blob: 56331f5124baf98da49905d1e5d5048dbba22180 [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
David Rientjes35979ef2014-06-04 16:08:27 -070092 zone->compact_cached_migrate_pfn[0] = start_pfn;
93 zone->compact_cached_migrate_pfn[1] = start_pfn;
Mel Gormanc89511a2012-10-08 16:32:45 -070094 zone->compact_cached_free_pfn = end_pfn;
Mel Gorman62997022012-10-08 16:32:47 -070095 zone->compact_blockskip_flush = false;
Mel Gormanbb13ffe2012-10-08 16:32:41 -070096
97 /* Walk the zone and mark every pageblock as suitable for isolation */
98 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
99 struct page *page;
100
101 cond_resched();
102
103 if (!pfn_valid(pfn))
104 continue;
105
106 page = pfn_to_page(pfn);
107 if (zone != page_zone(page))
108 continue;
109
110 clear_pageblock_skip(page);
111 }
112}
113
Mel Gorman62997022012-10-08 16:32:47 -0700114void reset_isolation_suitable(pg_data_t *pgdat)
115{
116 int zoneid;
117
118 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
119 struct zone *zone = &pgdat->node_zones[zoneid];
120 if (!populated_zone(zone))
121 continue;
122
123 /* Only flush if a full compaction finished recently */
124 if (zone->compact_blockskip_flush)
125 __reset_isolation_suitable(zone);
126 }
127}
128
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700129/*
130 * If no pages were isolated then mark this pageblock to be skipped in the
Mel Gorman62997022012-10-08 16:32:47 -0700131 * future. The information is later cleared by __reset_isolation_suitable().
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700132 */
Mel Gormanc89511a2012-10-08 16:32:45 -0700133static void update_pageblock_skip(struct compact_control *cc,
134 struct page *page, unsigned long nr_isolated,
David Rientjes35979ef2014-06-04 16:08:27 -0700135 bool set_unsuitable, bool migrate_scanner)
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700136{
Mel Gormanc89511a2012-10-08 16:32:45 -0700137 struct zone *zone = cc->zone;
David Rientjes35979ef2014-06-04 16:08:27 -0700138 unsigned long pfn;
Joonsoo Kim6815bf32013-12-18 17:08:52 -0800139
140 if (cc->ignore_skip_hint)
141 return;
142
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700143 if (!page)
144 return;
145
David Rientjes35979ef2014-06-04 16:08:27 -0700146 if (nr_isolated)
147 return;
148
149 /*
150 * Only skip pageblocks when all forms of compaction will be known to
151 * fail in the near future.
152 */
153 if (set_unsuitable)
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700154 set_pageblock_skip(page);
Mel Gormanc89511a2012-10-08 16:32:45 -0700155
David Rientjes35979ef2014-06-04 16:08:27 -0700156 pfn = page_to_pfn(page);
157
158 /* Update where async and sync compaction should restart */
159 if (migrate_scanner) {
160 if (cc->finished_update_migrate)
161 return;
162 if (pfn > zone->compact_cached_migrate_pfn[0])
163 zone->compact_cached_migrate_pfn[0] = pfn;
David Rientjese0b9dae2014-06-04 16:08:28 -0700164 if (cc->mode != MIGRATE_ASYNC &&
165 pfn > zone->compact_cached_migrate_pfn[1])
David Rientjes35979ef2014-06-04 16:08:27 -0700166 zone->compact_cached_migrate_pfn[1] = pfn;
167 } else {
168 if (cc->finished_update_free)
169 return;
170 if (pfn < zone->compact_cached_free_pfn)
171 zone->compact_cached_free_pfn = pfn;
Mel Gormanc89511a2012-10-08 16:32:45 -0700172 }
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700173}
174#else
175static inline bool isolation_suitable(struct compact_control *cc,
176 struct page *page)
177{
178 return true;
179}
180
Mel Gormanc89511a2012-10-08 16:32:45 -0700181static void update_pageblock_skip(struct compact_control *cc,
182 struct page *page, unsigned long nr_isolated,
David Rientjes35979ef2014-06-04 16:08:27 -0700183 bool set_unsuitable, bool migrate_scanner)
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700184{
185}
186#endif /* CONFIG_COMPACTION */
187
Mel Gorman2a1402a2012-10-08 16:32:33 -0700188static inline bool should_release_lock(spinlock_t *lock)
189{
190 return need_resched() || spin_is_contended(lock);
191}
192
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100193/*
Mel Gormanc67fe372012-08-21 16:16:17 -0700194 * Compaction requires the taking of some coarse locks that are potentially
195 * very heavily contended. Check if the process needs to be scheduled or
196 * if the lock is contended. For async compaction, back out in the event
197 * if contention is severe. For sync compaction, schedule.
198 *
199 * Returns true if the lock is held.
200 * Returns false if the lock is released and compaction should abort
201 */
202static bool compact_checklock_irqsave(spinlock_t *lock, unsigned long *flags,
203 bool locked, struct compact_control *cc)
204{
Mel Gorman2a1402a2012-10-08 16:32:33 -0700205 if (should_release_lock(lock)) {
Mel Gormanc67fe372012-08-21 16:16:17 -0700206 if (locked) {
207 spin_unlock_irqrestore(lock, *flags);
208 locked = false;
209 }
210
211 /* async aborts if taking too long or contended */
David Rientjese0b9dae2014-06-04 16:08:28 -0700212 if (cc->mode == MIGRATE_ASYNC) {
Shaohua Lie64c5232012-10-08 16:32:27 -0700213 cc->contended = true;
Mel Gormanc67fe372012-08-21 16:16:17 -0700214 return false;
215 }
216
217 cond_resched();
Mel Gormanc67fe372012-08-21 16:16:17 -0700218 }
219
220 if (!locked)
221 spin_lock_irqsave(lock, *flags);
222 return true;
223}
224
Mel Gormanf40d1e42012-10-08 16:32:36 -0700225/* Returns true if the page is within a block suitable for migration to */
226static bool suitable_migration_target(struct page *page)
227{
Joonsoo Kim7d348b92014-04-07 15:37:03 -0700228 /* If the page is a large free page, then disallow migration */
Mel Gormanf40d1e42012-10-08 16:32:36 -0700229 if (PageBuddy(page) && page_order(page) >= pageblock_order)
Joonsoo Kim7d348b92014-04-07 15:37:03 -0700230 return false;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700231
232 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
Joonsoo Kim7d348b92014-04-07 15:37:03 -0700233 if (migrate_async_suitable(get_pageblock_migratetype(page)))
Mel Gormanf40d1e42012-10-08 16:32:36 -0700234 return true;
235
236 /* Otherwise skip the block */
237 return false;
238}
239
Mel Gormanc67fe372012-08-21 16:16:17 -0700240/*
Jerome Marchand9e4be472013-11-12 15:07:12 -0800241 * Isolate free pages onto a private freelist. If @strict is true, will abort
242 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
243 * (even though it may still end up isolating some pages).
