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 /* Do not skip compaction more than 64 times */
85 #define COMPACT_MAX_DEFER_SHIFT 6
88 * Compaction is deferred when compaction fails to result in a page
89 * allocation success. 1 << compact_defer_limit compactions are skipped up
90 * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
92 void defer_compaction(struct zone
*zone
, int order
)
94 zone
->compact_considered
= 0;
95 zone
->compact_defer_shift
++;
97 if (order
< zone
->compact_order_failed
)
98 zone
->compact_order_failed
= order
;
100 if (zone
->compact_defer_shift
> COMPACT_MAX_DEFER_SHIFT
)
101 zone
->compact_defer_shift
= COMPACT_MAX_DEFER_SHIFT
;
103 trace_mm_compaction_defer_compaction(zone
, order
);
106 /* Returns true if compaction should be skipped this time */
107 bool compaction_deferred(struct zone
*zone
, int order
)
109 unsigned long defer_limit
= 1UL << zone
->compact_defer_shift
;
111 if (order
< zone
->compact_order_failed
)
114 /* Avoid possible overflow */
115 if (++zone
->compact_considered
> defer_limit
)
116 zone
->compact_considered
= defer_limit
;
118 if (zone
->compact_considered
>= defer_limit
)
121 trace_mm_compaction_deferred(zone
, order
);
127 * Update defer tracking counters after successful compaction of given order,
128 * which means an allocation either succeeded (alloc_success == true) or is
129 * expected to succeed.
131 void compaction_defer_reset(struct zone
*zone
, int order
,
135 zone
->compact_considered
= 0;
136 zone
->compact_defer_shift
= 0;
138 if (order
>= zone
->compact_order_failed
)
139 zone
->compact_order_failed
= order
+ 1;
141 trace_mm_compaction_defer_reset(zone
, order
);
144 /* Returns true if restarting compaction after many failures */
145 bool compaction_restarting(struct zone
*zone
, int order
)
147 if (order
< zone
->compact_order_failed
)
150 return zone
->compact_defer_shift
== COMPACT_MAX_DEFER_SHIFT
&&
151 zone
->compact_considered
>= 1UL << zone
->compact_defer_shift
;
154 /* Returns true if the pageblock should be scanned for pages to isolate. */
155 static inline bool isolation_suitable(struct compact_control
*cc
,
158 if (cc
->ignore_skip_hint
)
161 return !get_pageblock_skip(page
);
164 static void reset_cached_positions(struct zone
*zone
)
166 zone
->compact_cached_migrate_pfn
[0] = zone
->zone_start_pfn
;
167 zone
->compact_cached_migrate_pfn
[1] = zone
->zone_start_pfn
;
168 zone
->compact_cached_free_pfn
=
169 pageblock_start_pfn(zone_end_pfn(zone
) - 1);
173 * This function is called to clear all cached information on pageblocks that
174 * should be skipped for page isolation when the migrate and free page scanner
177 static void __reset_isolation_suitable(struct zone
*zone
)
179 unsigned long start_pfn
= zone
->zone_start_pfn
;
180 unsigned long end_pfn
= zone_end_pfn(zone
);
183 zone
->compact_blockskip_flush
= false;
185 /* Walk the zone and mark every pageblock as suitable for isolation */
186 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
194 page
= pfn_to_page(pfn
);
195 if (zone
!= page_zone(page
))
198 clear_pageblock_skip(page
);
201 reset_cached_positions(zone
);
204 void reset_isolation_suitable(pg_data_t
*pgdat
)
208 for (zoneid
= 0; zoneid
< MAX_NR_ZONES
; zoneid
++) {
209 struct zone
*zone
= &pgdat
->node_zones
[zoneid
];
210 if (!populated_zone(zone
))
213 /* Only flush if a full compaction finished recently */
214 if (zone
->compact_blockskip_flush
)
215 __reset_isolation_suitable(zone
);
220 * If no pages were isolated then mark this pageblock to be skipped in the
221 * future. The information is later cleared by __reset_isolation_suitable().
223 static void update_pageblock_skip(struct compact_control
*cc
,
224 struct page
*page
, unsigned long nr_isolated
,
225 bool migrate_scanner
)
227 struct zone
*zone
= cc
->zone
;
230 if (cc
->ignore_skip_hint
)
239 set_pageblock_skip(page
);
241 pfn
= page_to_pfn(page
);
243 /* Update where async and sync compaction should restart */
244 if (migrate_scanner
) {
245 if (pfn
> zone
->compact_cached_migrate_pfn
[0])
246 zone
->compact_cached_migrate_pfn
[0] = pfn
;
247 if (cc
->mode
!= MIGRATE_ASYNC
&&
248 pfn
> zone
->compact_cached_migrate_pfn
[1])
249 zone
->compact_cached_migrate_pfn
[1] = pfn
;
251 if (pfn
< zone
->compact_cached_free_pfn
)
252 zone
->compact_cached_free_pfn
= pfn
;
256 static inline bool isolation_suitable(struct compact_control
*cc
,
262 static void update_pageblock_skip(struct compact_control
*cc
,
263 struct page
*page
, unsigned long nr_isolated
,
264 bool migrate_scanner
)
267 #endif /* CONFIG_COMPACTION */
270 * Compaction requires the taking of some coarse locks that are potentially
271 * very heavily contended. For async compaction, back out if the lock cannot
272 * be taken immediately. For sync compaction, spin on the lock if needed.
274 * Returns true if the lock is held
275 * Returns false if the lock is not held and compaction should abort
277 static bool compact_trylock_irqsave(spinlock_t
*lock
, unsigned long *flags
,
278 struct compact_control
*cc
)
280 if (cc
->mode
== MIGRATE_ASYNC
) {
281 if (!spin_trylock_irqsave(lock
, *flags
)) {
282 cc
->contended
= COMPACT_CONTENDED_LOCK
;
286 spin_lock_irqsave(lock
, *flags
);
293 * Compaction requires the taking of some coarse locks that are potentially
294 * very heavily contended. The lock should be periodically unlocked to avoid
295 * having disabled IRQs for a long time, even when there is nobody waiting on
296 * the lock. It might also be that allowing the IRQs will result in
297 * need_resched() becoming true. If scheduling is needed, async compaction
298 * aborts. Sync compaction schedules.
299 * Either compaction type will also abort if a fatal signal is pending.
300 * In either case if the lock was locked, it is dropped and not regained.
302 * Returns true if compaction should abort due to fatal signal pending, or
303 * async compaction due to need_resched()
304 * Returns false when compaction can continue (sync compaction might have
307 static bool compact_unlock_should_abort(spinlock_t
*lock
,
308 unsigned long flags
, bool *locked
, struct compact_control
*cc
)
311 spin_unlock_irqrestore(lock
, flags
);
315 if (fatal_signal_pending(current
)) {
316 cc
->contended
= COMPACT_CONTENDED_SCHED
;
320 if (need_resched()) {
321 if (cc
->mode
== MIGRATE_ASYNC
) {
322 cc
->contended
= COMPACT_CONTENDED_SCHED
;
332 * Aside from avoiding lock contention, compaction also periodically checks
333 * need_resched() and either schedules in sync compaction or aborts async
334 * compaction. This is similar to what compact_unlock_should_abort() does, but
335 * is used where no lock is concerned.
337 * Returns false when no scheduling was needed, or sync compaction scheduled.
338 * Returns true when async compaction should abort.
