blob: 3a1828541dc0ba4861f0827e1479d4fef5cb31cf [file] [log] [blame]
Mel Gorman748446b2010-05-24 14:32:27 -07001/*
2 * linux/mm/compaction.c
3 *
4 * Memory compaction for the reduction of external fragmentation. Note that
5 * this heavily depends upon page migration to do all the real heavy
6 * lifting
7 *
8 * Copyright IBM Corp. 2007-2010 Mel Gorman <mel@csn.ul.ie>
9 */
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>
Mel Gorman76ab0f52010-05-24 14:32:28 -070015#include <linux/sysctl.h>
Mel Gormaned4a6d72010-05-24 14:32:29 -070016#include <linux/sysfs.h>
Rafael Aquinibf6bddf2012-12-11 16:02:42 -080017#include <linux/balloon_compaction.h>
Minchan Kim194159f2013-02-22 16:33:58 -080018#include <linux/page-isolation.h>
Mel Gorman748446b2010-05-24 14:32:27 -070019#include "internal.h"
20
Minchan Kim010fc292012-12-20 15:05:06 -080021#ifdef CONFIG_COMPACTION
22static inline void count_compact_event(enum vm_event_item item)
23{
24 count_vm_event(item);
25}
26
27static inline void count_compact_events(enum vm_event_item item, long delta)
28{
29 count_vm_events(item, delta);
30}
31#else
32#define count_compact_event(item) do { } while (0)
33#define count_compact_events(item, delta) do { } while (0)
34#endif
35
Michal Nazarewiczff9543f2011-12-29 13:09:50 +010036#if defined CONFIG_COMPACTION || defined CONFIG_CMA
37
Mel Gormanb7aba692011-01-13 15:45:54 -080038#define CREATE_TRACE_POINTS
39#include <trace/events/compaction.h>
40
Mel Gorman748446b2010-05-24 14:32:27 -070041static unsigned long release_freepages(struct list_head *freelist)
42{
43 struct page *page, *next;
44 unsigned long count = 0;
45
46 list_for_each_entry_safe(page, next, freelist, lru) {
47 list_del(&page->lru);
48 __free_page(page);
49 count++;
50 }
51
52 return count;
53}
54
Michal Nazarewiczff9543f2011-12-29 13:09:50 +010055static void map_pages(struct list_head *list)
56{
57 struct page *page;
58
59 list_for_each_entry(page, list, lru) {
60 arch_alloc_page(page, 0);
61 kernel_map_pages(page, 1, 1);
62 }
63}
64
Michal Nazarewicz47118af2011-12-29 13:09:50 +010065static inline bool migrate_async_suitable(int migratetype)
66{
67 return is_migrate_cma(migratetype) || migratetype == MIGRATE_MOVABLE;
68}
69
Mel Gormanbb13ffe2012-10-08 16:32:41 -070070#ifdef CONFIG_COMPACTION
71/* Returns true if the pageblock should be scanned for pages to isolate. */
72static inline bool isolation_suitable(struct compact_control *cc,
73 struct page *page)
74{
75 if (cc->ignore_skip_hint)
76 return true;
77
78 return !get_pageblock_skip(page);
79}
80
81/*
82 * This function is called to clear all cached information on pageblocks that
83 * should be skipped for page isolation when the migrate and free page scanner
84 * meet.
85 */
Mel Gorman62997022012-10-08 16:32:47 -070086static void __reset_isolation_suitable(struct zone *zone)
Mel Gormanbb13ffe2012-10-08 16:32:41 -070087{
88 unsigned long start_pfn = zone->zone_start_pfn;
Cody P Schafer108bcc92013-02-22 16:35:23 -080089 unsigned long end_pfn = zone_end_pfn(zone);
Mel Gormanbb13ffe2012-10-08 16:32:41 -070090 unsigned long pfn;
91
Mel Gormanc89511a2012-10-08 16:32:45 -070092 zone->compact_cached_migrate_pfn = start_pfn;
93 zone->compact_cached_free_pfn = end_pfn;
Mel Gorman62997022012-10-08 16:32:47 -070094 zone->compact_blockskip_flush = false;
Mel Gormanbb13ffe2012-10-08 16:32:41 -070095
96 /* Walk the zone and mark every pageblock as suitable for isolation */
97 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
98 struct page *page;
99
100 cond_resched();
101
102 if (!pfn_valid(pfn))
103 continue;
104
105 page = pfn_to_page(pfn);
106 if (zone != page_zone(page))
107 continue;
108
109 clear_pageblock_skip(page);
110 }
111}
112
Mel Gorman62997022012-10-08 16:32:47 -0700113void reset_isolation_suitable(pg_data_t *pgdat)
114{
115 int zoneid;
116
117 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
118 struct zone *zone = &pgdat->node_zones[zoneid];
119 if (!populated_zone(zone))
120 continue;
121
122 /* Only flush if a full compaction finished recently */
123 if (zone->compact_blockskip_flush)
124 __reset_isolation_suitable(zone);
125 }
126}
127
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700128/*
129 * If no pages were isolated then mark this pageblock to be skipped in the
Mel Gorman62997022012-10-08 16:32:47 -0700130 * future. The information is later cleared by __reset_isolation_suitable().
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700131 */
Mel Gormanc89511a2012-10-08 16:32:45 -0700132static void update_pageblock_skip(struct compact_control *cc,
133 struct page *page, unsigned long nr_isolated,
134 bool migrate_scanner)
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700135{
Mel Gormanc89511a2012-10-08 16:32:45 -0700136 struct zone *zone = cc->zone;
Joonsoo Kim6815bf32013-12-18 17:08:52 -0800137
138 if (cc->ignore_skip_hint)
139 return;
140
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700141 if (!page)
142 return;
143
Mel Gormanc89511a2012-10-08 16:32:45 -0700144 if (!nr_isolated) {
145 unsigned long pfn = page_to_pfn(page);
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700146 set_pageblock_skip(page);
Mel Gormanc89511a2012-10-08 16:32:45 -0700147
148 /* Update where compaction should restart */
149 if (migrate_scanner) {
150 if (!cc->finished_update_migrate &&
151 pfn > zone->compact_cached_migrate_pfn)
152 zone->compact_cached_migrate_pfn = pfn;
153 } else {
154 if (!cc->finished_update_free &&
155 pfn < zone->compact_cached_free_pfn)
156 zone->compact_cached_free_pfn = pfn;
157 }
158 }
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700159}
160#else
161static inline bool isolation_suitable(struct compact_control *cc,
162 struct page *page)
163{
164 return true;
165}
166
Mel Gormanc89511a2012-10-08 16:32:45 -0700167static void update_pageblock_skip(struct compact_control *cc,
168 struct page *page, unsigned long nr_isolated,
169 bool migrate_scanner)
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700170{
171}
172#endif /* CONFIG_COMPACTION */
173
Mel Gorman2a1402a2012-10-08 16:32:33 -0700174static inline bool should_release_lock(spinlock_t *lock)
175{
176 return need_resched() || spin_is_contended(lock);
177}
178
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100179/*
Mel Gormanc67fe372012-08-21 16:16:17 -0700180 * Compaction requires the taking of some coarse locks that are potentially
181 * very heavily contended. Check if the process needs to be scheduled or
182 * if the lock is contended. For async compaction, back out in the event
183 * if contention is severe. For sync compaction, schedule.
184 *
185 * Returns true if the lock is held.
