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 #define CREATE_TRACE_POINTS
20 #include <trace/events/compaction.h>
23 * compact_control is used to track pages being migrated and the free pages
24 * they are being migrated to during memory compaction. The free_pfn starts
25 * at the end of a zone and migrate_pfn begins at the start. Movable pages
26 * are moved to the end of a zone during a compaction run and the run
27 * completes when free_pfn <= migrate_pfn
29 struct compact_control
{
30 struct list_head freepages
; /* List of free pages to migrate to */
31 struct list_head migratepages
; /* List of pages being migrated */
32 unsigned long nr_freepages
; /* Number of isolated free pages */
33 unsigned long nr_migratepages
; /* Number of pages to migrate */
34 unsigned long free_pfn
; /* isolate_freepages search base */
35 unsigned long migrate_pfn
; /* isolate_migratepages search base */
36 bool sync
; /* Synchronous migration */
38 /* Account for isolated anon and file pages */
39 unsigned long nr_anon
;
40 unsigned long nr_file
;
42 unsigned int order
; /* order a direct compactor needs */
43 int migratetype
; /* MOVABLE, RECLAIMABLE etc */
47 static unsigned long release_freepages(struct list_head
*freelist
)
49 struct page
*page
, *next
;
50 unsigned long count
= 0;
52 list_for_each_entry_safe(page
, next
, freelist
, lru
) {
61 /* Isolate free pages onto a private freelist. Must hold zone->lock */
62 static unsigned long isolate_freepages_block(struct zone
*zone
,
63 unsigned long blockpfn
,
64 struct list_head
*freelist
)
66 unsigned long zone_end_pfn
, end_pfn
;
67 int nr_scanned
= 0, total_isolated
= 0;
70 /* Get the last PFN we should scan for free pages at */
71 zone_end_pfn
= zone
->zone_start_pfn
+ zone
->spanned_pages
;
72 end_pfn
= min(blockpfn
+ pageblock_nr_pages
, zone_end_pfn
);
74 /* Find the first usable PFN in the block to initialse page cursor */
75 for (; blockpfn
< end_pfn
; blockpfn
++) {
76 if (pfn_valid_within(blockpfn
))
79 cursor
= pfn_to_page(blockpfn
);
81 /* Isolate free pages. This assumes the block is valid */
82 for (; blockpfn
< end_pfn
; blockpfn
++, cursor
++) {
84 struct page
*page
= cursor
;
86 if (!pfn_valid_within(blockpfn
))
93 /* Found a free page, break it into order-0 pages */
94 isolated
= split_free_page(page
);
95 total_isolated
+= isolated
;
96 for (i
= 0; i
< isolated
; i
++) {
97 list_add(&page
->lru
, freelist
);
101 /* If a page was split, advance to the end of it */
103 blockpfn
+= isolated
- 1;
104 cursor
+= isolated
- 1;
108 trace_mm_compaction_isolate_freepages(nr_scanned
, total_isolated
);
109 return total_isolated
;
112 /* Returns true if the page is within a block suitable for migration to */
113 static bool suitable_migration_target(struct page
*page
)
116 int migratetype
= get_pageblock_migratetype(page
);
118 /* Don't interfere with memory hot-remove or the min_free_kbytes blocks */
119 if (migratetype
== MIGRATE_ISOLATE
|| migratetype
== MIGRATE_RESERVE
)
122 /* If the page is a large free page, then allow migration */
123 if (PageBuddy(page
) && page_order(page
) >= pageblock_order
)
126 /* If the block is MIGRATE_MOVABLE, allow migration */
127 if (migratetype
== MIGRATE_MOVABLE
)
130 /* Otherwise skip the block */
135 * Based on information in the current compact_control, find blocks
136 * suitable for isolating free pages from and then isolate them.
138 static void isolate_freepages(struct zone
*zone
,
139 struct compact_control
*cc
)
142 unsigned long high_pfn
, low_pfn
, pfn
;
144 int nr_freepages
= cc
->nr_freepages
;
145 struct list_head
*freelist
= &cc
->freepages
;
148 low_pfn
= cc
->migrate_pfn
+ pageblock_nr_pages
;
152 * Isolate free pages until enough are available to migrate the
153 * pages on cc->migratepages. We stop searching if the migrate
154 * and free page scanners meet or enough free pages are isolated.
