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/swap.h>
11 #include <linux/migrate.h>
12 #include <linux/compaction.h>
13 #include <linux/mm_inline.h>
14 #include <linux/backing-dev.h>
15 #include <linux/sysctl.h>
16 #include <linux/sysfs.h>
17 #include <linux/balloon_compaction.h>
18 #include <linux/page-isolation.h>
21 #ifdef CONFIG_COMPACTION
22 static inline void count_compact_event(enum vm_event_item item
)
27 static inline void count_compact_events(enum vm_event_item item
, long delta
)
29 count_vm_events(item
, delta
);
32 #define count_compact_event(item) do { } while (0)
33 #define count_compact_events(item, delta) do { } while (0)
36 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
38 #define CREATE_TRACE_POINTS
39 #include <trace/events/compaction.h>
41 static unsigned long release_freepages(struct list_head
*freelist
)
43 struct page
*page
, *next
;
44 unsigned long count
= 0;
46 list_for_each_entry_safe(page
, next
, freelist
, lru
) {
55 static void map_pages(struct list_head
*list
)
59 list_for_each_entry(page
, list
, lru
) {
60 arch_alloc_page(page
, 0);
61 kernel_map_pages(page
, 1, 1);
65 static inline bool migrate_async_suitable(int migratetype
)
67 return is_migrate_cma(migratetype
) || migratetype
== MIGRATE_MOVABLE
;
70 #ifdef CONFIG_COMPACTION
71 /* Returns true if the pageblock should be scanned for pages to isolate. */
72 static inline bool isolation_suitable(struct compact_control
*cc
,
75 if (cc
->ignore_skip_hint
)
78 return !get_pageblock_skip(page
);
82 * This function is called to clear all cached information on pageblocks that
83 * should be skipped for page isolation when the migrate and free page scanner
86 static void __reset_isolation_suitable(struct zone
*zone
)
88 unsigned long start_pfn
= zone
->zone_start_pfn
;
89 unsigned long end_pfn
= zone_end_pfn(zone
);
92 zone
->compact_cached_migrate_pfn
[0] = start_pfn
;
93 zone
->compact_cached_migrate_pfn
[1] = start_pfn
;
94 zone
->compact_cached_free_pfn
= end_pfn
;
95 zone
->compact_blockskip_flush
= false;
97 /* Walk the zone and mark every pageblock as suitable for isolation */
98 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
106 page
= pfn_to_page(pfn
);
107 if (zone
!= page_zone(page
))
110 clear_pageblock_skip(page
);
114 void reset_isolation_suitable(pg_data_t
*pgdat
)
118 for (zoneid
= 0; zoneid
< MAX_NR_ZONES
; zoneid
++) {
119 struct zone
*zone
= &pgdat
->node_zones
[zoneid
];
120 if (!populated_zone(zone
))
123 /* Only flush if a full compaction finished recently */
124 if (zone
->compact_blockskip_flush
)
125 __reset_isolation_suitable(zone
);
130 * If no pages were isolated then mark this pageblock to be skipped in the
131 * future. The information is later cleared by __reset_isolation_suitable().
133 static void update_pageblock_skip(struct compact_control
*cc
,
134 struct page
*page
, unsigned long nr_isolated
,
135 bool set_unsuitable
, bool migrate_scanner
)
137 struct zone
*zone
= cc
->zone
;
140 if (cc
->ignore_skip_hint
)
150 * Only skip pageblocks when all forms of compaction will be known to
151 * fail in the near future.
154 set_pageblock_skip(page
);
156 pfn
= page_to_pfn(page
);
158 /* Update where async and sync compaction should restart */
159 if (migrate_scanner
) {
160 if (cc
->finished_update_migrate
)
162 if (pfn
> zone
->compact_cached_migrate_pfn
[0])
163 zone
->compact_cached_migrate_pfn
[0] = pfn
;
164 if (cc
->mode
!= MIGRATE_ASYNC
&&
165 pfn
> zone
->compact_cached_migrate_pfn
[1])
166 zone
->compact_cached_migrate_pfn
[1] = pfn
;
168 if (cc
->finished_update_free
)
170 if (pfn
< zone
->compact_cached_free_pfn
)
171 zone
->compact_cached_free_pfn
= pfn
;
175 static inline bool isolation_suitable(struct compact_control
*cc
,
181 static void update_pageblock_skip(struct compact_control
*cc
,
182 struct page
*page
, unsigned long nr_isolated
,
183 bool set_unsuitable
, bool migrate_scanner
)
186 #endif /* CONFIG_COMPACTION */
188 static inline bool should_release_lock(spinlock_t
*lock
)
190 return need_resched() || spin_is_contended(lock
);
194 * Compaction requires the taking of some coarse locks that are potentially
195 * very heavily contended. Check if the process needs to be scheduled or
196 * if the lock is contended. For async compaction, back out in the event
197 * if contention is severe. For sync compaction, schedule.