340 static inline bool compact_should_abort(struct compact_control
*cc
)
342 /* async compaction aborts if contended */
343 if (need_resched()) {
344 if (cc
->mode
== MIGRATE_ASYNC
) {
345 cc
->contended
= COMPACT_CONTENDED_SCHED
;
356 * Isolate free pages onto a private freelist. If @strict is true, will abort
357 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
358 * (even though it may still end up isolating some pages).
360 static unsigned long isolate_freepages_block(struct compact_control
*cc
,
361 unsigned long *start_pfn
,
362 unsigned long end_pfn
,
363 struct list_head
*freelist
,
366 int nr_scanned
= 0, total_isolated
= 0;
367 struct page
*cursor
, *valid_page
= NULL
;
368 unsigned long flags
= 0;
370 unsigned long blockpfn
= *start_pfn
;
372 cursor
= pfn_to_page(blockpfn
);
374 /* Isolate free pages. */
375 for (; blockpfn
< end_pfn
; blockpfn
++, cursor
++) {
377 struct page
*page
= cursor
;
380 * Periodically drop the lock (if held) regardless of its
381 * contention, to give chance to IRQs. Abort if fatal signal
382 * pending or async compaction detects need_resched()
384 if (!(blockpfn
% SWAP_CLUSTER_MAX
)
385 && compact_unlock_should_abort(&cc
->zone
->lock
, flags
,
390 if (!pfn_valid_within(blockpfn
))
397 * For compound pages such as THP and hugetlbfs, we can save
398 * potentially a lot of iterations if we skip them at once.
399 * The check is racy, but we can consider only valid values
400 * and the only danger is skipping too much.
402 if (PageCompound(page
)) {
403 unsigned int comp_order
= compound_order(page
);
405 if (likely(comp_order
< MAX_ORDER
)) {
406 blockpfn
+= (1UL << comp_order
) - 1;
407 cursor
+= (1UL << comp_order
) - 1;
413 if (!PageBuddy(page
))
417 * If we already hold the lock, we can skip some rechecking.
418 * Note that if we hold the lock now, checked_pageblock was
419 * already set in some previous iteration (or strict is true),
420 * so it is correct to skip the suitable migration target
425 * The zone lock must be held to isolate freepages.
426 * Unfortunately this is a very coarse lock and can be
427 * heavily contended if there are parallel allocations
428 * or parallel compactions. For async compaction do not
429 * spin on the lock and we acquire the lock as late as
432 locked
= compact_trylock_irqsave(&cc
->zone
->lock
,
437 /* Recheck this is a buddy page under lock */
438 if (!PageBuddy(page
))
442 /* Found a free page, break it into order-0 pages */
443 isolated
= split_free_page(page
);
444 total_isolated
+= isolated
;
445 for (i
= 0; i
< isolated
; i
++) {
446 list_add(&page
->lru
, freelist
);
450 /* If a page was split, advance to the end of it */
452 cc
->nr_freepages
+= isolated
;
454 cc
->nr_migratepages
<= cc
->nr_freepages
) {
455 blockpfn
+= isolated
;
459 blockpfn
+= isolated
- 1;
460 cursor
+= isolated
- 1;
473 * There is a tiny chance that we have read bogus compound_order(),
474 * so be careful to not go outside of the pageblock.
476 if (unlikely(blockpfn
> end_pfn
))
479 trace_mm_compaction_isolate_freepages(*start_pfn
, blockpfn
,
480 nr_scanned
, total_isolated
);
482 /* Record how far we have got within the block */
483 *start_pfn
= blockpfn
;
486 * If strict isolation is requested by CMA then check that all the
487 * pages requested were isolated. If there were any failures, 0 is
488 * returned and CMA will fail.
490 if (strict
&& blockpfn
< end_pfn
)
494 spin_unlock_irqrestore(&cc
->zone
->lock
, flags
);
496 /* Update the pageblock-skip if the whole pageblock was scanned */
497 if (blockpfn
== end_pfn
)
498 update_pageblock_skip(cc
, valid_page
, total_isolated
, false);
500 count_compact_events(COMPACTFREE_SCANNED
, nr_scanned
);
502 count_compact_events(COMPACTISOLATED
, total_isolated
);
503 return total_isolated
;
507 * isolate_freepages_range() - isolate free pages.
508 * @start_pfn: The first PFN to start isolating.
509 * @end_pfn: The one-past-last PFN.
511 * Non-free pages, invalid PFNs, or zone boundaries within the
512 * [start_pfn, end_pfn) range are considered errors, cause function to
513 * undo its actions and return zero.
515 * Otherwise, function returns one-past-the-last PFN of isolated page
516 * (which may be greater then end_pfn if end fell in a middle of
520 isolate_freepages_range(struct compact_control
*cc
,
521 unsigned long start_pfn
, unsigned long end_pfn
)
523 unsigned long isolated
, pfn
, block_start_pfn
, block_end_pfn
;
527 block_start_pfn
= pageblock_start_pfn(pfn
);
528 if (block_start_pfn
< cc
->zone
->zone_start_pfn
)
529 block_start_pfn
= cc
->zone
->zone_start_pfn
;
530 block_end_pfn
= pageblock_end_pfn(pfn
);
532 for (; pfn
< end_pfn
; pfn
+= isolated
,
533 block_start_pfn
= block_end_pfn
,
534 block_end_pfn
+= pageblock_nr_pages
) {
535 /* Protect pfn from changing by isolate_freepages_block */
536 unsigned long isolate_start_pfn
= pfn
;
538 block_end_pfn
= min(block_end_pfn
, end_pfn
);
541 * pfn could pass the block_end_pfn if isolated freepage
542 * is more than pageblock order. In this case, we adjust
543 * scanning range to right one.
545 if (pfn
>= block_end_pfn
) {
546 block_start_pfn
= pageblock_start_pfn(pfn
);
547 block_end_pfn
= pageblock_end_pfn(pfn
);
548 block_end_pfn
= min(block_end_pfn
, end_pfn
);
551 if (!pageblock_pfn_to_page(block_start_pfn
,
552 block_end_pfn
, cc
->zone
))
555 isolated
= isolate_freepages_block(cc
, &isolate_start_pfn
,
556 block_end_pfn
, &freelist
, true);
559 * In strict mode, isolate_freepages_block() returns 0 if
560 * there are any holes in the block (ie. invalid PFNs or
567 * If we managed to isolate pages, it is always (1 << n) *
568 * pageblock_nr_pages for some non-negative n. (Max order
569 * page may span two pageblocks).
573 /* split_free_page does not map the pages */
574 map_pages(&freelist
);
577 /* Loop terminated early, cleanup. */
578 release_freepages(&freelist
);
582 /* We don't use freelists for anything. */
586 /* Update the number of anon and file isolated pages in the zone */
587 static void acct_isolated(struct zone
*zone
, struct compact_control
*cc
)
590 unsigned int count
[2] = { 0, };
592 if (list_empty(&cc
->migratepages
))
595 list_for_each_entry(page
, &cc
->migratepages
, lru
)
596 count
[!!page_is_file_cache(page
)]++;
598 mod_zone_page_state(zone
, NR_ISOLATED_ANON
, count
[0]);
599 mod_zone_page_state(zone
, NR_ISOLATED_FILE
, count
[1]);
602 /* Similar to reclaim, but different enough that they don't share logic */
603 static bool too_many_isolated(struct zone
*zone
)
605 unsigned long active
, inactive
, isolated
;
607 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
) +
608 zone_page_state(zone
, NR_INACTIVE_ANON
);
609 active
= zone_page_state(zone
, NR_ACTIVE_FILE
) +
610 zone_page_state(zone
, NR_ACTIVE_ANON
);
611 isolated
= zone_page_state(zone
, NR_ISOLATED_FILE
) +
612 zone_page_state(zone
, NR_ISOLATED_ANON
);
614 return isolated
> (inactive
+ active
) / 2;
618 * isolate_migratepages_block() - isolate all migrate-able pages within
620 * @cc: Compaction control structure.