156 spin_lock_irqsave(&zone
->lock
, flags
);
157 for (; pfn
> low_pfn
&& cc
->nr_migratepages
> nr_freepages
;
158 pfn
-= pageblock_nr_pages
) {
159 unsigned long isolated
;
165 * Check for overlapping nodes/zones. It's possible on some
166 * configurations to have a setup like
168 * i.e. it's possible that all pages within a zones range of
169 * pages do not belong to a single zone.
171 page
= pfn_to_page(pfn
);
172 if (page_zone(page
) != zone
)
175 /* Check the block is suitable for migration */
176 if (!suitable_migration_target(page
))
179 /* Found a block suitable for isolating free pages from */
180 isolated
= isolate_freepages_block(zone
, pfn
, freelist
);
181 nr_freepages
+= isolated
;
184 * Record the highest PFN we isolated pages from. When next
185 * looking for free pages, the search will restart here as
186 * page migration may have returned some pages to the allocator
189 high_pfn
= max(high_pfn
, pfn
);
191 spin_unlock_irqrestore(&zone
->lock
, flags
);
193 /* split_free_page does not map the pages */
194 list_for_each_entry(page
, freelist
, lru
) {
195 arch_alloc_page(page
, 0);
196 kernel_map_pages(page
, 1, 1);
199 cc
->free_pfn
= high_pfn
;
200 cc
->nr_freepages
= nr_freepages
;
203 /* Update the number of anon and file isolated pages in the zone */
204 static void acct_isolated(struct zone
*zone
, struct compact_control
*cc
)
207 unsigned int count
[NR_LRU_LISTS
] = { 0, };
209 list_for_each_entry(page
, &cc
->migratepages
, lru
) {
210 int lru
= page_lru_base_type(page
);
214 cc
->nr_anon
= count
[LRU_ACTIVE_ANON
] + count
[LRU_INACTIVE_ANON
];
215 cc
->nr_file
= count
[LRU_ACTIVE_FILE
] + count
[LRU_INACTIVE_FILE
];
216 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
, cc
->nr_anon
);
217 __mod_zone_page_state(zone
, NR_ISOLATED_FILE
, cc
->nr_file
);
220 /* Similar to reclaim, but different enough that they don't share logic */
221 static bool too_many_isolated(struct zone
*zone
)
223 unsigned long active
, inactive
, isolated
;
225 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
) +
226 zone_page_state(zone
, NR_INACTIVE_ANON
);
227 active
= zone_page_state(zone
, NR_ACTIVE_FILE
) +
228 zone_page_state(zone
, NR_ACTIVE_ANON
);
229 isolated
= zone_page_state(zone
, NR_ISOLATED_FILE
) +
230 zone_page_state(zone
, NR_ISOLATED_ANON
);
232 return isolated
> (inactive
+ active
) / 2;
236 * Isolate all pages that can be migrated from the block pointed to by
237 * the migrate scanner within compact_control.
