked
, struct compact_control
*cc
)
521 spin_unlock_irqrestore(lock
, flags
);
525 if (fatal_signal_pending(current
)) {
526 cc
->contended
= true;
536 * Isolate free pages onto a private freelist. If @strict is true, will abort
537 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
538 * (even though it may still end up isolating some pages).
540 static unsigned long isolate_freepages_block(struct compact_control
*cc
,
541 unsigned long *start_pfn
,
542 unsigned long end_pfn
,
543 struct list_head
*freelist
,
547 int nr_scanned
= 0, total_isolated
= 0;
549 unsigned long flags
= 0;
551 unsigned long blockpfn
= *start_pfn
;
554 /* Strict mode is for isolation, speed is secondary */
558 cursor
= pfn_to_page(blockpfn
);
560 /* Isolate free pages. */
561 for (; blockpfn
< end_pfn
; blockpfn
+= stride
, cursor
+= stride
) {
563 struct page
*page
= cursor
;
566 * Periodically drop the lock (if held) regardless of its
567 * contention, to give chance to IRQs. Abort if fatal signal
570 if (!(blockpfn
% COMPACT_CLUSTER_MAX
)
571 && compact_unlock_should_abort(&cc
->zone
->lock
, flags
,
578 * For compound pages such as THP and hugetlbfs, we can save
579 * potentially a lot of iterations if we skip them at once.
580 * The check is racy, but we can consider only valid values
581 * and the only danger is skipping too much.
583 if (PageCompound(page
)) {
584 const unsigned int order
= compound_order(page
);
586 if (likely(order
<= MAX_ORDER
)) {
587 blockpfn
+= (1UL << order
) - 1;
588 cursor
+= (1UL << order
) - 1;
589 nr_scanned
+= (1UL << order
) - 1;
594 if (!PageBuddy(page
))
597 /* If we already hold the lock, we can skip some rechecking. */
599 locked
= compact_lock_irqsave(&cc
->zone
->lock
,
602 /* Recheck this is a buddy page under lock */
603 if (!PageBuddy(page
))
607 /* Found a free page, will break it into order-0 pages */
608 order
= buddy_order(page
);
609 isolated
= __isolate_free_page(page
, order
);
612 set_page_private(page
, order
);
614 nr_scanned
+= isolated
- 1;
615 total_isolated
+= isolated
;
616 cc
->nr_freepages
+= isolated
;
617 list_add_tail(&page
->lru
, freelist
);
619 if (!strict
&& cc
->nr_migratepages
<= cc
->nr_freepages
) {
620 blockpfn
+= isolated
;
623 /* Advance to the end of split page */
624 blockpfn
+= isolated
- 1;
625 cursor
+= isolated
- 1;
637 spin_unlock_irqrestore(&cc
->zone
->lock
, flags
);
640 * There is a tiny chance that we have read bogus compound_order(),
641 * so be careful to not go outside of the pageblock.
643 if (unlikely(blockpfn
> end_pfn
))
646 trace_mm_compaction_isolate_freepages(*start_pfn
, blockpfn
,
647 nr_scanned
, total_isolated
);
649 /* Record how far we have got within the block */
650 *start_pfn
= blockpfn
;
653 * If strict isolation is requested by CMA then check that all the
654 * pages requested were isolated. If there were any failures, 0 is
655 * returned and CMA will fail.
657 if (strict
&& blockpfn
< end_pfn
)
660 cc
->total_free_scanned
+= nr_scanned
;
662 count_compact_events(COMPACTISOLATED
, total_isolated
);
663 return total_isolated
;
667 * isolate_freepages_range() - isolate free pages.
668 * @cc: Compaction control structure.
669 * @start_pfn: The first PFN to start isolating.
670 * @end_pfn: The one-past-last PFN.
672 * Non-free pages, invalid PFNs, or zone boundaries within the
673 * [start_pfn, end_pfn) range are considered errors, cause function to
674 * undo its actions and return zero.
676 * Otherwise, function returns one-past-the-last PFN of isolated page
677 * (which may be greater then end_pfn if end fell in a middle of
681 isolate_freepages_range(struct compact_control
*cc
,
682 unsigned long start_pfn
, unsigned long end_pfn
)
684 unsigned long isolated
, pfn
, block_start_pfn
, block_end_pfn
;
688 block_start_pfn
= pageblock_start_pfn(pfn
);
689 if (block_start_pfn
< cc
->zone
->zone_start_pfn
)
690 block_start_pfn
= cc
->zone
->zone_start_pfn
;
691 block_end_pfn
= pageblock_end_pfn(pfn
);
693 for (; pfn
< end_pfn
; pfn
+= isolated
,
694 block_start_pfn
= block_end_pfn
,
695 block_end_pfn
+= pageblock_nr_pages
) {
696 /* Protect pfn from changing by isolate_freepages_block */
697 unsigned long isolate_start_pfn
= pfn
;
699 block_end_pfn
= min(block_end_pfn
, end_pfn
);
702 * pfn could pass the block_end_pfn if isolated freepage
703 * is more than pageblock order. In this case, we adjust
704 * scanning range to right one.
706 if (pfn
>= block_end_pfn
) {
707 block_start_pfn
= pageblock_start_pfn(pfn
);
708 block_end_pfn
= pageblock_end_pfn(pfn
);
709 block_end_pfn
= min(block_end_pfn
, end_pfn
);
712 if (!pageblock_pfn_to_page(block_start_pfn
,
713 block_end_pfn
, cc
->zone
))
716 isolated
= isolate_freepages_block(cc
, &isolate_start_pfn
,
717 block_end_pfn
, &freelist
, 0, true);
720 * In strict mode, isolate_freepages_block() returns 0 if
721 * there are any holes in the block (ie. invalid PFNs or
728 * If we managed to isolate pages, it is always (1 << n) *
729 * pageblock_nr_pages for some non-negative n. (Max order
730 * page may span two pageblocks).
734 /* __isolate_free_page() does not map the pages */
735 split_map_pages(&freelist
);
738 /* Loop terminated early, cleanup. */
739 release_freepages(&freelist
);
743 /* We don't use freelists for anything. */
747 /* Similar to reclaim, but different enough that they don't share logic */
748 static bool too_many_isolated(pg_data_t
*pgdat
)
752 unsigned long active
, inactive
, isolated
;
754 inactive
= node_page_state(pgdat
, NR_INACTIVE_FILE
) +
755 node_page_state(pgdat
, NR_INACTIVE_ANON
);
756 active
= node_page_state(pgdat
, NR_ACTIVE_FILE
) +
757 node_page_state(pgdat
, NR_ACTIVE_ANON
);
758 isolated
= node_page_state(pgdat
, NR_ISOLATED_FILE
) +
759 node_page_state(pgdat
, NR_ISOLATED_ANON
);
761 too_many
= isolated
> (inactive
+ active
) / 2;
763 wake_throttle_isolated(pgdat
);
769 * isolate_migratepages_block() - isolate all migrate-able pages within
771 * @cc: Compaction control structure.
772 * @low_pfn: The first PFN to isolate
773 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
774 * @mode: Isolation mode to be used.
776 * Isolate all pages that can be migrated from the range specified by
777 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
778 * Returns errno, like -EAGAIN or -EINTR in case e.g signal pending or congestion,
779 * -ENOMEM in case we could not allocate a page, or 0.
780 * cc->migrate_pfn will contain the next pfn to scan.
782 * The pages are isolated on cc->migratepages list (not required to be empty),
783 * and cc->nr_migratepages is updated accordingly.
786 isolate_migratepages_block(struct compact_control
*cc
, unsigned long low_pfn
,
787 unsigned long end_pfn
, isolate_mode_t mode
)
789 pg_data_t
*pgdat
= cc
->zone
->zone_pgdat
;
790 unsigned long nr_scanned
= 0, nr_isolated
= 0;
791 struct lruvec
*lruvec
;
792 unsigned long flags
= 0;
793 struct lruvec
*locked
= NULL
;
794 struct page
*page
= NULL
, *valid_page
= NULL
;
795 struct address_space
*mapping
;
796 unsigned long start_pfn
= low_pfn
;
797 bool skip_on_failure
= false;
798 unsigned long next_skip_pfn
= 0;
799 bool skip_updated
= false;
802 cc
->migrate_pfn
= low_pfn
;
805 * Ensure that there are not too many pages isolated from the LRU
806 * list by either parallel reclaimers or compaction. If there are,
807 * delay for some time until fewer pages are isolated
809 while (unlikely(too_many_isolated(pgdat
))) {
810 /* stop isolation if there are still pages not migrated */
811 if (cc
->nr_migratepages
)
814 /* async migration should just abort */
815 if (cc
->mode
== MIGRATE_ASYNC
)
818 reclaim_throttle(pgdat
, VMSCAN_THROTTLE_ISOLATED
);
820 if (fatal_signal_pending(current
))
826 if (cc
->direct_compaction
&& (cc
->mode
== MIGRATE_ASYNC
)) {
827 skip_on_failure
= true;
828 next_skip_pfn
= block_end_pfn(low_pfn
, cc
->order
);
831 /* Time to isolate some pages for migration */
832 for (; low_pfn
< end_pfn
; low_pfn
++) {
834 if (skip_on_failure
&& low_pfn
>= next_skip_pfn
) {
836 * We have isolated all migration candidates in the
837 * previous order-aligned block, and did not skip it due
838 * to failure. We should migrate the pages now and
839 * hopefully succeed compaction.
845 * We failed to isolate in the previous order-aligned
846 * block. Set the new boundary to the end of the
847 * current block. Note we can't simply increase
848 * next_skip_pfn by 1 << order, as low_pfn might have
849 * been incremented by a higher number due to skipping
850 * a compound or a high-order buddy page in the
851 * previous loop iteration.
853 next_skip_pfn
= block_end_pfn(low_pfn
, cc
->order
);
857 * Periodically drop the lock (if held) regardless of its
858 * contention, to give chance to IRQs. Abort completely if
859 * a fatal signal is pending.
861 if (!(low_pfn
% COMPACT_CLUSTER_MAX
)) {
863 unlock_page_lruvec_irqrestore(locked
, flags
);
867 if (fatal_signal_pending(current
)) {
868 cc
->contended
= true;
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
&& pageblock_aligned(low_pfn
)) {
888 if (!isolation_suitable(cc
, page
)) {
896 if (PageHuge(page
) && cc
->alloc_contig
) {
898 unlock_page_lruvec_irqrestore(locked
, flags
);
902 ret
= isolate_or_dissolve_huge_page(page
, &cc
->migratepages
);
905 * Fail isolation in case isolate_or_dissolve_huge_page()
906 * reports an error. In case of -ENOMEM, abort right away.
