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 * Initialise the free scanner. The starting point is where we last
149 * scanned from (or the end of the zone if starting). The low point
150 * is the end of the pageblock the migration scanner is using.
153 low_pfn = cc->migrate_pfn + pageblock_nr_pages;
156 * Take care that if the migration scanner is at the end of the zone
157 * that the free scanner does not accidentally move to the next zone
158 * in the next isolation cycle.
160 high_pfn = min(low_pfn, pfn);
163 * Isolate free pages until enough are available to migrate the
164 * pages on cc->migratepages. We stop searching if the migrate
165 * and free page scanners meet or enough free pages are isolated.
167 for (; pfn > low_pfn && cc->nr_migratepages > nr_freepages;
168 pfn -= pageblock_nr_pages) {
169 unsigned long isolated;
175 * Check for overlapping nodes/zones. It's possible on some
176 * configurations to have a setup like
178 * i.e. it's possible that all pages within a zones range of
179 * pages do not belong to a single zone.
181 page = pfn_to_page(pfn);
182 if (page_zone(page) != zone)
185 /* Check the block is suitable for migration */
186 if (!suitable_migration_target(page))
190 * Found a block suitable for isolating free pages from. Now
191 * we disabled interrupts, double check things are ok and
192 * isolate the pages. This is to minimise the time IRQs
196 spin_lock_irqsave(&zone->lock, flags);
197 if (suitable_migration_target(page)) {
198 isolated = isolate_freepages_block(zone, pfn, freelist);
199 nr_freepages += isolated;
201 spin_unlock_irqrestore(&zone->lock, flags);
204 * Record the highest PFN we isolated pages from. When next
205 * looking for free pages, the search will restart here as
206 * page migration may have returned some pages to the allocator
209 high_pfn = max(high_pfn, pfn);
212 /* split_free_page does not map the pages */
213 list_for_each_entry(page, freelist, lru) {
214 arch_alloc_page(page, 0);
215 kernel_map_pages(page, 1, 1);
218 cc->free_pfn = high_pfn;
219 cc->nr_freepages = nr_freepages;
222 /* Update the number of anon and file isolated pages in the zone */
223 static void acct_isolated(struct zone *zone, struct compact_control *cc)
226 unsigned int count[NR_LRU_LISTS] = { 0, };
228 list_for_each_entry(page, &cc->migratepages, lru) {
229 int lru = page_lru_base_type(page);
233 cc->nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
234 cc->nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
235 __mod_zone_page_state(zone, NR_ISOLATED_ANON, cc->nr_anon);
236 __mod_zone_page_state(zone, NR_ISOLATED_FILE, cc->nr_file);
239 /* Similar to reclaim, but different enough that they don't share logic */
240 static bool too_many_isolated(struct zone *zone)
242 unsigned long active, inactive, isolated;
244 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
245 zone_page_state(zone, NR_INACTIVE_ANON);
246 active = zone_page_state(zone, NR_ACTIVE_FILE) +
247 zone_page_state(zone, NR_ACTIVE_ANON);
248 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
249 zone_page_state(zone, NR_ISOLATED_ANON);
251 return isolated > (inactive + active) / 2;
255 * Isolate all pages that can be migrated from the block pointed to by
256 * the migrate scanner within compact_control.
258 static unsigned long isolate_migratepages(struct zone *zone,
259 struct compact_control *cc)
261 unsigned long low_pfn, end_pfn;
262 unsigned long last_pageblock_nr = 0, pageblock_nr;
263 unsigned long nr_scanned = 0, nr_isolated = 0;
264 struct list_head *migratelist = &cc->migratepages;
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 congestion_wait(BLK_RW_ASYNC, HZ/10);
286 if (fatal_signal_pending(current))
290 /* Time to isolate some pages for migration */
292 spin_lock_irq(&zone->lru_lock);
293 for (; low_pfn < end_pfn; low_pfn++) {
297 /* give a chance to irqs before checking need_resched() */
298 if (!((low_pfn+1) % SWAP_CLUSTER_MAX)) {
299 spin_unlock_irq(&zone->lru_lock);
302 if (need_resched() || spin_is_contended(&zone->lru_lock)) {
304 spin_unlock_irq(&zone->lru_lock);
306 spin_lock_irq(&zone->lru_lock);
307 if (fatal_signal_pending(current))
310 spin_lock_irq(&zone->lru_lock);
312 if (!pfn_valid_within(low_pfn))
316 /* Get the page and skip if free */
317 page = pfn_to_page(low_pfn);
322 * For async migration, also only scan in MOVABLE blocks. Async
323 * migration is optimistic to see if the minimum amount of work
324 * satisfies the allocation
326 pageblock_nr = low_pfn >> pageblock_order;
327 if (!cc->sync && last_pageblock_nr != pageblock_nr &&
328 get_pageblock_migratetype(page) != MIGRATE_MOVABLE) {
329 low_pfn += pageblock_nr_pages;
330 low_pfn = ALIGN(low_pfn, pageblock_nr_pages) - 1;
331 last_pageblock_nr = pageblock_nr;
339 * PageLRU is set, and lru_lock excludes isolation,
340 * splitting and collapsing (collapsing has already
341 * happened if PageLRU is set).