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100244 */
Mel Gormanf40d1e42012-10-08 16:32:36 -0700245static unsigned long isolate_freepages_block(struct compact_control *cc,
246 unsigned long blockpfn,
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100247 unsigned long end_pfn,
248 struct list_head *freelist,
249 bool strict)
Mel Gorman748446b2010-05-24 14:32:27 -0700250{
Mel Gormanb7aba692011-01-13 15:45:54 -0800251 int nr_scanned = 0, total_isolated = 0;
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700252 struct page *cursor, *valid_page = NULL;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700253 unsigned long flags;
254 bool locked = false;
Joonsoo Kim01ead532014-04-07 15:37:04 -0700255 bool checked_pageblock = false;
Mel Gorman748446b2010-05-24 14:32:27 -0700256
Mel Gorman748446b2010-05-24 14:32:27 -0700257 cursor = pfn_to_page(blockpfn);
258
Mel Gormanf40d1e42012-10-08 16:32:36 -0700259 /* Isolate free pages. */
Mel Gorman748446b2010-05-24 14:32:27 -0700260 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
261 int isolated, i;
262 struct page *page = cursor;
263
Mel Gormanb7aba692011-01-13 15:45:54 -0800264 nr_scanned++;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700265 if (!pfn_valid_within(blockpfn))
Laura Abbott2af120b2014-03-10 15:49:44 -0700266 goto isolate_fail;
267
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700268 if (!valid_page)
269 valid_page = page;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700270 if (!PageBuddy(page))
Laura Abbott2af120b2014-03-10 15:49:44 -0700271 goto isolate_fail;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700272
273 /*
274 * The zone lock must be held to isolate freepages.
275 * Unfortunately this is a very coarse lock and can be
276 * heavily contended if there are parallel allocations
277 * or parallel compactions. For async compaction do not
278 * spin on the lock and we acquire the lock as late as
279 * possible.
280 */
281 locked = compact_checklock_irqsave(&cc->zone->lock, &flags,
282 locked, cc);
283 if (!locked)
284 break;
285
286 /* Recheck this is a suitable migration target under lock */
Joonsoo Kim01ead532014-04-07 15:37:04 -0700287 if (!strict && !checked_pageblock) {
288 /*
289 * We need to check suitability of pageblock only once
290 * and this isolate_freepages_block() is called with
291 * pageblock range, so just check once is sufficient.
292 */
293 checked_pageblock = true;
294 if (!suitable_migration_target(page))
295 break;
296 }
Mel Gormanf40d1e42012-10-08 16:32:36 -0700297
298 /* Recheck this is a buddy page under lock */
299 if (!PageBuddy(page))
Laura Abbott2af120b2014-03-10 15:49:44 -0700300 goto isolate_fail;
Mel Gorman748446b2010-05-24 14:32:27 -0700301
302 /* Found a free page, break it into order-0 pages */
303 isolated = split_free_page(page);
304 total_isolated += isolated;
305 for (i = 0; i < isolated; i++) {
306 list_add(&page->lru, freelist);
307 page++;
308 }
309
310 /* If a page was split, advance to the end of it */
311 if (isolated) {
312 blockpfn += isolated - 1;
313 cursor += isolated - 1;
Laura Abbott2af120b2014-03-10 15:49:44 -0700314 continue;
Mel Gorman748446b2010-05-24 14:32:27 -0700315 }
Laura Abbott2af120b2014-03-10 15:49:44 -0700316
317isolate_fail:
318 if (strict)
319 break;
320 else
321 continue;
322
Mel Gorman748446b2010-05-24 14:32:27 -0700323 }
324
Mel Gormanb7aba692011-01-13 15:45:54 -0800325 trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
Mel Gormanf40d1e42012-10-08 16:32:36 -0700326
327 /*
328 * If strict isolation is requested by CMA then check that all the
329 * pages requested were isolated. If there were any failures, 0 is
330 * returned and CMA will fail.
331 */
Laura Abbott2af120b2014-03-10 15:49:44 -0700332 if (strict && blockpfn < end_pfn)
Mel Gormanf40d1e42012-10-08 16:32:36 -0700333 total_isolated = 0;
334
335 if (locked)
336 spin_unlock_irqrestore(&cc->zone->lock, flags);
337
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700338 /* Update the pageblock-skip if the whole pageblock was scanned */
339 if (blockpfn == end_pfn)
David Rientjes35979ef2014-06-04 16:08:27 -0700340 update_pageblock_skip(cc, valid_page, total_isolated, true,
341 false);
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700342
Minchan Kim010fc292012-12-20 15:05:06 -0800343 count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
Mel Gorman397487d2012-10-19 12:00:10 +0100344 if (total_isolated)
Minchan Kim010fc292012-12-20 15:05:06 -0800345 count_compact_events(COMPACTISOLATED, total_isolated);
Mel Gorman748446b2010-05-24 14:32:27 -0700346 return total_isolated;
347}
348
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100349/**
350 * isolate_freepages_range() - isolate free pages.
351 * @start_pfn: The first PFN to start isolating.
352 * @end_pfn: The one-past-last PFN.
353 *
354 * Non-free pages, invalid PFNs, or zone boundaries within the
355 * [start_pfn, end_pfn) range are considered errors, cause function to
356 * undo its actions and return zero.
357 *
358 * Otherwise, function returns one-past-the-last PFN of isolated page
359 * (which may be greater then end_pfn if end fell in a middle of
360 * a free page).
361 */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100362unsigned long
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700363isolate_freepages_range(struct compact_control *cc,
364 unsigned long start_pfn, unsigned long end_pfn)
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100365{
Mel Gormanf40d1e42012-10-08 16:32:36 -0700366 unsigned long isolated, pfn, block_end_pfn;
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100367 LIST_HEAD(freelist);
368
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100369 for (pfn = start_pfn; pfn < end_pfn; pfn += isolated) {
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700370 if (!pfn_valid(pfn) || cc->zone != page_zone(pfn_to_page(pfn)))
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100371 break;
372
373 /*
374 * On subsequent iterations ALIGN() is actually not needed,
375 * but we keep it that we not to complicate the code.
376 */
377 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
378 block_end_pfn = min(block_end_pfn, end_pfn);
379
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700380 isolated = isolate_freepages_block(cc, pfn, block_end_pfn,
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100381 &freelist, true);
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100382
383 /*
384 * In strict mode, isolate_freepages_block() returns 0 if
385 * there are any holes in the block (ie. invalid PFNs or
386 * non-free pages).
387 */
388 if (!isolated)
389 break;
390
391 /*
392 * If we managed to isolate pages, it is always (1 << n) *
393 * pageblock_nr_pages for some non-negative n. (Max order
394 * page may span two pageblocks).
395 */
396 }
397
398 /* split_free_page does not map the pages */
399 map_pages(&freelist);
400
401 if (pfn < end_pfn) {
402 /* Loop terminated early, cleanup. */
403 release_freepages(&freelist);
404 return 0;
405 }
406
407 /* We don't use freelists for anything. */
408 return pfn;
409}
410
Mel Gorman748446b2010-05-24 14:32:27 -0700411/* Update the number of anon and file isolated pages in the zone */
Mel Gormanc67fe372012-08-21 16:16:17 -0700412static void acct_isolated(struct zone *zone, bool locked, struct compact_control *cc)
Mel Gorman748446b2010-05-24 14:32:27 -0700413{
414 struct page *page;
Minchan Kimb9e84ac2011-10-31 17:06:44 -0700415 unsigned int count[2] = { 0, };
Mel Gorman748446b2010-05-24 14:32:27 -0700416
Minchan Kimb9e84ac2011-10-31 17:06:44 -0700417 list_for_each_entry(page, &cc->migratepages, lru)
418 count[!!page_is_file_cache(page)]++;
Mel Gorman748446b2010-05-24 14:32:27 -0700419
Mel Gormanc67fe372012-08-21 16:16:17 -0700420 /* If locked we can use the interrupt unsafe versions */
421 if (locked) {
422 __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
423 __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
424 } else {
425 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
426 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
427 }
Mel Gorman748446b2010-05-24 14:32:27 -0700428}
429
430/* Similar to reclaim, but different enough that they don't share logic */
431static bool too_many_isolated(struct zone *zone)
432{
Minchan Kimbc693042010-09-09 16:38:00 -0700433 unsigned long active, inactive, isolated;
Mel Gorman748446b2010-05-24 14:32:27 -0700434
435 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
436 zone_page_state(zone, NR_INACTIVE_ANON);
Minchan Kimbc693042010-09-09 16:38:00 -0700437 active = zone_page_state(zone, NR_ACTIVE_FILE) +
438 zone_page_state(zone, NR_ACTIVE_ANON);
Mel Gorman748446b2010-05-24 14:32:27 -0700439 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
440 zone_page_state(zone, NR_ISOLATED_ANON);
441
Minchan Kimbc693042010-09-09 16:38:00 -0700442 return isolated > (inactive + active) / 2;
Mel Gorman748446b2010-05-24 14:32:27 -0700443}
444
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100445/**
446 * isolate_migratepages_range() - isolate all migrate-able pages in range.