186 * Returns false if the lock is released and compaction should abort
187 */
188static bool compact_checklock_irqsave(spinlock_t *lock, unsigned long *flags,
189 bool locked, struct compact_control *cc)
190{
Mel Gorman2a1402a2012-10-08 16:32:33 -0700191 if (should_release_lock(lock)) {
Mel Gormanc67fe372012-08-21 16:16:17 -0700192 if (locked) {
193 spin_unlock_irqrestore(lock, *flags);
194 locked = false;
195 }
196
197 /* async aborts if taking too long or contended */
198 if (!cc->sync) {
Shaohua Lie64c5232012-10-08 16:32:27 -0700199 cc->contended = true;
Mel Gormanc67fe372012-08-21 16:16:17 -0700200 return false;
201 }
202
203 cond_resched();
Mel Gormanc67fe372012-08-21 16:16:17 -0700204 }
205
206 if (!locked)
207 spin_lock_irqsave(lock, *flags);
208 return true;
209}
210
211static inline bool compact_trylock_irqsave(spinlock_t *lock,
212 unsigned long *flags, struct compact_control *cc)
213{
214 return compact_checklock_irqsave(lock, flags, false, cc);
215}
216
Mel Gormanf40d1e42012-10-08 16:32:36 -0700217/* Returns true if the page is within a block suitable for migration to */
218static bool suitable_migration_target(struct page *page)
219{
Joonsoo Kim7d348b92014-04-07 15:37:03 -0700220 /* If the page is a large free page, then disallow migration */
Mel Gormanf40d1e42012-10-08 16:32:36 -0700221 if (PageBuddy(page) && page_order(page) >= pageblock_order)
Joonsoo Kim7d348b92014-04-07 15:37:03 -0700222 return false;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700223
224 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
Joonsoo Kim7d348b92014-04-07 15:37:03 -0700225 if (migrate_async_suitable(get_pageblock_migratetype(page)))
Mel Gormanf40d1e42012-10-08 16:32:36 -0700226 return true;
227
228 /* Otherwise skip the block */
229 return false;
230}
231
Mel Gormanc67fe372012-08-21 16:16:17 -0700232/*
Jerome Marchand9e4be472013-11-12 15:07:12 -0800233 * Isolate free pages onto a private freelist. If @strict is true, will abort
234 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
235 * (even though it may still end up isolating some pages).
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100236 */
Mel Gormanf40d1e42012-10-08 16:32:36 -0700237static unsigned long isolate_freepages_block(struct compact_control *cc,
238 unsigned long blockpfn,
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100239 unsigned long end_pfn,
240 struct list_head *freelist,
241 bool strict)
Mel Gorman748446b2010-05-24 14:32:27 -0700242{
Mel Gormanb7aba692011-01-13 15:45:54 -0800243 int nr_scanned = 0, total_isolated = 0;
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700244 struct page *cursor, *valid_page = NULL;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700245 unsigned long flags;
246 bool locked = false;
Joonsoo Kim01ead532014-04-07 15:37:04 -0700247 bool checked_pageblock = false;
Mel Gorman748446b2010-05-24 14:32:27 -0700248
Mel Gorman748446b2010-05-24 14:32:27 -0700249 cursor = pfn_to_page(blockpfn);
250
Mel Gormanf40d1e42012-10-08 16:32:36 -0700251 /* Isolate free pages. */
Mel Gorman748446b2010-05-24 14:32:27 -0700252 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
253 int isolated, i;
254 struct page *page = cursor;
255
Mel Gormanb7aba692011-01-13 15:45:54 -0800256 nr_scanned++;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700257 if (!pfn_valid_within(blockpfn))
Laura Abbott2af120b2014-03-10 15:49:44 -0700258 goto isolate_fail;
259
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700260 if (!valid_page)
261 valid_page = page;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700262 if (!PageBuddy(page))
Laura Abbott2af120b2014-03-10 15:49:44 -0700263 goto isolate_fail;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700264
265 /*
266 * The zone lock must be held to isolate freepages.
267 * Unfortunately this is a very coarse lock and can be
268 * heavily contended if there are parallel allocations
269 * or parallel compactions. For async compaction do not
270 * spin on the lock and we acquire the lock as late as
271 * possible.
272 */
273 locked = compact_checklock_irqsave(&cc->zone->lock, &flags,
274 locked, cc);
275 if (!locked)
276 break;
277
278 /* Recheck this is a suitable migration target under lock */
Joonsoo Kim01ead532014-04-07 15:37:04 -0700279 if (!strict && !checked_pageblock) {
280 /*
281 * We need to check suitability of pageblock only once
282 * and this isolate_freepages_block() is called with
283 * pageblock range, so just check once is sufficient.
284 */
285 checked_pageblock = true;
286 if (!suitable_migration_target(page))
287 break;
288 }
Mel Gormanf40d1e42012-10-08 16:32:36 -0700289
290 /* Recheck this is a buddy page under lock */
291 if (!PageBuddy(page))
Laura Abbott2af120b2014-03-10 15:49:44 -0700292 goto isolate_fail;
Mel Gorman748446b2010-05-24 14:32:27 -0700293
294 /* Found a free page, break it into order-0 pages */
295 isolated = split_free_page(page);
296 total_isolated += isolated;
297 for (i = 0; i < isolated; i++) {
298 list_add(&page->lru, freelist);
299 page++;
300 }
301
302 /* If a page was split, advance to the end of it */
303 if (isolated) {
304 blockpfn += isolated - 1;
305 cursor += isolated - 1;
Laura Abbott2af120b2014-03-10 15:49:44 -0700306 continue;
Mel Gorman748446b2010-05-24 14:32:27 -0700307 }
Laura Abbott2af120b2014-03-10 15:49:44 -0700308
309isolate_fail:
310 if (strict)
311 break;
312 else
313 continue;
314
Mel Gorman748446b2010-05-24 14:32:27 -0700315 }
316
Mel Gormanb7aba692011-01-13 15:45:54 -0800317 trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
Mel Gormanf40d1e42012-10-08 16:32:36 -0700318
319 /*
320 * If strict isolation is requested by CMA then check that all the
321 * pages requested were isolated. If there were any failures, 0 is
322 * returned and CMA will fail.
323 */
Laura Abbott2af120b2014-03-10 15:49:44 -0700324 if (strict && blockpfn < end_pfn)
Mel Gormanf40d1e42012-10-08 16:32:36 -0700325 total_isolated = 0;
326
327 if (locked)
328 spin_unlock_irqrestore(&cc->zone->lock, flags);
329
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700330 /* Update the pageblock-skip if the whole pageblock was scanned */
331 if (blockpfn == end_pfn)
Mel Gormanc89511a2012-10-08 16:32:45 -0700332 update_pageblock_skip(cc, valid_page, total_isolated, false);
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700333
Minchan Kim010fc292012-12-20 15:05:06 -0800334 count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
Mel Gorman397487d2012-10-19 12:00:10 +0100335 if (total_isolated)
Minchan Kim010fc292012-12-20 15:05:06 -0800336 count_compact_events(COMPACTISOLATED, total_isolated);
Mel Gorman748446b2010-05-24 14:32:27 -0700337 return total_isolated;
338}
339
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100340/**
341 * isolate_freepages_range() - isolate free pages.
342 * @start_pfn: The first PFN to start isolating.
343 * @end_pfn: The one-past-last PFN.
344 *
345 * Non-free pages, invalid PFNs, or zone boundaries within the
346 * [start_pfn, end_pfn) range are considered errors, cause function to
347 * undo its actions and return zero.
348 *
349 * Otherwise, function returns one-past-the-last PFN of isolated page
350 * (which may be greater then end_pfn if end fell in a middle of
351 * a free page).
352 */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100353unsigned long
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700354isolate_freepages_range(struct compact_control *cc,
355 unsigned long start_pfn, unsigned long end_pfn)
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100356{
Mel Gormanf40d1e42012-10-08 16:32:36 -0700357 unsigned long isolated, pfn, block_end_pfn;
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100358 LIST_HEAD(freelist);
359
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100360 for (pfn = start_pfn; pfn < end_pfn; pfn += isolated) {
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700361 if (!pfn_valid(pfn) || cc->zone != page_zone(pfn_to_page(pfn)))
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100362 break;
363
364 /*
365 * On subsequent iterations ALIGN() is actually not needed,
366 * but we keep it that we not to complicate the code.
