blob: 10cd757f1006b8646b6cf1a386b8b0e5b5d049f0 [file] [log] [blame]
Greg Kroah-Hartmanb2441312017-11-01 15:07:57 +01001// SPDX-License-Identifier: GPL-2.0
Mel Gorman748446b2010-05-24 14:32:27 -07002/*
3 * linux/mm/compaction.c
4 *
5 * Memory compaction for the reduction of external fragmentation. Note that
6 * this heavily depends upon page migration to do all the real heavy
7 * lifting
8 *
9 * Copyright IBM Corp. 2007-2010 Mel Gorman <mel@csn.ul.ie>
10 */
Vlastimil Babka698b1b32016-03-17 14:18:08 -070011#include <linux/cpu.h>
Mel Gorman748446b2010-05-24 14:32:27 -070012#include <linux/swap.h>
13#include <linux/migrate.h>
14#include <linux/compaction.h>
15#include <linux/mm_inline.h>
Ingo Molnar174cd4b2017-02-02 19:15:33 +010016#include <linux/sched/signal.h>
Mel Gorman748446b2010-05-24 14:32:27 -070017#include <linux/backing-dev.h>
Mel Gorman76ab0f52010-05-24 14:32:28 -070018#include <linux/sysctl.h>
Mel Gormaned4a6d72010-05-24 14:32:29 -070019#include <linux/sysfs.h>
Minchan Kim194159f2013-02-22 16:33:58 -080020#include <linux/page-isolation.h>
Andrey Ryabininb8c73fc2015-02-13 14:39:28 -080021#include <linux/kasan.h>
Vlastimil Babka698b1b32016-03-17 14:18:08 -070022#include <linux/kthread.h>
23#include <linux/freezer.h>
Joonsoo Kim83358ec2016-07-26 15:23:43 -070024#include <linux/page_owner.h>
Mel Gorman748446b2010-05-24 14:32:27 -070025#include "internal.h"
26
Minchan Kim010fc292012-12-20 15:05:06 -080027#ifdef CONFIG_COMPACTION
28static inline void count_compact_event(enum vm_event_item item)
29{
30 count_vm_event(item);
31}
32
33static inline void count_compact_events(enum vm_event_item item, long delta)
34{
35 count_vm_events(item, delta);
36}
37#else
38#define count_compact_event(item) do { } while (0)
39#define count_compact_events(item, delta) do { } while (0)
40#endif
41
Michal Nazarewiczff9543f2011-12-29 13:09:50 +010042#if defined CONFIG_COMPACTION || defined CONFIG_CMA
43
Mel Gormanb7aba692011-01-13 15:45:54 -080044#define CREATE_TRACE_POINTS
45#include <trace/events/compaction.h>
46
Vlastimil Babka06b66402016-05-19 17:11:48 -070047#define block_start_pfn(pfn, order) round_down(pfn, 1UL << (order))
48#define block_end_pfn(pfn, order) ALIGN((pfn) + 1, 1UL << (order))
49#define pageblock_start_pfn(pfn) block_start_pfn(pfn, pageblock_order)
50#define pageblock_end_pfn(pfn) block_end_pfn(pfn, pageblock_order)
51
Mel Gorman748446b2010-05-24 14:32:27 -070052static unsigned long release_freepages(struct list_head *freelist)
53{
54 struct page *page, *next;
Vlastimil Babka6bace092014-12-10 15:43:31 -080055 unsigned long high_pfn = 0;
Mel Gorman748446b2010-05-24 14:32:27 -070056
57 list_for_each_entry_safe(page, next, freelist, lru) {
Vlastimil Babka6bace092014-12-10 15:43:31 -080058 unsigned long pfn = page_to_pfn(page);
Mel Gorman748446b2010-05-24 14:32:27 -070059 list_del(&page->lru);
60 __free_page(page);
Vlastimil Babka6bace092014-12-10 15:43:31 -080061 if (pfn > high_pfn)
62 high_pfn = pfn;
Mel Gorman748446b2010-05-24 14:32:27 -070063 }
64
Vlastimil Babka6bace092014-12-10 15:43:31 -080065 return high_pfn;
Mel Gorman748446b2010-05-24 14:32:27 -070066}
67
Michal Nazarewiczff9543f2011-12-29 13:09:50 +010068static void map_pages(struct list_head *list)
69{
Joonsoo Kim66c64222016-07-26 15:23:40 -070070 unsigned int i, order, nr_pages;
71 struct page *page, *next;
72 LIST_HEAD(tmp_list);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +010073
Joonsoo Kim66c64222016-07-26 15:23:40 -070074 list_for_each_entry_safe(page, next, list, lru) {
75 list_del(&page->lru);
76
77 order = page_private(page);
78 nr_pages = 1 << order;
Joonsoo Kim66c64222016-07-26 15:23:40 -070079
Joonsoo Kim46f24fd2016-07-26 15:23:58 -070080 post_alloc_hook(page, order, __GFP_MOVABLE);
Joonsoo Kim66c64222016-07-26 15:23:40 -070081 if (order)
82 split_page(page, order);
83
84 for (i = 0; i < nr_pages; i++) {
85 list_add(&page->lru, &tmp_list);
86 page++;
87 }
Michal Nazarewiczff9543f2011-12-29 13:09:50 +010088 }
Joonsoo Kim66c64222016-07-26 15:23:40 -070089
90 list_splice(&tmp_list, list);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +010091}
92
Mel Gormanbb13ffe2012-10-08 16:32:41 -070093#ifdef CONFIG_COMPACTION
Joonsoo Kim24e27162015-02-11 15:27:09 -080094
Minchan Kimbda807d2016-07-26 15:23:05 -070095int PageMovable(struct page *page)
96{
97 struct address_space *mapping;
98
99 VM_BUG_ON_PAGE(!PageLocked(page), page);
100 if (!__PageMovable(page))
101 return 0;
102
103 mapping = page_mapping(page);
104 if (mapping && mapping->a_ops && mapping->a_ops->isolate_page)
105 return 1;
106
107 return 0;
108}
109EXPORT_SYMBOL(PageMovable);
110
111void __SetPageMovable(struct page *page, struct address_space *mapping)
112{
113 VM_BUG_ON_PAGE(!PageLocked(page), page);
114 VM_BUG_ON_PAGE((unsigned long)mapping & PAGE_MAPPING_MOVABLE, page);
115 page->mapping = (void *)((unsigned long)mapping | PAGE_MAPPING_MOVABLE);
116}
117EXPORT_SYMBOL(__SetPageMovable);
118
119void __ClearPageMovable(struct page *page)
120{
121 VM_BUG_ON_PAGE(!PageLocked(page), page);
122 VM_BUG_ON_PAGE(!PageMovable(page), page);
123 /*
124 * Clear registered address_space val with keeping PAGE_MAPPING_MOVABLE
125 * flag so that VM can catch up released page by driver after isolation.
126 * With it, VM migration doesn't try to put it back.
127 */
128 page->mapping = (void *)((unsigned long)page->mapping &
129 PAGE_MAPPING_MOVABLE);
130}
131EXPORT_SYMBOL(__ClearPageMovable);
132
Joonsoo Kim24e27162015-02-11 15:27:09 -0800133/* Do not skip compaction more than 64 times */
134#define COMPACT_MAX_DEFER_SHIFT 6
135
136/*
137 * Compaction is deferred when compaction fails to result in a page
138 * allocation success. 1 << compact_defer_limit compactions are skipped up
139 * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
140 */
141void defer_compaction(struct zone *zone, int order)
142{
143 zone->compact_considered = 0;
144 zone->compact_defer_shift++;
145
146 if (order < zone->compact_order_failed)
147 zone->compact_order_failed = order;
148
149 if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
150 zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
151
152 trace_mm_compaction_defer_compaction(zone, order);
153}
154
155/* Returns true if compaction should be skipped this time */
156bool compaction_deferred(struct zone *zone, int order)
157{
158 unsigned long defer_limit = 1UL << zone->compact_defer_shift;
159
160 if (order < zone->compact_order_failed)
161 return false;
162
163 /* Avoid possible overflow */
164 if (++zone->compact_considered > defer_limit)
165 zone->compact_considered = defer_limit;
166
167 if (zone->compact_considered >= defer_limit)
168 return false;
169
170 trace_mm_compaction_deferred(zone, order);
171
172 return true;
173}
174
175/*
176 * Update defer tracking counters after successful compaction of given order,
177 * which means an allocation either succeeded (alloc_success == true) or is
178 * expected to succeed.
179 */
180void compaction_defer_reset(struct zone *zone, int order,
181 bool alloc_success)
182{
183 if (alloc_success) {
184 zone->compact_considered = 0;
185 zone->compact_defer_shift = 0;
186 }
187 if (order >= zone->compact_order_failed)
188 zone->compact_order_failed = order + 1;
189
190 trace_mm_compaction_defer_reset(zone, order);
191}
192
193/* Returns true if restarting compaction after many failures */
194bool compaction_restarting(struct zone *zone, int order)
195{
196 if (order < zone->compact_order_failed)
197 return false;
198
199 return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
200 zone->compact_considered >= 1UL << zone->compact_defer_shift;
201}
202
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700203/* Returns true if the pageblock should be scanned for pages to isolate. */
204static inline bool isolation_suitable(struct compact_control *cc,
205 struct page *page)
206{
207 if (cc->ignore_skip_hint)
208 return true;
209
210 return !get_pageblock_skip(page);
211}
212
Vlastimil Babka023336412015-09-08 15:02:42 -0700213static void reset_cached_positions(struct zone *zone)
214{
215 zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
216 zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
Joonsoo Kim623446e2016-03-15 14:57:45 -0700217 zone->compact_cached_free_pfn =
Vlastimil Babka06b66402016-05-19 17:11:48 -0700218 pageblock_start_pfn(zone_end_pfn(zone) - 1);
Vlastimil Babka023336412015-09-08 15:02:42 -0700219}
220
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700221/*
Vlastimil Babkab527cfe2017-11-17 15:26:34 -0800222 * Compound pages of >= pageblock_order should consistenly be skipped until
223 * released. It is always pointless to compact pages of such order (if they are
224 * migratable), and the pageblocks they occupy cannot contain any free pages.
David Rientjes21dc7e02017-11-17 15:26:30 -0800225 */
Vlastimil Babkab527cfe2017-11-17 15:26:34 -0800226static bool pageblock_skip_persistent(struct page *page)
David Rientjes21dc7e02017-11-17 15:26:30 -0800227{
Vlastimil Babkab527cfe2017-11-17 15:26:34 -0800228 if (!PageCompound(page))
David Rientjes21dc7e02017-11-17 15:26:30 -0800229 return false;
Vlastimil Babkab527cfe2017-11-17 15:26:34 -0800230
231 page = compound_head(page);
232
233 if (compound_order(page) >= pageblock_order)
234 return true;
235
236 return false;
David Rientjes21dc7e02017-11-17 15:26:30 -0800237}
238
239/*
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700240 * This function is called to clear all cached information on pageblocks that
241 * should be skipped for page isolation when the migrate and free page scanner
242 * meet.
243 */
Mel Gorman62997022012-10-08 16:32:47 -0700244static void __reset_isolation_suitable(struct zone *zone)
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700245{
246 unsigned long start_pfn = zone->zone_start_pfn;
Cody P Schafer108bcc92013-02-22 16:35:23 -0800247 unsigned long end_pfn = zone_end_pfn(zone);
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700248 unsigned long pfn;
249
Mel Gorman62997022012-10-08 16:32:47 -0700250 zone->compact_blockskip_flush = false;
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700251
252 /* Walk the zone and mark every pageblock as suitable for isolation */
253 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
254 struct page *page;
255
256 cond_resched();
257
Michal Hockoccbe1e42017-07-06 15:38:00 -0700258 page = pfn_to_online_page(pfn);
259 if (!page)
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700260 continue;
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700261 if (zone != page_zone(page))
262 continue;
Vlastimil Babkab527cfe2017-11-17 15:26:34 -0800263 if (pageblock_skip_persistent(page))
David Rientjes21dc7e02017-11-17 15:26:30 -0800264 continue;
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700265
266 clear_pageblock_skip(page);
267 }
Vlastimil Babka023336412015-09-08 15:02:42 -0700268
269 reset_cached_positions(zone);
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700270}
271
Mel Gorman62997022012-10-08 16:32:47 -0700272void reset_isolation_suitable(pg_data_t *pgdat)
273{
274 int zoneid;
275
276 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
277 struct zone *zone = &pgdat->node_zones[zoneid];
278 if (!populated_zone(zone))
279 continue;
280
281 /* Only flush if a full compaction finished recently */
282 if (zone->compact_blockskip_flush)
283 __reset_isolation_suitable(zone);
284 }
285}
286
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700287/*
288 * If no pages were isolated then mark this pageblock to be skipped in the
Mel Gorman62997022012-10-08 16:32:47 -0700289 * future. The information is later cleared by __reset_isolation_suitable().