621 * @low_pfn: The first PFN to isolate
622 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
623 * @isolate_mode: Isolation mode to be used.
625 * Isolate all pages that can be migrated from the range specified by
626 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
627 * Returns zero if there is a fatal signal pending, otherwise PFN of the
628 * first page that was not scanned (which may be both less, equal to or more
631 * The pages are isolated on cc->migratepages list (not required to be empty),
632 * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
633 * is neither read nor updated.
636 isolate_migratepages_block(struct compact_control
*cc
, unsigned long low_pfn
,
637 unsigned long end_pfn
, isolate_mode_t isolate_mode
)
639 struct zone
*zone
= cc
->zone
;
640 unsigned long nr_scanned
= 0, nr_isolated
= 0;
641 struct lruvec
*lruvec
;
642 unsigned long flags
= 0;
644 struct page
*page
= NULL
, *valid_page
= NULL
;
645 unsigned long start_pfn
= low_pfn
;
646 bool skip_on_failure
= false;
647 unsigned long next_skip_pfn
= 0;
650 * Ensure that there are not too many pages isolated from the LRU
651 * list by either parallel reclaimers or compaction. If there are,
652 * delay for some time until fewer pages are isolated
654 while (unlikely(too_many_isolated(zone
))) {
655 /* async migration should just abort */
656 if (cc
->mode
== MIGRATE_ASYNC
)
659 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
661 if (fatal_signal_pending(current
))
665 if (compact_should_abort(cc
))
668 if (cc
->direct_compaction
&& (cc
->mode
== MIGRATE_ASYNC
)) {
669 skip_on_failure
= true;
670 next_skip_pfn
= block_end_pfn(low_pfn
, cc
->order
);
673 /* Time to isolate some pages for migration */
674 for (; low_pfn
< end_pfn
; low_pfn
++) {
677 if (skip_on_failure
&& low_pfn
>= next_skip_pfn
) {
679 * We have isolated all migration candidates in the
680 * previous order-aligned block, and did not skip it due
681 * to failure. We should migrate the pages now and
682 * hopefully succeed compaction.
688 * We failed to isolate in the previous order-aligned
689 * block. Set the new boundary to the end of the
690 * current block. Note we can't simply increase
691 * next_skip_pfn by 1 << order, as low_pfn might have
692 * been incremented by a higher number due to skipping
693 * a compound or a high-order buddy page in the
694 * previous loop iteration.
696 next_skip_pfn
= block_end_pfn(low_pfn
, cc
->order
);
700 * Periodically drop the lock (if held) regardless of its
701 * contention, to give chance to IRQs. Abort async compaction
704 if (!(low_pfn
% SWAP_CLUSTER_MAX
)
705 && compact_unlock_should_abort(&zone
->lru_lock
, flags
,
709 if (!pfn_valid_within(low_pfn
))
713 page
= pfn_to_page(low_pfn
);
719 * Skip if free. We read page order here without zone lock
720 * which is generally unsafe, but the race window is small and
721 * the worst thing that can happen is that we skip some
722 * potential isolation targets.
724 if (PageBuddy(page
)) {
725 unsigned long freepage_order
= page_order_unsafe(page
);
728 * Without lock, we cannot be sure that what we got is
729 * a valid page order. Consider only values in the
730 * valid order range to prevent low_pfn overflow.
732 if (freepage_order
> 0 && freepage_order
< MAX_ORDER
)
733 low_pfn
+= (1UL << freepage_order
) - 1;
738 * Check may be lockless but that's ok as we recheck later.
739 * It's possible to migrate LRU pages and balloon pages
740 * Skip any other type of page
742 is_lru
= PageLRU(page
);
744 if (unlikely(balloon_page_movable(page
))) {
745 if (balloon_page_isolate(page
)) {
746 /* Successfully isolated */
747 goto isolate_success
;
753 * Regardless of being on LRU, compound pages such as THP and
754 * hugetlbfs are not to be compacted. We can potentially save
755 * a lot of iterations if we skip them at once. The check is
756 * racy, but we can consider only valid values and the only
757 * danger is skipping too much.
759 if (PageCompound(page
)) {
760 unsigned int comp_order
= compound_order(page
);
762 if (likely(comp_order
< MAX_ORDER
))
763 low_pfn
+= (1UL << comp_order
) - 1;
772 * Migration will fail if an anonymous page is pinned in memory,
773 * so avoid taking lru_lock and isolating it unnecessarily in an
774 * admittedly racy check.
776 if (!page_mapping(page
) &&
777 page_count(page
) > page_mapcount(page
))
780 /* If we already hold the lock, we can skip some rechecking */
782 locked
= compact_trylock_irqsave(&zone
->lru_lock
,
787 /* Recheck PageLRU and PageCompound under lock */
792 * Page become compound since the non-locked check,
793 * and it's on LRU. It can only be a THP so the order
794 * is safe to read and it's 0 for tail pages.
796 if (unlikely(PageCompound(page
))) {
797 low_pfn
+= (1UL << compound_order(page
)) - 1;
802 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
804 /* Try isolate the page */
805 if (__isolate_lru_page(page
, isolate_mode
) != 0)
808 VM_BUG_ON_PAGE(PageCompound(page
), page
);
810 /* Successfully isolated */
811 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
814 list_add(&page
->lru
, &cc
->migratepages
);
815 cc
->nr_migratepages
++;
819 * Record where we could have freed pages by migration and not
820 * yet flushed them to buddy allocator.
821 * - this is the lowest page that was isolated and likely be
822 * then freed by migration.
824 if (!cc
->last_migrated_pfn
)
825 cc
->last_migrated_pfn
= low_pfn
;
827 /* Avoid isolating too much */
828 if (cc
->nr_migratepages
== COMPACT_CLUSTER_MAX
) {
835 if (!skip_on_failure
)
839 * We have isolated some pages, but then failed. Release them
840 * instead of migrating, as we cannot form the cc->order buddy
845 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
848 acct_isolated(zone
, cc
);
849 putback_movable_pages(&cc
->migratepages
);
850 cc
->nr_migratepages
= 0;
851 cc
->last_migrated_pfn
= 0;
855 if (low_pfn
< next_skip_pfn
) {
856 low_pfn
= next_skip_pfn
- 1;
858 * The check near the loop beginning would have updated
859 * next_skip_pfn too, but this is a bit simpler.
861 next_skip_pfn
+= 1UL << cc
->order
;
866 * The PageBuddy() check could have potentially brought us outside
867 * the range to be scanned.
869 if (unlikely(low_pfn
> end_pfn
))
873 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
876 * Update the pageblock-skip information and cached scanner pfn,
877 * if the whole pageblock was scanned without isolating any page.
879 if (low_pfn
== end_pfn
)
880 update_pageblock_skip(cc
, valid_page
, nr_isolated
, true);
882 trace_mm_compaction_isolate_migratepages(start_pfn
, low_pfn
,
883 nr_scanned
, nr_isolated
);
885 count_compact_events(COMPACTMIGRATE_SCANNED
, nr_scanned
);
887 count_compact_events(COMPACTISOLATED
, nr_isolated
);
893 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
894 * @cc: Compaction control structure.
895 * @start_pfn: The first PFN to start isolating.