199 * Returns true if the lock is held.
200 * Returns false if the lock is released and compaction should abort
202 static bool compact_checklock_irqsave(spinlock_t
*lock
, unsigned long *flags
,
203 bool locked
, struct compact_control
*cc
)
205 if (should_release_lock(lock
)) {
207 spin_unlock_irqrestore(lock
, *flags
);
211 /* async aborts if taking too long or contended */
212 if (cc
->mode
== MIGRATE_ASYNC
) {
213 cc
->contended
= true;
221 spin_lock_irqsave(lock
, *flags
);
226 * Aside from avoiding lock contention, compaction also periodically checks
227 * need_resched() and either schedules in sync compaction or aborts async
228 * compaction. This is similar to what compact_checklock_irqsave() does, but
229 * is used where no lock is concerned.
231 * Returns false when no scheduling was needed, or sync compaction scheduled.
232 * Returns true when async compaction should abort.
234 static inline bool compact_should_abort(struct compact_control
*cc
)
236 /* async compaction aborts if contended */
237 if (need_resched()) {
238 if (cc
->mode
== MIGRATE_ASYNC
) {
239 cc
->contended
= true;
249 /* Returns true if the page is within a block suitable for migration to */
250 static bool suitable_migration_target(struct page
*page
)
252 /* If the page is a large free page, then disallow migration */
253 if (PageBuddy(page
) && page_order(page
) >= pageblock_order
)
256 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
257 if (migrate_async_suitable(get_pageblock_migratetype(page
)))
260 /* Otherwise skip the block */
265 * Isolate free pages onto a private freelist. If @strict is true, will abort
266 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
267 * (even though it may still end up isolating some pages).
269 static unsigned long isolate_freepages_block(struct compact_control
*cc
,
270 unsigned long blockpfn
,
271 unsigned long end_pfn
,
272 struct list_head
*freelist
,
275 int nr_scanned
= 0, total_isolated
= 0;
276 struct page
*cursor
, *valid_page
= NULL
;
280 cursor
= pfn_to_page(blockpfn
);
282 /* Isolate free pages. */
283 for (; blockpfn
< end_pfn
; blockpfn
++, cursor
++) {
285 struct page
*page
= cursor
;
288 if (!pfn_valid_within(blockpfn
))
293 if (!PageBuddy(page
))
297 * The zone lock must be held to isolate freepages.
298 * Unfortunately this is a very coarse lock and can be
299 * heavily contended if there are parallel allocations
300 * or parallel compactions. For async compaction do not
301 * spin on the lock and we acquire the lock as late as
304 locked
= compact_checklock_irqsave(&cc
->zone
->lock
, &flags
,
309 /* Recheck this is a buddy page under lock */
310 if (!PageBuddy(page
))
313 /* Found a free page, break it into order-0 pages */
314 isolated
= split_free_page(page
);
315 total_isolated
+= isolated
;
316 for (i
= 0; i
< isolated
; i
++) {
317 list_add(&page
->lru
, freelist
);
321 /* If a page was split, advance to the end of it */
323 blockpfn
+= isolated
- 1;
324 cursor
+= isolated
- 1;
336 trace_mm_compaction_isolate_freepages(nr_scanned
, total_isolated
);
339 * If strict isolation is requested by CMA then check that all the
340 * pages requested were isolated. If there were any failures, 0 is
341 * returned and CMA will fail.
343 if (strict
&& blockpfn
< end_pfn
)
347 spin_unlock_irqrestore(&cc
->zone
->lock
, flags
);
349 /* Update the pageblock-skip if the whole pageblock was scanned */
350 if (blockpfn
== end_pfn
)
351 update_pageblock_skip(cc
, valid_page
, total_isolated
, true,
354 count_compact_events(COMPACTFREE_SCANNED
, nr_scanned
);
356 count_compact_events(COMPACTISOLATED
, total_isolated
);
357 return total_isolated
;
361 * isolate_freepages_range() - isolate free pages.
362 * @start_pfn: The first PFN to start isolating.
363 * @end_pfn: The one-past-last PFN.
365 * Non-free pages, invalid PFNs, or zone boundaries within the
366 * [start_pfn, end_pfn) range are considered errors, cause function to
367 * undo its actions and return zero.
369 * Otherwise, function returns one-past-the-last PFN of isolated page
370 * (which may be greater then end_pfn if end fell in a middle of
374 isolate_freepages_range(struct compact_control
*cc
,
375 unsigned long start_pfn
, unsigned long end_pfn
)
377 unsigned long isolated
, pfn
, block_end_pfn
;
380 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= isolated
) {
381 if (!pfn_valid(pfn
) || cc
->zone
!= page_zone(pfn_to_page(pfn
)))
385 * On subsequent iterations ALIGN() is actually not needed,
386 * but we keep it that we not to complicate the code.
