blob: e8f5b4e2cb05511c7b9c53d1351e8a17c93afb0a [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
298 if (cc->ignore_skip_hint)
299 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
David Rientjes21dc7e02017-11-17 15:26:30 -0800479 if (pageblock_skip_persistent(page, order)) {
480 set_pageblock_skip(page);
481 blockpfn = end_pfn;
482 } else if (likely(order < MAX_ORDER)) {
483 blockpfn += (1UL << order) - 1;
484 cursor += (1UL << order) - 1;
Vlastimil Babka9fcd6d22015-09-08 15:02:49 -0700485 }
Vlastimil Babka9fcd6d22015-09-08 15:02:49 -0700486 goto isolate_fail;
487 }
488
Mel Gormanf40d1e42012-10-08 16:32:36 -0700489 if (!PageBuddy(page))
Laura Abbott2af120b2014-03-10 15:49:44 -0700490 goto isolate_fail;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700491
492 /*
Vlastimil Babka69b71892014-10-09 15:27:18 -0700493 * If we already hold the lock, we can skip some rechecking.
494 * Note that if we hold the lock now, checked_pageblock was
495 * already set in some previous iteration (or strict is true),
496 * so it is correct to skip the suitable migration target
497 * recheck as well.
Mel Gormanf40d1e42012-10-08 16:32:36 -0700498 */
Vlastimil Babka69b71892014-10-09 15:27:18 -0700499 if (!locked) {
500 /*
501 * The zone lock must be held to isolate freepages.
502 * Unfortunately this is a very coarse lock and can be
503 * heavily contended if there are parallel allocations
504 * or parallel compactions. For async compaction do not
505 * spin on the lock and we acquire the lock as late as
506 * possible.
507 */
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700508 locked = compact_trylock_irqsave(&cc->zone->lock,
509 &flags, cc);
Vlastimil Babka69b71892014-10-09 15:27:18 -0700510 if (!locked)
511 break;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700512
Vlastimil Babka69b71892014-10-09 15:27:18 -0700513 /* Recheck this is a buddy page under lock */
514 if (!PageBuddy(page))
515 goto isolate_fail;
516 }
Mel Gorman748446b2010-05-24 14:32:27 -0700517
Joonsoo Kim66c64222016-07-26 15:23:40 -0700518 /* Found a free page, will break it into order-0 pages */
519 order = page_order(page);
520 isolated = __isolate_free_page(page, order);
David Rientjesa4f04f22016-06-24 14:50:10 -0700521 if (!isolated)
522 break;
Joonsoo Kim66c64222016-07-26 15:23:40 -0700523 set_page_private(page, order);
David Rientjesa4f04f22016-06-24 14:50:10 -0700524
Mel Gorman748446b2010-05-24 14:32:27 -0700525 total_isolated += isolated;
David Rientjesa4f04f22016-06-24 14:50:10 -0700526 cc->nr_freepages += isolated;
Joonsoo Kim66c64222016-07-26 15:23:40 -0700527 list_add_tail(&page->lru, freelist);
528
David Rientjesa4f04f22016-06-24 14:50:10 -0700529 if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
530 blockpfn += isolated;
531 break;
Mel Gorman748446b2010-05-24 14:32:27 -0700532 }
David Rientjesa4f04f22016-06-24 14:50:10 -0700533 /* Advance to the end of split page */
534 blockpfn += isolated - 1;
535 cursor += isolated - 1;
536 continue;
Laura Abbott2af120b2014-03-10 15:49:44 -0700537
538isolate_fail:
539 if (strict)
540 break;
541 else
542 continue;
543
Mel Gorman748446b2010-05-24 14:32:27 -0700544 }
545
David Rientjesa4f04f22016-06-24 14:50:10 -0700546 if (locked)
547 spin_unlock_irqrestore(&cc->zone->lock, flags);
548
Vlastimil Babka9fcd6d22015-09-08 15:02:49 -0700549 /*
550 * There is a tiny chance that we have read bogus compound_order(),
551 * so be careful to not go outside of the pageblock.
552 */
553 if (unlikely(blockpfn > end_pfn))
554 blockpfn = end_pfn;
555
Joonsoo Kime34d85f2015-02-11 15:27:04 -0800556 trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
557 nr_scanned, total_isolated);
558
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700559 /* Record how far we have got within the block */
560 *start_pfn = blockpfn;
561
Mel Gormanf40d1e42012-10-08 16:32:36 -0700562 /*
563 * If strict isolation is requested by CMA then check that all the
564 * pages requested were isolated. If there were any failures, 0 is
565 * returned and CMA will fail.
566 */
Laura Abbott2af120b2014-03-10 15:49:44 -0700567 if (strict && blockpfn < end_pfn)
Mel Gormanf40d1e42012-10-08 16:32:36 -0700568 total_isolated = 0;
569
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700570 /* Update the pageblock-skip if the whole pageblock was scanned */
571 if (blockpfn == end_pfn)
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700572 update_pageblock_skip(cc, valid_page, total_isolated, false);
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700573
David Rientjes7f354a52017-02-22 15:44:50 -0800574 cc->total_free_scanned += nr_scanned;
Mel Gorman397487d2012-10-19 12:00:10 +0100575 if (total_isolated)
Minchan Kim010fc292012-12-20 15:05:06 -0800576 count_compact_events(COMPACTISOLATED, total_isolated);
Mel Gorman748446b2010-05-24 14:32:27 -0700577 return total_isolated;
578}
579
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100580/**
581 * isolate_freepages_range() - isolate free pages.
