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 */
37 /* Account for isolated anon and file pages */
38 unsigned long nr_anon;
39 unsigned long nr_file;
41 unsigned int order; /* order a direct compactor needs */
42 int migratetype; /* MOVABLE, RECLAIMABLE etc */
46 static unsigned long release_freepages(struct list_head *freelist)
48 struct page *page, *next;
49 unsigned long count = 0;
51 list_for_each_entry_safe(page, next, freelist, lru) {
60 /* Isolate free pages onto a private freelist. Must hold zone->lock */
61 static unsigned long isolate_freepages_block(struct zone *zone,
62 unsigned long blockpfn,
63 struct list_head *freelist)
65 unsigned long zone_end_pfn, end_pfn;
66 int nr_scanned = 0, total_isolated = 0;
69 /* Get the last PFN we should scan for free pages at */
70 zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
71 end_pfn = min(blockpfn + pageblock_nr_pages, zone_end_pfn);
73 /* Find the first usable PFN in the block to initialse page cursor */
74 for (; blockpfn < end_pfn; blockpfn++) {
75 if (pfn_valid_within(blockpfn))
78 cursor = pfn_to_page(blockpfn);
80 /* Isolate free pages. This assumes the block is valid */
81 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
83 struct page *page = cursor;
85 if (!pfn_valid_within(blockpfn))
92 /* Found a free page, break it into order-0 pages */
93 isolated = split_free_page(page);
94 total_isolated += isolated;
95 for (i = 0; i < isolated; i++) {
96 list_add(&page->lru, freelist);
100 /* If a page was split, advance to the end of it */
102 blockpfn += isolated - 1;
103 cursor += isolated - 1;
107 trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
108 return total_isolated;
111 /* Returns true if the page is within a block suitable for migration to */
112 static bool suitable_migration_target(struct page *page)
115 int migratetype = get_pageblock_migratetype(page);
117 /* Don't interfere with memory hot-remove or the min_free_kbytes blocks */
118 if (migratetype == MIGRATE_ISOLATE || migratetype == MIGRATE_RESERVE)
121 /* If the page is a large free page, then allow migration */
122 if (PageBuddy(page) && page_order(page) >= pageblock_order)
125 /* If the block is MIGRATE_MOVABLE, allow migration */
126 if (migratetype == MIGRATE_MOVABLE)
129 /* Otherwise skip the block */
134 * Based on information in the current compact_control, find blocks
135 * suitable for isolating free pages from and then isolate them.
137 static void isolate_freepages(struct zone *zone,
138 struct compact_control *cc)
141 unsigned long high_pfn, low_pfn, pfn;
143 int nr_freepages = cc->nr_freepages;
144 struct list_head *freelist = &cc->freepages;
147 low_pfn = cc->migrate_pfn + pageblock_nr_pages;
151 * Isolate free pages until enough are available to migrate the
152 * pages on cc->migratepages. We stop searching if the migrate
153 * and free page scanners meet or enough free pages are isolated.
155 spin_lock_irqsave(&zone->lock, flags);
156 for (; pfn > low_pfn && cc->nr_migratepages > nr_freepages;
157 pfn -= pageblock_nr_pages) {
158 unsigned long isolated;
164 * Check for overlapping nodes/zones. It's possible on some
165 * configurations to have a setup like
167 * i.e. it's possible that all pages within a zones range of
168 * pages do not belong to a single zone.
