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 int order; /* order a direct compactor needs */
39 int migratetype; /* MOVABLE, RECLAIMABLE etc */
43 static unsigned long release_freepages(struct list_head *freelist)
45 struct page *page, *next;
46 unsigned long count = 0;
48 list_for_each_entry_safe(page, next, freelist, lru) {
57 /* Isolate free pages onto a private freelist. Must hold zone->lock */
58 static unsigned long isolate_freepages_block(struct zone *zone,
59 unsigned long blockpfn,
60 struct list_head *freelist)
62 unsigned long zone_end_pfn, end_pfn;
63 int nr_scanned = 0, total_isolated = 0;
66 /* Get the last PFN we should scan for free pages at */
67 zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
68 end_pfn = min(blockpfn + pageblock_nr_pages, zone_end_pfn);
70 /* Find the first usable PFN in the block to initialse page cursor */
71 for (; blockpfn < end_pfn; blockpfn++) {
72 if (pfn_valid_within(blockpfn))
75 cursor = pfn_to_page(blockpfn);
77 /* Isolate free pages. This assumes the block is valid */
78 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
80 struct page *page = cursor;
82 if (!pfn_valid_within(blockpfn))
89 /* Found a free page, break it into order-0 pages */
90 isolated = split_free_page(page);
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;
108 /* Returns true if the page is within a block suitable for migration to */
109 static bool suitable_migration_target(struct page *page)
112 int migratetype = get_pageblock_migratetype(page);
114 /* Don't interfere with memory hot-remove or the min_free_kbytes blocks */
115 if (migratetype == MIGRATE_ISOLATE || migratetype == MIGRATE_RESERVE)
118 /* If the page is a large free page, then allow migration */
119 if (PageBuddy(page) && page_order(page) >= pageblock_order)
122 /* If the block is MIGRATE_MOVABLE, allow migration */
123 if (migratetype == MIGRATE_MOVABLE)
126 /* Otherwise skip the block */
131 * Based on information in the current compact_control, find blocks
132 * suitable for isolating free pages from and then isolate them.
134 static void isolate_freepages(struct zone *zone,
135 struct compact_control *cc)
138 unsigned long high_pfn, low_pfn, pfn;
140 int nr_freepages = cc->nr_freepages;
141 struct list_head *freelist = &cc->freepages;
144 * Initialise the free scanner. The starting point is where we last
145 * scanned from (or the end of the zone if starting). The low point
146 * is the end of the pageblock the migration scanner is using.
149 low_pfn = cc->migrate_pfn + pageblock_nr_pages;
152 * Take care that if the migration scanner is at the end of the zone
153 * that the free scanner does not accidentally move to the next zone
154 * in the next isolation cycle.
156 high_pfn = min(low_pfn, pfn);
159 * Isolate free pages until enough are available to migrate the
160 * pages on cc->migratepages. We stop searching if the migrate
161 * and free page scanners meet or enough free pages are isolated.
163 for (; pfn > low_pfn && cc->nr_migratepages > nr_freepages;
164 pfn -= pageblock_nr_pages) {
165 unsigned long isolated;
171 * Check for overlapping nodes/zones. It's possible on some
172 * configurations to have a setup like
174 * i.e. it's possible that all pages within a zones range of
175 * pages do not belong to a single zone.
177 page = pfn_to_page(pfn);
178 if (page_zone(page) != zone)
181 /* Check the block is suitable for migration */
182 if (!suitable_migration_target(page))
186 * Found a block suitable for isolating free pages from. Now
187 * we disabled interrupts, double check things are ok and
188 * isolate the pages. This is to minimise the time IRQs
192 spin_lock_irqsave(&zone->lock, flags);
193 if (suitable_migration_target(page)) {
194 isolated = isolate_freepages_block(zone, pfn, freelist);
195 nr_freepages += isolated;
197 spin_unlock_irqrestore(&zone->lock, flags);
200 * Record the highest PFN we isolated pages from. When next
201 * looking for free pages, the search will restart here as
202 * page migration may have returned some pages to the allocator
205 high_pfn = max(high_pfn, pfn);
208 /* split_free_page does not map the pages */
209 list_for_each_entry(page, freelist, lru) {
210 arch_alloc_page(page, 0);
211 kernel_map_pages(page, 1, 1);
214 cc->free_pfn = high_pfn;
215 cc->nr_freepages = nr_freepages;
218 /* Update the number of anon and file isolated pages in the zone */
219 static void acct_isolated(struct zone *zone, struct compact_control *cc)
222 unsigned int count[2] = { 0, };
224 list_for_each_entry(page, &cc->migratepages, lru)
225 count[!!page_is_file_cache(page)]++;
227 __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
228 __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
231 /* Similar to reclaim, but different enough that they don't share logic */
232 static bool too_many_isolated(struct zone *zone)
234 unsigned long active, inactive, isolated;
236 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
237 zone_page_state(zone, NR_INACTIVE_ANON);
238 active = zone_page_state(zone, NR_ACTIVE_FILE) +
239 zone_page_state(zone, NR_ACTIVE_ANON);
240 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
241 zone_page_state(zone, NR_ISOLATED_ANON);
243 return isolated > (inactive + active) / 2;
246 /* possible outcome of isolate_migratepages */
248 ISOLATE_ABORT, /* Abort compaction now */
249 ISOLATE_NONE, /* No pages isolated, continue scanning */
250 ISOLATE_SUCCESS, /* Pages isolated, migrate */
254 * Isolate all pages that can be migrated from the block pointed to by
255 * the migrate scanner within compact_control.
