2 * linux/mm/compaction.c
4 * Memory compaction for the reduction of external fragmentation. Note that
5 * this heavily depends upon page migration to do all the real heavy
8 * Copyright IBM Corp. 2007-2010 Mel Gorman <mel@csn.ul.ie>
10 #include <linux/cpu.h>
11 #include <linux/swap.h>
12 #include <linux/migrate.h>
13 #include <linux/compaction.h>
14 #include <linux/mm_inline.h>
15 #include <linux/backing-dev.h>
16 #include <linux/sysctl.h>
17 #include <linux/sysfs.h>
18 #include <linux/balloon_compaction.h>
19 #include <linux/page-isolation.h>
20 #include <linux/kasan.h>
21 #include <linux/kthread.h>
22 #include <linux/freezer.h>
25 #ifdef CONFIG_COMPACTION
26 static inline void count_compact_event(enum vm_event_item item
)
31 static inline void count_compact_events(enum vm_event_item item
, long delta
)
33 count_vm_events(item
, delta
);
36 #define count_compact_event(item) do { } while (0)
37 #define count_compact_events(item, delta) do { } while (0)
40 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
42 #define CREATE_TRACE_POINTS
43 #include <trace/events/compaction.h>
45 #define block_start_pfn(pfn, order) round_down(pfn, 1UL << (order))
46 #define block_end_pfn(pfn, order) ALIGN((pfn) + 1, 1UL << (order))
47 #define pageblock_start_pfn(pfn) block_start_pfn(pfn, pageblock_order)
48 #define pageblock_end_pfn(pfn) block_end_pfn(pfn, pageblock_order)
50 static unsigned long release_freepages(struct list_head
*freelist
)
52 struct page
*page
, *next
;
53 unsigned long high_pfn
= 0;
55 list_for_each_entry_safe(page
, next
, freelist
, lru
) {
56 unsigned long pfn
= page_to_pfn(page
);
66 static void map_pages(struct list_head
*list
)
70 list_for_each_entry(page
, list
, lru
) {
71 arch_alloc_page(page
, 0);
72 kernel_map_pages(page
, 1, 1);
73 kasan_alloc_pages(page
, 0);
77 static inline bool migrate_async_suitable(int migratetype
)
79 return is_migrate_cma(migratetype
) || migratetype
== MIGRATE_MOVABLE
;
82 #ifdef CONFIG_COMPACTION
84 int PageMovable(struct page
*page
)
86 struct address_space
*mapping
;
88 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
89 if (!__PageMovable(page
))
92 mapping
= page_mapping(page
);
93 if (mapping
&& mapping
->a_ops
&& mapping
->a_ops
->isolate_page
)
98 EXPORT_SYMBOL(PageMovable
);
100 void __SetPageMovable(struct page
*page
, struct address_space
*mapping
)
102 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
103 VM_BUG_ON_PAGE((unsigned long)mapping
& PAGE_MAPPING_MOVABLE
, page
);
104 page
->mapping
= (void *)((unsigned long)mapping
| PAGE_MAPPING_MOVABLE
);
106 EXPORT_SYMBOL(__SetPageMovable
);
108 void __ClearPageMovable(struct page
*page
)
110 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
111 VM_BUG_ON_PAGE(!PageMovable(page
), page
);
113 * Clear registered address_space val with keeping PAGE_MAPPING_MOVABLE
114 * flag so that VM can catch up released page by driver after isolation.
115 * With it, VM migration doesn't try to put it back.
117 page
->mapping
= (void *)((unsigned long)page
->mapping
&
118 PAGE_MAPPING_MOVABLE
);
120 EXPORT_SYMBOL(__ClearPageMovable
);
122 /* Do not skip compaction more than 64 times */
123 #define COMPACT_MAX_DEFER_SHIFT 6
126 * Compaction is deferred when compaction fails to result in a page
127 * allocation success. 1 << compact_defer_limit compactions are skipped up
128 * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
130 void defer_compaction(struct zone
*zone
, int order
)
132 zone
->compact_considered
= 0;
133 zone
->compact_defer_shift
++;
135 if (order
< zone
->compact_order_failed
)
136 zone
->compact_order_failed
= order
;
138 if (zone
->compact_defer_shift
> COMPACT_MAX_DEFER_SHIFT
)
139 zone
->compact_defer_shift
= COMPACT_MAX_DEFER_SHIFT
;
141 trace_mm_compaction_defer_compaction(zone
, order
);
144 /* Returns true if compaction should be skipped this time */
145 bool compaction_deferred(struct zone
*zone
, int order
)
147 unsigned long defer_limit
= 1UL << zone
->compact_defer_shift
;
149 if (order
< zone
->compact_order_failed
)
152 /* Avoid possible overflow */
153 if (++zone
->compact_considered
> defer_limit
)
154 zone
->compact_considered
= defer_limit
;
156 if (zone
->compact_considered
>= defer_limit
)
159 trace_mm_compaction_deferred(zone
, order
);
165 * Update defer tracking counters after successful compaction of given order,
166 * which means an allocation either succeeded (alloc_success == true) or is
167 * expected to succeed.
169 void compaction_defer_reset(struct zone
*zone
, int order
,
173 zone
->compact_considered
= 0;
174 zone
->compact_defer_shift
= 0;
176 if (order
>= zone
->compact_order_failed
)
177 zone
->compact_order_failed
= order
+ 1;
179 trace_mm_compaction_defer_reset(zone
, order
);
182 /* Returns true if restarting compaction after many failures */
183 bool compaction_restarting(struct zone
*zone
, int order
)
185 if (order
< zone
->compact_order_failed
)
188 return zone
->compact_defer_shift
== COMPACT_MAX_DEFER_SHIFT
&&
189 zone
->compact_considered
>= 1UL << zone
->compact_defer_shift
;
192 /* Returns true if the pageblock should be scanned for pages to isolate. */
193 static inline bool isolation_suitable(struct compact_control
*cc
,
196 if (cc
->ignore_skip_hint
)
199 return !get_pageblock_skip(page
);
202 static void reset_cached_positions(struct zone
*zone
)
204 zone
->compact_cached_migrate_pfn
[0] = zone
->zone_start_pfn
;
205 zone
->compact_cached_migrate_pfn
[1] = zone
->zone_start_pfn
;
206 zone
->compact_cached_free_pfn
=
207 pageblock_start_pfn(zone_end_pfn(zone
) - 1);
211 * This function is called to clear all cached information on pageblocks that
212 * should be skipped for page isolation when the migrate and free page scanner
215 static void __reset_isolation_suitable(struct zone
*zone
)
217 unsigned long start_pfn
= zone
->zone_start_pfn
;
218 unsigned long end_pfn
= zone_end_pfn(zone
);
221 zone
->compact_blockskip_flush
= false;
223 /* Walk the zone and mark every pageblock as suitable for isolation */
224 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
232 page
= pfn_to_page(pfn
);
233 if (zone
!= page_zone(page
))
236 clear_pageblock_skip(page
);
239 reset_cached_positions(zone
);
242 void reset_isolation_suitable(pg_data_t
*pgdat
)
246 for (zoneid
= 0; zoneid
< MAX_NR_ZONES
; zoneid
++) {
247 struct zone
*zone
= &pgdat
->node_zones
[zoneid
];
248 if (!populated_zone(zone
))
251 /* Only flush if a full compaction finished recently */
252 if (zone
->compact_blockskip_flush
)
253 __reset_isolation_suitable(zone
);
258 * If no pages were isolated then mark this pageblock to be skipped in the
259 * future. The information is later cleared by __reset_isolation_suitable().