582 * @start_pfn: The first PFN to start isolating.
583 * @end_pfn: The one-past-last PFN.
584 *
585 * Non-free pages, invalid PFNs, or zone boundaries within the
586 * [start_pfn, end_pfn) range are considered errors, cause function to
587 * undo its actions and return zero.
588 *
589 * Otherwise, function returns one-past-the-last PFN of isolated page
590 * (which may be greater then end_pfn if end fell in a middle of
591 * a free page).
592 */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100593unsigned long
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700594isolate_freepages_range(struct compact_control *cc,
595 unsigned long start_pfn, unsigned long end_pfn)
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100596{
Joonsoo Kime1409c32016-03-15 14:57:48 -0700597 unsigned long isolated, pfn, block_start_pfn, block_end_pfn;
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100598 LIST_HEAD(freelist);
599
Vlastimil Babka7d49d882014-10-09 15:27:11 -0700600 pfn = start_pfn;
Vlastimil Babka06b66402016-05-19 17:11:48 -0700601 block_start_pfn = pageblock_start_pfn(pfn);
Joonsoo Kime1409c32016-03-15 14:57:48 -0700602 if (block_start_pfn < cc->zone->zone_start_pfn)
603 block_start_pfn = cc->zone->zone_start_pfn;
Vlastimil Babka06b66402016-05-19 17:11:48 -0700604 block_end_pfn = pageblock_end_pfn(pfn);
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100605
Vlastimil Babka7d49d882014-10-09 15:27:11 -0700606 for (; pfn < end_pfn; pfn += isolated,
Joonsoo Kime1409c32016-03-15 14:57:48 -0700607 block_start_pfn = block_end_pfn,
Vlastimil Babka7d49d882014-10-09 15:27:11 -0700608 block_end_pfn += pageblock_nr_pages) {
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700609 /* Protect pfn from changing by isolate_freepages_block */
610 unsigned long isolate_start_pfn = pfn;
Vlastimil Babka7d49d882014-10-09 15:27:11 -0700611
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100612 block_end_pfn = min(block_end_pfn, end_pfn);
613
Joonsoo Kim58420012014-11-13 15:19:07 -0800614 /*
615 * pfn could pass the block_end_pfn if isolated freepage
616 * is more than pageblock order. In this case, we adjust
617 * scanning range to right one.
618 */
619 if (pfn >= block_end_pfn) {
Vlastimil Babka06b66402016-05-19 17:11:48 -0700620 block_start_pfn = pageblock_start_pfn(pfn);
621 block_end_pfn = pageblock_end_pfn(pfn);
Joonsoo Kim58420012014-11-13 15:19:07 -0800622 block_end_pfn = min(block_end_pfn, end_pfn);
623 }
624
Joonsoo Kime1409c32016-03-15 14:57:48 -0700625 if (!pageblock_pfn_to_page(block_start_pfn,
626 block_end_pfn, cc->zone))
Vlastimil Babka7d49d882014-10-09 15:27:11 -0700627 break;
628
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700629 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
630 block_end_pfn, &freelist, true);
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100631
632 /*
633 * In strict mode, isolate_freepages_block() returns 0 if
634 * there are any holes in the block (ie. invalid PFNs or
635 * non-free pages).
636 */
637 if (!isolated)
638 break;
639
640 /*
641 * If we managed to isolate pages, it is always (1 << n) *
642 * pageblock_nr_pages for some non-negative n. (Max order
643 * page may span two pageblocks).
644 */
645 }
646
Joonsoo Kim66c64222016-07-26 15:23:40 -0700647 /* __isolate_free_page() does not map the pages */
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100648 map_pages(&freelist);
649
650 if (pfn < end_pfn) {
651 /* Loop terminated early, cleanup. */
652 release_freepages(&freelist);
653 return 0;
654 }
655
656 /* We don't use freelists for anything. */
657 return pfn;
658}
659
Mel Gorman748446b2010-05-24 14:32:27 -0700660/* Similar to reclaim, but different enough that they don't share logic */
661static bool too_many_isolated(struct zone *zone)
662{
Minchan Kimbc693042010-09-09 16:38:00 -0700663 unsigned long active, inactive, isolated;
Mel Gorman748446b2010-05-24 14:32:27 -0700664
Mel Gorman599d0c92016-07-28 15:45:31 -0700665 inactive = node_page_state(zone->zone_pgdat, NR_INACTIVE_FILE) +
666 node_page_state(zone->zone_pgdat, NR_INACTIVE_ANON);
667 active = node_page_state(zone->zone_pgdat, NR_ACTIVE_FILE) +
668 node_page_state(zone->zone_pgdat, NR_ACTIVE_ANON);
669 isolated = node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE) +
670 node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON);
Mel Gorman748446b2010-05-24 14:32:27 -0700671
Minchan Kimbc693042010-09-09 16:38:00 -0700672 return isolated > (inactive + active) / 2;
Mel Gorman748446b2010-05-24 14:32:27 -0700673}
674
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100675/**
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700676 * isolate_migratepages_block() - isolate all migrate-able pages within
677 * a single pageblock
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100678 * @cc: Compaction control structure.
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700679 * @low_pfn: The first PFN to isolate
680 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
681 * @isolate_mode: Isolation mode to be used.