447 * @zone: Zone pages are in.
448 * @cc: Compaction control structure.
449 * @low_pfn: The first PFN of the range.
450 * @end_pfn: The one-past-the-last PFN of the range.
Minchan Kime46a2872012-10-08 16:33:48 -0700451 * @unevictable: true if it allows to isolate unevictable pages
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100452 *
453 * Isolate all pages that can be migrated from the range specified by
454 * [low_pfn, end_pfn). Returns zero if there is a fatal signal
455 * pending), otherwise PFN of the first page that was not scanned
456 * (which may be both less, equal to or more then end_pfn).
457 *
458 * Assumes that cc->migratepages is empty and cc->nr_migratepages is
459 * zero.
460 *
461 * Apart from cc->migratepages and cc->nr_migratetypes this function
462 * does not modify any cc's fields, in particular it does not modify
463 * (or read for that matter) cc->migrate_pfn.
Mel Gorman748446b2010-05-24 14:32:27 -0700464 */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100465unsigned long
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100466isolate_migratepages_range(struct zone *zone, struct compact_control *cc,
Minchan Kime46a2872012-10-08 16:33:48 -0700467 unsigned long low_pfn, unsigned long end_pfn, bool unevictable)
Mel Gorman748446b2010-05-24 14:32:27 -0700468{
Mel Gorman9927af742011-01-13 15:45:59 -0800469 unsigned long last_pageblock_nr = 0, pageblock_nr;
Mel Gormanb7aba692011-01-13 15:45:54 -0800470 unsigned long nr_scanned = 0, nr_isolated = 0;
Mel Gorman748446b2010-05-24 14:32:27 -0700471 struct list_head *migratelist = &cc->migratepages;
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700472 struct lruvec *lruvec;
Mel Gormanc67fe372012-08-21 16:16:17 -0700473 unsigned long flags;
Mel Gorman2a1402a2012-10-08 16:32:33 -0700474 bool locked = false;
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700475 struct page *page = NULL, *valid_page = NULL;
David Rientjes35979ef2014-06-04 16:08:27 -0700476 bool set_unsuitable = true;
David Rientjese0b9dae2014-06-04 16:08:28 -0700477 const isolate_mode_t mode = (cc->mode == MIGRATE_ASYNC ?
478 ISOLATE_ASYNC_MIGRATE : 0) |
David Rientjesda1c67a2014-04-07 15:37:34 -0700479 (unevictable ? ISOLATE_UNEVICTABLE : 0);
Mel Gorman748446b2010-05-24 14:32:27 -0700480
Mel Gorman748446b2010-05-24 14:32:27 -0700481 /*
482 * Ensure that there are not too many pages isolated from the LRU
483 * list by either parallel reclaimers or compaction. If there are,
484 * delay for some time until fewer pages are isolated
485 */
486 while (unlikely(too_many_isolated(zone))) {
Mel Gormanf9e35b32011-06-15 15:08:52 -0700487 /* async migration should just abort */
David Rientjese0b9dae2014-06-04 16:08:28 -0700488 if (cc->mode == MIGRATE_ASYNC)
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100489 return 0;
Mel Gormanf9e35b32011-06-15 15:08:52 -0700490
Mel Gorman748446b2010-05-24 14:32:27 -0700491 congestion_wait(BLK_RW_ASYNC, HZ/10);
492
493 if (fatal_signal_pending(current))
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100494 return 0;
Mel Gorman748446b2010-05-24 14:32:27 -0700495 }
496
David Rientjesaeef4b82014-06-04 16:08:31 -0700497 if (cond_resched()) {
498 /* Async terminates prematurely on need_resched() */
499 if (cc->mode == MIGRATE_ASYNC)
500 return 0;
501 }
502
Mel Gorman748446b2010-05-24 14:32:27 -0700503 /* Time to isolate some pages for migration */
Mel Gorman748446b2010-05-24 14:32:27 -0700504 for (; low_pfn < end_pfn; low_pfn++) {
Andrea Arcangelib2eef8c2011-03-22 16:33:10 -0700505 /* give a chance to irqs before checking need_resched() */
Joonsoo Kimbe1aa032014-04-07 15:37:05 -0700506 if (locked && !(low_pfn % SWAP_CLUSTER_MAX)) {
Mel Gorman2a1402a2012-10-08 16:32:33 -0700507 if (should_release_lock(&zone->lru_lock)) {
508 spin_unlock_irqrestore(&zone->lru_lock, flags);
509 locked = false;
510 }
Andrea Arcangelib2eef8c2011-03-22 16:33:10 -0700511 }
Mel Gormanc67fe372012-08-21 16:16:17 -0700512
Mel Gorman0bf380b2012-02-03 15:37:18 -0800513 /*
514 * migrate_pfn does not necessarily start aligned to a
515 * pageblock. Ensure that pfn_valid is called when moving
516 * into a new MAX_ORDER_NR_PAGES range in case of large
517 * memory holes within the zone
518 */
519 if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
520 if (!pfn_valid(low_pfn)) {
521 low_pfn += MAX_ORDER_NR_PAGES - 1;
522 continue;
523 }
524 }
525
Mel Gorman748446b2010-05-24 14:32:27 -0700526 if (!pfn_valid_within(low_pfn))
527 continue;
Mel Gormanb7aba692011-01-13 15:45:54 -0800528 nr_scanned++;
Mel Gorman748446b2010-05-24 14:32:27 -0700529
Mel Gormandc908602012-02-08 17:13:38 -0800530 /*
531 * Get the page and ensure the page is within the same zone.
532 * See the comment in isolate_freepages about overlapping
533 * nodes. It is deliberate that the new zone lock is not taken
534 * as memory compaction should not move pages between nodes.
535 */
Mel Gorman748446b2010-05-24 14:32:27 -0700536 page = pfn_to_page(low_pfn);
Mel Gormandc908602012-02-08 17:13:38 -0800537 if (page_zone(page) != zone)
538 continue;
539
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700540 if (!valid_page)
541 valid_page = page;
542
543 /* If isolation recently failed, do not retry */
544 pageblock_nr = low_pfn >> pageblock_order;
Joonsoo Kimc122b202014-04-07 15:37:06 -0700545 if (last_pageblock_nr != pageblock_nr) {
546 int mt;
547
548 last_pageblock_nr = pageblock_nr;
549 if (!isolation_suitable(cc, page))
550 goto next_pageblock;
551
552 /*
553 * For async migration, also only scan in MOVABLE
554 * blocks. Async migration is optimistic to see if
555 * the minimum amount of work satisfies the allocation
556 */
557 mt = get_pageblock_migratetype(page);
David Rientjese0b9dae2014-06-04 16:08:28 -0700558 if (cc->mode == MIGRATE_ASYNC &&
559 !migrate_async_suitable(mt)) {
David Rientjes35979ef2014-06-04 16:08:27 -0700560 set_unsuitable = false;
Joonsoo Kimc122b202014-04-07 15:37:06 -0700561 goto next_pageblock;
562 }
563 }
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700564
Mel Gorman6c144662014-01-23 15:53:38 -0800565 /*
566 * Skip if free. page_order cannot be used without zone->lock
567 * as nothing prevents parallel allocations or buddy merging.