343 if (PageTransHuge(page)) {
344 low_pfn += (1 << compound_order(page)) - 1;
348 /* Try isolate the page */
349 if (__isolate_lru_page(page, ISOLATE_BOTH, 0) != 0)
352 VM_BUG_ON(PageTransCompound(page));
354 /* Successfully isolated */
355 del_page_from_lru_list(zone, page, page_lru(page));
356 list_add(&page->lru, migratelist);
357 cc->nr_migratepages++;
360 /* Avoid isolating too much */
361 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
365 acct_isolated(zone, cc);
367 spin_unlock_irq(&zone->lru_lock);
368 cc->migrate_pfn = low_pfn;
370 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
372 return cc->nr_migratepages;
376 * This is a migrate-callback that "allocates" freepages by taking pages
377 * from the isolated freelists in the block we are migrating to.
379 static struct page *compaction_alloc(struct page *migratepage,
383 struct compact_control *cc = (struct compact_control *)data;
384 struct page *freepage;
386 /* Isolate free pages if necessary */
387 if (list_empty(&cc->freepages)) {
388 isolate_freepages(cc->zone, cc);
390 if (list_empty(&cc->freepages))
394 freepage = list_entry(cc->freepages.next, struct page, lru);
395 list_del(&freepage->lru);
402 * We cannot control nr_migratepages and nr_freepages fully when migration is
403 * running as migrate_pages() has no knowledge of compact_control. When
404 * migration is complete, we count the number of pages on the lists by hand.
406 static void update_nr_listpages(struct compact_control *cc)
408 int nr_migratepages = 0;
409 int nr_freepages = 0;
412 list_for_each_entry(page, &cc->migratepages, lru)
414 list_for_each_entry(page, &cc->freepages, lru)
417 cc->nr_migratepages = nr_migratepages;
418 cc->nr_freepages = nr_freepages;
421 static int compact_finished(struct zone *zone,
422 struct compact_control *cc)
425 unsigned long watermark;
427 if (fatal_signal_pending(current))
428 return COMPACT_PARTIAL;
430 /* Compaction run completes if the migrate and free scanner meet */
431 if (cc->free_pfn <= cc->migrate_pfn)
432 return COMPACT_COMPLETE;
435 * order == -1 is expected when compacting via
436 * /proc/sys/vm/compact_memory
439 return COMPACT_CONTINUE;
441 /* Compaction run is not finished if the watermark is not met */
442 watermark = low_wmark_pages(zone);
443 watermark += (1 << cc->order);
445 if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
446 return COMPACT_CONTINUE;
448 /* Direct compactor: Is a suitable page free? */
449 for (order = cc->order; order < MAX_ORDER; order++) {
450 /* Job done if page is free of the right migratetype */
451 if (!list_empty(&zone->free_area[order].free_list[cc->migratetype]))
452 return COMPACT_PARTIAL;
454 /* Job done if allocation would set block type */
455 if (order >= pageblock_order && zone->free_area[order].nr_free)
456 return COMPACT_PARTIAL;
459 return COMPACT_CONTINUE;
463 * compaction_suitable: Is this suitable to run compaction on this zone now?
465 * COMPACT_SKIPPED - If there are too few free pages for compaction
466 * COMPACT_PARTIAL - If the allocation would succeed without compaction
467 * COMPACT_CONTINUE - If compaction should run now
469 unsigned long compaction_suitable(struct zone *zone, int order)
472 unsigned long watermark;
475 * order == -1 is expected when compacting via
476 * /proc/sys/vm/compact_memory
479 return COMPACT_CONTINUE;
482 * Watermarks for order-0 must be met for compaction. Note the 2UL.
483 * This is because during migration, copies of pages need to be
484 * allocated and for a short time, the footprint is higher
486 watermark = low_wmark_pages(zone) + (2UL << order);
487 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
488 return COMPACT_SKIPPED;
491 * fragmentation index determines if allocation failures are due to
492 * low memory or external fragmentation
494 * index of -1000 implies allocations might succeed depending on
496 * index towards 0 implies failure is due to lack of memory
497 * index towards 1000 implies failure is due to fragmentation
499 * Only compact if a failure would be due to fragmentation.