261 static void update_pageblock_skip(struct compact_control
*cc
,
262 struct page
*page
, unsigned long nr_isolated
,
263 bool migrate_scanner
)
265 struct zone
*zone
= cc
->zone
;
268 if (cc
->ignore_skip_hint
)
277 set_pageblock_skip(page
);
279 pfn
= page_to_pfn(page
);
281 /* Update where async and sync compaction should restart */
282 if (migrate_scanner
) {
283 if (pfn
> zone
->compact_cached_migrate_pfn
[0])
284 zone
->compact_cached_migrate_pfn
[0] = pfn
;
285 if (cc
->mode
!= MIGRATE_ASYNC
&&
286 pfn
> zone
->compact_cached_migrate_pfn
[1])
287 zone
->compact_cached_migrate_pfn
[1] = pfn
;
289 if (pfn
< zone
->compact_cached_free_pfn
)
290 zone
->compact_cached_free_pfn
= pfn
;
294 static inline bool isolation_suitable(struct compact_control
*cc
,
300 static void update_pageblock_skip(struct compact_control
*cc
,
301 struct page
*page
, unsigned long nr_isolated
,
302 bool migrate_scanner
)
305 #endif /* CONFIG_COMPACTION */
308 * Compaction requires the taking of some coarse locks that are potentially
309 * very heavily contended. For async compaction, back out if the lock cannot
310 * be taken immediately. For sync compaction, spin on the lock if needed.
312 * Returns true if the lock is held
313 * Returns false if the lock is not held and compaction should abort
315 static bool compact_trylock_irqsave(spinlock_t
*lock
, unsigned long *flags
,
316 struct compact_control
*cc
)
318 if (cc
->mode
== MIGRATE_ASYNC
) {
319 if (!spin_trylock_irqsave(lock
, *flags
)) {
320 cc
->contended
= COMPACT_CONTENDED_LOCK
;
324 spin_lock_irqsave(lock
, *flags
);
331 * Compaction requires the taking of some coarse locks that are potentially
332 * very heavily contended. The lock should be periodically unlocked to avoid
333 * having disabled IRQs for a long time, even when there is nobody waiting on
334 * the lock. It might also be that allowing the IRQs will result in
335 * need_resched() becoming true. If scheduling is needed, async compaction
336 * aborts. Sync compaction schedules.
337 * Either compaction type will also abort if a fatal signal is pending.
338 * In either case if the lock was locked, it is dropped and not regained.
340 * Returns true if compaction should abort due to fatal signal pending, or
341 * async compaction due to need_resched()
342 * Returns false when compaction can continue (sync compaction might have
345 static bool compact_unlock_should_abort(spinlock_t
*lock
,
346 unsigned long flags
, bool *locked
, struct compact_control
*cc
)
349 spin_unlock_irqrestore(lock
, flags
);
353 if (fatal_signal_pending(current
)) {
354 cc
->contended
= COMPACT_CONTENDED_SCHED
;
358 if (need_resched()) {
359 if (cc
->mode
== MIGRATE_ASYNC
) {
360 cc
->contended
= COMPACT_CONTENDED_SCHED
;
370 * Aside from avoiding lock contention, compaction also periodically checks
371 * need_resched() and either schedules in sync compaction or aborts async
372 * compaction. This is similar to what compact_unlock_should_abort() does, but
373 * is used where no lock is concerned.
375 * Returns false when no scheduling was needed, or sync compaction scheduled.
376 * Returns true when async compaction should abort.
378 static inline bool compact_should_abort(struct compact_control
*cc
)
380 /* async compaction aborts if contended */
381 if (need_resched()) {
382 if (cc
->mode
== MIGRATE_ASYNC
) {
383 cc
->contended
= COMPACT_CONTENDED_SCHED
;
394 * Isolate free pages onto a private freelist. If @strict is true, will abort
395 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
396 * (even though it may still end up isolating some pages).
398 static unsigned long isolate_freepages_block(struct compact_control
*cc
,
399 unsigned long *start_pfn
,
400 unsigned long end_pfn
,
401 struct list_head
*freelist
,
404 int nr_scanned
= 0, total_isolated
= 0;
405 struct page
*cursor
, *valid_page
= NULL
;
406 unsigned long flags
= 0;
408 unsigned long blockpfn
= *start_pfn
;
410 cursor
= pfn_to_page(blockpfn
);
412 /* Isolate free pages. */
413 for (; blockpfn
< end_pfn
; blockpfn
++, cursor
++) {
415 struct page
*page
= cursor
;
418 * Periodically drop the lock (if held) regardless of its
419 * contention, to give chance to IRQs. Abort if fatal signal
420 * pending or async compaction detects need_resched()
422 if (!(blockpfn
% SWAP_CLUSTER_MAX
)
423 && compact_unlock_should_abort(&cc
->zone
->lock
, flags
,
428 if (!pfn_valid_within(blockpfn
))
435 * For compound pages such as THP and hugetlbfs, we can save
436 * potentially a lot of iterations if we skip them at once.
437 * The check is racy, but we can consider only valid values
438 * and the only danger is skipping too much.
440 if (PageCompound(page
)) {
441 unsigned int comp_order
= compound_order(page
);
443 if (likely(comp_order
< MAX_ORDER
)) {
444 blockpfn
+= (1UL << comp_order
) - 1;
445 cursor
+= (1UL << comp_order
) - 1;
451 if (!PageBuddy(page
))
455 * If we already hold the lock, we can skip some rechecking.
456 * Note that if we hold the lock now, checked_pageblock was
457 * already set in some previous iteration (or strict is true),
458 * so it is correct to skip the suitable migration target
463 * The zone lock must be held to isolate freepages.
464 * Unfortunately this is a very coarse lock and can be
465 * heavily contended if there are parallel allocations
466 * or parallel compactions. For async compaction do not
467 * spin on the lock and we acquire the lock as late as
470 locked
= compact_trylock_irqsave(&cc
->zone
->lock
,
475 /* Recheck this is a buddy page under lock */
476 if (!PageBuddy(page
))
480 /* Found a free page, break it into order-0 pages */
481 isolated
= split_free_page(page
);
485 total_isolated
+= isolated
;
486 cc
->nr_freepages
+= isolated
;
487 for (i
= 0; i
< isolated
; i
++) {
488 list_add(&page
->lru
, freelist
);
491 if (!strict
&& cc
->nr_migratepages
<= cc
->nr_freepages
) {
492 blockpfn
+= isolated
;
495 /* Advance to the end of split page */
496 blockpfn
+= isolated
- 1;
497 cursor
+= isolated
- 1;
509 spin_unlock_irqrestore(&cc
->zone
->lock
, flags
);
512 * There is a tiny chance that we have read bogus compound_order(),
513 * so be careful to not go outside of the pageblock.
515 if (unlikely(blockpfn
> end_pfn
))
518 trace_mm_compaction_isolate_freepages(*start_pfn
, blockpfn
,
519 nr_scanned
, total_isolated
);
521 /* Record how far we have got within the block */
522 *start_pfn
= blockpfn
;
525 * If strict isolation is requested by CMA then check that all the
526 * pages requested were isolated. If there were any failures, 0 is
527 * returned and CMA will fail.
529 if (strict
&& blockpfn
< end_pfn
)
532 /* Update the pageblock-skip if the whole pageblock was scanned */
533 if (blockpfn
== end_pfn
)
534 update_pageblock_skip(cc
, valid_page
, total_isolated
, false);
536 count_compact_events(COMPACTFREE_SCANNED
, nr_scanned
);
538 count_compact_events(COMPACTISOLATED
, total_isolated
);
539 return total_isolated
;
543 * isolate_freepages_range() - isolate free pages.
544 * @start_pfn: The first PFN to start isolating.
545 * @end_pfn: The one-past-last PFN.
547 * Non-free pages, invalid PFNs, or zone boundaries within the
548 * [start_pfn, end_pfn) range are considered errors, cause function to
549 * undo its actions and return zero.