568 */
Mel Gorman748446b2010-05-24 14:32:27 -0700569 if (PageBuddy(page))
570 continue;
571
Mel Gorman9927af742011-01-13 15:45:59 -0800572 /*
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800573 * Check may be lockless but that's ok as we recheck later.
574 * It's possible to migrate LRU pages and balloon pages
575 * Skip any other type of page
576 */
577 if (!PageLRU(page)) {
578 if (unlikely(balloon_page_movable(page))) {
579 if (locked && balloon_page_isolate(page)) {
580 /* Successfully isolated */
Joonsoo Kimb6c75012014-04-07 15:37:07 -0700581 goto isolate_success;
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800582 }
583 }
Andrea Arcangelibc835012011-01-13 15:47:08 -0800584 continue;
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800585 }
Andrea Arcangelibc835012011-01-13 15:47:08 -0800586
587 /*
Mel Gorman2a1402a2012-10-08 16:32:33 -0700588 * PageLRU is set. lru_lock normally excludes isolation
589 * splitting and collapsing (collapsing has already happened
590 * if PageLRU is set) but the lock is not necessarily taken
591 * here and it is wasteful to take it just to check transhuge.
592 * Check TransHuge without lock and skip the whole pageblock if
593 * it's either a transhuge or hugetlbfs page, as calling
594 * compound_order() without preventing THP from splitting the
595 * page underneath us may return surprising results.
Andrea Arcangelibc835012011-01-13 15:47:08 -0800596 */
597 if (PageTransHuge(page)) {
Mel Gorman2a1402a2012-10-08 16:32:33 -0700598 if (!locked)
599 goto next_pageblock;
600 low_pfn += (1 << compound_order(page)) - 1;
601 continue;
602 }
603
David Rientjes119d6d52014-04-03 14:48:00 -0700604 /*
605 * Migration will fail if an anonymous page is pinned in memory,
606 * so avoid taking lru_lock and isolating it unnecessarily in an
607 * admittedly racy check.
608 */
609 if (!page_mapping(page) &&
610 page_count(page) > page_mapcount(page))
611 continue;
612
Mel Gorman2a1402a2012-10-08 16:32:33 -0700613 /* Check if it is ok to still hold the lock */
614 locked = compact_checklock_irqsave(&zone->lru_lock, &flags,
615 locked, cc);
616 if (!locked || fatal_signal_pending(current))
617 break;
618
619 /* Recheck PageLRU and PageTransHuge under lock */
620 if (!PageLRU(page))
621 continue;
622 if (PageTransHuge(page)) {
Andrea Arcangelibc835012011-01-13 15:47:08 -0800623 low_pfn += (1 << compound_order(page)) - 1;
624 continue;
625 }
626
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700627 lruvec = mem_cgroup_page_lruvec(page, zone);
628
Mel Gorman748446b2010-05-24 14:32:27 -0700629 /* Try isolate the page */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -0700630 if (__isolate_lru_page(page, mode) != 0)
Mel Gorman748446b2010-05-24 14:32:27 -0700631 continue;
632
Sasha Levin309381fea2014-01-23 15:52:54 -0800633 VM_BUG_ON_PAGE(PageTransCompound(page), page);
Andrea Arcangelibc835012011-01-13 15:47:08 -0800634
Mel Gorman748446b2010-05-24 14:32:27 -0700635 /* Successfully isolated */
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700636 del_page_from_lru_list(page, lruvec, page_lru(page));
Joonsoo Kimb6c75012014-04-07 15:37:07 -0700637
638isolate_success:
639 cc->finished_update_migrate = true;
Mel Gorman748446b2010-05-24 14:32:27 -0700640 list_add(&page->lru, migratelist);
Mel Gorman748446b2010-05-24 14:32:27 -0700641 cc->nr_migratepages++;
Mel Gormanb7aba692011-01-13 15:45:54 -0800642 nr_isolated++;
Mel Gorman748446b2010-05-24 14:32:27 -0700643
644 /* Avoid isolating too much */
Hillf Danton31b83842012-01-10 15:07:59 -0800645 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
646 ++low_pfn;
Mel Gorman748446b2010-05-24 14:32:27 -0700647 break;
Hillf Danton31b83842012-01-10 15:07:59 -0800648 }
Mel Gorman2a1402a2012-10-08 16:32:33 -0700649
650 continue;
651
652next_pageblock:
Mel Gormana9aacbc2013-02-22 16:32:25 -0800653 low_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages) - 1;
Mel Gorman748446b2010-05-24 14:32:27 -0700654 }
655
Mel Gormanc67fe372012-08-21 16:16:17 -0700656 acct_isolated(zone, locked, cc);
Mel Gorman748446b2010-05-24 14:32:27 -0700657
Mel Gormanc67fe372012-08-21 16:16:17 -0700658 if (locked)
659 spin_unlock_irqrestore(&zone->lru_lock, flags);
Mel Gorman748446b2010-05-24 14:32:27 -0700660
Vlastimil Babka50b5b092014-01-21 15:51:10 -0800661 /*
662 * Update the pageblock-skip information and cached scanner pfn,
663 * if the whole pageblock was scanned without isolating any page.
Vlastimil Babka50b5b092014-01-21 15:51:10 -0800664 */
David Rientjes35979ef2014-06-04 16:08:27 -0700665 if (low_pfn == end_pfn)
666 update_pageblock_skip(cc, valid_page, nr_isolated,
667 set_unsuitable, true);
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700668
Mel Gormanb7aba692011-01-13 15:45:54 -0800669 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
670
Minchan Kim010fc292012-12-20 15:05:06 -0800671 count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
Mel Gorman397487d2012-10-19 12:00:10 +0100672 if (nr_isolated)
Minchan Kim010fc292012-12-20 15:05:06 -0800673 count_compact_events(COMPACTISOLATED, nr_isolated);
Mel Gorman397487d2012-10-19 12:00:10 +0100674
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100675 return low_pfn;
676}
677
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100678#endif /* CONFIG_COMPACTION || CONFIG_CMA */
679#ifdef CONFIG_COMPACTION
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100680/*
681 * Based on information in the current compact_control, find blocks
682 * suitable for isolating free pages from and then isolate them.
683 */
684static void isolate_freepages(struct zone *zone,
685 struct compact_control *cc)
686{
687 struct page *page;
Vlastimil Babkac96b9e52014-06-04 16:07:26 -0700688 unsigned long block_start_pfn; /* start of current pageblock */
689 unsigned long block_end_pfn; /* end of current pageblock */
690 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
691 unsigned long next_free_pfn; /* start pfn for scaning at next round */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100692 int nr_freepages = cc->nr_freepages;
693 struct list_head *freelist = &cc->freepages;
694
695 /*
696 * Initialise the free scanner. The starting point is where we last
Vlastimil Babka49e068f2014-05-06 12:50:03 -0700697 * successfully isolated from, zone-cached value, or the end of the
698 * zone when isolating for the first time. We need this aligned to
Vlastimil Babkac96b9e52014-06-04 16:07:26 -0700699 * the pageblock boundary, because we do
700 * block_start_pfn -= pageblock_nr_pages in the for loop.
701 * For ending point, take care when isolating in last pageblock of a
702 * a zone which ends in the middle of a pageblock.
Vlastimil Babka49e068f2014-05-06 12:50:03 -0700703 * The low boundary is the end of the pageblock the migration scanner
704 * is using.