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100682 *
683 * Isolate all pages that can be migrated from the range specified by
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700684 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
685 * Returns zero if there is a fatal signal pending, otherwise PFN of the
686 * first page that was not scanned (which may be both less, equal to or more
687 * than end_pfn).
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100688 *
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700689 * The pages are isolated on cc->migratepages list (not required to be empty),
690 * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
691 * is neither read nor updated.
Mel Gorman748446b2010-05-24 14:32:27 -0700692 */
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700693static unsigned long
694isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
695 unsigned long end_pfn, isolate_mode_t isolate_mode)
Mel Gorman748446b2010-05-24 14:32:27 -0700696{
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700697 struct zone *zone = cc->zone;
Mel Gormanb7aba692011-01-13 15:45:54 -0800698 unsigned long nr_scanned = 0, nr_isolated = 0;
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700699 struct lruvec *lruvec;
Xiubo Lib8b2d822014-10-09 15:28:21 -0700700 unsigned long flags = 0;
Mel Gorman2a1402a2012-10-08 16:32:33 -0700701 bool locked = false;
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700702 struct page *page = NULL, *valid_page = NULL;
Joonsoo Kime34d85f2015-02-11 15:27:04 -0800703 unsigned long start_pfn = low_pfn;
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700704 bool skip_on_failure = false;
705 unsigned long next_skip_pfn = 0;
Mel Gorman748446b2010-05-24 14:32:27 -0700706
Mel Gorman748446b2010-05-24 14:32:27 -0700707 /*
708 * Ensure that there are not too many pages isolated from the LRU
709 * list by either parallel reclaimers or compaction. If there are,
710 * delay for some time until fewer pages are isolated
711 */
712 while (unlikely(too_many_isolated(zone))) {
Mel Gormanf9e35b32011-06-15 15:08:52 -0700713 /* async migration should just abort */
David Rientjese0b9dae2014-06-04 16:08:28 -0700714 if (cc->mode == MIGRATE_ASYNC)
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100715 return 0;
Mel Gormanf9e35b32011-06-15 15:08:52 -0700716
Mel Gorman748446b2010-05-24 14:32:27 -0700717 congestion_wait(BLK_RW_ASYNC, HZ/10);
718
719 if (fatal_signal_pending(current))
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100720 return 0;
Mel Gorman748446b2010-05-24 14:32:27 -0700721 }
722
Vlastimil Babkabe976572014-06-04 16:10:41 -0700723 if (compact_should_abort(cc))
724 return 0;
David Rientjesaeef4b82014-06-04 16:08:31 -0700725
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700726 if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
727 skip_on_failure = true;
728 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
729 }
730
Mel Gorman748446b2010-05-24 14:32:27 -0700731 /* Time to isolate some pages for migration */
Mel Gorman748446b2010-05-24 14:32:27 -0700732 for (; low_pfn < end_pfn; low_pfn++) {
Vlastimil Babka29c0dde2015-09-08 15:02:46 -0700733
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700734 if (skip_on_failure && low_pfn >= next_skip_pfn) {
735 /*
736 * We have isolated all migration candidates in the
737 * previous order-aligned block, and did not skip it due
738 * to failure. We should migrate the pages now and
739 * hopefully succeed compaction.
740 */
741 if (nr_isolated)
742 break;
743
744 /*
745 * We failed to isolate in the previous order-aligned
746 * block. Set the new boundary to the end of the
747 * current block. Note we can't simply increase
748 * next_skip_pfn by 1 << order, as low_pfn might have
749 * been incremented by a higher number due to skipping
750 * a compound or a high-order buddy page in the
751 * previous loop iteration.
752 */
753 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
754 }
755
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700756 /*
757 * Periodically drop the lock (if held) regardless of its
758 * contention, to give chance to IRQs. Abort async compaction
759 * if contended.
760 */
761 if (!(low_pfn % SWAP_CLUSTER_MAX)
Mel Gormana52633d2016-07-28 15:45:28 -0700762 && compact_unlock_should_abort(zone_lru_lock(zone), flags,
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700763 &locked, cc))
764 break;
Mel Gormanc67fe372012-08-21 16:16:17 -0700765
Mel Gorman748446b2010-05-24 14:32:27 -0700766 if (!pfn_valid_within(low_pfn))
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700767 goto isolate_fail;
Mel Gormanb7aba692011-01-13 15:45:54 -0800768 nr_scanned++;
Mel Gorman748446b2010-05-24 14:32:27 -0700769
Mel Gorman748446b2010-05-24 14:32:27 -0700770 page = pfn_to_page(low_pfn);
Mel Gormandc908602012-02-08 17:13:38 -0800771
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700772 if (!valid_page)
773 valid_page = page;
774
Mel Gorman6c144662014-01-23 15:53:38 -0800775 /*
Vlastimil Babka99c0fd52014-10-09 15:27:23 -0700776 * Skip if free. We read page order here without zone lock
777 * which is generally unsafe, but the race window is small and
778 * the worst thing that can happen is that we skip some
779 * potential isolation targets.
Mel Gorman6c144662014-01-23 15:53:38 -0800780 */
Vlastimil Babka99c0fd52014-10-09 15:27:23 -0700781 if (PageBuddy(page)) {
782 unsigned long freepage_order = page_order_unsafe(page);
783
784 /*
785 * Without lock, we cannot be sure that what we got is
786 * a valid page order. Consider only values in the
787 * valid order range to prevent low_pfn overflow.