367 */
368 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
369 block_end_pfn = min(block_end_pfn, end_pfn);
370
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700371 isolated = isolate_freepages_block(cc, pfn, block_end_pfn,
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100372 &freelist, true);
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100373
374 /*
375 * In strict mode, isolate_freepages_block() returns 0 if
376 * there are any holes in the block (ie. invalid PFNs or
377 * non-free pages).
378 */
379 if (!isolated)
380 break;
381
382 /*
383 * If we managed to isolate pages, it is always (1 << n) *
384 * pageblock_nr_pages for some non-negative n. (Max order
385 * page may span two pageblocks).
386 */
387 }
388
389 /* split_free_page does not map the pages */
390 map_pages(&freelist);
391
392 if (pfn < end_pfn) {
393 /* Loop terminated early, cleanup. */
394 release_freepages(&freelist);
395 return 0;
396 }
397
398 /* We don't use freelists for anything. */
399 return pfn;
400}
401
Mel Gorman748446b2010-05-24 14:32:27 -0700402/* Update the number of anon and file isolated pages in the zone */
Mel Gormanc67fe372012-08-21 16:16:17 -0700403static void acct_isolated(struct zone *zone, bool locked, struct compact_control *cc)
Mel Gorman748446b2010-05-24 14:32:27 -0700404{
405 struct page *page;
Minchan Kimb9e84ac2011-10-31 17:06:44 -0700406 unsigned int count[2] = { 0, };
Mel Gorman748446b2010-05-24 14:32:27 -0700407
Minchan Kimb9e84ac2011-10-31 17:06:44 -0700408 list_for_each_entry(page, &cc->migratepages, lru)
409 count[!!page_is_file_cache(page)]++;
Mel Gorman748446b2010-05-24 14:32:27 -0700410
Mel Gormanc67fe372012-08-21 16:16:17 -0700411 /* If locked we can use the interrupt unsafe versions */
412 if (locked) {
413 __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
414 __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
415 } else {
416 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
417 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
418 }
Mel Gorman748446b2010-05-24 14:32:27 -0700419}
420
421/* Similar to reclaim, but different enough that they don't share logic */
422static bool too_many_isolated(struct zone *zone)
423{
Minchan Kimbc693042010-09-09 16:38:00 -0700424 unsigned long active, inactive, isolated;
Mel Gorman748446b2010-05-24 14:32:27 -0700425
426 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
427 zone_page_state(zone, NR_INACTIVE_ANON);
Minchan Kimbc693042010-09-09 16:38:00 -0700428 active = zone_page_state(zone, NR_ACTIVE_FILE) +
429 zone_page_state(zone, NR_ACTIVE_ANON);
Mel Gorman748446b2010-05-24 14:32:27 -0700430 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
431 zone_page_state(zone, NR_ISOLATED_ANON);
432
Minchan Kimbc693042010-09-09 16:38:00 -0700433 return isolated > (inactive + active) / 2;
Mel Gorman748446b2010-05-24 14:32:27 -0700434}
435
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100436/**
437 * isolate_migratepages_range() - isolate all migrate-able pages in range.
438 * @zone: Zone pages are in.
439 * @cc: Compaction control structure.
440 * @low_pfn: The first PFN of the range.
441 * @end_pfn: The one-past-the-last PFN of the range.
Minchan Kime46a2872012-10-08 16:33:48 -0700442 * @unevictable: true if it allows to isolate unevictable pages
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100443 *
444 * Isolate all pages that can be migrated from the range specified by
445 * [low_pfn, end_pfn). Returns zero if there is a fatal signal
446 * pending), otherwise PFN of the first page that was not scanned
447 * (which may be both less, equal to or more then end_pfn).
448 *
449 * Assumes that cc->migratepages is empty and cc->nr_migratepages is
450 * zero.
451 *
452 * Apart from cc->migratepages and cc->nr_migratetypes this function
453 * does not modify any cc's fields, in particular it does not modify
454 * (or read for that matter) cc->migrate_pfn.
Mel Gorman748446b2010-05-24 14:32:27 -0700455 */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100456unsigned long
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100457isolate_migratepages_range(struct zone *zone, struct compact_control *cc,
Minchan Kime46a2872012-10-08 16:33:48 -0700458 unsigned long low_pfn, unsigned long end_pfn, bool unevictable)
Mel Gorman748446b2010-05-24 14:32:27 -0700459{
Mel Gorman9927af742011-01-13 15:45:59 -0800460 unsigned long last_pageblock_nr = 0, pageblock_nr;
Mel Gormanb7aba692011-01-13 15:45:54 -0800461 unsigned long nr_scanned = 0, nr_isolated = 0;
Mel Gorman748446b2010-05-24 14:32:27 -0700462 struct list_head *migratelist = &cc->migratepages;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -0700463 isolate_mode_t mode = 0;
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700464 struct lruvec *lruvec;
Mel Gormanc67fe372012-08-21 16:16:17 -0700465 unsigned long flags;
Mel Gorman2a1402a2012-10-08 16:32:33 -0700466 bool locked = false;
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700467 struct page *page = NULL, *valid_page = NULL;
Vlastimil Babka50b5b092014-01-21 15:51:10 -0800468 bool skipped_async_unsuitable = false;
Mel Gorman748446b2010-05-24 14:32:27 -0700469
Mel Gorman748446b2010-05-24 14:32:27 -0700470 /*
471 * Ensure that there are not too many pages isolated from the LRU
472 * list by either parallel reclaimers or compaction. If there are,
473 * delay for some time until fewer pages are isolated
474 */
475 while (unlikely(too_many_isolated(zone))) {
Mel Gormanf9e35b32011-06-15 15:08:52 -0700476 /* async migration should just abort */
Linus Torvalds68e3e922012-06-03 20:05:57 -0700477 if (!cc->sync)
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100478 return 0;
Mel Gormanf9e35b32011-06-15 15:08:52 -0700479
Mel Gorman748446b2010-05-24 14:32:27 -0700480 congestion_wait(BLK_RW_ASYNC, HZ/10);
481
482 if (fatal_signal_pending(current))
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100483 return 0;
Mel Gorman748446b2010-05-24 14:32:27 -0700484 }
485
486 /* Time to isolate some pages for migration */
Andrea Arcangelib2eef8c2011-03-22 16:33:10 -0700487 cond_resched();
Mel Gorman748446b2010-05-24 14:32:27 -0700488 for (; low_pfn < end_pfn; low_pfn++) {
Andrea Arcangelib2eef8c2011-03-22 16:33:10 -0700489 /* give a chance to irqs before checking need_resched() */
Mel Gorman2a1402a2012-10-08 16:32:33 -0700490 if (locked && !((low_pfn+1) % SWAP_CLUSTER_MAX)) {
491 if (should_release_lock(&zone->lru_lock)) {
492 spin_unlock_irqrestore(&zone->lru_lock, flags);
493 locked = false;
494 }
Andrea Arcangelib2eef8c2011-03-22 16:33:10 -0700495 }
Mel Gormanc67fe372012-08-21 16:16:17 -0700496
Mel Gorman0bf380b2012-02-03 15:37:18 -0800497 /*
498 * migrate_pfn does not necessarily start aligned to a
499 * pageblock. Ensure that pfn_valid is called when moving
500 * into a new MAX_ORDER_NR_PAGES range in case of large
501 * memory holes within the zone
502 */
503 if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
504 if (!pfn_valid(low_pfn)) {
505 low_pfn += MAX_ORDER_NR_PAGES - 1;
506 continue;
507 }
508 }
509
Mel Gorman748446b2010-05-24 14:32:27 -0700510 if (!pfn_valid_within(low_pfn))
511 continue;
Mel Gormanb7aba692011-01-13 15:45:54 -0800512 nr_scanned++;
Mel Gorman748446b2010-05-24 14:32:27 -0700513
Mel Gormandc908602012-02-08 17:13:38 -0800514 /*
515 * Get the page and ensure the page is within the same zone.
