blob: 9a03fdb1fd848bca383a315f35ff875029c65de8 [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;
Mel Gorman748446b2010-05-24 14:32:27 -0700247
Mel Gorman748446b2010-05-24 14:32:27 -0700248 cursor = pfn_to_page(blockpfn);
249
Mel Gormanf40d1e42012-10-08 16:32:36 -0700250 /* Isolate free pages. */
Mel Gorman748446b2010-05-24 14:32:27 -0700251 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
252 int isolated, i;
253 struct page *page = cursor;
254
Mel Gormanb7aba692011-01-13 15:45:54 -0800255 nr_scanned++;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700256 if (!pfn_valid_within(blockpfn))
Laura Abbott2af120b2014-03-10 15:49:44 -0700257 goto isolate_fail;
258
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700259 if (!valid_page)
260 valid_page = page;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700261 if (!PageBuddy(page))
Laura Abbott2af120b2014-03-10 15:49:44 -0700262 goto isolate_fail;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700263
264 /*
265 * The zone lock must be held to isolate freepages.
266 * Unfortunately this is a very coarse lock and can be
267 * heavily contended if there are parallel allocations
268 * or parallel compactions. For async compaction do not
269 * spin on the lock and we acquire the lock as late as
270 * possible.
271 */
272 locked = compact_checklock_irqsave(&cc->zone->lock, &flags,
273 locked, cc);
274 if (!locked)
275 break;
276
277 /* Recheck this is a suitable migration target under lock */
278 if (!strict && !suitable_migration_target(page))
279 break;
280
281 /* Recheck this is a buddy page under lock */
282 if (!PageBuddy(page))
Laura Abbott2af120b2014-03-10 15:49:44 -0700283 goto isolate_fail;
Mel Gorman748446b2010-05-24 14:32:27 -0700284
285 /* Found a free page, break it into order-0 pages */
286 isolated = split_free_page(page);
287 total_isolated += isolated;
288 for (i = 0; i < isolated; i++) {
289 list_add(&page->lru, freelist);
290 page++;
291 }
292
293 /* If a page was split, advance to the end of it */
294 if (isolated) {
295 blockpfn += isolated - 1;
296 cursor += isolated - 1;
Laura Abbott2af120b2014-03-10 15:49:44 -0700297 continue;
Mel Gorman748446b2010-05-24 14:32:27 -0700298 }
Laura Abbott2af120b2014-03-10 15:49:44 -0700299
300isolate_fail:
301 if (strict)
302 break;
303 else
304 continue;
305
Mel Gorman748446b2010-05-24 14:32:27 -0700306 }
307
Mel Gormanb7aba692011-01-13 15:45:54 -0800308 trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
Mel Gormanf40d1e42012-10-08 16:32:36 -0700309
310 /*
311 * If strict isolation is requested by CMA then check that all the
312 * pages requested were isolated. If there were any failures, 0 is
313 * returned and CMA will fail.
314 */
Laura Abbott2af120b2014-03-10 15:49:44 -0700315 if (strict && blockpfn < end_pfn)
Mel Gormanf40d1e42012-10-08 16:32:36 -0700316 total_isolated = 0;
317
318 if (locked)
319 spin_unlock_irqrestore(&cc->zone->lock, flags);
320
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700321 /* Update the pageblock-skip if the whole pageblock was scanned */
322 if (blockpfn == end_pfn)
Mel Gormanc89511a2012-10-08 16:32:45 -0700323 update_pageblock_skip(cc, valid_page, total_isolated, false);
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700324
Minchan Kim010fc292012-12-20 15:05:06 -0800325 count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
Mel Gorman397487d2012-10-19 12:00:10 +0100326 if (total_isolated)
Minchan Kim010fc292012-12-20 15:05:06 -0800327 count_compact_events(COMPACTISOLATED, total_isolated);
Mel Gorman748446b2010-05-24 14:32:27 -0700328 return total_isolated;
329}
330
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100331/**
332 * isolate_freepages_range() - isolate free pages.
333 * @start_pfn: The first PFN to start isolating.
334 * @end_pfn: The one-past-last PFN.
335 *
336 * Non-free pages, invalid PFNs, or zone boundaries within the
337 * [start_pfn, end_pfn) range are considered errors, cause function to
338 * undo its actions and return zero.
339 *
340 * Otherwise, function returns one-past-the-last PFN of isolated page
341 * (which may be greater then end_pfn if end fell in a middle of
342 * a free page).
343 */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100344unsigned long
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700345isolate_freepages_range(struct compact_control *cc,
346 unsigned long start_pfn, unsigned long end_pfn)
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100347{
Mel Gormanf40d1e42012-10-08 16:32:36 -0700348 unsigned long isolated, pfn, block_end_pfn;
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100349 LIST_HEAD(freelist);
350
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100351 for (pfn = start_pfn; pfn < end_pfn; pfn += isolated) {
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700352 if (!pfn_valid(pfn) || cc->zone != page_zone(pfn_to_page(pfn)))
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100353 break;
354
355 /*
356 * On subsequent iterations ALIGN() is actually not needed,
357 * but we keep it that we not to complicate the code.
358 */
359 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
360 block_end_pfn = min(block_end_pfn, end_pfn);
361
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700362 isolated = isolate_freepages_block(cc, pfn, block_end_pfn,
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100363 &freelist, true);
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100364
365 /*
366 * In strict mode, isolate_freepages_block() returns 0 if
367 * there are any holes in the block (ie. invalid PFNs or
368 * non-free pages).
369 */
370 if (!isolated)
371 break;
372
373 /*
374 * If we managed to isolate pages, it is always (1 << n) *
375 * pageblock_nr_pages for some non-negative n. (Max order
376 * page may span two pageblocks).