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700290 */
Mel Gormanc89511a2012-10-08 16:32:45 -0700291static void update_pageblock_skip(struct compact_control *cc,
292 struct page *page, unsigned long nr_isolated,
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700293 bool migrate_scanner)
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700294{
Mel Gormanc89511a2012-10-08 16:32:45 -0700295 struct zone *zone = cc->zone;
David Rientjes35979ef2014-06-04 16:08:27 -0700296 unsigned long pfn;
Joonsoo Kim6815bf32013-12-18 17:08:52 -0800297
Vlastimil Babka2583d672017-11-17 15:26:38 -0800298 if (cc->no_set_skip_hint)
Joonsoo Kim6815bf32013-12-18 17:08:52 -0800299 return;
300
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700301 if (!page)
302 return;
303
David Rientjes35979ef2014-06-04 16:08:27 -0700304 if (nr_isolated)
305 return;
306
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700307 set_pageblock_skip(page);
Mel Gormanc89511a2012-10-08 16:32:45 -0700308
David Rientjes35979ef2014-06-04 16:08:27 -0700309 pfn = page_to_pfn(page);
310
311 /* Update where async and sync compaction should restart */
312 if (migrate_scanner) {
David Rientjes35979ef2014-06-04 16:08:27 -0700313 if (pfn > zone->compact_cached_migrate_pfn[0])
314 zone->compact_cached_migrate_pfn[0] = pfn;
David Rientjese0b9dae2014-06-04 16:08:28 -0700315 if (cc->mode != MIGRATE_ASYNC &&
316 pfn > zone->compact_cached_migrate_pfn[1])
David Rientjes35979ef2014-06-04 16:08:27 -0700317 zone->compact_cached_migrate_pfn[1] = pfn;
318 } else {
David Rientjes35979ef2014-06-04 16:08:27 -0700319 if (pfn < zone->compact_cached_free_pfn)
320 zone->compact_cached_free_pfn = pfn;
Mel Gormanc89511a2012-10-08 16:32:45 -0700321 }
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700322}
323#else
324static inline bool isolation_suitable(struct compact_control *cc,
325 struct page *page)
326{
327 return true;
328}
329
Vlastimil Babkab527cfe2017-11-17 15:26:34 -0800330static inline bool pageblock_skip_persistent(struct page *page)
David Rientjes21dc7e02017-11-17 15:26:30 -0800331{
332 return false;
333}
334
335static inline void update_pageblock_skip(struct compact_control *cc,
Mel Gormanc89511a2012-10-08 16:32:45 -0700336 struct page *page, unsigned long nr_isolated,
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700337 bool migrate_scanner)
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700338{
339}
340#endif /* CONFIG_COMPACTION */
341
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700342/*
343 * Compaction requires the taking of some coarse locks that are potentially
344 * very heavily contended. For async compaction, back out if the lock cannot
345 * be taken immediately. For sync compaction, spin on the lock if needed.
346 *
347 * Returns true if the lock is held
348 * Returns false if the lock is not held and compaction should abort
349 */
350static bool compact_trylock_irqsave(spinlock_t *lock, unsigned long *flags,
351 struct compact_control *cc)
Mel Gorman2a1402a2012-10-08 16:32:33 -0700352{
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700353 if (cc->mode == MIGRATE_ASYNC) {
354 if (!spin_trylock_irqsave(lock, *flags)) {
Vlastimil Babkac3486f52016-07-28 15:49:30 -0700355 cc->contended = true;
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700356 return false;
357 }
358 } else {
359 spin_lock_irqsave(lock, *flags);
360 }
Vlastimil Babka1f9efde2014-10-09 15:27:14 -0700361
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700362 return true;
Mel Gorman2a1402a2012-10-08 16:32:33 -0700363}
364
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100365/*
Mel Gormanc67fe372012-08-21 16:16:17 -0700366 * Compaction requires the taking of some coarse locks that are potentially
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700367 * very heavily contended. The lock should be periodically unlocked to avoid
368 * having disabled IRQs for a long time, even when there is nobody waiting on
369 * the lock. It might also be that allowing the IRQs will result in
370 * need_resched() becoming true. If scheduling is needed, async compaction
371 * aborts. Sync compaction schedules.
372 * Either compaction type will also abort if a fatal signal is pending.
373 * In either case if the lock was locked, it is dropped and not regained.
Mel Gormanc67fe372012-08-21 16:16:17 -0700374 *
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700375 * Returns true if compaction should abort due to fatal signal pending, or
376 * async compaction due to need_resched()
377 * Returns false when compaction can continue (sync compaction might have
378 * scheduled)
Mel Gormanc67fe372012-08-21 16:16:17 -0700379 */
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700380static bool compact_unlock_should_abort(spinlock_t *lock,
381 unsigned long flags, bool *locked, struct compact_control *cc)
Mel Gormanc67fe372012-08-21 16:16:17 -0700382{
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700383 if (*locked) {
384 spin_unlock_irqrestore(lock, flags);
385 *locked = false;
386 }
Vlastimil Babka1f9efde2014-10-09 15:27:14 -0700387
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700388 if (fatal_signal_pending(current)) {
Vlastimil Babkac3486f52016-07-28 15:49:30 -0700389 cc->contended = true;
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700390 return true;
391 }
Mel Gormanc67fe372012-08-21 16:16:17 -0700392
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700393 if (need_resched()) {
David Rientjese0b9dae2014-06-04 16:08:28 -0700394 if (cc->mode == MIGRATE_ASYNC) {
Vlastimil Babkac3486f52016-07-28 15:49:30 -0700395 cc->contended = true;
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700396 return true;
Mel Gormanc67fe372012-08-21 16:16:17 -0700397 }
Mel Gormanc67fe372012-08-21 16:16:17 -0700398 cond_resched();
Mel Gormanc67fe372012-08-21 16:16:17 -0700399 }
400
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700401 return false;
Mel Gormanc67fe372012-08-21 16:16:17 -0700402}
403
Vlastimil Babkabe976572014-06-04 16:10:41 -0700404/*
405 * Aside from avoiding lock contention, compaction also periodically checks
406 * need_resched() and either schedules in sync compaction or aborts async
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700407 * compaction. This is similar to what compact_unlock_should_abort() does, but
Vlastimil Babkabe976572014-06-04 16:10:41 -0700408 * is used where no lock is concerned.
409 *
410 * Returns false when no scheduling was needed, or sync compaction scheduled.
411 * Returns true when async compaction should abort.
412 */
413static inline bool compact_should_abort(struct compact_control *cc)
414{
415 /* async compaction aborts if contended */
416 if (need_resched()) {
417 if (cc->mode == MIGRATE_ASYNC) {
Vlastimil Babkac3486f52016-07-28 15:49:30 -0700418 cc->contended = true;
Vlastimil Babkabe976572014-06-04 16:10:41 -0700419 return true;
420 }
421
422 cond_resched();
423 }
424
425 return false;
426}
427
Mel Gormanc67fe372012-08-21 16:16:17 -0700428/*
Jerome Marchand9e4be472013-11-12 15:07:12 -0800429 * Isolate free pages onto a private freelist. If @strict is true, will abort
430 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
431 * (even though it may still end up isolating some pages).
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100432 */
Mel Gormanf40d1e42012-10-08 16:32:36 -0700433static unsigned long isolate_freepages_block(struct compact_control *cc,
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700434 unsigned long *start_pfn,
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100435 unsigned long end_pfn,
436 struct list_head *freelist,
437 bool strict)
Mel Gorman748446b2010-05-24 14:32:27 -0700438{
Mel Gormanb7aba692011-01-13 15:45:54 -0800439 int nr_scanned = 0, total_isolated = 0;
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700440 struct page *cursor, *valid_page = NULL;
Xiubo Lib8b2d822014-10-09 15:28:21 -0700441 unsigned long flags = 0;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700442 bool locked = false;
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700443 unsigned long blockpfn = *start_pfn;
Joonsoo Kim66c64222016-07-26 15:23:40 -0700444 unsigned int order;
Mel Gorman748446b2010-05-24 14:32:27 -0700445
Mel Gorman748446b2010-05-24 14:32:27 -0700446 cursor = pfn_to_page(blockpfn);
447
Mel Gormanf40d1e42012-10-08 16:32:36 -0700448 /* Isolate free pages. */
Mel Gorman748446b2010-05-24 14:32:27 -0700449 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
Joonsoo Kim66c64222016-07-26 15:23:40 -0700450 int isolated;
Mel Gorman748446b2010-05-24 14:32:27 -0700451 struct page *page = cursor;
452
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700453 /*
454 * Periodically drop the lock (if held) regardless of its
455 * contention, to give chance to IRQs. Abort if fatal signal
456 * pending or async compaction detects need_resched()
457 */
458 if (!(blockpfn % SWAP_CLUSTER_MAX)
459 && compact_unlock_should_abort(&cc->zone->lock, flags,
460 &locked, cc))
461 break;
462
Mel Gormanb7aba692011-01-13 15:45:54 -0800463 nr_scanned++;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700464 if (!pfn_valid_within(blockpfn))
Laura Abbott2af120b2014-03-10 15:49:44 -0700465 goto isolate_fail;
466
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700467 if (!valid_page)
468 valid_page = page;
Vlastimil Babka9fcd6d22015-09-08 15:02:49 -0700469
470 /*
471 * For compound pages such as THP and hugetlbfs, we can save
472 * potentially a lot of iterations if we skip them at once.
473 * The check is racy, but we can consider only valid values
474 * and the only danger is skipping too much.
475 */
476 if (PageCompound(page)) {
David Rientjes21dc7e02017-11-17 15:26:30 -0800477 const unsigned int order = compound_order(page);
Vlastimil Babka9fcd6d22015-09-08 15:02:49 -0700478
Vlastimil Babkad3c85ba2017-11-17 15:26:41 -0800479 if (likely(order < MAX_ORDER)) {
David Rientjes21dc7e02017-11-17 15:26:30 -0800480 blockpfn += (1UL << order) - 1;
481 cursor += (1UL << order) - 1;
Vlastimil Babka9fcd6d22015-09-08 15:02:49 -0700482 }
Vlastimil Babka9fcd6d22015-09-08 15:02:49 -0700483 goto isolate_fail;
484 }
485
Mel Gormanf40d1e42012-10-08 16:32:36 -0700486 if (!PageBuddy(page))
Laura Abbott2af120b2014-03-10 15:49:44 -0700487 goto isolate_fail;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700488
489 /*
Vlastimil Babka69b71892014-10-09 15:27:18 -0700490 * If we already hold the lock, we can skip some rechecking.
491 * Note that if we hold the lock now, checked_pageblock was
492 * already set in some previous iteration (or strict is true),
493 * so it is correct to skip the suitable migration target
494 * recheck as well.
Mel Gormanf40d1e42012-10-08 16:32:36 -0700495 */
Vlastimil Babka69b71892014-10-09 15:27:18 -0700496 if (!locked) {
497 /*
498 * The zone lock must be held to isolate freepages.
499 * Unfortunately this is a very coarse lock and can be
500 * heavily contended if there are parallel allocations
501 * or parallel compactions. For async compaction do not
502 * spin on the lock and we acquire the lock as late as
503 * possible.
504 */
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700505 locked = compact_trylock_irqsave(&cc->zone->lock,
506 &flags, cc);
Vlastimil Babka69b71892014-10-09 15:27:18 -0700507 if (!locked)
508 break;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700509
Vlastimil Babka69b71892014-10-09 15:27:18 -0700510 /* Recheck this is a buddy page under lock */
511 if (!PageBuddy(page))
512 goto isolate_fail;
513 }
Mel Gorman748446b2010-05-24 14:32:27 -0700514
Joonsoo Kim66c64222016-07-26 15:23:40 -0700515 /* Found a free page, will break it into order-0 pages */
516 order = page_order(page);
517 isolated = __isolate_free_page(page, order);
David Rientjesa4f04f22016-06-24 14:50:10 -0700518 if (!isolated)
519 break;
Joonsoo Kim66c64222016-07-26 15:23:40 -0700520 set_page_private(page, order);
David Rientjesa4f04f22016-06-24 14:50:10 -0700521
Mel Gorman748446b2010-05-24 14:32:27 -0700522 total_isolated += isolated;
David Rientjesa4f04f22016-06-24 14:50:10 -0700523 cc->nr_freepages += isolated;
Joonsoo Kim66c64222016-07-26 15:23:40 -0700524 list_add_tail(&page->lru, freelist);
525
David Rientjesa4f04f22016-06-24 14:50:10 -0700526 if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
527 blockpfn += isolated;
528 break;
Mel Gorman748446b2010-05-24 14:32:27 -0700529 }
David Rientjesa4f04f22016-06-24 14:50:10 -0700530 /* Advance to the end of split page */
531 blockpfn += isolated - 1;
532 cursor += isolated - 1;
533 continue;
Laura Abbott2af120b2014-03-10 15:49:44 -0700534
535isolate_fail:
536 if (strict)
537 break;
538 else
539 continue;
540
Mel Gorman748446b2010-05-24 14:32:27 -0700541 }
542
David Rientjesa4f04f22016-06-24 14:50:10 -0700543 if (locked)
544 spin_unlock_irqrestore(&cc->zone->lock, flags);
545
Vlastimil Babka9fcd6d22015-09-08 15:02:49 -0700546 /*
547 * There is a tiny chance that we have read bogus compound_order(),
548 * so be careful to not go outside of the pageblock.
549 */
550 if (unlikely(blockpfn > end_pfn))
551 blockpfn = end_pfn;
552
Joonsoo Kime34d85f2015-02-11 15:27:04 -0800553 trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
554 nr_scanned, total_isolated);
555
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700556 /* Record how far we have got within the block */
557 *start_pfn = blockpfn;
558
Mel Gormanf40d1e42012-10-08 16:32:36 -0700559 /*
560 * If strict isolation is requested by CMA then check that all the
561 * pages requested were isolated. If there were any failures, 0 is
562 * returned and CMA will fail.
563 */
Laura Abbott2af120b2014-03-10 15:49:44 -0700564 if (strict && blockpfn < end_pfn)
Mel Gormanf40d1e42012-10-08 16:32:36 -0700565 total_isolated = 0;
566
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700567 /* Update the pageblock-skip if the whole pageblock was scanned */
568 if (blockpfn == end_pfn)
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700569 update_pageblock_skip(cc, valid_page, total_isolated, false);
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700570
David Rientjes7f354a52017-02-22 15:44:50 -0800571 cc->total_free_scanned += nr_scanned;
Mel Gorman397487d2012-10-19 12:00:10 +0100572 if (total_isolated)
Minchan Kim010fc292012-12-20 15:05:06 -0800573 count_compact_events(COMPACTISOLATED, total_isolated);
Mel Gorman748446b2010-05-24 14:32:27 -0700574 return total_isolated;
575}
576
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100577/**
578 * isolate_freepages_range() - isolate free pages.
579 * @start_pfn: The first PFN to start isolating.
580 * @end_pfn: The one-past-last PFN.
581 *
582 * Non-free pages, invalid PFNs, or zone boundaries within the
583 * [start_pfn, end_pfn) range are considered errors, cause function to
584 * undo its actions and return zero.
585 *
586 * Otherwise, function returns one-past-the-last PFN of isolated page
587 * (which may be greater then end_pfn if end fell in a middle of
588 * a free page).