257 static isolate_migrate_t isolate_migratepages(struct zone *zone,
258 struct compact_control *cc)
260 unsigned long low_pfn, end_pfn;
261 unsigned long last_pageblock_nr = 0, pageblock_nr;
262 unsigned long nr_scanned = 0, nr_isolated = 0;
263 struct list_head *migratelist = &cc->migratepages;
264 isolate_mode_t mode = ISOLATE_ACTIVE|ISOLATE_INACTIVE;
266 /* Do not scan outside zone boundaries */
267 low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
269 /* Only scan within a pageblock boundary */
270 end_pfn = ALIGN(low_pfn + pageblock_nr_pages, pageblock_nr_pages);
272 /* Do not cross the free scanner or scan within a memory hole */
273 if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
274 cc->migrate_pfn = end_pfn;
279 * Ensure that there are not too many pages isolated from the LRU
280 * list by either parallel reclaimers or compaction. If there are,
281 * delay for some time until fewer pages are isolated
283 while (unlikely(too_many_isolated(zone))) {
284 /* async migration should just abort */
286 return ISOLATE_ABORT;
288 congestion_wait(BLK_RW_ASYNC, HZ/10);
290 if (fatal_signal_pending(current))
291 return ISOLATE_ABORT;
294 /* Time to isolate some pages for migration */
296 spin_lock_irq(&zone->lru_lock);
297 for (; low_pfn < end_pfn; low_pfn++) {
301 /* give a chance to irqs before checking need_resched() */
302 if (!((low_pfn+1) % SWAP_CLUSTER_MAX)) {
303 spin_unlock_irq(&zone->lru_lock);
306 if (need_resched() || spin_is_contended(&zone->lru_lock)) {
308 spin_unlock_irq(&zone->lru_lock);
310 spin_lock_irq(&zone->lru_lock);
311 if (fatal_signal_pending(current))
314 spin_lock_irq(&zone->lru_lock);
317 * migrate_pfn does not necessarily start aligned to a
318 * pageblock. Ensure that pfn_valid is called when moving
319 * into a new MAX_ORDER_NR_PAGES range in case of large
320 * memory holes within the zone
322 if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
323 if (!pfn_valid(low_pfn)) {
324 low_pfn += MAX_ORDER_NR_PAGES - 1;
329 if (!pfn_valid_within(low_pfn))
334 * Get the page and ensure the page is within the same zone.
335 * See the comment in isolate_freepages about overlapping
336 * nodes. It is deliberate that the new zone lock is not taken
337 * as memory compaction should not move pages between nodes.
339 page = pfn_to_page(low_pfn);
340 if (page_zone(page) != zone)
348 * For async migration, also only scan in MOVABLE blocks. Async
349 * migration is optimistic to see if the minimum amount of work
350 * satisfies the allocation
352 pageblock_nr = low_pfn >> pageblock_order;
353 if (!cc->sync && last_pageblock_nr != pageblock_nr &&
354 get_pageblock_migratetype(page) != MIGRATE_MOVABLE) {
355 low_pfn += pageblock_nr_pages;
356 low_pfn = ALIGN(low_pfn, pageblock_nr_pages) - 1;
357 last_pageblock_nr = pageblock_nr;
365 * PageLRU is set, and lru_lock excludes isolation,
366 * splitting and collapsing (collapsing has already
367 * happened if PageLRU is set).
369 if (PageTransHuge(page)) {
370 low_pfn += (1 << compound_order(page)) - 1;
375 mode |= ISOLATE_ASYNC_MIGRATE;
377 /* Try isolate the page */
378 if (__isolate_lru_page(page, mode, 0) != 0)
381 VM_BUG_ON(PageTransCompound(page));
383 /* Successfully isolated */
384 del_page_from_lru_list(zone, page, page_lru(page));
385 list_add(&page->lru, migratelist);
386 cc->nr_migratepages++;
389 /* Avoid isolating too much */
390 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
396 acct_isolated(zone, cc);
398 spin_unlock_irq(&zone->lru_lock);
399 cc->migrate_pfn = low_pfn;
401 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
403 return ISOLATE_SUCCESS;
407 * This is a migrate-callback that "allocates" freepages by taking pages
408 * from the isolated freelists in the block we are migrating to.