516 * See the comment in isolate_freepages about overlapping
517 * nodes. It is deliberate that the new zone lock is not taken
518 * as memory compaction should not move pages between nodes.
519 */
Mel Gorman748446b2010-05-24 14:32:27 -0700520 page = pfn_to_page(low_pfn);
Mel Gormandc908602012-02-08 17:13:38 -0800521 if (page_zone(page) != zone)
522 continue;
523
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700524 if (!valid_page)
525 valid_page = page;
526
527 /* If isolation recently failed, do not retry */
528 pageblock_nr = low_pfn >> pageblock_order;
529 if (!isolation_suitable(cc, page))
530 goto next_pageblock;
531
Mel Gorman6c144662014-01-23 15:53:38 -0800532 /*
533 * Skip if free. page_order cannot be used without zone->lock
534 * as nothing prevents parallel allocations or buddy merging.
535 */
Mel Gorman748446b2010-05-24 14:32:27 -0700536 if (PageBuddy(page))
537 continue;
538
Mel Gorman9927af742011-01-13 15:45:59 -0800539 /*
540 * For async migration, also only scan in MOVABLE blocks. Async
541 * migration is optimistic to see if the minimum amount of work
542 * satisfies the allocation
543 */
Linus Torvalds68e3e922012-06-03 20:05:57 -0700544 if (!cc->sync && last_pageblock_nr != pageblock_nr &&
Michal Nazarewicz47118af2011-12-29 13:09:50 +0100545 !migrate_async_suitable(get_pageblock_migratetype(page))) {
Mel Gormanc89511a2012-10-08 16:32:45 -0700546 cc->finished_update_migrate = true;
Vlastimil Babka50b5b092014-01-21 15:51:10 -0800547 skipped_async_unsuitable = true;
Mel Gorman2a1402a2012-10-08 16:32:33 -0700548 goto next_pageblock;
Mel Gorman9927af742011-01-13 15:45:59 -0800549 }
550
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800551 /*
552 * Check may be lockless but that's ok as we recheck later.
553 * It's possible to migrate LRU pages and balloon pages
554 * Skip any other type of page
555 */
556 if (!PageLRU(page)) {
557 if (unlikely(balloon_page_movable(page))) {
558 if (locked && balloon_page_isolate(page)) {
559 /* Successfully isolated */
560 cc->finished_update_migrate = true;
561 list_add(&page->lru, migratelist);
562 cc->nr_migratepages++;
563 nr_isolated++;
564 goto check_compact_cluster;
565 }
566 }
Andrea Arcangelibc835012011-01-13 15:47:08 -0800567 continue;
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800568 }
Andrea Arcangelibc835012011-01-13 15:47:08 -0800569
570 /*
Mel Gorman2a1402a2012-10-08 16:32:33 -0700571 * PageLRU is set. lru_lock normally excludes isolation
572 * splitting and collapsing (collapsing has already happened
573 * if PageLRU is set) but the lock is not necessarily taken
574 * here and it is wasteful to take it just to check transhuge.
575 * Check TransHuge without lock and skip the whole pageblock if
576 * it's either a transhuge or hugetlbfs page, as calling
577 * compound_order() without preventing THP from splitting the
578 * page underneath us may return surprising results.
Andrea Arcangelibc835012011-01-13 15:47:08 -0800579 */
580 if (PageTransHuge(page)) {
Mel Gorman2a1402a2012-10-08 16:32:33 -0700581 if (!locked)
582 goto next_pageblock;
583 low_pfn += (1 << compound_order(page)) - 1;
584 continue;
585 }
586
David Rientjes119d6d52014-04-03 14:48:00 -0700587 /*
588 * Migration will fail if an anonymous page is pinned in memory,
589 * so avoid taking lru_lock and isolating it unnecessarily in an
590 * admittedly racy check.
591 */
592 if (!page_mapping(page) &&
593 page_count(page) > page_mapcount(page))
594 continue;
595
Mel Gorman2a1402a2012-10-08 16:32:33 -0700596 /* Check if it is ok to still hold the lock */
597 locked = compact_checklock_irqsave(&zone->lru_lock, &flags,
598 locked, cc);
599 if (!locked || fatal_signal_pending(current))
600 break;
601
602 /* Recheck PageLRU and PageTransHuge under lock */
603 if (!PageLRU(page))
604 continue;
605 if (PageTransHuge(page)) {
Andrea Arcangelibc835012011-01-13 15:47:08 -0800606 low_pfn += (1 << compound_order(page)) - 1;
607 continue;
608 }
609
Linus Torvalds68e3e922012-06-03 20:05:57 -0700610 if (!cc->sync)
Mel Gormanc8244932012-01-12 17:19:38 -0800611 mode |= ISOLATE_ASYNC_MIGRATE;
612
Minchan Kime46a2872012-10-08 16:33:48 -0700613 if (unevictable)
614 mode |= ISOLATE_UNEVICTABLE;
615
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700616 lruvec = mem_cgroup_page_lruvec(page, zone);
617
Mel Gorman748446b2010-05-24 14:32:27 -0700618 /* Try isolate the page */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -0700619 if (__isolate_lru_page(page, mode) != 0)
Mel Gorman748446b2010-05-24 14:32:27 -0700620 continue;
621
Sasha Levin309381fea2014-01-23 15:52:54 -0800622 VM_BUG_ON_PAGE(PageTransCompound(page), page);
Andrea Arcangelibc835012011-01-13 15:47:08 -0800623
Mel Gorman748446b2010-05-24 14:32:27 -0700624 /* Successfully isolated */
Mel Gormanc89511a2012-10-08 16:32:45 -0700625 cc->finished_update_migrate = true;
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700626 del_page_from_lru_list(page, lruvec, page_lru(page));
Mel Gorman748446b2010-05-24 14:32:27 -0700627 list_add(&page->lru, migratelist);
Mel Gorman748446b2010-05-24 14:32:27 -0700628 cc->nr_migratepages++;
Mel Gormanb7aba692011-01-13 15:45:54 -0800629 nr_isolated++;
Mel Gorman748446b2010-05-24 14:32:27 -0700630
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800631check_compact_cluster:
Mel Gorman748446b2010-05-24 14:32:27 -0700632 /* Avoid isolating too much */
Hillf Danton31b83842012-01-10 15:07:59 -0800633 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
634 ++low_pfn;
Mel Gorman748446b2010-05-24 14:32:27 -0700635 break;
Hillf Danton31b83842012-01-10 15:07:59 -0800636 }
Mel Gorman2a1402a2012-10-08 16:32:33 -0700637
638 continue;
639
640next_pageblock:
Mel Gormana9aacbc2013-02-22 16:32:25 -0800641 low_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages) - 1;
Mel Gorman2a1402a2012-10-08 16:32:33 -0700642 last_pageblock_nr = pageblock_nr;
Mel Gorman748446b2010-05-24 14:32:27 -0700643 }
644
Mel Gormanc67fe372012-08-21 16:16:17 -0700645 acct_isolated(zone, locked, cc);
Mel Gorman748446b2010-05-24 14:32:27 -0700646
Mel Gormanc67fe372012-08-21 16:16:17 -0700647 if (locked)
648 spin_unlock_irqrestore(&zone->lru_lock, flags);
Mel Gorman748446b2010-05-24 14:32:27 -0700649
Vlastimil Babka50b5b092014-01-21 15:51:10 -0800650 /*
651 * Update the pageblock-skip information and cached scanner pfn,
652 * if the whole pageblock was scanned without isolating any page.
653 * This is not done when pageblock was skipped due to being unsuitable
654 * for async compaction, so that eventual sync compaction can try.
