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>
19 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
21 #define CREATE_TRACE_POINTS
22 #include <trace/events/compaction.h>
24 static unsigned long release_freepages(struct list_head
*freelist
)
26 struct page
*page
, *next
;
27 unsigned long count
= 0;
29 list_for_each_entry_safe(page
, next
, freelist
, lru
) {
38 static void map_pages(struct list_head
*list
)
42 list_for_each_entry(page
, list
, lru
) {
43 arch_alloc_page(page
, 0);
44 kernel_map_pages(page
, 1, 1);
48 static inline bool migrate_async_suitable(int migratetype
)
50 return is_migrate_cma(migratetype
) || migratetype
== MIGRATE_MOVABLE
;
53 #ifdef CONFIG_COMPACTION
54 /* Returns true if the pageblock should be scanned for pages to isolate. */
55 static inline bool isolation_suitable(struct compact_control
*cc
,
58 if (cc
->ignore_skip_hint
)
61 return !get_pageblock_skip(page
);
65 * This function is called to clear all cached information on pageblocks that
66 * should be skipped for page isolation when the migrate and free page scanner
69 static void __reset_isolation_suitable(struct zone
*zone
)
71 unsigned long start_pfn
= zone
->zone_start_pfn
;
72 unsigned long end_pfn
= zone
->zone_start_pfn
+ zone
->spanned_pages
;
75 zone
->compact_cached_migrate_pfn
= start_pfn
;
76 zone
->compact_cached_free_pfn
= end_pfn
;
77 zone
->compact_blockskip_flush
= false;
79 /* Walk the zone and mark every pageblock as suitable for isolation */
80 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
88 page
= pfn_to_page(pfn
);
89 if (zone
!= page_zone(page
))
92 clear_pageblock_skip(page
);
96 void reset_isolation_suitable(pg_data_t
*pgdat
)
100 for (zoneid
= 0; zoneid
< MAX_NR_ZONES
; zoneid
++) {
101 struct zone
*zone
= &pgdat
->node_zones
[zoneid
];
102 if (!populated_zone(zone
))
105 /* Only flush if a full compaction finished recently */
106 if (zone
->compact_blockskip_flush
)
107 __reset_isolation_suitable(zone
);
112 * If no pages were isolated then mark this pageblock to be skipped in the
113 * future. The information is later cleared by __reset_isolation_suitable().
115 static void update_pageblock_skip(struct compact_control
*cc
,
116 struct page
*page
, unsigned long nr_isolated
,
117 bool migrate_scanner
)
119 struct zone
*zone
= cc
->zone
;
124 unsigned long pfn
= page_to_pfn(page
);
125 set_pageblock_skip(page
);
127 /* Update where compaction should restart */
128 if (migrate_scanner
) {
129 if (!cc
->finished_update_migrate
&&
130 pfn
> zone
->compact_cached_migrate_pfn
)
131 zone
->compact_cached_migrate_pfn
= pfn
;
133 if (!cc
->finished_update_free
&&
134 pfn
< zone
->compact_cached_free_pfn
)
135 zone
->compact_cached_free_pfn
= pfn
;
140 static inline bool isolation_suitable(struct compact_control
*cc
,
146 static void update_pageblock_skip(struct compact_control
*cc
,
147 struct page
*page
, unsigned long nr_isolated
,
148 bool migrate_scanner
)
151 #endif /* CONFIG_COMPACTION */
153 static inline bool should_release_lock(spinlock_t
*lock
)
155 return need_resched() || spin_is_contended(lock
);
159 * Compaction requires the taking of some coarse locks that are potentially
160 * very heavily contended. Check if the process needs to be scheduled or
161 * if the lock is contended. For async compaction, back out in the event
162 * if contention is severe. For sync compaction, schedule.
164 * Returns true if the lock is held.