377 */
378 }
379
380 /* split_free_page does not map the pages */
381 map_pages(&freelist);
382
383 if (pfn < end_pfn) {
384 /* Loop terminated early, cleanup. */
385 release_freepages(&freelist);
386 return 0;
387 }
388
389 /* We don't use freelists for anything. */
390 return pfn;
391}
392
Mel Gorman748446b2010-05-24 14:32:27 -0700393/* Update the number of anon and file isolated pages in the zone */
Mel Gormanc67fe372012-08-21 16:16:17 -0700394static void acct_isolated(struct zone *zone, bool locked, struct compact_control *cc)
Mel Gorman748446b2010-05-24 14:32:27 -0700395{
396 struct page *page;
Minchan Kimb9e84ac2011-10-31 17:06:44 -0700397 unsigned int count[2] = { 0, };
Mel Gorman748446b2010-05-24 14:32:27 -0700398
Minchan Kimb9e84ac2011-10-31 17:06:44 -0700399 list_for_each_entry(page, &cc->migratepages, lru)
400 count[!!page_is_file_cache(page)]++;
Mel Gorman748446b2010-05-24 14:32:27 -0700401
Mel Gormanc67fe372012-08-21 16:16:17 -0700402 /* If locked we can use the interrupt unsafe versions */
403 if (locked) {
404 __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
405 __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
406 } else {
407 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
408 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
409 }
Mel Gorman748446b2010-05-24 14:32:27 -0700410}
411
412/* Similar to reclaim, but different enough that they don't share logic */
413static bool too_many_isolated(struct zone *zone)
414{
Minchan Kimbc693042010-09-09 16:38:00 -0700415 unsigned long active, inactive, isolated;
Mel Gorman748446b2010-05-24 14:32:27 -0700416
417 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
418 zone_page_state(zone, NR_INACTIVE_ANON);
Minchan Kimbc693042010-09-09 16:38:00 -0700419 active = zone_page_state(zone, NR_ACTIVE_FILE) +
420 zone_page_state(zone, NR_ACTIVE_ANON);
Mel Gorman748446b2010-05-24 14:32:27 -0700421 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
422 zone_page_state(zone, NR_ISOLATED_ANON);
423
Minchan Kimbc693042010-09-09 16:38:00 -0700424 return isolated > (inactive + active) / 2;
Mel Gorman748446b2010-05-24 14:32:27 -0700425}
426
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100427/**
428 * isolate_migratepages_range() - isolate all migrate-able pages in range.
429 * @zone: Zone pages are in.
430 * @cc: Compaction control structure.
431 * @low_pfn: The first PFN of the range.
432 * @end_pfn: The one-past-the-last PFN of the range.
Minchan Kime46a2872012-10-08 16:33:48 -0700433 * @unevictable: true if it allows to isolate unevictable pages
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100434 *
435 * Isolate all pages that can be migrated from the range specified by
436 * [low_pfn, end_pfn). Returns zero if there is a fatal signal
437 * pending), otherwise PFN of the first page that was not scanned
438 * (which may be both less, equal to or more then end_pfn).
439 *
440 * Assumes that cc->migratepages is empty and cc->nr_migratepages is
441 * zero.
442 *
443 * Apart from cc->migratepages and cc->nr_migratetypes this function
444 * does not modify any cc's fields, in particular it does not modify
445 * (or read for that matter) cc->migrate_pfn.
Mel Gorman748446b2010-05-24 14:32:27 -0700446 */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100447unsigned long
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100448isolate_migratepages_range(struct zone *zone, struct compact_control *cc,
Minchan Kime46a2872012-10-08 16:33:48 -0700449 unsigned long low_pfn, unsigned long end_pfn, bool unevictable)
Mel Gorman748446b2010-05-24 14:32:27 -0700450{
Mel Gorman9927af742011-01-13 15:45:59 -0800451 unsigned long last_pageblock_nr = 0, pageblock_nr;
Mel Gormanb7aba692011-01-13 15:45:54 -0800452 unsigned long nr_scanned = 0, nr_isolated = 0;
Mel Gorman748446b2010-05-24 14:32:27 -0700453 struct list_head *migratelist = &cc->migratepages;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -0700454 isolate_mode_t mode = 0;
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700455 struct lruvec *lruvec;
Mel Gormanc67fe372012-08-21 16:16:17 -0700456 unsigned long flags;
Mel Gorman2a1402a2012-10-08 16:32:33 -0700457 bool locked = false;
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700458 struct page *page = NULL, *valid_page = NULL;
Vlastimil Babka50b5b092014-01-21 15:51:10 -0800459 bool skipped_async_unsuitable = false;
Mel Gorman748446b2010-05-24 14:32:27 -0700460
Mel Gorman748446b2010-05-24 14:32:27 -0700461 /*
462 * Ensure that there are not too many pages isolated from the LRU
463 * list by either parallel reclaimers or compaction. If there are,
464 * delay for some time until fewer pages are isolated
465 */
466 while (unlikely(too_many_isolated(zone))) {
Mel Gormanf9e35b32011-06-15 15:08:52 -0700467 /* async migration should just abort */
Linus Torvalds68e3e922012-06-03 20:05:57 -0700468 if (!cc->sync)
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100469 return 0;
Mel Gormanf9e35b32011-06-15 15:08:52 -0700470
Mel Gorman748446b2010-05-24 14:32:27 -0700471 congestion_wait(BLK_RW_ASYNC, HZ/10);
472
473 if (fatal_signal_pending(current))
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100474 return 0;
Mel Gorman748446b2010-05-24 14:32:27 -0700475 }
476
477 /* Time to isolate some pages for migration */
Andrea Arcangelib2eef8c2011-03-22 16:33:10 -0700478 cond_resched();
Mel Gorman748446b2010-05-24 14:32:27 -0700479 for (; low_pfn < end_pfn; low_pfn++) {
Andrea Arcangelib2eef8c2011-03-22 16:33:10 -0700480 /* give a chance to irqs before checking need_resched() */
Mel Gorman2a1402a2012-10-08 16:32:33 -0700481 if (locked && !((low_pfn+1) % SWAP_CLUSTER_MAX)) {
482 if (should_release_lock(&zone->lru_lock)) {
483 spin_unlock_irqrestore(&zone->lru_lock, flags);
484 locked = false;
485 }
Andrea Arcangelib2eef8c2011-03-22 16:33:10 -0700486 }
Mel Gormanc67fe372012-08-21 16:16:17 -0700487
Mel Gorman0bf380b2012-02-03 15:37:18 -0800488 /*
489 * migrate_pfn does not necessarily start aligned to a
490 * pageblock. Ensure that pfn_valid is called when moving
491 * into a new MAX_ORDER_NR_PAGES range in case of large
492 * memory holes within the zone
493 */
494 if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
495 if (!pfn_valid(low_pfn)) {
496 low_pfn += MAX_ORDER_NR_PAGES - 1;
497 continue;
498 }
499 }
500
Mel Gorman748446b2010-05-24 14:32:27 -0700501 if (!pfn_valid_within(low_pfn))
502 continue;
Mel Gormanb7aba692011-01-13 15:45:54 -0800503 nr_scanned++;
Mel Gorman748446b2010-05-24 14:32:27 -0700504
Mel Gormandc908602012-02-08 17:13:38 -0800505 /*
506 * Get the page and ensure the page is within the same zone.
507 * See the comment in isolate_freepages about overlapping
508 * nodes. It is deliberate that the new zone lock is not taken
509 * as memory compaction should not move pages between nodes.