388 block_end_pfn
= ALIGN(pfn
+ 1, pageblock_nr_pages
);
389 block_end_pfn
= min(block_end_pfn
, end_pfn
);
391 isolated
= isolate_freepages_block(cc
, pfn
, block_end_pfn
,
395 * In strict mode, isolate_freepages_block() returns 0 if
396 * there are any holes in the block (ie. invalid PFNs or
403 * If we managed to isolate pages, it is always (1 << n) *
404 * pageblock_nr_pages for some non-negative n. (Max order
405 * page may span two pageblocks).
409 /* split_free_page does not map the pages */
410 map_pages(&freelist
);
413 /* Loop terminated early, cleanup. */
414 release_freepages(&freelist
);
418 /* We don't use freelists for anything. */
422 /* Update the number of anon and file isolated pages in the zone */
423 static void acct_isolated(struct zone
*zone
, bool locked
, struct compact_control
*cc
)
426 unsigned int count
[2] = { 0, };
428 list_for_each_entry(page
, &cc
->migratepages
, lru
)
429 count
[!!page_is_file_cache(page
)]++;
431 /* If locked we can use the interrupt unsafe versions */
433 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
, count
[0]);
434 __mod_zone_page_state(zone
, NR_ISOLATED_FILE
, count
[1]);
436 mod_zone_page_state(zone
, NR_ISOLATED_ANON
, count
[0]);
437 mod_zone_page_state(zone
, NR_ISOLATED_FILE
, count
[1]);
441 /* Similar to reclaim, but different enough that they don't share logic */
442 static bool too_many_isolated(struct zone
*zone
)
444 unsigned long active
, inactive
, isolated
;
446 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
) +
447 zone_page_state(zone
, NR_INACTIVE_ANON
);
448 active
= zone_page_state(zone
, NR_ACTIVE_FILE
) +
449 zone_page_state(zone
, NR_ACTIVE_ANON
);
450 isolated
= zone_page_state(zone
, NR_ISOLATED_FILE
) +
451 zone_page_state(zone
, NR_ISOLATED_ANON
);
453 return isolated
> (inactive
+ active
) / 2;
457 * isolate_migratepages_range() - isolate all migrate-able pages in range.
458 * @zone: Zone pages are in.
459 * @cc: Compaction control structure.
460 * @low_pfn: The first PFN of the range.
461 * @end_pfn: The one-past-the-last PFN of the range.
462 * @unevictable: true if it allows to isolate unevictable pages
464 * Isolate all pages that can be migrated from the range specified by
465 * [low_pfn, end_pfn). Returns zero if there is a fatal signal
466 * pending), otherwise PFN of the first page that was not scanned
467 * (which may be both less, equal to or more then end_pfn).
469 * Assumes that cc->migratepages is empty and cc->nr_migratepages is
472 * Apart from cc->migratepages and cc->nr_migratetypes this function
473 * does not modify any cc's fields, in particular it does not modify
474 * (or read for that matter) cc->migrate_pfn.
477 isolate_migratepages_range(struct zone
*zone
, struct compact_control
*cc
,
478 unsigned long low_pfn
, unsigned long end_pfn
, bool unevictable
)
480 unsigned long last_pageblock_nr
= 0, pageblock_nr
;
481 unsigned long nr_scanned
= 0, nr_isolated
= 0;
482 struct list_head
*migratelist
= &cc
->migratepages
;
483 struct lruvec
*lruvec
;
486 struct page
*page
= NULL
, *valid_page
= NULL
;
487 bool set_unsuitable
= true;
488 const isolate_mode_t mode
= (cc
->mode
== MIGRATE_ASYNC
?
489 ISOLATE_ASYNC_MIGRATE
: 0) |
490 (unevictable
? ISOLATE_UNEVICTABLE
: 0);
493 * Ensure that there are not too many pages isolated from the LRU
494 * list by either parallel reclaimers or compaction. If there are,
495 * delay for some time until fewer pages are isolated
497 while (unlikely(too_many_isolated(zone
))) {
498 /* async migration should just abort */
499 if (cc
->mode
== MIGRATE_ASYNC
)
502 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
504 if (fatal_signal_pending(current
))
508 if (compact_should_abort(cc
))
511 /* Time to isolate some pages for migration */
512 for (; low_pfn
< end_pfn
; low_pfn
++) {
513 /* give a chance to irqs before checking need_resched() */
514 if (locked
&& !(low_pfn
% SWAP_CLUSTER_MAX
)) {
515 if (should_release_lock(&zone
->lru_lock
)) {
516 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
522 * migrate_pfn does not necessarily start aligned to a
523 * pageblock. Ensure that pfn_valid is called when moving
524 * into a new MAX_ORDER_NR_PAGES range in case of large
525 * memory holes within the zone
527 if ((low_pfn
& (MAX_ORDER_NR_PAGES
- 1)) == 0) {
528 if (!pfn_valid(low_pfn
)) {
529 low_pfn
+= MAX_ORDER_NR_PAGES
- 1;
534 if (!pfn_valid_within(low_pfn
))
539 * Get the page and ensure the page is within the same zone.