170 page = pfn_to_page(pfn);
171 if (page_zone(page) != zone)
174 /* Check the block is suitable for migration */
175 if (!suitable_migration_target(page))
178 /* Found a block suitable for isolating free pages from */
179 isolated = isolate_freepages_block(zone, pfn, freelist);
180 nr_freepages += isolated;
183 * Record the highest PFN we isolated pages from. When next
184 * looking for free pages, the search will restart here as
185 * page migration may have returned some pages to the allocator
188 high_pfn = max(high_pfn, pfn);
190 spin_unlock_irqrestore(&zone->lock, flags);
192 /* split_free_page does not map the pages */
193 list_for_each_entry(page, freelist, lru) {
194 arch_alloc_page(page, 0);
195 kernel_map_pages(page, 1, 1);
198 cc->free_pfn = high_pfn;
199 cc->nr_freepages = nr_freepages;
202 /* Update the number of anon and file isolated pages in the zone */
203 static void acct_isolated(struct zone *zone, struct compact_control *cc)
206 unsigned int count[NR_LRU_LISTS] = { 0, };
208 list_for_each_entry(page, &cc->migratepages, lru) {
209 int lru = page_lru_base_type(page);
213 cc->nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
214 cc->nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
215 __mod_zone_page_state(zone, NR_ISOLATED_ANON, cc->nr_anon);
216 __mod_zone_page_state(zone, NR_ISOLATED_FILE, cc->nr_file);
219 /* Similar to reclaim, but different enough that they don't share logic */
220 static bool too_many_isolated(struct zone *zone)
222 unsigned long active, inactive, isolated;
224 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
225 zone_page_state(zone, NR_INACTIVE_ANON);
226 active = zone_page_state(zone, NR_ACTIVE_FILE) +
227 zone_page_state(zone, NR_ACTIVE_ANON);
228 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
229 zone_page_state(zone, NR_ISOLATED_ANON);
231 return isolated > (inactive + active) / 2;
235 * Isolate all pages that can be migrated from the block pointed to by
236 * the migrate scanner within compact_control.
238 static unsigned long isolate_migratepages(struct zone *zone,
239 struct compact_control *cc)
241 unsigned long low_pfn, end_pfn;
242 unsigned long nr_scanned = 0, nr_isolated = 0;
243 struct list_head *migratelist = &cc->migratepages;
245 /* Do not scan outside zone boundaries */
246 low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
248 /* Only scan within a pageblock boundary */
249 end_pfn = ALIGN(low_pfn + pageblock_nr_pages, pageblock_nr_pages);
251 /* Do not cross the free scanner or scan within a memory hole */
252 if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
253 cc->migrate_pfn = end_pfn;
258 * Ensure that there are not too many pages isolated from the LRU
259 * list by either parallel reclaimers or compaction. If there are,
260 * delay for some time until fewer pages are isolated
262 while (unlikely(too_many_isolated(zone))) {
263 congestion_wait(BLK_RW_ASYNC, HZ/10);
265 if (fatal_signal_pending(current))
269 /* Time to isolate some pages for migration */
270 spin_lock_irq(&zone->lru_lock);
271 for (; low_pfn < end_pfn; low_pfn++) {
273 if (!pfn_valid_within(low_pfn))
277 /* Get the page and skip if free */
278 page = pfn_to_page(low_pfn);
282 /* Try isolate the page */
283 if (__isolate_lru_page(page, ISOLATE_BOTH, 0) != 0)
286 /* Successfully isolated */
287 del_page_from_lru_list(zone, page, page_lru(page));
288 list_add(&page->lru, migratelist);
289 cc->nr_migratepages++;
292 /* Avoid isolating too much */
293 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
297 acct_isolated(zone, cc);
299 spin_unlock_irq(&zone->lru_lock);
300 cc->migrate_pfn = low_pfn;
302 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
304 return cc->nr_migratepages;
308 * This is a migrate-callback that "allocates" freepages by taking pages
309 * from the isolated freelists in the block we are migrating to.
311 static struct page *compaction_alloc(struct page *migratepage,
315 struct compact_control *cc = (struct compact_control *)data;
316 struct page *freepage;
318 /* Isolate free pages if necessary */
319 if (list_empty(&cc->freepages)) {
320 isolate_freepages(cc->zone, cc);
322 if (list_empty(&cc->freepages))
326 freepage = list_entry(cc->freepages.next, struct page, lru);
327 list_del(&freepage->lru);
334 * We cannot control nr_migratepages and nr_freepages fully when migration is
335 * running as migrate_pages() has no knowledge of compact_control. When
336 * migration is complete, we count the number of pages on the lists by hand.