165 * Returns false if the lock is released and compaction should abort
167 static bool compact_checklock_irqsave(spinlock_t
*lock
, unsigned long *flags
,
168 bool locked
, struct compact_control
*cc
)
170 if (should_release_lock(lock
)) {
172 spin_unlock_irqrestore(lock
, *flags
);
176 /* async aborts if taking too long or contended */
178 cc
->contended
= true;
186 spin_lock_irqsave(lock
, *flags
);
190 static inline bool compact_trylock_irqsave(spinlock_t
*lock
,
191 unsigned long *flags
, struct compact_control
*cc
)
193 return compact_checklock_irqsave(lock
, flags
, false, cc
);
196 /* Returns true if the page is within a block suitable for migration to */
197 static bool suitable_migration_target(struct page
*page
)
199 int migratetype
= get_pageblock_migratetype(page
);
201 /* Don't interfere with memory hot-remove or the min_free_kbytes blocks */
202 if (migratetype
== MIGRATE_ISOLATE
|| migratetype
== MIGRATE_RESERVE
)
205 /* If the page is a large free page, then allow migration */
206 if (PageBuddy(page
) && page_order(page
) >= pageblock_order
)
209 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
210 if (migrate_async_suitable(migratetype
))
213 /* Otherwise skip the block */
217 static void compact_capture_page(struct compact_control
*cc
)
220 int mtype
, mtype_low
, mtype_high
;
222 if (!cc
->page
|| *cc
->page
)
226 * For MIGRATE_MOVABLE allocations we capture a suitable page ASAP
227 * regardless of the migratetype of the freelist is is captured from.
228 * This is fine because the order for a high-order MIGRATE_MOVABLE
229 * allocation is typically at least a pageblock size and overall
230 * fragmentation is not impaired. Other allocation types must
231 * capture pages from their own migratelist because otherwise they
232 * could pollute other pageblocks like MIGRATE_MOVABLE with
233 * difficult to move pages and making fragmentation worse overall.
235 if (cc
->migratetype
== MIGRATE_MOVABLE
) {
237 mtype_high
= MIGRATE_PCPTYPES
;
239 mtype_low
= cc
->migratetype
;
240 mtype_high
= cc
->migratetype
+ 1;
243 /* Speculatively examine the free lists without zone lock */
244 for (mtype
= mtype_low
; mtype
< mtype_high
; mtype
++) {
246 for (order
= cc
->order
; order
< MAX_ORDER
; order
++) {
248 struct free_area
*area
;
249 area
= &(cc
->zone
->free_area
[order
]);
250 if (list_empty(&area
->free_list
[mtype
]))
253 /* Take the lock and attempt capture of the page */
254 if (!compact_trylock_irqsave(&cc
->zone
->lock
, &flags
, cc
))
256 if (!list_empty(&area
->free_list
[mtype
])) {
257 page
= list_entry(area
->free_list
[mtype
].next
,
259 if (capture_free_page(page
, cc
->order
, mtype
)) {
260 spin_unlock_irqrestore(&cc
->zone
->lock
,
266 spin_unlock_irqrestore(&cc
->zone
->lock
, flags
);
272 * Isolate free pages onto a private freelist. Caller must hold zone->lock.
273 * If @strict is true, will abort returning 0 on any invalid PFNs or non-free
274 * pages inside of the pageblock (even though it may still end up isolating
277 static unsigned long isolate_freepages_block(struct compact_control
*cc
,
278 unsigned long blockpfn
,
279 unsigned long end_pfn
,
280 struct list_head
*freelist
,
283 int nr_scanned
= 0, total_isolated
= 0;
284 struct page
*cursor
, *valid_page
= NULL
;
285 unsigned long nr_strict_required
= end_pfn
- blockpfn
;
289 cursor
= pfn_to_page(blockpfn
);
291 /* Isolate free pages. */
292 for (; blockpfn
< end_pfn
; blockpfn
++, cursor
++) {
294 struct page
*page
= cursor
;
297 if (!pfn_valid_within(blockpfn
))
301 if (!PageBuddy(page
))
305 * The zone lock must be held to isolate freepages.
306 * Unfortunately this is a very coarse lock and can be
307 * heavily contended if there are parallel allocations
308 * or parallel compactions. For async compaction do not
309 * spin on the lock and we acquire the lock as late as
312 locked
= compact_checklock_irqsave(&cc
->zone
->lock
, &flags
,
317 /* Recheck this is a suitable migration target under lock */
318 if (!strict
&& !suitable_migration_target(page
))
321 /* Recheck this is a buddy page under lock */
322 if (!PageBuddy(page
))
325 /* Found a free page, break it into order-0 pages */
326 isolated
= split_free_page(page
);
327 if (!isolated
&& strict
)
329 total_isolated
+= isolated
;
330 for (i
= 0; i
< isolated
; i
++) {
331 list_add(&page
->lru
, freelist
);
335 /* If a page was split, advance to the end of it */
337 blockpfn
+= isolated
- 1;
338 cursor
+= isolated
- 1;
342 trace_mm_compaction_isolate_freepages(nr_scanned
, total_isolated
);
345 * If strict isolation is requested by CMA then check that all the
346 * pages requested were isolated. If there were any failures, 0 is
347 * returned and CMA will fail.