788 */
789 if (freepage_order > 0 && freepage_order < MAX_ORDER)
790 low_pfn += (1UL << freepage_order) - 1;
Mel Gorman748446b2010-05-24 14:32:27 -0700791 continue;
Vlastimil Babka99c0fd52014-10-09 15:27:23 -0700792 }
Mel Gorman748446b2010-05-24 14:32:27 -0700793
Mel Gorman9927af742011-01-13 15:45:59 -0800794 /*
Vlastimil Babka29c0dde2015-09-08 15:02:46 -0700795 * Regardless of being on LRU, compound pages such as THP and
796 * hugetlbfs are not to be compacted. We can potentially save
797 * a lot of iterations if we skip them at once. The check is
798 * racy, but we can consider only valid values and the only
799 * danger is skipping too much.
Andrea Arcangelibc835012011-01-13 15:47:08 -0800800 */
Vlastimil Babka29c0dde2015-09-08 15:02:46 -0700801 if (PageCompound(page)) {
David Rientjes21dc7e02017-11-17 15:26:30 -0800802 const unsigned int order = compound_order(page);
Vlastimil Babka29c0dde2015-09-08 15:02:46 -0700803
David Rientjes21dc7e02017-11-17 15:26:30 -0800804 if (pageblock_skip_persistent(page, order)) {
805 set_pageblock_skip(page);
806 low_pfn = end_pfn;
807 } else if (likely(order < MAX_ORDER))
808 low_pfn += (1UL << order) - 1;
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700809 goto isolate_fail;
Mel Gorman2a1402a2012-10-08 16:32:33 -0700810 }
811
Minchan Kimbda807d2016-07-26 15:23:05 -0700812 /*
813 * Check may be lockless but that's ok as we recheck later.
814 * It's possible to migrate LRU and non-lru movable pages.
815 * Skip any other type of page
816 */
817 if (!PageLRU(page)) {
Minchan Kimbda807d2016-07-26 15:23:05 -0700818 /*
819 * __PageMovable can return false positive so we need
820 * to verify it under page_lock.
821 */
822 if (unlikely(__PageMovable(page)) &&
823 !PageIsolated(page)) {
824 if (locked) {
Mel Gormana52633d2016-07-28 15:45:28 -0700825 spin_unlock_irqrestore(zone_lru_lock(zone),
Minchan Kimbda807d2016-07-26 15:23:05 -0700826 flags);
827 locked = false;
828 }
829
Yisheng Xie9e5bcd62017-02-24 14:57:29 -0800830 if (!isolate_movable_page(page, isolate_mode))
Minchan Kimbda807d2016-07-26 15:23:05 -0700831 goto isolate_success;
832 }
833
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700834 goto isolate_fail;
Minchan Kimbda807d2016-07-26 15:23:05 -0700835 }
Vlastimil Babka29c0dde2015-09-08 15:02:46 -0700836
David Rientjes119d6d52014-04-03 14:48:00 -0700837 /*
838 * Migration will fail if an anonymous page is pinned in memory,
839 * so avoid taking lru_lock and isolating it unnecessarily in an
840 * admittedly racy check.
841 */
842 if (!page_mapping(page) &&
843 page_count(page) > page_mapcount(page))
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700844 goto isolate_fail;
David Rientjes119d6d52014-04-03 14:48:00 -0700845
Michal Hocko73e64c52016-12-14 15:04:07 -0800846 /*
847 * Only allow to migrate anonymous pages in GFP_NOFS context
848 * because those do not depend on fs locks.
849 */
850 if (!(cc->gfp_mask & __GFP_FS) && page_mapping(page))
851 goto isolate_fail;
852
Vlastimil Babka69b71892014-10-09 15:27:18 -0700853 /* If we already hold the lock, we can skip some rechecking */
854 if (!locked) {
Mel Gormana52633d2016-07-28 15:45:28 -0700855 locked = compact_trylock_irqsave(zone_lru_lock(zone),
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700856 &flags, cc);
Vlastimil Babka69b71892014-10-09 15:27:18 -0700857 if (!locked)
858 break;
Mel Gorman2a1402a2012-10-08 16:32:33 -0700859
Vlastimil Babka29c0dde2015-09-08 15:02:46 -0700860 /* Recheck PageLRU and PageCompound under lock */
Vlastimil Babka69b71892014-10-09 15:27:18 -0700861 if (!PageLRU(page))
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700862 goto isolate_fail;
Vlastimil Babka29c0dde2015-09-08 15:02:46 -0700863
864 /*
865 * Page become compound since the non-locked check,
866 * and it's on LRU. It can only be a THP so the order
867 * is safe to read and it's 0 for tail pages.
