blob: 3a91a2ea3d343d5bb22fdc1a5e67660a133de5e0 [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{
220 int migratetype = get_pageblock_migratetype(page);
221
222 /* Don't interfere with memory hot-remove or the min_free_kbytes blocks */
Minchan Kim194159f2013-02-22 16:33:58 -0800223 if (migratetype == MIGRATE_RESERVE)
224 return false;
225
226 if (is_migrate_isolate(migratetype))
Mel Gormanf40d1e42012-10-08 16:32:36 -0700227 return false;
228
229 /* If the page is a large free page, then allow migration */
230 if (PageBuddy(page) && page_order(page) >= pageblock_order)
231 return true;
232
233 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
234 if (migrate_async_suitable(migratetype))
235 return true;
236
237 /* Otherwise skip the block */
238 return false;
239}
240
Mel Gormanc67fe372012-08-21 16:16:17 -0700241/*
Jerome Marchand9e4be472013-11-12 15:07:12 -0800242 * Isolate free pages onto a private freelist. If @strict is true, will abort
243 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
244 * (even though it may still end up isolating some pages).
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100245 */
Mel Gormanf40d1e42012-10-08 16:32:36 -0700246static unsigned long isolate_freepages_block(struct compact_control *cc,
247 unsigned long blockpfn,
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100248 unsigned long end_pfn,
249 struct list_head *freelist,
250 bool strict)
Mel Gorman748446b2010-05-24 14:32:27 -0700251{
Mel Gormanb7aba692011-01-13 15:45:54 -0800252 int nr_scanned = 0, total_isolated = 0;
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700253 struct page *cursor, *valid_page = NULL;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700254 unsigned long nr_strict_required = end_pfn - blockpfn;
255 unsigned long flags;
256 bool locked = false;
Mel Gorman748446b2010-05-24 14:32:27 -0700257
Mel Gorman748446b2010-05-24 14:32:27 -0700258 cursor = pfn_to_page(blockpfn);
259
Mel Gormanf40d1e42012-10-08 16:32:36 -0700260 /* Isolate free pages. */
Mel Gorman748446b2010-05-24 14:32:27 -0700261 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
262 int isolated, i;
263 struct page *page = cursor;
264
Mel Gormanb7aba692011-01-13 15:45:54 -0800265 nr_scanned++;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700266 if (!pfn_valid_within(blockpfn))
Mel Gorman748446b2010-05-24 14:32:27 -0700267 continue;
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))
271 continue;
272
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 */
287 if (!strict && !suitable_migration_target(page))
288 break;
289
290 /* Recheck this is a buddy page under lock */
291 if (!PageBuddy(page))
292 continue;
Mel Gorman748446b2010-05-24 14:32:27 -0700293
294 /* Found a free page, break it into order-0 pages */
295 isolated = split_free_page(page);
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100296 if (!isolated && strict)
Mel Gormanf40d1e42012-10-08 16:32:36 -0700297 break;
Mel Gorman748446b2010-05-24 14:32:27 -0700298 total_isolated += isolated;
299 for (i = 0; i < isolated; i++) {
300 list_add(&page->lru, freelist);
301 page++;
302 }
303
304 /* If a page was split, advance to the end of it */
305 if (isolated) {
306 blockpfn += isolated - 1;
307 cursor += isolated - 1;
308 }
309 }
310
Mel Gormanb7aba692011-01-13 15:45:54 -0800311 trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
Mel Gormanf40d1e42012-10-08 16:32:36 -0700312
313 /*
314 * If strict isolation is requested by CMA then check that all the
315 * pages requested were isolated. If there were any failures, 0 is
316 * returned and CMA will fail.
317 */
Mel Gorman0db63d72012-10-19 13:56:57 -0700318 if (strict && nr_strict_required > total_isolated)
Mel Gormanf40d1e42012-10-08 16:32:36 -0700319 total_isolated = 0;
320
321 if (locked)
322 spin_unlock_irqrestore(&cc->zone->lock, flags);
323
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700324 /* Update the pageblock-skip if the whole pageblock was scanned */
325 if (blockpfn == end_pfn)
Mel Gormanc89511a2012-10-08 16:32:45 -0700326 update_pageblock_skip(cc, valid_page, total_isolated, false);
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700327
Minchan Kim010fc292012-12-20 15:05:06 -0800328 count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
Mel Gorman397487d2012-10-19 12:00:10 +0100329 if (total_isolated)
Minchan Kim010fc292012-12-20 15:05:06 -0800330 count_compact_events(COMPACTISOLATED, total_isolated);
Mel Gorman748446b2010-05-24 14:32:27 -0700331 return total_isolated;
332}
333
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100334/**
335 * isolate_freepages_range() - isolate free pages.
336 * @start_pfn: The first PFN to start isolating.
337 * @end_pfn: The one-past-last PFN.
338 *
339 * Non-free pages, invalid PFNs, or zone boundaries within the
340 * [start_pfn, end_pfn) range are considered errors, cause function to
341 * undo its actions and return zero.
342 *
343 * Otherwise, function returns one-past-the-last PFN of isolated page
344 * (which may be greater then end_pfn if end fell in a middle of
345 * a free page).
346 */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100347unsigned long
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700348isolate_freepages_range(struct compact_control *cc,
349 unsigned long start_pfn, unsigned long end_pfn)
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100350{
Mel Gormanf40d1e42012-10-08 16:32:36 -0700351 unsigned long isolated, pfn, block_end_pfn;
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100352 LIST_HEAD(freelist);
353
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100354 for (pfn = start_pfn; pfn < end_pfn; pfn += isolated) {
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700355 if (!pfn_valid(pfn) || cc->zone != page_zone(pfn_to_page(pfn)))
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100356 break;
357
358 /*
359 * On subsequent iterations ALIGN() is actually not needed,
360 * but we keep it that we not to complicate the code.
361 */
362 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
363 block_end_pfn = min(block_end_pfn, end_pfn);
364
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700365 isolated = isolate_freepages_block(cc, pfn, block_end_pfn,
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100366 &freelist, true);
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100367
368 /*
369 * In strict mode, isolate_freepages_block() returns 0 if
370 * there are any holes in the block (ie. invalid PFNs or
371 * non-free pages).