349 if (strict
&& nr_strict_required
!= total_isolated
)
353 spin_unlock_irqrestore(&cc
->zone
->lock
, flags
);
355 /* Update the pageblock-skip if the whole pageblock was scanned */
356 if (blockpfn
== end_pfn
)
357 update_pageblock_skip(cc
, valid_page
, total_isolated
, false);
359 return total_isolated
;
363 * isolate_freepages_range() - isolate free pages.
364 * @start_pfn: The first PFN to start isolating.
365 * @end_pfn: The one-past-last PFN.
367 * Non-free pages, invalid PFNs, or zone boundaries within the
368 * [start_pfn, end_pfn) range are considered errors, cause function to
369 * undo its actions and return zero.
371 * Otherwise, function returns one-past-the-last PFN of isolated page
372 * (which may be greater then end_pfn if end fell in a middle of
376 isolate_freepages_range(struct compact_control
*cc
,
377 unsigned long start_pfn
, unsigned long end_pfn
)
379 unsigned long isolated
, pfn
, block_end_pfn
;
382 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= isolated
) {
383 if (!pfn_valid(pfn
) || cc
->zone
!= page_zone(pfn_to_page(pfn
)))
387 * On subsequent iterations ALIGN() is actually not needed,
388 * but we keep it that we not to complicate the code.
390 block_end_pfn
= ALIGN(pfn
+ 1, pageblock_nr_pages
);
391 block_end_pfn
= min(block_end_pfn
, end_pfn
);
393 isolated
= isolate_freepages_block(cc
, pfn
, block_end_pfn
,
397 * In strict mode, isolate_freepages_block() returns 0 if
398 * there are any holes in the block (ie. invalid PFNs or
405 * If we managed to isolate pages, it is always (1 << n) *
406 * pageblock_nr_pages for some non-negative n. (Max order
407 * page may span two pageblocks).
411 /* split_free_page does not map the pages */
412 map_pages(&freelist
);
415 /* Loop terminated early, cleanup. */
416 release_freepages(&freelist
);
420 /* We don't use freelists for anything. */
424 /* Update the number of anon and file isolated pages in the zone */
425 static void acct_isolated(struct zone
*zone
, bool locked
, struct compact_control
*cc
)
428 unsigned int count
[2] = { 0, };
430 list_for_each_entry(page
, &cc
->migratepages
, lru
)
431 count
[!!page_is_file_cache(page
)]++;
433 /* If locked we can use the interrupt unsafe versions */
435 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
, count
[0]);
436 __mod_zone_page_state(zone
, NR_ISOLATED_FILE
, count
[1]);
438 mod_zone_page_state(zone
, NR_ISOLATED_ANON
, count
[0]);
439 mod_zone_page_state(zone
, NR_ISOLATED_FILE
, count
[1]);
443 /* Similar to reclaim, but different enough that they don't share logic */
444 static bool too_many_isolated(struct zone
*zone
)
446 unsigned long active
, inactive
, isolated
;
448 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
) +
449 zone_page_state(zone
, NR_INACTIVE_ANON
);
450 active
= zone_page_state(zone
, NR_ACTIVE_FILE
) +
451 zone_page_state(zone
, NR_ACTIVE_ANON
);
452 isolated
= zone_page_state(zone
, NR_ISOLATED_FILE
) +
453 zone_page_state(zone
, NR_ISOLATED_ANON
);
455 return isolated
> (inactive
+ active
) / 2;
459 * isolate_migratepages_range() - isolate all migrate-able pages in range.
460 * @zone: Zone pages are in.
461 * @cc: Compaction control structure.
462 * @low_pfn: The first PFN of the range.
463 * @end_pfn: The one-past-the-last PFN of the range.
465 * Isolate all pages that can be migrated from the range specified by
466 * [low_pfn, end_pfn). Returns zero if there is a fatal signal
467 * pending), otherwise PFN of the first page that was not scanned
468 * (which may be both less, equal to or more then end_pfn).
470 * Assumes that cc->migratepages is empty and cc->nr_migratepages is
473 * Apart from cc->migratepages and cc->nr_migratetypes this function
474 * does not modify any cc's fields, in particular it does not modify
475 * (or read for that matter) cc->migrate_pfn.
