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;
54 * Isolate free pages onto a private freelist. Caller must hold zone->lock.
55 * If @strict is true, will abort returning 0 on any invalid PFNs or non-free
56 * pages inside of the pageblock (even though it may still end up isolating
59 static unsigned long isolate_freepages_block(unsigned long blockpfn,
60 unsigned long end_pfn,
61 struct list_head *freelist,
64 int nr_scanned = 0, total_isolated = 0;
67 cursor = pfn_to_page(blockpfn);
69 /* Isolate free pages. This assumes the block is valid */
70 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
72 struct page *page = cursor;
74 if (!pfn_valid_within(blockpfn)) {
81 if (!PageBuddy(page)) {
87 /* Found a free page, break it into order-0 pages */
88 isolated = split_free_page(page);
89 if (!isolated && strict)
91 total_isolated += isolated;
92 for (i = 0; i < isolated; i++) {
93 list_add(&page->lru, freelist);
97 /* If a page was split, advance to the end of it */
99 blockpfn += isolated - 1;
100 cursor += isolated - 1;
104 trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
105 return total_isolated;
109 * isolate_freepages_range() - isolate free pages.
110 * @start_pfn: The first PFN to start isolating.
111 * @end_pfn: The one-past-last PFN.
113 * Non-free pages, invalid PFNs, or zone boundaries within the
114 * [start_pfn, end_pfn) range are considered errors, cause function to
115 * undo its actions and return zero.
117 * Otherwise, function returns one-past-the-last PFN of isolated page
118 * (which may be greater then end_pfn if end fell in a middle of
122 isolate_freepages_range(unsigned long start_pfn, unsigned long end_pfn)
124 unsigned long isolated, pfn, block_end_pfn, flags;
125 struct zone *zone = NULL;
128 if (pfn_valid(start_pfn))
129 zone = page_zone(pfn_to_page(start_pfn));
131 for (pfn = start_pfn; pfn < end_pfn; pfn += isolated) {
132 if (!pfn_valid(pfn) || zone != page_zone(pfn_to_page(pfn)))
136 * On subsequent iterations ALIGN() is actually not needed,
137 * but we keep it that we not to complicate the code.
139 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
140 block_end_pfn = min(block_end_pfn, end_pfn);
142 spin_lock_irqsave(&zone->lock, flags);
143 isolated = isolate_freepages_block(pfn, block_end_pfn,
145 spin_unlock_irqrestore(&zone->lock, flags);
148 * In strict mode, isolate_freepages_block() returns 0 if
149 * there are any holes in the block (ie. invalid PFNs or
156 * If we managed to isolate pages, it is always (1 << n) *
157 * pageblock_nr_pages for some non-negative n. (Max order
158 * page may span two pageblocks).
162 /* split_free_page does not map the pages */
163 map_pages(&freelist);
166 /* Loop terminated early, cleanup. */
167 release_freepages(&freelist);
171 /* We don't use freelists for anything. */
175 /* Update the number of anon and file isolated pages in the zone */
176 static void acct_isolated(struct zone *zone, struct compact_control *cc)
179 unsigned int count[2] = { 0, };
181 list_for_each_entry(page, &cc->migratepages, lru)
182 count[!!page_is_file_cache(page)]++;
184 __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
185 __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
188 /* Similar to reclaim, but different enough that they don't share logic */
189 static bool too_many_isolated(struct zone *zone)
191 unsigned long active, inactive, isolated;
193 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
194 zone_page_state(zone, NR_INACTIVE_ANON);
195 active = zone_page_state(zone, NR_ACTIVE_FILE) +
196 zone_page_state(zone, NR_ACTIVE_ANON);
197 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
198 zone_page_state(zone, NR_ISOLATED_ANON);
200 return isolated > (inactive + active) / 2;
204 * isolate_migratepages_range() - isolate all migrate-able pages in range.
205 * @zone: Zone pages are in.
206 * @cc: Compaction control structure.
207 * @low_pfn: The first PFN of the range.
208 * @end_pfn: The one-past-the-last PFN of the range.
210 * Isolate all pages that can be migrated from the range specified by
211 * [low_pfn, end_pfn). Returns zero if there is a fatal signal
212 * pending), otherwise PFN of the first page that was not scanned
213 * (which may be both less, equal to or more then end_pfn).