655 */
656 if (low_pfn == end_pfn && !skipped_async_unsuitable)
Mel Gormanc89511a2012-10-08 16:32:45 -0700657 update_pageblock_skip(cc, valid_page, nr_isolated, true);
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700658
Mel Gormanb7aba692011-01-13 15:45:54 -0800659 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
660
Minchan Kim010fc292012-12-20 15:05:06 -0800661 count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
Mel Gorman397487d2012-10-19 12:00:10 +0100662 if (nr_isolated)
Minchan Kim010fc292012-12-20 15:05:06 -0800663 count_compact_events(COMPACTISOLATED, nr_isolated);
Mel Gorman397487d2012-10-19 12:00:10 +0100664
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100665 return low_pfn;
666}
667
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100668#endif /* CONFIG_COMPACTION || CONFIG_CMA */
669#ifdef CONFIG_COMPACTION
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100670/*
671 * Based on information in the current compact_control, find blocks
672 * suitable for isolating free pages from and then isolate them.
673 */
674static void isolate_freepages(struct zone *zone,
675 struct compact_control *cc)
676{
677 struct page *page;
Cody P Schafer108bcc92013-02-22 16:35:23 -0800678 unsigned long high_pfn, low_pfn, pfn, z_end_pfn, end_pfn;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100679 int nr_freepages = cc->nr_freepages;
680 struct list_head *freelist = &cc->freepages;
681
682 /*
683 * Initialise the free scanner. The starting point is where we last
684 * scanned from (or the end of the zone if starting). The low point
685 * is the end of the pageblock the migration scanner is using.
686 */
687 pfn = cc->free_pfn;
Vlastimil Babka7ed695e2014-01-21 15:51:09 -0800688 low_pfn = ALIGN(cc->migrate_pfn + 1, pageblock_nr_pages);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100689
690 /*
691 * Take care that if the migration scanner is at the end of the zone
692 * that the free scanner does not accidentally move to the next zone
693 * in the next isolation cycle.
694 */
695 high_pfn = min(low_pfn, pfn);
696
Cody P Schafer108bcc92013-02-22 16:35:23 -0800697 z_end_pfn = zone_end_pfn(zone);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100698
699 /*
700 * Isolate free pages until enough are available to migrate the
701 * pages on cc->migratepages. We stop searching if the migrate
702 * and free page scanners meet or enough free pages are isolated.
703 */
Vlastimil Babka7ed695e2014-01-21 15:51:09 -0800704 for (; pfn >= low_pfn && cc->nr_migratepages > nr_freepages;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100705 pfn -= pageblock_nr_pages) {
706 unsigned long isolated;
707
David Rientjesf6ea3ad2013-09-30 13:45:03 -0700708 /*
709 * This can iterate a massively long zone without finding any
710 * suitable migration targets, so periodically check if we need
711 * to schedule.
712 */
713 cond_resched();
714
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100715 if (!pfn_valid(pfn))
716 continue;
717
718 /*
719 * Check for overlapping nodes/zones. It's possible on some
720 * configurations to have a setup like
721 * node0 node1 node0
722 * i.e. it's possible that all pages within a zones range of
723 * pages do not belong to a single zone.
724 */
725 page = pfn_to_page(pfn);
726 if (page_zone(page) != zone)
727 continue;
728
729 /* Check the block is suitable for migration */
Linus Torvalds68e3e922012-06-03 20:05:57 -0700730 if (!suitable_migration_target(page))
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100731 continue;
Linus Torvalds68e3e922012-06-03 20:05:57 -0700732
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700733 /* If isolation recently failed, do not retry */
734 if (!isolation_suitable(cc, page))
735 continue;
736
Mel Gormanf40d1e42012-10-08 16:32:36 -0700737 /* Found a block suitable for isolating free pages from */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100738 isolated = 0;
Mel Gorman60177d32012-12-06 19:01:14 +0000739
740 /*
741 * As pfn may not start aligned, pfn+pageblock_nr_page
742 * may cross a MAX_ORDER_NR_PAGES boundary and miss
743 * a pfn_valid check. Ensure isolate_freepages_block()
744 * only scans within a pageblock
745 */
746 end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
Cody P Schafer108bcc92013-02-22 16:35:23 -0800747 end_pfn = min(end_pfn, z_end_pfn);
Mel Gormanf40d1e42012-10-08 16:32:36 -0700748 isolated = isolate_freepages_block(cc, pfn, end_pfn,
749 freelist, false);
750 nr_freepages += isolated;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100751
752 /*
753 * Record the highest PFN we isolated pages from. When next
754 * looking for free pages, the search will restart here as
755 * page migration may have returned some pages to the allocator
756 */
Mel Gormanc89511a2012-10-08 16:32:45 -0700757 if (isolated) {
758 cc->finished_update_free = true;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100759 high_pfn = max(high_pfn, pfn);
Mel Gormanc89511a2012-10-08 16:32:45 -0700760 }
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100761 }
762
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100763 /* split_free_page does not map the pages */
764 map_pages(freelist);
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100765
Vlastimil Babka7ed695e2014-01-21 15:51:09 -0800766 /*
767 * If we crossed the migrate scanner, we want to keep it that way
768 * so that compact_finished() may detect this
769 */
770 if (pfn < low_pfn)
771 cc->free_pfn = max(pfn, zone->zone_start_pfn);
772 else
773 cc->free_pfn = high_pfn;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100774 cc->nr_freepages = nr_freepages;
Mel Gorman748446b2010-05-24 14:32:27 -0700775}
776
777/*
778 * This is a migrate-callback that "allocates" freepages by taking pages
779 * from the isolated freelists in the block we are migrating to.
780 */
781static struct page *compaction_alloc(struct page *migratepage,
782 unsigned long data,
783 int **result)
784{
785 struct compact_control *cc = (struct compact_control *)data;
786 struct page *freepage;
787
788 /* Isolate free pages if necessary */
789 if (list_empty(&cc->freepages)) {
790 isolate_freepages(cc->zone, cc);
791
792 if (list_empty(&cc->freepages))
793 return NULL;
794 }
795
796 freepage = list_entry(cc->freepages.next, struct page, lru);
797 list_del(&freepage->lru);
798 cc->nr_freepages--;
799
800 return freepage;
801}
802
803/*
804 * We cannot control nr_migratepages and nr_freepages fully when migration is
805 * running as migrate_pages() has no knowledge of compact_control. When
806 * migration is complete, we count the number of pages on the lists by hand.
807 */
808static void update_nr_listpages(struct compact_control *cc)
809{
810 int nr_migratepages = 0;
811 int nr_freepages = 0;
812 struct page *page;
813
814 list_for_each_entry(page, &cc->migratepages, lru)
815 nr_migratepages++;
816 list_for_each_entry(page, &cc->freepages, lru)
817 nr_freepages++;
818
819 cc->nr_migratepages = nr_migratepages;
820 cc->nr_freepages = nr_freepages;
821}
822
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100823/* possible outcome of isolate_migratepages */
824typedef enum {
825 ISOLATE_ABORT, /* Abort compaction now */
826 ISOLATE_NONE, /* No pages isolated, continue scanning */
827 ISOLATE_SUCCESS, /* Pages isolated, migrate */
828} isolate_migrate_t;
829
830/*
831 * Isolate all pages that can be migrated from the block pointed to by
832 * the migrate scanner within compact_control.