478 isolate_migratepages_range(struct zone
*zone
, struct compact_control
*cc
,
479 unsigned long low_pfn
, unsigned long end_pfn
)
481 unsigned long last_pageblock_nr
= 0, pageblock_nr
;
482 unsigned long nr_scanned
= 0, nr_isolated
= 0;
483 struct list_head
*migratelist
= &cc
->migratepages
;
484 isolate_mode_t mode
= 0;
485 struct lruvec
*lruvec
;
488 struct page
*page
= NULL
, *valid_page
= NULL
;
491 * Ensure that there are not too many pages isolated from the LRU
492 * list by either parallel reclaimers or compaction. If there are,
493 * delay for some time until fewer pages are isolated
495 while (unlikely(too_many_isolated(zone
))) {
496 /* async migration should just abort */
500 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
502 if (fatal_signal_pending(current
))
506 /* Time to isolate some pages for migration */
508 for (; low_pfn
< end_pfn
; low_pfn
++) {
509 /* give a chance to irqs before checking need_resched() */
510 if (locked
&& !((low_pfn
+1) % SWAP_CLUSTER_MAX
)) {
511 if (should_release_lock(&zone
->lru_lock
)) {
512 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
518 * migrate_pfn does not necessarily start aligned to a
519 * pageblock. Ensure that pfn_valid is called when moving
520 * into a new MAX_ORDER_NR_PAGES range in case of large
521 * memory holes within the zone
523 if ((low_pfn
& (MAX_ORDER_NR_PAGES
- 1)) == 0) {
524 if (!pfn_valid(low_pfn
)) {
525 low_pfn
+= MAX_ORDER_NR_PAGES
- 1;
530 if (!pfn_valid_within(low_pfn
))
535 * Get the page and ensure the page is within the same zone.
536 * See the comment in isolate_freepages about overlapping
537 * nodes. It is deliberate that the new zone lock is not taken
538 * as memory compaction should not move pages between nodes.
540 page
= pfn_to_page(low_pfn
);
541 if (page_zone(page
) != zone
)
547 /* If isolation recently failed, do not retry */
548 pageblock_nr
= low_pfn
>> pageblock_order
;
549 if (!isolation_suitable(cc
, page
))
557 * For async migration, also only scan in MOVABLE blocks. Async
558 * migration is optimistic to see if the minimum amount of work
559 * satisfies the allocation
561 if (!cc
->sync
&& last_pageblock_nr
!= pageblock_nr
&&
562 !migrate_async_suitable(get_pageblock_migratetype(page
))) {
563 cc
->finished_update_migrate
= true;
567 /* Check may be lockless but that's ok as we recheck later */
572 * PageLRU is set. lru_lock normally excludes isolation
573 * splitting and collapsing (collapsing has already happened
574 * if PageLRU is set) but the lock is not necessarily taken
575 * here and it is wasteful to take it just to check transhuge.
576 * Check TransHuge without lock and skip the whole pageblock if
577 * it's either a transhuge or hugetlbfs page, as calling
578 * compound_order() without preventing THP from splitting the
579 * page underneath us may return surprising results.
581 if (PageTransHuge(page
)) {
584 low_pfn
+= (1 << compound_order(page
)) - 1;
588 /* Check if it is ok to still hold the lock */
589 locked
= compact_checklock_irqsave(&zone
->lru_lock
, &flags
,
591 if (!locked
|| fatal_signal_pending(current
))
594 /* Recheck PageLRU and PageTransHuge under lock */
597 if (PageTransHuge(page
)) {
598 low_pfn
+= (1 << compound_order(page
)) - 1;
603 mode
|= ISOLATE_ASYNC_MIGRATE
;
605 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
607 /* Try isolate the page */
608 if (__isolate_lru_page(page
, mode
) != 0)
611 VM_BUG_ON(PageTransCompound(page
));
613 /* Successfully isolated */
614 cc
->finished_update_migrate
= true;
615 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
616 list_add(&page
->lru
, migratelist
);
617 cc
->nr_migratepages
++;
620 /* Avoid isolating too much */
621 if (cc
->nr_migratepages
== COMPACT_CLUSTER_MAX
) {
629 low_pfn
+= pageblock_nr_pages
;
630 low_pfn
= ALIGN(low_pfn
, pageblock_nr_pages
) - 1;
631 last_pageblock_nr
= pageblock_nr
;
634 acct_isolated(zone
, locked
, cc
);
637 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
639 /* Update the pageblock-skip if the whole pageblock was scanned */
640 if (low_pfn
== end_pfn
)
641 update_pageblock_skip(cc
, valid_page
, nr_isolated
, true);
643 trace_mm_compaction_isolate_migratepages(nr_scanned
, nr_isolated
);
648 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
649 #ifdef CONFIG_COMPACTION
651 * Based on information in the current compact_control, find blocks
652 * suitable for isolating free pages from and then isolate them.