909 /* Do not report -EBUSY down the chain */
912 low_pfn
+= compound_nr(page
) - 1;
913 nr_scanned
+= compound_nr(page
) - 1;
917 if (PageHuge(page
)) {
919 * Hugepage was successfully isolated and placed
920 * on the cc->migratepages list.
922 low_pfn
+= compound_nr(page
) - 1;
923 goto isolate_success_no_list
;
927 * Ok, the hugepage was dissolved. Now these pages are
928 * Buddy and cannot be re-allocated because they are
929 * isolated. Fall-through as the check below handles
935 * Skip if free. We read page order here without zone lock
936 * which is generally unsafe, but the race window is small and
937 * the worst thing that can happen is that we skip some
938 * potential isolation targets.
940 if (PageBuddy(page
)) {
941 unsigned long freepage_order
= buddy_order_unsafe(page
);
944 * Without lock, we cannot be sure that what we got is
945 * a valid page order. Consider only values in the
946 * valid order range to prevent low_pfn overflow.
948 if (freepage_order
> 0 && freepage_order
<= MAX_ORDER
) {
949 low_pfn
+= (1UL << freepage_order
) - 1;
950 nr_scanned
+= (1UL << freepage_order
) - 1;
956 * Regardless of being on LRU, compound pages such as THP and
957 * hugetlbfs are not to be compacted unless we are attempting
958 * an allocation much larger than the huge page size (eg CMA).
959 * We can potentially save a lot of iterations if we skip them
960 * at once. The check is racy, but we can consider only valid
961 * values and the only danger is skipping too much.
963 if (PageCompound(page
) && !cc
->alloc_contig
) {
964 const unsigned int order
= compound_order(page
);
966 if (likely(order
<= MAX_ORDER
)) {
967 low_pfn
+= (1UL << order
) - 1;
968 nr_scanned
+= (1UL << order
) - 1;
974 * Check may be lockless but that's ok as we recheck later.
975 * It's possible to migrate LRU and non-lru movable pages.
976 * Skip any other type of page
978 if (!PageLRU(page
)) {
980 * __PageMovable can return false positive so we need
981 * to verify it under page_lock.
983 if (unlikely(__PageMovable(page
)) &&
984 !PageIsolated(page
)) {
986 unlock_page_lruvec_irqrestore(locked
, flags
);
990 if (isolate_movable_page(page
, mode
))
991 goto isolate_success
;
998 * Be careful not to clear PageLRU until after we're
999 * sure the page is not being freed elsewhere -- the
1000 * page release code relies on it.
1002 if (unlikely(!get_page_unless_zero(page
)))
1006 * Migration will fail if an anonymous page is pinned in memory,
1007 * so avoid taking lru_lock and isolating it unnecessarily in an
1008 * admittedly racy check.
1010 mapping
= page_mapping(page
);
1011 if (!mapping
&& (page_count(page
) - 1) > total_mapcount(page
))
1012 goto isolate_fail_put
;
1015 * Only allow to migrate anonymous pages in GFP_NOFS context
1016 * because those do not depend on fs locks.
1018 if (!(cc
->gfp_mask
& __GFP_FS
) && mapping
)
1019 goto isolate_fail_put
;
1021 /* Only take pages on LRU: a check now makes later tests safe */
1023 goto isolate_fail_put
;
1025 /* Compaction might skip unevictable pages but CMA takes them */
1026 if (!(mode
& ISOLATE_UNEVICTABLE
) && PageUnevictable(page
))
1027 goto isolate_fail_put
;
1030 * To minimise LRU disruption, the caller can indicate with
1031 * ISOLATE_ASYNC_MIGRATE that it only wants to isolate pages
1032 * it will be able to migrate without blocking - clean pages
1033 * for the most part. PageWriteback would require blocking.
1035 if ((mode
& ISOLATE_ASYNC_MIGRATE
) && PageWriteback(page
))
1036 goto isolate_fail_put
;
1038 if ((mode
& ISOLATE_ASYNC_MIGRATE
) && PageDirty(page
)) {
1042 * Only pages without mappings or that have a
1043 * ->migrate_folio callback are possible to migrate
1044 * without blocking. However, we can be racing with
1045 * truncation so it's necessary to lock the page
1046 * to stabilise the mapping as truncation holds
1047 * the page lock until after the page is removed
1048 * from the page cache.
1050 if (!trylock_page(page
))
1051 goto isolate_fail_put
;
1053 mapping
= page_mapping(page
);
1054 migrate_dirty
= !mapping
||
1055 mapping
->a_ops
->migrate_folio
;
1058 goto isolate_fail_put
;
1061 /* Try isolate the page */
1062 if (!TestClearPageLRU(page
))
1063 goto isolate_fail_put
;
1065 lruvec
= folio_lruvec(page_folio(page
));
1067 /* If we already hold the lock, we can skip some rechecking */
1068 if (lruvec
!= locked
) {
1070 unlock_page_lruvec_irqrestore(locked
, flags
);
1072 compact_lock_irqsave(&lruvec
->lru_lock
, &flags
, cc
);
1075 lruvec_memcg_debug(lruvec
, page_folio(page
));
1077 /* Try get exclusive access under lock */
1078 if (!skip_updated
) {
1079 skip_updated
= true;
1080 if (test_and_set_skip(cc
, page
, low_pfn
))
1085 * Page become compound since the non-locked check,
1086 * and it's on LRU. It can only be a THP so the order
1087 * is safe to read and it's 0 for tail pages.
1089 if (unlikely(PageCompound(page
) && !cc
->alloc_contig
)) {
1090 low_pfn
+= compound_nr(page
) - 1;
1091 nr_scanned
+= compound_nr(page
) - 1;
1093 goto isolate_fail_put
;
1097 /* The whole page is taken off the LRU; skip the tail pages. */
1098 if (PageCompound(page
))
1099 low_pfn
+= compound_nr(page
) - 1;
1101 /* Successfully isolated */
1102 del_page_from_lru_list(page
, lruvec
);
1103 mod_node_page_state(page_pgdat(page
),
1104 NR_ISOLATED_ANON
+ page_is_file_lru(page
),
1105 thp_nr_pages(page
));
1108 list_add(&page
->lru
, &cc
->migratepages
);
1109 isolate_success_no_list
:
1110 cc
->nr_migratepages
+= compound_nr(page
);
1111 nr_isolated
+= compound_nr(page
);
1112 nr_scanned
+= compound_nr(page
) - 1;
1115 * Avoid isolating too much unless this block is being
1116 * fully scanned (e.g. dirty/writeback pages, parallel allocation)
1117 * or a lock is contended. For contention, isolate quickly to
1118 * potentially remove one source of contention.
1120 if (cc
->nr_migratepages
>= COMPACT_CLUSTER_MAX
&&
1121 !cc
->finish_pageblock
&& !cc
->contended
) {
1129 /* Avoid potential deadlock in freeing page under lru_lock */
1131 unlock_page_lruvec_irqrestore(locked
, flags
);
1137 if (!skip_on_failure
&& ret
!= -ENOMEM
)
1141 * We have isolated some pages, but then failed. Release them
1142 * instead of migrating, as we cannot form the cc->order buddy
1147 unlock_page_lruvec_irqrestore(locked
, flags
);
1150 putback_movable_pages(&cc
->migratepages
);
1151 cc
->nr_migratepages
= 0;
1155 if (low_pfn
< next_skip_pfn
) {
1156 low_pfn
= next_skip_pfn
- 1;
1158 * The check near the loop beginning would have updated
1159 * next_skip_pfn too, but this is a bit simpler.
1161 next_skip_pfn
+= 1UL << cc
->order
;
1169 * The PageBuddy() check could have potentially brought us outside
1170 * the range to be scanned.
1172 if (unlikely(low_pfn
> end_pfn
))
1179 unlock_page_lruvec_irqrestore(locked
, flags
);
1186 * Update the cached scanner pfn once the pageblock has been scanned.
1187 * Pages will either be migrated in which case there is no point
1188 * scanning in the near future or migration failed in which case the
1189 * failure reason may persist. The block is marked for skipping if
1190 * there were no pages isolated in the block or if the block is
1191 * rescanned twice in a row.
1193 if (low_pfn
== end_pfn
&& (!nr_isolated
|| cc
->finish_pageblock
)) {
1194 if (valid_page
&& !skip_updated
)
1195 set_pageblock_skip(valid_page
);
1196 update_cached_migrate(cc
, low_pfn
);
1199 trace_mm_compaction_isolate_migratepages(start_pfn
, low_pfn
,
1200 nr_scanned
, nr_isolated
);
1203 cc
->total_migrate_scanned
+= nr_scanned
;
1205 count_compact_events(COMPACTISOLATED
, nr_isolated
);
1207 cc
->migrate_pfn
= low_pfn
;
1213 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
1214 * @cc: Compaction control structure.
1215 * @start_pfn: The first PFN to start isolating.
1216 * @end_pfn: The one-past-last PFN.
1218 * Returns -EAGAIN when contented, -EINTR in case of a signal pending, -ENOMEM
1219 * in case we could not allocate a page, or 0.