410 static struct page *compaction_alloc(struct page *migratepage,
414 struct compact_control *cc = (struct compact_control *)data;
415 struct page *freepage;
417 /* Isolate free pages if necessary */
418 if (list_empty(&cc->freepages)) {
419 isolate_freepages(cc->zone, cc);
421 if (list_empty(&cc->freepages))
425 freepage = list_entry(cc->freepages.next, struct page, lru);
426 list_del(&freepage->lru);
433 * We cannot control nr_migratepages and nr_freepages fully when migration is
434 * running as migrate_pages() has no knowledge of compact_control. When
435 * migration is complete, we count the number of pages on the lists by hand.
437 static void update_nr_listpages(struct compact_control *cc)
439 int nr_migratepages = 0;
440 int nr_freepages = 0;
443 list_for_each_entry(page, &cc->migratepages, lru)
445 list_for_each_entry(page, &cc->freepages, lru)
448 cc->nr_migratepages = nr_migratepages;
449 cc->nr_freepages = nr_freepages;
452 static int compact_finished(struct zone *zone,
453 struct compact_control *cc)
456 unsigned long watermark;
458 if (fatal_signal_pending(current))
459 return COMPACT_PARTIAL;
461 /* Compaction run completes if the migrate and free scanner meet */
462 if (cc->free_pfn <= cc->migrate_pfn)
463 return COMPACT_COMPLETE;
466 * order == -1 is expected when compacting via
467 * /proc/sys/vm/compact_memory
470 return COMPACT_CONTINUE;
472 /* Compaction run is not finished if the watermark is not met */
473 watermark = low_wmark_pages(zone);
474 watermark += (1 << cc->order);
476 if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
477 return COMPACT_CONTINUE;
479 /* Direct compactor: Is a suitable page free? */
480 for (order = cc->order; order < MAX_ORDER; order++) {
481 /* Job done if page is free of the right migratetype */
482 if (!list_empty(&zone->free_area[order].free_list[cc->migratetype]))
483 return COMPACT_PARTIAL;
485 /* Job done if allocation would set block type */
486 if (order >= pageblock_order && zone->free_area[order].nr_free)
487 return COMPACT_PARTIAL;
490 return COMPACT_CONTINUE;
494 * compaction_suitable: Is this suitable to run compaction on this zone now?
496 * COMPACT_SKIPPED - If there are too few free pages for compaction
497 * COMPACT_PARTIAL - If the allocation would succeed without compaction
498 * COMPACT_CONTINUE - If compaction should run now
500 unsigned long compaction_suitable(struct zone *zone, int order)
503 unsigned long watermark;
506 * order == -1 is expected when compacting via
507 * /proc/sys/vm/compact_memory
510 return COMPACT_CONTINUE;
513 * Watermarks for order-0 must be met for compaction. Note the 2UL.
514 * This is because during migration, copies of pages need to be
515 * allocated and for a short time, the footprint is higher
517 watermark = low_wmark_pages(zone) + (2UL << order);
518 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
519 return COMPACT_SKIPPED;
522 * fragmentation index determines if allocation failures are due to
523 * low memory or external fragmentation
525 * index of -1000 implies allocations might succeed depending on
527 * index towards 0 implies failure is due to lack of memory
528 * index towards 1000 implies failure is due to fragmentation
530 * Only compact if a failure would be due to fragmentation.