372 */
373 if (!isolated)
374 break;
375
376 /*
377 * If we managed to isolate pages, it is always (1 << n) *
378 * pageblock_nr_pages for some non-negative n. (Max order
379 * page may span two pageblocks).
380 */
381 }
382
383 /* split_free_page does not map the pages */
384 map_pages(&freelist);
385
386 if (pfn < end_pfn) {
387 /* Loop terminated early, cleanup. */
388 release_freepages(&freelist);
389 return 0;
390 }
391
392 /* We don't use freelists for anything. */
393 return pfn;
394}
395
Mel Gorman748446b2010-05-24 14:32:27 -0700396/* Update the number of anon and file isolated pages in the zone */
Mel Gormanc67fe372012-08-21 16:16:17 -0700397static void acct_isolated(struct zone *zone, bool locked, struct compact_control *cc)
Mel Gorman748446b2010-05-24 14:32:27 -0700398{
399 struct page *page;
Minchan Kimb9e84ac2011-10-31 17:06:44 -0700400 unsigned int count[2] = { 0, };
Mel Gorman748446b2010-05-24 14:32:27 -0700401
Minchan Kimb9e84ac2011-10-31 17:06:44 -0700402 list_for_each_entry(page, &cc->migratepages, lru)
403 count[!!page_is_file_cache(page)]++;
Mel Gorman748446b2010-05-24 14:32:27 -0700404
Mel Gormanc67fe372012-08-21 16:16:17 -0700405 /* If locked we can use the interrupt unsafe versions */
406 if (locked) {
407 __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
408 __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
409 } else {
410 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
411 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
412 }
Mel Gorman748446b2010-05-24 14:32:27 -0700413}
414
415/* Similar to reclaim, but different enough that they don't share logic */
416static bool too_many_isolated(struct zone *zone)
417{
Minchan Kimbc693042010-09-09 16:38:00 -0700418 unsigned long active, inactive, isolated;
Mel Gorman748446b2010-05-24 14:32:27 -0700419
420 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
421 zone_page_state(zone, NR_INACTIVE_ANON);
Minchan Kimbc693042010-09-09 16:38:00 -0700422 active = zone_page_state(zone, NR_ACTIVE_FILE) +
423 zone_page_state(zone, NR_ACTIVE_ANON);
Mel Gorman748446b2010-05-24 14:32:27 -0700424 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
425 zone_page_state(zone, NR_ISOLATED_ANON);
426
Minchan Kimbc693042010-09-09 16:38:00 -0700427 return isolated > (inactive + active) / 2;
Mel Gorman748446b2010-05-24 14:32:27 -0700428}
429
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100430/**
431 * isolate_migratepages_range() - isolate all migrate-able pages in range.
432 * @zone: Zone pages are in.
433 * @cc: Compaction control structure.
434 * @low_pfn: The first PFN of the range.
435 * @end_pfn: The one-past-the-last PFN of the range.
Minchan Kime46a2872012-10-08 16:33:48 -0700436 * @unevictable: true if it allows to isolate unevictable pages
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100437 *
438 * Isolate all pages that can be migrated from the range specified by
439 * [low_pfn, end_pfn). Returns zero if there is a fatal signal
440 * pending), otherwise PFN of the first page that was not scanned
441 * (which may be both less, equal to or more then end_pfn).
442 *
443 * Assumes that cc->migratepages is empty and cc->nr_migratepages is
444 * zero.
445 *
446 * Apart from cc->migratepages and cc->nr_migratetypes this function
447 * does not modify any cc's fields, in particular it does not modify
448 * (or read for that matter) cc->migrate_pfn.
Mel Gorman748446b2010-05-24 14:32:27 -0700449 */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100450unsigned long
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100451isolate_migratepages_range(struct zone *zone, struct compact_control *cc,
Minchan Kime46a2872012-10-08 16:33:48 -0700452 unsigned long low_pfn, unsigned long end_pfn, bool unevictable)
Mel Gorman748446b2010-05-24 14:32:27 -0700453{
Mel Gorman9927af742011-01-13 15:45:59 -0800454 unsigned long last_pageblock_nr = 0, pageblock_nr;
Mel Gormanb7aba692011-01-13 15:45:54 -0800455 unsigned long nr_scanned = 0, nr_isolated = 0;
Mel Gorman748446b2010-05-24 14:32:27 -0700456 struct list_head *migratelist = &cc->migratepages;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -0700457 isolate_mode_t mode = 0;
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700458 struct lruvec *lruvec;
Mel Gormanc67fe372012-08-21 16:16:17 -0700459 unsigned long flags;
Mel Gorman2a1402a2012-10-08 16:32:33 -0700460 bool locked = false;
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700461 struct page *page = NULL, *valid_page = NULL;
Vlastimil Babka50b5b092014-01-21 15:51:10 -0800462 bool skipped_async_unsuitable = false;
Mel Gorman748446b2010-05-24 14:32:27 -0700463
Mel Gorman748446b2010-05-24 14:32:27 -0700464 /*
465 * Ensure that there are not too many pages isolated from the LRU
466 * list by either parallel reclaimers or compaction. If there are,
467 * delay for some time until fewer pages are isolated
468 */
469 while (unlikely(too_many_isolated(zone))) {
Mel Gormanf9e35b32011-06-15 15:08:52 -0700470 /* async migration should just abort */
Linus Torvalds68e3e922012-06-03 20:05:57 -0700471 if (!cc->sync)
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100472 return 0;
Mel Gormanf9e35b32011-06-15 15:08:52 -0700473
Mel Gorman748446b2010-05-24 14:32:27 -0700474 congestion_wait(BLK_RW_ASYNC, HZ/10);
475
476 if (fatal_signal_pending(current))
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100477 return 0;
Mel Gorman748446b2010-05-24 14:32:27 -0700478 }
479
480 /* Time to isolate some pages for migration */
Andrea Arcangelib2eef8c2011-03-22 16:33:10 -0700481 cond_resched();
Mel Gorman748446b2010-05-24 14:32:27 -0700482 for (; low_pfn < end_pfn; low_pfn++) {
Andrea Arcangelib2eef8c2011-03-22 16:33:10 -0700483 /* give a chance to irqs before checking need_resched() */
Mel Gorman2a1402a2012-10-08 16:32:33 -0700484 if (locked && !((low_pfn+1) % SWAP_CLUSTER_MAX)) {
485 if (should_release_lock(&zone->lru_lock)) {
486 spin_unlock_irqrestore(&zone->lru_lock, flags);
487 locked = false;
488 }
Andrea Arcangelib2eef8c2011-03-22 16:33:10 -0700489 }
Mel Gormanc67fe372012-08-21 16:16:17 -0700490
Mel Gorman0bf380b2012-02-03 15:37:18 -0800491 /*
492 * migrate_pfn does not necessarily start aligned to a
493 * pageblock. Ensure that pfn_valid is called when moving
494 * into a new MAX_ORDER_NR_PAGES range in case of large
495 * memory holes within the zone
496 */
497 if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
498 if (!pfn_valid(low_pfn)) {
499 low_pfn += MAX_ORDER_NR_PAGES - 1;
500 continue;
501 }
502 }
503
Mel Gorman748446b2010-05-24 14:32:27 -0700504 if (!pfn_valid_within(low_pfn))
505 continue;
Mel Gormanb7aba692011-01-13 15:45:54 -0800506 nr_scanned++;
Mel Gorman748446b2010-05-24 14:32:27 -0700507
Mel Gormandc908602012-02-08 17:13:38 -0800508 /*
509 * Get the page and ensure the page is within the same zone.