896 * @end_pfn: The one-past-last PFN.
898 * Returns zero if isolation fails fatally due to e.g. pending signal.
899 * Otherwise, function returns one-past-the-last PFN of isolated page
900 * (which may be greater than end_pfn if end fell in a middle of a THP page).
903 isolate_migratepages_range(struct compact_control
*cc
, unsigned long start_pfn
,
904 unsigned long end_pfn
)
906 unsigned long pfn
, block_start_pfn
, block_end_pfn
;
908 /* Scan block by block. First and last block may be incomplete */
910 block_start_pfn
= pageblock_start_pfn(pfn
);
911 if (block_start_pfn
< cc
->zone
->zone_start_pfn
)
912 block_start_pfn
= cc
->zone
->zone_start_pfn
;
913 block_end_pfn
= pageblock_end_pfn(pfn
);
915 for (; pfn
< end_pfn
; pfn
= block_end_pfn
,
916 block_start_pfn
= block_end_pfn
,
917 block_end_pfn
+= pageblock_nr_pages
) {
919 block_end_pfn
= min(block_end_pfn
, end_pfn
);
921 if (!pageblock_pfn_to_page(block_start_pfn
,
922 block_end_pfn
, cc
->zone
))
925 pfn
= isolate_migratepages_block(cc
, pfn
, block_end_pfn
,
926 ISOLATE_UNEVICTABLE
);
931 if (cc
->nr_migratepages
== COMPACT_CLUSTER_MAX
)
934 acct_isolated(cc
->zone
, cc
);
939 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
940 #ifdef CONFIG_COMPACTION
942 /* Returns true if the page is within a block suitable for migration to */
943 static bool suitable_migration_target(struct page
*page
)
945 /* If the page is a large free page, then disallow migration */
946 if (PageBuddy(page
)) {
948 * We are checking page_order without zone->lock taken. But
949 * the only small danger is that we skip a potentially suitable
950 * pageblock, so it's not worth to check order for valid range.
952 if (page_order_unsafe(page
) >= pageblock_order
)
956 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
957 if (migrate_async_suitable(get_pageblock_migratetype(page
)))
960 /* Otherwise skip the block */
965 * Test whether the free scanner has reached the same or lower pageblock than
966 * the migration scanner, and compaction should thus terminate.
968 static inline bool compact_scanners_met(struct compact_control
*cc
)
970 return (cc
->free_pfn
>> pageblock_order
)
971 <= (cc
->migrate_pfn
>> pageblock_order
);
975 * Based on information in the current compact_control, find blocks
976 * suitable for isolating free pages from and then isolate them.
978 static void isolate_freepages(struct compact_control
*cc
)
980 struct zone
*zone
= cc
->zone
;
982 unsigned long block_start_pfn
; /* start of current pageblock */
983 unsigned long isolate_start_pfn
; /* exact pfn we start at */
984 unsigned long block_end_pfn
; /* end of current pageblock */
985 unsigned long low_pfn
; /* lowest pfn scanner is able to scan */
986 struct list_head
*freelist
= &cc
->freepages
;
989 * Initialise the free scanner. The starting point is where we last
990 * successfully isolated from, zone-cached value, or the end of the
991 * zone when isolating for the first time. For looping we also need
992 * this pfn aligned down to the pageblock boundary, because we do
993 * block_start_pfn -= pageblock_nr_pages in the for loop.
994 * For ending point, take care when isolating in last pageblock of a
995 * a zone which ends in the middle of a pageblock.
996 * The low boundary is the end of the pageblock the migration scanner
999 isolate_start_pfn
= cc
->free_pfn
;
1000 block_start_pfn
= pageblock_start_pfn(cc
->free_pfn
);
1001 block_end_pfn
= min(block_start_pfn
+ pageblock_nr_pages
,
1002 zone_end_pfn(zone
));
1003 low_pfn
= pageblock_end_pfn(cc
->migrate_pfn
);
1006 * Isolate free pages until enough are available to migrate the
1007 * pages on cc->migratepages. We stop searching if the migrate
1008 * and free page scanners meet or enough free pages are isolated.
1010 for (; block_start_pfn
>= low_pfn
;
1011 block_end_pfn
= block_start_pfn
,
1012 block_start_pfn
-= pageblock_nr_pages
,
1013 isolate_start_pfn
= block_start_pfn
) {
1016 * This can iterate a massively long zone without finding any
1017 * suitable migration targets, so periodically check if we need
1018 * to schedule, or even abort async compaction.
1020 if (!(block_start_pfn
% (SWAP_CLUSTER_MAX
* pageblock_nr_pages
))
1021 && compact_should_abort(cc
))
1024 page
= pageblock_pfn_to_page(block_start_pfn
, block_end_pfn
,
1029 /* Check the block is suitable for migration */
1030 if (!suitable_migration_target(page
))
1033 /* If isolation recently failed, do not retry */
1034 if (!isolation_suitable(cc
, page
))
1037 /* Found a block suitable for isolating free pages from. */
1038 isolate_freepages_block(cc
, &isolate_start_pfn
,
1039 block_end_pfn
, freelist
, false);
1042 * If we isolated enough freepages, or aborted due to async
1043 * compaction being contended, terminate the loop.
1044 * Remember where the free scanner should restart next time,
1045 * which is where isolate_freepages_block() left off.
1046 * But if it scanned the whole pageblock, isolate_start_pfn
1047 * now points at block_end_pfn, which is the start of the next
1049 * In that case we will however want to restart at the start
1050 * of the previous pageblock.
1052 if ((cc
->nr_freepages
>= cc
->nr_migratepages
)
1054 if (isolate_start_pfn
>= block_end_pfn
)
1056 block_start_pfn
- pageblock_nr_pages
;
1060 * isolate_freepages_block() should not terminate
1061 * prematurely unless contended, or isolated enough
1063 VM_BUG_ON(isolate_start_pfn
< block_end_pfn
);
1067 /* split_free_page does not map the pages */
1068 map_pages(freelist
);
1071 * Record where the free scanner will restart next time. Either we
1072 * broke from the loop and set isolate_start_pfn based on the last
1073 * call to isolate_freepages_block(), or we met the migration scanner
1074 * and the loop terminated due to isolate_start_pfn < low_pfn
1076 cc
->free_pfn
= isolate_start_pfn
;
1080 * This is a migrate-callback that "allocates" freepages by taking pages
1081 * from the isolated freelists in the block we are migrating to.
1083 static struct page
*compaction_alloc(struct page
*migratepage
,
1087 struct compact_control
*cc
= (struct compact_control
*)data
;
1088 struct page
*freepage
;
1091 * Isolate free pages if necessary, and if we are not aborting due to
1094 if (list_empty(&cc
->freepages
)) {
1096 isolate_freepages(cc
);
1098 if (list_empty(&cc
->freepages
))
1102 freepage
= list_entry(cc
->freepages
.next
, struct page
, lru
);
1103 list_del(&freepage
->lru
);
1110 * This is a migrate-callback that "frees" freepages back to the isolated
1111 * freelist. All pages on the freelist are from the same zone, so there is no
1112 * special handling needed for NUMA.
1114 static void compaction_free(struct page
*page
, unsigned long data
)
1116 struct compact_control
*cc
= (struct compact_control
*)data
;
1118 list_add(&page
->lru
, &cc
->freepages
);
1122 /* possible outcome of isolate_migratepages */
1124 ISOLATE_ABORT
, /* Abort compaction now */
1125 ISOLATE_NONE
, /* No pages isolated, continue scanning */
1126 ISOLATE_SUCCESS
, /* Pages isolated, migrate */
1127 } isolate_migrate_t
;
1130 * Allow userspace to control policy on scanning the unevictable LRU for
1131 * compactable pages.