551 * Otherwise, function returns one-past-the-last PFN of isolated page
552 * (which may be greater then end_pfn if end fell in a middle of
556 isolate_freepages_range(struct compact_control
*cc
,
557 unsigned long start_pfn
, unsigned long end_pfn
)
559 unsigned long isolated
, pfn
, block_start_pfn
, block_end_pfn
;
563 block_start_pfn
= pageblock_start_pfn(pfn
);
564 if (block_start_pfn
< cc
->zone
->zone_start_pfn
)
565 block_start_pfn
= cc
->zone
->zone_start_pfn
;
566 block_end_pfn
= pageblock_end_pfn(pfn
);
568 for (; pfn
< end_pfn
; pfn
+= isolated
,
569 block_start_pfn
= block_end_pfn
,
570 block_end_pfn
+= pageblock_nr_pages
) {
571 /* Protect pfn from changing by isolate_freepages_block */
572 unsigned long isolate_start_pfn
= pfn
;
574 block_end_pfn
= min(block_end_pfn
, end_pfn
);
577 * pfn could pass the block_end_pfn if isolated freepage
578 * is more than pageblock order. In this case, we adjust
579 * scanning range to right one.
581 if (pfn
>= block_end_pfn
) {
582 block_start_pfn
= pageblock_start_pfn(pfn
);
583 block_end_pfn
= pageblock_end_pfn(pfn
);
584 block_end_pfn
= min(block_end_pfn
, end_pfn
);
587 if (!pageblock_pfn_to_page(block_start_pfn
,
588 block_end_pfn
, cc
->zone
))
591 isolated
= isolate_freepages_block(cc
, &isolate_start_pfn
,
592 block_end_pfn
, &freelist
, true);
595 * In strict mode, isolate_freepages_block() returns 0 if
596 * there are any holes in the block (ie. invalid PFNs or
603 * If we managed to isolate pages, it is always (1 << n) *
604 * pageblock_nr_pages for some non-negative n. (Max order
605 * page may span two pageblocks).
609 /* split_free_page does not map the pages */
610 map_pages(&freelist
);
613 /* Loop terminated early, cleanup. */
614 release_freepages(&freelist
);
618 /* We don't use freelists for anything. */
622 /* Update the number of anon and file isolated pages in the zone */
623 static void acct_isolated(struct zone
*zone
, struct compact_control
*cc
)
626 unsigned int count
[2] = { 0, };
628 if (list_empty(&cc
->migratepages
))
631 list_for_each_entry(page
, &cc
->migratepages
, lru
)
632 count
[!!page_is_file_cache(page
)]++;
634 mod_zone_page_state(zone
, NR_ISOLATED_ANON
, count
[0]);
635 mod_zone_page_state(zone
, NR_ISOLATED_FILE
, count
[1]);
638 /* Similar to reclaim, but different enough that they don't share logic */
639 static bool too_many_isolated(struct zone
*zone
)
641 unsigned long active
, inactive
, isolated
;
643 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
) +
644 zone_page_state(zone
, NR_INACTIVE_ANON
);
645 active
= zone_page_state(zone
, NR_ACTIVE_FILE
) +
646 zone_page_state(zone
, NR_ACTIVE_ANON
);
647 isolated
= zone_page_state(zone
, NR_ISOLATED_FILE
) +
648 zone_page_state(zone
, NR_ISOLATED_ANON
);
650 return isolated
> (inactive
+ active
) / 2;
654 * isolate_migratepages_block() - isolate all migrate-able pages within
656 * @cc: Compaction control structure.
657 * @low_pfn: The first PFN to isolate
658 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
659 * @isolate_mode: Isolation mode to be used.
661 * Isolate all pages that can be migrated from the range specified by
662 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
663 * Returns zero if there is a fatal signal pending, otherwise PFN of the
664 * first page that was not scanned (which may be both less, equal to or more
667 * The pages are isolated on cc->migratepages list (not required to be empty),
668 * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
669 * is neither read nor updated.
672 isolate_migratepages_block(struct compact_control
*cc
, unsigned long low_pfn
,
673 unsigned long end_pfn
, isolate_mode_t isolate_mode
)
675 struct zone
*zone
= cc
->zone
;
676 unsigned long nr_scanned
= 0, nr_isolated
= 0;
677 struct lruvec
*lruvec
;
678 unsigned long flags
= 0;
680 struct page
*page
= NULL
, *valid_page
= NULL
;
681 unsigned long start_pfn
= low_pfn
;
682 bool skip_on_failure
= false;
683 unsigned long next_skip_pfn
= 0;
686 * Ensure that there are not too many pages isolated from the LRU
687 * list by either parallel reclaimers or compaction. If there are,
688 * delay for some time until fewer pages are isolated
690 while (unlikely(too_many_isolated(zone
))) {
691 /* async migration should just abort */
692 if (cc
->mode
== MIGRATE_ASYNC
)
695 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
697 if (fatal_signal_pending(current
))
701 if (compact_should_abort(cc
))
704 if (cc
->direct_compaction
&& (cc
->mode
== MIGRATE_ASYNC
)) {
705 skip_on_failure
= true;
706 next_skip_pfn
= block_end_pfn(low_pfn
, cc
->order
);
709 /* Time to isolate some pages for migration */
710 for (; low_pfn
< end_pfn
; low_pfn
++) {
712 if (skip_on_failure
&& low_pfn
>= next_skip_pfn
) {
714 * We have isolated all migration candidates in the
715 * previous order-aligned block, and did not skip it due
716 * to failure. We should migrate the pages now and
717 * hopefully succeed compaction.
723 * We failed to isolate in the previous order-aligned
724 * block. Set the new boundary to the end of the
725 * current block. Note we can't simply increase
726 * next_skip_pfn by 1 << order, as low_pfn might have
727 * been incremented by a higher number due to skipping
728 * a compound or a high-order buddy page in the
729 * previous loop iteration.
731 next_skip_pfn
= block_end_pfn(low_pfn
, cc
->order
);
735 * Periodically drop the lock (if held) regardless of its
736 * contention, to give chance to IRQs. Abort async compaction
739 if (!(low_pfn
% SWAP_CLUSTER_MAX
)
740 && compact_unlock_should_abort(&zone
->lru_lock
, flags
,
744 if (!pfn_valid_within(low_pfn
))
748 page
= pfn_to_page(low_pfn
);
754 * Skip if free. We read page order here without zone lock
755 * which is generally unsafe, but the race window is small and
756 * the worst thing that can happen is that we skip some
757 * potential isolation targets.
759 if (PageBuddy(page
)) {
760 unsigned long freepage_order
= page_order_unsafe(page
);
763 * Without lock, we cannot be sure that what we got is
764 * a valid page order. Consider only values in the
765 * valid order range to prevent low_pfn overflow.
767 if (freepage_order
> 0 && freepage_order
< MAX_ORDER
)
768 low_pfn
+= (1UL << freepage_order
) - 1;
773 * Regardless of being on LRU, compound pages such as THP and
774 * hugetlbfs are not to be compacted. We can potentially save
775 * a lot of iterations if we skip them at once. The check is
776 * racy, but we can consider only valid values and the only
777 * danger is skipping too much.
779 if (PageCompound(page
)) {
780 unsigned int comp_order
= compound_order(page
);
782 if (likely(comp_order
< MAX_ORDER
))
783 low_pfn
+= (1UL << comp_order
) - 1;
789 * Check may be lockless but that's ok as we recheck later.
790 * It's possible to migrate LRU and non-lru movable pages.
791 * Skip any other type of page
793 if (!PageLRU(page
)) {
794 if (unlikely(balloon_page_movable(page
))) {
795 if (balloon_page_isolate(page
)) {
796 /* Successfully isolated */
797 goto isolate_success
;
802 * __PageMovable can return false positive so we need
803 * to verify it under page_lock.
805 if (unlikely(__PageMovable(page
)) &&
806 !PageIsolated(page
)) {
808 spin_unlock_irqrestore(&zone
->lru_lock
,
813 if (isolate_movable_page(page
, isolate_mode
))
814 goto isolate_success
;
821 * Migration will fail if an anonymous page is pinned in memory,
822 * so avoid taking lru_lock and isolating it unnecessarily in an
823 * admittedly racy check.