1222 isolate_migratepages_range(struct compact_control
*cc
, unsigned long start_pfn
,
1223 unsigned long end_pfn
)
1225 unsigned long pfn
, block_start_pfn
, block_end_pfn
;
1228 /* Scan block by block. First and last block may be incomplete */
1230 block_start_pfn
= pageblock_start_pfn(pfn
);
1231 if (block_start_pfn
< cc
->zone
->zone_start_pfn
)
1232 block_start_pfn
= cc
->zone
->zone_start_pfn
;
1233 block_end_pfn
= pageblock_end_pfn(pfn
);
1235 for (; pfn
< end_pfn
; pfn
= block_end_pfn
,
1236 block_start_pfn
= block_end_pfn
,
1237 block_end_pfn
+= pageblock_nr_pages
) {
1239 block_end_pfn
= min(block_end_pfn
, end_pfn
);
1241 if (!pageblock_pfn_to_page(block_start_pfn
,
1242 block_end_pfn
, cc
->zone
))
1245 ret
= isolate_migratepages_block(cc
, pfn
, block_end_pfn
,
1246 ISOLATE_UNEVICTABLE
);
1251 if (cc
->nr_migratepages
>= COMPACT_CLUSTER_MAX
)
1258 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
1259 #ifdef CONFIG_COMPACTION
1261 static bool suitable_migration_source(struct compact_control
*cc
,
1266 if (pageblock_skip_persistent(page
))
1269 if ((cc
->mode
!= MIGRATE_ASYNC
) || !cc
->direct_compaction
)
1272 block_mt
= get_pageblock_migratetype(page
);
1274 if (cc
->migratetype
== MIGRATE_MOVABLE
)
1275 return is_migrate_movable(block_mt
);
1277 return block_mt
== cc
->migratetype
;
1280 /* Returns true if the page is within a block suitable for migration to */
1281 static bool suitable_migration_target(struct compact_control
*cc
,
1284 /* If the page is a large free page, then disallow migration */
1285 if (PageBuddy(page
)) {
1287 * We are checking page_order without zone->lock taken. But
1288 * the only small danger is that we skip a potentially suitable
1289 * pageblock, so it's not worth to check order for valid range.
1291 if (buddy_order_unsafe(page
) >= pageblock_order
)
1295 if (cc
->ignore_block_suitable
)
1298 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1299 if (is_migrate_movable(get_pageblock_migratetype(page
)))
1302 /* Otherwise skip the block */
1306 static inline unsigned int
1307 freelist_scan_limit(struct compact_control
*cc
)
1309 unsigned short shift
= BITS_PER_LONG
- 1;
1311 return (COMPACT_CLUSTER_MAX
>> min(shift
, cc
->fast_search_fail
)) + 1;
1315 * Test whether the free scanner has reached the same or lower pageblock than
1316 * the migration scanner, and compaction should thus terminate.
1318 static inline bool compact_scanners_met(struct compact_control
*cc
)
1320 return (cc
->free_pfn
>> pageblock_order
)
1321 <= (cc
->migrate_pfn
>> pageblock_order
);
1325 * Used when scanning for a suitable migration target which scans freelists
1326 * in reverse. Reorders the list such as the unscanned pages are scanned
1327 * first on the next iteration of the free scanner
1330 move_freelist_head(struct list_head
*freelist
, struct page
*freepage
)
1334 if (!list_is_last(freelist
, &freepage
->lru
)) {
1335 list_cut_before(&sublist
, freelist
, &freepage
->lru
);
1336 list_splice_tail(&sublist
, freelist
);
1341 * Similar to move_freelist_head except used by the migration scanner
1342 * when scanning forward. It's possible for these list operations to
1343 * move against each other if they search the free list exactly in
1347 move_freelist_tail(struct list_head
*freelist
, struct page
*freepage
)
1351 if (!list_is_first(freelist
, &freepage
->lru
)) {
1352 list_cut_position(&sublist
, freelist
, &freepage
->lru
);
1353 list_splice_tail(&sublist
, freelist
);
1358 fast_isolate_around(struct compact_control
*cc
, unsigned long pfn
)
1360 unsigned long start_pfn
, end_pfn
;
1363 /* Do not search around if there are enough pages already */
1364 if (cc
->nr_freepages
>= cc
->nr_migratepages
)
1367 /* Minimise scanning during async compaction */
1368 if (cc
->direct_compaction
&& cc
->mode
== MIGRATE_ASYNC
)
1371 /* Pageblock boundaries */
1372 start_pfn
= max(pageblock_start_pfn(pfn
), cc
->zone
->zone_start_pfn
);
1373 end_pfn
= min(pageblock_end_pfn(pfn
), zone_end_pfn(cc
->zone
));
1375 page
= pageblock_pfn_to_page(start_pfn
, end_pfn
, cc
->zone
);
1379 isolate_freepages_block(cc
, &start_pfn
, end_pfn
, &cc
->freepages
, 1, false);
1381 /* Skip this pageblock in the future as it's full or nearly full */
1382 if (cc
->nr_freepages
< cc
->nr_migratepages
)
1383 set_pageblock_skip(page
);
1388 /* Search orders in round-robin fashion */
1389 static int next_search_order(struct compact_control
*cc
, int order
)
1393 order
= cc
->order
- 1;
1395 /* Search wrapped around? */
1396 if (order
== cc
->search_order
) {
1398 if (cc
->search_order
< 0)
1399 cc
->search_order
= cc
->order
- 1;
1406 static unsigned long
1407 fast_isolate_freepages(struct compact_control
*cc
)
1409 unsigned int limit
= max(1U, freelist_scan_limit(cc
) >> 1);
1410 unsigned int nr_scanned
= 0;
1411 unsigned long low_pfn
, min_pfn
, highest
= 0;
1412 unsigned long nr_isolated
= 0;
1413 unsigned long distance
;
1414 struct page
*page
= NULL
;
1415 bool scan_start
= false;
1418 /* Full compaction passes in a negative order */
1420 return cc
->free_pfn
;
1423 * If starting the scan, use a deeper search and use the highest
1424 * PFN found if a suitable one is not found.
1426 if (cc
->free_pfn
>= cc
->zone
->compact_init_free_pfn
) {
1427 limit
= pageblock_nr_pages
>> 1;
1432 * Preferred point is in the top quarter of the scan space but take
1433 * a pfn from the top half if the search is problematic.
1435 distance
= (cc
->free_pfn
- cc
->migrate_pfn
);
1436 low_pfn
= pageblock_start_pfn(cc
->free_pfn
- (distance
>> 2));
1437 min_pfn
= pageblock_start_pfn(cc
->free_pfn
- (distance
>> 1));
1439 if (WARN_ON_ONCE(min_pfn
> low_pfn
))
1443 * Search starts from the last successful isolation order or the next
1444 * order to search after a previous failure
1446 cc
->search_order
= min_t(unsigned int, cc
->order
- 1, cc
->search_order
);
1448 for (order
= cc
->search_order
;
1449 !page
&& order
>= 0;
1450 order
= next_search_order(cc
, order
)) {
1451 struct free_area
*area
= &cc
->zone
->free_area
[order
];
1452 struct list_head
*freelist
;
1453 struct page
*freepage
;
1454 unsigned long flags
;
1455 unsigned int order_scanned
= 0;
1456 unsigned long high_pfn
= 0;
1461 spin_lock_irqsave(&cc
->zone
->lock
, flags
);
1462 freelist
= &area
->free_list
[MIGRATE_MOVABLE
];
1463 list_for_each_entry_reverse(freepage
, freelist
, lru
) {
1468 pfn
= page_to_pfn(freepage
);
1471 highest
= max(pageblock_start_pfn(pfn
),
1472 cc
->zone
->zone_start_pfn
);
1474 if (pfn
>= low_pfn
) {
1475 cc
->fast_search_fail
= 0;
1476 cc
->search_order
= order
;
1481 if (pfn
>= min_pfn
&& pfn
> high_pfn
) {
1484 /* Shorten the scan if a candidate is found */
1488 if (order_scanned
>= limit
)
1492 /* Use a minimum pfn if a preferred one was not found */
1493 if (!page
&& high_pfn
) {
1494 page
= pfn_to_page(high_pfn
);
1496 /* Update freepage for the list reorder below */
1500 /* Reorder to so a future search skips recent pages */
1501 move_freelist_head(freelist
, freepage
);
1503 /* Isolate the page if available */
1505 if (__isolate_free_page(page
, order
)) {
1506 set_page_private(page
, order
);
1507 nr_isolated
= 1 << order
;
1508 nr_scanned
+= nr_isolated
- 1;
1509 cc
->nr_freepages
+= nr_isolated
;
1510 list_add_tail(&page
->lru
, &cc
->freepages
);
1511 count_compact_events(COMPACTISOLATED
, nr_isolated
);
1513 /* If isolation fails, abort the search */
1514 order
= cc
->search_order
+ 1;
1519 spin_unlock_irqrestore(&cc
->zone
->lock
, flags
);
1522 * Smaller scan on next order so the total scan is related
1523 * to freelist_scan_limit.
1525 if (order_scanned
>= limit
)
1526 limit
= max(1U, limit
>> 1);
1530 cc
->fast_search_fail
++;
1533 * Use the highest PFN found above min. If one was
1534 * not found, be pessimistic for direct compaction
1535 * and use the min mark.
1537 if (highest
>= min_pfn
) {
1538 page
= pfn_to_page(highest
);
1539 cc
->free_pfn
= highest
;
1541 if (cc
->direct_compaction
&& pfn_valid(min_pfn
)) {
1542 page
= pageblock_pfn_to_page(min_pfn
,
1543 min(pageblock_end_pfn(min_pfn
),
1544 zone_end_pfn(cc
->zone
)),
1546 cc
->free_pfn
= min_pfn
;
1552 if (highest
&& highest
>= cc
->zone
->compact_cached_free_pfn
) {
1553 highest
-= pageblock_nr_pages
;
1554 cc
->zone
->compact_cached_free_pfn
= highest
;
1557 cc
->total_free_scanned
+= nr_scanned
;
1559 return cc
->free_pfn
;
1561 low_pfn
= page_to_pfn(page
);
1562 fast_isolate_around(cc
, low_pfn
);
1567 * Based on information in the current compact_control, find blocks
1568 * suitable for isolating free pages from and then isolate them.
1570 static void isolate_freepages(struct compact_control
*cc
)
1572 struct zone
*zone
= cc
->zone
;
1574 unsigned long block_start_pfn
; /* start of current pageblock */
1575 unsigned long isolate_start_pfn
; /* exact pfn we start at */
1576 unsigned long block_end_pfn
; /* end of current pageblock */
1577 unsigned long low_pfn
; /* lowest pfn scanner is able to scan */
1578 struct list_head
*freelist
= &cc
->freepages
;
1579 unsigned int stride
;
1581 /* Try a small search of the free lists for a candidate */
1582 fast_isolate_freepages(cc
);
1583 if (cc
->nr_freepages
)
1587 * Initialise the free scanner. The starting point is where we last
1588 * successfully isolated from, zone-cached value, or the end of the
1589 * zone when isolating for the first time. For looping we also need
1590 * this pfn aligned down to the pageblock boundary, because we do
1591 * block_start_pfn -= pageblock_nr_pages in the for loop.
1592 * For ending point, take care when isolating in last pageblock of a
1593 * zone which ends in the middle of a pageblock.