833 */
834static isolate_migrate_t isolate_migratepages(struct zone *zone,
835 struct compact_control *cc)
836{
837 unsigned long low_pfn, end_pfn;
838
839 /* Do not scan outside zone boundaries */
840 low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
841
842 /* Only scan within a pageblock boundary */
Mel Gormana9aacbc2013-02-22 16:32:25 -0800843 end_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100844
845 /* Do not cross the free scanner or scan within a memory hole */
846 if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
847 cc->migrate_pfn = end_pfn;
848 return ISOLATE_NONE;
849 }
850
851 /* Perform the isolation */
Minchan Kime46a2872012-10-08 16:33:48 -0700852 low_pfn = isolate_migratepages_range(zone, cc, low_pfn, end_pfn, false);
Shaohua Lie64c5232012-10-08 16:32:27 -0700853 if (!low_pfn || cc->contended)
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100854 return ISOLATE_ABORT;
855
856 cc->migrate_pfn = low_pfn;
857
858 return ISOLATE_SUCCESS;
859}
860
Mel Gorman748446b2010-05-24 14:32:27 -0700861static int compact_finished(struct zone *zone,
Andrea Arcangeli5a03b052011-01-13 15:47:11 -0800862 struct compact_control *cc)
Mel Gorman748446b2010-05-24 14:32:27 -0700863{
Mel Gorman8fb74b92013-01-11 14:32:16 -0800864 unsigned int order;
Andrea Arcangeli5a03b052011-01-13 15:47:11 -0800865 unsigned long watermark;
Mel Gorman56de7262010-05-24 14:32:30 -0700866
Mel Gorman748446b2010-05-24 14:32:27 -0700867 if (fatal_signal_pending(current))
868 return COMPACT_PARTIAL;
869
Mel Gorman753341a2012-10-08 16:32:40 -0700870 /* Compaction run completes if the migrate and free scanner meet */
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700871 if (cc->free_pfn <= cc->migrate_pfn) {
Vlastimil Babka55b7c4c2014-01-21 15:51:11 -0800872 /* Let the next compaction start anew. */
873 zone->compact_cached_migrate_pfn = zone->zone_start_pfn;
874 zone->compact_cached_free_pfn = zone_end_pfn(zone);
875
Mel Gorman62997022012-10-08 16:32:47 -0700876 /*
877 * Mark that the PG_migrate_skip information should be cleared
878 * by kswapd when it goes to sleep. kswapd does not set the
879 * flag itself as the decision to be clear should be directly
880 * based on an allocation request.
881 */
882 if (!current_is_kswapd())
883 zone->compact_blockskip_flush = true;
884
Mel Gorman748446b2010-05-24 14:32:27 -0700885 return COMPACT_COMPLETE;
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700886 }
Mel Gorman748446b2010-05-24 14:32:27 -0700887
Johannes Weiner82478fb2011-01-20 14:44:21 -0800888 /*
889 * order == -1 is expected when compacting via
890 * /proc/sys/vm/compact_memory
891 */
Mel Gorman56de7262010-05-24 14:32:30 -0700892 if (cc->order == -1)
893 return COMPACT_CONTINUE;
894
Michal Hocko3957c772011-06-15 15:08:25 -0700895 /* Compaction run is not finished if the watermark is not met */
896 watermark = low_wmark_pages(zone);
897 watermark += (1 << cc->order);
898
899 if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
900 return COMPACT_CONTINUE;
901
Mel Gorman56de7262010-05-24 14:32:30 -0700902 /* Direct compactor: Is a suitable page free? */
Mel Gorman8fb74b92013-01-11 14:32:16 -0800903 for (order = cc->order; order < MAX_ORDER; order++) {
904 struct free_area *area = &zone->free_area[order];
Mel Gorman56de7262010-05-24 14:32:30 -0700905
Mel Gorman8fb74b92013-01-11 14:32:16 -0800906 /* Job done if page is free of the right migratetype */
907 if (!list_empty(&area->free_list[cc->migratetype]))
908 return COMPACT_PARTIAL;
909
910 /* Job done if allocation would set block type */
911 if (cc->order >= pageblock_order && area->nr_free)
912 return COMPACT_PARTIAL;
Mel Gorman56de7262010-05-24 14:32:30 -0700913 }
914
Mel Gorman748446b2010-05-24 14:32:27 -0700915 return COMPACT_CONTINUE;
916}
917
Mel Gorman3e7d3442011-01-13 15:45:56 -0800918/*
919 * compaction_suitable: Is this suitable to run compaction on this zone now?
920 * Returns
921 * COMPACT_SKIPPED - If there are too few free pages for compaction
922 * COMPACT_PARTIAL - If the allocation would succeed without compaction
923 * COMPACT_CONTINUE - If compaction should run now
924 */
925unsigned long compaction_suitable(struct zone *zone, int order)
926{
927 int fragindex;
928 unsigned long watermark;
929
930 /*
Michal Hocko3957c772011-06-15 15:08:25 -0700931 * order == -1 is expected when compacting via
932 * /proc/sys/vm/compact_memory
933 */
934 if (order == -1)
935 return COMPACT_CONTINUE;
936
937 /*
Mel Gorman3e7d3442011-01-13 15:45:56 -0800938 * Watermarks for order-0 must be met for compaction. Note the 2UL.
939 * This is because during migration, copies of pages need to be
940 * allocated and for a short time, the footprint is higher
941 */
942 watermark = low_wmark_pages(zone) + (2UL << order);
943 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
944 return COMPACT_SKIPPED;
945
946 /*
947 * fragmentation index determines if allocation failures are due to
948 * low memory or external fragmentation
949 *
Shaohua Lia582a732011-06-15 15:08:49 -0700950 * index of -1000 implies allocations might succeed depending on
951 * watermarks
Mel Gorman3e7d3442011-01-13 15:45:56 -0800952 * index towards 0 implies failure is due to lack of memory
953 * index towards 1000 implies failure is due to fragmentation
954 *
955 * Only compact if a failure would be due to fragmentation.
956 */
957 fragindex = fragmentation_index(zone, order);
958 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
959 return COMPACT_SKIPPED;
960
Shaohua Lia582a732011-06-15 15:08:49 -0700961 if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
962 0, 0))
Mel Gorman3e7d3442011-01-13 15:45:56 -0800963 return COMPACT_PARTIAL;
964
965 return COMPACT_CONTINUE;
966}
967
Mel Gorman748446b2010-05-24 14:32:27 -0700968static int compact_zone(struct zone *zone, struct compact_control *cc)
969{
970 int ret;
Mel Gormanc89511a2012-10-08 16:32:45 -0700971 unsigned long start_pfn = zone->zone_start_pfn;
Cody P Schafer108bcc92013-02-22 16:35:23 -0800972 unsigned long end_pfn = zone_end_pfn(zone);
Mel Gorman748446b2010-05-24 14:32:27 -0700973
Mel Gorman3e7d3442011-01-13 15:45:56 -0800974 ret = compaction_suitable(zone, cc->order);
975 switch (ret) {
976 case COMPACT_PARTIAL:
977 case COMPACT_SKIPPED:
978 /* Compaction is likely to fail */
979 return ret;
980 case COMPACT_CONTINUE:
981 /* Fall through to compaction */
982 ;
983 }
984
Mel Gormanc89511a2012-10-08 16:32:45 -0700985 /*
Vlastimil Babkad3132e42014-01-21 15:51:08 -0800986 * Clear pageblock skip if there were failures recently and compaction
987 * is about to be retried after being deferred. kswapd does not do
988 * this reset as it'll reset the cached information when going to sleep.
989 */
990 if (compaction_restarting(zone, cc->order) && !current_is_kswapd())
991 __reset_isolation_suitable(zone);
992
993 /*
Mel Gormanc89511a2012-10-08 16:32:45 -0700994 * Setup to move all movable pages to the end of the zone. Used cached
995 * information on where the scanners should start but check that it
996 * is initialised by ensuring the values are within zone boundaries.