589 */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100590unsigned long
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700591isolate_freepages_range(struct compact_control *cc,
592 unsigned long start_pfn, unsigned long end_pfn)
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100593{
Joonsoo Kime1409c32016-03-15 14:57:48 -0700594 unsigned long isolated, pfn, block_start_pfn, block_end_pfn;
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100595 LIST_HEAD(freelist);
596
Vlastimil Babka7d49d882014-10-09 15:27:11 -0700597 pfn = start_pfn;
Vlastimil Babka06b66402016-05-19 17:11:48 -0700598 block_start_pfn = pageblock_start_pfn(pfn);
Joonsoo Kime1409c32016-03-15 14:57:48 -0700599 if (block_start_pfn < cc->zone->zone_start_pfn)
600 block_start_pfn = cc->zone->zone_start_pfn;
Vlastimil Babka06b66402016-05-19 17:11:48 -0700601 block_end_pfn = pageblock_end_pfn(pfn);
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100602
Vlastimil Babka7d49d882014-10-09 15:27:11 -0700603 for (; pfn < end_pfn; pfn += isolated,
Joonsoo Kime1409c32016-03-15 14:57:48 -0700604 block_start_pfn = block_end_pfn,
Vlastimil Babka7d49d882014-10-09 15:27:11 -0700605 block_end_pfn += pageblock_nr_pages) {
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700606 /* Protect pfn from changing by isolate_freepages_block */
607 unsigned long isolate_start_pfn = pfn;
Vlastimil Babka7d49d882014-10-09 15:27:11 -0700608
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100609 block_end_pfn = min(block_end_pfn, end_pfn);
610
Joonsoo Kim58420012014-11-13 15:19:07 -0800611 /*
612 * pfn could pass the block_end_pfn if isolated freepage
613 * is more than pageblock order. In this case, we adjust
614 * scanning range to right one.
615 */
616 if (pfn >= block_end_pfn) {
Vlastimil Babka06b66402016-05-19 17:11:48 -0700617 block_start_pfn = pageblock_start_pfn(pfn);
618 block_end_pfn = pageblock_end_pfn(pfn);
Joonsoo Kim58420012014-11-13 15:19:07 -0800619 block_end_pfn = min(block_end_pfn, end_pfn);
620 }
621
Joonsoo Kime1409c32016-03-15 14:57:48 -0700622 if (!pageblock_pfn_to_page(block_start_pfn,
623 block_end_pfn, cc->zone))
Vlastimil Babka7d49d882014-10-09 15:27:11 -0700624 break;
625
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700626 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
627 block_end_pfn, &freelist, true);
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100628
629 /*
630 * In strict mode, isolate_freepages_block() returns 0 if
631 * there are any holes in the block (ie. invalid PFNs or
632 * non-free pages).
633 */
634 if (!isolated)
635 break;
636
637 /*
638 * If we managed to isolate pages, it is always (1 << n) *
639 * pageblock_nr_pages for some non-negative n. (Max order
640 * page may span two pageblocks).
641 */
642 }
643
Joonsoo Kim66c64222016-07-26 15:23:40 -0700644 /* __isolate_free_page() does not map the pages */
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100645 map_pages(&freelist);
646
647 if (pfn < end_pfn) {
648 /* Loop terminated early, cleanup. */
649 release_freepages(&freelist);
650 return 0;
651 }
652
653 /* We don't use freelists for anything. */
654 return pfn;
655}
656
Mel Gorman748446b2010-05-24 14:32:27 -0700657/* Similar to reclaim, but different enough that they don't share logic */
658static bool too_many_isolated(struct zone *zone)
659{
Minchan Kimbc693042010-09-09 16:38:00 -0700660 unsigned long active, inactive, isolated;
Mel Gorman748446b2010-05-24 14:32:27 -0700661
Mel Gorman599d0c92016-07-28 15:45:31 -0700662 inactive = node_page_state(zone->zone_pgdat, NR_INACTIVE_FILE) +
663 node_page_state(zone->zone_pgdat, NR_INACTIVE_ANON);
664 active = node_page_state(zone->zone_pgdat, NR_ACTIVE_FILE) +
665 node_page_state(zone->zone_pgdat, NR_ACTIVE_ANON);
666 isolated = node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE) +
667 node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON);
Mel Gorman748446b2010-05-24 14:32:27 -0700668
Minchan Kimbc693042010-09-09 16:38:00 -0700669 return isolated > (inactive + active) / 2;
Mel Gorman748446b2010-05-24 14:32:27 -0700670}
671
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100672/**
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700673 * isolate_migratepages_block() - isolate all migrate-able pages within
674 * a single pageblock
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100675 * @cc: Compaction control structure.
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700676 * @low_pfn: The first PFN to isolate
677 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
678 * @isolate_mode: Isolation mode to be used.
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100679 *
680 * Isolate all pages that can be migrated from the range specified by
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700681 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
682 * Returns zero if there is a fatal signal pending, otherwise PFN of the
683 * first page that was not scanned (which may be both less, equal to or more
684 * than end_pfn).
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100685 *
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700686 * The pages are isolated on cc->migratepages list (not required to be empty),
687 * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
688 * is neither read nor updated.
Mel Gorman748446b2010-05-24 14:32:27 -0700689 */
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700690static unsigned long
691isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
692 unsigned long end_pfn, isolate_mode_t isolate_mode)
Mel Gorman748446b2010-05-24 14:32:27 -0700693{
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700694 struct zone *zone = cc->zone;
Mel Gormanb7aba692011-01-13 15:45:54 -0800695 unsigned long nr_scanned = 0, nr_isolated = 0;
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700696 struct lruvec *lruvec;
Xiubo Lib8b2d822014-10-09 15:28:21 -0700697 unsigned long flags = 0;
Mel Gorman2a1402a2012-10-08 16:32:33 -0700698 bool locked = false;
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700699 struct page *page = NULL, *valid_page = NULL;
Joonsoo Kime34d85f2015-02-11 15:27:04 -0800700 unsigned long start_pfn = low_pfn;
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700701 bool skip_on_failure = false;
702 unsigned long next_skip_pfn = 0;
Mel Gorman748446b2010-05-24 14:32:27 -0700703
Mel Gorman748446b2010-05-24 14:32:27 -0700704 /*
705 * Ensure that there are not too many pages isolated from the LRU
706 * list by either parallel reclaimers or compaction. If there are,
707 * delay for some time until fewer pages are isolated
708 */
709 while (unlikely(too_many_isolated(zone))) {
Mel Gormanf9e35b32011-06-15 15:08:52 -0700710 /* async migration should just abort */
David Rientjese0b9dae2014-06-04 16:08:28 -0700711 if (cc->mode == MIGRATE_ASYNC)
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100712 return 0;
Mel Gormanf9e35b32011-06-15 15:08:52 -0700713
Mel Gorman748446b2010-05-24 14:32:27 -0700714 congestion_wait(BLK_RW_ASYNC, HZ/10);
715
716 if (fatal_signal_pending(current))
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100717 return 0;
Mel Gorman748446b2010-05-24 14:32:27 -0700718 }
719
Vlastimil Babkabe976572014-06-04 16:10:41 -0700720 if (compact_should_abort(cc))
721 return 0;
David Rientjesaeef4b82014-06-04 16:08:31 -0700722
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700723 if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
724 skip_on_failure = true;
725 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
726 }
727
Mel Gorman748446b2010-05-24 14:32:27 -0700728 /* Time to isolate some pages for migration */
Mel Gorman748446b2010-05-24 14:32:27 -0700729 for (; low_pfn < end_pfn; low_pfn++) {
Vlastimil Babka29c0dde2015-09-08 15:02:46 -0700730
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700731 if (skip_on_failure && low_pfn >= next_skip_pfn) {
732 /*
733 * We have isolated all migration candidates in the
734 * previous order-aligned block, and did not skip it due
735 * to failure. We should migrate the pages now and
736 * hopefully succeed compaction.
737 */
738 if (nr_isolated)
739 break;
740
741 /*
742 * We failed to isolate in the previous order-aligned
743 * block. Set the new boundary to the end of the
744 * current block. Note we can't simply increase
745 * next_skip_pfn by 1 << order, as low_pfn might have
746 * been incremented by a higher number due to skipping
747 * a compound or a high-order buddy page in the
748 * previous loop iteration.
749 */
750 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
751 }
752
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700753 /*
754 * Periodically drop the lock (if held) regardless of its
755 * contention, to give chance to IRQs. Abort async compaction
756 * if contended.
757 */
758 if (!(low_pfn % SWAP_CLUSTER_MAX)
Mel Gormana52633d2016-07-28 15:45:28 -0700759 && compact_unlock_should_abort(zone_lru_lock(zone), flags,
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700760 &locked, cc))
761 break;
Mel Gormanc67fe372012-08-21 16:16:17 -0700762
Mel Gorman748446b2010-05-24 14:32:27 -0700763 if (!pfn_valid_within(low_pfn))
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700764 goto isolate_fail;
Mel Gormanb7aba692011-01-13 15:45:54 -0800765 nr_scanned++;
Mel Gorman748446b2010-05-24 14:32:27 -0700766
Mel Gorman748446b2010-05-24 14:32:27 -0700767 page = pfn_to_page(low_pfn);
Mel Gormandc908602012-02-08 17:13:38 -0800768
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700769 if (!valid_page)
770 valid_page = page;
771
Mel Gorman6c144662014-01-23 15:53:38 -0800772 /*
Vlastimil Babka99c0fd52014-10-09 15:27:23 -0700773 * Skip if free. We read page order here without zone lock
774 * which is generally unsafe, but the race window is small and
775 * the worst thing that can happen is that we skip some
776 * potential isolation targets.
Mel Gorman6c144662014-01-23 15:53:38 -0800777 */
Vlastimil Babka99c0fd52014-10-09 15:27:23 -0700778 if (PageBuddy(page)) {
779 unsigned long freepage_order = page_order_unsafe(page);
780
781 /*
782 * Without lock, we cannot be sure that what we got is
783 * a valid page order. Consider only values in the
784 * valid order range to prevent low_pfn overflow.
785 */
786 if (freepage_order > 0 && freepage_order < MAX_ORDER)
787 low_pfn += (1UL << freepage_order) - 1;
Mel Gorman748446b2010-05-24 14:32:27 -0700788 continue;
Vlastimil Babka99c0fd52014-10-09 15:27:23 -0700789 }
Mel Gorman748446b2010-05-24 14:32:27 -0700790
Mel Gorman9927af742011-01-13 15:45:59 -0800791 /*
Vlastimil Babka29c0dde2015-09-08 15:02:46 -0700792 * Regardless of being on LRU, compound pages such as THP and
793 * hugetlbfs are not to be compacted. We can potentially save
794 * a lot of iterations if we skip them at once. The check is
795 * racy, but we can consider only valid values and the only
796 * danger is skipping too much.
Andrea Arcangelibc835012011-01-13 15:47:08 -0800797 */
Vlastimil Babka29c0dde2015-09-08 15:02:46 -0700798 if (PageCompound(page)) {
David Rientjes21dc7e02017-11-17 15:26:30 -0800799 const unsigned int order = compound_order(page);
Vlastimil Babka29c0dde2015-09-08 15:02:46 -0700800
Vlastimil Babkad3c85ba2017-11-17 15:26:41 -0800801 if (likely(order < MAX_ORDER))
David Rientjes21dc7e02017-11-17 15:26:30 -0800802 low_pfn += (1UL << order) - 1;
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700803 goto isolate_fail;
Mel Gorman2a1402a2012-10-08 16:32:33 -0700804 }
805
Minchan Kimbda807d2016-07-26 15:23:05 -0700806 /*
807 * Check may be lockless but that's ok as we recheck later.
808 * It's possible to migrate LRU and non-lru movable pages.
809 * Skip any other type of page
810 */
811 if (!PageLRU(page)) {
Minchan Kimbda807d2016-07-26 15:23:05 -0700812 /*
813 * __PageMovable can return false positive so we need
814 * to verify it under page_lock.
815 */
816 if (unlikely(__PageMovable(page)) &&
817 !PageIsolated(page)) {
818 if (locked) {
Mel Gormana52633d2016-07-28 15:45:28 -0700819 spin_unlock_irqrestore(zone_lru_lock(zone),
Minchan Kimbda807d2016-07-26 15:23:05 -0700820 flags);
821 locked = false;
822 }
823
Yisheng Xie9e5bcd62017-02-24 14:57:29 -0800824 if (!isolate_movable_page(page, isolate_mode))
Minchan Kimbda807d2016-07-26 15:23:05 -0700825 goto isolate_success;
826 }
827
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700828 goto isolate_fail;
Minchan Kimbda807d2016-07-26 15:23:05 -0700829 }
Vlastimil Babka29c0dde2015-09-08 15:02:46 -0700830
David Rientjes119d6d52014-04-03 14:48:00 -0700831 /*
832 * Migration will fail if an anonymous page is pinned in memory,
833 * so avoid taking lru_lock and isolating it unnecessarily in an
834 * admittedly racy check.
835 */
836 if (!page_mapping(page) &&
837 page_count(page) > page_mapcount(page))
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700838 goto isolate_fail;
David Rientjes119d6d52014-04-03 14:48:00 -0700839
Michal Hocko73e64c52016-12-14 15:04:07 -0800840 /*
841 * Only allow to migrate anonymous pages in GFP_NOFS context
842 * because those do not depend on fs locks.
843 */
844 if (!(cc->gfp_mask & __GFP_FS) && page_mapping(page))
845 goto isolate_fail;
846
Vlastimil Babka69b71892014-10-09 15:27:18 -0700847 /* If we already hold the lock, we can skip some rechecking */
848 if (!locked) {
Mel Gormana52633d2016-07-28 15:45:28 -0700849 locked = compact_trylock_irqsave(zone_lru_lock(zone),
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700850 &flags, cc);
Vlastimil Babka69b71892014-10-09 15:27:18 -0700851 if (!locked)
852 break;
Mel Gorman2a1402a2012-10-08 16:32:33 -0700853
Vlastimil Babka29c0dde2015-09-08 15:02:46 -0700854 /* Recheck PageLRU and PageCompound under lock */
Vlastimil Babka69b71892014-10-09 15:27:18 -0700855 if (!PageLRU(page))
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700856 goto isolate_fail;
Vlastimil Babka29c0dde2015-09-08 15:02:46 -0700857
858 /*
859 * Page become compound since the non-locked check,
860 * and it's on LRU. It can only be a THP so the order
861 * is safe to read and it's 0 for tail pages.