1594 * The low boundary is the end of the pageblock the migration scanner
1597 isolate_start_pfn
= cc
->free_pfn
;
1598 block_start_pfn
= pageblock_start_pfn(isolate_start_pfn
);
1599 block_end_pfn
= min(block_start_pfn
+ pageblock_nr_pages
,
1600 zone_end_pfn(zone
));
1601 low_pfn
= pageblock_end_pfn(cc
->migrate_pfn
);
1602 stride
= cc
->mode
== MIGRATE_ASYNC
? COMPACT_CLUSTER_MAX
: 1;
1605 * Isolate free pages until enough are available to migrate the
1606 * pages on cc->migratepages. We stop searching if the migrate
1607 * and free page scanners meet or enough free pages are isolated.
1609 for (; block_start_pfn
>= low_pfn
;
1610 block_end_pfn
= block_start_pfn
,
1611 block_start_pfn
-= pageblock_nr_pages
,
1612 isolate_start_pfn
= block_start_pfn
) {
1613 unsigned long nr_isolated
;
1616 * This can iterate a massively long zone without finding any
1617 * suitable migration targets, so periodically check resched.
1619 if (!(block_start_pfn
% (COMPACT_CLUSTER_MAX
* pageblock_nr_pages
)))
1622 page
= pageblock_pfn_to_page(block_start_pfn
, block_end_pfn
,
1627 /* Check the block is suitable for migration */
1628 if (!suitable_migration_target(cc
, page
))
1631 /* If isolation recently failed, do not retry */
1632 if (!isolation_suitable(cc
, page
))
1635 /* Found a block suitable for isolating free pages from. */
1636 nr_isolated
= isolate_freepages_block(cc
, &isolate_start_pfn
,
1637 block_end_pfn
, freelist
, stride
, false);
1639 /* Update the skip hint if the full pageblock was scanned */
1640 if (isolate_start_pfn
== block_end_pfn
)
1641 update_pageblock_skip(cc
, page
, block_start_pfn
);
1643 /* Are enough freepages isolated? */
1644 if (cc
->nr_freepages
>= cc
->nr_migratepages
) {
1645 if (isolate_start_pfn
>= block_end_pfn
) {
1647 * Restart at previous pageblock if more
1648 * freepages can be isolated next time.
1651 block_start_pfn
- pageblock_nr_pages
;
1654 } else if (isolate_start_pfn
< block_end_pfn
) {
1656 * If isolation failed early, do not continue
1662 /* Adjust stride depending on isolation */
1667 stride
= min_t(unsigned int, COMPACT_CLUSTER_MAX
, stride
<< 1);
1671 * Record where the free scanner will restart next time. Either we
1672 * broke from the loop and set isolate_start_pfn based on the last
1673 * call to isolate_freepages_block(), or we met the migration scanner
1674 * and the loop terminated due to isolate_start_pfn < low_pfn
1676 cc
->free_pfn
= isolate_start_pfn
;
1679 /* __isolate_free_page() does not map the pages */
1680 split_map_pages(freelist
);
1684 * This is a migrate-callback that "allocates" freepages by taking pages
1685 * from the isolated freelists in the block we are migrating to.
1687 static struct page
*compaction_alloc(struct page
*migratepage
,
1690 struct compact_control
*cc
= (struct compact_control
*)data
;
1691 struct page
*freepage
;
1693 if (list_empty(&cc
->freepages
)) {
1694 isolate_freepages(cc
);
1696 if (list_empty(&cc
->freepages
))
1700 freepage
= list_entry(cc
->freepages
.next
, struct page
, lru
);
1701 list_del(&freepage
->lru
);
1708 * This is a migrate-callback that "frees" freepages back to the isolated
1709 * freelist. All pages on the freelist are from the same zone, so there is no
1710 * special handling needed for NUMA.
1712 static void compaction_free(struct page
*page
, unsigned long data
)
1714 struct compact_control
*cc
= (struct compact_control
*)data
;
1716 list_add(&page
->lru
, &cc
->freepages
);
1720 /* possible outcome of isolate_migratepages */
1722 ISOLATE_ABORT
, /* Abort compaction now */
1723 ISOLATE_NONE
, /* No pages isolated, continue scanning */
1724 ISOLATE_SUCCESS
, /* Pages isolated, migrate */
1725 } isolate_migrate_t
;
1728 * Allow userspace to control policy on scanning the unevictable LRU for
1729 * compactable pages.
1731 static int sysctl_compact_unevictable_allowed __read_mostly
= CONFIG_COMPACT_UNEVICTABLE_DEFAULT
;
1733 * Tunable for proactive compaction. It determines how
1734 * aggressively the kernel should compact memory in the
1735 * background. It takes values in the range [0, 100].
1737 static unsigned int __read_mostly sysctl_compaction_proactiveness
= 20;
1738 static int sysctl_extfrag_threshold
= 500;
1741 update_fast_start_pfn(struct compact_control
*cc
, unsigned long pfn
)
1743 if (cc
->fast_start_pfn
== ULONG_MAX
)
1746 if (!cc
->fast_start_pfn
)
1747 cc
->fast_start_pfn
= pfn
;
1749 cc
->fast_start_pfn
= min(cc
->fast_start_pfn
, pfn
);
1752 static inline unsigned long
1753 reinit_migrate_pfn(struct compact_control
*cc
)
1755 if (!cc
->fast_start_pfn
|| cc
->fast_start_pfn
== ULONG_MAX
)
1756 return cc
->migrate_pfn
;
1758 cc
->migrate_pfn
= cc
->fast_start_pfn
;
1759 cc
->fast_start_pfn
= ULONG_MAX
;
1761 return cc
->migrate_pfn
;
1765 * Briefly search the free lists for a migration source that already has
1766 * some free pages to reduce the number of pages that need migration
1767 * before a pageblock is free.
1769 static unsigned long fast_find_migrateblock(struct compact_control
*cc
)
1771 unsigned int limit
= freelist_scan_limit(cc
);
1772 unsigned int nr_scanned
= 0;
1773 unsigned long distance
;
1774 unsigned long pfn
= cc
->migrate_pfn
;
1775 unsigned long high_pfn
;
1777 bool found_block
= false;
1779 /* Skip hints are relied on to avoid repeats on the fast search */
1780 if (cc
->ignore_skip_hint
)
1784 * If the pageblock should be finished then do not select a different
1787 if (cc
->finish_pageblock
)
1791 * If the migrate_pfn is not at the start of a zone or the start
1792 * of a pageblock then assume this is a continuation of a previous
1793 * scan restarted due to COMPACT_CLUSTER_MAX.
1795 if (pfn
!= cc
->zone
->zone_start_pfn
&& pfn
!= pageblock_start_pfn(pfn
))
1799 * For smaller orders, just linearly scan as the number of pages
1800 * to migrate should be relatively small and does not necessarily
1801 * justify freeing up a large block for a small allocation.
1803 if (cc
->order
<= PAGE_ALLOC_COSTLY_ORDER
)
1807 * Only allow kcompactd and direct requests for movable pages to
1808 * quickly clear out a MOVABLE pageblock for allocation. This
1809 * reduces the risk that a large movable pageblock is freed for
1810 * an unmovable/reclaimable small allocation.
1812 if (cc
->direct_compaction
&& cc
->migratetype
!= MIGRATE_MOVABLE
)
1816 * When starting the migration scanner, pick any pageblock within the
1817 * first half of the search space. Otherwise try and pick a pageblock
1818 * within the first eighth to reduce the chances that a migration
1819 * target later becomes a source.
1821 distance
= (cc
->free_pfn
- cc
->migrate_pfn
) >> 1;
1822 if (cc
->migrate_pfn
!= cc
->zone
->zone_start_pfn
)
1824 high_pfn
= pageblock_start_pfn(cc
->migrate_pfn
+ distance
);
1826 for (order
= cc
->order
- 1;
1827 order
>= PAGE_ALLOC_COSTLY_ORDER
&& !found_block
&& nr_scanned
< limit
;
1829 struct free_area
*area
= &cc
->zone
->free_area
[order
];
1830 struct list_head
*freelist
;
1831 unsigned long flags
;
1832 struct page
*freepage
;
1837 spin_lock_irqsave(&cc
->zone
->lock
, flags
);
1838 freelist
= &area
->free_list
[MIGRATE_MOVABLE
];
1839 list_for_each_entry(freepage
, freelist
, lru
) {
1840 unsigned long free_pfn
;
1842 if (nr_scanned
++ >= limit
) {
1843 move_freelist_tail(freelist
, freepage
);
1847 free_pfn
= page_to_pfn(freepage
);
1848 if (free_pfn
< high_pfn
) {
1850 * Avoid if skipped recently. Ideally it would
1851 * move to the tail but even safe iteration of
1852 * the list assumes an entry is deleted, not
1855 if (get_pageblock_skip(freepage
))
1858 /* Reorder to so a future search skips recent pages */
1859 move_freelist_tail(freelist
, freepage
);
1861 update_fast_start_pfn(cc
, free_pfn
);
1862 pfn
= pageblock_start_pfn(free_pfn
);
1863 if (pfn
< cc
->zone
->zone_start_pfn
)
1864 pfn
= cc
->zone
->zone_start_pfn
;
1865 cc
->fast_search_fail
= 0;
1867 set_pageblock_skip(freepage
);
1871 spin_unlock_irqrestore(&cc
->zone
->lock
, flags
);
1874 cc
->total_migrate_scanned
+= nr_scanned
;
1877 * If fast scanning failed then use a cached entry for a page block
1878 * that had free pages as the basis for starting a linear scan.
1881 cc
->fast_search_fail
++;
1882 pfn
= reinit_migrate_pfn(cc
);
1888 * Isolate all pages that can be migrated from the first suitable block,
1889 * starting at the block pointed to by the migrate scanner pfn within
1892 static isolate_migrate_t
isolate_migratepages(struct compact_control
*cc
)
1894 unsigned long block_start_pfn
;
1895 unsigned long block_end_pfn
;
1896 unsigned long low_pfn
;
1898 const isolate_mode_t isolate_mode
=
1899 (sysctl_compact_unevictable_allowed
? ISOLATE_UNEVICTABLE
: 0) |
1900 (cc
->mode
!= MIGRATE_SYNC
? ISOLATE_ASYNC_MIGRATE
: 0);
1901 bool fast_find_block
;
1904 * Start at where we last stopped, or beginning of the zone as
1905 * initialized by compact_zone(). The first failure will use
1906 * the lowest PFN as the starting point for linear scanning.