997 */
998 cc->migrate_pfn = zone->compact_cached_migrate_pfn;
999 cc->free_pfn = zone->compact_cached_free_pfn;
1000 if (cc->free_pfn < start_pfn || cc->free_pfn > end_pfn) {
1001 cc->free_pfn = end_pfn & ~(pageblock_nr_pages-1);
1002 zone->compact_cached_free_pfn = cc->free_pfn;
1003 }
1004 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn > end_pfn) {
1005 cc->migrate_pfn = start_pfn;
1006 zone->compact_cached_migrate_pfn = cc->migrate_pfn;
1007 }
Mel Gorman748446b2010-05-24 14:32:27 -07001008
Mel Gorman0eb927c2014-01-21 15:51:05 -08001009 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn, cc->free_pfn, end_pfn);
1010
Mel Gorman748446b2010-05-24 14:32:27 -07001011 migrate_prep_local();
1012
1013 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
1014 unsigned long nr_migrate, nr_remaining;
Minchan Kim9d502c12011-03-22 16:30:39 -07001015 int err;
Mel Gorman748446b2010-05-24 14:32:27 -07001016
Mel Gormanf9e35b32011-06-15 15:08:52 -07001017 switch (isolate_migratepages(zone, cc)) {
1018 case ISOLATE_ABORT:
1019 ret = COMPACT_PARTIAL;
Rafael Aquini5733c7d2012-12-11 16:02:47 -08001020 putback_movable_pages(&cc->migratepages);
Shaohua Lie64c5232012-10-08 16:32:27 -07001021 cc->nr_migratepages = 0;
Mel Gormanf9e35b32011-06-15 15:08:52 -07001022 goto out;
1023 case ISOLATE_NONE:
Mel Gorman748446b2010-05-24 14:32:27 -07001024 continue;
Mel Gormanf9e35b32011-06-15 15:08:52 -07001025 case ISOLATE_SUCCESS:
1026 ;
1027 }
Mel Gorman748446b2010-05-24 14:32:27 -07001028
1029 nr_migrate = cc->nr_migratepages;
Minchan Kim9d502c12011-03-22 16:30:39 -07001030 err = migrate_pages(&cc->migratepages, compaction_alloc,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001031 (unsigned long)cc,
Mel Gorman7b2a2d42012-10-19 14:07:31 +01001032 cc->sync ? MIGRATE_SYNC_LIGHT : MIGRATE_ASYNC,
1033 MR_COMPACTION);
Mel Gorman748446b2010-05-24 14:32:27 -07001034 update_nr_listpages(cc);
1035 nr_remaining = cc->nr_migratepages;
1036
Mel Gormanb7aba692011-01-13 15:45:54 -08001037 trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
1038 nr_remaining);
Mel Gorman748446b2010-05-24 14:32:27 -07001039
Rafael Aquini5733c7d2012-12-11 16:02:47 -08001040 /* Release isolated pages not migrated */
Minchan Kim9d502c12011-03-22 16:30:39 -07001041 if (err) {
Rafael Aquini5733c7d2012-12-11 16:02:47 -08001042 putback_movable_pages(&cc->migratepages);
Mel Gorman748446b2010-05-24 14:32:27 -07001043 cc->nr_migratepages = 0;
Vlastimil Babka7ed695e2014-01-21 15:51:09 -08001044 /*
1045 * migrate_pages() may return -ENOMEM when scanners meet
1046 * and we want compact_finished() to detect it
1047 */
1048 if (err == -ENOMEM && cc->free_pfn > cc->migrate_pfn) {
David Rientjes4bf2bba2012-07-11 14:02:13 -07001049 ret = COMPACT_PARTIAL;
1050 goto out;
1051 }
Mel Gorman748446b2010-05-24 14:32:27 -07001052 }
Mel Gorman748446b2010-05-24 14:32:27 -07001053 }
1054
Mel Gormanf9e35b32011-06-15 15:08:52 -07001055out:
Mel Gorman748446b2010-05-24 14:32:27 -07001056 /* Release free pages and check accounting */
1057 cc->nr_freepages -= release_freepages(&cc->freepages);
1058 VM_BUG_ON(cc->nr_freepages != 0);
1059
Mel Gorman0eb927c2014-01-21 15:51:05 -08001060 trace_mm_compaction_end(ret);
1061
Mel Gorman748446b2010-05-24 14:32:27 -07001062 return ret;
1063}
Mel Gorman76ab0f52010-05-24 14:32:28 -07001064
Kyungmin Parkd43a87e2011-10-31 17:09:08 -07001065static unsigned long compact_zone_order(struct zone *zone,
Andrea Arcangeli5a03b052011-01-13 15:47:11 -08001066 int order, gfp_t gfp_mask,
Mel Gorman8fb74b92013-01-11 14:32:16 -08001067 bool sync, bool *contended)
Mel Gorman56de7262010-05-24 14:32:30 -07001068{
Shaohua Lie64c5232012-10-08 16:32:27 -07001069 unsigned long ret;
Mel Gorman56de7262010-05-24 14:32:30 -07001070 struct compact_control cc = {
1071 .nr_freepages = 0,
1072 .nr_migratepages = 0,
1073 .order = order,
1074 .migratetype = allocflags_to_migratetype(gfp_mask),
1075 .zone = zone,
Linus Torvalds68e3e922012-06-03 20:05:57 -07001076 .sync = sync,
Mel Gorman56de7262010-05-24 14:32:30 -07001077 };
1078 INIT_LIST_HEAD(&cc.freepages);
1079 INIT_LIST_HEAD(&cc.migratepages);
1080
Shaohua Lie64c5232012-10-08 16:32:27 -07001081 ret = compact_zone(zone, &cc);
1082
1083 VM_BUG_ON(!list_empty(&cc.freepages));
1084 VM_BUG_ON(!list_empty(&cc.migratepages));
1085
1086 *contended = cc.contended;
1087 return ret;
Mel Gorman56de7262010-05-24 14:32:30 -07001088}
1089
Mel Gorman5e771902010-05-24 14:32:31 -07001090int sysctl_extfrag_threshold = 500;
1091
Mel Gorman56de7262010-05-24 14:32:30 -07001092/**
1093 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1094 * @zonelist: The zonelist used for the current allocation
1095 * @order: The order of the current allocation
1096 * @gfp_mask: The GFP mask of the current allocation
1097 * @nodemask: The allowed nodes to allocate from
Mel Gorman77f1fe62011-01-13 15:45:57 -08001098 * @sync: Whether migration is synchronous or not
Mel Gorman661c4cb2012-10-08 16:32:31 -07001099 * @contended: Return value that is true if compaction was aborted due to lock contention
1100 * @page: Optionally capture a free page of the requested order during compaction
Mel Gorman56de7262010-05-24 14:32:30 -07001101 *
1102 * This is the main entry point for direct page compaction.