862 */
863 if (unlikely(PageCompound(page))) {
Vlastimil Babkad3c85ba2017-11-17 15:26:41 -0800864 low_pfn += (1UL << compound_order(page)) - 1;
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700865 goto isolate_fail;
Vlastimil Babka69b71892014-10-09 15:27:18 -0700866 }
Andrea Arcangelibc835012011-01-13 15:47:08 -0800867 }
868
Mel Gorman599d0c92016-07-28 15:45:31 -0700869 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700870
Mel Gorman748446b2010-05-24 14:32:27 -0700871 /* Try isolate the page */
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700872 if (__isolate_lru_page(page, isolate_mode) != 0)
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700873 goto isolate_fail;
Mel Gorman748446b2010-05-24 14:32:27 -0700874
Vlastimil Babka29c0dde2015-09-08 15:02:46 -0700875 VM_BUG_ON_PAGE(PageCompound(page), page);
Andrea Arcangelibc835012011-01-13 15:47:08 -0800876
Mel Gorman748446b2010-05-24 14:32:27 -0700877 /* Successfully isolated */
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700878 del_page_from_lru_list(page, lruvec, page_lru(page));
Ming Ling6afcf8e2016-12-12 16:42:26 -0800879 inc_node_page_state(page,
880 NR_ISOLATED_ANON + page_is_file_cache(page));
Joonsoo Kimb6c75012014-04-07 15:37:07 -0700881
882isolate_success:
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700883 list_add(&page->lru, &cc->migratepages);
Mel Gorman748446b2010-05-24 14:32:27 -0700884 cc->nr_migratepages++;
Mel Gormanb7aba692011-01-13 15:45:54 -0800885 nr_isolated++;
Mel Gorman748446b2010-05-24 14:32:27 -0700886
Vlastimil Babkaa34753d2016-05-19 17:11:51 -0700887 /*
888 * Record where we could have freed pages by migration and not
889 * yet flushed them to buddy allocator.
890 * - this is the lowest page that was isolated and likely be
891 * then freed by migration.
892 */
893 if (!cc->last_migrated_pfn)
894 cc->last_migrated_pfn = low_pfn;
895
Mel Gorman748446b2010-05-24 14:32:27 -0700896 /* Avoid isolating too much */
Hillf Danton31b83842012-01-10 15:07:59 -0800897 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
898 ++low_pfn;
Mel Gorman748446b2010-05-24 14:32:27 -0700899 break;
Hillf Danton31b83842012-01-10 15:07:59 -0800900 }
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700901
902 continue;
903isolate_fail:
904 if (!skip_on_failure)
905 continue;
906
907 /*
908 * We have isolated some pages, but then failed. Release them
909 * instead of migrating, as we cannot form the cc->order buddy
910 * page anyway.
911 */
912 if (nr_isolated) {
913 if (locked) {
Mel Gormana52633d2016-07-28 15:45:28 -0700914 spin_unlock_irqrestore(zone_lru_lock(zone), flags);
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700915 locked = false;
916 }
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700917 putback_movable_pages(&cc->migratepages);
918 cc->nr_migratepages = 0;
919 cc->last_migrated_pfn = 0;
920 nr_isolated = 0;
921 }
922
923 if (low_pfn < next_skip_pfn) {
924 low_pfn = next_skip_pfn - 1;
925 /*
926 * The check near the loop beginning would have updated
927 * next_skip_pfn too, but this is a bit simpler.
928 */
929 next_skip_pfn += 1UL << cc->order;
930 }
Mel Gorman748446b2010-05-24 14:32:27 -0700931 }
932
Vlastimil Babka99c0fd52014-10-09 15:27:23 -0700933 /*
934 * The PageBuddy() check could have potentially brought us outside
935 * the range to be scanned.
936 */
937 if (unlikely(low_pfn > end_pfn))
938 low_pfn = end_pfn;
939
Mel Gormanc67fe372012-08-21 16:16:17 -0700940 if (locked)
Mel Gormana52633d2016-07-28 15:45:28 -0700941 spin_unlock_irqrestore(zone_lru_lock(zone), flags);
Mel Gorman748446b2010-05-24 14:32:27 -0700942
Vlastimil Babka50b5b092014-01-21 15:51:10 -0800943 /*
944 * Update the pageblock-skip information and cached scanner pfn,
945 * if the whole pageblock was scanned without isolating any page.
Vlastimil Babka50b5b092014-01-21 15:51:10 -0800946 */
David Rientjes35979ef2014-06-04 16:08:27 -0700947 if (low_pfn == end_pfn)
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700948 update_pageblock_skip(cc, valid_page, nr_isolated, true);
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700949
Joonsoo Kime34d85f2015-02-11 15:27:04 -0800950 trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
951 nr_scanned, nr_isolated);
Mel Gormanb7aba692011-01-13 15:45:54 -0800952
David Rientjes7f354a52017-02-22 15:44:50 -0800953 cc->total_migrate_scanned += nr_scanned;
Mel Gorman397487d2012-10-19 12:00:10 +0100954 if (nr_isolated)
Minchan Kim010fc292012-12-20 15:05:06 -0800955 count_compact_events(COMPACTISOLATED, nr_isolated);
Mel Gorman397487d2012-10-19 12:00:10 +0100956
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100957 return low_pfn;
958}
959
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700960/**
961 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
962 * @cc: Compaction control structure.
963 * @start_pfn: The first PFN to start isolating.
964 * @end_pfn: The one-past-last PFN.
965 *
966 * Returns zero if isolation fails fatally due to e.g. pending signal.
967 * Otherwise, function returns one-past-the-last PFN of isolated page
968 * (which may be greater than end_pfn if end fell in a middle of a THP page).
969 */
970unsigned long
971isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
972 unsigned long end_pfn)
973{
Joonsoo Kime1409c32016-03-15 14:57:48 -0700974 unsigned long pfn, block_start_pfn, block_end_pfn;
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700975
976 /* Scan block by block. First and last block may be incomplete */
977 pfn = start_pfn;
Vlastimil Babka06b66402016-05-19 17:11:48 -0700978 block_start_pfn = pageblock_start_pfn(pfn);
Joonsoo Kime1409c32016-03-15 14:57:48 -0700979 if (block_start_pfn < cc->zone->zone_start_pfn)
980 block_start_pfn = cc->zone->zone_start_pfn;
Vlastimil Babka06b66402016-05-19 17:11:48 -0700981 block_end_pfn = pageblock_end_pfn(pfn);
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700982
983 for (; pfn < end_pfn; pfn = block_end_pfn,
Joonsoo Kime1409c32016-03-15 14:57:48 -0700984 block_start_pfn = block_end_pfn,
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700985 block_end_pfn += pageblock_nr_pages) {
986
987 block_end_pfn = min(block_end_pfn, end_pfn);
988
Joonsoo Kime1409c32016-03-15 14:57:48 -0700989 if (!pageblock_pfn_to_page(block_start_pfn,
990 block_end_pfn, cc->zone))
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700991 continue;
992
993 pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
994 ISOLATE_UNEVICTABLE);
995
Hugh Dickins14af4a52016-05-05 16:22:15 -0700996 if (!pfn)
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700997 break;
Joonsoo Kim6ea41c02014-10-29 14:50:20 -0700998
999 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
1000 break;
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001001 }
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001002
1003 return pfn;
1004}
1005
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001006#endif /* CONFIG_COMPACTION || CONFIG_CMA */
1007#ifdef CONFIG_COMPACTION
Andrew Morton018e9a42015-04-15 16:15:20 -07001008
Vlastimil Babkab682deb2017-05-08 15:54:43 -07001009static bool suitable_migration_source(struct compact_control *cc,
1010 struct page *page)
1011{
Vlastimil Babka282722b2017-05-08 15:54:49 -07001012 int block_mt;
1013
1014 if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction)
Vlastimil Babkab682deb2017-05-08 15:54:43 -07001015 return true;
1016
Vlastimil Babka282722b2017-05-08 15:54:49 -07001017 block_mt = get_pageblock_migratetype(page);
1018
1019 if (cc->migratetype == MIGRATE_MOVABLE)
1020 return is_migrate_movable(block_mt);
1021 else
1022 return block_mt == cc->migratetype;
Vlastimil Babkab682deb2017-05-08 15:54:43 -07001023}
1024
Andrew Morton018e9a42015-04-15 16:15:20 -07001025/* Returns true if the page is within a block suitable for migration to */
Vlastimil Babka9f7e3382016-10-07 17:00:37 -07001026static bool suitable_migration_target(struct compact_control *cc,
1027 struct page *page)
Andrew Morton018e9a42015-04-15 16:15:20 -07001028{
1029 /* If the page is a large free page, then disallow migration */
1030 if (PageBuddy(page)) {
1031 /*
1032 * We are checking page_order without zone->lock taken. But
1033 * the only small danger is that we skip a potentially suitable
1034 * pageblock, so it's not worth to check order for valid range.
1035 */
1036 if (page_order_unsafe(page) >= pageblock_order)
1037 return false;
1038 }
1039
Yisheng Xie1ef36db2017-05-03 14:53:54 -07001040 if (cc->ignore_block_suitable)
1041 return true;
1042
Andrew Morton018e9a42015-04-15 16:15:20 -07001043 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
Vlastimil Babkab682deb2017-05-08 15:54:43 -07001044 if (is_migrate_movable(get_pageblock_migratetype(page)))
Andrew Morton018e9a42015-04-15 16:15:20 -07001045 return true;
1046
1047 /* Otherwise skip the block */
1048 return false;
1049}
1050
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001051/*
Vlastimil Babkaf2849aa2015-09-08 15:02:36 -07001052 * Test whether the free scanner has reached the same or lower pageblock than
1053 * the migration scanner, and compaction should thus terminate.
1054 */
1055static inline bool compact_scanners_met(struct compact_control *cc)
1056{
1057 return (cc->free_pfn >> pageblock_order)
1058 <= (cc->migrate_pfn >> pageblock_order);
1059}
1060
1061/*
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001062 * Based on information in the current compact_control, find blocks
1063 * suitable for isolating free pages from and then isolate them.
1064 */
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001065static void isolate_freepages(struct compact_control *cc)
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001066{
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001067 struct zone *zone = cc->zone;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001068 struct page *page;
Vlastimil Babkac96b9e52014-06-04 16:07:26 -07001069 unsigned long block_start_pfn; /* start of current pageblock */
Vlastimil Babkae14c7202014-10-09 15:27:20 -07001070 unsigned long isolate_start_pfn; /* exact pfn we start at */
Vlastimil Babkac96b9e52014-06-04 16:07:26 -07001071 unsigned long block_end_pfn; /* end of current pageblock */
1072 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001073 struct list_head *freelist = &cc->freepages;
1074
1075 /*
1076 * Initialise the free scanner. The starting point is where we last
Vlastimil Babka49e068f2014-05-06 12:50:03 -07001077 * successfully isolated from, zone-cached value, or the end of the
Vlastimil Babkae14c7202014-10-09 15:27:20 -07001078 * zone when isolating for the first time. For looping we also need
1079 * this pfn aligned down to the pageblock boundary, because we do
Vlastimil Babkac96b9e52014-06-04 16:07:26 -07001080 * block_start_pfn -= pageblock_nr_pages in the for loop.
1081 * For ending point, take care when isolating in last pageblock of a
1082 * a zone which ends in the middle of a pageblock.
Vlastimil Babka49e068f2014-05-06 12:50:03 -07001083 * The low boundary is the end of the pageblock the migration scanner
1084 * is using.
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001085 */
Vlastimil Babkae14c7202014-10-09 15:27:20 -07001086 isolate_start_pfn = cc->free_pfn;
Vlastimil Babka06b66402016-05-19 17:11:48 -07001087 block_start_pfn = pageblock_start_pfn(cc->free_pfn);
Vlastimil Babkac96b9e52014-06-04 16:07:26 -07001088 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1089 zone_end_pfn(zone));
Vlastimil Babka06b66402016-05-19 17:11:48 -07001090 low_pfn = pageblock_end_pfn(cc->migrate_pfn);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001091
1092 /*
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001093 * Isolate free pages until enough are available to migrate the
1094 * pages on cc->migratepages. We stop searching if the migrate
1095 * and free page scanners meet or enough free pages are isolated.
1096 */
Vlastimil Babkaf5f61a32015-09-08 15:02:39 -07001097 for (; block_start_pfn >= low_pfn;
Vlastimil Babkac96b9e52014-06-04 16:07:26 -07001098 block_end_pfn = block_start_pfn,
Vlastimil Babkae14c7202014-10-09 15:27:20 -07001099 block_start_pfn -= pageblock_nr_pages,
1100 isolate_start_pfn = block_start_pfn) {
David Rientjesf6ea3ad2013-09-30 13:45:03 -07001101 /*
1102 * This can iterate a massively long zone without finding any
1103 * suitable migration targets, so periodically check if we need
Vlastimil Babkabe976572014-06-04 16:10:41 -07001104 * to schedule, or even abort async compaction.
David Rientjesf6ea3ad2013-09-30 13:45:03 -07001105 */
Vlastimil Babkabe976572014-06-04 16:10:41 -07001106 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1107 && compact_should_abort(cc))
1108 break;
David Rientjesf6ea3ad2013-09-30 13:45:03 -07001109
Vlastimil Babka7d49d882014-10-09 15:27:11 -07001110 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1111 zone);
1112 if (!page)
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001113 continue;
1114
1115 /* Check the block is suitable for migration */
Vlastimil Babka9f7e3382016-10-07 17:00:37 -07001116 if (!suitable_migration_target(cc, page))
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001117 continue;
Linus Torvalds68e3e922012-06-03 20:05:57 -07001118
Mel Gormanbb13ffe2012-10-08 16:32:41 -07001119 /* If isolation recently failed, do not retry */
1120 if (!isolation_suitable(cc, page))
1121 continue;
1122
Vlastimil Babkae14c7202014-10-09 15:27:20 -07001123 /* Found a block suitable for isolating free pages from. */
David Rientjesa46cbf32016-07-14 12:06:50 -07001124 isolate_freepages_block(cc, &isolate_start_pfn, block_end_pfn,
1125 freelist, false);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001126
1127 /*
David Rientjesa46cbf32016-07-14 12:06:50 -07001128 * If we isolated enough freepages, or aborted due to lock
1129 * contention, terminate.