338 static void update_nr_listpages(struct compact_control *cc)
340 int nr_migratepages = 0;
341 int nr_freepages = 0;
344 list_for_each_entry(page, &cc->migratepages, lru)
346 list_for_each_entry(page, &cc->freepages, lru)
349 cc->nr_migratepages = nr_migratepages;
350 cc->nr_freepages = nr_freepages;
353 static int compact_finished(struct zone *zone,
354 struct compact_control *cc)
357 unsigned long watermark = low_wmark_pages(zone) + (1 << cc->order);
359 if (fatal_signal_pending(current))
360 return COMPACT_PARTIAL;
362 /* Compaction run completes if the migrate and free scanner meet */
363 if (cc->free_pfn <= cc->migrate_pfn)
364 return COMPACT_COMPLETE;
366 /* Compaction run is not finished if the watermark is not met */
367 if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
368 return COMPACT_CONTINUE;
371 return COMPACT_CONTINUE;
373 /* Direct compactor: Is a suitable page free? */
374 for (order = cc->order; order < MAX_ORDER; order++) {
375 /* Job done if page is free of the right migratetype */
376 if (!list_empty(&zone->free_area[order].free_list[cc->migratetype]))
377 return COMPACT_PARTIAL;
379 /* Job done if allocation would set block type */
380 if (order >= pageblock_order && zone->free_area[order].nr_free)
381 return COMPACT_PARTIAL;
384 return COMPACT_CONTINUE;
388 * compaction_suitable: Is this suitable to run compaction on this zone now?
390 * COMPACT_SKIPPED - If there are too few free pages for compaction
391 * COMPACT_PARTIAL - If the allocation would succeed without compaction
392 * COMPACT_CONTINUE - If compaction should run now
394 unsigned long compaction_suitable(struct zone *zone, int order)
397 unsigned long watermark;
400 * Watermarks for order-0 must be met for compaction. Note the 2UL.
401 * This is because during migration, copies of pages need to be
402 * allocated and for a short time, the footprint is higher
404 watermark = low_wmark_pages(zone) + (2UL << order);
405 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
406 return COMPACT_SKIPPED;
409 * fragmentation index determines if allocation failures are due to
410 * low memory or external fragmentation
412 * index of -1 implies allocations might succeed dependingon watermarks
413 * index towards 0 implies failure is due to lack of memory
414 * index towards 1000 implies failure is due to fragmentation
416 * Only compact if a failure would be due to fragmentation.
418 fragindex = fragmentation_index(zone, order);
419 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
420 return COMPACT_SKIPPED;
422 if (fragindex == -1 && zone_watermark_ok(zone, order, watermark, 0, 0))
423 return COMPACT_PARTIAL;
425 return COMPACT_CONTINUE;
428 static int compact_zone(struct zone *zone, struct compact_control *cc)
432 ret = compaction_suitable(zone, cc->order);
434 case COMPACT_PARTIAL:
435 case COMPACT_SKIPPED:
436 /* Compaction is likely to fail */
438 case COMPACT_CONTINUE:
439 /* Fall through to compaction */
443 /* Setup to move all movable pages to the end of the zone */
444 cc->migrate_pfn = zone->zone_start_pfn;
445 cc->free_pfn = cc->migrate_pfn + zone->spanned_pages;
446 cc->free_pfn &= ~(pageblock_nr_pages-1);
448 migrate_prep_local();
450 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
451 unsigned long nr_migrate, nr_remaining;
453 if (!isolate_migratepages(zone, cc))
456 nr_migrate = cc->nr_migratepages;
457 migrate_pages(&cc->migratepages, compaction_alloc,
458 (unsigned long)cc, 0);
459 update_nr_listpages(cc);
460 nr_remaining = cc->nr_migratepages;
462 count_vm_event(COMPACTBLOCKS);
463 count_vm_events(COMPACTPAGES, nr_migrate - nr_remaining);
465 count_vm_events(COMPACTPAGEFAILED, nr_remaining);
466 trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
469 /* Release LRU pages not migrated */
470 if (!list_empty(&cc->migratepages)) {
471 putback_lru_pages(&cc->migratepages);
472 cc->nr_migratepages = 0;
477 /* Release free pages and check accounting */
478 cc->nr_freepages -= release_freepages(&cc->freepages);
479 VM_BUG_ON(cc->nr_freepages != 0);
484 unsigned long compact_zone_order(struct zone *zone,
485 int order, gfp_t gfp_mask)
487 struct compact_control cc = {
489 .nr_migratepages = 0,
491 .migratetype = allocflags_to_migratetype(gfp_mask),
494 INIT_LIST_HEAD(&cc.freepages);
495 INIT_LIST_HEAD(&cc.migratepages);
497 return compact_zone(zone, &cc);
500 int sysctl_extfrag_threshold = 500;
503 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
504 * @zonelist: The zonelist used for the current allocation
505 * @order: The order of the current allocation
506 * @gfp_mask: The GFP mask of the current allocation
507 * @nodemask: The allowed nodes to allocate from
509 * This is the main entry point for direct page compaction.