825 if (!page_mapping(page
) &&
826 page_count(page
) > page_mapcount(page
))
829 /* If we already hold the lock, we can skip some rechecking */
831 locked
= compact_trylock_irqsave(&zone
->lru_lock
,
836 /* Recheck PageLRU and PageCompound under lock */
841 * Page become compound since the non-locked check,
842 * and it's on LRU. It can only be a THP so the order
843 * is safe to read and it's 0 for tail pages.
845 if (unlikely(PageCompound(page
))) {
846 low_pfn
+= (1UL << compound_order(page
)) - 1;
851 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
853 /* Try isolate the page */
854 if (__isolate_lru_page(page
, isolate_mode
) != 0)
857 VM_BUG_ON_PAGE(PageCompound(page
), page
);
859 /* Successfully isolated */
860 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
863 list_add(&page
->lru
, &cc
->migratepages
);
864 cc
->nr_migratepages
++;
868 * Record where we could have freed pages by migration and not
869 * yet flushed them to buddy allocator.
870 * - this is the lowest page that was isolated and likely be
871 * then freed by migration.
873 if (!cc
->last_migrated_pfn
)
874 cc
->last_migrated_pfn
= low_pfn
;
876 /* Avoid isolating too much */
877 if (cc
->nr_migratepages
== COMPACT_CLUSTER_MAX
) {
884 if (!skip_on_failure
)
888 * We have isolated some pages, but then failed. Release them
889 * instead of migrating, as we cannot form the cc->order buddy
894 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
897 acct_isolated(zone
, cc
);
898 putback_movable_pages(&cc
->migratepages
);
899 cc
->nr_migratepages
= 0;
900 cc
->last_migrated_pfn
= 0;
904 if (low_pfn
< next_skip_pfn
) {
905 low_pfn
= next_skip_pfn
- 1;
907 * The check near the loop beginning would have updated
908 * next_skip_pfn too, but this is a bit simpler.
910 next_skip_pfn
+= 1UL << cc
->order
;
915 * The PageBuddy() check could have potentially brought us outside
916 * the range to be scanned.
918 if (unlikely(low_pfn
> end_pfn
))
922 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
925 * Update the pageblock-skip information and cached scanner pfn,
926 * if the whole pageblock was scanned without isolating any page.
928 if (low_pfn
== end_pfn
)
929 update_pageblock_skip(cc
, valid_page
, nr_isolated
, true);
931 trace_mm_compaction_isolate_migratepages(start_pfn
, low_pfn
,
932 nr_scanned
, nr_isolated
);
934 count_compact_events(COMPACTMIGRATE_SCANNED
, nr_scanned
);
936 count_compact_events(COMPACTISOLATED
, nr_isolated
);
942 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
943 * @cc: Compaction control structure.
944 * @start_pfn: The first PFN to start isolating.
945 * @end_pfn: The one-past-last PFN.
947 * Returns zero if isolation fails fatally due to e.g. pending signal.
948 * Otherwise, function returns one-past-the-last PFN of isolated page
949 * (which may be greater than end_pfn if end fell in a middle of a THP page).
952 isolate_migratepages_range(struct compact_control
*cc
, unsigned long start_pfn
,
953 unsigned long end_pfn
)
955 unsigned long pfn
, block_start_pfn
, block_end_pfn
;
957 /* Scan block by block. First and last block may be incomplete */
959 block_start_pfn
= pageblock_start_pfn(pfn
);
960 if (block_start_pfn
< cc
->zone
->zone_start_pfn
)
961 block_start_pfn
= cc
->zone
->zone_start_pfn
;
962 block_end_pfn
= pageblock_end_pfn(pfn
);
964 for (; pfn
< end_pfn
; pfn
= block_end_pfn
,
965 block_start_pfn
= block_end_pfn
,
966 block_end_pfn
+= pageblock_nr_pages
) {
968 block_end_pfn
= min(block_end_pfn
, end_pfn
);
970 if (!pageblock_pfn_to_page(block_start_pfn
,
971 block_end_pfn
, cc
->zone
))
974 pfn
= isolate_migratepages_block(cc
, pfn
, block_end_pfn
,
975 ISOLATE_UNEVICTABLE
);
980 if (cc
->nr_migratepages
== COMPACT_CLUSTER_MAX
)
983 acct_isolated(cc
->zone
, cc
);
988 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
989 #ifdef CONFIG_COMPACTION
991 /* Returns true if the page is within a block suitable for migration to */
992 static bool suitable_migration_target(struct page
*page
)
994 /* If the page is a large free page, then disallow migration */
995 if (PageBuddy(page
)) {
997 * We are checking page_order without zone->lock taken. But
998 * the only small danger is that we skip a potentially suitable
999 * pageblock, so it's not worth to check order for valid range.
1001 if (page_order_unsafe(page
) >= pageblock_order
)
1005 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1006 if (migrate_async_suitable(get_pageblock_migratetype(page
)))
1009 /* Otherwise skip the block */
1014 * Test whether the free scanner has reached the same or lower pageblock than
1015 * the migration scanner, and compaction should thus terminate.
1017 static inline bool compact_scanners_met(struct compact_control
*cc
)
1019 return (cc
->free_pfn
>> pageblock_order
)
1020 <= (cc
->migrate_pfn
>> pageblock_order
);
1024 * Based on information in the current compact_control, find blocks
1025 * suitable for isolating free pages from and then isolate them.
1027 static void isolate_freepages(struct compact_control
*cc
)
1029 struct zone
*zone
= cc
->zone
;
1031 unsigned long block_start_pfn
; /* start of current pageblock */
1032 unsigned long isolate_start_pfn
; /* exact pfn we start at */
1033 unsigned long block_end_pfn
; /* end of current pageblock */
1034 unsigned long low_pfn
; /* lowest pfn scanner is able to scan */
1035 struct list_head
*freelist
= &cc
->freepages
;
1038 * Initialise the free scanner. The starting point is where we last
1039 * successfully isolated from, zone-cached value, or the end of the
1040 * zone when isolating for the first time. For looping we also need
1041 * this pfn aligned down to the pageblock boundary, because we do
1042 * block_start_pfn -= pageblock_nr_pages in the for loop.
1043 * For ending point, take care when isolating in last pageblock of a
1044 * a zone which ends in the middle of a pageblock.
1045 * The low boundary is the end of the pageblock the migration scanner
1048 isolate_start_pfn
= cc
->free_pfn
;
1049 block_start_pfn
= pageblock_start_pfn(cc
->free_pfn
);
1050 block_end_pfn
= min(block_start_pfn
+ pageblock_nr_pages
,
1051 zone_end_pfn(zone
));
1052 low_pfn
= pageblock_end_pfn(cc
->migrate_pfn
);
1055 * Isolate free pages until enough are available to migrate the
1056 * pages on cc->migratepages. We stop searching if the migrate
1057 * and free page scanners meet or enough free pages are isolated.
1059 for (; block_start_pfn
>= low_pfn
;
1060 block_end_pfn
= block_start_pfn
,
1061 block_start_pfn
-= pageblock_nr_pages
,
1062 isolate_start_pfn
= block_start_pfn
) {
1064 * This can iterate a massively long zone without finding any
1065 * suitable migration targets, so periodically check if we need
1066 * to schedule, or even abort async compaction.
1068 if (!(block_start_pfn
% (SWAP_CLUSTER_MAX
* pageblock_nr_pages
))
1069 && compact_should_abort(cc
))
1072 page
= pageblock_pfn_to_page(block_start_pfn
, block_end_pfn
,
1077 /* Check the block is suitable for migration */
1078 if (!suitable_migration_target(page
))
1081 /* If isolation recently failed, do not retry */
1082 if (!isolation_suitable(cc
, page
))
1085 /* Found a block suitable for isolating free pages from. */
1086 isolate_freepages_block(cc
, &isolate_start_pfn
, block_end_pfn
,
1090 * If we isolated enough freepages, or aborted due to lock
1091 * contention, terminate.
1093 if ((cc
->nr_freepages
>= cc
->nr_migratepages
)
1095 if (isolate_start_pfn
>= block_end_pfn
) {
1097 * Restart at previous pageblock if more
1098 * freepages can be isolated next time.