Vlastimil Babkae14c7202014-10-09 15:27:20 -07001130 */
Vlastimil Babkaf5f61a32015-09-08 15:02:39 -07001131 if ((cc->nr_freepages >= cc->nr_migratepages)
1132 || cc->contended) {
David Rientjesa46cbf32016-07-14 12:06:50 -07001133 if (isolate_start_pfn >= block_end_pfn) {
1134 /*
1135 * Restart at previous pageblock if more
1136 * freepages can be isolated next time.
1137 */
Vlastimil Babkaf5f61a32015-09-08 15:02:39 -07001138 isolate_start_pfn =
1139 block_start_pfn - pageblock_nr_pages;
David Rientjesa46cbf32016-07-14 12:06:50 -07001140 }
Vlastimil Babkabe976572014-06-04 16:10:41 -07001141 break;
David Rientjesa46cbf32016-07-14 12:06:50 -07001142 } else if (isolate_start_pfn < block_end_pfn) {
Vlastimil Babkaf5f61a32015-09-08 15:02:39 -07001143 /*
David Rientjesa46cbf32016-07-14 12:06:50 -07001144 * If isolation failed early, do not continue
1145 * needlessly.
Vlastimil Babkaf5f61a32015-09-08 15:02:39 -07001146 */
David Rientjesa46cbf32016-07-14 12:06:50 -07001147 break;
Vlastimil Babkaf5f61a32015-09-08 15:02:39 -07001148 }
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +01001149 }
1150
Joonsoo Kim66c64222016-07-26 15:23:40 -07001151 /* __isolate_free_page() does not map the pages */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001152 map_pages(freelist);
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +01001153
Vlastimil Babka7ed695e2014-01-21 15:51:09 -08001154 /*
Vlastimil Babkaf5f61a32015-09-08 15:02:39 -07001155 * Record where the free scanner will restart next time. Either we
1156 * broke from the loop and set isolate_start_pfn based on the last
1157 * call to isolate_freepages_block(), or we met the migration scanner
1158 * and the loop terminated due to isolate_start_pfn < low_pfn
Vlastimil Babka7ed695e2014-01-21 15:51:09 -08001159 */
Vlastimil Babkaf5f61a32015-09-08 15:02:39 -07001160 cc->free_pfn = isolate_start_pfn;
Mel Gorman748446b2010-05-24 14:32:27 -07001161}
1162
1163/*
1164 * This is a migrate-callback that "allocates" freepages by taking pages
1165 * from the isolated freelists in the block we are migrating to.
1166 */
1167static struct page *compaction_alloc(struct page *migratepage,
1168 unsigned long data,
1169 int **result)
1170{
1171 struct compact_control *cc = (struct compact_control *)data;
1172 struct page *freepage;
1173
Vlastimil Babkabe976572014-06-04 16:10:41 -07001174 /*
1175 * Isolate free pages if necessary, and if we are not aborting due to
1176 * contention.
1177 */
Mel Gorman748446b2010-05-24 14:32:27 -07001178 if (list_empty(&cc->freepages)) {
Vlastimil Babkabe976572014-06-04 16:10:41 -07001179 if (!cc->contended)
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001180 isolate_freepages(cc);
Mel Gorman748446b2010-05-24 14:32:27 -07001181
1182 if (list_empty(&cc->freepages))
1183 return NULL;
1184 }
1185
1186 freepage = list_entry(cc->freepages.next, struct page, lru);
1187 list_del(&freepage->lru);
1188 cc->nr_freepages--;
1189
1190 return freepage;
1191}
1192
1193/*
David Rientjesd53aea32014-06-04 16:08:26 -07001194 * This is a migrate-callback that "frees" freepages back to the isolated
1195 * freelist. All pages on the freelist are from the same zone, so there is no
1196 * special handling needed for NUMA.
1197 */
1198static void compaction_free(struct page *page, unsigned long data)
1199{
1200 struct compact_control *cc = (struct compact_control *)data;
1201
1202 list_add(&page->lru, &cc->freepages);
1203 cc->nr_freepages++;
1204}
1205
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001206/* possible outcome of isolate_migratepages */
1207typedef enum {
1208 ISOLATE_ABORT, /* Abort compaction now */
1209 ISOLATE_NONE, /* No pages isolated, continue scanning */
1210 ISOLATE_SUCCESS, /* Pages isolated, migrate */
1211} isolate_migrate_t;
1212
1213/*
Eric B Munson5bbe3542015-04-15 16:13:20 -07001214 * Allow userspace to control policy on scanning the unevictable LRU for
1215 * compactable pages.
1216 */
1217int sysctl_compact_unevictable_allowed __read_mostly = 1;
1218
1219/*
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001220 * Isolate all pages that can be migrated from the first suitable block,
1221 * starting at the block pointed to by the migrate scanner pfn within
1222 * compact_control.
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001223 */
1224static isolate_migrate_t isolate_migratepages(struct zone *zone,
1225 struct compact_control *cc)
1226{
Joonsoo Kime1409c32016-03-15 14:57:48 -07001227 unsigned long block_start_pfn;
1228 unsigned long block_end_pfn;
1229 unsigned long low_pfn;
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001230 struct page *page;
1231 const isolate_mode_t isolate_mode =
Eric B Munson5bbe3542015-04-15 16:13:20 -07001232 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
Hugh Dickins1d2047f2016-07-28 15:48:41 -07001233 (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001234
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001235 /*
1236 * Start at where we last stopped, or beginning of the zone as
1237 * initialized by compact_zone()
1238 */
1239 low_pfn = cc->migrate_pfn;
Vlastimil Babka06b66402016-05-19 17:11:48 -07001240 block_start_pfn = pageblock_start_pfn(low_pfn);
Joonsoo Kime1409c32016-03-15 14:57:48 -07001241 if (block_start_pfn < zone->zone_start_pfn)
1242 block_start_pfn = zone->zone_start_pfn;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001243
1244 /* Only scan within a pageblock boundary */
Vlastimil Babka06b66402016-05-19 17:11:48 -07001245 block_end_pfn = pageblock_end_pfn(low_pfn);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001246
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001247 /*
1248 * Iterate over whole pageblocks until we find the first suitable.
1249 * Do not cross the free scanner.
1250 */
Joonsoo Kime1409c32016-03-15 14:57:48 -07001251 for (; block_end_pfn <= cc->free_pfn;
1252 low_pfn = block_end_pfn,
1253 block_start_pfn = block_end_pfn,
1254 block_end_pfn += pageblock_nr_pages) {
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001255
1256 /*
1257 * This can potentially iterate a massively long zone with
1258 * many pageblocks unsuitable, so periodically check if we
1259 * need to schedule, or even abort async compaction.
1260 */
1261 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1262 && compact_should_abort(cc))
1263 break;
1264
Joonsoo Kime1409c32016-03-15 14:57:48 -07001265 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1266 zone);
Vlastimil Babka7d49d882014-10-09 15:27:11 -07001267 if (!page)
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001268 continue;
1269
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001270 /* If isolation recently failed, do not retry */
1271 if (!isolation_suitable(cc, page))
1272 continue;
1273
1274 /*
1275 * For async compaction, also only scan in MOVABLE blocks.
1276 * Async compaction is optimistic to see if the minimum amount
1277 * of work satisfies the allocation.
1278 */
Vlastimil Babkab682deb2017-05-08 15:54:43 -07001279 if (!suitable_migration_source(cc, page))
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001280 continue;
1281
1282 /* Perform the isolation */
Joonsoo Kime1409c32016-03-15 14:57:48 -07001283 low_pfn = isolate_migratepages_block(cc, low_pfn,
1284 block_end_pfn, isolate_mode);
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001285
Ming Ling6afcf8e2016-12-12 16:42:26 -08001286 if (!low_pfn || cc->contended)
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001287 return ISOLATE_ABORT;
1288
1289 /*
1290 * Either we isolated something and proceed with migration. Or
1291 * we failed and compact_zone should decide if we should
1292 * continue or not.
1293 */
1294 break;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001295 }
1296
Vlastimil Babkaf2849aa2015-09-08 15:02:36 -07001297 /* Record where migration scanner will be restarted. */
1298 cc->migrate_pfn = low_pfn;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001299
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001300 return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001301}
1302
Yaowei Bai21c527a2015-11-05 18:47:20 -08001303/*
1304 * order == -1 is expected when compacting via
1305 * /proc/sys/vm/compact_memory
1306 */
1307static inline bool is_via_compact_memory(int order)
1308{
1309 return order == -1;
1310}
1311
Vlastimil Babkad39773a2017-05-08 15:54:46 -07001312static enum compact_result __compact_finished(struct zone *zone,
1313 struct compact_control *cc)
Mel Gorman748446b2010-05-24 14:32:27 -07001314{
Mel Gorman8fb74b92013-01-11 14:32:16 -08001315 unsigned int order;
Vlastimil Babkad39773a2017-05-08 15:54:46 -07001316 const int migratetype = cc->migratetype;
Mel Gorman56de7262010-05-24 14:32:30 -07001317
Vlastimil Babkabe976572014-06-04 16:10:41 -07001318 if (cc->contended || fatal_signal_pending(current))
Vlastimil Babka2d1e1042015-11-05 18:48:02 -08001319 return COMPACT_CONTENDED;
Mel Gorman748446b2010-05-24 14:32:27 -07001320
Mel Gorman753341a2012-10-08 16:32:40 -07001321 /* Compaction run completes if the migrate and free scanner meet */
Vlastimil Babkaf2849aa2015-09-08 15:02:36 -07001322 if (compact_scanners_met(cc)) {
Vlastimil Babka55b7c4c2014-01-21 15:51:11 -08001323 /* Let the next compaction start anew. */
Vlastimil Babka023336412015-09-08 15:02:42 -07001324 reset_cached_positions(zone);
Vlastimil Babka55b7c4c2014-01-21 15:51:11 -08001325
Mel Gorman62997022012-10-08 16:32:47 -07001326 /*
1327 * Mark that the PG_migrate_skip information should be cleared
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07001328 * by kswapd when it goes to sleep. kcompactd does not set the
Mel Gorman62997022012-10-08 16:32:47 -07001329 * flag itself as the decision to be clear should be directly
1330 * based on an allocation request.
1331 */
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07001332 if (cc->direct_compaction)
Mel Gorman62997022012-10-08 16:32:47 -07001333 zone->compact_blockskip_flush = true;
1334
Michal Hockoc8f7de02016-05-20 16:56:47 -07001335 if (cc->whole_zone)
1336 return COMPACT_COMPLETE;
1337 else
1338 return COMPACT_PARTIAL_SKIPPED;
Mel Gormanbb13ffe2012-10-08 16:32:41 -07001339 }
Mel Gorman748446b2010-05-24 14:32:27 -07001340
Yaowei Bai21c527a2015-11-05 18:47:20 -08001341 if (is_via_compact_memory(cc->order))
Mel Gorman56de7262010-05-24 14:32:30 -07001342 return COMPACT_CONTINUE;
1343
Vlastimil Babkabaf6a9a2017-05-08 15:54:52 -07001344 if (cc->finishing_block) {
1345 /*
1346 * We have finished the pageblock, but better check again that
1347 * we really succeeded.
1348 */
1349 if (IS_ALIGNED(cc->migrate_pfn, pageblock_nr_pages))
1350 cc->finishing_block = false;
1351 else
1352 return COMPACT_CONTINUE;
1353 }
1354
Mel Gorman56de7262010-05-24 14:32:30 -07001355 /* Direct compactor: Is a suitable page free? */
Mel Gorman8fb74b92013-01-11 14:32:16 -08001356 for (order = cc->order; order < MAX_ORDER; order++) {
1357 struct free_area *area = &zone->free_area[order];
Joonsoo Kim2149cda2015-04-14 15:45:21 -07001358 bool can_steal;
Mel Gorman56de7262010-05-24 14:32:30 -07001359
Mel Gorman8fb74b92013-01-11 14:32:16 -08001360 /* Job done if page is free of the right migratetype */
David Rientjes6d7ce552014-10-09 15:27:27 -07001361 if (!list_empty(&area->free_list[migratetype]))
Vlastimil Babkacf378312016-10-07 16:57:41 -07001362 return COMPACT_SUCCESS;
Mel Gorman8fb74b92013-01-11 14:32:16 -08001363
Joonsoo Kim2149cda2015-04-14 15:45:21 -07001364#ifdef CONFIG_CMA
1365 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
1366 if (migratetype == MIGRATE_MOVABLE &&
1367 !list_empty(&area->free_list[MIGRATE_CMA]))
Vlastimil Babkacf378312016-10-07 16:57:41 -07001368 return COMPACT_SUCCESS;
Joonsoo Kim2149cda2015-04-14 15:45:21 -07001369#endif
1370 /*
1371 * Job done if allocation would steal freepages from
1372 * other migratetype buddy lists.
1373 */
1374 if (find_suitable_fallback(area, order, migratetype,
Vlastimil Babkabaf6a9a2017-05-08 15:54:52 -07001375 true, &can_steal) != -1) {
1376
1377 /* movable pages are OK in any pageblock */
1378 if (migratetype == MIGRATE_MOVABLE)
1379 return COMPACT_SUCCESS;
1380
1381 /*
1382 * We are stealing for a non-movable allocation. Make
1383 * sure we finish compacting the current pageblock
1384 * first so it is as free as possible and we won't
1385 * have to steal another one soon. This only applies
1386 * to sync compaction, as async compaction operates
1387 * on pageblocks of the same migratetype.