1133 int sysctl_compact_unevictable_allowed __read_mostly
= 1;
1136 * Isolate all pages that can be migrated from the first suitable block,
1137 * starting at the block pointed to by the migrate scanner pfn within
1140 static isolate_migrate_t
isolate_migratepages(struct zone
*zone
,
1141 struct compact_control
*cc
)
1143 unsigned long block_start_pfn
;
1144 unsigned long block_end_pfn
;
1145 unsigned long low_pfn
;
1147 const isolate_mode_t isolate_mode
=
1148 (sysctl_compact_unevictable_allowed
? ISOLATE_UNEVICTABLE
: 0) |
1149 (cc
->mode
== MIGRATE_ASYNC
? ISOLATE_ASYNC_MIGRATE
: 0);
1152 * Start at where we last stopped, or beginning of the zone as
1153 * initialized by compact_zone()
1155 low_pfn
= cc
->migrate_pfn
;
1156 block_start_pfn
= pageblock_start_pfn(low_pfn
);
1157 if (block_start_pfn
< zone
->zone_start_pfn
)
1158 block_start_pfn
= zone
->zone_start_pfn
;
1160 /* Only scan within a pageblock boundary */
1161 block_end_pfn
= pageblock_end_pfn(low_pfn
);
1164 * Iterate over whole pageblocks until we find the first suitable.
1165 * Do not cross the free scanner.
1167 for (; block_end_pfn
<= cc
->free_pfn
;
1168 low_pfn
= block_end_pfn
,
1169 block_start_pfn
= block_end_pfn
,
1170 block_end_pfn
+= pageblock_nr_pages
) {
1173 * This can potentially iterate a massively long zone with
1174 * many pageblocks unsuitable, so periodically check if we
1175 * need to schedule, or even abort async compaction.
1177 if (!(low_pfn
% (SWAP_CLUSTER_MAX
* pageblock_nr_pages
))
1178 && compact_should_abort(cc
))
1181 page
= pageblock_pfn_to_page(block_start_pfn
, block_end_pfn
,
1186 /* If isolation recently failed, do not retry */
1187 if (!isolation_suitable(cc
, page
))
1191 * For async compaction, also only scan in MOVABLE blocks.
1192 * Async compaction is optimistic to see if the minimum amount
1193 * of work satisfies the allocation.
1195 if (cc
->mode
== MIGRATE_ASYNC
&&
1196 !migrate_async_suitable(get_pageblock_migratetype(page
)))
1199 /* Perform the isolation */
1200 low_pfn
= isolate_migratepages_block(cc
, low_pfn
,
1201 block_end_pfn
, isolate_mode
);
1203 if (!low_pfn
|| cc
->contended
) {
1204 acct_isolated(zone
, cc
);
1205 return ISOLATE_ABORT
;
1209 * Either we isolated something and proceed with migration. Or
1210 * we failed and compact_zone should decide if we should
1216 acct_isolated(zone
, cc
);
1217 /* Record where migration scanner will be restarted. */
1218 cc
->migrate_pfn
= low_pfn
;
1220 return cc
->nr_migratepages
? ISOLATE_SUCCESS
: ISOLATE_NONE
;
1224 * order == -1 is expected when compacting via
1225 * /proc/sys/vm/compact_memory
1227 static inline bool is_via_compact_memory(int order
)
1232 static enum compact_result
__compact_finished(struct zone
*zone
, struct compact_control
*cc
,
1233 const int migratetype
)
1236 unsigned long watermark
;
1238 if (cc
->contended
|| fatal_signal_pending(current
))
1239 return COMPACT_CONTENDED
;
1241 /* Compaction run completes if the migrate and free scanner meet */
1242 if (compact_scanners_met(cc
)) {
1243 /* Let the next compaction start anew. */
1244 reset_cached_positions(zone
);
1247 * Mark that the PG_migrate_skip information should be cleared
1248 * by kswapd when it goes to sleep. kcompactd does not set the
1249 * flag itself as the decision to be clear should be directly
1250 * based on an allocation request.
1252 if (cc
->direct_compaction
)
1253 zone
->compact_blockskip_flush
= true;
1256 return COMPACT_COMPLETE
;
1258 return COMPACT_PARTIAL_SKIPPED
;
1261 if (is_via_compact_memory(cc
->order
))
1262 return COMPACT_CONTINUE
;
1264 /* Compaction run is not finished if the watermark is not met */
1265 watermark
= low_wmark_pages(zone
);
1267 if (!zone_watermark_ok(zone
, cc
->order
, watermark
, cc
->classzone_idx
,
1269 return COMPACT_CONTINUE
;
1271 /* Direct compactor: Is a suitable page free? */
1272 for (order
= cc
->order
; order
< MAX_ORDER
; order
++) {
1273 struct free_area
*area
= &zone
->free_area
[order
];
1276 /* Job done if page is free of the right migratetype */
1277 if (!list_empty(&area
->free_list
[migratetype
]))
1278 return COMPACT_PARTIAL
;
1281 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
1282 if (migratetype
== MIGRATE_MOVABLE
&&
1283 !list_empty(&area
->free_list
[MIGRATE_CMA
]))
1284 return COMPACT_PARTIAL
;
1287 * Job done if allocation would steal freepages from
1288 * other migratetype buddy lists.
1290 if (find_suitable_fallback(area
, order
, migratetype
,
1291 true, &can_steal
) != -1)
1292 return COMPACT_PARTIAL
;
1295 return COMPACT_NO_SUITABLE_PAGE
;
1298 static enum compact_result
compact_finished(struct zone
*zone
,
1299 struct compact_control
*cc
,
1300 const int migratetype
)
1304 ret
= __compact_finished(zone
, cc
, migratetype
);
1305 trace_mm_compaction_finished(zone
, cc
->order
, ret
);
1306 if (ret
== COMPACT_NO_SUITABLE_PAGE
)
1307 ret
= COMPACT_CONTINUE
;
1313 * compaction_suitable: Is this suitable to run compaction on this zone now?
1315 * COMPACT_SKIPPED - If there are too few free pages for compaction
1316 * COMPACT_PARTIAL - If the allocation would succeed without compaction
1317 * COMPACT_CONTINUE - If compaction should run now
1319 static enum compact_result
__compaction_suitable(struct zone
*zone
, int order
,
1320 unsigned int alloc_flags
,
1322 unsigned long wmark_target
)
1325 unsigned long watermark
;
1327 if (is_via_compact_memory(order
))
1328 return COMPACT_CONTINUE
;
1330 watermark
= low_wmark_pages(zone
);
1332 * If watermarks for high-order allocation are already met, there
1333 * should be no need for compaction at all.
1335 if (zone_watermark_ok(zone
, order
, watermark
, classzone_idx
,
1337 return COMPACT_PARTIAL
;
1340 * Watermarks for order-0 must be met for compaction. Note the 2UL.
1341 * This is because during migration, copies of pages need to be
1342 * allocated and for a short time, the footprint is higher
1344 watermark
+= (2UL << order
);
1345 if (!__zone_watermark_ok(zone
, 0, watermark
, classzone_idx
,
1346 alloc_flags
, wmark_target
))
1347 return COMPACT_SKIPPED
;
1350 * fragmentation index determines if allocation failures are due to
1351 * low memory or external fragmentation
1353 * index of -1000 would imply allocations might succeed depending on
1354 * watermarks, but we already failed the high-order watermark check
1355 * index towards 0 implies failure is due to lack of memory
1356 * index towards 1000 implies failure is due to fragmentation
1358 * Only compact if a failure would be due to fragmentation.