654 static void isolate_freepages(struct zone
*zone
,
655 struct compact_control
*cc
)
658 unsigned long high_pfn
, low_pfn
, pfn
, zone_end_pfn
, end_pfn
;
659 int nr_freepages
= cc
->nr_freepages
;
660 struct list_head
*freelist
= &cc
->freepages
;
663 * Initialise the free scanner. The starting point is where we last
664 * scanned from (or the end of the zone if starting). The low point
665 * is the end of the pageblock the migration scanner is using.
668 low_pfn
= cc
->migrate_pfn
+ pageblock_nr_pages
;
671 * Take care that if the migration scanner is at the end of the zone
672 * that the free scanner does not accidentally move to the next zone
673 * in the next isolation cycle.
675 high_pfn
= min(low_pfn
, pfn
);
677 zone_end_pfn
= zone
->zone_start_pfn
+ zone
->spanned_pages
;
680 * Isolate free pages until enough are available to migrate the
681 * pages on cc->migratepages. We stop searching if the migrate
682 * and free page scanners meet or enough free pages are isolated.
684 for (; pfn
> low_pfn
&& cc
->nr_migratepages
> nr_freepages
;
685 pfn
-= pageblock_nr_pages
) {
686 unsigned long isolated
;
692 * Check for overlapping nodes/zones. It's possible on some
693 * configurations to have a setup like
695 * i.e. it's possible that all pages within a zones range of
696 * pages do not belong to a single zone.
698 page
= pfn_to_page(pfn
);
699 if (page_zone(page
) != zone
)
702 /* Check the block is suitable for migration */
703 if (!suitable_migration_target(page
))
706 /* If isolation recently failed, do not retry */
707 if (!isolation_suitable(cc
, page
))
710 /* Found a block suitable for isolating free pages from */
712 end_pfn
= min(pfn
+ pageblock_nr_pages
, zone_end_pfn
);
713 isolated
= isolate_freepages_block(cc
, pfn
, end_pfn
,
715 nr_freepages
+= isolated
;
718 * Record the highest PFN we isolated pages from. When next
719 * looking for free pages, the search will restart here as
720 * page migration may have returned some pages to the allocator
723 cc
->finished_update_free
= true;
724 high_pfn
= max(high_pfn
, pfn
);
728 /* split_free_page does not map the pages */
731 cc
->free_pfn
= high_pfn
;
732 cc
->nr_freepages
= nr_freepages
;
736 * This is a migrate-callback that "allocates" freepages by taking pages
737 * from the isolated freelists in the block we are migrating to.
739 static struct page
*compaction_alloc(struct page
*migratepage
,
743 struct compact_control
*cc
= (struct compact_control
*)data
;
744 struct page
*freepage
;
746 /* Isolate free pages if necessary */
747 if (list_empty(&cc
->freepages
)) {
748 isolate_freepages(cc
->zone
, cc
);
750 if (list_empty(&cc
->freepages
))
754 freepage
= list_entry(cc
->freepages
.next
, struct page
, lru
);
755 list_del(&freepage
->lru
);
762 * We cannot control nr_migratepages and nr_freepages fully when migration is
763 * running as migrate_pages() has no knowledge of compact_control. When
764 * migration is complete, we count the number of pages on the lists by hand.
766 static void update_nr_listpages(struct compact_control
*cc
)
768 int nr_migratepages
= 0;
769 int nr_freepages
= 0;
772 list_for_each_entry(page
, &cc
->migratepages
, lru
)
774 list_for_each_entry(page
, &cc
->freepages
, lru
)
777 cc
->nr_migratepages
= nr_migratepages
;
778 cc
->nr_freepages
= nr_freepages
;
781 /* possible outcome of isolate_migratepages */
783 ISOLATE_ABORT
, /* Abort compaction now */
784 ISOLATE_NONE
, /* No pages isolated, continue scanning */
785 ISOLATE_SUCCESS
, /* Pages isolated, migrate */
789 * Isolate all pages that can be migrated from the block pointed to by
790 * the migrate scanner within compact_control.