510 */
Mel Gorman748446b2010-05-24 14:32:27 -0700511 page = pfn_to_page(low_pfn);
Mel Gormandc908602012-02-08 17:13:38 -0800512 if (page_zone(page) != zone)
513 continue;
514
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700515 if (!valid_page)
516 valid_page = page;
517
518 /* If isolation recently failed, do not retry */
519 pageblock_nr = low_pfn >> pageblock_order;
520 if (!isolation_suitable(cc, page))
521 goto next_pageblock;
522
Mel Gorman6c144662014-01-23 15:53:38 -0800523 /*
524 * Skip if free. page_order cannot be used without zone->lock
525 * as nothing prevents parallel allocations or buddy merging.
526 */
Mel Gorman748446b2010-05-24 14:32:27 -0700527 if (PageBuddy(page))
528 continue;
529
Mel Gorman9927af742011-01-13 15:45:59 -0800530 /*
531 * For async migration, also only scan in MOVABLE blocks. Async
532 * migration is optimistic to see if the minimum amount of work
533 * satisfies the allocation
534 */
Linus Torvalds68e3e922012-06-03 20:05:57 -0700535 if (!cc->sync && last_pageblock_nr != pageblock_nr &&
Michal Nazarewicz47118af2011-12-29 13:09:50 +0100536 !migrate_async_suitable(get_pageblock_migratetype(page))) {
Mel Gormanc89511a2012-10-08 16:32:45 -0700537 cc->finished_update_migrate = true;
Vlastimil Babka50b5b092014-01-21 15:51:10 -0800538 skipped_async_unsuitable = true;
Mel Gorman2a1402a2012-10-08 16:32:33 -0700539 goto next_pageblock;
Mel Gorman9927af742011-01-13 15:45:59 -0800540 }
541
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800542 /*
543 * Check may be lockless but that's ok as we recheck later.
544 * It's possible to migrate LRU pages and balloon pages
545 * Skip any other type of page
546 */
547 if (!PageLRU(page)) {
548 if (unlikely(balloon_page_movable(page))) {
549 if (locked && balloon_page_isolate(page)) {
550 /* Successfully isolated */
551 cc->finished_update_migrate = true;
552 list_add(&page->lru, migratelist);
553 cc->nr_migratepages++;
554 nr_isolated++;
555 goto check_compact_cluster;
556 }
557 }
Andrea Arcangelibc835012011-01-13 15:47:08 -0800558 continue;
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800559 }
Andrea Arcangelibc835012011-01-13 15:47:08 -0800560
561 /*
Mel Gorman2a1402a2012-10-08 16:32:33 -0700562 * PageLRU is set. lru_lock normally excludes isolation
563 * splitting and collapsing (collapsing has already happened
564 * if PageLRU is set) but the lock is not necessarily taken
565 * here and it is wasteful to take it just to check transhuge.
566 * Check TransHuge without lock and skip the whole pageblock if
567 * it's either a transhuge or hugetlbfs page, as calling
568 * compound_order() without preventing THP from splitting the
569 * page underneath us may return surprising results.
Andrea Arcangelibc835012011-01-13 15:47:08 -0800570 */
571 if (PageTransHuge(page)) {
Mel Gorman2a1402a2012-10-08 16:32:33 -0700572 if (!locked)
573 goto next_pageblock;
574 low_pfn += (1 << compound_order(page)) - 1;
575 continue;
576 }
577
David Rientjes119d6d52014-04-03 14:48:00 -0700578 /*
579 * Migration will fail if an anonymous page is pinned in memory,
580 * so avoid taking lru_lock and isolating it unnecessarily in an
581 * admittedly racy check.
582 */
583 if (!page_mapping(page) &&
584 page_count(page) > page_mapcount(page))
585 continue;
586
Mel Gorman2a1402a2012-10-08 16:32:33 -0700587 /* Check if it is ok to still hold the lock */
588 locked = compact_checklock_irqsave(&zone->lru_lock, &flags,
589 locked, cc);
590 if (!locked || fatal_signal_pending(current))
591 break;
592
593 /* Recheck PageLRU and PageTransHuge under lock */
594 if (!PageLRU(page))
595 continue;
596 if (PageTransHuge(page)) {
Andrea Arcangelibc835012011-01-13 15:47:08 -0800597 low_pfn += (1 << compound_order(page)) - 1;
598 continue;
599 }
600
Linus Torvalds68e3e922012-06-03 20:05:57 -0700601 if (!cc->sync)
Mel Gormanc8244932012-01-12 17:19:38 -0800602 mode |= ISOLATE_ASYNC_MIGRATE;
603
Minchan Kime46a2872012-10-08 16:33:48 -0700604 if (unevictable)
605 mode |= ISOLATE_UNEVICTABLE;
606
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700607 lruvec = mem_cgroup_page_lruvec(page, zone);
608
Mel Gorman748446b2010-05-24 14:32:27 -0700609 /* Try isolate the page */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -0700610 if (__isolate_lru_page(page, mode) != 0)
Mel Gorman748446b2010-05-24 14:32:27 -0700611 continue;
612
Sasha Levin309381fea2014-01-23 15:52:54 -0800613 VM_BUG_ON_PAGE(PageTransCompound(page), page);
Andrea Arcangelibc835012011-01-13 15:47:08 -0800614
Mel Gorman748446b2010-05-24 14:32:27 -0700615 /* Successfully isolated */
Mel Gormanc89511a2012-10-08 16:32:45 -0700616 cc->finished_update_migrate = true;
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700617 del_page_from_lru_list(page, lruvec, page_lru(page));
Mel Gorman748446b2010-05-24 14:32:27 -0700618 list_add(&page->lru, migratelist);
Mel Gorman748446b2010-05-24 14:32:27 -0700619 cc->nr_migratepages++;
Mel Gormanb7aba692011-01-13 15:45:54 -0800620 nr_isolated++;
Mel Gorman748446b2010-05-24 14:32:27 -0700621
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800622check_compact_cluster:
Mel Gorman748446b2010-05-24 14:32:27 -0700623 /* Avoid isolating too much */
Hillf Danton31b83842012-01-10 15:07:59 -0800624 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
625 ++low_pfn;
Mel Gorman748446b2010-05-24 14:32:27 -0700626 break;
Hillf Danton31b83842012-01-10 15:07:59 -0800627 }
Mel Gorman2a1402a2012-10-08 16:32:33 -0700628
629 continue;
630
631next_pageblock:
Mel Gormana9aacbc2013-02-22 16:32:25 -0800632 low_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages) - 1;
Mel Gorman2a1402a2012-10-08 16:32:33 -0700633 last_pageblock_nr = pageblock_nr;
Mel Gorman748446b2010-05-24 14:32:27 -0700634 }
635
Mel Gormanc67fe372012-08-21 16:16:17 -0700636 acct_isolated(zone, locked, cc);
Mel Gorman748446b2010-05-24 14:32:27 -0700637
Mel Gormanc67fe372012-08-21 16:16:17 -0700638 if (locked)
639 spin_unlock_irqrestore(&zone->lru_lock, flags);
Mel Gorman748446b2010-05-24 14:32:27 -0700640
Vlastimil Babka50b5b092014-01-21 15:51:10 -0800641 /*
642 * Update the pageblock-skip information and cached scanner pfn,
643 * if the whole pageblock was scanned without isolating any page.