540 * See the comment in isolate_freepages about overlapping
541 * nodes. It is deliberate that the new zone lock is not taken
542 * as memory compaction should not move pages between nodes.
544 page
= pfn_to_page(low_pfn
);
545 if (page_zone(page
) != zone
)
551 /* If isolation recently failed, do not retry */
552 pageblock_nr
= low_pfn
>> pageblock_order
;
553 if (last_pageblock_nr
!= pageblock_nr
) {
556 last_pageblock_nr
= pageblock_nr
;
557 if (!isolation_suitable(cc
, page
))
561 * For async migration, also only scan in MOVABLE
562 * blocks. Async migration is optimistic to see if
563 * the minimum amount of work satisfies the allocation
565 mt
= get_pageblock_migratetype(page
);
566 if (cc
->mode
== MIGRATE_ASYNC
&&
567 !migrate_async_suitable(mt
)) {
568 set_unsuitable
= false;
574 * Skip if free. page_order cannot be used without zone->lock
575 * as nothing prevents parallel allocations or buddy merging.
581 * Check may be lockless but that's ok as we recheck later.
582 * It's possible to migrate LRU pages and balloon pages
583 * Skip any other type of page
585 if (!PageLRU(page
)) {
586 if (unlikely(balloon_page_movable(page
))) {
587 if (locked
&& balloon_page_isolate(page
)) {
588 /* Successfully isolated */
589 goto isolate_success
;
596 * PageLRU is set. lru_lock normally excludes isolation
597 * splitting and collapsing (collapsing has already happened
598 * if PageLRU is set) but the lock is not necessarily taken
599 * here and it is wasteful to take it just to check transhuge.
600 * Check TransHuge without lock and skip the whole pageblock if
601 * it's either a transhuge or hugetlbfs page, as calling
602 * compound_order() without preventing THP from splitting the
603 * page underneath us may return surprising results.
605 if (PageTransHuge(page
)) {
608 low_pfn
+= (1 << compound_order(page
)) - 1;
613 * Migration will fail if an anonymous page is pinned in memory,
614 * so avoid taking lru_lock and isolating it unnecessarily in an
615 * admittedly racy check.
617 if (!page_mapping(page
) &&
618 page_count(page
) > page_mapcount(page
))
621 /* Check if it is ok to still hold the lock */
622 locked
= compact_checklock_irqsave(&zone
->lru_lock
, &flags
,
624 if (!locked
|| fatal_signal_pending(current
))
627 /* Recheck PageLRU and PageTransHuge under lock */
630 if (PageTransHuge(page
)) {
631 low_pfn
+= (1 << compound_order(page
)) - 1;
635 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
637 /* Try isolate the page */
638 if (__isolate_lru_page(page
, mode
) != 0)
641 VM_BUG_ON_PAGE(PageTransCompound(page
), page
);
643 /* Successfully isolated */
644 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
647 cc
->finished_update_migrate
= true;
648 list_add(&page
->lru
, migratelist
);
649 cc
->nr_migratepages
++;
652 /* Avoid isolating too much */
653 if (cc
->nr_migratepages
== COMPACT_CLUSTER_MAX
) {
661 low_pfn
= ALIGN(low_pfn
+ 1, pageblock_nr_pages
) - 1;
664 acct_isolated(zone
, locked
, cc
);
667 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
670 * Update the pageblock-skip information and cached scanner pfn,
671 * if the whole pageblock was scanned without isolating any page.
673 if (low_pfn
== end_pfn
)
674 update_pageblock_skip(cc
, valid_page
, nr_isolated
,
675 set_unsuitable
, true);
677 trace_mm_compaction_isolate_migratepages(nr_scanned
, nr_isolated
);
679 count_compact_events(COMPACTMIGRATE_SCANNED
, nr_scanned
);
681 count_compact_events(COMPACTISOLATED
, nr_isolated
);
686 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
687 #ifdef CONFIG_COMPACTION
689 * Based on information in the current compact_control, find blocks
690 * suitable for isolating free pages from and then isolate them.