510 * See the comment in isolate_freepages about overlapping
511 * nodes. It is deliberate that the new zone lock is not taken
512 * as memory compaction should not move pages between nodes.
513 */
Mel Gorman748446b2010-05-24 14:32:27 -0700514 page = pfn_to_page(low_pfn);
Mel Gormandc908602012-02-08 17:13:38 -0800515 if (page_zone(page) != zone)
516 continue;
517
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700518 if (!valid_page)
519 valid_page = page;
520
521 /* If isolation recently failed, do not retry */
522 pageblock_nr = low_pfn >> pageblock_order;
523 if (!isolation_suitable(cc, page))
524 goto next_pageblock;
525
Mel Gormandc908602012-02-08 17:13:38 -0800526 /* Skip if free */
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
578 /* Check if it is ok to still hold the lock */
579 locked = compact_checklock_irqsave(&zone->lru_lock, &flags,
580 locked, cc);
581 if (!locked || fatal_signal_pending(current))
582 break;
583
584 /* Recheck PageLRU and PageTransHuge under lock */
585 if (!PageLRU(page))
586 continue;
587 if (PageTransHuge(page)) {
Andrea Arcangelibc835012011-01-13 15:47:08 -0800588 low_pfn += (1 << compound_order(page)) - 1;
589 continue;
590 }
591
Linus Torvalds68e3e922012-06-03 20:05:57 -0700592 if (!cc->sync)
Mel Gormanc8244932012-01-12 17:19:38 -0800593 mode |= ISOLATE_ASYNC_MIGRATE;
594
Minchan Kime46a2872012-10-08 16:33:48 -0700595 if (unevictable)
596 mode |= ISOLATE_UNEVICTABLE;
597
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700598 lruvec = mem_cgroup_page_lruvec(page, zone);
599
Mel Gorman748446b2010-05-24 14:32:27 -0700600 /* Try isolate the page */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -0700601 if (__isolate_lru_page(page, mode) != 0)
Mel Gorman748446b2010-05-24 14:32:27 -0700602 continue;
603
Andrea Arcangelibc835012011-01-13 15:47:08 -0800604 VM_BUG_ON(PageTransCompound(page));
605
Mel Gorman748446b2010-05-24 14:32:27 -0700606 /* Successfully isolated */
Mel Gormanc89511a2012-10-08 16:32:45 -0700607 cc->finished_update_migrate = true;
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700608 del_page_from_lru_list(page, lruvec, page_lru(page));
Mel Gorman748446b2010-05-24 14:32:27 -0700609 list_add(&page->lru, migratelist);
Mel Gorman748446b2010-05-24 14:32:27 -0700610 cc->nr_migratepages++;
Mel Gormanb7aba692011-01-13 15:45:54 -0800611 nr_isolated++;
Mel Gorman748446b2010-05-24 14:32:27 -0700612
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800613check_compact_cluster:
Mel Gorman748446b2010-05-24 14:32:27 -0700614 /* Avoid isolating too much */
Hillf Danton31b83842012-01-10 15:07:59 -0800615 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
616 ++low_pfn;
Mel Gorman748446b2010-05-24 14:32:27 -0700617 break;
Hillf Danton31b83842012-01-10 15:07:59 -0800618 }
Mel Gorman2a1402a2012-10-08 16:32:33 -0700619
620 continue;
621
622next_pageblock:
Mel Gormana9aacbc2013-02-22 16:32:25 -0800623 low_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages) - 1;
Mel Gorman2a1402a2012-10-08 16:32:33 -0700624 last_pageblock_nr = pageblock_nr;
Mel Gorman748446b2010-05-24 14:32:27 -0700625 }
626
Mel Gormanc67fe372012-08-21 16:16:17 -0700627 acct_isolated(zone, locked, cc);
Mel Gorman748446b2010-05-24 14:32:27 -0700628
Mel Gormanc67fe372012-08-21 16:16:17 -0700629 if (locked)
630 spin_unlock_irqrestore(&zone->lru_lock, flags);
Mel Gorman748446b2010-05-24 14:32:27 -0700631
Vlastimil Babka50b5b092014-01-21 15:51:10 -0800632 /*
633 * Update the pageblock-skip information and cached scanner pfn,
634 * if the whole pageblock was scanned without isolating any page.
635 * This is not done when pageblock was skipped due to being unsuitable
636 * for async compaction, so that eventual sync compaction can try.
637 */
638 if (low_pfn == end_pfn && !skipped_async_unsuitable)
Mel Gormanc89511a2012-10-08 16:32:45 -0700639 update_pageblock_skip(cc, valid_page, nr_isolated, true);
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700640
Mel Gormanb7aba692011-01-13 15:45:54 -0800641 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
642
Minchan Kim010fc292012-12-20 15:05:06 -0800643 count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
Mel Gorman397487d2012-10-19 12:00:10 +0100644 if (nr_isolated)
Minchan Kim010fc292012-12-20 15:05:06 -0800645 count_compact_events(COMPACTISOLATED, nr_isolated);
Mel Gorman397487d2012-10-19 12:00:10 +0100646
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100647 return low_pfn;
648}
649
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100650#endif /* CONFIG_COMPACTION || CONFIG_CMA */
651#ifdef CONFIG_COMPACTION
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100652/*
653 * Based on information in the current compact_control, find blocks
654 * suitable for isolating free pages from and then isolate them.