644 * This is not done when pageblock was skipped due to being unsuitable
645 * for async compaction, so that eventual sync compaction can try.
646 */
647 if (low_pfn == end_pfn && !skipped_async_unsuitable)
Mel Gormanc89511a2012-10-08 16:32:45 -0700648 update_pageblock_skip(cc, valid_page, nr_isolated, true);
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700649
Mel Gormanb7aba692011-01-13 15:45:54 -0800650 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
651
Minchan Kim010fc292012-12-20 15:05:06 -0800652 count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
Mel Gorman397487d2012-10-19 12:00:10 +0100653 if (nr_isolated)
Minchan Kim010fc292012-12-20 15:05:06 -0800654 count_compact_events(COMPACTISOLATED, nr_isolated);
Mel Gorman397487d2012-10-19 12:00:10 +0100655
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100656 return low_pfn;
657}
658
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100659#endif /* CONFIG_COMPACTION || CONFIG_CMA */
660#ifdef CONFIG_COMPACTION
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100661/*
662 * Based on information in the current compact_control, find blocks
663 * suitable for isolating free pages from and then isolate them.
664 */
665static void isolate_freepages(struct zone *zone,
666 struct compact_control *cc)
667{
668 struct page *page;
Cody P Schafer108bcc92013-02-22 16:35:23 -0800669 unsigned long high_pfn, low_pfn, pfn, z_end_pfn, end_pfn;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100670 int nr_freepages = cc->nr_freepages;
671 struct list_head *freelist = &cc->freepages;
672
673 /*
674 * Initialise the free scanner. The starting point is where we last
675 * scanned from (or the end of the zone if starting). The low point
676 * is the end of the pageblock the migration scanner is using.
677 */
678 pfn = cc->free_pfn;
Vlastimil Babka7ed695e2014-01-21 15:51:09 -0800679 low_pfn = ALIGN(cc->migrate_pfn + 1, pageblock_nr_pages);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100680
681 /*
682 * Take care that if the migration scanner is at the end of the zone
683 * that the free scanner does not accidentally move to the next zone
684 * in the next isolation cycle.
685 */
686 high_pfn = min(low_pfn, pfn);
687
Cody P Schafer108bcc92013-02-22 16:35:23 -0800688 z_end_pfn = zone_end_pfn(zone);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100689
690 /*
691 * Isolate free pages until enough are available to migrate the
692 * pages on cc->migratepages. We stop searching if the migrate
693 * and free page scanners meet or enough free pages are isolated.
694 */
Vlastimil Babka7ed695e2014-01-21 15:51:09 -0800695 for (; pfn >= low_pfn && cc->nr_migratepages > nr_freepages;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100696 pfn -= pageblock_nr_pages) {
697 unsigned long isolated;
698
David Rientjesf6ea3ad2013-09-30 13:45:03 -0700699 /*
700 * This can iterate a massively long zone without finding any
701 * suitable migration targets, so periodically check if we need
702 * to schedule.
703 */
704 cond_resched();
705
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100706 if (!pfn_valid(pfn))
707 continue;
708
709 /*
710 * Check for overlapping nodes/zones. It's possible on some
711 * configurations to have a setup like
712 * node0 node1 node0
713 * i.e. it's possible that all pages within a zones range of
714 * pages do not belong to a single zone.
715 */
716 page = pfn_to_page(pfn);
717 if (page_zone(page) != zone)
718 continue;
719
720 /* Check the block is suitable for migration */
Linus Torvalds68e3e922012-06-03 20:05:57 -0700721 if (!suitable_migration_target(page))
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100722 continue;
Linus Torvalds68e3e922012-06-03 20:05:57 -0700723
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700724 /* If isolation recently failed, do not retry */
725 if (!isolation_suitable(cc, page))
726 continue;
727
Mel Gormanf40d1e42012-10-08 16:32:36 -0700728 /* Found a block suitable for isolating free pages from */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100729 isolated = 0;
Mel Gorman60177d32012-12-06 19:01:14 +0000730
731 /*
732 * As pfn may not start aligned, pfn+pageblock_nr_page
733 * may cross a MAX_ORDER_NR_PAGES boundary and miss
734 * a pfn_valid check. Ensure isolate_freepages_block()
735 * only scans within a pageblock
736 */
737 end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
Cody P Schafer108bcc92013-02-22 16:35:23 -0800738 end_pfn = min(end_pfn, z_end_pfn);
Mel Gormanf40d1e42012-10-08 16:32:36 -0700739 isolated = isolate_freepages_block(cc, pfn, end_pfn,
740 freelist, false);
741 nr_freepages += isolated;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100742
743 /*
744 * Record the highest PFN we isolated pages from. When next
745 * looking for free pages, the search will restart here as
746 * page migration may have returned some pages to the allocator
747 */
Mel Gormanc89511a2012-10-08 16:32:45 -0700748 if (isolated) {
749 cc->finished_update_free = true;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100750 high_pfn = max(high_pfn, pfn);
Mel Gormanc89511a2012-10-08 16:32:45 -0700751 }
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100752 }
753
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100754 /* split_free_page does not map the pages */
755 map_pages(freelist);
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100756
Vlastimil Babka7ed695e2014-01-21 15:51:09 -0800757 /*
758 * If we crossed the migrate scanner, we want to keep it that way
759 * so that compact_finished() may detect this
760 */
761 if (pfn < low_pfn)
762 cc->free_pfn = max(pfn, zone->zone_start_pfn);
763 else
764 cc->free_pfn = high_pfn;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100765 cc->nr_freepages = nr_freepages;
Mel Gorman748446b2010-05-24 14:32:27 -0700766}
767
768/*
769 * This is a migrate-callback that "allocates" freepages by taking pages
770 * from the isolated freelists in the block we are migrating to.