692 static void isolate_freepages(struct zone
*zone
,
693 struct compact_control
*cc
)
696 unsigned long block_start_pfn
; /* start of current pageblock */
697 unsigned long block_end_pfn
; /* end of current pageblock */
698 unsigned long low_pfn
; /* lowest pfn scanner is able to scan */
699 int nr_freepages
= cc
->nr_freepages
;
700 struct list_head
*freelist
= &cc
->freepages
;
703 * Initialise the free scanner. The starting point is where we last
704 * successfully isolated from, zone-cached value, or the end of the
705 * zone when isolating for the first time. We need this aligned to
706 * the pageblock boundary, because we do
707 * block_start_pfn -= pageblock_nr_pages in the for loop.
708 * For ending point, take care when isolating in last pageblock of a
709 * a zone which ends in the middle of a pageblock.
710 * The low boundary is the end of the pageblock the migration scanner
713 block_start_pfn
= cc
->free_pfn
& ~(pageblock_nr_pages
-1);
714 block_end_pfn
= min(block_start_pfn
+ pageblock_nr_pages
,
716 low_pfn
= ALIGN(cc
->migrate_pfn
+ 1, pageblock_nr_pages
);
719 * Isolate free pages until enough are available to migrate the
720 * pages on cc->migratepages. We stop searching if the migrate
721 * and free page scanners meet or enough free pages are isolated.
723 for (; block_start_pfn
>= low_pfn
&& cc
->nr_migratepages
> nr_freepages
;
724 block_end_pfn
= block_start_pfn
,
725 block_start_pfn
-= pageblock_nr_pages
) {
726 unsigned long isolated
;
729 * This can iterate a massively long zone without finding any
730 * suitable migration targets, so periodically check if we need
731 * to schedule, or even abort async compaction.
733 if (!(block_start_pfn
% (SWAP_CLUSTER_MAX
* pageblock_nr_pages
))
734 && compact_should_abort(cc
))
737 if (!pfn_valid(block_start_pfn
))
741 * Check for overlapping nodes/zones. It's possible on some
742 * configurations to have a setup like
744 * i.e. it's possible that all pages within a zones range of
745 * pages do not belong to a single zone.
747 page
= pfn_to_page(block_start_pfn
);
748 if (page_zone(page
) != zone
)
751 /* Check the block is suitable for migration */
752 if (!suitable_migration_target(page
))
755 /* If isolation recently failed, do not retry */
756 if (!isolation_suitable(cc
, page
))
759 /* Found a block suitable for isolating free pages from */
760 cc
->free_pfn
= block_start_pfn
;
761 isolated
= isolate_freepages_block(cc
, block_start_pfn
,
762 block_end_pfn
, freelist
, false);
763 nr_freepages
+= isolated
;
766 * Set a flag that we successfully isolated in this pageblock.
767 * In the next loop iteration, zone->compact_cached_free_pfn
768 * will not be updated and thus it will effectively contain the
769 * highest pageblock we isolated pages from.
772 cc
->finished_update_free
= true;
775 * isolate_freepages_block() might have aborted due to async
776 * compaction being contended
782 /* split_free_page does not map the pages */
786 * If we crossed the migrate scanner, we want to keep it that way
787 * so that compact_finished() may detect this
789 if (block_start_pfn
< low_pfn
)
790 cc
->free_pfn
= cc
->migrate_pfn
;
792 cc
->nr_freepages
= nr_freepages
;
796 * This is a migrate-callback that "allocates" freepages by taking pages
797 * from the isolated freelists in the block we are migrating to.
799 static struct page
*compaction_alloc(struct page
*migratepage
,
803 struct compact_control
*cc
= (struct compact_control
*)data
;
804 struct page
*freepage
;
807 * Isolate free pages if necessary, and if we are not aborting due to
810 if (list_empty(&cc
->freepages
)) {
812 isolate_freepages(cc
->zone
, cc
);
814 if (list_empty(&cc
->freepages
))
818 freepage
= list_entry(cc
->freepages
.next
, struct page
, lru
);
819 list_del(&freepage
->lru
);
826 * This is a migrate-callback that "frees" freepages back to the isolated
827 * freelist. All pages on the freelist are from the same zone, so there is no
828 * special handling needed for NUMA.
830 static void compaction_free(struct page
*page
, unsigned long data
)
832 struct compact_control
*cc
= (struct compact_control
*)data
;
834 list_add(&page
->lru
, &cc
->freepages
);
838 /* possible outcome of isolate_migratepages */
840 ISOLATE_ABORT
, /* Abort compaction now */
841 ISOLATE_NONE
, /* No pages isolated, continue scanning */
842 ISOLATE_SUCCESS
, /* Pages isolated, migrate */
846 * Isolate all pages that can be migrated from the block pointed to by
847 * the migrate scanner within compact_control.