239 static unsigned long isolate_migratepages(struct zone
*zone
,
240 struct compact_control
*cc
)
242 unsigned long low_pfn
, end_pfn
;
243 unsigned long last_pageblock_nr
= 0, pageblock_nr
;
244 unsigned long nr_scanned
= 0, nr_isolated
= 0;
245 struct list_head
*migratelist
= &cc
->migratepages
;
247 /* Do not scan outside zone boundaries */
248 low_pfn
= max(cc
->migrate_pfn
, zone
->zone_start_pfn
);
250 /* Only scan within a pageblock boundary */
251 end_pfn
= ALIGN(low_pfn
+ pageblock_nr_pages
, pageblock_nr_pages
);
253 /* Do not cross the free scanner or scan within a memory hole */
254 if (end_pfn
> cc
->free_pfn
|| !pfn_valid(low_pfn
)) {
255 cc
->migrate_pfn
= end_pfn
;
260 * Ensure that there are not too many pages isolated from the LRU
261 * list by either parallel reclaimers or compaction. If there are,
262 * delay for some time until fewer pages are isolated
264 while (unlikely(too_many_isolated(zone
))) {
265 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
267 if (fatal_signal_pending(current
))
271 /* Time to isolate some pages for migration */
272 spin_lock_irq(&zone
->lru_lock
);
273 for (; low_pfn
< end_pfn
; low_pfn
++) {
275 if (!pfn_valid_within(low_pfn
))
279 /* Get the page and skip if free */
280 page
= pfn_to_page(low_pfn
);
285 * For async migration, also only scan in MOVABLE blocks. Async
286 * migration is optimistic to see if the minimum amount of work
287 * satisfies the allocation
289 pageblock_nr
= low_pfn
>> pageblock_order
;
290 if (!cc
->sync
&& last_pageblock_nr
!= pageblock_nr
&&
291 get_pageblock_migratetype(page
) != MIGRATE_MOVABLE
) {
292 low_pfn
+= pageblock_nr_pages
;
293 low_pfn
= ALIGN(low_pfn
, pageblock_nr_pages
) - 1;
294 last_pageblock_nr
= pageblock_nr
;
302 * PageLRU is set, and lru_lock excludes isolation,
303 * splitting and collapsing (collapsing has already
304 * happened if PageLRU is set).
306 if (PageTransHuge(page
)) {
307 low_pfn
+= (1 << compound_order(page
)) - 1;
311 /* Try isolate the page */
312 if (__isolate_lru_page(page
, ISOLATE_BOTH
, 0) != 0)
315 VM_BUG_ON(PageTransCompound(page
));
317 /* Successfully isolated */
318 del_page_from_lru_list(zone
, page
, page_lru(page
));
319 list_add(&page
->lru
, migratelist
);
320 cc
->nr_migratepages
++;
323 /* Avoid isolating too much */
324 if (cc
->nr_migratepages
== COMPACT_CLUSTER_MAX
)
328 acct_isolated(zone
, cc
);
330 spin_unlock_irq(&zone
->lru_lock
);
331 cc
->migrate_pfn
= low_pfn
;
333 trace_mm_compaction_isolate_migratepages(nr_scanned
, nr_isolated
);
335 return cc
->nr_migratepages
;
339 * This is a migrate-callback that "allocates" freepages by taking pages
340 * from the isolated freelists in the block we are migrating to.
342 static struct page
*compaction_alloc(struct page
*migratepage
,
346 struct compact_control
*cc
= (struct compact_control
*)data
;
347 struct page
*freepage
;
349 /* Isolate free pages if necessary */
350 if (list_empty(&cc
->freepages
)) {
351 isolate_freepages(cc
->zone
, cc
);
353 if (list_empty(&cc
->freepages
))
357 freepage
= list_entry(cc
->freepages
.next
, struct page
, lru
);
358 list_del(&freepage
->lru
);
365 * We cannot control nr_migratepages and nr_freepages fully when migration is
366 * running as migrate_pages() has no knowledge of compact_control. When
367 * migration is complete, we count the number of pages on the lists by hand.
369 static void update_nr_listpages(struct compact_control
*cc
)
371 int nr_migratepages
= 0;
372 int nr_freepages
= 0;
375 list_for_each_entry(page
, &cc
->migratepages
, lru
)
377 list_for_each_entry(page
, &cc
->freepages
, lru
)
380 cc
->nr_migratepages
= nr_migratepages
;
381 cc
->nr_freepages
= nr_freepages
;
384 static int compact_finished(struct zone
*zone
,
385 struct compact_control
*cc
)
388 unsigned long watermark
;
390 if (fatal_signal_pending(current
))
391 return COMPACT_PARTIAL
;
393 /* Compaction run completes if the migrate and free scanner meet */
394 if (cc
->free_pfn
<= cc
->migrate_pfn
)
395 return COMPACT_COMPLETE
;
397 /* Compaction run is not finished if the watermark is not met */
398 watermark
= low_wmark_pages(zone
);
399 watermark
+= (1 << cc
->order
);
401 if (!zone_watermark_ok(zone
, cc
->order
, watermark
, 0, 0))
402 return COMPACT_CONTINUE
;
405 * order == -1 is expected when compacting via
406 * /proc/sys/vm/compact_memory
409 return COMPACT_CONTINUE
;
411 /* Direct compactor: Is a suitable page free? */
412 for (order
= cc
->order
; order
< MAX_ORDER
; order
++) {
413 /* Job done if page is free of the right migratetype */
414 if (!list_empty(&zone
->free_area
[order
].free_list
[cc
->migratetype
]))
415 return COMPACT_PARTIAL
;
417 /* Job done if allocation would set block type */
418 if (order
>= pageblock_order
&& zone
->free_area
[order
].nr_free
)
419 return COMPACT_PARTIAL
;
422 return COMPACT_CONTINUE
;
426 * compaction_suitable: Is this suitable to run compaction on this zone now?