771 */
772static struct page *compaction_alloc(struct page *migratepage,
773 unsigned long data,
774 int **result)
775{
776 struct compact_control *cc = (struct compact_control *)data;
777 struct page *freepage;
778
779 /* Isolate free pages if necessary */
780 if (list_empty(&cc->freepages)) {
781 isolate_freepages(cc->zone, cc);
782
783 if (list_empty(&cc->freepages))
784 return NULL;
785 }
786
787 freepage = list_entry(cc->freepages.next, struct page, lru);
788 list_del(&freepage->lru);
789 cc->nr_freepages--;
790
791 return freepage;
792}
793
794/*
795 * We cannot control nr_migratepages and nr_freepages fully when migration is
796 * running as migrate_pages() has no knowledge of compact_control. When
797 * migration is complete, we count the number of pages on the lists by hand.
798 */
799static void update_nr_listpages(struct compact_control *cc)
800{
801 int nr_migratepages = 0;
802 int nr_freepages = 0;
803 struct page *page;
804
805 list_for_each_entry(page, &cc->migratepages, lru)
806 nr_migratepages++;
807 list_for_each_entry(page, &cc->freepages, lru)
808 nr_freepages++;
809
810 cc->nr_migratepages = nr_migratepages;
811 cc->nr_freepages = nr_freepages;
812}
813
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100814/* possible outcome of isolate_migratepages */
815typedef enum {
816 ISOLATE_ABORT, /* Abort compaction now */
817 ISOLATE_NONE, /* No pages isolated, continue scanning */
818 ISOLATE_SUCCESS, /* Pages isolated, migrate */
819} isolate_migrate_t;
820
821/*
822 * Isolate all pages that can be migrated from the block pointed to by
823 * the migrate scanner within compact_control.
824 */
825static isolate_migrate_t isolate_migratepages(struct zone *zone,
826 struct compact_control *cc)
827{
828 unsigned long low_pfn, end_pfn;
829
830 /* Do not scan outside zone boundaries */
831 low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
832
833 /* Only scan within a pageblock boundary */
Mel Gormana9aacbc2013-02-22 16:32:25 -0800834 end_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100835
836 /* Do not cross the free scanner or scan within a memory hole */
837 if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
838 cc->migrate_pfn = end_pfn;
839 return ISOLATE_NONE;
840 }
841
842 /* Perform the isolation */
Minchan Kime46a2872012-10-08 16:33:48 -0700843 low_pfn = isolate_migratepages_range(zone, cc, low_pfn, end_pfn, false);
Shaohua Lie64c5232012-10-08 16:32:27 -0700844 if (!low_pfn || cc->contended)
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100845 return ISOLATE_ABORT;
846
847 cc->migrate_pfn = low_pfn;
848
849 return ISOLATE_SUCCESS;
850}
851
Mel Gorman748446b2010-05-24 14:32:27 -0700852static int compact_finished(struct zone *zone,
Andrea Arcangeli5a03b052011-01-13 15:47:11 -0800853 struct compact_control *cc)
Mel Gorman748446b2010-05-24 14:32:27 -0700854{
Mel Gorman8fb74b92013-01-11 14:32:16 -0800855 unsigned int order;
Andrea Arcangeli5a03b052011-01-13 15:47:11 -0800856 unsigned long watermark;
Mel Gorman56de7262010-05-24 14:32:30 -0700857
Mel Gorman748446b2010-05-24 14:32:27 -0700858 if (fatal_signal_pending(current))
859 return COMPACT_PARTIAL;
860
Mel Gorman753341a2012-10-08 16:32:40 -0700861 /* Compaction run completes if the migrate and free scanner meet */
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700862 if (cc->free_pfn <= cc->migrate_pfn) {
Vlastimil Babka55b7c4c2014-01-21 15:51:11 -0800863 /* Let the next compaction start anew. */
864 zone->compact_cached_migrate_pfn = zone->zone_start_pfn;
865 zone->compact_cached_free_pfn = zone_end_pfn(zone);
866
Mel Gorman62997022012-10-08 16:32:47 -0700867 /*
868 * Mark that the PG_migrate_skip information should be cleared
869 * by kswapd when it goes to sleep. kswapd does not set the
870 * flag itself as the decision to be clear should be directly
871 * based on an allocation request.
872 */
873 if (!current_is_kswapd())
874 zone->compact_blockskip_flush = true;
875
Mel Gorman748446b2010-05-24 14:32:27 -0700876 return COMPACT_COMPLETE;
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700877 }
Mel Gorman748446b2010-05-24 14:32:27 -0700878
Johannes Weiner82478fb2011-01-20 14:44:21 -0800879 /*
880 * order == -1 is expected when compacting via
881 * /proc/sys/vm/compact_memory
882 */
Mel Gorman56de7262010-05-24 14:32:30 -0700883 if (cc->order == -1)
884 return COMPACT_CONTINUE;
885
Michal Hocko3957c772011-06-15 15:08:25 -0700886 /* Compaction run is not finished if the watermark is not met */
887 watermark = low_wmark_pages(zone);
888 watermark += (1 << cc->order);
889
890 if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
891 return COMPACT_CONTINUE;
892
Mel Gorman56de7262010-05-24 14:32:30 -0700893 /* Direct compactor: Is a suitable page free? */
Mel Gorman8fb74b92013-01-11 14:32:16 -0800894 for (order = cc->order; order < MAX_ORDER; order++) {
895 struct free_area *area = &zone->free_area[order];
Mel Gorman56de7262010-05-24 14:32:30 -0700896
Mel Gorman8fb74b92013-01-11 14:32:16 -0800897 /* Job done if page is free of the right migratetype */
898 if (!list_empty(&area->free_list[cc->migratetype]))
899 return COMPACT_PARTIAL;
900
901 /* Job done if allocation would set block type */
902 if (cc->order >= pageblock_order && area->nr_free)
903 return COMPACT_PARTIAL;
Mel Gorman56de7262010-05-24 14:32:30 -0700904 }
905
Mel Gorman748446b2010-05-24 14:32:27 -0700906 return COMPACT_CONTINUE;
907}
908
Mel Gorman3e7d3442011-01-13 15:45:56 -0800909/*
910 * compaction_suitable: Is this suitable to run compaction on this zone now?