1908 low_pfn
= fast_find_migrateblock(cc
);
1909 block_start_pfn
= pageblock_start_pfn(low_pfn
);
1910 if (block_start_pfn
< cc
->zone
->zone_start_pfn
)
1911 block_start_pfn
= cc
->zone
->zone_start_pfn
;
1914 * fast_find_migrateblock marks a pageblock skipped so to avoid
1915 * the isolation_suitable check below, check whether the fast
1916 * search was successful.
1918 fast_find_block
= low_pfn
!= cc
->migrate_pfn
&& !cc
->fast_search_fail
;
1920 /* Only scan within a pageblock boundary */
1921 block_end_pfn
= pageblock_end_pfn(low_pfn
);
1924 * Iterate over whole pageblocks until we find the first suitable.
1925 * Do not cross the free scanner.
1927 for (; block_end_pfn
<= cc
->free_pfn
;
1928 fast_find_block
= false,
1929 cc
->migrate_pfn
= low_pfn
= block_end_pfn
,
1930 block_start_pfn
= block_end_pfn
,
1931 block_end_pfn
+= pageblock_nr_pages
) {
1934 * This can potentially iterate a massively long zone with
1935 * many pageblocks unsuitable, so periodically check if we
1938 if (!(low_pfn
% (COMPACT_CLUSTER_MAX
* pageblock_nr_pages
)))
1941 page
= pageblock_pfn_to_page(block_start_pfn
,
1942 block_end_pfn
, cc
->zone
);
1947 * If isolation recently failed, do not retry. Only check the
1948 * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock
1949 * to be visited multiple times. Assume skip was checked
1950 * before making it "skip" so other compaction instances do
1951 * not scan the same block.
1953 if (pageblock_aligned(low_pfn
) &&
1954 !fast_find_block
&& !isolation_suitable(cc
, page
))
1958 * For async direct compaction, only scan the pageblocks of the
1959 * same migratetype without huge pages. Async direct compaction
1960 * is optimistic to see if the minimum amount of work satisfies
1961 * the allocation. The cached PFN is updated as it's possible
1962 * that all remaining blocks between source and target are
1963 * unsuitable and the compaction scanners fail to meet.
1965 if (!suitable_migration_source(cc
, page
)) {
1966 update_cached_migrate(cc
, block_end_pfn
);
1970 /* Perform the isolation */
1971 if (isolate_migratepages_block(cc
, low_pfn
, block_end_pfn
,
1973 return ISOLATE_ABORT
;
1976 * Either we isolated something and proceed with migration. Or
1977 * we failed and compact_zone should decide if we should
1983 return cc
->nr_migratepages
? ISOLATE_SUCCESS
: ISOLATE_NONE
;
1987 * order == -1 is expected when compacting via
1988 * /proc/sys/vm/compact_memory
1990 static inline bool is_via_compact_memory(int order
)
1996 * Determine whether kswapd is (or recently was!) running on this node.
1998 * pgdat_kswapd_lock() pins pgdat->kswapd, so a concurrent kswapd_stop() can't
2001 static bool kswapd_is_running(pg_data_t
*pgdat
)
2005 pgdat_kswapd_lock(pgdat
);
2006 running
= pgdat
->kswapd
&& task_is_running(pgdat
->kswapd
);
2007 pgdat_kswapd_unlock(pgdat
);
2013 * A zone's fragmentation score is the external fragmentation wrt to the
2014 * COMPACTION_HPAGE_ORDER. It returns a value in the range [0, 100].
2016 static unsigned int fragmentation_score_zone(struct zone
*zone
)
2018 return extfrag_for_order(zone
, COMPACTION_HPAGE_ORDER
);
2022 * A weighted zone's fragmentation score is the external fragmentation
2023 * wrt to the COMPACTION_HPAGE_ORDER scaled by the zone's size. It
2024 * returns a value in the range [0, 100].
2026 * The scaling factor ensures that proactive compaction focuses on larger
2027 * zones like ZONE_NORMAL, rather than smaller, specialized zones like
2028 * ZONE_DMA32. For smaller zones, the score value remains close to zero,
2029 * and thus never exceeds the high threshold for proactive compaction.
2031 static unsigned int fragmentation_score_zone_weighted(struct zone
*zone
)
2033 unsigned long score
;
2035 score
= zone
->present_pages
* fragmentation_score_zone(zone
);
2036 return div64_ul(score
, zone
->zone_pgdat
->node_present_pages
+ 1);
2040 * The per-node proactive (background) compaction process is started by its
2041 * corresponding kcompactd thread when the node's fragmentation score
2042 * exceeds the high threshold. The compaction process remains active till
2043 * the node's score falls below the low threshold, or one of the back-off
2044 * conditions is met.
2046 static unsigned int fragmentation_score_node(pg_data_t
*pgdat
)
2048 unsigned int score
= 0;
2051 for (zoneid
= 0; zoneid
< MAX_NR_ZONES
; zoneid
++) {
2054 zone
= &pgdat
->node_zones
[zoneid
];
2055 if (!populated_zone(zone
))
2057 score
+= fragmentation_score_zone_weighted(zone
);
2063 static unsigned int fragmentation_score_wmark(pg_data_t
*pgdat
, bool low
)
2065 unsigned int wmark_low
;
2068 * Cap the low watermark to avoid excessive compaction
2069 * activity in case a user sets the proactiveness tunable
2070 * close to 100 (maximum).
2072 wmark_low
= max(100U - sysctl_compaction_proactiveness
, 5U);
2073 return low
? wmark_low
: min(wmark_low
+ 10, 100U);
2076 static bool should_proactive_compact_node(pg_data_t
*pgdat
)
2080 if (!sysctl_compaction_proactiveness
|| kswapd_is_running(pgdat
))
2083 wmark_high
= fragmentation_score_wmark(pgdat
, false);
2084 return fragmentation_score_node(pgdat
) > wmark_high
;
2087 static enum compact_result
__compact_finished(struct compact_control
*cc
)
2090 const int migratetype
= cc
->migratetype
;
2093 /* Compaction run completes if the migrate and free scanner meet */
2094 if (compact_scanners_met(cc
)) {
2095 /* Let the next compaction start anew. */
2096 reset_cached_positions(cc
->zone
);
2099 * Mark that the PG_migrate_skip information should be cleared
2100 * by kswapd when it goes to sleep. kcompactd does not set the
2101 * flag itself as the decision to be clear should be directly
2102 * based on an allocation request.
2104 if (cc
->direct_compaction
)
2105 cc
->zone
->compact_blockskip_flush
= true;
2108 return COMPACT_COMPLETE
;
2110 return COMPACT_PARTIAL_SKIPPED
;
2113 if (cc
->proactive_compaction
) {
2114 int score
, wmark_low
;
2117 pgdat
= cc
->zone
->zone_pgdat
;
2118 if (kswapd_is_running(pgdat
))
2119 return COMPACT_PARTIAL_SKIPPED
;
2121 score
= fragmentation_score_zone(cc
->zone
);
2122 wmark_low
= fragmentation_score_wmark(pgdat
, true);
2124 if (score
> wmark_low
)
2125 ret
= COMPACT_CONTINUE
;
2127 ret
= COMPACT_SUCCESS
;
2132 if (is_via_compact_memory(cc
->order
))
2133 return COMPACT_CONTINUE
;
2136 * Always finish scanning a pageblock to reduce the possibility of
2137 * fallbacks in the future. This is particularly important when
2138 * migration source is unmovable/reclaimable but it's not worth
2141 if (!pageblock_aligned(cc
->migrate_pfn
))
2142 return COMPACT_CONTINUE
;
2144 /* Direct compactor: Is a suitable page free? */
2145 ret
= COMPACT_NO_SUITABLE_PAGE
;
2146 for (order
= cc
->order
; order
<= MAX_ORDER
; order
++) {
2147 struct free_area
*area
= &cc
->zone
->free_area
[order
];
2150 /* Job done if page is free of the right migratetype */
2151 if (!free_area_empty(area
, migratetype
))
2152 return COMPACT_SUCCESS
;
2155 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
2156 if (migratetype
== MIGRATE_MOVABLE
&&
2157 !free_area_empty(area
, MIGRATE_CMA
))
2158 return COMPACT_SUCCESS
;
2161 * Job done if allocation would steal freepages from
2162 * other migratetype buddy lists.
2164 if (find_suitable_fallback(area
, order
, migratetype
,
2165 true, &can_steal
) != -1)
2167 * Movable pages are OK in any pageblock. If we are
2168 * stealing for a non-movable allocation, make sure
2169 * we finish compacting the current pageblock first
2170 * (which is assured by the above migrate_pfn align
2171 * check) so it is as free as possible and we won't
2172 * have to steal another one soon.
2174 return COMPACT_SUCCESS
;
2178 if (cc
->contended
|| fatal_signal_pending(current
))
2179 ret
= COMPACT_CONTENDED
;
2184 static enum compact_result
compact_finished(struct compact_control
*cc
)
2188 ret
= __compact_finished(cc
);
2189 trace_mm_compaction_finished(cc
->zone
, cc
->order
, ret
);
2190 if (ret
== COMPACT_NO_SUITABLE_PAGE
)
2191 ret
= COMPACT_CONTINUE
;
2196 static enum compact_result
__compaction_suitable(struct zone
*zone
, int order
,
2197 unsigned int alloc_flags
,
2198 int highest_zoneidx
,
2199 unsigned long wmark_target
)
2201 unsigned long watermark
;
2203 if (is_via_compact_memory(order
))
2204 return COMPACT_CONTINUE
;
2206 watermark
= wmark_pages(zone
, alloc_flags
& ALLOC_WMARK_MASK
);
2208 * If watermarks for high-order allocation are already met, there
2209 * should be no need for compaction at all.
2211 if (zone_watermark_ok(zone
, order
, watermark
, highest_zoneidx
,
2213 return COMPACT_SUCCESS
;
2216 * Watermarks for order-0 must be met for compaction to be able to
2217 * isolate free pages for migration targets. This means that the
2218 * watermark and alloc_flags have to match, or be more pessimistic than
2219 * the check in __isolate_free_page(). We don't use the direct
2220 * compactor's alloc_flags, as they are not relevant for freepage
2221 * isolation. We however do use the direct compactor's highest_zoneidx
2222 * to skip over zones where lowmem reserves would prevent allocation
2223 * even if compaction succeeds.