1388 */
1389 if (cc->mode == MIGRATE_ASYNC ||
1390 IS_ALIGNED(cc->migrate_pfn,
1391 pageblock_nr_pages)) {
1392 return COMPACT_SUCCESS;
1393 }
1394
1395 cc->finishing_block = true;
1396 return COMPACT_CONTINUE;
1397 }
Mel Gorman56de7262010-05-24 14:32:30 -07001398 }
1399
Joonsoo Kim837d0262015-02-11 15:27:06 -08001400 return COMPACT_NO_SUITABLE_PAGE;
1401}
1402
Michal Hockoea7ab982016-05-20 16:56:38 -07001403static enum compact_result compact_finished(struct zone *zone,
Vlastimil Babkad39773a2017-05-08 15:54:46 -07001404 struct compact_control *cc)
Joonsoo Kim837d0262015-02-11 15:27:06 -08001405{
1406 int ret;
1407
Vlastimil Babkad39773a2017-05-08 15:54:46 -07001408 ret = __compact_finished(zone, cc);
Joonsoo Kim837d0262015-02-11 15:27:06 -08001409 trace_mm_compaction_finished(zone, cc->order, ret);
1410 if (ret == COMPACT_NO_SUITABLE_PAGE)
1411 ret = COMPACT_CONTINUE;
1412
1413 return ret;
Mel Gorman748446b2010-05-24 14:32:27 -07001414}
1415
Mel Gorman3e7d3442011-01-13 15:45:56 -08001416/*
1417 * compaction_suitable: Is this suitable to run compaction on this zone now?
1418 * Returns
1419 * COMPACT_SKIPPED - If there are too few free pages for compaction
Vlastimil Babkacf378312016-10-07 16:57:41 -07001420 * COMPACT_SUCCESS - If the allocation would succeed without compaction
Mel Gorman3e7d3442011-01-13 15:45:56 -08001421 * COMPACT_CONTINUE - If compaction should run now
1422 */
Michal Hockoea7ab982016-05-20 16:56:38 -07001423static enum compact_result __compaction_suitable(struct zone *zone, int order,
Mel Gormanc6038442016-05-19 17:13:38 -07001424 unsigned int alloc_flags,
Michal Hocko86a294a2016-05-20 16:57:12 -07001425 int classzone_idx,
1426 unsigned long wmark_target)
Mel Gorman3e7d3442011-01-13 15:45:56 -08001427{
Mel Gorman3e7d3442011-01-13 15:45:56 -08001428 unsigned long watermark;
1429
Yaowei Bai21c527a2015-11-05 18:47:20 -08001430 if (is_via_compact_memory(order))
Michal Hocko3957c772011-06-15 15:08:25 -07001431 return COMPACT_CONTINUE;
1432
Vlastimil Babkaf2b82282016-10-07 16:57:50 -07001433 watermark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
Vlastimil Babkaebff3982014-12-10 15:43:22 -08001434 /*
1435 * If watermarks for high-order allocation are already met, there
1436 * should be no need for compaction at all.
1437 */
1438 if (zone_watermark_ok(zone, order, watermark, classzone_idx,
1439 alloc_flags))
Vlastimil Babkacf378312016-10-07 16:57:41 -07001440 return COMPACT_SUCCESS;
Vlastimil Babkaebff3982014-12-10 15:43:22 -08001441
Michal Hocko3957c772011-06-15 15:08:25 -07001442 /*
Vlastimil Babka9861a622016-10-07 16:57:53 -07001443 * Watermarks for order-0 must be met for compaction to be able to
Vlastimil Babka984fdba2016-10-07 16:57:57 -07001444 * isolate free pages for migration targets. This means that the
1445 * watermark and alloc_flags have to match, or be more pessimistic than
1446 * the check in __isolate_free_page(). We don't use the direct
1447 * compactor's alloc_flags, as they are not relevant for freepage
1448 * isolation. We however do use the direct compactor's classzone_idx to
1449 * skip over zones where lowmem reserves would prevent allocation even
1450 * if compaction succeeds.
Vlastimil Babka8348faf2016-10-07 16:58:00 -07001451 * For costly orders, we require low watermark instead of min for
1452 * compaction to proceed to increase its chances.
Vlastimil Babka984fdba2016-10-07 16:57:57 -07001453 * ALLOC_CMA is used, as pages in CMA pageblocks are considered
1454 * suitable migration targets
Mel Gorman3e7d3442011-01-13 15:45:56 -08001455 */
Vlastimil Babka8348faf2016-10-07 16:58:00 -07001456 watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
1457 low_wmark_pages(zone) : min_wmark_pages(zone);
1458 watermark += compact_gap(order);
Michal Hocko86a294a2016-05-20 16:57:12 -07001459 if (!__zone_watermark_ok(zone, 0, watermark, classzone_idx,
Vlastimil Babka984fdba2016-10-07 16:57:57 -07001460 ALLOC_CMA, wmark_target))
Mel Gorman3e7d3442011-01-13 15:45:56 -08001461 return COMPACT_SKIPPED;
1462
Vlastimil Babkacc5c9f02016-10-07 17:00:43 -07001463 return COMPACT_CONTINUE;
1464}
1465
1466enum compact_result compaction_suitable(struct zone *zone, int order,
1467 unsigned int alloc_flags,
1468 int classzone_idx)
1469{
1470 enum compact_result ret;
1471 int fragindex;
1472
1473 ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx,
1474 zone_page_state(zone, NR_FREE_PAGES));
Mel Gorman3e7d3442011-01-13 15:45:56 -08001475 /*
1476 * fragmentation index determines if allocation failures are due to
1477 * low memory or external fragmentation
1478 *
Vlastimil Babkaebff3982014-12-10 15:43:22 -08001479 * index of -1000 would imply allocations might succeed depending on
1480 * watermarks, but we already failed the high-order watermark check
Mel Gorman3e7d3442011-01-13 15:45:56 -08001481 * index towards 0 implies failure is due to lack of memory
1482 * index towards 1000 implies failure is due to fragmentation
1483 *
Vlastimil Babka20311422016-10-07 17:00:46 -07001484 * Only compact if a failure would be due to fragmentation. Also
1485 * ignore fragindex for non-costly orders where the alternative to
1486 * a successful reclaim/compaction is OOM. Fragindex and the
1487 * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
1488 * excessive compaction for costly orders, but it should not be at the
1489 * expense of system stability.
Mel Gorman3e7d3442011-01-13 15:45:56 -08001490 */
Vlastimil Babka20311422016-10-07 17:00:46 -07001491 if (ret == COMPACT_CONTINUE && (order > PAGE_ALLOC_COSTLY_ORDER)) {
Vlastimil Babkacc5c9f02016-10-07 17:00:43 -07001492 fragindex = fragmentation_index(zone, order);
1493 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
1494 ret = COMPACT_NOT_SUITABLE_ZONE;
1495 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08001496
Joonsoo Kim837d0262015-02-11 15:27:06 -08001497 trace_mm_compaction_suitable(zone, order, ret);
1498 if (ret == COMPACT_NOT_SUITABLE_ZONE)
1499 ret = COMPACT_SKIPPED;
1500
1501 return ret;
1502}
1503
Michal Hocko86a294a2016-05-20 16:57:12 -07001504bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
1505 int alloc_flags)
1506{
1507 struct zone *zone;
1508 struct zoneref *z;
1509
1510 /*
1511 * Make sure at least one zone would pass __compaction_suitable if we continue
1512 * retrying the reclaim.
1513 */
1514 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1515 ac->nodemask) {
1516 unsigned long available;
1517 enum compact_result compact_result;
1518
1519 /*
1520 * Do not consider all the reclaimable memory because we do not
1521 * want to trash just for a single high order allocation which
1522 * is even not guaranteed to appear even if __compaction_suitable
1523 * is happy about the watermark check.
1524 */
Mel Gorman5a1c84b2016-07-28 15:47:31 -07001525 available = zone_reclaimable_pages(zone) / order;
Michal Hocko86a294a2016-05-20 16:57:12 -07001526 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
1527 compact_result = __compaction_suitable(zone, order, alloc_flags,
1528 ac_classzone_idx(ac), available);
Vlastimil Babkacc5c9f02016-10-07 17:00:43 -07001529 if (compact_result != COMPACT_SKIPPED)
Michal Hocko86a294a2016-05-20 16:57:12 -07001530 return true;
1531 }
1532
1533 return false;
1534}
1535
Michal Hockoea7ab982016-05-20 16:56:38 -07001536static enum compact_result compact_zone(struct zone *zone, struct compact_control *cc)
Mel Gorman748446b2010-05-24 14:32:27 -07001537{
Michal Hockoea7ab982016-05-20 16:56:38 -07001538 enum compact_result ret;
Mel Gormanc89511a2012-10-08 16:32:45 -07001539 unsigned long start_pfn = zone->zone_start_pfn;
Cody P Schafer108bcc92013-02-22 16:35:23 -08001540 unsigned long end_pfn = zone_end_pfn(zone);
David Rientjese0b9dae2014-06-04 16:08:28 -07001541 const bool sync = cc->mode != MIGRATE_ASYNC;
Mel Gorman748446b2010-05-24 14:32:27 -07001542
Vlastimil Babkad39773a2017-05-08 15:54:46 -07001543 cc->migratetype = gfpflags_to_migratetype(cc->gfp_mask);
Vlastimil Babkaebff3982014-12-10 15:43:22 -08001544 ret = compaction_suitable(zone, cc->order, cc->alloc_flags,
1545 cc->classzone_idx);
Michal Hockoc46649d2016-05-20 16:56:41 -07001546 /* Compaction is likely to fail */
Vlastimil Babkacf378312016-10-07 16:57:41 -07001547 if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED)
Mel Gorman3e7d3442011-01-13 15:45:56 -08001548 return ret;
Michal Hockoc46649d2016-05-20 16:56:41 -07001549
1550 /* huh, compaction_suitable is returning something unexpected */
1551 VM_BUG_ON(ret != COMPACT_CONTINUE);
Mel Gorman3e7d3442011-01-13 15:45:56 -08001552
Mel Gormanc89511a2012-10-08 16:32:45 -07001553 /*
Vlastimil Babkad3132e42014-01-21 15:51:08 -08001554 * Clear pageblock skip if there were failures recently and compaction
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07001555 * is about to be retried after being deferred.
Vlastimil Babkad3132e42014-01-21 15:51:08 -08001556 */
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07001557 if (compaction_restarting(zone, cc->order))
Vlastimil Babkad3132e42014-01-21 15:51:08 -08001558 __reset_isolation_suitable(zone);
1559
1560 /*
Mel Gormanc89511a2012-10-08 16:32:45 -07001561 * Setup to move all movable pages to the end of the zone. Used cached
Vlastimil Babka06ed2992016-10-07 16:57:35 -07001562 * information on where the scanners should start (unless we explicitly
1563 * want to compact the whole zone), but check that it is initialised
1564 * by ensuring the values are within zone boundaries.
Mel Gormanc89511a2012-10-08 16:32:45 -07001565 */
Vlastimil Babka06ed2992016-10-07 16:57:35 -07001566 if (cc->whole_zone) {
Mel Gormanc89511a2012-10-08 16:32:45 -07001567 cc->migrate_pfn = start_pfn;
Vlastimil Babka06ed2992016-10-07 16:57:35 -07001568 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
1569 } else {
1570 cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync];
1571 cc->free_pfn = zone->compact_cached_free_pfn;
1572 if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
1573 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
1574 zone->compact_cached_free_pfn = cc->free_pfn;
1575 }
1576 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
1577 cc->migrate_pfn = start_pfn;
1578 zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
1579 zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
1580 }
Michal Hockoc8f7de02016-05-20 16:56:47 -07001581
Vlastimil Babka06ed2992016-10-07 16:57:35 -07001582 if (cc->migrate_pfn == start_pfn)
1583 cc->whole_zone = true;
1584 }
Michal Hockoc8f7de02016-05-20 16:56:47 -07001585
Joonsoo Kim1a167182015-09-08 15:03:59 -07001586 cc->last_migrated_pfn = 0;
Mel Gorman748446b2010-05-24 14:32:27 -07001587
Joonsoo Kim16c4a092015-02-11 15:27:01 -08001588 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
1589 cc->free_pfn, end_pfn, sync);
Mel Gorman0eb927c2014-01-21 15:51:05 -08001590
Mel Gorman748446b2010-05-24 14:32:27 -07001591 migrate_prep_local();
1592
Vlastimil Babkad39773a2017-05-08 15:54:46 -07001593 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
Minchan Kim9d502c12011-03-22 16:30:39 -07001594 int err;
Mel Gorman748446b2010-05-24 14:32:27 -07001595
Mel Gormanf9e35b32011-06-15 15:08:52 -07001596 switch (isolate_migratepages(zone, cc)) {
1597 case ISOLATE_ABORT:
Vlastimil Babka2d1e1042015-11-05 18:48:02 -08001598 ret = COMPACT_CONTENDED;
Rafael Aquini5733c7d2012-12-11 16:02:47 -08001599 putback_movable_pages(&cc->migratepages);
Shaohua Lie64c5232012-10-08 16:32:27 -07001600 cc->nr_migratepages = 0;
Mel Gormanf9e35b32011-06-15 15:08:52 -07001601 goto out;
1602 case ISOLATE_NONE:
Vlastimil Babkafdaf7f52014-12-10 15:43:34 -08001603 /*
1604 * We haven't isolated and migrated anything, but
1605 * there might still be unflushed migrations from
1606 * previous cc->order aligned block.