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100705 */
Vlastimil Babkac96b9e52014-06-04 16:07:26 -0700706 block_start_pfn = cc->free_pfn & ~(pageblock_nr_pages-1);
707 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
708 zone_end_pfn(zone));
Vlastimil Babka7ed695e2014-01-21 15:51:09 -0800709 low_pfn = ALIGN(cc->migrate_pfn + 1, pageblock_nr_pages);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100710
711 /*
Vlastimil Babkac96b9e52014-06-04 16:07:26 -0700712 * If no pages are isolated, the block_start_pfn < low_pfn check
713 * will kick in.
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100714 */
Vlastimil Babkac96b9e52014-06-04 16:07:26 -0700715 next_free_pfn = 0;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100716
717 /*
718 * Isolate free pages until enough are available to migrate the
719 * pages on cc->migratepages. We stop searching if the migrate
720 * and free page scanners meet or enough free pages are isolated.
721 */
Vlastimil Babkac96b9e52014-06-04 16:07:26 -0700722 for (; block_start_pfn >= low_pfn && cc->nr_migratepages > nr_freepages;
723 block_end_pfn = block_start_pfn,
724 block_start_pfn -= pageblock_nr_pages) {
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100725 unsigned long isolated;
726
David Rientjesf6ea3ad2013-09-30 13:45:03 -0700727 /*
728 * This can iterate a massively long zone without finding any
729 * suitable migration targets, so periodically check if we need
730 * to schedule.
731 */
732 cond_resched();
733
Vlastimil Babkac96b9e52014-06-04 16:07:26 -0700734 if (!pfn_valid(block_start_pfn))
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100735 continue;
736
737 /*
738 * Check for overlapping nodes/zones. It's possible on some
739 * configurations to have a setup like
740 * node0 node1 node0
741 * i.e. it's possible that all pages within a zones range of
742 * pages do not belong to a single zone.
743 */
Vlastimil Babkac96b9e52014-06-04 16:07:26 -0700744 page = pfn_to_page(block_start_pfn);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100745 if (page_zone(page) != zone)
746 continue;
747
748 /* Check the block is suitable for migration */
Linus Torvalds68e3e922012-06-03 20:05:57 -0700749 if (!suitable_migration_target(page))
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100750 continue;
Linus Torvalds68e3e922012-06-03 20:05:57 -0700751
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700752 /* If isolation recently failed, do not retry */
753 if (!isolation_suitable(cc, page))
754 continue;
755
Mel Gormanf40d1e42012-10-08 16:32:36 -0700756 /* Found a block suitable for isolating free pages from */
Vlastimil Babkac96b9e52014-06-04 16:07:26 -0700757 isolated = isolate_freepages_block(cc, block_start_pfn,
758 block_end_pfn, freelist, false);
Mel Gormanf40d1e42012-10-08 16:32:36 -0700759 nr_freepages += isolated;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100760
761 /*
762 * Record the highest PFN we isolated pages from. When next
763 * looking for free pages, the search will restart here as
764 * page migration may have returned some pages to the allocator
765 */
Vlastimil Babkac96b9e52014-06-04 16:07:26 -0700766 if (isolated && next_free_pfn == 0) {
Mel Gormanc89511a2012-10-08 16:32:45 -0700767 cc->finished_update_free = true;
Vlastimil Babkac96b9e52014-06-04 16:07:26 -0700768 next_free_pfn = block_start_pfn;
Mel Gormanc89511a2012-10-08 16:32:45 -0700769 }
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100770 }
771
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100772 /* split_free_page does not map the pages */
773 map_pages(freelist);
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100774
Vlastimil Babka7ed695e2014-01-21 15:51:09 -0800775 /*
776 * If we crossed the migrate scanner, we want to keep it that way
777 * so that compact_finished() may detect this
778 */
Vlastimil Babkac96b9e52014-06-04 16:07:26 -0700779 if (block_start_pfn < low_pfn)
780 next_free_pfn = cc->migrate_pfn;
781
782 cc->free_pfn = next_free_pfn;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100783 cc->nr_freepages = nr_freepages;
Mel Gorman748446b2010-05-24 14:32:27 -0700784}
785
786/*
787 * This is a migrate-callback that "allocates" freepages by taking pages
788 * from the isolated freelists in the block we are migrating to.
789 */
790static struct page *compaction_alloc(struct page *migratepage,
791 unsigned long data,
792 int **result)
793{
794 struct compact_control *cc = (struct compact_control *)data;
795 struct page *freepage;
796
797 /* Isolate free pages if necessary */
798 if (list_empty(&cc->freepages)) {
799 isolate_freepages(cc->zone, cc);
800
801 if (list_empty(&cc->freepages))
802 return NULL;
803 }
804
805 freepage = list_entry(cc->freepages.next, struct page, lru);
806 list_del(&freepage->lru);
807 cc->nr_freepages--;
808
809 return freepage;
810}
811
812/*
David Rientjesd53aea32014-06-04 16:08:26 -0700813 * This is a migrate-callback that "frees" freepages back to the isolated
814 * freelist. All pages on the freelist are from the same zone, so there is no
815 * special handling needed for NUMA.
816 */
817static void compaction_free(struct page *page, unsigned long data)
818{
819 struct compact_control *cc = (struct compact_control *)data;
820
821 list_add(&page->lru, &cc->freepages);
822 cc->nr_freepages++;
823}
824
825/*
826 * We cannot control nr_migratepages fully when migration is running as
827 * migrate_pages() has no knowledge of of compact_control. When migration is
828 * complete, we count the number of pages on the list by hand.
Mel Gorman748446b2010-05-24 14:32:27 -0700829 */
830static void update_nr_listpages(struct compact_control *cc)
831{
832 int nr_migratepages = 0;
Mel Gorman748446b2010-05-24 14:32:27 -0700833 struct page *page;
834
835 list_for_each_entry(page, &cc->migratepages, lru)
836 nr_migratepages++;
Mel Gorman748446b2010-05-24 14:32:27 -0700837
838 cc->nr_migratepages = nr_migratepages;
Mel Gorman748446b2010-05-24 14:32:27 -0700839}
840
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100841/* possible outcome of isolate_migratepages */
842typedef enum {
843 ISOLATE_ABORT, /* Abort compaction now */
844 ISOLATE_NONE, /* No pages isolated, continue scanning */
845 ISOLATE_SUCCESS, /* Pages isolated, migrate */
846} isolate_migrate_t;
847
848/*
849 * Isolate all pages that can be migrated from the block pointed to by
850 * the migrate scanner within compact_control.
851 */
852static isolate_migrate_t isolate_migratepages(struct zone *zone,
853 struct compact_control *cc)
854{
855 unsigned long low_pfn, end_pfn;
856
857 /* Do not scan outside zone boundaries */
858 low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
859
860 /* Only scan within a pageblock boundary */
Mel Gormana9aacbc2013-02-22 16:32:25 -0800861 end_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100862
863 /* Do not cross the free scanner or scan within a memory hole */
864 if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
865 cc->migrate_pfn = end_pfn;
866 return ISOLATE_NONE;
867 }
868
869 /* Perform the isolation */
Minchan Kime46a2872012-10-08 16:33:48 -0700870 low_pfn = isolate_migratepages_range(zone, cc, low_pfn, end_pfn, false);
Shaohua Lie64c5232012-10-08 16:32:27 -0700871 if (!low_pfn || cc->contended)
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100872 return ISOLATE_ABORT;
873
874 cc->migrate_pfn = low_pfn;
875
876 return ISOLATE_SUCCESS;
877}
878
Mel Gorman748446b2010-05-24 14:32:27 -0700879static int compact_finished(struct zone *zone,
Andrea Arcangeli5a03b052011-01-13 15:47:11 -0800880 struct compact_control *cc)
Mel Gorman748446b2010-05-24 14:32:27 -0700881{
Mel Gorman8fb74b92013-01-11 14:32:16 -0800882 unsigned int order;
Andrea Arcangeli5a03b052011-01-13 15:47:11 -0800883 unsigned long watermark;
Mel Gorman56de7262010-05-24 14:32:30 -0700884
Mel Gorman748446b2010-05-24 14:32:27 -0700885 if (fatal_signal_pending(current))
886 return COMPACT_PARTIAL;
887
Mel Gorman753341a2012-10-08 16:32:40 -0700888 /* Compaction run completes if the migrate and free scanner meet */
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700889 if (cc->free_pfn <= cc->migrate_pfn) {
Vlastimil Babka55b7c4c2014-01-21 15:51:11 -0800890 /* Let the next compaction start anew. */
David Rientjes35979ef2014-06-04 16:08:27 -0700891 zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
892 zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
Vlastimil Babka55b7c4c2014-01-21 15:51:11 -0800893 zone->compact_cached_free_pfn = zone_end_pfn(zone);
894
Mel Gorman62997022012-10-08 16:32:47 -0700895 /*
896 * Mark that the PG_migrate_skip information should be cleared
897 * by kswapd when it goes to sleep. kswapd does not set the
898 * flag itself as the decision to be clear should be directly
899 * based on an allocation request.