911 * Returns
912 * COMPACT_SKIPPED - If there are too few free pages for compaction
913 * COMPACT_PARTIAL - If the allocation would succeed without compaction
914 * COMPACT_CONTINUE - If compaction should run now
915 */
916unsigned long compaction_suitable(struct zone *zone, int order)
917{
918 int fragindex;
919 unsigned long watermark;
920
921 /*
Michal Hocko3957c772011-06-15 15:08:25 -0700922 * order == -1 is expected when compacting via
923 * /proc/sys/vm/compact_memory
924 */
925 if (order == -1)
926 return COMPACT_CONTINUE;
927
928 /*
Mel Gorman3e7d3442011-01-13 15:45:56 -0800929 * Watermarks for order-0 must be met for compaction. Note the 2UL.
930 * This is because during migration, copies of pages need to be
931 * allocated and for a short time, the footprint is higher
932 */
933 watermark = low_wmark_pages(zone) + (2UL << order);
934 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
935 return COMPACT_SKIPPED;
936
937 /*
938 * fragmentation index determines if allocation failures are due to
939 * low memory or external fragmentation
940 *
Shaohua Lia582a732011-06-15 15:08:49 -0700941 * index of -1000 implies allocations might succeed depending on
942 * watermarks
Mel Gorman3e7d3442011-01-13 15:45:56 -0800943 * index towards 0 implies failure is due to lack of memory
944 * index towards 1000 implies failure is due to fragmentation
945 *
946 * Only compact if a failure would be due to fragmentation.
947 */
948 fragindex = fragmentation_index(zone, order);
949 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
950 return COMPACT_SKIPPED;
951
Shaohua Lia582a732011-06-15 15:08:49 -0700952 if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
953 0, 0))
Mel Gorman3e7d3442011-01-13 15:45:56 -0800954 return COMPACT_PARTIAL;
955
956 return COMPACT_CONTINUE;
957}
958
Mel Gorman748446b2010-05-24 14:32:27 -0700959static int compact_zone(struct zone *zone, struct compact_control *cc)
960{
961 int ret;
Mel Gormanc89511a2012-10-08 16:32:45 -0700962 unsigned long start_pfn = zone->zone_start_pfn;
Cody P Schafer108bcc92013-02-22 16:35:23 -0800963 unsigned long end_pfn = zone_end_pfn(zone);
Mel Gorman748446b2010-05-24 14:32:27 -0700964
Mel Gorman3e7d3442011-01-13 15:45:56 -0800965 ret = compaction_suitable(zone, cc->order);
966 switch (ret) {
967 case COMPACT_PARTIAL:
968 case COMPACT_SKIPPED:
969 /* Compaction is likely to fail */
970 return ret;
971 case COMPACT_CONTINUE:
972 /* Fall through to compaction */
973 ;
974 }
975
Mel Gormanc89511a2012-10-08 16:32:45 -0700976 /*
Vlastimil Babkad3132e42014-01-21 15:51:08 -0800977 * Clear pageblock skip if there were failures recently and compaction
978 * is about to be retried after being deferred. kswapd does not do
979 * this reset as it'll reset the cached information when going to sleep.
980 */
981 if (compaction_restarting(zone, cc->order) && !current_is_kswapd())
982 __reset_isolation_suitable(zone);
983
984 /*
Mel Gormanc89511a2012-10-08 16:32:45 -0700985 * Setup to move all movable pages to the end of the zone. Used cached
986 * information on where the scanners should start but check that it
987 * is initialised by ensuring the values are within zone boundaries.
988 */
989 cc->migrate_pfn = zone->compact_cached_migrate_pfn;
990 cc->free_pfn = zone->compact_cached_free_pfn;
991 if (cc->free_pfn < start_pfn || cc->free_pfn > end_pfn) {
992 cc->free_pfn = end_pfn & ~(pageblock_nr_pages-1);
993 zone->compact_cached_free_pfn = cc->free_pfn;
994 }
995 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn > end_pfn) {
996 cc->migrate_pfn = start_pfn;
997 zone->compact_cached_migrate_pfn = cc->migrate_pfn;
998 }
Mel Gorman748446b2010-05-24 14:32:27 -0700999
Mel Gorman0eb927c2014-01-21 15:51:05 -08001000 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn, cc->free_pfn, end_pfn);
1001
Mel Gorman748446b2010-05-24 14:32:27 -07001002 migrate_prep_local();
1003
1004 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
1005 unsigned long nr_migrate, nr_remaining;
Minchan Kim9d502c12011-03-22 16:30:39 -07001006 int err;
Mel Gorman748446b2010-05-24 14:32:27 -07001007
Mel Gormanf9e35b32011-06-15 15:08:52 -07001008 switch (isolate_migratepages(zone, cc)) {
1009 case ISOLATE_ABORT:
1010 ret = COMPACT_PARTIAL;
Rafael Aquini5733c7d2012-12-11 16:02:47 -08001011 putback_movable_pages(&cc->migratepages);
Shaohua Lie64c5232012-10-08 16:32:27 -07001012 cc->nr_migratepages = 0;
Mel Gormanf9e35b32011-06-15 15:08:52 -07001013 goto out;
1014 case ISOLATE_NONE:
Mel Gorman748446b2010-05-24 14:32:27 -07001015 continue;
Mel Gormanf9e35b32011-06-15 15:08:52 -07001016 case ISOLATE_SUCCESS:
1017 ;
1018 }
Mel Gorman748446b2010-05-24 14:32:27 -07001019
1020 nr_migrate = cc->nr_migratepages;
Minchan Kim9d502c12011-03-22 16:30:39 -07001021 err = migrate_pages(&cc->migratepages, compaction_alloc,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001022 (unsigned long)cc,
Mel Gorman7b2a2d42012-10-19 14:07:31 +01001023 cc->sync ? MIGRATE_SYNC_LIGHT : MIGRATE_ASYNC,
1024 MR_COMPACTION);
Mel Gorman748446b2010-05-24 14:32:27 -07001025 update_nr_listpages(cc);
1026 nr_remaining = cc->nr_migratepages;
1027
Mel Gormanb7aba692011-01-13 15:45:54 -08001028 trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
1029 nr_remaining);
Mel Gorman748446b2010-05-24 14:32:27 -07001030
Rafael Aquini5733c7d2012-12-11 16:02:47 -08001031 /* Release isolated pages not migrated */