655 */
656static void isolate_freepages(struct zone *zone,
657 struct compact_control *cc)
658{
659 struct page *page;
Cody P Schafer108bcc92013-02-22 16:35:23 -0800660 unsigned long high_pfn, low_pfn, pfn, z_end_pfn, end_pfn;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100661 int nr_freepages = cc->nr_freepages;
662 struct list_head *freelist = &cc->freepages;
663
664 /*
665 * Initialise the free scanner. The starting point is where we last
666 * scanned from (or the end of the zone if starting). The low point
667 * is the end of the pageblock the migration scanner is using.
668 */
669 pfn = cc->free_pfn;
Vlastimil Babka7ed695e2014-01-21 15:51:09 -0800670 low_pfn = ALIGN(cc->migrate_pfn + 1, pageblock_nr_pages);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100671
672 /*
673 * Take care that if the migration scanner is at the end of the zone
674 * that the free scanner does not accidentally move to the next zone
675 * in the next isolation cycle.
676 */
677 high_pfn = min(low_pfn, pfn);
678
Cody P Schafer108bcc92013-02-22 16:35:23 -0800679 z_end_pfn = zone_end_pfn(zone);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100680
681 /*
682 * Isolate free pages until enough are available to migrate the
683 * pages on cc->migratepages. We stop searching if the migrate
684 * and free page scanners meet or enough free pages are isolated.
685 */
Vlastimil Babka7ed695e2014-01-21 15:51:09 -0800686 for (; pfn >= low_pfn && cc->nr_migratepages > nr_freepages;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100687 pfn -= pageblock_nr_pages) {
688 unsigned long isolated;
689
David Rientjesf6ea3ad2013-09-30 13:45:03 -0700690 /*
691 * This can iterate a massively long zone without finding any
692 * suitable migration targets, so periodically check if we need
693 * to schedule.
694 */
695 cond_resched();
696
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100697 if (!pfn_valid(pfn))
698 continue;
699
700 /*
701 * Check for overlapping nodes/zones. It's possible on some
702 * configurations to have a setup like
703 * node0 node1 node0
704 * i.e. it's possible that all pages within a zones range of
705 * pages do not belong to a single zone.
706 */
707 page = pfn_to_page(pfn);
708 if (page_zone(page) != zone)
709 continue;
710
711 /* Check the block is suitable for migration */
Linus Torvalds68e3e922012-06-03 20:05:57 -0700712 if (!suitable_migration_target(page))
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100713 continue;
Linus Torvalds68e3e922012-06-03 20:05:57 -0700714
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700715 /* If isolation recently failed, do not retry */
716 if (!isolation_suitable(cc, page))
717 continue;
718
Mel Gormanf40d1e42012-10-08 16:32:36 -0700719 /* Found a block suitable for isolating free pages from */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100720 isolated = 0;
Mel Gorman60177d32012-12-06 19:01:14 +0000721
722 /*
723 * As pfn may not start aligned, pfn+pageblock_nr_page
724 * may cross a MAX_ORDER_NR_PAGES boundary and miss
725 * a pfn_valid check. Ensure isolate_freepages_block()
726 * only scans within a pageblock
727 */
728 end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
Cody P Schafer108bcc92013-02-22 16:35:23 -0800729 end_pfn = min(end_pfn, z_end_pfn);
Mel Gormanf40d1e42012-10-08 16:32:36 -0700730 isolated = isolate_freepages_block(cc, pfn, end_pfn,
731 freelist, false);
732 nr_freepages += isolated;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100733
734 /*
735 * Record the highest PFN we isolated pages from. When next
736 * looking for free pages, the search will restart here as
737 * page migration may have returned some pages to the allocator
738 */
Mel Gormanc89511a2012-10-08 16:32:45 -0700739 if (isolated) {
740 cc->finished_update_free = true;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100741 high_pfn = max(high_pfn, pfn);
Mel Gormanc89511a2012-10-08 16:32:45 -0700742 }
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100743 }
744
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100745 /* split_free_page does not map the pages */
746 map_pages(freelist);
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100747
Vlastimil Babka7ed695e2014-01-21 15:51:09 -0800748 /*
749 * If we crossed the migrate scanner, we want to keep it that way
750 * so that compact_finished() may detect this
751 */
752 if (pfn < low_pfn)
753 cc->free_pfn = max(pfn, zone->zone_start_pfn);
754 else
755 cc->free_pfn = high_pfn;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100756 cc->nr_freepages = nr_freepages;
Mel Gorman748446b2010-05-24 14:32:27 -0700757}
758
759/*
760 * This is a migrate-callback that "allocates" freepages by taking pages
761 * from the isolated freelists in the block we are migrating to.
762 */
763static struct page *compaction_alloc(struct page *migratepage,
764 unsigned long data,
765 int **result)
766{
767 struct compact_control *cc = (struct compact_control *)data;
768 struct page *freepage;
769
770 /* Isolate free pages if necessary */
771 if (list_empty(&cc->freepages)) {
772 isolate_freepages(cc->zone, cc);
773
774 if (list_empty(&cc->freepages))
775 return NULL;
776 }
777
778 freepage = list_entry(cc->freepages.next, struct page, lru);
779 list_del(&freepage->lru);
780 cc->nr_freepages--;
781
782 return freepage;
783}
784
785/*
786 * We cannot control nr_migratepages and nr_freepages fully when migration is
787 * running as migrate_pages() has no knowledge of compact_control. When
788 * migration is complete, we count the number of pages on the lists by hand.