868 */
869 if (unlikely(PageCompound(page))) {
David Rientjes21dc7e02017-11-17 15:26:30 -0800870 const unsigned int order = compound_order(page);
871
872 if (pageblock_skip_persistent(page, order)) {
873 set_pageblock_skip(page);
874 low_pfn = end_pfn;
875 } else
876 low_pfn += (1UL << order) - 1;
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700877 goto isolate_fail;
Vlastimil Babka69b71892014-10-09 15:27:18 -0700878 }
Andrea Arcangelibc835012011-01-13 15:47:08 -0800879 }
880
Mel Gorman599d0c92016-07-28 15:45:31 -0700881 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700882
Mel Gorman748446b2010-05-24 14:32:27 -0700883 /* Try isolate the page */
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700884 if (__isolate_lru_page(page, isolate_mode) != 0)
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700885 goto isolate_fail;
Mel Gorman748446b2010-05-24 14:32:27 -0700886
Vlastimil Babka29c0dde2015-09-08 15:02:46 -0700887 VM_BUG_ON_PAGE(PageCompound(page), page);
Andrea Arcangelibc835012011-01-13 15:47:08 -0800888
Mel Gorman748446b2010-05-24 14:32:27 -0700889 /* Successfully isolated */
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700890 del_page_from_lru_list(page, lruvec, page_lru(page));
Ming Ling6afcf8e2016-12-12 16:42:26 -0800891 inc_node_page_state(page,
892 NR_ISOLATED_ANON + page_is_file_cache(page));
Joonsoo Kimb6c75012014-04-07 15:37:07 -0700893
894isolate_success:
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700895 list_add(&page->lru, &cc->migratepages);
Mel Gorman748446b2010-05-24 14:32:27 -0700896 cc->nr_migratepages++;
Mel Gormanb7aba692011-01-13 15:45:54 -0800897 nr_isolated++;
Mel Gorman748446b2010-05-24 14:32:27 -0700898
Vlastimil Babkaa34753d2016-05-19 17:11:51 -0700899 /*
900 * Record where we could have freed pages by migration and not
901 * yet flushed them to buddy allocator.
902 * - this is the lowest page that was isolated and likely be
903 * then freed by migration.
904 */
905 if (!cc->last_migrated_pfn)
906 cc->last_migrated_pfn = low_pfn;
907
Mel Gorman748446b2010-05-24 14:32:27 -0700908 /* Avoid isolating too much */
Hillf Danton31b83842012-01-10 15:07:59 -0800909 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
910 ++low_pfn;
Mel Gorman748446b2010-05-24 14:32:27 -0700911 break;
Hillf Danton31b83842012-01-10 15:07:59 -0800912 }
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700913
914 continue;
915isolate_fail:
916 if (!skip_on_failure)
917 continue;
918
919 /*
920 * We have isolated some pages, but then failed. Release them
921 * instead of migrating, as we cannot form the cc->order buddy
922 * page anyway.
923 */
924 if (nr_isolated) {
925 if (locked) {
Mel Gormana52633d2016-07-28 15:45:28 -0700926 spin_unlock_irqrestore(zone_lru_lock(zone), flags);
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700927 locked = false;
928 }
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700929 putback_movable_pages(&cc->migratepages);
930 cc->nr_migratepages = 0;
931 cc->last_migrated_pfn = 0;
932 nr_isolated = 0;
933 }
934
935 if (low_pfn < next_skip_pfn) {
936 low_pfn = next_skip_pfn - 1;
937 /*
938 * The check near the loop beginning would have updated
939 * next_skip_pfn too, but this is a bit simpler.
940 */
941 next_skip_pfn += 1UL << cc->order;
942 }
Mel Gorman748446b2010-05-24 14:32:27 -0700943 }
944
Vlastimil Babka99c0fd52014-10-09 15:27:23 -0700945 /*
946 * The PageBuddy() check could have potentially brought us outside
947 * the range to be scanned.
948 */
949 if (unlikely(low_pfn > end_pfn))
950 low_pfn = end_pfn;
951
Mel Gormanc67fe372012-08-21 16:16:17 -0700952 if (locked)
Mel Gormana52633d2016-07-28 15:45:28 -0700953 spin_unlock_irqrestore(zone_lru_lock(zone), flags);
Mel Gorman748446b2010-05-24 14:32:27 -0700954
Vlastimil Babka50b5b092014-01-21 15:51:10 -0800955 /*
956 * Update the pageblock-skip information and cached scanner pfn,
957 * if the whole pageblock was scanned without isolating any page.
Vlastimil Babka50b5b092014-01-21 15:51:10 -0800958 */
David Rientjes35979ef2014-06-04 16:08:27 -0700959 if (low_pfn == end_pfn)
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700960 update_pageblock_skip(cc, valid_page, nr_isolated, true);
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700961
Joonsoo Kime34d85f2015-02-11 15:27:04 -0800962 trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
963 nr_scanned, nr_isolated);
Mel Gormanb7aba692011-01-13 15:45:54 -0800964
David Rientjes7f354a52017-02-22 15:44:50 -0800965 cc->total_migrate_scanned += nr_scanned;
Mel Gorman397487d2012-10-19 12:00:10 +0100966 if (nr_isolated)
Minchan Kim010fc292012-12-20 15:05:06 -0800967 count_compact_events(COMPACTISOLATED, nr_isolated);
Mel Gorman397487d2012-10-19 12:00:10 +0100968
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100969 return low_pfn;
970}
971
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700972/**
973 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
974 * @cc: Compaction control structure.
975 * @start_pfn: The first PFN to start isolating.
976 * @end_pfn: The one-past-last PFN.
977 *
978 * Returns zero if isolation fails fatally due to e.g. pending signal.
979 * Otherwise, function returns one-past-the-last PFN of isolated page
980 * (which may be greater than end_pfn if end fell in a middle of a THP page).