1103 */
1104unsigned long try_to_compact_pages(struct zonelist *zonelist,
Mel Gorman77f1fe62011-01-13 15:45:57 -08001105 int order, gfp_t gfp_mask, nodemask_t *nodemask,
Mel Gorman8fb74b92013-01-11 14:32:16 -08001106 bool sync, bool *contended)
Mel Gorman56de7262010-05-24 14:32:30 -07001107{
1108 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1109 int may_enter_fs = gfp_mask & __GFP_FS;
1110 int may_perform_io = gfp_mask & __GFP_IO;
Mel Gorman56de7262010-05-24 14:32:30 -07001111 struct zoneref *z;
1112 struct zone *zone;
1113 int rc = COMPACT_SKIPPED;
Bartlomiej Zolnierkiewiczd95ea5d2012-10-08 16:32:05 -07001114 int alloc_flags = 0;
Mel Gorman56de7262010-05-24 14:32:30 -07001115
Mel Gorman4ffb6332012-10-08 16:29:09 -07001116 /* Check if the GFP flags allow compaction */
Andrea Arcangelic5a73c32011-01-13 15:47:11 -08001117 if (!order || !may_enter_fs || !may_perform_io)
Mel Gorman56de7262010-05-24 14:32:30 -07001118 return rc;
1119
Minchan Kim010fc292012-12-20 15:05:06 -08001120 count_compact_event(COMPACTSTALL);
Mel Gorman56de7262010-05-24 14:32:30 -07001121
Bartlomiej Zolnierkiewiczd95ea5d2012-10-08 16:32:05 -07001122#ifdef CONFIG_CMA
1123 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
1124 alloc_flags |= ALLOC_CMA;
1125#endif
Mel Gorman56de7262010-05-24 14:32:30 -07001126 /* Compact each zone in the list */
1127 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1128 nodemask) {
Mel Gorman56de7262010-05-24 14:32:30 -07001129 int status;
1130
Mel Gormanc67fe372012-08-21 16:16:17 -07001131 status = compact_zone_order(zone, order, gfp_mask, sync,
Mel Gorman8fb74b92013-01-11 14:32:16 -08001132 contended);
Mel Gorman56de7262010-05-24 14:32:30 -07001133 rc = max(status, rc);
1134
Mel Gorman3e7d3442011-01-13 15:45:56 -08001135 /* If a normal allocation would succeed, stop compacting */
Bartlomiej Zolnierkiewiczd95ea5d2012-10-08 16:32:05 -07001136 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0,
1137 alloc_flags))
Mel Gorman56de7262010-05-24 14:32:30 -07001138 break;
1139 }
1140
1141 return rc;
1142}
1143
1144
Mel Gorman76ab0f52010-05-24 14:32:28 -07001145/* Compact all zones within a node */
Andrew Morton7103f162013-02-22 16:32:33 -08001146static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
Mel Gorman76ab0f52010-05-24 14:32:28 -07001147{
1148 int zoneid;
Mel Gorman76ab0f52010-05-24 14:32:28 -07001149 struct zone *zone;
1150
Mel Gorman76ab0f52010-05-24 14:32:28 -07001151 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
Mel Gorman76ab0f52010-05-24 14:32:28 -07001152
1153 zone = &pgdat->node_zones[zoneid];
1154 if (!populated_zone(zone))
1155 continue;
1156
Rik van Riel7be62de2012-03-21 16:33:52 -07001157 cc->nr_freepages = 0;
1158 cc->nr_migratepages = 0;
1159 cc->zone = zone;
1160 INIT_LIST_HEAD(&cc->freepages);
1161 INIT_LIST_HEAD(&cc->migratepages);
Mel Gorman76ab0f52010-05-24 14:32:28 -07001162
Dan Carpenteraad6ec32012-03-21 16:33:54 -07001163 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
Rik van Riel7be62de2012-03-21 16:33:52 -07001164 compact_zone(zone, cc);
Mel Gorman76ab0f52010-05-24 14:32:28 -07001165
Rik van Rielaff62242012-03-21 16:33:52 -07001166 if (cc->order > 0) {
Vlastimil Babkade6c60a2014-01-21 15:51:07 -08001167 if (zone_watermark_ok(zone, cc->order,
1168 low_wmark_pages(zone), 0, 0))
1169 compaction_defer_reset(zone, cc->order, false);
Rik van Rielaff62242012-03-21 16:33:52 -07001170 /* Currently async compaction is never deferred. */
Vlastimil Babkade6c60a2014-01-21 15:51:07 -08001171 else if (cc->sync)
Rik van Rielaff62242012-03-21 16:33:52 -07001172 defer_compaction(zone, cc->order);
1173 }
1174
Rik van Riel7be62de2012-03-21 16:33:52 -07001175 VM_BUG_ON(!list_empty(&cc->freepages));
1176 VM_BUG_ON(!list_empty(&cc->migratepages));
Mel Gorman76ab0f52010-05-24 14:32:28 -07001177 }
Mel Gorman76ab0f52010-05-24 14:32:28 -07001178}
1179
Andrew Morton7103f162013-02-22 16:32:33 -08001180void compact_pgdat(pg_data_t *pgdat, int order)
Rik van Riel7be62de2012-03-21 16:33:52 -07001181{
1182 struct compact_control cc = {
1183 .order = order,
Linus Torvalds68e3e922012-06-03 20:05:57 -07001184 .sync = false,
Rik van Riel7be62de2012-03-21 16:33:52 -07001185 };
1186
Mel Gorman3a7200a2013-09-11 14:22:19 -07001187 if (!order)
1188 return;
1189
Andrew Morton7103f162013-02-22 16:32:33 -08001190 __compact_pgdat(pgdat, &cc);
Rik van Riel7be62de2012-03-21 16:33:52 -07001191}
1192
Andrew Morton7103f162013-02-22 16:32:33 -08001193static void compact_node(int nid)
Rik van Riel7be62de2012-03-21 16:33:52 -07001194{
Rik van Riel7be62de2012-03-21 16:33:52 -07001195 struct compact_control cc = {
1196 .order = -1,
Linus Torvalds68e3e922012-06-03 20:05:57 -07001197 .sync = true,
David Rientjes91ca9182014-04-03 14:47:23 -07001198 .ignore_skip_hint = true,
Rik van Riel7be62de2012-03-21 16:33:52 -07001199 };
1200
Andrew Morton7103f162013-02-22 16:32:33 -08001201 __compact_pgdat(NODE_DATA(nid), &cc);
Rik van Riel7be62de2012-03-21 16:33:52 -07001202}
1203
Mel Gorman76ab0f52010-05-24 14:32:28 -07001204/* Compact all nodes in the system */
Jason Liu7964c062013-01-11 14:31:47 -08001205static void compact_nodes(void)
Mel Gorman76ab0f52010-05-24 14:32:28 -07001206{
1207 int nid;
1208
Hugh Dickins8575ec22012-03-21 16:33:53 -07001209 /* Flush pending updates to the LRU lists */
1210 lru_add_drain_all();
1211
Mel Gorman76ab0f52010-05-24 14:32:28 -07001212 for_each_online_node(nid)
1213 compact_node(nid);
Mel Gorman76ab0f52010-05-24 14:32:28 -07001214}
1215
1216/* The written value is actually unused, all memory is compacted */
1217int sysctl_compact_memory;
1218
1219/* This is the entry point for compacting all nodes via /proc/sys/vm */
1220int sysctl_compaction_handler(struct ctl_table *table, int write,
1221 void __user *buffer, size_t *length, loff_t *ppos)
1222{
1223 if (write)
Jason Liu7964c062013-01-11 14:31:47 -08001224 compact_nodes();
Mel Gorman76ab0f52010-05-24 14:32:28 -07001225
1226 return 0;
1227}
Mel Gormaned4a6d72010-05-24 14:32:29 -07001228
Mel Gorman5e771902010-05-24 14:32:31 -07001229int sysctl_extfrag_handler(struct ctl_table *table, int write,
1230 void __user *buffer, size_t *length, loff_t *ppos)
1231{
1232 proc_dointvec_minmax(table, write, buffer, length, ppos);
1233
1234 return 0;
1235}
1236
Mel Gormaned4a6d72010-05-24 14:32:29 -07001237#if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
Rashika Kheria74e77fb2014-04-03 14:48:01 -07001238static ssize_t sysfs_compact_node(struct device *dev,
Kay Sievers10fbcf42011-12-21 14:48:43 -08001239 struct device_attribute *attr,
Mel Gormaned4a6d72010-05-24 14:32:29 -07001240 const char *buf, size_t count)
1241{
Hugh Dickins8575ec22012-03-21 16:33:53 -07001242 int nid = dev->id;
1243
1244 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1245 /* Flush pending updates to the LRU lists */
1246 lru_add_drain_all();
1247
1248 compact_node(nid);
1249 }
Mel Gormaned4a6d72010-05-24 14:32:29 -07001250
1251 return count;
1252}
Kay Sievers10fbcf42011-12-21 14:48:43 -08001253static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
Mel Gormaned4a6d72010-05-24 14:32:29 -07001254
1255int compaction_register_node(struct node *node)
1256{
Kay Sievers10fbcf42011-12-21 14:48:43 -08001257 return device_create_file(&node->dev, &dev_attr_compact);
Mel Gormaned4a6d72010-05-24 14:32:29 -07001258}
1259
1260void compaction_unregister_node(struct node *node)
1261{
Kay Sievers10fbcf42011-12-21 14:48:43 -08001262 return device_remove_file(&node->dev, &dev_attr_compact);
Mel Gormaned4a6d72010-05-24 14:32:29 -07001263}
1264#endif /* CONFIG_SYSFS && CONFIG_NUMA */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001265
1266#endif /* CONFIG_COMPACTION */