1607 */
1608 goto check_drain;
Mel Gormanf9e35b32011-06-15 15:08:52 -07001609 case ISOLATE_SUCCESS:
1610 ;
1611 }
Mel Gorman748446b2010-05-24 14:32:27 -07001612
David Rientjesd53aea32014-06-04 16:08:26 -07001613 err = migrate_pages(&cc->migratepages, compaction_alloc,
David Rientjese0b9dae2014-06-04 16:08:28 -07001614 compaction_free, (unsigned long)cc, cc->mode,
Mel Gorman7b2a2d42012-10-19 14:07:31 +01001615 MR_COMPACTION);
Mel Gorman748446b2010-05-24 14:32:27 -07001616
Vlastimil Babkaf8c93012014-06-04 16:08:32 -07001617 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
1618 &cc->migratepages);
Mel Gorman748446b2010-05-24 14:32:27 -07001619
Vlastimil Babkaf8c93012014-06-04 16:08:32 -07001620 /* All pages were either migrated or will be released */
1621 cc->nr_migratepages = 0;
Minchan Kim9d502c12011-03-22 16:30:39 -07001622 if (err) {
Rafael Aquini5733c7d2012-12-11 16:02:47 -08001623 putback_movable_pages(&cc->migratepages);
Vlastimil Babka7ed695e2014-01-21 15:51:09 -08001624 /*
1625 * migrate_pages() may return -ENOMEM when scanners meet
1626 * and we want compact_finished() to detect it
1627 */
Vlastimil Babkaf2849aa2015-09-08 15:02:36 -07001628 if (err == -ENOMEM && !compact_scanners_met(cc)) {
Vlastimil Babka2d1e1042015-11-05 18:48:02 -08001629 ret = COMPACT_CONTENDED;
David Rientjes4bf2bba2012-07-11 14:02:13 -07001630 goto out;
1631 }
Vlastimil Babkafdd048e2016-05-19 17:11:55 -07001632 /*
1633 * We failed to migrate at least one page in the current
1634 * order-aligned block, so skip the rest of it.
1635 */
1636 if (cc->direct_compaction &&
1637 (cc->mode == MIGRATE_ASYNC)) {
1638 cc->migrate_pfn = block_end_pfn(
1639 cc->migrate_pfn - 1, cc->order);
1640 /* Draining pcplists is useless in this case */
1641 cc->last_migrated_pfn = 0;
1642
1643 }
Mel Gorman748446b2010-05-24 14:32:27 -07001644 }
Vlastimil Babkafdaf7f52014-12-10 15:43:34 -08001645
Vlastimil Babkafdaf7f52014-12-10 15:43:34 -08001646check_drain:
1647 /*
1648 * Has the migration scanner moved away from the previous
1649 * cc->order aligned block where we migrated from? If yes,
1650 * flush the pages that were freed, so that they can merge and
1651 * compact_finished() can detect immediately if allocation
1652 * would succeed.
1653 */
Joonsoo Kim1a167182015-09-08 15:03:59 -07001654 if (cc->order > 0 && cc->last_migrated_pfn) {
Vlastimil Babkafdaf7f52014-12-10 15:43:34 -08001655 int cpu;
1656 unsigned long current_block_start =
Vlastimil Babka06b66402016-05-19 17:11:48 -07001657 block_start_pfn(cc->migrate_pfn, cc->order);
Vlastimil Babkafdaf7f52014-12-10 15:43:34 -08001658
Joonsoo Kim1a167182015-09-08 15:03:59 -07001659 if (cc->last_migrated_pfn < current_block_start) {
Vlastimil Babkafdaf7f52014-12-10 15:43:34 -08001660 cpu = get_cpu();
1661 lru_add_drain_cpu(cpu);
1662 drain_local_pages(zone);
1663 put_cpu();
1664 /* No more flushing until we migrate again */
Joonsoo Kim1a167182015-09-08 15:03:59 -07001665 cc->last_migrated_pfn = 0;
Vlastimil Babkafdaf7f52014-12-10 15:43:34 -08001666 }
1667 }
1668
Mel Gorman748446b2010-05-24 14:32:27 -07001669 }
1670
Mel Gormanf9e35b32011-06-15 15:08:52 -07001671out:
Vlastimil Babka6bace092014-12-10 15:43:31 -08001672 /*
1673 * Release free pages and update where the free scanner should restart,
1674 * so we don't leave any returned pages behind in the next attempt.
1675 */
1676 if (cc->nr_freepages > 0) {
1677 unsigned long free_pfn = release_freepages(&cc->freepages);
1678
1679 cc->nr_freepages = 0;
1680 VM_BUG_ON(free_pfn == 0);
1681 /* The cached pfn is always the first in a pageblock */
Vlastimil Babka06b66402016-05-19 17:11:48 -07001682 free_pfn = pageblock_start_pfn(free_pfn);
Vlastimil Babka6bace092014-12-10 15:43:31 -08001683 /*
1684 * Only go back, not forward. The cached pfn might have been
1685 * already reset to zone end in compact_finished()
1686 */
1687 if (free_pfn > zone->compact_cached_free_pfn)
1688 zone->compact_cached_free_pfn = free_pfn;
1689 }
Mel Gorman748446b2010-05-24 14:32:27 -07001690
David Rientjes7f354a52017-02-22 15:44:50 -08001691 count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned);
1692 count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned);
1693
Joonsoo Kim16c4a092015-02-11 15:27:01 -08001694 trace_mm_compaction_end(start_pfn, cc->migrate_pfn,
1695 cc->free_pfn, end_pfn, sync, ret);
Mel Gorman0eb927c2014-01-21 15:51:05 -08001696
Mel Gorman748446b2010-05-24 14:32:27 -07001697 return ret;
1698}
Mel Gorman76ab0f52010-05-24 14:32:28 -07001699
Michal Hockoea7ab982016-05-20 16:56:38 -07001700static enum compact_result compact_zone_order(struct zone *zone, int order,
Vlastimil Babkac3486f52016-07-28 15:49:30 -07001701 gfp_t gfp_mask, enum compact_priority prio,
Mel Gormanc6038442016-05-19 17:13:38 -07001702 unsigned int alloc_flags, int classzone_idx)
Mel Gorman56de7262010-05-24 14:32:30 -07001703{
Michal Hockoea7ab982016-05-20 16:56:38 -07001704 enum compact_result ret;
Mel Gorman56de7262010-05-24 14:32:30 -07001705 struct compact_control cc = {
1706 .nr_freepages = 0,
1707 .nr_migratepages = 0,
David Rientjes7f354a52017-02-22 15:44:50 -08001708 .total_migrate_scanned = 0,
1709 .total_free_scanned = 0,
Mel Gorman56de7262010-05-24 14:32:30 -07001710 .order = order,
David Rientjes6d7ce552014-10-09 15:27:27 -07001711 .gfp_mask = gfp_mask,
Mel Gorman56de7262010-05-24 14:32:30 -07001712 .zone = zone,
Vlastimil Babkaa5508cd2016-07-28 15:49:28 -07001713 .mode = (prio == COMPACT_PRIO_ASYNC) ?
1714 MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT,
Vlastimil Babkaebff3982014-12-10 15:43:22 -08001715 .alloc_flags = alloc_flags,
1716 .classzone_idx = classzone_idx,
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07001717 .direct_compaction = true,
Vlastimil Babkaa8e025e2016-10-07 16:57:47 -07001718 .whole_zone = (prio == MIN_COMPACT_PRIORITY),
Vlastimil Babka9f7e3382016-10-07 17:00:37 -07001719 .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
1720 .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
Mel Gorman56de7262010-05-24 14:32:30 -07001721 };
1722 INIT_LIST_HEAD(&cc.freepages);
1723 INIT_LIST_HEAD(&cc.migratepages);
1724
Shaohua Lie64c5232012-10-08 16:32:27 -07001725 ret = compact_zone(zone, &cc);
1726
1727 VM_BUG_ON(!list_empty(&cc.freepages));
1728 VM_BUG_ON(!list_empty(&cc.migratepages));
1729
Shaohua Lie64c5232012-10-08 16:32:27 -07001730 return ret;
Mel Gorman56de7262010-05-24 14:32:30 -07001731}
1732
Mel Gorman5e771902010-05-24 14:32:31 -07001733int sysctl_extfrag_threshold = 500;
1734
Mel Gorman56de7262010-05-24 14:32:30 -07001735/**
1736 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
Mel Gorman56de7262010-05-24 14:32:30 -07001737 * @gfp_mask: The GFP mask of the current allocation
Vlastimil Babka1a6d53a2015-02-11 15:25:44 -08001738 * @order: The order of the current allocation
1739 * @alloc_flags: The allocation flags of the current allocation
1740 * @ac: The context of current allocation
David Rientjese0b9dae2014-06-04 16:08:28 -07001741 * @mode: The migration mode for async, sync light, or sync migration
Mel Gorman56de7262010-05-24 14:32:30 -07001742 *
1743 * This is the main entry point for direct page compaction.
1744 */
Michal Hockoea7ab982016-05-20 16:56:38 -07001745enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
Mel Gormanc6038442016-05-19 17:13:38 -07001746 unsigned int alloc_flags, const struct alloc_context *ac,
Vlastimil Babkac3486f52016-07-28 15:49:30 -07001747 enum compact_priority prio)
Mel Gorman56de7262010-05-24 14:32:30 -07001748{
Mel Gorman56de7262010-05-24 14:32:30 -07001749 int may_perform_io = gfp_mask & __GFP_IO;
Mel Gorman56de7262010-05-24 14:32:30 -07001750 struct zoneref *z;
1751 struct zone *zone;
Michal Hocko1d4746d2016-05-20 16:56:44 -07001752 enum compact_result rc = COMPACT_SKIPPED;
Mel Gorman56de7262010-05-24 14:32:30 -07001753
Michal Hocko73e64c52016-12-14 15:04:07 -08001754 /*
1755 * Check if the GFP flags allow compaction - GFP_NOIO is really
1756 * tricky context because the migration might require IO
1757 */
1758 if (!may_perform_io)
Vlastimil Babka53853e22014-10-09 15:27:02 -07001759 return COMPACT_SKIPPED;
Mel Gorman56de7262010-05-24 14:32:30 -07001760
Vlastimil Babkaa5508cd2016-07-28 15:49:28 -07001761 trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
Joonsoo Kim837d0262015-02-11 15:27:06 -08001762
Mel Gorman56de7262010-05-24 14:32:30 -07001763 /* Compact each zone in the list */
Vlastimil Babka1a6d53a2015-02-11 15:25:44 -08001764 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1765 ac->nodemask) {
Michal Hockoea7ab982016-05-20 16:56:38 -07001766 enum compact_result status;
Mel Gorman56de7262010-05-24 14:32:30 -07001767
Vlastimil Babkaa8e025e2016-10-07 16:57:47 -07001768 if (prio > MIN_COMPACT_PRIORITY
1769 && compaction_deferred(zone, order)) {
Michal Hocko1d4746d2016-05-20 16:56:44 -07001770 rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
Vlastimil Babka53853e22014-10-09 15:27:02 -07001771 continue;
Michal Hocko1d4746d2016-05-20 16:56:44 -07001772 }
Vlastimil Babka53853e22014-10-09 15:27:02 -07001773
Vlastimil Babkaa5508cd2016-07-28 15:49:28 -07001774 status = compact_zone_order(zone, order, gfp_mask, prio,
Vlastimil Babkac3486f52016-07-28 15:49:30 -07001775 alloc_flags, ac_classzone_idx(ac));
Mel Gorman56de7262010-05-24 14:32:30 -07001776 rc = max(status, rc);
1777
Vlastimil Babka7ceb0092016-10-07 16:57:44 -07001778 /* The allocation should succeed, stop compacting */
1779 if (status == COMPACT_SUCCESS) {
Vlastimil Babka53853e22014-10-09 15:27:02 -07001780 /*
1781 * We think the allocation will succeed in this zone,
1782 * but it is not certain, hence the false. The caller
1783 * will repeat this with true if allocation indeed
1784 * succeeds in this zone.
1785 */
1786 compaction_defer_reset(zone, order, false);
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001787
Vlastimil Babkac3486f52016-07-28 15:49:30 -07001788 break;
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001789 }
1790
Vlastimil Babkaa5508cd2016-07-28 15:49:28 -07001791 if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
Vlastimil Babkac3486f52016-07-28 15:49:30 -07001792 status == COMPACT_PARTIAL_SKIPPED))
Vlastimil Babka53853e22014-10-09 15:27:02 -07001793 /*
1794 * We think that allocation won't succeed in this zone
1795 * so we defer compaction there. If it ends up
1796 * succeeding after all, it will be reset.