792 static isolate_migrate_t
isolate_migratepages(struct zone
*zone
,
793 struct compact_control
*cc
)
795 unsigned long low_pfn
, end_pfn
;
797 /* Do not scan outside zone boundaries */
798 low_pfn
= max(cc
->migrate_pfn
, zone
->zone_start_pfn
);
800 /* Only scan within a pageblock boundary */
801 end_pfn
= ALIGN(low_pfn
+ pageblock_nr_pages
, pageblock_nr_pages
);
803 /* Do not cross the free scanner or scan within a memory hole */
804 if (end_pfn
> cc
->free_pfn
|| !pfn_valid(low_pfn
)) {
805 cc
->migrate_pfn
= end_pfn
;
809 /* Perform the isolation */
810 low_pfn
= isolate_migratepages_range(zone
, cc
, low_pfn
, end_pfn
);
811 if (!low_pfn
|| cc
->contended
)
812 return ISOLATE_ABORT
;
814 cc
->migrate_pfn
= low_pfn
;
816 return ISOLATE_SUCCESS
;
819 static int compact_finished(struct zone
*zone
,
820 struct compact_control
*cc
)
822 unsigned long watermark
;
824 if (fatal_signal_pending(current
))
825 return COMPACT_PARTIAL
;
827 /* Compaction run completes if the migrate and free scanner meet */
828 if (cc
->free_pfn
<= cc
->migrate_pfn
) {
830 * Mark that the PG_migrate_skip information should be cleared
831 * by kswapd when it goes to sleep. kswapd does not set the
832 * flag itself as the decision to be clear should be directly
833 * based on an allocation request.
835 if (!current_is_kswapd())
836 zone
->compact_blockskip_flush
= true;
838 return COMPACT_COMPLETE
;
842 * order == -1 is expected when compacting via
843 * /proc/sys/vm/compact_memory
846 return COMPACT_CONTINUE
;
848 /* Compaction run is not finished if the watermark is not met */
849 watermark
= low_wmark_pages(zone
);
850 watermark
+= (1 << cc
->order
);
852 if (!zone_watermark_ok(zone
, cc
->order
, watermark
, 0, 0))
853 return COMPACT_CONTINUE
;
855 /* Direct compactor: Is a suitable page free? */
857 /* Was a suitable page captured? */
859 return COMPACT_PARTIAL
;
862 for (order
= cc
->order
; order
< MAX_ORDER
; order
++) {
863 struct free_area
*area
= &zone
->free_area
[cc
->order
];
864 /* Job done if page is free of the right migratetype */
865 if (!list_empty(&area
->free_list
[cc
->migratetype
]))
866 return COMPACT_PARTIAL
;
868 /* Job done if allocation would set block type */
869 if (cc
->order
>= pageblock_order
&& area
->nr_free
)
870 return COMPACT_PARTIAL
;
874 return COMPACT_CONTINUE
;
878 * compaction_suitable: Is this suitable to run compaction on this zone now?
880 * COMPACT_SKIPPED - If there are too few free pages for compaction
881 * COMPACT_PARTIAL - If the allocation would succeed without compaction
882 * COMPACT_CONTINUE - If compaction should run now
884 unsigned long compaction_suitable(struct zone
*zone
, int order
)
887 unsigned long watermark
;
890 * order == -1 is expected when compacting via
891 * /proc/sys/vm/compact_memory
894 return COMPACT_CONTINUE
;
897 * Watermarks for order-0 must be met for compaction. Note the 2UL.
898 * This is because during migration, copies of pages need to be
899 * allocated and for a short time, the footprint is higher
901 watermark
= low_wmark_pages(zone
) + (2UL << order
);
902 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
903 return COMPACT_SKIPPED
;
906 * fragmentation index determines if allocation failures are due to
907 * low memory or external fragmentation
909 * index of -1000 implies allocations might succeed depending on
911 * index towards 0 implies failure is due to lack of memory
912 * index towards 1000 implies failure is due to fragmentation
914 * Only compact if a failure would be due to fragmentation.