900 */
901 if (!current_is_kswapd())
902 zone->compact_blockskip_flush = true;
903
Mel Gorman748446b2010-05-24 14:32:27 -0700904 return COMPACT_COMPLETE;
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700905 }
Mel Gorman748446b2010-05-24 14:32:27 -0700906
Johannes Weiner82478fb2011-01-20 14:44:21 -0800907 /*
908 * order == -1 is expected when compacting via
909 * /proc/sys/vm/compact_memory
910 */
Mel Gorman56de7262010-05-24 14:32:30 -0700911 if (cc->order == -1)
912 return COMPACT_CONTINUE;
913
Michal Hocko3957c772011-06-15 15:08:25 -0700914 /* Compaction run is not finished if the watermark is not met */
915 watermark = low_wmark_pages(zone);
916 watermark += (1 << cc->order);
917
918 if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
919 return COMPACT_CONTINUE;
920
Mel Gorman56de7262010-05-24 14:32:30 -0700921 /* Direct compactor: Is a suitable page free? */
Mel Gorman8fb74b92013-01-11 14:32:16 -0800922 for (order = cc->order; order < MAX_ORDER; order++) {
923 struct free_area *area = &zone->free_area[order];
Mel Gorman56de7262010-05-24 14:32:30 -0700924
Mel Gorman8fb74b92013-01-11 14:32:16 -0800925 /* Job done if page is free of the right migratetype */
926 if (!list_empty(&area->free_list[cc->migratetype]))
927 return COMPACT_PARTIAL;
928
929 /* Job done if allocation would set block type */
930 if (cc->order >= pageblock_order && area->nr_free)
931 return COMPACT_PARTIAL;
Mel Gorman56de7262010-05-24 14:32:30 -0700932 }
933
Mel Gorman748446b2010-05-24 14:32:27 -0700934 return COMPACT_CONTINUE;
935}
936
Mel Gorman3e7d3442011-01-13 15:45:56 -0800937/*
938 * compaction_suitable: Is this suitable to run compaction on this zone now?
939 * Returns
940 * COMPACT_SKIPPED - If there are too few free pages for compaction
941 * COMPACT_PARTIAL - If the allocation would succeed without compaction
942 * COMPACT_CONTINUE - If compaction should run now
943 */
944unsigned long compaction_suitable(struct zone *zone, int order)
945{
946 int fragindex;
947 unsigned long watermark;
948
949 /*
Michal Hocko3957c772011-06-15 15:08:25 -0700950 * order == -1 is expected when compacting via
951 * /proc/sys/vm/compact_memory
952 */
953 if (order == -1)
954 return COMPACT_CONTINUE;
955
956 /*
Mel Gorman3e7d3442011-01-13 15:45:56 -0800957 * Watermarks for order-0 must be met for compaction. Note the 2UL.
958 * This is because during migration, copies of pages need to be
959 * allocated and for a short time, the footprint is higher
960 */
961 watermark = low_wmark_pages(zone) + (2UL << order);
962 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
963 return COMPACT_SKIPPED;
964
965 /*
966 * fragmentation index determines if allocation failures are due to
967 * low memory or external fragmentation
968 *
Shaohua Lia582a732011-06-15 15:08:49 -0700969 * index of -1000 implies allocations might succeed depending on
970 * watermarks
Mel Gorman3e7d3442011-01-13 15:45:56 -0800971 * index towards 0 implies failure is due to lack of memory
972 * index towards 1000 implies failure is due to fragmentation
973 *
974 * Only compact if a failure would be due to fragmentation.
975 */
976 fragindex = fragmentation_index(zone, order);
977 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
978 return COMPACT_SKIPPED;
979
Shaohua Lia582a732011-06-15 15:08:49 -0700980 if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
981 0, 0))
Mel Gorman3e7d3442011-01-13 15:45:56 -0800982 return COMPACT_PARTIAL;
983
984 return COMPACT_CONTINUE;
985}
986
Mel Gorman748446b2010-05-24 14:32:27 -0700987static int compact_zone(struct zone *zone, struct compact_control *cc)
988{
989 int ret;
Mel Gormanc89511a2012-10-08 16:32:45 -0700990 unsigned long start_pfn = zone->zone_start_pfn;
Cody P Schafer108bcc92013-02-22 16:35:23 -0800991 unsigned long end_pfn = zone_end_pfn(zone);
David Rientjese0b9dae2014-06-04 16:08:28 -0700992 const bool sync = cc->mode != MIGRATE_ASYNC;
Mel Gorman748446b2010-05-24 14:32:27 -0700993
Mel Gorman3e7d3442011-01-13 15:45:56 -0800994 ret = compaction_suitable(zone, cc->order);
995 switch (ret) {
996 case COMPACT_PARTIAL:
997 case COMPACT_SKIPPED:
998 /* Compaction is likely to fail */
999 return ret;
1000 case COMPACT_CONTINUE:
1001 /* Fall through to compaction */
1002 ;
1003 }
1004
Mel Gormanc89511a2012-10-08 16:32:45 -07001005 /*
Vlastimil Babkad3132e42014-01-21 15:51:08 -08001006 * Clear pageblock skip if there were failures recently and compaction
1007 * is about to be retried after being deferred. kswapd does not do
1008 * this reset as it'll reset the cached information when going to sleep.
1009 */
1010 if (compaction_restarting(zone, cc->order) && !current_is_kswapd())
1011 __reset_isolation_suitable(zone);
1012
1013 /*
Mel Gormanc89511a2012-10-08 16:32:45 -07001014 * Setup to move all movable pages to the end of the zone. Used cached
1015 * information on where the scanners should start but check that it
1016 * is initialised by ensuring the values are within zone boundaries.