1797 */
1798 defer_compaction(zone, order);
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001799
1800 /*
1801 * We might have stopped compacting due to need_resched() in
1802 * async compaction, or due to a fatal signal detected. In that
Vlastimil Babkac3486f52016-07-28 15:49:30 -07001803 * case do not try further zones
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001804 */
Vlastimil Babkac3486f52016-07-28 15:49:30 -07001805 if ((prio == COMPACT_PRIO_ASYNC && need_resched())
1806 || fatal_signal_pending(current))
1807 break;
Mel Gorman56de7262010-05-24 14:32:30 -07001808 }
1809
1810 return rc;
1811}
1812
1813
Mel Gorman76ab0f52010-05-24 14:32:28 -07001814/* Compact all zones within a node */
Andrew Morton7103f162013-02-22 16:32:33 -08001815static void compact_node(int nid)
Rik van Riel7be62de2012-03-21 16:33:52 -07001816{
Vlastimil Babka791cae92016-10-07 16:57:38 -07001817 pg_data_t *pgdat = NODE_DATA(nid);
1818 int zoneid;
1819 struct zone *zone;
Rik van Riel7be62de2012-03-21 16:33:52 -07001820 struct compact_control cc = {
1821 .order = -1,
David Rientjes7f354a52017-02-22 15:44:50 -08001822 .total_migrate_scanned = 0,
1823 .total_free_scanned = 0,
David Rientjese0b9dae2014-06-04 16:08:28 -07001824 .mode = MIGRATE_SYNC,
David Rientjes91ca9182014-04-03 14:47:23 -07001825 .ignore_skip_hint = true,
Vlastimil Babka06ed2992016-10-07 16:57:35 -07001826 .whole_zone = true,
Michal Hocko73e64c52016-12-14 15:04:07 -08001827 .gfp_mask = GFP_KERNEL,
Rik van Riel7be62de2012-03-21 16:33:52 -07001828 };
1829
Vlastimil Babka791cae92016-10-07 16:57:38 -07001830
1831 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1832
1833 zone = &pgdat->node_zones[zoneid];
1834 if (!populated_zone(zone))
1835 continue;
1836
1837 cc.nr_freepages = 0;
1838 cc.nr_migratepages = 0;
1839 cc.zone = zone;
1840 INIT_LIST_HEAD(&cc.freepages);
1841 INIT_LIST_HEAD(&cc.migratepages);
1842
1843 compact_zone(zone, &cc);
1844
1845 VM_BUG_ON(!list_empty(&cc.freepages));
1846 VM_BUG_ON(!list_empty(&cc.migratepages));
1847 }
Rik van Riel7be62de2012-03-21 16:33:52 -07001848}
1849
Mel Gorman76ab0f52010-05-24 14:32:28 -07001850/* Compact all nodes in the system */
Jason Liu7964c062013-01-11 14:31:47 -08001851static void compact_nodes(void)
Mel Gorman76ab0f52010-05-24 14:32:28 -07001852{
1853 int nid;
1854
Hugh Dickins8575ec22012-03-21 16:33:53 -07001855 /* Flush pending updates to the LRU lists */
1856 lru_add_drain_all();
1857
Mel Gorman76ab0f52010-05-24 14:32:28 -07001858 for_each_online_node(nid)
1859 compact_node(nid);
Mel Gorman76ab0f52010-05-24 14:32:28 -07001860}
1861
1862/* The written value is actually unused, all memory is compacted */
1863int sysctl_compact_memory;
1864
Yaowei Baifec4eb22016-01-14 15:20:09 -08001865/*
1866 * This is the entry point for compacting all nodes via
1867 * /proc/sys/vm/compact_memory
1868 */
Mel Gorman76ab0f52010-05-24 14:32:28 -07001869int sysctl_compaction_handler(struct ctl_table *table, int write,
1870 void __user *buffer, size_t *length, loff_t *ppos)
1871{
1872 if (write)
Jason Liu7964c062013-01-11 14:31:47 -08001873 compact_nodes();
Mel Gorman76ab0f52010-05-24 14:32:28 -07001874
1875 return 0;
1876}
Mel Gormaned4a6d72010-05-24 14:32:29 -07001877
Mel Gorman5e771902010-05-24 14:32:31 -07001878int sysctl_extfrag_handler(struct ctl_table *table, int write,
1879 void __user *buffer, size_t *length, loff_t *ppos)
1880{
1881 proc_dointvec_minmax(table, write, buffer, length, ppos);
1882
1883 return 0;
1884}
1885
Mel Gormaned4a6d72010-05-24 14:32:29 -07001886#if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
Rashika Kheria74e77fb2014-04-03 14:48:01 -07001887static ssize_t sysfs_compact_node(struct device *dev,
Kay Sievers10fbcf42011-12-21 14:48:43 -08001888 struct device_attribute *attr,
Mel Gormaned4a6d72010-05-24 14:32:29 -07001889 const char *buf, size_t count)
1890{
Hugh Dickins8575ec22012-03-21 16:33:53 -07001891 int nid = dev->id;
1892
1893 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1894 /* Flush pending updates to the LRU lists */
1895 lru_add_drain_all();
1896
1897 compact_node(nid);
1898 }
Mel Gormaned4a6d72010-05-24 14:32:29 -07001899
1900 return count;
1901}
Kay Sievers10fbcf42011-12-21 14:48:43 -08001902static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
Mel Gormaned4a6d72010-05-24 14:32:29 -07001903
1904int compaction_register_node(struct node *node)
1905{
Kay Sievers10fbcf42011-12-21 14:48:43 -08001906 return device_create_file(&node->dev, &dev_attr_compact);
Mel Gormaned4a6d72010-05-24 14:32:29 -07001907}
1908
1909void compaction_unregister_node(struct node *node)
1910{
Kay Sievers10fbcf42011-12-21 14:48:43 -08001911 return device_remove_file(&node->dev, &dev_attr_compact);
Mel Gormaned4a6d72010-05-24 14:32:29 -07001912}
1913#endif /* CONFIG_SYSFS && CONFIG_NUMA */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001914
Vlastimil Babka698b1b32016-03-17 14:18:08 -07001915static inline bool kcompactd_work_requested(pg_data_t *pgdat)
1916{
Vlastimil Babka172400c2016-05-05 16:22:32 -07001917 return pgdat->kcompactd_max_order > 0 || kthread_should_stop();
Vlastimil Babka698b1b32016-03-17 14:18:08 -07001918}
1919
1920static bool kcompactd_node_suitable(pg_data_t *pgdat)
1921{
1922 int zoneid;
1923 struct zone *zone;
1924 enum zone_type classzone_idx = pgdat->kcompactd_classzone_idx;
1925
Chen Feng6cd9dc32016-05-20 16:59:02 -07001926 for (zoneid = 0; zoneid <= classzone_idx; zoneid++) {
Vlastimil Babka698b1b32016-03-17 14:18:08 -07001927 zone = &pgdat->node_zones[zoneid];
1928
1929 if (!populated_zone(zone))
1930 continue;
1931
1932 if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0,
1933 classzone_idx) == COMPACT_CONTINUE)
1934 return true;
1935 }
1936
1937 return false;
1938}
1939
1940static void kcompactd_do_work(pg_data_t *pgdat)
1941{
1942 /*
1943 * With no special task, compact all zones so that a page of requested
1944 * order is allocatable.
1945 */
1946 int zoneid;
1947 struct zone *zone;
1948 struct compact_control cc = {
1949 .order = pgdat->kcompactd_max_order,
David Rientjes7f354a52017-02-22 15:44:50 -08001950 .total_migrate_scanned = 0,
1951 .total_free_scanned = 0,
Vlastimil Babka698b1b32016-03-17 14:18:08 -07001952 .classzone_idx = pgdat->kcompactd_classzone_idx,
1953 .mode = MIGRATE_SYNC_LIGHT,
David Rientjesa0647dc2017-11-17 15:26:27 -08001954 .ignore_skip_hint = false,
Michal Hocko73e64c52016-12-14 15:04:07 -08001955 .gfp_mask = GFP_KERNEL,
Vlastimil Babka698b1b32016-03-17 14:18:08 -07001956 };
Vlastimil Babka698b1b32016-03-17 14:18:08 -07001957 trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
1958 cc.classzone_idx);
David Rientjes7f354a52017-02-22 15:44:50 -08001959 count_compact_event(KCOMPACTD_WAKE);
Vlastimil Babka698b1b32016-03-17 14:18:08 -07001960
Chen Feng6cd9dc32016-05-20 16:59:02 -07001961 for (zoneid = 0; zoneid <= cc.classzone_idx; zoneid++) {
Vlastimil Babka698b1b32016-03-17 14:18:08 -07001962 int status;
1963
1964 zone = &pgdat->node_zones[zoneid];
1965 if (!populated_zone(zone))
1966 continue;
1967
1968 if (compaction_deferred(zone, cc.order))
1969 continue;
1970
1971 if (compaction_suitable(zone, cc.order, 0, zoneid) !=
1972 COMPACT_CONTINUE)
1973 continue;
1974
1975 cc.nr_freepages = 0;
1976 cc.nr_migratepages = 0;
David Rientjes7f354a52017-02-22 15:44:50 -08001977 cc.total_migrate_scanned = 0;
1978 cc.total_free_scanned = 0;
Vlastimil Babka698b1b32016-03-17 14:18:08 -07001979 cc.zone = zone;
1980 INIT_LIST_HEAD(&cc.freepages);
1981 INIT_LIST_HEAD(&cc.migratepages);
1982
Vlastimil Babka172400c2016-05-05 16:22:32 -07001983 if (kthread_should_stop())
1984 return;
Vlastimil Babka698b1b32016-03-17 14:18:08 -07001985 status = compact_zone(zone, &cc);
1986
Vlastimil Babka7ceb0092016-10-07 16:57:44 -07001987 if (status == COMPACT_SUCCESS) {
Vlastimil Babka698b1b32016-03-17 14:18:08 -07001988 compaction_defer_reset(zone, cc.order, false);
Michal Hockoc8f7de02016-05-20 16:56:47 -07001989 } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
Vlastimil Babka698b1b32016-03-17 14:18:08 -07001990 /*
1991 * We use sync migration mode here, so we defer like
1992 * sync direct compaction does.
1993 */
1994 defer_compaction(zone, cc.order);
1995 }
1996
David Rientjes7f354a52017-02-22 15:44:50 -08001997 count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
1998 cc.total_migrate_scanned);
1999 count_compact_events(KCOMPACTD_FREE_SCANNED,
2000 cc.total_free_scanned);
2001
Vlastimil Babka698b1b32016-03-17 14:18:08 -07002002 VM_BUG_ON(!list_empty(&cc.freepages));
2003 VM_BUG_ON(!list_empty(&cc.migratepages));
2004 }
2005
2006 /*
2007 * Regardless of success, we are done until woken up next. But remember
2008 * the requested order/classzone_idx in case it was higher/tighter than
2009 * our current ones
2010 */
2011 if (pgdat->kcompactd_max_order <= cc.order)
2012 pgdat->kcompactd_max_order = 0;
2013 if (pgdat->kcompactd_classzone_idx >= cc.classzone_idx)
2014 pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
2015}
2016
2017void wakeup_kcompactd(pg_data_t *pgdat, int order, int classzone_idx)
2018{
2019 if (!order)
2020 return;
2021
2022 if (pgdat->kcompactd_max_order < order)
2023 pgdat->kcompactd_max_order = order;
2024
2025 if (pgdat->kcompactd_classzone_idx > classzone_idx)
2026 pgdat->kcompactd_classzone_idx = classzone_idx;
2027
Davidlohr Bueso68186002017-10-03 16:15:03 -07002028 /*
2029 * Pairs with implicit barrier in wait_event_freezable()
2030 * such that wakeups are not missed.
2031 */
2032 if (!wq_has_sleeper(&pgdat->kcompactd_wait))
Vlastimil Babka698b1b32016-03-17 14:18:08 -07002033 return;
2034
2035 if (!kcompactd_node_suitable(pgdat))
2036 return;
2037
2038 trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
2039 classzone_idx);
2040 wake_up_interruptible(&pgdat->kcompactd_wait);
2041}
2042
2043/*
2044 * The background compaction daemon, started as a kernel thread
2045 * from the init process.
2046 */
2047static int kcompactd(void *p)
2048{
2049 pg_data_t *pgdat = (pg_data_t*)p;
2050 struct task_struct *tsk = current;
2051
2052 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2053
2054 if (!cpumask_empty(cpumask))
2055 set_cpus_allowed_ptr(tsk, cpumask);
2056
2057 set_freezable();
2058
2059 pgdat->kcompactd_max_order = 0;
2060 pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
2061
2062 while (!kthread_should_stop()) {
2063 trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
2064 wait_event_freezable(pgdat->kcompactd_wait,
2065 kcompactd_work_requested(pgdat));
2066
2067 kcompactd_do_work(pgdat);
2068 }
2069
2070 return 0;
2071}
2072
2073/*
2074 * This kcompactd start function will be called by init and node-hot-add.
2075 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
2076 */
2077int kcompactd_run(int nid)
2078{
2079 pg_data_t *pgdat = NODE_DATA(nid);
2080 int ret = 0;
2081
2082 if (pgdat->kcompactd)
2083 return 0;
2084
2085 pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
2086 if (IS_ERR(pgdat->kcompactd)) {
2087 pr_err("Failed to start kcompactd on node %d\n", nid);
2088 ret = PTR_ERR(pgdat->kcompactd);
2089 pgdat->kcompactd = NULL;
2090 }
2091 return ret;
2092}
2093
2094/*
2095 * Called by memory hotplug when all memory in a node is offlined. Caller must
2096 * hold mem_hotplug_begin/end().
2097 */
2098void kcompactd_stop(int nid)
2099{
2100 struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
2101
2102 if (kcompactd) {
2103 kthread_stop(kcompactd);
2104 NODE_DATA(nid)->kcompactd = NULL;
2105 }
2106}
2107
2108/*
2109 * It's optimal to keep kcompactd on the same CPUs as their memory, but
2110 * not required for correctness. So if the last cpu in a node goes
2111 * away, we get changed to run anywhere: as the first one comes back,
2112 * restore their cpu bindings.
2113 */
Anna-Maria Gleixnere46b1db2016-11-27 00:13:42 +01002114static int kcompactd_cpu_online(unsigned int cpu)
Vlastimil Babka698b1b32016-03-17 14:18:08 -07002115{
2116 int nid;
2117
Anna-Maria Gleixnere46b1db2016-11-27 00:13:42 +01002118 for_each_node_state(nid, N_MEMORY) {
2119 pg_data_t *pgdat = NODE_DATA(nid);
2120 const struct cpumask *mask;
Vlastimil Babka698b1b32016-03-17 14:18:08 -07002121
Anna-Maria Gleixnere46b1db2016-11-27 00:13:42 +01002122 mask = cpumask_of_node(pgdat->node_id);
Vlastimil Babka698b1b32016-03-17 14:18:08 -07002123
Anna-Maria Gleixnere46b1db2016-11-27 00:13:42 +01002124 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2125 /* One of our CPUs online: restore mask */
2126 set_cpus_allowed_ptr(pgdat->kcompactd, mask);
Vlastimil Babka698b1b32016-03-17 14:18:08 -07002127 }
Anna-Maria Gleixnere46b1db2016-11-27 00:13:42 +01002128 return 0;
Vlastimil Babka698b1b32016-03-17 14:18:08 -07002129}
2130
2131static int __init kcompactd_init(void)
2132{
2133 int nid;
Anna-Maria Gleixnere46b1db2016-11-27 00:13:42 +01002134 int ret;
2135
2136 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
2137 "mm/compaction:online",
2138 kcompactd_cpu_online, NULL);
2139 if (ret < 0) {
2140 pr_err("kcompactd: failed to register hotplug callbacks.\n");
2141 return ret;
2142 }
Vlastimil Babka698b1b32016-03-17 14:18:08 -07002143
2144 for_each_node_state(nid, N_MEMORY)
2145 kcompactd_run(nid);
Vlastimil Babka698b1b32016-03-17 14:18:08 -07002146 return 0;
2147}
2148subsys_initcall(kcompactd_init)
2149
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01002150#endif /* CONFIG_COMPACTION */