1360 fragindex
= fragmentation_index(zone
, order
);
1361 if (fragindex
>= 0 && fragindex
<= sysctl_extfrag_threshold
)
1362 return COMPACT_NOT_SUITABLE_ZONE
;
1364 return COMPACT_CONTINUE
;
1367 enum compact_result
compaction_suitable(struct zone
*zone
, int order
,
1368 unsigned int alloc_flags
,
1371 enum compact_result ret
;
1373 ret
= __compaction_suitable(zone
, order
, alloc_flags
, classzone_idx
,
1374 zone_page_state(zone
, NR_FREE_PAGES
));
1375 trace_mm_compaction_suitable(zone
, order
, ret
);
1376 if (ret
== COMPACT_NOT_SUITABLE_ZONE
)
1377 ret
= COMPACT_SKIPPED
;
1382 bool compaction_zonelist_suitable(struct alloc_context
*ac
, int order
,
1389 * Make sure at least one zone would pass __compaction_suitable if we continue
1390 * retrying the reclaim.
1392 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
1394 unsigned long available
;
1395 enum compact_result compact_result
;
1398 * Do not consider all the reclaimable memory because we do not
1399 * want to trash just for a single high order allocation which
1400 * is even not guaranteed to appear even if __compaction_suitable
1401 * is happy about the watermark check.
1403 available
= zone_reclaimable_pages(zone
) / order
;
1404 available
+= zone_page_state_snapshot(zone
, NR_FREE_PAGES
);
1405 compact_result
= __compaction_suitable(zone
, order
, alloc_flags
,
1406 ac_classzone_idx(ac
), available
);
1407 if (compact_result
!= COMPACT_SKIPPED
&&
1408 compact_result
!= COMPACT_NOT_SUITABLE_ZONE
)
1415 static enum compact_result
compact_zone(struct zone
*zone
, struct compact_control
*cc
)
1417 enum compact_result ret
;
1418 unsigned long start_pfn
= zone
->zone_start_pfn
;
1419 unsigned long end_pfn
= zone_end_pfn(zone
);
1420 const int migratetype
= gfpflags_to_migratetype(cc
->gfp_mask
);
1421 const bool sync
= cc
->mode
!= MIGRATE_ASYNC
;
1423 ret
= compaction_suitable(zone
, cc
->order
, cc
->alloc_flags
,
1425 /* Compaction is likely to fail */
1426 if (ret
== COMPACT_PARTIAL
|| ret
== COMPACT_SKIPPED
)
1429 /* huh, compaction_suitable is returning something unexpected */
1430 VM_BUG_ON(ret
!= COMPACT_CONTINUE
);
1433 * Clear pageblock skip if there were failures recently and compaction
1434 * is about to be retried after being deferred.
1436 if (compaction_restarting(zone
, cc
->order
))
1437 __reset_isolation_suitable(zone
);
1440 * Setup to move all movable pages to the end of the zone. Used cached
1441 * information on where the scanners should start but check that it
1442 * is initialised by ensuring the values are within zone boundaries.
1444 cc
->migrate_pfn
= zone
->compact_cached_migrate_pfn
[sync
];
1445 cc
->free_pfn
= zone
->compact_cached_free_pfn
;
1446 if (cc
->free_pfn
< start_pfn
|| cc
->free_pfn
>= end_pfn
) {
1447 cc
->free_pfn
= pageblock_start_pfn(end_pfn
- 1);
1448 zone
->compact_cached_free_pfn
= cc
->free_pfn
;
1450 if (cc
->migrate_pfn
< start_pfn
|| cc
->migrate_pfn
>= end_pfn
) {
1451 cc
->migrate_pfn
= start_pfn
;
1452 zone
->compact_cached_migrate_pfn
[0] = cc
->migrate_pfn
;
1453 zone
->compact_cached_migrate_pfn
[1] = cc
->migrate_pfn
;
1456 if (cc
->migrate_pfn
== start_pfn
)
1457 cc
->whole_zone
= true;
1459 cc
->last_migrated_pfn
= 0;
1461 trace_mm_compaction_begin(start_pfn
, cc
->migrate_pfn
,
1462 cc
->free_pfn
, end_pfn
, sync
);
1464 migrate_prep_local();
1466 while ((ret
= compact_finished(zone
, cc
, migratetype
)) ==
1470 switch (isolate_migratepages(zone
, cc
)) {
1472 ret
= COMPACT_CONTENDED
;
1473 putback_movable_pages(&cc
->migratepages
);
1474 cc
->nr_migratepages
= 0;
1478 * We haven't isolated and migrated anything, but
1479 * there might still be unflushed migrations from
1480 * previous cc->order aligned block.
1483 case ISOLATE_SUCCESS
:
1487 err
= migrate_pages(&cc
->migratepages
, compaction_alloc
,
1488 compaction_free
, (unsigned long)cc
, cc
->mode
,
1491 trace_mm_compaction_migratepages(cc
->nr_migratepages
, err
,
1494 /* All pages were either migrated or will be released */
1495 cc
->nr_migratepages
= 0;
1497 putback_movable_pages(&cc
->migratepages
);
1499 * migrate_pages() may return -ENOMEM when scanners meet
1500 * and we want compact_finished() to detect it
1502 if (err
== -ENOMEM
&& !compact_scanners_met(cc
)) {
1503 ret
= COMPACT_CONTENDED
;
1507 * We failed to migrate at least one page in the current
1508 * order-aligned block, so skip the rest of it.
1510 if (cc
->direct_compaction
&&
1511 (cc
->mode
== MIGRATE_ASYNC
)) {
1512 cc
->migrate_pfn
= block_end_pfn(
1513 cc
->migrate_pfn
- 1, cc
->order
);
1514 /* Draining pcplists is useless in this case */
1515 cc
->last_migrated_pfn
= 0;
1522 * Has the migration scanner moved away from the previous
1523 * cc->order aligned block where we migrated from? If yes,
1524 * flush the pages that were freed, so that they can merge and
1525 * compact_finished() can detect immediately if allocation
1528 if (cc
->order
> 0 && cc
->last_migrated_pfn
) {
1530 unsigned long current_block_start
=
1531 block_start_pfn(cc
->migrate_pfn
, cc
->order
);
1533 if (cc
->last_migrated_pfn
< current_block_start
) {
1535 lru_add_drain_cpu(cpu
);
1536 drain_local_pages(zone
);
1538 /* No more flushing until we migrate again */
1539 cc
->last_migrated_pfn
= 0;
1547 * Release free pages and update where the free scanner should restart,
1548 * so we don't leave any returned pages behind in the next attempt.