916 fragindex
= fragmentation_index(zone
, order
);
917 if (fragindex
>= 0 && fragindex
<= sysctl_extfrag_threshold
)
918 return COMPACT_SKIPPED
;
920 if (fragindex
== -1000 && zone_watermark_ok(zone
, order
, watermark
,
922 return COMPACT_PARTIAL
;
924 return COMPACT_CONTINUE
;
927 static int compact_zone(struct zone
*zone
, struct compact_control
*cc
)
930 unsigned long start_pfn
= zone
->zone_start_pfn
;
931 unsigned long end_pfn
= zone
->zone_start_pfn
+ zone
->spanned_pages
;
933 ret
= compaction_suitable(zone
, cc
->order
);
935 case COMPACT_PARTIAL
:
936 case COMPACT_SKIPPED
:
937 /* Compaction is likely to fail */
939 case COMPACT_CONTINUE
:
940 /* Fall through to compaction */
945 * Setup to move all movable pages to the end of the zone. Used cached
946 * information on where the scanners should start but check that it
947 * is initialised by ensuring the values are within zone boundaries.
949 cc
->migrate_pfn
= zone
->compact_cached_migrate_pfn
;
950 cc
->free_pfn
= zone
->compact_cached_free_pfn
;
951 if (cc
->free_pfn
< start_pfn
|| cc
->free_pfn
> end_pfn
) {
952 cc
->free_pfn
= end_pfn
& ~(pageblock_nr_pages
-1);
953 zone
->compact_cached_free_pfn
= cc
->free_pfn
;
955 if (cc
->migrate_pfn
< start_pfn
|| cc
->migrate_pfn
> end_pfn
) {
956 cc
->migrate_pfn
= start_pfn
;
957 zone
->compact_cached_migrate_pfn
= cc
->migrate_pfn
;
961 * Clear pageblock skip if there were failures recently and compaction
962 * is about to be retried after being deferred. kswapd does not do
963 * this reset as it'll reset the cached information when going to sleep.
965 if (compaction_restarting(zone
, cc
->order
) && !current_is_kswapd())
966 __reset_isolation_suitable(zone
);
968 migrate_prep_local();
970 while ((ret
= compact_finished(zone
, cc
)) == COMPACT_CONTINUE
) {
971 unsigned long nr_migrate
, nr_remaining
;
974 switch (isolate_migratepages(zone
, cc
)) {
976 ret
= COMPACT_PARTIAL
;
977 putback_lru_pages(&cc
->migratepages
);
978 cc
->nr_migratepages
= 0;
982 case ISOLATE_SUCCESS
:
986 nr_migrate
= cc
->nr_migratepages
;
987 err
= migrate_pages(&cc
->migratepages
, compaction_alloc
,
988 (unsigned long)cc
, false,
989 cc
->sync
? MIGRATE_SYNC_LIGHT
: MIGRATE_ASYNC
);
990 update_nr_listpages(cc
);
991 nr_remaining
= cc
->nr_migratepages
;
993 count_vm_event(COMPACTBLOCKS
);
994 count_vm_events(COMPACTPAGES
, nr_migrate
- nr_remaining
);
996 count_vm_events(COMPACTPAGEFAILED
, nr_remaining
);
997 trace_mm_compaction_migratepages(nr_migrate
- nr_remaining
,
1000 /* Release LRU pages not migrated */
1002 putback_lru_pages(&cc
->migratepages
);
1003 cc
->nr_migratepages
= 0;
1004 if (err
== -ENOMEM
) {
1005 ret
= COMPACT_PARTIAL
;
1010 /* Capture a page now if it is a suitable size */
1011 compact_capture_page(cc
);
1015 /* Release free pages and check accounting */
1016 cc
->nr_freepages
-= release_freepages(&cc
->freepages
);
1017 VM_BUG_ON(cc
->nr_freepages
!= 0);
1022 static unsigned long compact_zone_order(struct zone
*zone
,
1023 int order
, gfp_t gfp_mask
,
1024 bool sync
, bool *contended
,
1028 struct compact_control cc
= {
1030 .nr_migratepages
= 0,
1032 .migratetype
= allocflags_to_migratetype(gfp_mask
),
1037 INIT_LIST_HEAD(&cc
.freepages
);
1038 INIT_LIST_HEAD(&cc
.migratepages
);
1040 ret
= compact_zone(zone
, &cc
);
1042 VM_BUG_ON(!list_empty(&cc
.freepages
));
1043 VM_BUG_ON(!list_empty(&cc
.migratepages
));
1045 *contended
= cc
.contended
;
1049 int sysctl_extfrag_threshold
= 500;
1052 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1053 * @zonelist: The zonelist used for the current allocation
1054 * @order: The order of the current allocation
1055 * @gfp_mask: The GFP mask of the current allocation
1056 * @nodemask: The allowed nodes to allocate from
1057 * @sync: Whether migration is synchronous or not
1058 * @contended: Return value that is true if compaction was aborted due to lock contention
1059 * @page: Optionally capture a free page of the requested order during compaction
1061 * This is the main entry point for direct page compaction.