2224 * For costly orders, we require low watermark instead of min for
2225 * compaction to proceed to increase its chances.
2226 * ALLOC_CMA is used, as pages in CMA pageblocks are considered
2227 * suitable migration targets
2229 watermark
= (order
> PAGE_ALLOC_COSTLY_ORDER
) ?
2230 low_wmark_pages(zone
) : min_wmark_pages(zone
);
2231 watermark
+= compact_gap(order
);
2232 if (!__zone_watermark_ok(zone
, 0, watermark
, highest_zoneidx
,
2233 ALLOC_CMA
, wmark_target
))
2234 return COMPACT_SKIPPED
;
2236 return COMPACT_CONTINUE
;
2240 * compaction_suitable: Is this suitable to run compaction on this zone now?
2242 * COMPACT_SKIPPED - If there are too few free pages for compaction
2243 * COMPACT_SUCCESS - If the allocation would succeed without compaction
2244 * COMPACT_CONTINUE - If compaction should run now
2246 enum compact_result
compaction_suitable(struct zone
*zone
, int order
,
2247 unsigned int alloc_flags
,
2248 int highest_zoneidx
)
2250 enum compact_result ret
;
2253 ret
= __compaction_suitable(zone
, order
, alloc_flags
, highest_zoneidx
,
2254 zone_page_state(zone
, NR_FREE_PAGES
));
2256 * fragmentation index determines if allocation failures are due to
2257 * low memory or external fragmentation
2259 * index of -1000 would imply allocations might succeed depending on
2260 * watermarks, but we already failed the high-order watermark check
2261 * index towards 0 implies failure is due to lack of memory
2262 * index towards 1000 implies failure is due to fragmentation
2264 * Only compact if a failure would be due to fragmentation. Also
2265 * ignore fragindex for non-costly orders where the alternative to
2266 * a successful reclaim/compaction is OOM. Fragindex and the
2267 * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
2268 * excessive compaction for costly orders, but it should not be at the
2269 * expense of system stability.
2271 if (ret
== COMPACT_CONTINUE
&& (order
> PAGE_ALLOC_COSTLY_ORDER
)) {
2272 fragindex
= fragmentation_index(zone
, order
);
2273 if (fragindex
>= 0 && fragindex
<= sysctl_extfrag_threshold
)
2274 ret
= COMPACT_NOT_SUITABLE_ZONE
;
2277 trace_mm_compaction_suitable(zone
, order
, ret
);
2278 if (ret
== COMPACT_NOT_SUITABLE_ZONE
)
2279 ret
= COMPACT_SKIPPED
;
2284 bool compaction_zonelist_suitable(struct alloc_context
*ac
, int order
,
2291 * Make sure at least one zone would pass __compaction_suitable if we continue
2292 * retrying the reclaim.
2294 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
2295 ac
->highest_zoneidx
, ac
->nodemask
) {
2296 unsigned long available
;
2297 enum compact_result compact_result
;
2300 * Do not consider all the reclaimable memory because we do not
2301 * want to trash just for a single high order allocation which
2302 * is even not guaranteed to appear even if __compaction_suitable
2303 * is happy about the watermark check.
2305 available
= zone_reclaimable_pages(zone
) / order
;
2306 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
2307 compact_result
= __compaction_suitable(zone
, order
, alloc_flags
,
2308 ac
->highest_zoneidx
, available
);
2309 if (compact_result
== COMPACT_CONTINUE
)
2316 static enum compact_result
2317 compact_zone(struct compact_control
*cc
, struct capture_control
*capc
)
2319 enum compact_result ret
;
2320 unsigned long start_pfn
= cc
->zone
->zone_start_pfn
;
2321 unsigned long end_pfn
= zone_end_pfn(cc
->zone
);
2322 unsigned long last_migrated_pfn
;
2323 const bool sync
= cc
->mode
!= MIGRATE_ASYNC
;
2325 unsigned int nr_succeeded
= 0;
2328 * These counters track activities during zone compaction. Initialize
2329 * them before compacting a new zone.
2331 cc
->total_migrate_scanned
= 0;
2332 cc
->total_free_scanned
= 0;
2333 cc
->nr_migratepages
= 0;
2334 cc
->nr_freepages
= 0;
2335 INIT_LIST_HEAD(&cc
->freepages
);
2336 INIT_LIST_HEAD(&cc
->migratepages
);
2338 cc
->migratetype
= gfp_migratetype(cc
->gfp_mask
);
2339 ret
= compaction_suitable(cc
->zone
, cc
->order
, cc
->alloc_flags
,
2340 cc
->highest_zoneidx
);
2341 /* Compaction is likely to fail */
2342 if (ret
== COMPACT_SUCCESS
|| ret
== COMPACT_SKIPPED
)
2346 * Clear pageblock skip if there were failures recently and compaction
2347 * is about to be retried after being deferred.
2349 if (compaction_restarting(cc
->zone
, cc
->order
))
2350 __reset_isolation_suitable(cc
->zone
);
2353 * Setup to move all movable pages to the end of the zone. Used cached
2354 * information on where the scanners should start (unless we explicitly
2355 * want to compact the whole zone), but check that it is initialised
2356 * by ensuring the values are within zone boundaries.
2358 cc
->fast_start_pfn
= 0;
2359 if (cc
->whole_zone
) {
2360 cc
->migrate_pfn
= start_pfn
;
2361 cc
->free_pfn
= pageblock_start_pfn(end_pfn
- 1);
2363 cc
->migrate_pfn
= cc
->zone
->compact_cached_migrate_pfn
[sync
];
2364 cc
->free_pfn
= cc
->zone
->compact_cached_free_pfn
;
2365 if (cc
->free_pfn
< start_pfn
|| cc
->free_pfn
>= end_pfn
) {
2366 cc
->free_pfn
= pageblock_start_pfn(end_pfn
- 1);
2367 cc
->zone
->compact_cached_free_pfn
= cc
->free_pfn
;
2369 if (cc
->migrate_pfn
< start_pfn
|| cc
->migrate_pfn
>= end_pfn
) {
2370 cc
->migrate_pfn
= start_pfn
;
2371 cc
->zone
->compact_cached_migrate_pfn
[0] = cc
->migrate_pfn
;
2372 cc
->zone
->compact_cached_migrate_pfn
[1] = cc
->migrate_pfn
;
2375 if (cc
->migrate_pfn
<= cc
->zone
->compact_init_migrate_pfn
)
2376 cc
->whole_zone
= true;
2379 last_migrated_pfn
= 0;
2382 * Migrate has separate cached PFNs for ASYNC and SYNC* migration on
2383 * the basis that some migrations will fail in ASYNC mode. However,
2384 * if the cached PFNs match and pageblocks are skipped due to having
2385 * no isolation candidates, then the sync state does not matter.
2386 * Until a pageblock with isolation candidates is found, keep the
2387 * cached PFNs in sync to avoid revisiting the same blocks.
2389 update_cached
= !sync
&&
2390 cc
->zone
->compact_cached_migrate_pfn
[0] == cc
->zone
->compact_cached_migrate_pfn
[1];
2392 trace_mm_compaction_begin(cc
, start_pfn
, end_pfn
, sync
);
2394 /* lru_add_drain_all could be expensive with involving other CPUs */
2397 while ((ret
= compact_finished(cc
)) == COMPACT_CONTINUE
) {
2399 unsigned long iteration_start_pfn
= cc
->migrate_pfn
;
2402 * Avoid multiple rescans of the same pageblock which can
2403 * happen if a page cannot be isolated (dirty/writeback in
2404 * async mode) or if the migrated pages are being allocated
2405 * before the pageblock is cleared. The first rescan will
2406 * capture the entire pageblock for migration. If it fails,
2407 * it'll be marked skip and scanning will proceed as normal.
2409 cc
->finish_pageblock
= false;
2410 if (pageblock_start_pfn(last_migrated_pfn
) ==
2411 pageblock_start_pfn(iteration_start_pfn
)) {
2412 cc
->finish_pageblock
= true;
2416 switch (isolate_migratepages(cc
)) {
2418 ret
= COMPACT_CONTENDED
;
2419 putback_movable_pages(&cc
->migratepages
);
2420 cc
->nr_migratepages
= 0;
2423 if (update_cached
) {
2424 cc
->zone
->compact_cached_migrate_pfn
[1] =
2425 cc
->zone
->compact_cached_migrate_pfn
[0];
2429 * We haven't isolated and migrated anything, but
2430 * there might still be unflushed migrations from
2431 * previous cc->order aligned block.
2434 case ISOLATE_SUCCESS
:
2435 update_cached
= false;
2436 last_migrated_pfn
= iteration_start_pfn
;
2439 err
= migrate_pages(&cc
->migratepages
, compaction_alloc
,
2440 compaction_free
, (unsigned long)cc
, cc
->mode
,
2441 MR_COMPACTION
, &nr_succeeded
);
2443 trace_mm_compaction_migratepages(cc
, nr_succeeded
);
2445 /* All pages were either migrated or will be released */
2446 cc
->nr_migratepages
= 0;
2448 putback_movable_pages(&cc
->migratepages
);
2450 * migrate_pages() may return -ENOMEM when scanners meet
2451 * and we want compact_finished() to detect it
2453 if (err
== -ENOMEM
&& !compact_scanners_met(cc
)) {
2454 ret
= COMPACT_CONTENDED
;
2458 * If an ASYNC or SYNC_LIGHT fails to migrate a page
2459 * within the current order-aligned block, scan the
2460 * remainder of the pageblock. This will mark the
2461 * pageblock "skip" to avoid rescanning in the near
2462 * future. This will isolate more pages than necessary
2463 * for the request but avoid loops due to
2464 * fast_find_migrateblock revisiting blocks that were
2465 * recently partially scanned.
2467 if (cc
->direct_compaction
&& !cc
->finish_pageblock
&&
2468 (cc
->mode
< MIGRATE_SYNC
)) {
2469 cc
->finish_pageblock
= true;
2472 * Draining pcplists does not help THP if
2473 * any page failed to migrate. Even after
2474 * drain, the pageblock will not be free.