1550 if (cc
->nr_freepages
> 0) {
1551 unsigned long free_pfn
= release_freepages(&cc
->freepages
);
1553 cc
->nr_freepages
= 0;
1554 VM_BUG_ON(free_pfn
== 0);
1555 /* The cached pfn is always the first in a pageblock */
1556 free_pfn
= pageblock_start_pfn(free_pfn
);
1558 * Only go back, not forward. The cached pfn might have been
1559 * already reset to zone end in compact_finished()
1561 if (free_pfn
> zone
->compact_cached_free_pfn
)
1562 zone
->compact_cached_free_pfn
= free_pfn
;
1565 trace_mm_compaction_end(start_pfn
, cc
->migrate_pfn
,
1566 cc
->free_pfn
, end_pfn
, sync
, ret
);
1568 if (ret
== COMPACT_CONTENDED
)
1569 ret
= COMPACT_PARTIAL
;
1574 static enum compact_result
compact_zone_order(struct zone
*zone
, int order
,
1575 gfp_t gfp_mask
, enum migrate_mode mode
, int *contended
,
1576 unsigned int alloc_flags
, int classzone_idx
)
1578 enum compact_result ret
;
1579 struct compact_control cc
= {
1581 .nr_migratepages
= 0,
1583 .gfp_mask
= gfp_mask
,
1586 .alloc_flags
= alloc_flags
,
1587 .classzone_idx
= classzone_idx
,
1588 .direct_compaction
= true,
1590 INIT_LIST_HEAD(&cc
.freepages
);
1591 INIT_LIST_HEAD(&cc
.migratepages
);
1593 ret
= compact_zone(zone
, &cc
);
1595 VM_BUG_ON(!list_empty(&cc
.freepages
));
1596 VM_BUG_ON(!list_empty(&cc
.migratepages
));
1598 *contended
= cc
.contended
;
1602 int sysctl_extfrag_threshold
= 500;
1605 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1606 * @gfp_mask: The GFP mask of the current allocation
1607 * @order: The order of the current allocation
1608 * @alloc_flags: The allocation flags of the current allocation
1609 * @ac: The context of current allocation
1610 * @mode: The migration mode for async, sync light, or sync migration
1611 * @contended: Return value that determines if compaction was aborted due to
1612 * need_resched() or lock contention
1614 * This is the main entry point for direct page compaction.
1616 enum compact_result
try_to_compact_pages(gfp_t gfp_mask
, unsigned int order
,
1617 unsigned int alloc_flags
, const struct alloc_context
*ac
,
1618 enum migrate_mode mode
, int *contended
)
1620 int may_enter_fs
= gfp_mask
& __GFP_FS
;
1621 int may_perform_io
= gfp_mask
& __GFP_IO
;
1624 enum compact_result rc
= COMPACT_SKIPPED
;
1625 int all_zones_contended
= COMPACT_CONTENDED_LOCK
; /* init for &= op */
1627 *contended
= COMPACT_CONTENDED_NONE
;
1629 /* Check if the GFP flags allow compaction */
1630 if (!order
|| !may_enter_fs
|| !may_perform_io
)
1631 return COMPACT_SKIPPED
;
1633 trace_mm_compaction_try_to_compact_pages(order
, gfp_mask
, mode
);
1635 /* Compact each zone in the list */
1636 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
, ac
->high_zoneidx
,
1638 enum compact_result status
;
1641 if (compaction_deferred(zone
, order
)) {
1642 rc
= max_t(enum compact_result
, COMPACT_DEFERRED
, rc
);
1646 status
= compact_zone_order(zone
, order
, gfp_mask
, mode
,
1647 &zone_contended
, alloc_flags
,
1648 ac_classzone_idx(ac
));
1649 rc
= max(status
, rc
);
1651 * It takes at least one zone that wasn't lock contended
1652 * to clear all_zones_contended.
1654 all_zones_contended
&= zone_contended
;
1656 /* If a normal allocation would succeed, stop compacting */
1657 if (zone_watermark_ok(zone
, order
, low_wmark_pages(zone
),
1658 ac_classzone_idx(ac
), alloc_flags
)) {
1660 * We think the allocation will succeed in this zone,
1661 * but it is not certain, hence the false. The caller
1662 * will repeat this with true if allocation indeed
1663 * succeeds in this zone.
1665 compaction_defer_reset(zone
, order
, false);
1667 * It is possible that async compaction aborted due to
1668 * need_resched() and the watermarks were ok thanks to
1669 * somebody else freeing memory. The allocation can
1670 * however still fail so we better signal the
1671 * need_resched() contention anyway (this will not
1672 * prevent the allocation attempt).
1674 if (zone_contended
== COMPACT_CONTENDED_SCHED
)
1675 *contended
= COMPACT_CONTENDED_SCHED
;
1680 if (mode
!= MIGRATE_ASYNC
&& (status
== COMPACT_COMPLETE
||
1681 status
== COMPACT_PARTIAL_SKIPPED
)) {
1683 * We think that allocation won't succeed in this zone
1684 * so we defer compaction there. If it ends up
1685 * succeeding after all, it will be reset.
1687 defer_compaction(zone
, order
);
1691 * We might have stopped compacting due to need_resched() in
1692 * async compaction, or due to a fatal signal detected. In that
1693 * case do not try further zones and signal need_resched()
1696 if ((zone_contended
== COMPACT_CONTENDED_SCHED
)
1697 || fatal_signal_pending(current
)) {
1698 *contended
= COMPACT_CONTENDED_SCHED
;
1705 * We might not have tried all the zones, so be conservative
1706 * and assume they are not all lock contended.
1708 all_zones_contended
= 0;
1713 * If at least one zone wasn't deferred or skipped, we report if all
1714 * zones that were tried were lock contended.
1716 if (rc
> COMPACT_INACTIVE
&& all_zones_contended
)
1717 *contended
= COMPACT_CONTENDED_LOCK
;
1723 /* Compact all zones within a node */
1724 static void __compact_pgdat(pg_data_t
*pgdat
, struct compact_control
*cc
)
1729 for (zoneid
= 0; zoneid
< MAX_NR_ZONES
; zoneid
++) {
1731 zone
= &pgdat
->node_zones
[zoneid
];
1732 if (!populated_zone(zone
))
1735 cc
->nr_freepages
= 0;
1736 cc
->nr_migratepages
= 0;
1738 INIT_LIST_HEAD(&cc
->freepages
);
1739 INIT_LIST_HEAD(&cc
->migratepages
);
1742 * When called via /proc/sys/vm/compact_memory
1743 * this makes sure we compact the whole zone regardless of
1744 * cached scanner positions.
1746 if (is_via_compact_memory(cc
->order
))
1747 __reset_isolation_suitable(zone
);
1749 if (is_via_compact_memory(cc
->order
) ||
1750 !compaction_deferred(zone
, cc
->order
))
1751 compact_zone(zone
, cc
);
1753 VM_BUG_ON(!list_empty(&cc
->freepages
));
1754 VM_BUG_ON(!list_empty(&cc
->migratepages
));
1756 if (is_via_compact_memory(cc
->order
))
1759 if (zone_watermark_ok(zone
, cc
->order
,
1760 low_wmark_pages(zone
), 0, 0))
1761 compaction_defer_reset(zone
, cc
->order
, false);
1765 void compact_pgdat(pg_data_t
*pgdat
, int order
)
1767 struct compact_control cc
= {
1769 .mode
= MIGRATE_ASYNC
,
1775 __compact_pgdat(pgdat
, &cc
);
1778 static void compact_node(int nid
)
1780 struct compact_control cc
= {
1782 .mode
= MIGRATE_SYNC
,
1783 .ignore_skip_hint
= true,
1786 __compact_pgdat(NODE_DATA(nid
), &cc
);
1789 /* Compact all nodes in the system */
1790 static void compact_nodes(void)
1794 /* Flush pending updates to the LRU lists */
1795 lru_add_drain_all();
1797 for_each_online_node(nid
)
1801 /* The written value is actually unused, all memory is compacted */
1802 int sysctl_compact_memory
;
1805 * This is the entry point for compacting all nodes via
1806 * /proc/sys/vm/compact_memory
1808 int sysctl_compaction_handler(struct ctl_table
*table
, int write
,
1809 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
1817 int sysctl_extfrag_handler(struct ctl_table
*table
, int write
,
1818 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
1820 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
1825 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1826 static ssize_t
sysfs_compact_node(struct device
*dev
,
1827 struct device_attribute
*attr
,
1828 const char *buf
, size_t count
)
1832 if (nid
>= 0 && nid
< nr_node_ids
&& node_online(nid
)) {
1833 /* Flush pending updates to the LRU lists */
1834 lru_add_drain_all();
1841 static DEVICE_ATTR(compact
, S_IWUSR
, NULL
, sysfs_compact_node
);
1843 int compaction_register_node(struct node
*node
)
1845 return device_create_file(&node
->dev
, &dev_attr_compact
);
1848 void compaction_unregister_node(struct node
*node
)
1850 return device_remove_file(&node
->dev
, &dev_attr_compact
);
1852 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1854 static inline bool kcompactd_work_requested(pg_data_t
*pgdat
)
1856 return pgdat
->kcompactd_max_order
> 0 || kthread_should_stop();
1859 static bool kcompactd_node_suitable(pg_data_t
*pgdat
)
1863 enum zone_type classzone_idx
= pgdat
->kcompactd_classzone_idx
;
1865 for (zoneid
= 0; zoneid
<= classzone_idx
; zoneid
++) {
1866 zone
= &pgdat
->node_zones
[zoneid
];
1868 if (!populated_zone(zone
))
1871 if (compaction_suitable(zone
, pgdat
->kcompactd_max_order
, 0,
1872 classzone_idx
) == COMPACT_CONTINUE
)
1879 static void kcompactd_do_work(pg_data_t
*pgdat
)
1882 * With no special task, compact all zones so that a page of requested
1883 * order is allocatable.