849 static isolate_migrate_t
isolate_migratepages(struct zone
*zone
,
850 struct compact_control
*cc
)
852 unsigned long low_pfn
, end_pfn
;
854 /* Do not scan outside zone boundaries */
855 low_pfn
= max(cc
->migrate_pfn
, zone
->zone_start_pfn
);
857 /* Only scan within a pageblock boundary */
858 end_pfn
= ALIGN(low_pfn
+ 1, pageblock_nr_pages
);
860 /* Do not cross the free scanner or scan within a memory hole */
861 if (end_pfn
> cc
->free_pfn
|| !pfn_valid(low_pfn
)) {
862 cc
->migrate_pfn
= end_pfn
;
866 /* Perform the isolation */
867 low_pfn
= isolate_migratepages_range(zone
, cc
, low_pfn
, end_pfn
, false);
868 if (!low_pfn
|| cc
->contended
)
869 return ISOLATE_ABORT
;
871 cc
->migrate_pfn
= low_pfn
;
873 return ISOLATE_SUCCESS
;
876 static int compact_finished(struct zone
*zone
,
877 struct compact_control
*cc
)
880 unsigned long watermark
;
882 if (cc
->contended
|| fatal_signal_pending(current
))
883 return COMPACT_PARTIAL
;
885 /* Compaction run completes if the migrate and free scanner meet */
886 if (cc
->free_pfn
<= cc
->migrate_pfn
) {
887 /* Let the next compaction start anew. */
888 zone
->compact_cached_migrate_pfn
[0] = zone
->zone_start_pfn
;
889 zone
->compact_cached_migrate_pfn
[1] = zone
->zone_start_pfn
;
890 zone
->compact_cached_free_pfn
= zone_end_pfn(zone
);
893 * Mark that the PG_migrate_skip information should be cleared
894 * by kswapd when it goes to sleep. kswapd does not set the
895 * flag itself as the decision to be clear should be directly
896 * based on an allocation request.
898 if (!current_is_kswapd())
899 zone
->compact_blockskip_flush
= true;
901 return COMPACT_COMPLETE
;
905 * order == -1 is expected when compacting via
906 * /proc/sys/vm/compact_memory
909 return COMPACT_CONTINUE
;
911 /* Compaction run is not finished if the watermark is not met */
912 watermark
= low_wmark_pages(zone
);
913 watermark
+= (1 << cc
->order
);
915 if (!zone_watermark_ok(zone
, cc
->order
, watermark
, 0, 0))
916 return COMPACT_CONTINUE
;
918 /* Direct compactor: Is a suitable page free? */
919 for (order
= cc
->order
; order
< MAX_ORDER
; order
++) {
920 struct free_area
*area
= &zone
->free_area
[order
];
922 /* Job done if page is free of the right migratetype */
923 if (!list_empty(&area
->free_list
[cc
->migratetype
]))
924 return COMPACT_PARTIAL
;
926 /* Job done if allocation would set block type */
927 if (cc
->order
>= pageblock_order
&& area
->nr_free
)
928 return COMPACT_PARTIAL
;
931 return COMPACT_CONTINUE
;
935 * compaction_suitable: Is this suitable to run compaction on this zone now?
937 * COMPACT_SKIPPED - If there are too few free pages for compaction
938 * COMPACT_PARTIAL - If the allocation would succeed without compaction
939 * COMPACT_CONTINUE - If compaction should run now
941 unsigned long compaction_suitable(struct zone
*zone
, int order
)
944 unsigned long watermark
;
947 * order == -1 is expected when compacting via
948 * /proc/sys/vm/compact_memory
951 return COMPACT_CONTINUE
;
954 * Watermarks for order-0 must be met for compaction. Note the 2UL.
955 * This is because during migration, copies of pages need to be
956 * allocated and for a short time, the footprint is higher
958 watermark
= low_wmark_pages(zone
) + (2UL << order
);
959 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
960 return COMPACT_SKIPPED
;
963 * fragmentation index determines if allocation failures are due to
964 * low memory or external fragmentation
966 * index of -1000 implies allocations might succeed depending on
968 * index towards 0 implies failure is due to lack of memory
969 * index towards 1000 implies failure is due to fragmentation
971 * Only compact if a failure would be due to fragmentation.