215 * Assumes that cc->migratepages is empty and cc->nr_migratepages is
218 * Apart from cc->migratepages and cc->nr_migratetypes this function
219 * does not modify any cc's fields, in particular it does not modify
220 * (or read for that matter) cc->migrate_pfn.
223 isolate_migratepages_range(struct zone *zone, struct compact_control *cc,
224 unsigned long low_pfn, unsigned long end_pfn)
226 unsigned long last_pageblock_nr = 0, pageblock_nr;
227 unsigned long nr_scanned = 0, nr_isolated = 0;
228 struct list_head *migratelist = &cc->migratepages;
229 isolate_mode_t mode = 0;
230 struct lruvec *lruvec;
233 * Ensure that there are not too many pages isolated from the LRU
234 * list by either parallel reclaimers or compaction. If there are,
235 * delay for some time until fewer pages are isolated
237 while (unlikely(too_many_isolated(zone))) {
238 /* async migration should just abort */
242 congestion_wait(BLK_RW_ASYNC, HZ/10);
244 if (fatal_signal_pending(current))
248 /* Time to isolate some pages for migration */
250 spin_lock_irq(&zone->lru_lock);
251 for (; low_pfn < end_pfn; low_pfn++) {
255 /* give a chance to irqs before checking need_resched() */
256 if (!((low_pfn+1) % SWAP_CLUSTER_MAX)) {
257 spin_unlock_irq(&zone->lru_lock);
260 if (need_resched() || spin_is_contended(&zone->lru_lock)) {
262 spin_unlock_irq(&zone->lru_lock);
264 spin_lock_irq(&zone->lru_lock);
265 if (fatal_signal_pending(current))
268 spin_lock_irq(&zone->lru_lock);
271 * migrate_pfn does not necessarily start aligned to a
272 * pageblock. Ensure that pfn_valid is called when moving
273 * into a new MAX_ORDER_NR_PAGES range in case of large
274 * memory holes within the zone
276 if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
277 if (!pfn_valid(low_pfn)) {
278 low_pfn += MAX_ORDER_NR_PAGES - 1;
283 if (!pfn_valid_within(low_pfn))
288 * Get the page and ensure the page is within the same zone.
289 * See the comment in isolate_freepages about overlapping
290 * nodes. It is deliberate that the new zone lock is not taken
291 * as memory compaction should not move pages between nodes.
293 page = pfn_to_page(low_pfn);
294 if (page_zone(page) != zone)
302 * For async migration, also only scan in MOVABLE blocks. Async
303 * migration is optimistic to see if the minimum amount of work
304 * satisfies the allocation
306 pageblock_nr = low_pfn >> pageblock_order;
307 if (!cc->sync && last_pageblock_nr != pageblock_nr &&
308 !migrate_async_suitable(get_pageblock_migratetype(page))) {
309 low_pfn += pageblock_nr_pages;
310 low_pfn = ALIGN(low_pfn, pageblock_nr_pages) - 1;
311 last_pageblock_nr = pageblock_nr;
319 * PageLRU is set, and lru_lock excludes isolation,
320 * splitting and collapsing (collapsing has already
321 * happened if PageLRU is set).
323 if (PageTransHuge(page)) {
324 low_pfn += (1 << compound_order(page)) - 1;
329 mode |= ISOLATE_ASYNC_MIGRATE;
331 lruvec = mem_cgroup_page_lruvec(page, zone);
333 /* Try isolate the page */
334 if (__isolate_lru_page(page, mode) != 0)
337 VM_BUG_ON(PageTransCompound(page));
339 /* Successfully isolated */
340 del_page_from_lru_list(page, lruvec, page_lru(page));
341 list_add(&page->lru, migratelist);
342 cc->nr_migratepages++;
345 /* Avoid isolating too much */
346 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
352 acct_isolated(zone, cc);
354 spin_unlock_irq(&zone->lru_lock);
356 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
361 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
362 #ifdef CONFIG_COMPACTION
364 /* Returns true if the page is within a block suitable for migration to */
365 static bool suitable_migration_target(struct page *page)
368 int migratetype = get_pageblock_migratetype(page);
370 /* Don't interfere with memory hot-remove or the min_free_kbytes blocks */
371 if (migratetype == MIGRATE_ISOLATE || migratetype == MIGRATE_RESERVE)
374 /* If the page is a large free page, then allow migration */
375 if (PageBuddy(page) && page_order(page) >= pageblock_order)
378 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
379 if (migrate_async_suitable(migratetype))
382 /* Otherwise skip the block */
387 * Based on information in the current compact_control, find blocks
388 * suitable for isolating free pages from and then isolate them.