2476 if (cc
->order
== COMPACTION_HPAGE_ORDER
)
2477 last_migrated_pfn
= 0;
2483 /* Stop if a page has been captured */
2484 if (capc
&& capc
->page
) {
2485 ret
= COMPACT_SUCCESS
;
2491 * Has the migration scanner moved away from the previous
2492 * cc->order aligned block where we migrated from? If yes,
2493 * flush the pages that were freed, so that they can merge and
2494 * compact_finished() can detect immediately if allocation
2497 if (cc
->order
> 0 && last_migrated_pfn
) {
2498 unsigned long current_block_start
=
2499 block_start_pfn(cc
->migrate_pfn
, cc
->order
);
2501 if (last_migrated_pfn
< current_block_start
) {
2502 lru_add_drain_cpu_zone(cc
->zone
);
2503 /* No more flushing until we migrate again */
2504 last_migrated_pfn
= 0;
2511 * Release free pages and update where the free scanner should restart,
2512 * so we don't leave any returned pages behind in the next attempt.
2514 if (cc
->nr_freepages
> 0) {
2515 unsigned long free_pfn
= release_freepages(&cc
->freepages
);
2517 cc
->nr_freepages
= 0;
2518 VM_BUG_ON(free_pfn
== 0);
2519 /* The cached pfn is always the first in a pageblock */
2520 free_pfn
= pageblock_start_pfn(free_pfn
);
2522 * Only go back, not forward. The cached pfn might have been
2523 * already reset to zone end in compact_finished()
2525 if (free_pfn
> cc
->zone
->compact_cached_free_pfn
)
2526 cc
->zone
->compact_cached_free_pfn
= free_pfn
;
2529 count_compact_events(COMPACTMIGRATE_SCANNED
, cc
->total_migrate_scanned
);
2530 count_compact_events(COMPACTFREE_SCANNED
, cc
->total_free_scanned
);
2532 trace_mm_compaction_end(cc
, start_pfn
, end_pfn
, sync
, ret
);
2534 VM_BUG_ON(!list_empty(&cc
->freepages
));
2535 VM_BUG_ON(!list_empty(&cc
->migratepages
));
2540 static enum compact_result
compact_zone_order(struct zone
*zone
, int order
,
2541 gfp_t gfp_mask
, enum compact_priority prio
,
2542 unsigned int alloc_flags
, int highest_zoneidx
,
2543 struct page
**capture
)
2545 enum compact_result ret
;
2546 struct compact_control cc
= {
2548 .search_order
= order
,
2549 .gfp_mask
= gfp_mask
,
2551 .mode
= (prio
== COMPACT_PRIO_ASYNC
) ?
2552 MIGRATE_ASYNC
: MIGRATE_SYNC_LIGHT
,
2553 .alloc_flags
= alloc_flags
,
2554 .highest_zoneidx
= highest_zoneidx
,
2555 .direct_compaction
= true,
2556 .whole_zone
= (prio
== MIN_COMPACT_PRIORITY
),
2557 .ignore_skip_hint
= (prio
== MIN_COMPACT_PRIORITY
),
2558 .ignore_block_suitable
= (prio
== MIN_COMPACT_PRIORITY
)
2560 struct capture_control capc
= {
2566 * Make sure the structs are really initialized before we expose the
2567 * capture control, in case we are interrupted and the interrupt handler
2571 WRITE_ONCE(current
->capture_control
, &capc
);
2573 ret
= compact_zone(&cc
, &capc
);
2576 * Make sure we hide capture control first before we read the captured
2577 * page pointer, otherwise an interrupt could free and capture a page
2578 * and we would leak it.
2580 WRITE_ONCE(current
->capture_control
, NULL
);
2581 *capture
= READ_ONCE(capc
.page
);
2583 * Technically, it is also possible that compaction is skipped but
2584 * the page is still captured out of luck(IRQ came and freed the page).
2585 * Returning COMPACT_SUCCESS in such cases helps in properly accounting
2586 * the COMPACT[STALL|FAIL] when compaction is skipped.
2589 ret
= COMPACT_SUCCESS
;
2595 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
2596 * @gfp_mask: The GFP mask of the current allocation
2597 * @order: The order of the current allocation
2598 * @alloc_flags: The allocation flags of the current allocation
2599 * @ac: The context of current allocation
2600 * @prio: Determines how hard direct compaction should try to succeed
2601 * @capture: Pointer to free page created by compaction will be stored here
2603 * This is the main entry point for direct page compaction.
2605 enum compact_result
try_to_compact_pages(gfp_t gfp_mask
, unsigned int order
,
2606 unsigned int alloc_flags
, const struct alloc_context
*ac
,
2607 enum compact_priority prio
, struct page
**capture
)
2609 int may_perform_io
= (__force
int)(gfp_mask
& __GFP_IO
);
2612 enum compact_result rc
= COMPACT_SKIPPED
;
2615 * Check if the GFP flags allow compaction - GFP_NOIO is really
2616 * tricky context because the migration might require IO
2618 if (!may_perform_io
)
2619 return COMPACT_SKIPPED
;
2621 trace_mm_compaction_try_to_compact_pages(order
, gfp_mask
, prio
);
2623 /* Compact each zone in the list */
2624 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
2625 ac
->highest_zoneidx
, ac
->nodemask
) {
2626 enum compact_result status
;
2628 if (prio
> MIN_COMPACT_PRIORITY
2629 && compaction_deferred(zone
, order
)) {
2630 rc
= max_t(enum compact_result
, COMPACT_DEFERRED
, rc
);
2634 status
= compact_zone_order(zone
, order
, gfp_mask
, prio
,
2635 alloc_flags
, ac
->highest_zoneidx
, capture
);
2636 rc
= max(status
, rc
);
2638 /* The allocation should succeed, stop compacting */
2639 if (status
== COMPACT_SUCCESS
) {
2641 * We think the allocation will succeed in this zone,
2642 * but it is not certain, hence the false. The caller
2643 * will repeat this with true if allocation indeed
2644 * succeeds in this zone.
2646 compaction_defer_reset(zone
, order
, false);
2651 if (prio
!= COMPACT_PRIO_ASYNC
&& (status
== COMPACT_COMPLETE
||
2652 status
== COMPACT_PARTIAL_SKIPPED
))
2654 * We think that allocation won't succeed in this zone
2655 * so we defer compaction there. If it ends up
2656 * succeeding after all, it will be reset.
2658 defer_compaction(zone
, order
);
2661 * We might have stopped compacting due to need_resched() in
2662 * async compaction, or due to a fatal signal detected. In that
2663 * case do not try further zones
2665 if ((prio
== COMPACT_PRIO_ASYNC
&& need_resched())
2666 || fatal_signal_pending(current
))
2674 * Compact all zones within a node till each zone's fragmentation score
2675 * reaches within proactive compaction thresholds (as determined by the
2676 * proactiveness tunable).
2678 * It is possible that the function returns before reaching score targets
2679 * due to various back-off conditions, such as, contention on per-node or
2682 static void proactive_compact_node(pg_data_t
*pgdat
)
2686 struct compact_control cc
= {
2688 .mode
= MIGRATE_SYNC_LIGHT
,
2689 .ignore_skip_hint
= true,
2691 .gfp_mask
= GFP_KERNEL
,
2692 .proactive_compaction
= true,
2695 for (zoneid
= 0; zoneid
< MAX_NR_ZONES
; zoneid
++) {
2696 zone
= &pgdat
->node_zones
[zoneid
];
2697 if (!populated_zone(zone
))
2702 compact_zone(&cc
, NULL
);
2704 count_compact_events(KCOMPACTD_MIGRATE_SCANNED
,
2705 cc
.total_migrate_scanned
);
2706 count_compact_events(KCOMPACTD_FREE_SCANNED
,
2707 cc
.total_free_scanned
);
2711 /* Compact all zones within a node */
2712 static void compact_node(int nid
)
2714 pg_data_t
*pgdat
= NODE_DATA(nid
);
2717 struct compact_control cc
= {
2719 .mode
= MIGRATE_SYNC
,
2720 .ignore_skip_hint
= true,
2722 .gfp_mask
= GFP_KERNEL
,
2726 for (zoneid
= 0; zoneid
< MAX_NR_ZONES
; zoneid
++) {
2728 zone
= &pgdat
->node_zones
[zoneid
];
2729 if (!populated_zone(zone
))
2734 compact_zone(&cc
, NULL
);
2738 /* Compact all nodes in the system */
2739 static void compact_nodes(void)
2743 /* Flush pending updates to the LRU lists */
2744 lru_add_drain_all();
2746 for_each_online_node(nid
)
2750 static int compaction_proactiveness_sysctl_handler(struct ctl_table
*table
, int write
,
2751 void *buffer
, size_t *length
, loff_t
*ppos
)
2755 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
2759 if (write
&& sysctl_compaction_proactiveness
) {
2760 for_each_online_node(nid
) {
2761 pg_data_t
*pgdat
= NODE_DATA(nid
);
2763 if (pgdat
->proactive_compact_trigger
)
2766 pgdat
->proactive_compact_trigger
= true;
2767 trace_mm_compaction_wakeup_kcompactd(pgdat
->node_id
, -1,
2768 pgdat
->nr_zones
- 1);
2769 wake_up_interruptible(&pgdat
->kcompactd_wait
);
2777 * This is the entry point for compacting all nodes via
2778 * /proc/sys/vm/compact_memory
2780 static int sysctl_compaction_handler(struct ctl_table
*table
, int write
,
2781 void *buffer
, size_t *length
, loff_t
*ppos
)
2789 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
2790 static ssize_t
compact_store(struct device
*dev
,
2791 struct device_attribute
*attr
,
2792 const char *buf
, size_t count
)
2796 if (nid
>= 0 && nid
< nr_node_ids
&& node_online(nid
)) {
2797 /* Flush pending updates to the LRU lists */
2798 lru_add_drain_all();
2805 static DEVICE_ATTR_WO(compact
);
2807 int compaction_register_node(struct node
*node
)
2809 return device_create_file(&node
->dev
, &dev_attr_compact
);
2812 void compaction_unregister_node(struct node
*node
)
2814 return device_remove_file(&node
->dev
, &dev_attr_compact
);
2816 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
2818 static inline bool kcompactd_work_requested(pg_data_t
*pgdat
)
2820 return pgdat
->kcompactd_max_order
> 0 || kthread_should_stop() ||
2821 pgdat
->proactive_compact_trigger
;
2824 static bool kcompactd_node_suitable(pg_data_t
*pgdat
)
2828 enum zone_type highest_zoneidx
= pgdat
->kcompactd_highest_zoneidx
;
2830 for (zoneid
= 0; zoneid
<= highest_zoneidx
; zoneid
++) {
2831 zone
= &pgdat
->node_zones
[zoneid
];
2833 if (!populated_zone(zone
))
2836 if (compaction_suitable(zone
, pgdat
->kcompactd_max_order
, 0,
2837 highest_zoneidx
) == COMPACT_CONTINUE
)
2844 static void kcompactd_do_work(pg_data_t
*pgdat
)
2847 * With no special task, compact all zones so that a page of requested
2848 * order is allocatable.