973 fragindex
= fragmentation_index(zone
, order
);
974 if (fragindex
>= 0 && fragindex
<= sysctl_extfrag_threshold
)
975 return COMPACT_SKIPPED
;
977 if (fragindex
== -1000 && zone_watermark_ok(zone
, order
, watermark
,
979 return COMPACT_PARTIAL
;
981 return COMPACT_CONTINUE
;
984 static int compact_zone(struct zone
*zone
, struct compact_control
*cc
)
987 unsigned long start_pfn
= zone
->zone_start_pfn
;
988 unsigned long end_pfn
= zone_end_pfn(zone
);
989 const bool sync
= cc
->mode
!= MIGRATE_ASYNC
;
991 ret
= compaction_suitable(zone
, cc
->order
);
993 case COMPACT_PARTIAL
:
994 case COMPACT_SKIPPED
:
995 /* Compaction is likely to fail */
997 case COMPACT_CONTINUE
:
998 /* Fall through to compaction */
1003 * Clear pageblock skip if there were failures recently and compaction
1004 * is about to be retried after being deferred. kswapd does not do
1005 * this reset as it'll reset the cached information when going to sleep.
1007 if (compaction_restarting(zone
, cc
->order
) && !current_is_kswapd())
1008 __reset_isolation_suitable(zone
);
1011 * Setup to move all movable pages to the end of the zone. Used cached
1012 * information on where the scanners should start but check that it
1013 * is initialised by ensuring the values are within zone boundaries.
1015 cc
->migrate_pfn
= zone
->compact_cached_migrate_pfn
[sync
];
1016 cc
->free_pfn
= zone
->compact_cached_free_pfn
;
1017 if (cc
->free_pfn
< start_pfn
|| cc
->free_pfn
> end_pfn
) {
1018 cc
->free_pfn
= end_pfn
& ~(pageblock_nr_pages
-1);
1019 zone
->compact_cached_free_pfn
= cc
->free_pfn
;
1021 if (cc
->migrate_pfn
< start_pfn
|| cc
->migrate_pfn
> end_pfn
) {
1022 cc
->migrate_pfn
= start_pfn
;
1023 zone
->compact_cached_migrate_pfn
[0] = cc
->migrate_pfn
;
1024 zone
->compact_cached_migrate_pfn
[1] = cc
->migrate_pfn
;
1027 trace_mm_compaction_begin(start_pfn
, cc
->migrate_pfn
, cc
->free_pfn
, end_pfn
);
1029 migrate_prep_local();
1031 while ((ret
= compact_finished(zone
, cc
)) == COMPACT_CONTINUE
) {
1034 switch (isolate_migratepages(zone
, cc
)) {
1036 ret
= COMPACT_PARTIAL
;
1037 putback_movable_pages(&cc
->migratepages
);
1038 cc
->nr_migratepages
= 0;
1042 case ISOLATE_SUCCESS
:
1046 if (!cc
->nr_migratepages
)
1049 err
= migrate_pages(&cc
->migratepages
, compaction_alloc
,
1050 compaction_free
, (unsigned long)cc
, cc
->mode
,
1053 trace_mm_compaction_migratepages(cc
->nr_migratepages
, err
,
1056 /* All pages were either migrated or will be released */
1057 cc
->nr_migratepages
= 0;
1059 putback_movable_pages(&cc
->migratepages
);
1061 * migrate_pages() may return -ENOMEM when scanners meet
1062 * and we want compact_finished() to detect it
1064 if (err
== -ENOMEM
&& cc
->free_pfn
> cc
->migrate_pfn
) {
1065 ret
= COMPACT_PARTIAL
;
1072 /* Release free pages and check accounting */
1073 cc
->nr_freepages
-= release_freepages(&cc
->freepages
);
1074 VM_BUG_ON(cc
->nr_freepages
!= 0);
1076 trace_mm_compaction_end(ret
);
1081 static unsigned long compact_zone_order(struct zone
*zone
, int order
,
1082 gfp_t gfp_mask
, enum migrate_mode mode
, bool *contended
)
1085 struct compact_control cc
= {
1087 .nr_migratepages
= 0,
1089 .migratetype
= allocflags_to_migratetype(gfp_mask
),
1093 INIT_LIST_HEAD(&cc
.freepages
);
1094 INIT_LIST_HEAD(&cc
.migratepages
);
1096 ret
= compact_zone(zone
, &cc
);
1098 VM_BUG_ON(!list_empty(&cc
.freepages
));
1099 VM_BUG_ON(!list_empty(&cc
.migratepages
));
1101 *contended
= cc
.contended
;
1105 int sysctl_extfrag_threshold
= 500;
1108 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1109 * @zonelist: The zonelist used for the current allocation
1110 * @order: The order of the current allocation
1111 * @gfp_mask: The GFP mask of the current allocation
1112 * @nodemask: The allowed nodes to allocate from
1113 * @mode: The migration mode for async, sync light, or sync migration
1114 * @contended: Return value that is true if compaction was aborted due to lock contention
1115 * @candidate_zone: Return the zone where we think allocation should succeed
1117 * This is the main entry point for direct page compaction.