390 static void isolate_freepages(struct zone *zone,
391 struct compact_control *cc)
394 unsigned long high_pfn, low_pfn, pfn, zone_end_pfn, end_pfn;
396 int nr_freepages = cc->nr_freepages;
397 struct list_head *freelist = &cc->freepages;
400 * Initialise the free scanner. The starting point is where we last
401 * scanned from (or the end of the zone if starting). The low point
402 * is the end of the pageblock the migration scanner is using.
405 low_pfn = cc->migrate_pfn + pageblock_nr_pages;
408 * Take care that if the migration scanner is at the end of the zone
409 * that the free scanner does not accidentally move to the next zone
410 * in the next isolation cycle.
412 high_pfn = min(low_pfn, pfn);
414 zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
417 * Isolate free pages until enough are available to migrate the
418 * pages on cc->migratepages. We stop searching if the migrate
419 * and free page scanners meet or enough free pages are isolated.
421 for (; pfn > low_pfn && cc->nr_migratepages > nr_freepages;
422 pfn -= pageblock_nr_pages) {
423 unsigned long isolated;
429 * Check for overlapping nodes/zones. It's possible on some
430 * configurations to have a setup like
432 * i.e. it's possible that all pages within a zones range of
433 * pages do not belong to a single zone.
435 page = pfn_to_page(pfn);
436 if (page_zone(page) != zone)
439 /* Check the block is suitable for migration */
440 if (!suitable_migration_target(page))
444 * Found a block suitable for isolating free pages from. Now
445 * we disabled interrupts, double check things are ok and
446 * isolate the pages. This is to minimise the time IRQs
450 spin_lock_irqsave(&zone->lock, flags);
451 if (suitable_migration_target(page)) {
452 end_pfn = min(pfn + pageblock_nr_pages, zone_end_pfn);
453 isolated = isolate_freepages_block(pfn, end_pfn,
455 nr_freepages += isolated;
457 spin_unlock_irqrestore(&zone->lock, flags);
460 * Record the highest PFN we isolated pages from. When next
461 * looking for free pages, the search will restart here as
462 * page migration may have returned some pages to the allocator
465 high_pfn = max(high_pfn, pfn);
468 /* split_free_page does not map the pages */
471 cc->free_pfn = high_pfn;
472 cc->nr_freepages = nr_freepages;
476 * This is a migrate-callback that "allocates" freepages by taking pages
477 * from the isolated freelists in the block we are migrating to.
479 static struct page *compaction_alloc(struct page *migratepage,
483 struct compact_control *cc = (struct compact_control *)data;
484 struct page *freepage;
486 /* Isolate free pages if necessary */
487 if (list_empty(&cc->freepages)) {
488 isolate_freepages(cc->zone, cc);
490 if (list_empty(&cc->freepages))
494 freepage = list_entry(cc->freepages.next, struct page, lru);
495 list_del(&freepage->lru);
502 * We cannot control nr_migratepages and nr_freepages fully when migration is
503 * running as migrate_pages() has no knowledge of compact_control. When
504 * migration is complete, we count the number of pages on the lists by hand.
506 static void update_nr_listpages(struct compact_control *cc)
508 int nr_migratepages = 0;
509 int nr_freepages = 0;
512 list_for_each_entry(page, &cc->migratepages, lru)
514 list_for_each_entry(page, &cc->freepages, lru)
517 cc->nr_migratepages = nr_migratepages;
518 cc->nr_freepages = nr_freepages;
521 /* possible outcome of isolate_migratepages */
523 ISOLATE_ABORT, /* Abort compaction now */
524 ISOLATE_NONE, /* No pages isolated, continue scanning */
525 ISOLATE_SUCCESS, /* Pages isolated, migrate */
529 * Isolate all pages that can be migrated from the block pointed to by
530 * the migrate scanner within compact_control.