1017 */
David Rientjese0b9dae2014-06-04 16:08:28 -07001018 cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync];
Mel Gormanc89511a2012-10-08 16:32:45 -07001019 cc->free_pfn = zone->compact_cached_free_pfn;
1020 if (cc->free_pfn < start_pfn || cc->free_pfn > end_pfn) {
1021 cc->free_pfn = end_pfn & ~(pageblock_nr_pages-1);
1022 zone->compact_cached_free_pfn = cc->free_pfn;
1023 }
1024 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn > end_pfn) {
1025 cc->migrate_pfn = start_pfn;
David Rientjes35979ef2014-06-04 16:08:27 -07001026 zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
1027 zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
Mel Gormanc89511a2012-10-08 16:32:45 -07001028 }
Mel Gorman748446b2010-05-24 14:32:27 -07001029
Mel Gorman0eb927c2014-01-21 15:51:05 -08001030 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn, cc->free_pfn, end_pfn);
1031
Mel Gorman748446b2010-05-24 14:32:27 -07001032 migrate_prep_local();
1033
1034 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
1035 unsigned long nr_migrate, nr_remaining;
Minchan Kim9d502c12011-03-22 16:30:39 -07001036 int err;
Mel Gorman748446b2010-05-24 14:32:27 -07001037
Mel Gormanf9e35b32011-06-15 15:08:52 -07001038 switch (isolate_migratepages(zone, cc)) {
1039 case ISOLATE_ABORT:
1040 ret = COMPACT_PARTIAL;
Rafael Aquini5733c7d2012-12-11 16:02:47 -08001041 putback_movable_pages(&cc->migratepages);
Shaohua Lie64c5232012-10-08 16:32:27 -07001042 cc->nr_migratepages = 0;
Mel Gormanf9e35b32011-06-15 15:08:52 -07001043 goto out;
1044 case ISOLATE_NONE:
Mel Gorman748446b2010-05-24 14:32:27 -07001045 continue;
Mel Gormanf9e35b32011-06-15 15:08:52 -07001046 case ISOLATE_SUCCESS:
1047 ;
1048 }
Mel Gorman748446b2010-05-24 14:32:27 -07001049
1050 nr_migrate = cc->nr_migratepages;
David Rientjesd53aea32014-06-04 16:08:26 -07001051 err = migrate_pages(&cc->migratepages, compaction_alloc,
David Rientjese0b9dae2014-06-04 16:08:28 -07001052 compaction_free, (unsigned long)cc, cc->mode,
Mel Gorman7b2a2d42012-10-19 14:07:31 +01001053 MR_COMPACTION);
Mel Gorman748446b2010-05-24 14:32:27 -07001054 update_nr_listpages(cc);
1055 nr_remaining = cc->nr_migratepages;
1056
Mel Gormanb7aba692011-01-13 15:45:54 -08001057 trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
1058 nr_remaining);
Mel Gorman748446b2010-05-24 14:32:27 -07001059
Rafael Aquini5733c7d2012-12-11 16:02:47 -08001060 /* Release isolated pages not migrated */
Minchan Kim9d502c12011-03-22 16:30:39 -07001061 if (err) {
Rafael Aquini5733c7d2012-12-11 16:02:47 -08001062 putback_movable_pages(&cc->migratepages);
Mel Gorman748446b2010-05-24 14:32:27 -07001063 cc->nr_migratepages = 0;
Vlastimil Babka7ed695e2014-01-21 15:51:09 -08001064 /*
1065 * migrate_pages() may return -ENOMEM when scanners meet
1066 * and we want compact_finished() to detect it
1067 */
1068 if (err == -ENOMEM && cc->free_pfn > cc->migrate_pfn) {
David Rientjes4bf2bba2012-07-11 14:02:13 -07001069 ret = COMPACT_PARTIAL;
1070 goto out;
1071 }
Mel Gorman748446b2010-05-24 14:32:27 -07001072 }
Mel Gorman748446b2010-05-24 14:32:27 -07001073 }
1074
Mel Gormanf9e35b32011-06-15 15:08:52 -07001075out:
Mel Gorman748446b2010-05-24 14:32:27 -07001076 /* Release free pages and check accounting */
1077 cc->nr_freepages -= release_freepages(&cc->freepages);
1078 VM_BUG_ON(cc->nr_freepages != 0);
1079
Mel Gorman0eb927c2014-01-21 15:51:05 -08001080 trace_mm_compaction_end(ret);
1081
Mel Gorman748446b2010-05-24 14:32:27 -07001082 return ret;
1083}
Mel Gorman76ab0f52010-05-24 14:32:28 -07001084
David Rientjese0b9dae2014-06-04 16:08:28 -07001085static unsigned long compact_zone_order(struct zone *zone, int order,
1086 gfp_t gfp_mask, enum migrate_mode mode, bool *contended)
Mel Gorman56de7262010-05-24 14:32:30 -07001087{
Shaohua Lie64c5232012-10-08 16:32:27 -07001088 unsigned long ret;
Mel Gorman56de7262010-05-24 14:32:30 -07001089 struct compact_control cc = {
1090 .nr_freepages = 0,
1091 .nr_migratepages = 0,
1092 .order = order,
1093 .migratetype = allocflags_to_migratetype(gfp_mask),
1094 .zone = zone,
David Rientjese0b9dae2014-06-04 16:08:28 -07001095 .mode = mode,
Mel Gorman56de7262010-05-24 14:32:30 -07001096 };
1097 INIT_LIST_HEAD(&cc.freepages);
1098 INIT_LIST_HEAD(&cc.migratepages);
1099
Shaohua Lie64c5232012-10-08 16:32:27 -07001100 ret = compact_zone(zone, &cc);
1101
1102 VM_BUG_ON(!list_empty(&cc.freepages));
1103 VM_BUG_ON(!list_empty(&cc.migratepages));
1104
1105 *contended = cc.contended;
1106 return ret;
Mel Gorman56de7262010-05-24 14:32:30 -07001107}
1108
Mel Gorman5e771902010-05-24 14:32:31 -07001109int sysctl_extfrag_threshold = 500;
1110
Mel Gorman56de7262010-05-24 14:32:30 -07001111/**
1112 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1113 * @zonelist: The zonelist used for the current allocation
1114 * @order: The order of the current allocation
1115 * @gfp_mask: The GFP mask of the current allocation
1116 * @nodemask: The allowed nodes to allocate from
David Rientjese0b9dae2014-06-04 16:08:28 -07001117 * @mode: The migration mode for async, sync light, or sync migration
Mel Gorman661c4cb2012-10-08 16:32:31 -07001118 * @contended: Return value that is true if compaction was aborted due to lock contention
1119 * @page: Optionally capture a free page of the requested order during compaction
Mel Gorman56de7262010-05-24 14:32:30 -07001120 *
1121 * This is the main entry point for direct page compaction.