1119 unsigned long try_to_compact_pages(struct zonelist
*zonelist
,
1120 int order
, gfp_t gfp_mask
, nodemask_t
*nodemask
,
1121 enum migrate_mode mode
, bool *contended
,
1122 struct zone
**candidate_zone
)
1124 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
1125 int may_enter_fs
= gfp_mask
& __GFP_FS
;
1126 int may_perform_io
= gfp_mask
& __GFP_IO
;
1129 int rc
= COMPACT_DEFERRED
;
1130 int alloc_flags
= 0;
1132 /* Check if the GFP flags allow compaction */
1133 if (!order
|| !may_enter_fs
|| !may_perform_io
)
1134 return COMPACT_SKIPPED
;
1137 if (allocflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
1138 alloc_flags
|= ALLOC_CMA
;
1140 /* Compact each zone in the list */
1141 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, high_zoneidx
,
1145 if (compaction_deferred(zone
, order
))
1148 status
= compact_zone_order(zone
, order
, gfp_mask
, mode
,
1150 rc
= max(status
, rc
);
1152 /* If a normal allocation would succeed, stop compacting */
1153 if (zone_watermark_ok(zone
, order
, low_wmark_pages(zone
), 0,
1155 *candidate_zone
= zone
;
1157 * We think the allocation will succeed in this zone,
1158 * but it is not certain, hence the false. The caller
1159 * will repeat this with true if allocation indeed
1160 * succeeds in this zone.
1162 compaction_defer_reset(zone
, order
, false);
1164 } else if (mode
!= MIGRATE_ASYNC
) {
1166 * We think that allocation won't succeed in this zone
1167 * so we defer compaction there. If it ends up
1168 * succeeding after all, it will be reset.
1170 defer_compaction(zone
, order
);
1178 /* Compact all zones within a node */
1179 static void __compact_pgdat(pg_data_t
*pgdat
, struct compact_control
*cc
)
1184 for (zoneid
= 0; zoneid
< MAX_NR_ZONES
; zoneid
++) {
1186 zone
= &pgdat
->node_zones
[zoneid
];
1187 if (!populated_zone(zone
))
1190 cc
->nr_freepages
= 0;
1191 cc
->nr_migratepages
= 0;
1193 INIT_LIST_HEAD(&cc
->freepages
);
1194 INIT_LIST_HEAD(&cc
->migratepages
);
1196 if (cc
->order
== -1 || !compaction_deferred(zone
, cc
->order
))
1197 compact_zone(zone
, cc
);
1199 if (cc
->order
> 0) {
1200 if (zone_watermark_ok(zone
, cc
->order
,
1201 low_wmark_pages(zone
), 0, 0))
1202 compaction_defer_reset(zone
, cc
->order
, false);
1205 VM_BUG_ON(!list_empty(&cc
->freepages
));
1206 VM_BUG_ON(!list_empty(&cc
->migratepages
));
1210 void compact_pgdat(pg_data_t
*pgdat
, int order
)
1212 struct compact_control cc
= {
1214 .mode
= MIGRATE_ASYNC
,
1220 __compact_pgdat(pgdat
, &cc
);
1223 static void compact_node(int nid
)
1225 struct compact_control cc
= {
1227 .mode
= MIGRATE_SYNC
,
1228 .ignore_skip_hint
= true,
1231 __compact_pgdat(NODE_DATA(nid
), &cc
);
1234 /* Compact all nodes in the system */
1235 static void compact_nodes(void)
1239 /* Flush pending updates to the LRU lists */
1240 lru_add_drain_all();
1242 for_each_online_node(nid
)
1246 /* The written value is actually unused, all memory is compacted */
1247 int sysctl_compact_memory
;
1249 /* This is the entry point for compacting all nodes via /proc/sys/vm */
1250 int sysctl_compaction_handler(struct ctl_table
*table
, int write
,
1251 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
1259 int sysctl_extfrag_handler(struct ctl_table
*table
, int write
,
1260 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
1262 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
1267 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1268 static ssize_t
sysfs_compact_node(struct device
*dev
,
1269 struct device_attribute
*attr
,
1270 const char *buf
, size_t count
)
1274 if (nid
>= 0 && nid
< nr_node_ids
&& node_online(nid
)) {
1275 /* Flush pending updates to the LRU lists */
1276 lru_add_drain_all();
1283 static DEVICE_ATTR(compact
, S_IWUSR
, NULL
, sysfs_compact_node
);
1285 int compaction_register_node(struct node
*node
)
1287 return device_create_file(&node
->dev
, &dev_attr_compact
);
1290 void compaction_unregister_node(struct node
*node
)
1292 return device_remove_file(&node
->dev
, &dev_attr_compact
);
1294 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1296 #endif /* CONFIG_COMPACTION */