1887 struct compact_control cc
= {
1888 .order
= pgdat
->kcompactd_max_order
,
1889 .classzone_idx
= pgdat
->kcompactd_classzone_idx
,
1890 .mode
= MIGRATE_SYNC_LIGHT
,
1891 .ignore_skip_hint
= true,
1894 bool success
= false;
1896 trace_mm_compaction_kcompactd_wake(pgdat
->node_id
, cc
.order
,
1898 count_vm_event(KCOMPACTD_WAKE
);
1900 for (zoneid
= 0; zoneid
<= cc
.classzone_idx
; zoneid
++) {
1903 zone
= &pgdat
->node_zones
[zoneid
];
1904 if (!populated_zone(zone
))
1907 if (compaction_deferred(zone
, cc
.order
))
1910 if (compaction_suitable(zone
, cc
.order
, 0, zoneid
) !=
1914 cc
.nr_freepages
= 0;
1915 cc
.nr_migratepages
= 0;
1917 INIT_LIST_HEAD(&cc
.freepages
);
1918 INIT_LIST_HEAD(&cc
.migratepages
);
1920 if (kthread_should_stop())
1922 status
= compact_zone(zone
, &cc
);
1924 if (zone_watermark_ok(zone
, cc
.order
, low_wmark_pages(zone
),
1925 cc
.classzone_idx
, 0)) {
1927 compaction_defer_reset(zone
, cc
.order
, false);
1928 } else if (status
== COMPACT_PARTIAL_SKIPPED
|| status
== COMPACT_COMPLETE
) {
1930 * We use sync migration mode here, so we defer like
1931 * sync direct compaction does.
1933 defer_compaction(zone
, cc
.order
);
1936 VM_BUG_ON(!list_empty(&cc
.freepages
));
1937 VM_BUG_ON(!list_empty(&cc
.migratepages
));
1941 * Regardless of success, we are done until woken up next. But remember
1942 * the requested order/classzone_idx in case it was higher/tighter than
1945 if (pgdat
->kcompactd_max_order
<= cc
.order
)
1946 pgdat
->kcompactd_max_order
= 0;
1947 if (pgdat
->kcompactd_classzone_idx
>= cc
.classzone_idx
)
1948 pgdat
->kcompactd_classzone_idx
= pgdat
->nr_zones
- 1;
1951 void wakeup_kcompactd(pg_data_t
*pgdat
, int order
, int classzone_idx
)
1956 if (pgdat
->kcompactd_max_order
< order
)
1957 pgdat
->kcompactd_max_order
= order
;
1959 if (pgdat
->kcompactd_classzone_idx
> classzone_idx
)
1960 pgdat
->kcompactd_classzone_idx
= classzone_idx
;
1962 if (!waitqueue_active(&pgdat
->kcompactd_wait
))
1965 if (!kcompactd_node_suitable(pgdat
))
1968 trace_mm_compaction_wakeup_kcompactd(pgdat
->node_id
, order
,
1970 wake_up_interruptible(&pgdat
->kcompactd_wait
);
1974 * The background compaction daemon, started as a kernel thread
1975 * from the init process.
1977 static int kcompactd(void *p
)
1979 pg_data_t
*pgdat
= (pg_data_t
*)p
;
1980 struct task_struct
*tsk
= current
;
1982 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
1984 if (!cpumask_empty(cpumask
))
1985 set_cpus_allowed_ptr(tsk
, cpumask
);
1989 pgdat
->kcompactd_max_order
= 0;
1990 pgdat
->kcompactd_classzone_idx
= pgdat
->nr_zones
- 1;
1992 while (!kthread_should_stop()) {
1993 trace_mm_compaction_kcompactd_sleep(pgdat
->node_id
);
1994 wait_event_freezable(pgdat
->kcompactd_wait
,
1995 kcompactd_work_requested(pgdat
));
1997 kcompactd_do_work(pgdat
);
2004 * This kcompactd start function will be called by init and node-hot-add.
2005 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
2007 int kcompactd_run(int nid
)
2009 pg_data_t
*pgdat
= NODE_DATA(nid
);
2012 if (pgdat
->kcompactd
)
2015 pgdat
->kcompactd
= kthread_run(kcompactd
, pgdat
, "kcompactd%d", nid
);
2016 if (IS_ERR(pgdat
->kcompactd
)) {
2017 pr_err("Failed to start kcompactd on node %d\n", nid
);
2018 ret
= PTR_ERR(pgdat
->kcompactd
);
2019 pgdat
->kcompactd
= NULL
;
2025 * Called by memory hotplug when all memory in a node is offlined. Caller must
2026 * hold mem_hotplug_begin/end().
2028 void kcompactd_stop(int nid
)
2030 struct task_struct
*kcompactd
= NODE_DATA(nid
)->kcompactd
;
2033 kthread_stop(kcompactd
);
2034 NODE_DATA(nid
)->kcompactd
= NULL
;
2039 * It's optimal to keep kcompactd on the same CPUs as their memory, but
2040 * not required for correctness. So if the last cpu in a node goes
2041 * away, we get changed to run anywhere: as the first one comes back,
2042 * restore their cpu bindings.
2044 static int cpu_callback(struct notifier_block
*nfb
, unsigned long action
,
2049 if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
) {
2050 for_each_node_state(nid
, N_MEMORY
) {
2051 pg_data_t
*pgdat
= NODE_DATA(nid
);
2052 const struct cpumask
*mask
;
2054 mask
= cpumask_of_node(pgdat
->node_id
);
2056 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
2057 /* One of our CPUs online: restore mask */
2058 set_cpus_allowed_ptr(pgdat
->kcompactd
, mask
);
2064 static int __init
kcompactd_init(void)
2068 for_each_node_state(nid
, N_MEMORY
)
2070 hotcpu_notifier(cpu_callback
, 0);
2073 subsys_initcall(kcompactd_init
)
2075 #endif /* CONFIG_COMPACTION */