789 */
790static void update_nr_listpages(struct compact_control *cc)
791{
792 int nr_migratepages = 0;
793 int nr_freepages = 0;
794 struct page *page;
795
796 list_for_each_entry(page, &cc->migratepages, lru)
797 nr_migratepages++;
798 list_for_each_entry(page, &cc->freepages, lru)
799 nr_freepages++;
800
801 cc->nr_migratepages = nr_migratepages;
802 cc->nr_freepages = nr_freepages;
803}
804
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100805/* possible outcome of isolate_migratepages */
806typedef enum {
807 ISOLATE_ABORT, /* Abort compaction now */
808 ISOLATE_NONE, /* No pages isolated, continue scanning */
809 ISOLATE_SUCCESS, /* Pages isolated, migrate */
810} isolate_migrate_t;
811
812/*
813 * Isolate all pages that can be migrated from the block pointed to by
814 * the migrate scanner within compact_control.
815 */
816static isolate_migrate_t isolate_migratepages(struct zone *zone,
817 struct compact_control *cc)
818{
819 unsigned long low_pfn, end_pfn;
820
821 /* Do not scan outside zone boundaries */
822 low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
823
824 /* Only scan within a pageblock boundary */
Mel Gormana9aacbc2013-02-22 16:32:25 -0800825 end_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100826
827 /* Do not cross the free scanner or scan within a memory hole */
828 if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
829 cc->migrate_pfn = end_pfn;
830 return ISOLATE_NONE;
831 }
832
833 /* Perform the isolation */
Minchan Kime46a2872012-10-08 16:33:48 -0700834 low_pfn = isolate_migratepages_range(zone, cc, low_pfn, end_pfn, false);
Shaohua Lie64c5232012-10-08 16:32:27 -0700835 if (!low_pfn || cc->contended)
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100836 return ISOLATE_ABORT;
837
838 cc->migrate_pfn = low_pfn;
839
840 return ISOLATE_SUCCESS;
841}
842
Mel Gorman748446b2010-05-24 14:32:27 -0700843static int compact_finished(struct zone *zone,
Andrea Arcangeli5a03b052011-01-13 15:47:11 -0800844 struct compact_control *cc)
Mel Gorman748446b2010-05-24 14:32:27 -0700845{
Mel Gorman8fb74b92013-01-11 14:32:16 -0800846 unsigned int order;
Andrea Arcangeli5a03b052011-01-13 15:47:11 -0800847 unsigned long watermark;
Mel Gorman56de7262010-05-24 14:32:30 -0700848
Mel Gorman748446b2010-05-24 14:32:27 -0700849 if (fatal_signal_pending(current))
850 return COMPACT_PARTIAL;
851
Mel Gorman753341a2012-10-08 16:32:40 -0700852 /* Compaction run completes if the migrate and free scanner meet */
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700853 if (cc->free_pfn <= cc->migrate_pfn) {
Vlastimil Babka55b7c4c2014-01-21 15:51:11 -0800854 /* Let the next compaction start anew. */
855 zone->compact_cached_migrate_pfn = zone->zone_start_pfn;
856 zone->compact_cached_free_pfn = zone_end_pfn(zone);
857
Mel Gorman62997022012-10-08 16:32:47 -0700858 /*
859 * Mark that the PG_migrate_skip information should be cleared
860 * by kswapd when it goes to sleep. kswapd does not set the
861 * flag itself as the decision to be clear should be directly
862 * based on an allocation request.
863 */
864 if (!current_is_kswapd())
865 zone->compact_blockskip_flush = true;
866
Mel Gorman748446b2010-05-24 14:32:27 -0700867 return COMPACT_COMPLETE;
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700868 }
Mel Gorman748446b2010-05-24 14:32:27 -0700869
Johannes Weiner82478fb2011-01-20 14:44:21 -0800870 /*
871 * order == -1 is expected when compacting via
872 * /proc/sys/vm/compact_memory
873 */
Mel Gorman56de7262010-05-24 14:32:30 -0700874 if (cc->order == -1)
875 return COMPACT_CONTINUE;
876
Michal Hocko3957c772011-06-15 15:08:25 -0700877 /* Compaction run is not finished if the watermark is not met */
878 watermark = low_wmark_pages(zone);
879 watermark += (1 << cc->order);
880
881 if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
882 return COMPACT_CONTINUE;
883
Mel Gorman56de7262010-05-24 14:32:30 -0700884 /* Direct compactor: Is a suitable page free? */
Mel Gorman8fb74b92013-01-11 14:32:16 -0800885 for (order = cc->order; order < MAX_ORDER; order++) {
886 struct free_area *area = &zone->free_area[order];
Mel Gorman56de7262010-05-24 14:32:30 -0700887
Mel Gorman8fb74b92013-01-11 14:32:16 -0800888 /* Job done if page is free of the right migratetype */
889 if (!list_empty(&area->free_list[cc->migratetype]))
890 return COMPACT_PARTIAL;
891
892 /* Job done if allocation would set block type */
893 if (cc->order >= pageblock_order && area->nr_free)
894 return COMPACT_PARTIAL;
Mel Gorman56de7262010-05-24 14:32:30 -0700895 }
896
Mel Gorman748446b2010-05-24 14:32:27 -0700897 return COMPACT_CONTINUE;
898}
899
Mel Gorman3e7d3442011-01-13 15:45:56 -0800900/*
901 * compaction_suitable: Is this suitable to run compaction on this zone now?
902 * Returns
903 * COMPACT_SKIPPED - If there are too few free pages for compaction
904 * COMPACT_PARTIAL - If the allocation would succeed without compaction
905 * COMPACT_CONTINUE - If compaction should run now
906 */
907unsigned long compaction_suitable(struct zone *zone, int order)
908{
909 int fragindex;
910 unsigned long watermark;
911
912 /*
Michal Hocko3957c772011-06-15 15:08:25 -0700913 * order == -1 is expected when compacting via
914 * /proc/sys/vm/compact_memory
915 */
916 if (order == -1)
917 return COMPACT_CONTINUE;
918
919 /*
Mel Gorman3e7d3442011-01-13 15:45:56 -0800920 * Watermarks for order-0 must be met for compaction. Note the 2UL.
921 * This is because during migration, copies of pages need to be
922 * allocated and for a short time, the footprint is higher
923 */
924 watermark = low_wmark_pages(zone) + (2UL << order);
925 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
926 return COMPACT_SKIPPED;
927
928 /*
929 * fragmentation index determines if allocation failures are due to
930 * low memory or external fragmentation
931 *
Shaohua Lia582a732011-06-15 15:08:49 -0700932 * index of -1000 implies allocations might succeed depending on
933 * watermarks
Mel Gorman3e7d3442011-01-13 15:45:56 -0800934 * index towards 0 implies failure is due to lack of memory
935 * index towards 1000 implies failure is due to fragmentation
936 *
937 * Only compact if a failure would be due to fragmentation.