1101 block_start_pfn
- pageblock_nr_pages
;
1104 } else if (isolate_start_pfn
< block_end_pfn
) {
1106 * If isolation failed early, do not continue
1113 /* split_free_page does not map the pages */
1114 map_pages(freelist
);
1117 * Record where the free scanner will restart next time. Either we
1118 * broke from the loop and set isolate_start_pfn based on the last
1119 * call to isolate_freepages_block(), or we met the migration scanner
1120 * and the loop terminated due to isolate_start_pfn < low_pfn
1122 cc
->free_pfn
= isolate_start_pfn
;
1126 * This is a migrate-callback that "allocates" freepages by taking pages
1127 * from the isolated freelists in the block we are migrating to.
1129 static struct page
*compaction_alloc(struct page
*migratepage
,
1133 struct compact_control
*cc
= (struct compact_control
*)data
;
1134 struct page
*freepage
;
1137 * Isolate free pages if necessary, and if we are not aborting due to
1140 if (list_empty(&cc
->freepages
)) {
1142 isolate_freepages(cc
);
1144 if (list_empty(&cc
->freepages
))
1148 freepage
= list_entry(cc
->freepages
.next
, struct page
, lru
);
1149 list_del(&freepage
->lru
);
1156 * This is a migrate-callback that "frees" freepages back to the isolated
1157 * freelist. All pages on the freelist are from the same zone, so there is no
1158 * special handling needed for NUMA.
1160 static void compaction_free(struct page
*page
, unsigned long data
)
1162 struct compact_control
*cc
= (struct compact_control
*)data
;
1164 list_add(&page
->lru
, &cc
->freepages
);
1168 /* possible outcome of isolate_migratepages */
1170 ISOLATE_ABORT
, /* Abort compaction now */
1171 ISOLATE_NONE
, /* No pages isolated, continue scanning */
1172 ISOLATE_SUCCESS
, /* Pages isolated, migrate */
1173 } isolate_migrate_t
;
1176 * Allow userspace to control policy on scanning the unevictable LRU for
1177 * compactable pages.
1179 int sysctl_compact_unevictable_allowed __read_mostly
= 1;
1182 * Isolate all pages that can be migrated from the first suitable block,
1183 * starting at the block pointed to by the migrate scanner pfn within
1186 static isolate_migrate_t
isolate_migratepages(struct zone
*zone
,
1187 struct compact_control
*cc
)
1189 unsigned long block_start_pfn
;
1190 unsigned long block_end_pfn
;
1191 unsigned long low_pfn
;
1193 const isolate_mode_t isolate_mode
=
1194 (sysctl_compact_unevictable_allowed
? ISOLATE_UNEVICTABLE
: 0) |
1195 (cc
->mode
== MIGRATE_ASYNC
? ISOLATE_ASYNC_MIGRATE
: 0);
1198 * Start at where we last stopped, or beginning of the zone as
1199 * initialized by compact_zone()
1201 low_pfn
= cc
->migrate_pfn
;
1202 block_start_pfn
= pageblock_start_pfn(low_pfn
);
1203 if (block_start_pfn
< zone
->zone_start_pfn
)
1204 block_start_pfn
= zone
->zone_start_pfn
;
1206 /* Only scan within a pageblock boundary */
1207 block_end_pfn
= pageblock_end_pfn(low_pfn
);
1210 * Iterate over whole pageblocks until we find the first suitable.
1211 * Do not cross the free scanner.
1213 for (; block_end_pfn
<= cc
->free_pfn
;
1214 low_pfn
= block_end_pfn
,
1215 block_start_pfn
= block_end_pfn
,
1216 block_end_pfn
+= pageblock_nr_pages
) {
1219 * This can potentially iterate a massively long zone with
1220 * many pageblocks unsuitable, so periodically check if we
1221 * need to schedule, or even abort async compaction.
1223 if (!(low_pfn
% (SWAP_CLUSTER_MAX
* pageblock_nr_pages
))
1224 && compact_should_abort(cc
))
1227 page
= pageblock_pfn_to_page(block_start_pfn
, block_end_pfn
,
1232 /* If isolation recently failed, do not retry */
1233 if (!isolation_suitable(cc
, page
))
1237 * For async compaction, also only scan in MOVABLE blocks.
1238 * Async compaction is optimistic to see if the minimum amount
1239 * of work satisfies the allocation.
1241 if (cc
->mode
== MIGRATE_ASYNC
&&
1242 !migrate_async_suitable(get_pageblock_migratetype(page
)))
1245 /* Perform the isolation */
1246 low_pfn
= isolate_migratepages_block(cc
, low_pfn
,
1247 block_end_pfn
, isolate_mode
);
1249 if (!low_pfn
|| cc
->contended
) {
1250 acct_isolated(zone
, cc
);
1251 return ISOLATE_ABORT
;
1255 * Either we isolated something and proceed with migration. Or
1256 * we failed and compact_zone should decide if we should
1262 acct_isolated(zone
, cc
);
1263 /* Record where migration scanner will be restarted. */
1264 cc
->migrate_pfn
= low_pfn
;
1266 return cc
->nr_migratepages
? ISOLATE_SUCCESS
: ISOLATE_NONE
;
1270 * order == -1 is expected when compacting via
1271 * /proc/sys/vm/compact_memory
1273 static inline bool is_via_compact_memory(int order
)
1278 static enum compact_result
__compact_finished(struct zone
*zone
, struct compact_control
*cc
,
1279 const int migratetype
)
1282 unsigned long watermark
;
1284 if (cc
->contended
|| fatal_signal_pending(current
))
1285 return COMPACT_CONTENDED
;
1287 /* Compaction run completes if the migrate and free scanner meet */
1288 if (compact_scanners_met(cc
)) {
1289 /* Let the next compaction start anew. */
1290 reset_cached_positions(zone
);
1293 * Mark that the PG_migrate_skip information should be cleared
1294 * by kswapd when it goes to sleep. kcompactd does not set the
1295 * flag itself as the decision to be clear should be directly
1296 * based on an allocation request.
1298 if (cc
->direct_compaction
)
1299 zone
->compact_blockskip_flush
= true;
1302 return COMPACT_COMPLETE
;
1304 return COMPACT_PARTIAL_SKIPPED
;
1307 if (is_via_compact_memory(cc
->order
))
1308 return COMPACT_CONTINUE
;
1310 /* Compaction run is not finished if the watermark is not met */
1311 watermark
= low_wmark_pages(zone
);
1313 if (!zone_watermark_ok(zone
, cc
->order
, watermark
, cc
->classzone_idx
,
1315 return COMPACT_CONTINUE
;
1317 /* Direct compactor: Is a suitable page free? */
1318 for (order
= cc
->order
; order
< MAX_ORDER
; order
++) {
1319 struct free_area
*area
= &zone
->free_area
[order
];
1322 /* Job done if page is free of the right migratetype */
1323 if (!list_empty(&area
->free_list
[migratetype
]))
1324 return COMPACT_PARTIAL
;
1327 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
1328 if (migratetype
== MIGRATE_MOVABLE
&&
1329 !list_empty(&area
->free_list
[MIGRATE_CMA
]))
1330 return COMPACT_PARTIAL
;
1333 * Job done if allocation would steal freepages from
1334 * other migratetype buddy lists.
1336 if (find_suitable_fallback(area
, order
, migratetype
,
1337 true, &can_steal
) != -1)
1338 return COMPACT_PARTIAL
;
1341 return COMPACT_NO_SUITABLE_PAGE
;
1344 static enum compact_result
compact_finished(struct zone
*zone
,
1345 struct compact_control
*cc
,
1346 const int migratetype
)
1350 ret
= __compact_finished(zone
, cc
, migratetype
);
1351 trace_mm_compaction_finished(zone
, cc
->order
, ret
);
1352 if (ret
== COMPACT_NO_SUITABLE_PAGE
)
1353 ret
= COMPACT_CONTINUE
;
1359 * compaction_suitable: Is this suitable to run compaction on this zone now?