1122 */
1123unsigned long try_to_compact_pages(struct zonelist *zonelist,
Mel Gorman77f1fe62011-01-13 15:45:57 -08001124 int order, gfp_t gfp_mask, nodemask_t *nodemask,
David Rientjese0b9dae2014-06-04 16:08:28 -07001125 enum migrate_mode mode, bool *contended)
Mel Gorman56de7262010-05-24 14:32:30 -07001126{
1127 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1128 int may_enter_fs = gfp_mask & __GFP_FS;
1129 int may_perform_io = gfp_mask & __GFP_IO;
Mel Gorman56de7262010-05-24 14:32:30 -07001130 struct zoneref *z;
1131 struct zone *zone;
1132 int rc = COMPACT_SKIPPED;
Bartlomiej Zolnierkiewiczd95ea5d2012-10-08 16:32:05 -07001133 int alloc_flags = 0;
Mel Gorman56de7262010-05-24 14:32:30 -07001134
Mel Gorman4ffb6332012-10-08 16:29:09 -07001135 /* Check if the GFP flags allow compaction */
Andrea Arcangelic5a73c32011-01-13 15:47:11 -08001136 if (!order || !may_enter_fs || !may_perform_io)
Mel Gorman56de7262010-05-24 14:32:30 -07001137 return rc;
1138
Minchan Kim010fc292012-12-20 15:05:06 -08001139 count_compact_event(COMPACTSTALL);
Mel Gorman56de7262010-05-24 14:32:30 -07001140
Bartlomiej Zolnierkiewiczd95ea5d2012-10-08 16:32:05 -07001141#ifdef CONFIG_CMA
1142 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
1143 alloc_flags |= ALLOC_CMA;
1144#endif
Mel Gorman56de7262010-05-24 14:32:30 -07001145 /* Compact each zone in the list */
1146 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1147 nodemask) {
Mel Gorman56de7262010-05-24 14:32:30 -07001148 int status;
1149
David Rientjese0b9dae2014-06-04 16:08:28 -07001150 status = compact_zone_order(zone, order, gfp_mask, mode,
Mel Gorman8fb74b92013-01-11 14:32:16 -08001151 contended);
Mel Gorman56de7262010-05-24 14:32:30 -07001152 rc = max(status, rc);
1153
Mel Gorman3e7d3442011-01-13 15:45:56 -08001154 /* If a normal allocation would succeed, stop compacting */
Bartlomiej Zolnierkiewiczd95ea5d2012-10-08 16:32:05 -07001155 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0,
1156 alloc_flags))
Mel Gorman56de7262010-05-24 14:32:30 -07001157 break;
1158 }
1159
1160 return rc;
1161}
1162
1163
Mel Gorman76ab0f52010-05-24 14:32:28 -07001164/* Compact all zones within a node */
Andrew Morton7103f162013-02-22 16:32:33 -08001165static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
Mel Gorman76ab0f52010-05-24 14:32:28 -07001166{
1167 int zoneid;
Mel Gorman76ab0f52010-05-24 14:32:28 -07001168 struct zone *zone;
1169
Mel Gorman76ab0f52010-05-24 14:32:28 -07001170 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
Mel Gorman76ab0f52010-05-24 14:32:28 -07001171
1172 zone = &pgdat->node_zones[zoneid];
1173 if (!populated_zone(zone))
1174 continue;
1175
Rik van Riel7be62de2012-03-21 16:33:52 -07001176 cc->nr_freepages = 0;
1177 cc->nr_migratepages = 0;
1178 cc->zone = zone;
1179 INIT_LIST_HEAD(&cc->freepages);
1180 INIT_LIST_HEAD(&cc->migratepages);
Mel Gorman76ab0f52010-05-24 14:32:28 -07001181
Dan Carpenteraad6ec32012-03-21 16:33:54 -07001182 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
Rik van Riel7be62de2012-03-21 16:33:52 -07001183 compact_zone(zone, cc);
Mel Gorman76ab0f52010-05-24 14:32:28 -07001184
Rik van Rielaff62242012-03-21 16:33:52 -07001185 if (cc->order > 0) {
Vlastimil Babkade6c60a2014-01-21 15:51:07 -08001186 if (zone_watermark_ok(zone, cc->order,
1187 low_wmark_pages(zone), 0, 0))
1188 compaction_defer_reset(zone, cc->order, false);
Rik van Rielaff62242012-03-21 16:33:52 -07001189 }
1190
Rik van Riel7be62de2012-03-21 16:33:52 -07001191 VM_BUG_ON(!list_empty(&cc->freepages));
1192 VM_BUG_ON(!list_empty(&cc->migratepages));
Mel Gorman76ab0f52010-05-24 14:32:28 -07001193 }
Mel Gorman76ab0f52010-05-24 14:32:28 -07001194}
1195
Andrew Morton7103f162013-02-22 16:32:33 -08001196void compact_pgdat(pg_data_t *pgdat, int order)
Rik van Riel7be62de2012-03-21 16:33:52 -07001197{
1198 struct compact_control cc = {
1199 .order = order,
David Rientjese0b9dae2014-06-04 16:08:28 -07001200 .mode = MIGRATE_ASYNC,
Rik van Riel7be62de2012-03-21 16:33:52 -07001201 };
1202
Mel Gorman3a7200a2013-09-11 14:22:19 -07001203 if (!order)
1204 return;
1205
Andrew Morton7103f162013-02-22 16:32:33 -08001206 __compact_pgdat(pgdat, &cc);
Rik van Riel7be62de2012-03-21 16:33:52 -07001207}
1208
Andrew Morton7103f162013-02-22 16:32:33 -08001209static void compact_node(int nid)
Rik van Riel7be62de2012-03-21 16:33:52 -07001210{
Rik van Riel7be62de2012-03-21 16:33:52 -07001211 struct compact_control cc = {
1212 .order = -1,
David Rientjese0b9dae2014-06-04 16:08:28 -07001213 .mode = MIGRATE_SYNC,
David Rientjes91ca9182014-04-03 14:47:23 -07001214 .ignore_skip_hint = true,
Rik van Riel7be62de2012-03-21 16:33:52 -07001215 };
1216
Andrew Morton7103f162013-02-22 16:32:33 -08001217 __compact_pgdat(NODE_DATA(nid), &cc);
Rik van Riel7be62de2012-03-21 16:33:52 -07001218}
1219
Mel Gorman76ab0f52010-05-24 14:32:28 -07001220/* Compact all nodes in the system */
Jason Liu7964c062013-01-11 14:31:47 -08001221static void compact_nodes(void)
Mel Gorman76ab0f52010-05-24 14:32:28 -07001222{
1223 int nid;
1224
Hugh Dickins8575ec22012-03-21 16:33:53 -07001225 /* Flush pending updates to the LRU lists */
1226 lru_add_drain_all();
1227
Mel Gorman76ab0f52010-05-24 14:32:28 -07001228 for_each_online_node(nid)
1229 compact_node(nid);
Mel Gorman76ab0f52010-05-24 14:32:28 -07001230}
1231
1232/* The written value is actually unused, all memory is compacted */
1233int sysctl_compact_memory;
1234
1235/* This is the entry point for compacting all nodes via /proc/sys/vm */
1236int sysctl_compaction_handler(struct ctl_table *table, int write,
1237 void __user *buffer, size_t *length, loff_t *ppos)
1238{
1239 if (write)
Jason Liu7964c062013-01-11 14:31:47 -08001240 compact_nodes();
Mel Gorman76ab0f52010-05-24 14:32:28 -07001241
1242 return 0;
1243}
Mel Gormaned4a6d72010-05-24 14:32:29 -07001244
Mel Gorman5e771902010-05-24 14:32:31 -07001245int sysctl_extfrag_handler(struct ctl_table *table, int write,
1246 void __user *buffer, size_t *length, loff_t *ppos)
1247{
1248 proc_dointvec_minmax(table, write, buffer, length, ppos);
1249
1250 return 0;
1251}
1252
Mel Gormaned4a6d72010-05-24 14:32:29 -07001253#if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
Rashika Kheria74e77fb2014-04-03 14:48:01 -07001254static ssize_t sysfs_compact_node(struct device *dev,
Kay Sievers10fbcf42011-12-21 14:48:43 -08001255 struct device_attribute *attr,
Mel Gormaned4a6d72010-05-24 14:32:29 -07001256 const char *buf, size_t count)
1257{
Hugh Dickins8575ec22012-03-21 16:33:53 -07001258 int nid = dev->id;
1259
1260 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1261 /* Flush pending updates to the LRU lists */
1262 lru_add_drain_all();
1263
1264 compact_node(nid);
1265 }
Mel Gormaned4a6d72010-05-24 14:32:29 -07001266
1267 return count;
1268}
Kay Sievers10fbcf42011-12-21 14:48:43 -08001269static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
Mel Gormaned4a6d72010-05-24 14:32:29 -07001270
1271int compaction_register_node(struct node *node)
1272{
Kay Sievers10fbcf42011-12-21 14:48:43 -08001273 return device_create_file(&node->dev, &dev_attr_compact);
Mel Gormaned4a6d72010-05-24 14:32:29 -07001274}
1275
1276void compaction_unregister_node(struct node *node)
1277{
Kay Sievers10fbcf42011-12-21 14:48:43 -08001278 return device_remove_file(&node->dev, &dev_attr_compact);
Mel Gormaned4a6d72010-05-24 14:32:29 -07001279}
1280#endif /* CONFIG_SYSFS && CONFIG_NUMA */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001281
1282#endif /* CONFIG_COMPACTION */