981 */
982unsigned long
983isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
984 unsigned long end_pfn)
985{
Joonsoo Kime1409c32016-03-15 14:57:48 -0700986 unsigned long pfn, block_start_pfn, block_end_pfn;
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700987
988 /* Scan block by block. First and last block may be incomplete */
989 pfn = start_pfn;
Vlastimil Babka06b66402016-05-19 17:11:48 -0700990 block_start_pfn = pageblock_start_pfn(pfn);
Joonsoo Kime1409c32016-03-15 14:57:48 -0700991 if (block_start_pfn < cc->zone->zone_start_pfn)
992 block_start_pfn = cc->zone->zone_start_pfn;
Vlastimil Babka06b66402016-05-19 17:11:48 -0700993 block_end_pfn = pageblock_end_pfn(pfn);
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700994
995 for (; pfn < end_pfn; pfn = block_end_pfn,
Joonsoo Kime1409c32016-03-15 14:57:48 -0700996 block_start_pfn = block_end_pfn,
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700997 block_end_pfn += pageblock_nr_pages) {
998
999 block_end_pfn = min(block_end_pfn, end_pfn);
1000
Joonsoo Kime1409c32016-03-15 14:57:48 -07001001 if (!pageblock_pfn_to_page(block_start_pfn,
1002 block_end_pfn, cc->zone))
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001003 continue;
1004
1005 pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
1006 ISOLATE_UNEVICTABLE);
1007
Hugh Dickins14af4a52016-05-05 16:22:15 -07001008 if (!pfn)
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001009 break;
Joonsoo Kim6ea41c02014-10-29 14:50:20 -07001010
1011 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
1012 break;
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001013 }
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001014
1015 return pfn;
1016}
1017
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001018#endif /* CONFIG_COMPACTION || CONFIG_CMA */
1019#ifdef CONFIG_COMPACTION
Andrew Morton018e9a42015-04-15 16:15:20 -07001020
Vlastimil Babkab682deb2017-05-08 15:54:43 -07001021static bool suitable_migration_source(struct compact_control *cc,
1022 struct page *page)
1023{
Vlastimil Babka282722b2017-05-08 15:54:49 -07001024 int block_mt;
1025
1026 if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction)
Vlastimil Babkab682deb2017-05-08 15:54:43 -07001027 return true;
1028
Vlastimil Babka282722b2017-05-08 15:54:49 -07001029 block_mt = get_pageblock_migratetype(page);
1030
1031 if (cc->migratetype == MIGRATE_MOVABLE)
1032 return is_migrate_movable(block_mt);
1033 else
1034 return block_mt == cc->migratetype;
Vlastimil Babkab682deb2017-05-08 15:54:43 -07001035}
1036
Andrew Morton018e9a42015-04-15 16:15:20 -07001037/* Returns true if the page is within a block suitable for migration to */
Vlastimil Babka9f7e3382016-10-07 17:00:37 -07001038static bool suitable_migration_target(struct compact_control *cc,
1039 struct page *page)
Andrew Morton018e9a42015-04-15 16:15:20 -07001040{
1041 /* If the page is a large free page, then disallow migration */
1042 if (PageBuddy(page)) {
1043 /*
1044 * We are checking page_order without zone->lock taken. But
1045 * the only small danger is that we skip a potentially suitable
1046 * pageblock, so it's not worth to check order for valid range.
1047 */
1048 if (page_order_unsafe(page) >= pageblock_order)
1049 return false;
1050 }
1051
Yisheng Xie1ef36db2017-05-03 14:53:54 -07001052 if (cc->ignore_block_suitable)
1053 return true;
1054
Andrew Morton018e9a42015-04-15 16:15:20 -07001055 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
Vlastimil Babkab682deb2017-05-08 15:54:43 -07001056 if (is_migrate_movable(get_pageblock_migratetype(page)))
Andrew Morton018e9a42015-04-15 16:15:20 -07001057 return true;
1058
1059 /* Otherwise skip the block */
1060 return false;
1061}
1062
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001063/*
Vlastimil Babkaf2849aa2015-09-08 15:02:36 -07001064 * Test whether the free scanner has reached the same or lower pageblock than
1065 * the migration scanner, and compaction should thus terminate.
1066 */
1067static inline bool compact_scanners_met(struct compact_control *cc)
1068{
1069 return (cc->free_pfn >> pageblock_order)
1070 <= (cc->migrate_pfn >> pageblock_order);
1071}
1072
1073/*
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001074 * Based on information in the current compact_control, find blocks
1075 * suitable for isolating free pages from and then isolate them.
1076 */
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001077static void isolate_freepages(struct compact_control *cc)
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001078{
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001079 struct zone *zone = cc->zone;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001080 struct page *page;
Vlastimil Babkac96b9e52014-06-04 16:07:26 -07001081 unsigned long block_start_pfn; /* start of current pageblock */
Vlastimil Babkae14c7202014-10-09 15:27:20 -07001082 unsigned long isolate_start_pfn; /* exact pfn we start at */
Vlastimil Babkac96b9e52014-06-04 16:07:26 -07001083 unsigned long block_end_pfn; /* end of current pageblock */
1084 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001085 struct list_head *freelist = &cc->freepages;
1086
1087 /*
1088 * Initialise the free scanner. The starting point is where we last
Vlastimil Babka49e068f2014-05-06 12:50:03 -07001089 * successfully isolated from, zone-cached value, or the end of the
Vlastimil Babkae14c7202014-10-09 15:27:20 -07001090 * zone when isolating for the first time. For looping we also need
1091 * this pfn aligned down to the pageblock boundary, because we do
Vlastimil Babkac96b9e52014-06-04 16:07:26 -07001092 * block_start_pfn -= pageblock_nr_pages in the for loop.
1093 * For ending point, take care when isolating in last pageblock of a
1094 * a zone which ends in the middle of a pageblock.