Minchan Kim9d502c12011-03-22 16:30:39 -07001032 if (err) {
Rafael Aquini5733c7d2012-12-11 16:02:47 -08001033 putback_movable_pages(&cc->migratepages);
Mel Gorman748446b2010-05-24 14:32:27 -07001034 cc->nr_migratepages = 0;
Vlastimil Babka7ed695e2014-01-21 15:51:09 -08001035 /*
1036 * migrate_pages() may return -ENOMEM when scanners meet
1037 * and we want compact_finished() to detect it
1038 */
1039 if (err == -ENOMEM && cc->free_pfn > cc->migrate_pfn) {
David Rientjes4bf2bba2012-07-11 14:02:13 -07001040 ret = COMPACT_PARTIAL;
1041 goto out;
1042 }
Mel Gorman748446b2010-05-24 14:32:27 -07001043 }
Mel Gorman748446b2010-05-24 14:32:27 -07001044 }
1045
Mel Gormanf9e35b32011-06-15 15:08:52 -07001046out:
Mel Gorman748446b2010-05-24 14:32:27 -07001047 /* Release free pages and check accounting */
1048 cc->nr_freepages -= release_freepages(&cc->freepages);
1049 VM_BUG_ON(cc->nr_freepages != 0);
1050
Mel Gorman0eb927c2014-01-21 15:51:05 -08001051 trace_mm_compaction_end(ret);
1052
Mel Gorman748446b2010-05-24 14:32:27 -07001053 return ret;
1054}
Mel Gorman76ab0f52010-05-24 14:32:28 -07001055
Kyungmin Parkd43a87e2011-10-31 17:09:08 -07001056static unsigned long compact_zone_order(struct zone *zone,
Andrea Arcangeli5a03b052011-01-13 15:47:11 -08001057 int order, gfp_t gfp_mask,
Mel Gorman8fb74b92013-01-11 14:32:16 -08001058 bool sync, bool *contended)
Mel Gorman56de7262010-05-24 14:32:30 -07001059{
Shaohua Lie64c5232012-10-08 16:32:27 -07001060 unsigned long ret;
Mel Gorman56de7262010-05-24 14:32:30 -07001061 struct compact_control cc = {
1062 .nr_freepages = 0,
1063 .nr_migratepages = 0,
1064 .order = order,
1065 .migratetype = allocflags_to_migratetype(gfp_mask),
1066 .zone = zone,
Linus Torvalds68e3e922012-06-03 20:05:57 -07001067 .sync = sync,
Mel Gorman56de7262010-05-24 14:32:30 -07001068 };
1069 INIT_LIST_HEAD(&cc.freepages);
1070 INIT_LIST_HEAD(&cc.migratepages);
1071
Shaohua Lie64c5232012-10-08 16:32:27 -07001072 ret = compact_zone(zone, &cc);
1073
1074 VM_BUG_ON(!list_empty(&cc.freepages));
1075 VM_BUG_ON(!list_empty(&cc.migratepages));
1076
1077 *contended = cc.contended;
1078 return ret;
Mel Gorman56de7262010-05-24 14:32:30 -07001079}
1080
Mel Gorman5e771902010-05-24 14:32:31 -07001081int sysctl_extfrag_threshold = 500;
1082
Mel Gorman56de7262010-05-24 14:32:30 -07001083/**
1084 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1085 * @zonelist: The zonelist used for the current allocation
1086 * @order: The order of the current allocation
1087 * @gfp_mask: The GFP mask of the current allocation
1088 * @nodemask: The allowed nodes to allocate from
Mel Gorman77f1fe62011-01-13 15:45:57 -08001089 * @sync: Whether migration is synchronous or not
Mel Gorman661c4cb2012-10-08 16:32:31 -07001090 * @contended: Return value that is true if compaction was aborted due to lock contention
1091 * @page: Optionally capture a free page of the requested order during compaction
Mel Gorman56de7262010-05-24 14:32:30 -07001092 *
1093 * This is the main entry point for direct page compaction.
1094 */
1095unsigned long try_to_compact_pages(struct zonelist *zonelist,
Mel Gorman77f1fe62011-01-13 15:45:57 -08001096 int order, gfp_t gfp_mask, nodemask_t *nodemask,
Mel Gorman8fb74b92013-01-11 14:32:16 -08001097 bool sync, bool *contended)
Mel Gorman56de7262010-05-24 14:32:30 -07001098{
1099 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1100 int may_enter_fs = gfp_mask & __GFP_FS;
1101 int may_perform_io = gfp_mask & __GFP_IO;
Mel Gorman56de7262010-05-24 14:32:30 -07001102 struct zoneref *z;
1103 struct zone *zone;
1104 int rc = COMPACT_SKIPPED;
Bartlomiej Zolnierkiewiczd95ea5d2012-10-08 16:32:05 -07001105 int alloc_flags = 0;
Mel Gorman56de7262010-05-24 14:32:30 -07001106
Mel Gorman4ffb6332012-10-08 16:29:09 -07001107 /* Check if the GFP flags allow compaction */
Andrea Arcangelic5a73c32011-01-13 15:47:11 -08001108 if (!order || !may_enter_fs || !may_perform_io)
Mel Gorman56de7262010-05-24 14:32:30 -07001109 return rc;
1110
Minchan Kim010fc292012-12-20 15:05:06 -08001111 count_compact_event(COMPACTSTALL);
Mel Gorman56de7262010-05-24 14:32:30 -07001112
Bartlomiej Zolnierkiewiczd95ea5d2012-10-08 16:32:05 -07001113#ifdef CONFIG_CMA
1114 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
1115 alloc_flags |= ALLOC_CMA;
1116#endif
Mel Gorman56de7262010-05-24 14:32:30 -07001117 /* Compact each zone in the list */
1118 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1119 nodemask) {
Mel Gorman56de7262010-05-24 14:32:30 -07001120 int status;
1121
Mel Gormanc67fe372012-08-21 16:16:17 -07001122 status = compact_zone_order(zone, order, gfp_mask, sync,
Mel Gorman8fb74b92013-01-11 14:32:16 -08001123 contended);
Mel Gorman56de7262010-05-24 14:32:30 -07001124 rc = max(status, rc);
1125
Mel Gorman3e7d3442011-01-13 15:45:56 -08001126 /* If a normal allocation would succeed, stop compacting */
Bartlomiej Zolnierkiewiczd95ea5d2012-10-08 16:32:05 -07001127 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0,
1128 alloc_flags))
Mel Gorman56de7262010-05-24 14:32:30 -07001129 break;
1130 }
1131
1132 return rc;
1133}
1134
1135
Mel Gorman76ab0f52010-05-24 14:32:28 -07001136/* Compact all zones within a node */
Andrew Morton7103f162013-02-22 16:32:33 -08001137static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
Mel Gorman76ab0f52010-05-24 14:32:28 -07001138{