2852 struct compact_control cc
= {
2853 .order
= pgdat
->kcompactd_max_order
,
2854 .search_order
= pgdat
->kcompactd_max_order
,
2855 .highest_zoneidx
= pgdat
->kcompactd_highest_zoneidx
,
2856 .mode
= MIGRATE_SYNC_LIGHT
,
2857 .ignore_skip_hint
= false,
2858 .gfp_mask
= GFP_KERNEL
,
2860 trace_mm_compaction_kcompactd_wake(pgdat
->node_id
, cc
.order
,
2861 cc
.highest_zoneidx
);
2862 count_compact_event(KCOMPACTD_WAKE
);
2864 for (zoneid
= 0; zoneid
<= cc
.highest_zoneidx
; zoneid
++) {
2867 zone
= &pgdat
->node_zones
[zoneid
];
2868 if (!populated_zone(zone
))
2871 if (compaction_deferred(zone
, cc
.order
))
2874 if (compaction_suitable(zone
, cc
.order
, 0, zoneid
) !=
2878 if (kthread_should_stop())
2882 status
= compact_zone(&cc
, NULL
);
2884 if (status
== COMPACT_SUCCESS
) {
2885 compaction_defer_reset(zone
, cc
.order
, false);
2886 } else if (status
== COMPACT_PARTIAL_SKIPPED
|| status
== COMPACT_COMPLETE
) {
2888 * Buddy pages may become stranded on pcps that could
2889 * otherwise coalesce on the zone's free area for
2890 * order >= cc.order. This is ratelimited by the
2891 * upcoming deferral.
2893 drain_all_pages(zone
);
2896 * We use sync migration mode here, so we defer like
2897 * sync direct compaction does.
2899 defer_compaction(zone
, cc
.order
);
2902 count_compact_events(KCOMPACTD_MIGRATE_SCANNED
,
2903 cc
.total_migrate_scanned
);
2904 count_compact_events(KCOMPACTD_FREE_SCANNED
,
2905 cc
.total_free_scanned
);
2909 * Regardless of success, we are done until woken up next. But remember
2910 * the requested order/highest_zoneidx in case it was higher/tighter
2911 * than our current ones
2913 if (pgdat
->kcompactd_max_order
<= cc
.order
)
2914 pgdat
->kcompactd_max_order
= 0;
2915 if (pgdat
->kcompactd_highest_zoneidx
>= cc
.highest_zoneidx
)
2916 pgdat
->kcompactd_highest_zoneidx
= pgdat
->nr_zones
- 1;
2919 void wakeup_kcompactd(pg_data_t
*pgdat
, int order
, int highest_zoneidx
)
2924 if (pgdat
->kcompactd_max_order
< order
)
2925 pgdat
->kcompactd_max_order
= order
;
2927 if (pgdat
->kcompactd_highest_zoneidx
> highest_zoneidx
)
2928 pgdat
->kcompactd_highest_zoneidx
= highest_zoneidx
;
2931 * Pairs with implicit barrier in wait_event_freezable()
2932 * such that wakeups are not missed.
2934 if (!wq_has_sleeper(&pgdat
->kcompactd_wait
))
2937 if (!kcompactd_node_suitable(pgdat
))
2940 trace_mm_compaction_wakeup_kcompactd(pgdat
->node_id
, order
,
2942 wake_up_interruptible(&pgdat
->kcompactd_wait
);
2946 * The background compaction daemon, started as a kernel thread
2947 * from the init process.
2949 static int kcompactd(void *p
)
2951 pg_data_t
*pgdat
= (pg_data_t
*)p
;
2952 struct task_struct
*tsk
= current
;
2953 long default_timeout
= msecs_to_jiffies(HPAGE_FRAG_CHECK_INTERVAL_MSEC
);
2954 long timeout
= default_timeout
;
2956 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
2958 if (!cpumask_empty(cpumask
))
2959 set_cpus_allowed_ptr(tsk
, cpumask
);
2963 pgdat
->kcompactd_max_order
= 0;
2964 pgdat
->kcompactd_highest_zoneidx
= pgdat
->nr_zones
- 1;
2966 while (!kthread_should_stop()) {
2967 unsigned long pflags
;
2970 * Avoid the unnecessary wakeup for proactive compaction
2971 * when it is disabled.
2973 if (!sysctl_compaction_proactiveness
)
2974 timeout
= MAX_SCHEDULE_TIMEOUT
;
2975 trace_mm_compaction_kcompactd_sleep(pgdat
->node_id
);
2976 if (wait_event_freezable_timeout(pgdat
->kcompactd_wait
,
2977 kcompactd_work_requested(pgdat
), timeout
) &&
2978 !pgdat
->proactive_compact_trigger
) {
2980 psi_memstall_enter(&pflags
);
2981 kcompactd_do_work(pgdat
);
2982 psi_memstall_leave(&pflags
);
2984 * Reset the timeout value. The defer timeout from
2985 * proactive compaction is lost here but that is fine
2986 * as the condition of the zone changing substantionally
2987 * then carrying on with the previous defer interval is
2990 timeout
= default_timeout
;
2995 * Start the proactive work with default timeout. Based
2996 * on the fragmentation score, this timeout is updated.
2998 timeout
= default_timeout
;
2999 if (should_proactive_compact_node(pgdat
)) {
3000 unsigned int prev_score
, score
;
3002 prev_score
= fragmentation_score_node(pgdat
);
3003 proactive_compact_node(pgdat
);
3004 score
= fragmentation_score_node(pgdat
);
3006 * Defer proactive compaction if the fragmentation
3007 * score did not go down i.e. no progress made.
3009 if (unlikely(score
>= prev_score
))
3011 default_timeout
<< COMPACT_MAX_DEFER_SHIFT
;
3013 if (unlikely(pgdat
->proactive_compact_trigger
))
3014 pgdat
->proactive_compact_trigger
= false;
3021 * This kcompactd start function will be called by init and node-hot-add.
3022 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
3024 void kcompactd_run(int nid
)
3026 pg_data_t
*pgdat
= NODE_DATA(nid
);
3028 if (pgdat
->kcompactd
)
3031 pgdat
->kcompactd
= kthread_run(kcompactd
, pgdat
, "kcompactd%d", nid
);
3032 if (IS_ERR(pgdat
->kcompactd
)) {
3033 pr_err("Failed to start kcompactd on node %d\n", nid
);
3034 pgdat
->kcompactd
= NULL
;
3039 * Called by memory hotplug when all memory in a node is offlined. Caller must
3040 * be holding mem_hotplug_begin/done().
3042 void kcompactd_stop(int nid
)
3044 struct task_struct
*kcompactd
= NODE_DATA(nid
)->kcompactd
;
3047 kthread_stop(kcompactd
);
3048 NODE_DATA(nid
)->kcompactd
= NULL
;
3053 * It's optimal to keep kcompactd on the same CPUs as their memory, but
3054 * not required for correctness. So if the last cpu in a node goes
3055 * away, we get changed to run anywhere: as the first one comes back,
3056 * restore their cpu bindings.
3058 static int kcompactd_cpu_online(unsigned int cpu
)
3062 for_each_node_state(nid
, N_MEMORY
) {
3063 pg_data_t
*pgdat
= NODE_DATA(nid
);
3064 const struct cpumask
*mask
;
3066 mask
= cpumask_of_node(pgdat
->node_id
);
3068 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
3069 /* One of our CPUs online: restore mask */
3070 if (pgdat
->kcompactd
)
3071 set_cpus_allowed_ptr(pgdat
->kcompactd
, mask
);
3076 static int proc_dointvec_minmax_warn_RT_change(struct ctl_table
*table
,
3077 int write
, void *buffer
, size_t *lenp
, loff_t
*ppos
)
3081 if (!IS_ENABLED(CONFIG_PREEMPT_RT
) || !write
)
3082 return proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
3084 old
= *(int *)table
->data
;
3085 ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
3088 if (old
!= *(int *)table
->data
)
3089 pr_warn_once("sysctl attribute %s changed by %s[%d]\n",
3090 table
->procname
, current
->comm
,
3091 task_pid_nr(current
));
3095 static struct ctl_table vm_compaction
[] = {
3097 .procname
= "compact_memory",
3099 .maxlen
= sizeof(int),
3101 .proc_handler
= sysctl_compaction_handler
,
3104 .procname
= "compaction_proactiveness",
3105 .data
= &sysctl_compaction_proactiveness
,
3106 .maxlen
= sizeof(sysctl_compaction_proactiveness
),
3108 .proc_handler
= compaction_proactiveness_sysctl_handler
,
3109 .extra1
= SYSCTL_ZERO
,
3110 .extra2
= SYSCTL_ONE_HUNDRED
,
3113 .procname
= "extfrag_threshold",
3114 .data
= &sysctl_extfrag_threshold
,
3115 .maxlen
= sizeof(int),
3117 .proc_handler
= proc_dointvec_minmax
,
3118 .extra1
= SYSCTL_ZERO
,
3119 .extra2
= SYSCTL_ONE_THOUSAND
,
3122 .procname
= "compact_unevictable_allowed",
3123 .data
= &sysctl_compact_unevictable_allowed
,
3124 .maxlen
= sizeof(int),
3126 .proc_handler
= proc_dointvec_minmax_warn_RT_change
,
3127 .extra1
= SYSCTL_ZERO
,
3128 .extra2
= SYSCTL_ONE
,
3133 static int __init
kcompactd_init(void)
3138 ret
= cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN
,
3139 "mm/compaction:online",
3140 kcompactd_cpu_online
, NULL
);
3142 pr_err("kcompactd: failed to register hotplug callbacks.\n");
3146 for_each_node_state(nid
, N_MEMORY
)
3148 register_sysctl_init("vm", vm_compaction
);
3151 subsys_initcall(kcompactd_init
)
3153 #endif /* CONFIG_COMPACTION */