Vlastimil Babka49e068f2014-05-06 12:50:03 -07001095 * The low boundary is the end of the pageblock the migration scanner
1096 * is using.
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001097 */
Vlastimil Babkae14c7202014-10-09 15:27:20 -07001098 isolate_start_pfn = cc->free_pfn;
Vlastimil Babka06b66402016-05-19 17:11:48 -07001099 block_start_pfn = pageblock_start_pfn(cc->free_pfn);
Vlastimil Babkac96b9e52014-06-04 16:07:26 -07001100 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1101 zone_end_pfn(zone));
Vlastimil Babka06b66402016-05-19 17:11:48 -07001102 low_pfn = pageblock_end_pfn(cc->migrate_pfn);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001103
1104 /*
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001105 * Isolate free pages until enough are available to migrate the
1106 * pages on cc->migratepages. We stop searching if the migrate
1107 * and free page scanners meet or enough free pages are isolated.
1108 */
Vlastimil Babkaf5f61a32015-09-08 15:02:39 -07001109 for (; block_start_pfn >= low_pfn;
Vlastimil Babkac96b9e52014-06-04 16:07:26 -07001110 block_end_pfn = block_start_pfn,
Vlastimil Babkae14c7202014-10-09 15:27:20 -07001111 block_start_pfn -= pageblock_nr_pages,
1112 isolate_start_pfn = block_start_pfn) {
David Rientjesf6ea3ad2013-09-30 13:45:03 -07001113 /*
1114 * This can iterate a massively long zone without finding any
1115 * suitable migration targets, so periodically check if we need
Vlastimil Babkabe976572014-06-04 16:10:41 -07001116 * to schedule, or even abort async compaction.
David Rientjesf6ea3ad2013-09-30 13:45:03 -07001117 */
Vlastimil Babkabe976572014-06-04 16:10:41 -07001118 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1119 && compact_should_abort(cc))
1120 break;
David Rientjesf6ea3ad2013-09-30 13:45:03 -07001121
Vlastimil Babka7d49d882014-10-09 15:27:11 -07001122 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1123 zone);
1124 if (!page)
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001125 continue;
1126
1127 /* Check the block is suitable for migration */
Vlastimil Babka9f7e3382016-10-07 17:00:37 -07001128 if (!suitable_migration_target(cc, page))
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001129 continue;
Linus Torvalds68e3e922012-06-03 20:05:57 -07001130
Mel Gormanbb13ffe2012-10-08 16:32:41 -07001131 /* If isolation recently failed, do not retry */
1132 if (!isolation_suitable(cc, page))
1133 continue;
1134
Vlastimil Babkae14c7202014-10-09 15:27:20 -07001135 /* Found a block suitable for isolating free pages from. */
David Rientjesa46cbf32016-07-14 12:06:50 -07001136 isolate_freepages_block(cc, &isolate_start_pfn, block_end_pfn,
1137 freelist, false);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001138
1139 /*
David Rientjesa46cbf32016-07-14 12:06:50 -07001140 * If we isolated enough freepages, or aborted due to lock
1141 * contention, terminate.
Vlastimil Babkae14c7202014-10-09 15:27:20 -07001142 */
Vlastimil Babkaf5f61a32015-09-08 15:02:39 -07001143 if ((cc->nr_freepages >= cc->nr_migratepages)
1144 || cc->contended) {
David Rientjesa46cbf32016-07-14 12:06:50 -07001145 if (isolate_start_pfn >= block_end_pfn) {
1146 /*
1147 * Restart at previous pageblock if more
1148 * freepages can be isolated next time.
1149 */
Vlastimil Babkaf5f61a32015-09-08 15:02:39 -07001150 isolate_start_pfn =
1151 block_start_pfn - pageblock_nr_pages;
David Rientjesa46cbf32016-07-14 12:06:50 -07001152 }
Vlastimil Babkabe976572014-06-04 16:10:41 -07001153 break;
David Rientjesa46cbf32016-07-14 12:06:50 -07001154 } else if (isolate_start_pfn < block_end_pfn) {
Vlastimil Babkaf5f61a32015-09-08 15:02:39 -07001155 /*
David Rientjesa46cbf32016-07-14 12:06:50 -07001156 * If isolation failed early, do not continue
1157 * needlessly.
Vlastimil Babkaf5f61a32015-09-08 15:02:39 -07001158 */
David Rientjesa46cbf32016-07-14 12:06:50 -07001159 break;
Vlastimil Babkaf5f61a32015-09-08 15:02:39 -07001160 }
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +01001161 }
1162
Joonsoo Kim66c64222016-07-26 15:23:40 -07001163 /* __isolate_free_page() does not map the pages */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001164 map_pages(freelist);
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +01001165
Vlastimil Babka7ed695e2014-01-21 15:51:09 -08001166 /*
Vlastimil Babkaf5f61a32015-09-08 15:02:39 -07001167 * Record where the free scanner will restart next time. Either we
1168 * broke from the loop and set isolate_start_pfn based on the last
1169 * call to isolate_freepages_block(), or we met the migration scanner
1170 * and the loop terminated due to isolate_start_pfn < low_pfn
Vlastimil Babka7ed695e2014-01-21 15:51:09 -08001171 */
Vlastimil Babkaf5f61a32015-09-08 15:02:39 -07001172 cc->free_pfn = isolate_start_pfn;
Mel Gorman748446b2010-05-24 14:32:27 -07001173}
1174
1175/*
1176 * This is a migrate-callback that "allocates" freepages by taking pages
1177 * from the isolated freelists in the block we are migrating to.