1361 * COMPACT_SKIPPED - If there are too few free pages for compaction
1362 * COMPACT_PARTIAL - If the allocation would succeed without compaction
1363 * COMPACT_CONTINUE - If compaction should run now
1365 static enum compact_result
__compaction_suitable(struct zone
*zone
, int order
,
1366 unsigned int alloc_flags
,
1368 unsigned long wmark_target
)
1371 unsigned long watermark
;
1373 if (is_via_compact_memory(order
))
1374 return COMPACT_CONTINUE
;
1376 watermark
= low_wmark_pages(zone
);
1378 * If watermarks for high-order allocation are already met, there
1379 * should be no need for compaction at all.
1381 if (zone_watermark_ok(zone
, order
, watermark
, classzone_idx
,
1383 return COMPACT_PARTIAL
;
1386 * Watermarks for order-0 must be met for compaction. Note the 2UL.
1387 * This is because during migration, copies of pages need to be
1388 * allocated and for a short time, the footprint is higher
1390 watermark
+= (2UL << order
);
1391 if (!__zone_watermark_ok(zone
, 0, watermark
, classzone_idx
,
1392 alloc_flags
, wmark_target
))
1393 return COMPACT_SKIPPED
;
1396 * fragmentation index determines if allocation failures are due to
1397 * low memory or external fragmentation
1399 * index of -1000 would imply allocations might succeed depending on
1400 * watermarks, but we already failed the high-order watermark check
1401 * index towards 0 implies failure is due to lack of memory
1402 * index towards 1000 implies failure is due to fragmentation
1404 * Only compact if a failure would be due to fragmentation.
1406 fragindex
= fragmentation_index(zone
, order
);
1407 if (fragindex
>= 0 && fragindex
<= sysctl_extfrag_threshold
)
1408 return COMPACT_NOT_SUITABLE_ZONE
;
1410 return COMPACT_CONTINUE
;
1413 enum compact_result
compaction_suitable(struct zone
*zone
, int order
,
1414 unsigned int alloc_flags
,
1417 enum compact_result ret
;
1419 ret
= __compaction_suitable(zone
, order
, alloc_flags
, classzone_idx
,
1420 zone_page_state(zone
, NR_FREE_PAGES
));
1421 trace_mm_compaction_suitable(zone
, order
, ret
);
1422 if (ret
== COMPACT_NOT_SUITABLE_ZONE
)
1423 ret
= COMPACT_SKIPPED
;
1428 bool compaction_zonelist_suitable(struct alloc_context
*ac
, int order
,
1435 * Make sure at least one zone would pass __compaction_suitable if we continue
1436 * retrying the reclaim.
1438 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
1440 unsigned long available
;
1441 enum compact_result compact_result
;
1444 * Do not consider all the reclaimable memory because we do not
1445 * want to trash just for a single high order allocation which
1446 * is even not guaranteed to appear even if __compaction_suitable
1447 * is happy about the watermark check.
1449 available
= zone_reclaimable_pages(zone
) / order
;
1450 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
1451 compact_result
= __compaction_suitable(zone
, order
, alloc_flags
,
1452 ac_classzone_idx(ac
), available
);
1453 if (compact_result
!= COMPACT_SKIPPED
&&
1454 compact_result
!= COMPACT_NOT_SUITABLE_ZONE
)
1461 static enum compact_result
compact_zone(struct zone
*zone
, struct compact_control
*cc
)
1463 enum compact_result ret
;
1464 unsigned long start_pfn
= zone
->zone_start_pfn
;
1465 unsigned long end_pfn
= zone_end_pfn(zone
);
1466 const int migratetype
= gfpflags_to_migratetype(cc
->gfp_mask
);
1467 const bool sync
= cc
->mode
!= MIGRATE_ASYNC
;
1469 ret
= compaction_suitable(zone
, cc
->order
, cc
->alloc_flags
,
1471 /* Compaction is likely to fail */
1472 if (ret
== COMPACT_PARTIAL
|| ret
== COMPACT_SKIPPED
)
1475 /* huh, compaction_suitable is returning something unexpected */
1476 VM_BUG_ON(ret
!= COMPACT_CONTINUE
);
1479 * Clear pageblock skip if there were failures recently and compaction
1480 * is about to be retried after being deferred.
1482 if (compaction_restarting(zone
, cc
->order
))
1483 __reset_isolation_suitable(zone
);
1486 * Setup to move all movable pages to the end of the zone. Used cached
1487 * information on where the scanners should start but check that it
1488 * is initialised by ensuring the values are within zone boundaries.
1490 cc
->migrate_pfn
= zone
->compact_cached_migrate_pfn
[sync
];
1491 cc
->free_pfn
= zone
->compact_cached_free_pfn
;
1492 if (cc
->free_pfn
< start_pfn
|| cc
->free_pfn
>= end_pfn
) {
1493 cc
->free_pfn
= pageblock_start_pfn(end_pfn
- 1);
1494 zone
->compact_cached_free_pfn
= cc
->free_pfn
;
1496 if (cc
->migrate_pfn
< start_pfn
|| cc
->migrate_pfn
>= end_pfn
) {
1497 cc
->migrate_pfn
= start_pfn
;
1498 zone
->compact_cached_migrate_pfn
[0] = cc
->migrate_pfn
;
1499 zone
->compact_cached_migrate_pfn
[1] = cc
->migrate_pfn
;
1502 if (cc
->migrate_pfn
== start_pfn
)
1503 cc
->whole_zone
= true;
1505 cc
->last_migrated_pfn
= 0;
1507 trace_mm_compaction_begin(start_pfn
, cc
->migrate_pfn
,
1508 cc
->free_pfn
, end_pfn
, sync
);
1510 migrate_prep_local();
1512 while ((ret
= compact_finished(zone
, cc
, migratetype
)) ==
1516 switch (isolate_migratepages(zone
, cc
)) {
1518 ret
= COMPACT_CONTENDED
;
1519 putback_movable_pages(&cc
->migratepages
);
1520 cc
->nr_migratepages
= 0;
1524 * We haven't isolated and migrated anything, but
1525 * there might still be unflushed migrations from
1526 * previous cc->order aligned block.
1529 case ISOLATE_SUCCESS
:
1533 err
= migrate_pages(&cc
->migratepages
, compaction_alloc
,
1534 compaction_free
, (unsigned long)cc
, cc
->mode
,
1537 trace_mm_compaction_migratepages(cc
->nr_migratepages
, err
,
1540 /* All pages were either migrated or will be released */
1541 cc
->nr_migratepages
= 0;
1543 putback_movable_pages(&cc
->migratepages
);
1545 * migrate_pages() may return -ENOMEM when scanners meet
1546 * and we want compact_finished() to detect it
1548 if (err
== -ENOMEM
&& !compact_scanners_met(cc
)) {
1549 ret
= COMPACT_CONTENDED
;
1553 * We failed to migrate at least one page in the current
1554 * order-aligned block, so skip the rest of it.
1556 if (cc
->direct_compaction
&&
1557 (cc
->mode
== MIGRATE_ASYNC
)) {
1558 cc
->migrate_pfn
= block_end_pfn(
1559 cc
->migrate_pfn
- 1, cc
->order
);
1560 /* Draining pcplists is useless in this case */
1561 cc
->last_migrated_pfn
= 0;
1568 * Has the migration scanner moved away from the previous
1569 * cc->order aligned block where we migrated from? If yes,
1570 * flush the pages that were freed, so that they can merge and
1571 * compact_finished() can detect immediately if allocation
1574 if (cc
->order
> 0 && cc
->last_migrated_pfn
) {
1576 unsigned long current_block_start
=
1577 block_start_pfn(cc
->migrate_pfn
, cc
->order
);
1579 if (cc
->last_migrated_pfn
< current_block_start
) {
1581 lru_add_drain_cpu(cpu
);
1582 drain_local_pages(zone
);
1584 /* No more flushing until we migrate again */
1585 cc
->last_migrated_pfn
= 0;
1593 * Release free pages and update where the free scanner should restart,
1594 * so we don't leave any returned pages behind in the next attempt.