532 static isolate_migrate_t isolate_migratepages(struct zone *zone,
533 struct compact_control *cc)
535 unsigned long low_pfn, end_pfn;
537 /* Do not scan outside zone boundaries */
538 low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
540 /* Only scan within a pageblock boundary */
541 end_pfn = ALIGN(low_pfn + pageblock_nr_pages, pageblock_nr_pages);
543 /* Do not cross the free scanner or scan within a memory hole */
544 if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
545 cc->migrate_pfn = end_pfn;
549 /* Perform the isolation */
550 low_pfn = isolate_migratepages_range(zone, cc, low_pfn, end_pfn);
552 return ISOLATE_ABORT;
554 cc->migrate_pfn = low_pfn;
556 return ISOLATE_SUCCESS;
559 static int compact_finished(struct zone *zone,
560 struct compact_control *cc)
563 unsigned long watermark;
565 if (fatal_signal_pending(current))
566 return COMPACT_PARTIAL;
568 /* Compaction run completes if the migrate and free scanner meet */
569 if (cc->free_pfn <= cc->migrate_pfn)
570 return COMPACT_COMPLETE;
573 * order == -1 is expected when compacting via
574 * /proc/sys/vm/compact_memory
577 return COMPACT_CONTINUE;
579 /* Compaction run is not finished if the watermark is not met */
580 watermark = low_wmark_pages(zone);
581 watermark += (1 << cc->order);
583 if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
584 return COMPACT_CONTINUE;
586 /* Direct compactor: Is a suitable page free? */
587 for (order = cc->order; order < MAX_ORDER; order++) {
588 /* Job done if page is free of the right migratetype */
589 if (!list_empty(&zone->free_area[order].free_list[cc->migratetype]))
590 return COMPACT_PARTIAL;
592 /* Job done if allocation would set block type */
593 if (order >= pageblock_order && zone->free_area[order].nr_free)
594 return COMPACT_PARTIAL;
597 return COMPACT_CONTINUE;
601 * compaction_suitable: Is this suitable to run compaction on this zone now?
603 * COMPACT_SKIPPED - If there are too few free pages for compaction
604 * COMPACT_PARTIAL - If the allocation would succeed without compaction
605 * COMPACT_CONTINUE - If compaction should run now
607 unsigned long compaction_suitable(struct zone *zone, int order)
610 unsigned long watermark;
613 * order == -1 is expected when compacting via
614 * /proc/sys/vm/compact_memory
617 return COMPACT_CONTINUE;
620 * Watermarks for order-0 must be met for compaction. Note the 2UL.
621 * This is because during migration, copies of pages need to be
622 * allocated and for a short time, the footprint is higher
624 watermark = low_wmark_pages(zone) + (2UL << order);
625 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
626 return COMPACT_SKIPPED;
629 * fragmentation index determines if allocation failures are due to
630 * low memory or external fragmentation
632 * index of -1000 implies allocations might succeed depending on
634 * index towards 0 implies failure is due to lack of memory
635 * index towards 1000 implies failure is due to fragmentation
637 * Only compact if a failure would be due to fragmentation.
639 fragindex = fragmentation_index(zone, order);
640 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
641 return COMPACT_SKIPPED;
643 if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
645 return COMPACT_PARTIAL;
647 return COMPACT_CONTINUE;
650 static int compact_zone(struct zone *zone, struct compact_control *cc)
654 ret = compaction_suitable(zone, cc->order);
656 case COMPACT_PARTIAL:
657 case COMPACT_SKIPPED:
658 /* Compaction is likely to fail */
660 case COMPACT_CONTINUE:
661 /* Fall through to compaction */
665 /* Setup to move all movable pages to the end of the zone */
666 cc->migrate_pfn = zone->zone_start_pfn;
667 cc->free_pfn = cc->migrate_pfn + zone->spanned_pages;
668 cc->free_pfn &= ~(pageblock_nr_pages-1);
670 migrate_prep_local();
672 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
673 unsigned long nr_migrate, nr_remaining;
676 switch (isolate_migratepages(zone, cc)) {
678 ret = COMPACT_PARTIAL;
682 case ISOLATE_SUCCESS:
686 nr_migrate = cc->nr_migratepages;
687 err = migrate_pages(&cc->migratepages, compaction_alloc,
688 (unsigned long)cc, false,
689 cc->sync ? MIGRATE_SYNC_LIGHT : MIGRATE_ASYNC);
690 update_nr_listpages(cc);
691 nr_remaining = cc->nr_migratepages;
693 count_vm_event(COMPACTBLOCKS);
694 count_vm_events(COMPACTPAGES, nr_migrate - nr_remaining);
696 count_vm_events(COMPACTPAGEFAILED, nr_remaining);
697 trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
700 /* Release LRU pages not migrated */
702 putback_lru_pages(&cc->migratepages);
703 cc->nr_migratepages = 0;
704 if (err == -ENOMEM) {
705 ret = COMPACT_PARTIAL;
712 /* Release free pages and check accounting */
713 cc->nr_freepages -= release_freepages(&cc->freepages);
714 VM_BUG_ON(cc->nr_freepages != 0);
719 static unsigned long compact_zone_order(struct zone *zone,
720 int order, gfp_t gfp_mask,
723 struct compact_control cc = {
725 .nr_migratepages = 0,
727 .migratetype = allocflags_to_migratetype(gfp_mask),
731 INIT_LIST_HEAD(&cc.freepages);
732 INIT_LIST_HEAD(&cc.migratepages);
734 return compact_zone(zone, &cc);
737 int sysctl_extfrag_threshold = 500;
740 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
741 * @zonelist: The zonelist used for the current allocation
742 * @order: The order of the current allocation
743 * @gfp_mask: The GFP mask of the current allocation
744 * @nodemask: The allowed nodes to allocate from
745 * @sync: Whether migration is synchronous or not
747 * This is the main entry point for direct page compaction.