938 */
939 fragindex = fragmentation_index(zone, order);
940 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
941 return COMPACT_SKIPPED;
942
Shaohua Lia582a732011-06-15 15:08:49 -0700943 if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
944 0, 0))
Mel Gorman3e7d3442011-01-13 15:45:56 -0800945 return COMPACT_PARTIAL;
946
947 return COMPACT_CONTINUE;
948}
949
Mel Gorman748446b2010-05-24 14:32:27 -0700950static int compact_zone(struct zone *zone, struct compact_control *cc)
951{
952 int ret;
Mel Gormanc89511a2012-10-08 16:32:45 -0700953 unsigned long start_pfn = zone->zone_start_pfn;
Cody P Schafer108bcc92013-02-22 16:35:23 -0800954 unsigned long end_pfn = zone_end_pfn(zone);
Mel Gorman748446b2010-05-24 14:32:27 -0700955
Mel Gorman3e7d3442011-01-13 15:45:56 -0800956 ret = compaction_suitable(zone, cc->order);
957 switch (ret) {
958 case COMPACT_PARTIAL:
959 case COMPACT_SKIPPED:
960 /* Compaction is likely to fail */
961 return ret;
962 case COMPACT_CONTINUE:
963 /* Fall through to compaction */
964 ;
965 }
966
Mel Gormanc89511a2012-10-08 16:32:45 -0700967 /*
Vlastimil Babkad3132e42014-01-21 15:51:08 -0800968 * Clear pageblock skip if there were failures recently and compaction
969 * is about to be retried after being deferred. kswapd does not do
970 * this reset as it'll reset the cached information when going to sleep.
971 */
972 if (compaction_restarting(zone, cc->order) && !current_is_kswapd())
973 __reset_isolation_suitable(zone);
974
975 /*
Mel Gormanc89511a2012-10-08 16:32:45 -0700976 * Setup to move all movable pages to the end of the zone. Used cached
977 * information on where the scanners should start but check that it
978 * is initialised by ensuring the values are within zone boundaries.
979 */
980 cc->migrate_pfn = zone->compact_cached_migrate_pfn;
981 cc->free_pfn = zone->compact_cached_free_pfn;
982 if (cc->free_pfn < start_pfn || cc->free_pfn > end_pfn) {
983 cc->free_pfn = end_pfn & ~(pageblock_nr_pages-1);
984 zone->compact_cached_free_pfn = cc->free_pfn;
985 }
986 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn > end_pfn) {
987 cc->migrate_pfn = start_pfn;
988 zone->compact_cached_migrate_pfn = cc->migrate_pfn;
989 }
Mel Gorman748446b2010-05-24 14:32:27 -0700990
Mel Gorman0eb927c2014-01-21 15:51:05 -0800991 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn, cc->free_pfn, end_pfn);
992
Mel Gorman748446b2010-05-24 14:32:27 -0700993 migrate_prep_local();
994
995 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
996 unsigned long nr_migrate, nr_remaining;
Minchan Kim9d502c12011-03-22 16:30:39 -0700997 int err;
Mel Gorman748446b2010-05-24 14:32:27 -0700998
Mel Gormanf9e35b32011-06-15 15:08:52 -0700999 switch (isolate_migratepages(zone, cc)) {
1000 case ISOLATE_ABORT:
1001 ret = COMPACT_PARTIAL;
Rafael Aquini5733c7d2012-12-11 16:02:47 -08001002 putback_movable_pages(&cc->migratepages);
Shaohua Lie64c5232012-10-08 16:32:27 -07001003 cc->nr_migratepages = 0;
Mel Gormanf9e35b32011-06-15 15:08:52 -07001004 goto out;
1005 case ISOLATE_NONE:
Mel Gorman748446b2010-05-24 14:32:27 -07001006 continue;
Mel Gormanf9e35b32011-06-15 15:08:52 -07001007 case ISOLATE_SUCCESS:
1008 ;
1009 }
Mel Gorman748446b2010-05-24 14:32:27 -07001010
1011 nr_migrate = cc->nr_migratepages;
Minchan Kim9d502c12011-03-22 16:30:39 -07001012 err = migrate_pages(&cc->migratepages, compaction_alloc,
Hugh Dickins9c620e22013-02-22 16:35:14 -08001013 (unsigned long)cc,
Mel Gorman7b2a2d42012-10-19 14:07:31 +01001014 cc->sync ? MIGRATE_SYNC_LIGHT : MIGRATE_ASYNC,
1015 MR_COMPACTION);
Mel Gorman748446b2010-05-24 14:32:27 -07001016 update_nr_listpages(cc);
1017 nr_remaining = cc->nr_migratepages;
1018
Mel Gormanb7aba692011-01-13 15:45:54 -08001019 trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
1020 nr_remaining);
Mel Gorman748446b2010-05-24 14:32:27 -07001021
Rafael Aquini5733c7d2012-12-11 16:02:47 -08001022 /* Release isolated pages not migrated */
Minchan Kim9d502c12011-03-22 16:30:39 -07001023 if (err) {
Rafael Aquini5733c7d2012-12-11 16:02:47 -08001024 putback_movable_pages(&cc->migratepages);
Mel Gorman748446b2010-05-24 14:32:27 -07001025 cc->nr_migratepages = 0;
Vlastimil Babka7ed695e2014-01-21 15:51:09 -08001026 /*
1027 * migrate_pages() may return -ENOMEM when scanners meet
1028 * and we want compact_finished() to detect it
1029 */
1030 if (err == -ENOMEM && cc->free_pfn > cc->migrate_pfn) {
David Rientjes4bf2bba2012-07-11 14:02:13 -07001031 ret = COMPACT_PARTIAL;
1032 goto out;
1033 }
Mel Gorman748446b2010-05-24 14:32:27 -07001034 }
Mel Gorman748446b2010-05-24 14:32:27 -07001035 }
1036
Mel Gormanf9e35b32011-06-15 15:08:52 -07001037out:
Mel Gorman748446b2010-05-24 14:32:27 -07001038 /* Release free pages and check accounting */
1039 cc->nr_freepages -= release_freepages(&cc->freepages);
1040 VM_BUG_ON(cc->nr_freepages != 0);