532 fragindex = fragmentation_index(zone, order);
533 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
534 return COMPACT_SKIPPED;
536 if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
538 return COMPACT_PARTIAL;
540 return COMPACT_CONTINUE;
543 static int compact_zone(struct zone *zone, struct compact_control *cc)
547 ret = compaction_suitable(zone, cc->order);
549 case COMPACT_PARTIAL:
550 case COMPACT_SKIPPED:
551 /* Compaction is likely to fail */
553 case COMPACT_CONTINUE:
554 /* Fall through to compaction */
558 /* Setup to move all movable pages to the end of the zone */
559 cc->migrate_pfn = zone->zone_start_pfn;
560 cc->free_pfn = cc->migrate_pfn + zone->spanned_pages;
561 cc->free_pfn &= ~(pageblock_nr_pages-1);
563 migrate_prep_local();
565 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
566 unsigned long nr_migrate, nr_remaining;
569 switch (isolate_migratepages(zone, cc)) {
571 ret = COMPACT_PARTIAL;
575 case ISOLATE_SUCCESS:
579 nr_migrate = cc->nr_migratepages;
580 err = migrate_pages(&cc->migratepages, compaction_alloc,
581 (unsigned long)cc, false,
582 cc->sync ? MIGRATE_SYNC_LIGHT : MIGRATE_ASYNC);
583 update_nr_listpages(cc);
584 nr_remaining = cc->nr_migratepages;
586 count_vm_event(COMPACTBLOCKS);
587 count_vm_events(COMPACTPAGES, nr_migrate - nr_remaining);
589 count_vm_events(COMPACTPAGEFAILED, nr_remaining);
590 trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
593 /* Release LRU pages not migrated */
595 putback_lru_pages(&cc->migratepages);
596 cc->nr_migratepages = 0;
602 /* Release free pages and check accounting */
603 cc->nr_freepages -= release_freepages(&cc->freepages);
604 VM_BUG_ON(cc->nr_freepages != 0);
609 static unsigned long compact_zone_order(struct zone *zone,
610 int order, gfp_t gfp_mask,
613 struct compact_control cc = {
615 .nr_migratepages = 0,
617 .migratetype = allocflags_to_migratetype(gfp_mask),
621 INIT_LIST_HEAD(&cc.freepages);
622 INIT_LIST_HEAD(&cc.migratepages);
624 return compact_zone(zone, &cc);
627 int sysctl_extfrag_threshold = 500;
630 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
631 * @zonelist: The zonelist used for the current allocation
632 * @order: The order of the current allocation
633 * @gfp_mask: The GFP mask of the current allocation
634 * @nodemask: The allowed nodes to allocate from
635 * @sync: Whether migration is synchronous or not
637 * This is the main entry point for direct page compaction.
639 unsigned long try_to_compact_pages(struct zonelist *zonelist,
640 int order, gfp_t gfp_mask, nodemask_t *nodemask,
643 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
644 int may_enter_fs = gfp_mask & __GFP_FS;
645 int may_perform_io = gfp_mask & __GFP_IO;
648 int rc = COMPACT_SKIPPED;
651 * Check whether it is worth even starting compaction. The order check is
652 * made because an assumption is made that the page allocator can satisfy
653 * the "cheaper" orders without taking special steps
655 if (!order || !may_enter_fs || !may_perform_io)
658 count_vm_event(COMPACTSTALL);
660 /* Compact each zone in the list */
661 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
665 status = compact_zone_order(zone, order, gfp_mask, sync);
666 rc = max(status, rc);
668 /* If a normal allocation would succeed, stop compacting */
669 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
677 /* Compact all zones within a node */
678 static int __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
683 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
685 zone = &pgdat->node_zones[zoneid];
686 if (!populated_zone(zone))
689 cc->nr_freepages = 0;
690 cc->nr_migratepages = 0;
692 INIT_LIST_HEAD(&cc->freepages);
693 INIT_LIST_HEAD(&cc->migratepages);
695 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
696 compact_zone(zone, cc);
699 int ok = zone_watermark_ok(zone, cc->order,
700 low_wmark_pages(zone), 0, 0);
701 if (ok && cc->order > zone->compact_order_failed)
702 zone->compact_order_failed = cc->order + 1;
703 /* Currently async compaction is never deferred. */
704 else if (!ok && cc->sync)
705 defer_compaction(zone, cc->order);
708 VM_BUG_ON(!list_empty(&cc->freepages));
709 VM_BUG_ON(!list_empty(&cc->migratepages));
715 int compact_pgdat(pg_data_t *pgdat, int order)
717 struct compact_control cc = {
722 return __compact_pgdat(pgdat, &cc);
725 static int compact_node(int nid)
727 struct compact_control cc = {
732 return __compact_pgdat(NODE_DATA(nid), &cc);
735 /* Compact all nodes in the system */
736 static int compact_nodes(void)
740 /* Flush pending updates to the LRU lists */
743 for_each_online_node(nid)
746 return COMPACT_COMPLETE;
749 /* The written value is actually unused, all memory is compacted */
750 int sysctl_compact_memory;
752 /* This is the entry point for compacting all nodes via /proc/sys/vm */
753 int sysctl_compaction_handler(struct ctl_table *table, int write,
754 void __user *buffer, size_t *length, loff_t *ppos)
757 return compact_nodes();
762 int sysctl_extfrag_handler(struct ctl_table *table, int write,
763 void __user *buffer, size_t *length, loff_t *ppos)
765 proc_dointvec_minmax(table, write, buffer, length, ppos);
770 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
771 ssize_t sysfs_compact_node(struct device *dev,
772 struct device_attribute *attr,
773 const char *buf, size_t count)
777 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
778 /* Flush pending updates to the LRU lists */
786 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
788 int compaction_register_node(struct node *node)
790 return device_create_file(&node->dev, &dev_attr_compact);
793 void compaction_unregister_node(struct node *node)
795 return device_remove_file(&node->dev, &dev_attr_compact);
797 #endif /* CONFIG_SYSFS && CONFIG_NUMA */