749 unsigned long try_to_compact_pages(struct zonelist *zonelist,
750 int order, gfp_t gfp_mask, nodemask_t *nodemask,
753 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
754 int may_enter_fs = gfp_mask & __GFP_FS;
755 int may_perform_io = gfp_mask & __GFP_IO;
758 int rc = COMPACT_SKIPPED;
761 * Check whether it is worth even starting compaction. The order check is
762 * made because an assumption is made that the page allocator can satisfy
763 * the "cheaper" orders without taking special steps
765 if (!order || !may_enter_fs || !may_perform_io)
768 count_vm_event(COMPACTSTALL);
770 /* Compact each zone in the list */
771 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
775 status = compact_zone_order(zone, order, gfp_mask, sync);
776 rc = max(status, rc);
778 /* If a normal allocation would succeed, stop compacting */
779 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
787 /* Compact all zones within a node */
788 static int __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
793 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
795 zone = &pgdat->node_zones[zoneid];
796 if (!populated_zone(zone))
799 cc->nr_freepages = 0;
800 cc->nr_migratepages = 0;
802 INIT_LIST_HEAD(&cc->freepages);
803 INIT_LIST_HEAD(&cc->migratepages);
805 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
806 compact_zone(zone, cc);
809 int ok = zone_watermark_ok(zone, cc->order,
810 low_wmark_pages(zone), 0, 0);
811 if (ok && cc->order > zone->compact_order_failed)
812 zone->compact_order_failed = cc->order + 1;
813 /* Currently async compaction is never deferred. */
814 else if (!ok && cc->sync)
815 defer_compaction(zone, cc->order);
818 VM_BUG_ON(!list_empty(&cc->freepages));
819 VM_BUG_ON(!list_empty(&cc->migratepages));
825 int compact_pgdat(pg_data_t *pgdat, int order)
827 struct compact_control cc = {
832 return __compact_pgdat(pgdat, &cc);
835 static int compact_node(int nid)
837 struct compact_control cc = {
842 return __compact_pgdat(NODE_DATA(nid), &cc);
845 /* Compact all nodes in the system */
846 static int compact_nodes(void)
850 /* Flush pending updates to the LRU lists */
853 for_each_online_node(nid)
856 return COMPACT_COMPLETE;
859 /* The written value is actually unused, all memory is compacted */
860 int sysctl_compact_memory;
862 /* This is the entry point for compacting all nodes via /proc/sys/vm */
863 int sysctl_compaction_handler(struct ctl_table *table, int write,
864 void __user *buffer, size_t *length, loff_t *ppos)
867 return compact_nodes();
872 int sysctl_extfrag_handler(struct ctl_table *table, int write,
873 void __user *buffer, size_t *length, loff_t *ppos)
875 proc_dointvec_minmax(table, write, buffer, length, ppos);
880 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
881 ssize_t sysfs_compact_node(struct device *dev,
882 struct device_attribute *attr,
883 const char *buf, size_t count)
887 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
888 /* Flush pending updates to the LRU lists */
896 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
898 int compaction_register_node(struct node *node)
900 return device_create_file(&node->dev, &dev_attr_compact);
903 void compaction_unregister_node(struct node *node)
905 return device_remove_file(&node->dev, &dev_attr_compact);
907 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
909 #endif /* CONFIG_COMPACTION */