511 unsigned long try_to_compact_pages(struct zonelist *zonelist,
512 int order, gfp_t gfp_mask, nodemask_t *nodemask)
514 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
515 int may_enter_fs = gfp_mask & __GFP_FS;
516 int may_perform_io = gfp_mask & __GFP_IO;
519 int rc = COMPACT_SKIPPED;
522 * Check whether it is worth even starting compaction. The order check is
523 * made because an assumption is made that the page allocator can satisfy
524 * the "cheaper" orders without taking special steps
526 if (order <= PAGE_ALLOC_COSTLY_ORDER || !may_enter_fs || !may_perform_io)
529 count_vm_event(COMPACTSTALL);
531 /* Compact each zone in the list */
532 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
536 status = compact_zone_order(zone, order, gfp_mask);
537 rc = max(status, rc);
539 /* If a normal allocation would succeed, stop compacting */
540 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
548 /* Compact all zones within a node */
549 static int compact_node(int nid)
555 if (nid < 0 || nid >= nr_node_ids || !node_online(nid))
557 pgdat = NODE_DATA(nid);
559 /* Flush pending updates to the LRU lists */
562 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
563 struct compact_control cc = {
565 .nr_migratepages = 0,
569 zone = &pgdat->node_zones[zoneid];
570 if (!populated_zone(zone))
574 INIT_LIST_HEAD(&cc.freepages);
575 INIT_LIST_HEAD(&cc.migratepages);
577 compact_zone(zone, &cc);
579 VM_BUG_ON(!list_empty(&cc.freepages));
580 VM_BUG_ON(!list_empty(&cc.migratepages));
586 /* Compact all nodes in the system */
587 static int compact_nodes(void)
591 for_each_online_node(nid)
594 return COMPACT_COMPLETE;
597 /* The written value is actually unused, all memory is compacted */
598 int sysctl_compact_memory;
600 /* This is the entry point for compacting all nodes via /proc/sys/vm */
601 int sysctl_compaction_handler(struct ctl_table *table, int write,
602 void __user *buffer, size_t *length, loff_t *ppos)
605 return compact_nodes();
610 int sysctl_extfrag_handler(struct ctl_table *table, int write,
611 void __user *buffer, size_t *length, loff_t *ppos)
613 proc_dointvec_minmax(table, write, buffer, length, ppos);
618 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
619 ssize_t sysfs_compact_node(struct sys_device *dev,
620 struct sysdev_attribute *attr,
621 const char *buf, size_t count)
623 compact_node(dev->id);
627 static SYSDEV_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
629 int compaction_register_node(struct node *node)
631 return sysdev_create_file(&node->sysdev, &attr_compact);
634 void compaction_unregister_node(struct node *node)
636 return sysdev_remove_file(&node->sysdev, &attr_compact);
638 #endif /* CONFIG_SYSFS && CONFIG_NUMA */