1596 if (cc
->nr_freepages
> 0) {
1597 unsigned long free_pfn
= release_freepages(&cc
->freepages
);
1599 cc
->nr_freepages
= 0;
1600 VM_BUG_ON(free_pfn
== 0);
1601 /* The cached pfn is always the first in a pageblock */
1602 free_pfn
= pageblock_start_pfn(free_pfn
);
1604 * Only go back, not forward. The cached pfn might have been
1605 * already reset to zone end in compact_finished()
1607 if (free_pfn
> zone
->compact_cached_free_pfn
)
1608 zone
->compact_cached_free_pfn
= free_pfn
;
1611 trace_mm_compaction_end(start_pfn
, cc
->migrate_pfn
,
1612 cc
->free_pfn
, end_pfn
, sync
, ret
);
1614 if (ret
== COMPACT_CONTENDED
)
1615 ret
= COMPACT_PARTIAL
;
1620 static enum compact_result
compact_zone_order(struct zone
*zone
, int order
,
1621 gfp_t gfp_mask
, enum migrate_mode mode
, int *contended
,
1622 unsigned int alloc_flags
, int classzone_idx
)
1624 enum compact_result ret
;
1625 struct compact_control cc
= {
1627 .nr_migratepages
= 0,
1629 .gfp_mask
= gfp_mask
,
1632 .alloc_flags
= alloc_flags
,
1633 .classzone_idx
= classzone_idx
,
1634 .direct_compaction
= true,
1636 INIT_LIST_HEAD(&cc
.freepages
);
1637 INIT_LIST_HEAD(&cc
.migratepages
);
1639 ret
= compact_zone(zone
, &cc
);
1641 VM_BUG_ON(!list_empty(&cc
.freepages
));
1642 VM_BUG_ON(!list_empty(&cc
.migratepages
));
1644 *contended
= cc
.contended
;
1648 int sysctl_extfrag_threshold
= 500;
1651 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1652 * @gfp_mask: The GFP mask of the current allocation
1653 * @order: The order of the current allocation
1654 * @alloc_flags: The allocation flags of the current allocation
1655 * @ac: The context of current allocation
1656 * @mode: The migration mode for async, sync light, or sync migration
1657 * @contended: Return value that determines if compaction was aborted due to
1658 * need_resched() or lock contention
1660 * This is the main entry point for direct page compaction.
1662 enum compact_result
try_to_compact_pages(gfp_t gfp_mask
, unsigned int order
,
1663 unsigned int alloc_flags
, const struct alloc_context
*ac
,
1664 enum migrate_mode mode
, int *contended
)
1666 int may_enter_fs
= gfp_mask
& __GFP_FS
;
1667 int may_perform_io
= gfp_mask
& __GFP_IO
;
1670 enum compact_result rc
= COMPACT_SKIPPED
;
1671 int all_zones_contended
= COMPACT_CONTENDED_LOCK
; /* init for &= op */
1673 *contended
= COMPACT_CONTENDED_NONE
;
1675 /* Check if the GFP flags allow compaction */
1676 if (!order
|| !may_enter_fs
|| !may_perform_io
)
1677 return COMPACT_SKIPPED
;
1679 trace_mm_compaction_try_to_compact_pages(order
, gfp_mask
, mode
);
1681 /* Compact each zone in the list */
1682 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
1684 enum compact_result status
;
1687 if (compaction_deferred(zone
, order
)) {
1688 rc
= max_t(enum compact_result
, COMPACT_DEFERRED
, rc
);
1692 status
= compact_zone_order(zone
, order
, gfp_mask
, mode
,
1693 &zone_contended
, alloc_flags
,
1694 ac_classzone_idx(ac
));
1695 rc
= max(status
, rc
);
1697 * It takes at least one zone that wasn't lock contended
1698 * to clear all_zones_contended.
1700 all_zones_contended
&= zone_contended
;
1702 /* If a normal allocation would succeed, stop compacting */
1703 if (zone_watermark_ok(zone
, order
, low_wmark_pages(zone
),
1704 ac_classzone_idx(ac
), alloc_flags
)) {
1706 * We think the allocation will succeed in this zone,
1707 * but it is not certain, hence the false. The caller
1708 * will repeat this with true if allocation indeed
1709 * succeeds in this zone.
1711 compaction_defer_reset(zone
, order
, false);
1713 * It is possible that async compaction aborted due to
1714 * need_resched() and the watermarks were ok thanks to
1715 * somebody else freeing memory. The allocation can
1716 * however still fail so we better signal the
1717 * need_resched() contention anyway (this will not
1718 * prevent the allocation attempt).
1720 if (zone_contended
== COMPACT_CONTENDED_SCHED
)
1721 *contended
= COMPACT_CONTENDED_SCHED
;
1726 if (mode
!= MIGRATE_ASYNC
&& (status
== COMPACT_COMPLETE
||
1727 status
== COMPACT_PARTIAL_SKIPPED
)) {
1729 * We think that allocation won't succeed in this zone
1730 * so we defer compaction there. If it ends up
1731 * succeeding after all, it will be reset.
1733 defer_compaction(zone
, order
);
1737 * We might have stopped compacting due to need_resched() in
1738 * async compaction, or due to a fatal signal detected. In that
1739 * case do not try further zones and signal need_resched()
1742 if ((zone_contended
== COMPACT_CONTENDED_SCHED
)
1743 || fatal_signal_pending(current
)) {
1744 *contended
= COMPACT_CONTENDED_SCHED
;
1751 * We might not have tried all the zones, so be conservative
1752 * and assume they are not all lock contended.
1754 all_zones_contended
= 0;
1759 * If at least one zone wasn't deferred or skipped, we report if all
1760 * zones that were tried were lock contended.
1762 if (rc
> COMPACT_INACTIVE
&& all_zones_contended
)
1763 *contended
= COMPACT_CONTENDED_LOCK
;
1769 /* Compact all zones within a node */
1770 static void __compact_pgdat(pg_data_t
*pgdat
, struct compact_control
*cc
)
1775 for (zoneid
= 0; zoneid
< MAX_NR_ZONES
; zoneid
++) {
1777 zone
= &pgdat
->node_zones
[zoneid
];
1778 if (!populated_zone(zone
))
1781 cc
->nr_freepages
= 0;
1782 cc
->nr_migratepages
= 0;
1784 INIT_LIST_HEAD(&cc
->freepages
);
1785 INIT_LIST_HEAD(&cc
->migratepages
);
1788 * When called via /proc/sys/vm/compact_memory
1789 * this makes sure we compact the whole zone regardless of
1790 * cached scanner positions.