428 * COMPACT_SKIPPED - If there are too few free pages for compaction
429 * COMPACT_PARTIAL - If the allocation would succeed without compaction
430 * COMPACT_CONTINUE - If compaction should run now
432 unsigned long compaction_suitable(struct zone
*zone
, int order
)
435 unsigned long watermark
;
438 * Watermarks for order-0 must be met for compaction. Note the 2UL.
439 * This is because during migration, copies of pages need to be
440 * allocated and for a short time, the footprint is higher
442 watermark
= low_wmark_pages(zone
) + (2UL << order
);
443 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
444 return COMPACT_SKIPPED
;
447 * order == -1 is expected when compacting via
448 * /proc/sys/vm/compact_memory
451 return COMPACT_CONTINUE
;
454 * fragmentation index determines if allocation failures are due to
455 * low memory or external fragmentation
457 * index of -1 implies allocations might succeed dependingon watermarks
458 * index towards 0 implies failure is due to lack of memory
459 * index towards 1000 implies failure is due to fragmentation
461 * Only compact if a failure would be due to fragmentation.
463 fragindex
= fragmentation_index(zone
, order
);
464 if (fragindex
>= 0 && fragindex
<= sysctl_extfrag_threshold
)
465 return COMPACT_SKIPPED
;
467 if (fragindex
== -1 && zone_watermark_ok(zone
, order
, watermark
, 0, 0))
468 return COMPACT_PARTIAL
;
470 return COMPACT_CONTINUE
;
473 static int compact_zone(struct zone
*zone
, struct compact_control
*cc
)
477 ret
= compaction_suitable(zone
, cc
->order
);
479 case COMPACT_PARTIAL
:
480 case COMPACT_SKIPPED
:
481 /* Compaction is likely to fail */
483 case COMPACT_CONTINUE
:
484 /* Fall through to compaction */
488 /* Setup to move all movable pages to the end of the zone */
489 cc
->migrate_pfn
= zone
->zone_start_pfn
;
490 cc
->free_pfn
= cc
->migrate_pfn
+ zone
->spanned_pages
;
491 cc
->free_pfn
&= ~(pageblock_nr_pages
-1);
493 migrate_prep_local();
495 while ((ret
= compact_finished(zone
, cc
)) == COMPACT_CONTINUE
) {
496 unsigned long nr_migrate
, nr_remaining
;
498 if (!isolate_migratepages(zone
, cc
))
501 nr_migrate
= cc
->nr_migratepages
;
502 migrate_pages(&cc
->migratepages
, compaction_alloc
,
503 (unsigned long)cc
, false,
505 update_nr_listpages(cc
);
506 nr_remaining
= cc
->nr_migratepages
;
508 count_vm_event(COMPACTBLOCKS
);
509 count_vm_events(COMPACTPAGES
, nr_migrate
- nr_remaining
);
511 count_vm_events(COMPACTPAGEFAILED
, nr_remaining
);
512 trace_mm_compaction_migratepages(nr_migrate
- nr_remaining
,
515 /* Release LRU pages not migrated */
516 if (!list_empty(&cc
->migratepages
)) {
517 putback_lru_pages(&cc
->migratepages
);
518 cc
->nr_migratepages
= 0;
523 /* Release free pages and check accounting */
524 cc
->nr_freepages
-= release_freepages(&cc
->freepages
);
525 VM_BUG_ON(cc
->nr_freepages
!= 0);
530 unsigned long compact_zone_order(struct zone
*zone
,
531 int order
, gfp_t gfp_mask
,
534 struct compact_control cc
= {
536 .nr_migratepages
= 0,
538 .migratetype
= allocflags_to_migratetype(gfp_mask
),
542 INIT_LIST_HEAD(&cc
.freepages
);
543 INIT_LIST_HEAD(&cc
.migratepages
);
545 return compact_zone(zone
, &cc
);
548 int sysctl_extfrag_threshold
= 500;
551 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
552 * @zonelist: The zonelist used for the current allocation
553 * @order: The order of the current allocation
554 * @gfp_mask: The GFP mask of the current allocation
555 * @nodemask: The allowed nodes to allocate from
556 * @sync: Whether migration is synchronous or not
558 * This is the main entry point for direct page compaction.