1041
Mel Gorman0eb927c2014-01-21 15:51:05 -08001042 trace_mm_compaction_end(ret);
1043
Mel Gorman748446b2010-05-24 14:32:27 -07001044 return ret;
1045}
Mel Gorman76ab0f52010-05-24 14:32:28 -07001046
Kyungmin Parkd43a87e2011-10-31 17:09:08 -07001047static unsigned long compact_zone_order(struct zone *zone,
Andrea Arcangeli5a03b052011-01-13 15:47:11 -08001048 int order, gfp_t gfp_mask,
Mel Gorman8fb74b92013-01-11 14:32:16 -08001049 bool sync, bool *contended)
Mel Gorman56de7262010-05-24 14:32:30 -07001050{
Shaohua Lie64c5232012-10-08 16:32:27 -07001051 unsigned long ret;
Mel Gorman56de7262010-05-24 14:32:30 -07001052 struct compact_control cc = {
1053 .nr_freepages = 0,
1054 .nr_migratepages = 0,
1055 .order = order,
1056 .migratetype = allocflags_to_migratetype(gfp_mask),
1057 .zone = zone,
Linus Torvalds68e3e922012-06-03 20:05:57 -07001058 .sync = sync,
Mel Gorman56de7262010-05-24 14:32:30 -07001059 };
1060 INIT_LIST_HEAD(&cc.freepages);
1061 INIT_LIST_HEAD(&cc.migratepages);
1062
Shaohua Lie64c5232012-10-08 16:32:27 -07001063 ret = compact_zone(zone, &cc);
1064
1065 VM_BUG_ON(!list_empty(&cc.freepages));
1066 VM_BUG_ON(!list_empty(&cc.migratepages));
1067
1068 *contended = cc.contended;
1069 return ret;
Mel Gorman56de7262010-05-24 14:32:30 -07001070}
1071
Mel Gorman5e771902010-05-24 14:32:31 -07001072int sysctl_extfrag_threshold = 500;
1073
Mel Gorman56de7262010-05-24 14:32:30 -07001074/**
1075 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1076 * @zonelist: The zonelist used for the current allocation
1077 * @order: The order of the current allocation
1078 * @gfp_mask: The GFP mask of the current allocation
1079 * @nodemask: The allowed nodes to allocate from
Mel Gorman77f1fe62011-01-13 15:45:57 -08001080 * @sync: Whether migration is synchronous or not
Mel Gorman661c4cb2012-10-08 16:32:31 -07001081 * @contended: Return value that is true if compaction was aborted due to lock contention
1082 * @page: Optionally capture a free page of the requested order during compaction
Mel Gorman56de7262010-05-24 14:32:30 -07001083 *
1084 * This is the main entry point for direct page compaction.
1085 */
1086unsigned long try_to_compact_pages(struct zonelist *zonelist,
Mel Gorman77f1fe62011-01-13 15:45:57 -08001087 int order, gfp_t gfp_mask, nodemask_t *nodemask,
Mel Gorman8fb74b92013-01-11 14:32:16 -08001088 bool sync, bool *contended)
Mel Gorman56de7262010-05-24 14:32:30 -07001089{
1090 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1091 int may_enter_fs = gfp_mask & __GFP_FS;
1092 int may_perform_io = gfp_mask & __GFP_IO;
Mel Gorman56de7262010-05-24 14:32:30 -07001093 struct zoneref *z;
1094 struct zone *zone;
1095 int rc = COMPACT_SKIPPED;
Bartlomiej Zolnierkiewiczd95ea5d2012-10-08 16:32:05 -07001096 int alloc_flags = 0;
Mel Gorman56de7262010-05-24 14:32:30 -07001097
Mel Gorman4ffb6332012-10-08 16:29:09 -07001098 /* Check if the GFP flags allow compaction */
Andrea Arcangelic5a73c32011-01-13 15:47:11 -08001099 if (!order || !may_enter_fs || !may_perform_io)
Mel Gorman56de7262010-05-24 14:32:30 -07001100 return rc;
1101
Minchan Kim010fc292012-12-20 15:05:06 -08001102 count_compact_event(COMPACTSTALL);
Mel Gorman56de7262010-05-24 14:32:30 -07001103
Bartlomiej Zolnierkiewiczd95ea5d2012-10-08 16:32:05 -07001104#ifdef CONFIG_CMA
1105 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
1106 alloc_flags |= ALLOC_CMA;
1107#endif
Mel Gorman56de7262010-05-24 14:32:30 -07001108 /* Compact each zone in the list */
1109 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1110 nodemask) {
Mel Gorman56de7262010-05-24 14:32:30 -07001111 int status;
1112
Mel Gormanc67fe372012-08-21 16:16:17 -07001113 status = compact_zone_order(zone, order, gfp_mask, sync,
Mel Gorman8fb74b92013-01-11 14:32:16 -08001114 contended);
Mel Gorman56de7262010-05-24 14:32:30 -07001115 rc = max(status, rc);
1116
Mel Gorman3e7d3442011-01-13 15:45:56 -08001117 /* If a normal allocation would succeed, stop compacting */
Bartlomiej Zolnierkiewiczd95ea5d2012-10-08 16:32:05 -07001118 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0,
1119 alloc_flags))
Mel Gorman56de7262010-05-24 14:32:30 -07001120 break;
1121 }
1122
1123 return rc;
1124}
1125
1126
Mel Gorman76ab0f52010-05-24 14:32:28 -07001127/* Compact all zones within a node */
Andrew Morton7103f162013-02-22 16:32:33 -08001128static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
Mel Gorman76ab0f52010-05-24 14:32:28 -07001129{
1130 int zoneid;
Mel Gorman76ab0f52010-05-24 14:32:28 -07001131 struct zone *zone;
1132
Mel Gorman76ab0f52010-05-24 14:32:28 -07001133 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
Mel Gorman76ab0f52010-05-24 14:32:28 -07001134
1135 zone = &pgdat->node_zones[zoneid];
1136 if (!populated_zone(zone))
1137 continue;
1138
Rik van Riel7be62de2012-03-21 16:33:52 -07001139 cc->nr_freepages = 0;