501 fragindex = fragmentation_index(zone, order);
502 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
503 return COMPACT_SKIPPED;
505 if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
507 return COMPACT_PARTIAL;
509 return COMPACT_CONTINUE;
512 static int compact_zone(struct zone *zone, struct compact_control *cc)
516 ret = compaction_suitable(zone, cc->order);
518 case COMPACT_PARTIAL:
519 case COMPACT_SKIPPED:
520 /* Compaction is likely to fail */
522 case COMPACT_CONTINUE:
523 /* Fall through to compaction */
527 /* Setup to move all movable pages to the end of the zone */
528 cc->migrate_pfn = zone->zone_start_pfn;
529 cc->free_pfn = cc->migrate_pfn + zone->spanned_pages;
530 cc->free_pfn &= ~(pageblock_nr_pages-1);
532 migrate_prep_local();
534 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
535 unsigned long nr_migrate, nr_remaining;
538 if (!isolate_migratepages(zone, cc))
541 nr_migrate = cc->nr_migratepages;
542 err = migrate_pages(&cc->migratepages, compaction_alloc,
543 (unsigned long)cc, false,
545 update_nr_listpages(cc);
546 nr_remaining = cc->nr_migratepages;
548 count_vm_event(COMPACTBLOCKS);
549 count_vm_events(COMPACTPAGES, nr_migrate - nr_remaining);
551 count_vm_events(COMPACTPAGEFAILED, nr_remaining);
552 trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
555 /* Release LRU pages not migrated */
557 putback_lru_pages(&cc->migratepages);
558 cc->nr_migratepages = 0;
563 /* Release free pages and check accounting */
564 cc->nr_freepages -= release_freepages(&cc->freepages);
565 VM_BUG_ON(cc->nr_freepages != 0);
570 unsigned long compact_zone_order(struct zone *zone,
571 int order, gfp_t gfp_mask,
574 struct compact_control cc = {
576 .nr_migratepages = 0,
578 .migratetype = allocflags_to_migratetype(gfp_mask),
582 INIT_LIST_HEAD(&cc.freepages);
583 INIT_LIST_HEAD(&cc.migratepages);
585 return compact_zone(zone, &cc);
588 int sysctl_extfrag_threshold = 500;
591 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
592 * @zonelist: The zonelist used for the current allocation
593 * @order: The order of the current allocation
594 * @gfp_mask: The GFP mask of the current allocation
595 * @nodemask: The allowed nodes to allocate from
596 * @sync: Whether migration is synchronous or not
598 * This is the main entry point for direct page compaction.
600 unsigned long try_to_compact_pages(struct zonelist *zonelist,
601 int order, gfp_t gfp_mask, nodemask_t *nodemask,
604 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
605 int may_enter_fs = gfp_mask & __GFP_FS;
606 int may_perform_io = gfp_mask & __GFP_IO;
609 int rc = COMPACT_SKIPPED;
612 * Check whether it is worth even starting compaction. The order check is
613 * made because an assumption is made that the page allocator can satisfy
614 * the "cheaper" orders without taking special steps
616 if (!order || !may_enter_fs || !may_perform_io)
619 count_vm_event(COMPACTSTALL);
621 /* Compact each zone in the list */
622 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
626 status = compact_zone_order(zone, order, gfp_mask, sync);
627 rc = max(status, rc);
629 /* If a normal allocation would succeed, stop compacting */
630 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
638 /* Compact all zones within a node */
639 static int compact_node(int nid)
645 if (nid < 0 || nid >= nr_node_ids || !node_online(nid))
647 pgdat = NODE_DATA(nid);
649 /* Flush pending updates to the LRU lists */
652 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
653 struct compact_control cc = {
655 .nr_migratepages = 0,
659 zone = &pgdat->node_zones[zoneid];
660 if (!populated_zone(zone))
664 INIT_LIST_HEAD(&cc.freepages);
665 INIT_LIST_HEAD(&cc.migratepages);
667 compact_zone(zone, &cc);
669 VM_BUG_ON(!list_empty(&cc.freepages));
670 VM_BUG_ON(!list_empty(&cc.migratepages));
676 /* Compact all nodes in the system */
677 static int compact_nodes(void)
681 for_each_online_node(nid)
684 return COMPACT_COMPLETE;
687 /* The written value is actually unused, all memory is compacted */
688 int sysctl_compact_memory;
690 /* This is the entry point for compacting all nodes via /proc/sys/vm */
691 int sysctl_compaction_handler(struct ctl_table *table, int write,
692 void __user *buffer, size_t *length, loff_t *ppos)
695 return compact_nodes();
700 int sysctl_extfrag_handler(struct ctl_table *table, int write,
701 void __user *buffer, size_t *length, loff_t *ppos)
703 proc_dointvec_minmax(table, write, buffer, length, ppos);
708 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
709 ssize_t sysfs_compact_node(struct sys_device *dev,
710 struct sysdev_attribute *attr,
711 const char *buf, size_t count)
713 compact_node(dev->id);
717 static SYSDEV_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
719 int compaction_register_node(struct node *node)
721 return sysdev_create_file(&node->sysdev, &attr_compact);
724 void compaction_unregister_node(struct node *node)
726 return sysdev_remove_file(&node->sysdev, &attr_compact);
728 #endif /* CONFIG_SYSFS && CONFIG_NUMA */