560 unsigned long try_to_compact_pages(struct zonelist
*zonelist
,
561 int order
, gfp_t gfp_mask
, nodemask_t
*nodemask
,
564 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
565 int may_enter_fs
= gfp_mask
& __GFP_FS
;
566 int may_perform_io
= gfp_mask
& __GFP_IO
;
569 int rc
= COMPACT_SKIPPED
;
572 * Check whether it is worth even starting compaction. The order check is
573 * made because an assumption is made that the page allocator can satisfy
574 * the "cheaper" orders without taking special steps
576 if (!order
|| !may_enter_fs
|| !may_perform_io
)
579 count_vm_event(COMPACTSTALL
);
581 /* Compact each zone in the list */
582 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, high_zoneidx
,
586 status
= compact_zone_order(zone
, order
, gfp_mask
, sync
);
587 rc
= max(status
, rc
);
589 /* If a normal allocation would succeed, stop compacting */
590 if (zone_watermark_ok(zone
, order
, low_wmark_pages(zone
), 0, 0))
598 /* Compact all zones within a node */
599 static int compact_node(int nid
)
605 if (nid
< 0 || nid
>= nr_node_ids
|| !node_online(nid
))
607 pgdat
= NODE_DATA(nid
);
609 /* Flush pending updates to the LRU lists */
612 for (zoneid
= 0; zoneid
< MAX_NR_ZONES
; zoneid
++) {
613 struct compact_control cc
= {
615 .nr_migratepages
= 0,
619 zone
= &pgdat
->node_zones
[zoneid
];
620 if (!populated_zone(zone
))
624 INIT_LIST_HEAD(&cc
.freepages
);
625 INIT_LIST_HEAD(&cc
.migratepages
);
627 compact_zone(zone
, &cc
);
629 VM_BUG_ON(!list_empty(&cc
.freepages
));
630 VM_BUG_ON(!list_empty(&cc
.migratepages
));
636 /* Compact all nodes in the system */
637 static int compact_nodes(void)
641 for_each_online_node(nid
)
644 return COMPACT_COMPLETE
;
647 /* The written value is actually unused, all memory is compacted */
648 int sysctl_compact_memory
;
650 /* This is the entry point for compacting all nodes via /proc/sys/vm */
651 int sysctl_compaction_handler(struct ctl_table
*table
, int write
,
652 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
655 return compact_nodes();
660 int sysctl_extfrag_handler(struct ctl_table
*table
, int write
,
661 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
663 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
668 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
669 ssize_t
sysfs_compact_node(struct sys_device
*dev
,
670 struct sysdev_attribute
*attr
,
671 const char *buf
, size_t count
)
673 compact_node(dev
->id
);
677 static SYSDEV_ATTR(compact
, S_IWUSR
, NULL
, sysfs_compact_node
);
679 int compaction_register_node(struct node
*node
)
681 return sysdev_create_file(&node
->sysdev
, &attr_compact
);
684 void compaction_unregister_node(struct node
*node
)
686 return sysdev_remove_file(&node
->sysdev
, &attr_compact
);
688 #endif /* CONFIG_SYSFS && CONFIG_NUMA */