1140 cc->nr_migratepages = 0;
1141 cc->zone = zone;
1142 INIT_LIST_HEAD(&cc->freepages);
1143 INIT_LIST_HEAD(&cc->migratepages);
Mel Gorman76ab0f52010-05-24 14:32:28 -07001144
Dan Carpenteraad6ec32012-03-21 16:33:54 -07001145 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
Rik van Riel7be62de2012-03-21 16:33:52 -07001146 compact_zone(zone, cc);
Mel Gorman76ab0f52010-05-24 14:32:28 -07001147
Rik van Rielaff62242012-03-21 16:33:52 -07001148 if (cc->order > 0) {
Vlastimil Babkade6c60a2014-01-21 15:51:07 -08001149 if (zone_watermark_ok(zone, cc->order,
1150 low_wmark_pages(zone), 0, 0))
1151 compaction_defer_reset(zone, cc->order, false);
Rik van Rielaff62242012-03-21 16:33:52 -07001152 /* Currently async compaction is never deferred. */
Vlastimil Babkade6c60a2014-01-21 15:51:07 -08001153 else if (cc->sync)
Rik van Rielaff62242012-03-21 16:33:52 -07001154 defer_compaction(zone, cc->order);
1155 }
1156
Rik van Riel7be62de2012-03-21 16:33:52 -07001157 VM_BUG_ON(!list_empty(&cc->freepages));
1158 VM_BUG_ON(!list_empty(&cc->migratepages));
Mel Gorman76ab0f52010-05-24 14:32:28 -07001159 }
Mel Gorman76ab0f52010-05-24 14:32:28 -07001160}
1161
Andrew Morton7103f162013-02-22 16:32:33 -08001162void compact_pgdat(pg_data_t *pgdat, int order)
Rik van Riel7be62de2012-03-21 16:33:52 -07001163{
1164 struct compact_control cc = {
1165 .order = order,
Linus Torvalds68e3e922012-06-03 20:05:57 -07001166 .sync = false,
Rik van Riel7be62de2012-03-21 16:33:52 -07001167 };
1168
Mel Gorman3a7200a2013-09-11 14:22:19 -07001169 if (!order)
1170 return;
1171
Andrew Morton7103f162013-02-22 16:32:33 -08001172 __compact_pgdat(pgdat, &cc);
Rik van Riel7be62de2012-03-21 16:33:52 -07001173}
1174
Andrew Morton7103f162013-02-22 16:32:33 -08001175static void compact_node(int nid)
Rik van Riel7be62de2012-03-21 16:33:52 -07001176{
Rik van Riel7be62de2012-03-21 16:33:52 -07001177 struct compact_control cc = {
1178 .order = -1,
Linus Torvalds68e3e922012-06-03 20:05:57 -07001179 .sync = true,
Rik van Riel7be62de2012-03-21 16:33:52 -07001180 };
1181
Andrew Morton7103f162013-02-22 16:32:33 -08001182 __compact_pgdat(NODE_DATA(nid), &cc);
Rik van Riel7be62de2012-03-21 16:33:52 -07001183}
1184
Mel Gorman76ab0f52010-05-24 14:32:28 -07001185/* Compact all nodes in the system */
Jason Liu7964c062013-01-11 14:31:47 -08001186static void compact_nodes(void)
Mel Gorman76ab0f52010-05-24 14:32:28 -07001187{
1188 int nid;
1189
Hugh Dickins8575ec22012-03-21 16:33:53 -07001190 /* Flush pending updates to the LRU lists */
1191 lru_add_drain_all();
1192
Mel Gorman76ab0f52010-05-24 14:32:28 -07001193 for_each_online_node(nid)
1194 compact_node(nid);
Mel Gorman76ab0f52010-05-24 14:32:28 -07001195}
1196
1197/* The written value is actually unused, all memory is compacted */
1198int sysctl_compact_memory;
1199
1200/* This is the entry point for compacting all nodes via /proc/sys/vm */
1201int sysctl_compaction_handler(struct ctl_table *table, int write,
1202 void __user *buffer, size_t *length, loff_t *ppos)
1203{
1204 if (write)
Jason Liu7964c062013-01-11 14:31:47 -08001205 compact_nodes();
Mel Gorman76ab0f52010-05-24 14:32:28 -07001206
1207 return 0;
1208}
Mel Gormaned4a6d72010-05-24 14:32:29 -07001209
Mel Gorman5e771902010-05-24 14:32:31 -07001210int sysctl_extfrag_handler(struct ctl_table *table, int write,
1211 void __user *buffer, size_t *length, loff_t *ppos)
1212{
1213 proc_dointvec_minmax(table, write, buffer, length, ppos);
1214
1215 return 0;
1216}
1217
Mel Gormaned4a6d72010-05-24 14:32:29 -07001218#if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
Kay Sievers10fbcf42011-12-21 14:48:43 -08001219ssize_t sysfs_compact_node(struct device *dev,
1220 struct device_attribute *attr,
Mel Gormaned4a6d72010-05-24 14:32:29 -07001221 const char *buf, size_t count)
1222{
Hugh Dickins8575ec22012-03-21 16:33:53 -07001223 int nid = dev->id;
1224
1225 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1226 /* Flush pending updates to the LRU lists */
1227 lru_add_drain_all();
1228
1229 compact_node(nid);
1230 }
Mel Gormaned4a6d72010-05-24 14:32:29 -07001231
1232 return count;
1233}
Kay Sievers10fbcf42011-12-21 14:48:43 -08001234static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
Mel Gormaned4a6d72010-05-24 14:32:29 -07001235
1236int compaction_register_node(struct node *node)
1237{
Kay Sievers10fbcf42011-12-21 14:48:43 -08001238 return device_create_file(&node->dev, &dev_attr_compact);
Mel Gormaned4a6d72010-05-24 14:32:29 -07001239}
1240
1241void compaction_unregister_node(struct node *node)
1242{
Kay Sievers10fbcf42011-12-21 14:48:43 -08001243 return device_remove_file(&node->dev, &dev_attr_compact);
Mel Gormaned4a6d72010-05-24 14:32:29 -07001244}
1245#endif /* CONFIG_SYSFS && CONFIG_NUMA */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001246
1247#endif /* CONFIG_COMPACTION */