1792 if (is_via_compact_memory(cc
->order
))
1793 __reset_isolation_suitable(zone
);
1795 if (is_via_compact_memory(cc
->order
) ||
1796 !compaction_deferred(zone
, cc
->order
))
1797 compact_zone(zone
, cc
);
1799 VM_BUG_ON(!list_empty(&cc
->freepages
));
1800 VM_BUG_ON(!list_empty(&cc
->migratepages
));
1802 if (is_via_compact_memory(cc
->order
))
1805 if (zone_watermark_ok(zone
, cc
->order
,
1806 low_wmark_pages(zone
), 0, 0))
1807 compaction_defer_reset(zone
, cc
->order
, false);
1811 void compact_pgdat(pg_data_t
*pgdat
, int order
)
1813 struct compact_control cc
= {
1815 .mode
= MIGRATE_ASYNC
,
1821 __compact_pgdat(pgdat
, &cc
);
1824 static void compact_node(int nid
)
1826 struct compact_control cc
= {
1828 .mode
= MIGRATE_SYNC
,
1829 .ignore_skip_hint
= true,
1832 __compact_pgdat(NODE_DATA(nid
), &cc
);
1835 /* Compact all nodes in the system */
1836 static void compact_nodes(void)
1840 /* Flush pending updates to the LRU lists */
1841 lru_add_drain_all();
1843 for_each_online_node(nid
)
1847 /* The written value is actually unused, all memory is compacted */
1848 int sysctl_compact_memory
;
1851 * This is the entry point for compacting all nodes via
1852 * /proc/sys/vm/compact_memory
1854 int sysctl_compaction_handler(struct ctl_table
*table
, int write
,
1855 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
1863 int sysctl_extfrag_handler(struct ctl_table
*table
, int write
,
1864 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
1866 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
1871 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1872 static ssize_t
sysfs_compact_node(struct device
*dev
,
1873 struct device_attribute
*attr
,
1874 const char *buf
, size_t count
)
1878 if (nid
>= 0 && nid
< nr_node_ids
&& node_online(nid
)) {
1879 /* Flush pending updates to the LRU lists */
1880 lru_add_drain_all();
1887 static DEVICE_ATTR(compact
, S_IWUSR
, NULL
, sysfs_compact_node
);
1889 int compaction_register_node(struct node
*node
)
1891 return device_create_file(&node
->dev
, &dev_attr_compact
);
1894 void compaction_unregister_node(struct node
*node
)
1896 return device_remove_file(&node
->dev
, &dev_attr_compact
);
1898 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1900 static inline bool kcompactd_work_requested(pg_data_t
*pgdat
)
1902 return pgdat
->kcompactd_max_order
> 0 || kthread_should_stop();
1905 static bool kcompactd_node_suitable(pg_data_t
*pgdat
)
1909 enum zone_type classzone_idx
= pgdat
->kcompactd_classzone_idx
;
1911 for (zoneid
= 0; zoneid
<= classzone_idx
; zoneid
++) {
1912 zone
= &pgdat
->node_zones
[zoneid
];
1914 if (!populated_zone(zone
))
1917 if (compaction_suitable(zone
, pgdat
->kcompactd_max_order
, 0,
1918 classzone_idx
) == COMPACT_CONTINUE
)
1925 static void kcompactd_do_work(pg_data_t
*pgdat
)
1928 * With no special task, compact all zones so that a page of requested
1929 * order is allocatable.
1933 struct compact_control cc
= {
1934 .order
= pgdat
->kcompactd_max_order
,
1935 .classzone_idx
= pgdat
->kcompactd_classzone_idx
,
1936 .mode
= MIGRATE_SYNC_LIGHT
,
1937 .ignore_skip_hint
= true,
1940 bool success
= false;
1942 trace_mm_compaction_kcompactd_wake(pgdat
->node_id
, cc
.order
,
1944 count_vm_event(KCOMPACTD_WAKE
);
1946 for (zoneid
= 0; zoneid
<= cc
.classzone_idx
; zoneid
++) {
1949 zone
= &pgdat
->node_zones
[zoneid
];
1950 if (!populated_zone(zone
))
1953 if (compaction_deferred(zone
, cc
.order
))
1956 if (compaction_suitable(zone
, cc
.order
, 0, zoneid
) !=
1960 cc
.nr_freepages
= 0;
1961 cc
.nr_migratepages
= 0;
1963 INIT_LIST_HEAD(&cc
.freepages
);
1964 INIT_LIST_HEAD(&cc
.migratepages
);
1966 if (kthread_should_stop())
1968 status
= compact_zone(zone
, &cc
);
1970 if (zone_watermark_ok(zone
, cc
.order
, low_wmark_pages(zone
),
1971 cc
.classzone_idx
, 0)) {
1973 compaction_defer_reset(zone
, cc
.order
, false);
1974 } else if (status
== COMPACT_PARTIAL_SKIPPED
|| status
== COMPACT_COMPLETE
) {
1976 * We use sync migration mode here, so we defer like
1977 * sync direct compaction does.
1979 defer_compaction(zone
, cc
.order
);
1982 VM_BUG_ON(!list_empty(&cc
.freepages
));
1983 VM_BUG_ON(!list_empty(&cc
.migratepages
));
1987 * Regardless of success, we are done until woken up next. But remember
1988 * the requested order/classzone_idx in case it was higher/tighter than
1991 if (pgdat
->kcompactd_max_order
<= cc
.order
)
1992 pgdat
->kcompactd_max_order
= 0;
1993 if (pgdat
->kcompactd_classzone_idx
>= cc
.classzone_idx
)
1994 pgdat
->kcompactd_classzone_idx
= pgdat
->nr_zones
- 1;
1997 void wakeup_kcompactd(pg_data_t
*pgdat
, int order
, int classzone_idx
)
2002 if (pgdat
->kcompactd_max_order
< order
)
2003 pgdat
->kcompactd_max_order
= order
;
2005 if (pgdat
->kcompactd_classzone_idx
> classzone_idx
)
2006 pgdat
->kcompactd_classzone_idx
= classzone_idx
;
2008 if (!waitqueue_active(&pgdat
->kcompactd_wait
))
2011 if (!kcompactd_node_suitable(pgdat
))
2014 trace_mm_compaction_wakeup_kcompactd(pgdat
->node_id
, order
,
2016 wake_up_interruptible(&pgdat
->kcompactd_wait
);
2020 * The background compaction daemon, started as a kernel thread
2021 * from the init process.
2023 static int kcompactd(void *p
)
2025 pg_data_t
*pgdat
= (pg_data_t
*)p
;
2026 struct task_struct
*tsk
= current
;
2028 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
2030 if (!cpumask_empty(cpumask
))
2031 set_cpus_allowed_ptr(tsk
, cpumask
);
2035 pgdat
->kcompactd_max_order
= 0;
2036 pgdat
->kcompactd_classzone_idx
= pgdat
->nr_zones
- 1;
2038 while (!kthread_should_stop()) {
2039 trace_mm_compaction_kcompactd_sleep(pgdat
->node_id
);
2040 wait_event_freezable(pgdat
->kcompactd_wait
,
2041 kcompactd_work_requested(pgdat
));
2043 kcompactd_do_work(pgdat
);
2050 * This kcompactd start function will be called by init and node-hot-add.
2051 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
2053 int kcompactd_run(int nid
)
2055 pg_data_t
*pgdat
= NODE_DATA(nid
);
2058 if (pgdat
->kcompactd
)
2061 pgdat
->kcompactd
= kthread_run(kcompactd
, pgdat
, "kcompactd%d", nid
);
2062 if (IS_ERR(pgdat
->kcompactd
)) {
2063 pr_err("Failed to start kcompactd on node %d\n", nid
);
2064 ret
= PTR_ERR(pgdat
->kcompactd
);
2065 pgdat
->kcompactd
= NULL
;
2071 * Called by memory hotplug when all memory in a node is offlined. Caller must
2072 * hold mem_hotplug_begin/end().
2074 void kcompactd_stop(int nid
)
2076 struct task_struct
*kcompactd
= NODE_DATA(nid
)->kcompactd
;
2079 kthread_stop(kcompactd
);
2080 NODE_DATA(nid
)->kcompactd
= NULL
;
2085 * It's optimal to keep kcompactd on the same CPUs as their memory, but
2086 * not required for correctness. So if the last cpu in a node goes
2087 * away, we get changed to run anywhere: as the first one comes back,
2088 * restore their cpu bindings.
2090 static int cpu_callback(struct notifier_block
*nfb
, unsigned long action
,
2095 if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
) {
2096 for_each_node_state(nid
, N_MEMORY
) {
2097 pg_data_t
*pgdat
= NODE_DATA(nid
);
2098 const struct cpumask
*mask
;
2100 mask
= cpumask_of_node(pgdat
->node_id
);
2102 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
2103 /* One of our CPUs online: restore mask */
2104 set_cpus_allowed_ptr(pgdat
->kcompactd
, mask
);
2110 static int __init
kcompactd_init(void)
2114 for_each_node_state(nid
, N_MEMORY
)
2116 hotcpu_notifier(cpu_callback
, 0);
2119 subsys_initcall(kcompactd_init
)
2121 #endif /* CONFIG_COMPACTION */