1063 unsigned long try_to_compact_pages(struct zonelist
*zonelist
,
1064 int order
, gfp_t gfp_mask
, nodemask_t
*nodemask
,
1065 bool sync
, bool *contended
, struct page
**page
)
1067 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
1068 int may_enter_fs
= gfp_mask
& __GFP_FS
;
1069 int may_perform_io
= gfp_mask
& __GFP_IO
;
1072 int rc
= COMPACT_SKIPPED
;
1073 int alloc_flags
= 0;
1075 /* Check if the GFP flags allow compaction */
1076 if (!order
|| !may_enter_fs
|| !may_perform_io
)
1079 count_vm_event(COMPACTSTALL
);
1082 if (allocflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
1083 alloc_flags
|= ALLOC_CMA
;
1085 /* Compact each zone in the list */
1086 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, high_zoneidx
,
1090 status
= compact_zone_order(zone
, order
, gfp_mask
, sync
,
1092 rc
= max(status
, rc
);
1094 /* If a normal allocation would succeed, stop compacting */
1095 if (zone_watermark_ok(zone
, order
, low_wmark_pages(zone
), 0,
1104 /* Compact all zones within a node */
1105 static int __compact_pgdat(pg_data_t
*pgdat
, struct compact_control
*cc
)
1110 for (zoneid
= 0; zoneid
< MAX_NR_ZONES
; zoneid
++) {
1112 zone
= &pgdat
->node_zones
[zoneid
];
1113 if (!populated_zone(zone
))
1116 cc
->nr_freepages
= 0;
1117 cc
->nr_migratepages
= 0;
1119 INIT_LIST_HEAD(&cc
->freepages
);
1120 INIT_LIST_HEAD(&cc
->migratepages
);
1122 if (cc
->order
== -1 || !compaction_deferred(zone
, cc
->order
))
1123 compact_zone(zone
, cc
);
1125 if (cc
->order
> 0) {
1126 int ok
= zone_watermark_ok(zone
, cc
->order
,
1127 low_wmark_pages(zone
), 0, 0);
1128 if (ok
&& cc
->order
>= zone
->compact_order_failed
)
1129 zone
->compact_order_failed
= cc
->order
+ 1;
1130 /* Currently async compaction is never deferred. */
1131 else if (!ok
&& cc
->sync
)
1132 defer_compaction(zone
, cc
->order
);
1135 VM_BUG_ON(!list_empty(&cc
->freepages
));
1136 VM_BUG_ON(!list_empty(&cc
->migratepages
));
1142 int compact_pgdat(pg_data_t
*pgdat
, int order
)
1144 struct compact_control cc
= {
1150 return __compact_pgdat(pgdat
, &cc
);
1153 static int compact_node(int nid
)
1155 struct compact_control cc
= {
1161 return __compact_pgdat(NODE_DATA(nid
), &cc
);
1164 /* Compact all nodes in the system */
1165 static int compact_nodes(void)
1169 /* Flush pending updates to the LRU lists */
1170 lru_add_drain_all();
1172 for_each_online_node(nid
)
1175 return COMPACT_COMPLETE
;
1178 /* The written value is actually unused, all memory is compacted */
1179 int sysctl_compact_memory
;
1181 /* This is the entry point for compacting all nodes via /proc/sys/vm */
1182 int sysctl_compaction_handler(struct ctl_table
*table
, int write
,
1183 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
1186 return compact_nodes();
1191 int sysctl_extfrag_handler(struct ctl_table
*table
, int write
,
1192 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
1194 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
1199 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1200 ssize_t
sysfs_compact_node(struct device
*dev
,
1201 struct device_attribute
*attr
,
1202 const char *buf
, size_t count
)
1206 if (nid
>= 0 && nid
< nr_node_ids
&& node_online(nid
)) {
1207 /* Flush pending updates to the LRU lists */
1208 lru_add_drain_all();
1215 static DEVICE_ATTR(compact
, S_IWUSR
, NULL
, sysfs_compact_node
);
1217 int compaction_register_node(struct node
*node
)
1219 return device_create_file(&node
->dev
, &dev_attr_compact
);
1222 void compaction_unregister_node(struct node
*node
)
1224 return device_remove_file(&node
->dev
, &dev_attr_compact
);
1226 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1228 #endif /* CONFIG_COMPACTION */