1139 int zoneid;
Mel Gorman76ab0f52010-05-24 14:32:28 -07001140 struct zone *zone;
1141
Mel Gorman76ab0f52010-05-24 14:32:28 -07001142 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
Mel Gorman76ab0f52010-05-24 14:32:28 -07001143
1144 zone = &pgdat->node_zones[zoneid];
1145 if (!populated_zone(zone))
1146 continue;
1147
Rik van Riel7be62de2012-03-21 16:33:52 -07001148 cc->nr_freepages = 0;
1149 cc->nr_migratepages = 0;
1150 cc->zone = zone;
1151 INIT_LIST_HEAD(&cc->freepages);
1152 INIT_LIST_HEAD(&cc->migratepages);
Mel Gorman76ab0f52010-05-24 14:32:28 -07001153
Dan Carpenteraad6ec32012-03-21 16:33:54 -07001154 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
Rik van Riel7be62de2012-03-21 16:33:52 -07001155 compact_zone(zone, cc);
Mel Gorman76ab0f52010-05-24 14:32:28 -07001156
Rik van Rielaff62242012-03-21 16:33:52 -07001157 if (cc->order > 0) {
Vlastimil Babkade6c60a2014-01-21 15:51:07 -08001158 if (zone_watermark_ok(zone, cc->order,
1159 low_wmark_pages(zone), 0, 0))
1160 compaction_defer_reset(zone, cc->order, false);
Rik van Rielaff62242012-03-21 16:33:52 -07001161 /* Currently async compaction is never deferred. */
Vlastimil Babkade6c60a2014-01-21 15:51:07 -08001162 else if (cc->sync)
Rik van Rielaff62242012-03-21 16:33:52 -07001163 defer_compaction(zone, cc->order);
1164 }
1165
Rik van Riel7be62de2012-03-21 16:33:52 -07001166 VM_BUG_ON(!list_empty(&cc->freepages));
1167 VM_BUG_ON(!list_empty(&cc->migratepages));
Mel Gorman76ab0f52010-05-24 14:32:28 -07001168 }
Mel Gorman76ab0f52010-05-24 14:32:28 -07001169}
1170
Andrew Morton7103f162013-02-22 16:32:33 -08001171void compact_pgdat(pg_data_t *pgdat, int order)
Rik van Riel7be62de2012-03-21 16:33:52 -07001172{
1173 struct compact_control cc = {
1174 .order = order,
Linus Torvalds68e3e922012-06-03 20:05:57 -07001175 .sync = false,
Rik van Riel7be62de2012-03-21 16:33:52 -07001176 };
1177
Mel Gorman3a7200a2013-09-11 14:22:19 -07001178 if (!order)
1179 return;
1180
Andrew Morton7103f162013-02-22 16:32:33 -08001181 __compact_pgdat(pgdat, &cc);
Rik van Riel7be62de2012-03-21 16:33:52 -07001182}
1183
Andrew Morton7103f162013-02-22 16:32:33 -08001184static void compact_node(int nid)
Rik van Riel7be62de2012-03-21 16:33:52 -07001185{
Rik van Riel7be62de2012-03-21 16:33:52 -07001186 struct compact_control cc = {
1187 .order = -1,
Linus Torvalds68e3e922012-06-03 20:05:57 -07001188 .sync = true,
David Rientjes91ca9182014-04-03 14:47:23 -07001189 .ignore_skip_hint = true,
Rik van Riel7be62de2012-03-21 16:33:52 -07001190 };
1191
Andrew Morton7103f162013-02-22 16:32:33 -08001192 __compact_pgdat(NODE_DATA(nid), &cc);
Rik van Riel7be62de2012-03-21 16:33:52 -07001193}
1194
Mel Gorman76ab0f52010-05-24 14:32:28 -07001195/* Compact all nodes in the system */
Jason Liu7964c062013-01-11 14:31:47 -08001196static void compact_nodes(void)
Mel Gorman76ab0f52010-05-24 14:32:28 -07001197{
1198 int nid;
1199
Hugh Dickins8575ec22012-03-21 16:33:53 -07001200 /* Flush pending updates to the LRU lists */
1201 lru_add_drain_all();
1202
Mel Gorman76ab0f52010-05-24 14:32:28 -07001203 for_each_online_node(nid)
1204 compact_node(nid);
Mel Gorman76ab0f52010-05-24 14:32:28 -07001205}
1206
1207/* The written value is actually unused, all memory is compacted */
1208int sysctl_compact_memory;
1209
1210/* This is the entry point for compacting all nodes via /proc/sys/vm */
1211int sysctl_compaction_handler(struct ctl_table *table, int write,
1212 void __user *buffer, size_t *length, loff_t *ppos)
1213{
1214 if (write)
Jason Liu7964c062013-01-11 14:31:47 -08001215 compact_nodes();
Mel Gorman76ab0f52010-05-24 14:32:28 -07001216
1217 return 0;
1218}
Mel Gormaned4a6d72010-05-24 14:32:29 -07001219
Mel Gorman5e771902010-05-24 14:32:31 -07001220int sysctl_extfrag_handler(struct ctl_table *table, int write,
1221 void __user *buffer, size_t *length, loff_t *ppos)
1222{
1223 proc_dointvec_minmax(table, write, buffer, length, ppos);
1224
1225 return 0;
1226}
1227
Mel Gormaned4a6d72010-05-24 14:32:29 -07001228#if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
Rashika Kheria74e77fb2014-04-03 14:48:01 -07001229static ssize_t sysfs_compact_node(struct device *dev,
Kay Sievers10fbcf42011-12-21 14:48:43 -08001230 struct device_attribute *attr,
Mel Gormaned4a6d72010-05-24 14:32:29 -07001231 const char *buf, size_t count)
1232{
Hugh Dickins8575ec22012-03-21 16:33:53 -07001233 int nid = dev->id;
1234
1235 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1236 /* Flush pending updates to the LRU lists */
1237 lru_add_drain_all();
1238
1239 compact_node(nid);
1240 }
Mel Gormaned4a6d72010-05-24 14:32:29 -07001241
1242 return count;
1243}
Kay Sievers10fbcf42011-12-21 14:48:43 -08001244static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
Mel Gormaned4a6d72010-05-24 14:32:29 -07001245
1246int compaction_register_node(struct node *node)
1247{
Kay Sievers10fbcf42011-12-21 14:48:43 -08001248 return device_create_file(&node->dev, &dev_attr_compact);
Mel Gormaned4a6d72010-05-24 14:32:29 -07001249}
1250
1251void compaction_unregister_node(struct node *node)
1252{
Kay Sievers10fbcf42011-12-21 14:48:43 -08001253 return device_remove_file(&node->dev, &dev_attr_compact);
Mel Gormaned4a6d72010-05-24 14:32:29 -07001254}
1255#endif /* CONFIG_SYSFS && CONFIG_NUMA */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001256
1257#endif /* CONFIG_COMPACTION */