1178 */
1179static struct page *compaction_alloc(struct page *migratepage,
1180 unsigned long data,
1181 int **result)
1182{
1183 struct compact_control *cc = (struct compact_control *)data;
1184 struct page *freepage;
1185
Vlastimil Babkabe976572014-06-04 16:10:41 -07001186 /*
1187 * Isolate free pages if necessary, and if we are not aborting due to
1188 * contention.
1189 */
Mel Gorman748446b2010-05-24 14:32:27 -07001190 if (list_empty(&cc->freepages)) {
Vlastimil Babkabe976572014-06-04 16:10:41 -07001191 if (!cc->contended)
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001192 isolate_freepages(cc);
Mel Gorman748446b2010-05-24 14:32:27 -07001193
1194 if (list_empty(&cc->freepages))
1195 return NULL;
1196 }
1197
1198 freepage = list_entry(cc->freepages.next, struct page, lru);
1199 list_del(&freepage->lru);
1200 cc->nr_freepages--;
1201
1202 return freepage;
1203}
1204
1205/*
David Rientjesd53aea32014-06-04 16:08:26 -07001206 * This is a migrate-callback that "frees" freepages back to the isolated
1207 * freelist. All pages on the freelist are from the same zone, so there is no
1208 * special handling needed for NUMA.
1209 */
1210static void compaction_free(struct page *page, unsigned long data)
1211{
1212 struct compact_control *cc = (struct compact_control *)data;
1213
1214 list_add(&page->lru, &cc->freepages);
1215 cc->nr_freepages++;
1216}
1217
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001218/* possible outcome of isolate_migratepages */
1219typedef enum {
1220 ISOLATE_ABORT, /* Abort compaction now */
1221 ISOLATE_NONE, /* No pages isolated, continue scanning */
1222 ISOLATE_SUCCESS, /* Pages isolated, migrate */
1223} isolate_migrate_t;
1224
1225/*
Eric B Munson5bbe3542015-04-15 16:13:20 -07001226 * Allow userspace to control policy on scanning the unevictable LRU for
1227 * compactable pages.
1228 */
1229int sysctl_compact_unevictable_allowed __read_mostly = 1;
1230
1231/*
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001232 * Isolate all pages that can be migrated from the first suitable block,
1233 * starting at the block pointed to by the migrate scanner pfn within
1234 * compact_control.
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001235 */
1236static isolate_migrate_t isolate_migratepages(struct zone *zone,
1237 struct compact_control *cc)
1238{
Joonsoo Kime1409c32016-03-15 14:57:48 -07001239 unsigned long block_start_pfn;
1240 unsigned long block_end_pfn;
1241 unsigned long low_pfn;
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001242 struct page *page;
1243 const isolate_mode_t isolate_mode =
Eric B Munson5bbe3542015-04-15 16:13:20 -07001244 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
Hugh Dickins1d2047f2016-07-28 15:48:41 -07001245 (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001246
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001247 /*
1248 * Start at where we last stopped, or beginning of the zone as
1249 * initialized by compact_zone()
1250 */
1251 low_pfn = cc->migrate_pfn;
Vlastimil Babka06b66402016-05-19 17:11:48 -07001252 block_start_pfn = pageblock_start_pfn(low_pfn);
Joonsoo Kime1409c32016-03-15 14:57:48 -07001253 if (block_start_pfn < zone->zone_start_pfn)
1254 block_start_pfn = zone->zone_start_pfn;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001255
1256 /* Only scan within a pageblock boundary */
Vlastimil Babka06b66402016-05-19 17:11:48 -07001257 block_end_pfn = pageblock_end_pfn(low_pfn);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001258
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001259 /*
1260 * Iterate over whole pageblocks until we find the first suitable.
1261 * Do not cross the free scanner.
1262 */
Joonsoo Kime1409c32016-03-15 14:57:48 -07001263 for (; block_end_pfn <= cc->free_pfn;
1264 low_pfn = block_end_pfn,
1265 block_start_pfn = block_end_pfn,
1266 block_end_pfn += pageblock_nr_pages) {
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001267
1268 /*
1269 * This can potentially iterate a massively long zone with
1270 * many pageblocks unsuitable, so periodically check if we
1271 * need to schedule, or even abort async compaction.
1272 */
1273 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1274 && compact_should_abort(cc))
1275 break;
1276
Joonsoo Kime1409c32016-03-15 14:57:48 -07001277 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1278 zone);
Vlastimil Babka7d49d882014-10-09 15:27:11 -07001279 if (!page)
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001280 continue;
1281
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001282 /* If isolation recently failed, do not retry */
1283 if (!isolation_suitable(cc, page))
1284 continue;
1285
1286 /*
1287 * For async compaction, also only scan in MOVABLE blocks.
1288 * Async compaction is optimistic to see if the minimum amount
1289 * of work satisfies the allocation.
1290 */
Vlastimil Babkab682deb2017-05-08 15:54:43 -07001291 if (!suitable_migration_source(cc, page))
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001292 continue;
1293
1294 /* Perform the isolation */
Joonsoo Kime1409c32016-03-15 14:57:48 -07001295 low_pfn = isolate_migratepages_block(cc, low_pfn,
1296 block_end_pfn, isolate_mode);
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001297
Ming Ling6afcf8e2016-12-12 16:42:26 -08001298 if (!low_pfn || cc->contended)
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001299 return ISOLATE_ABORT;
1300
1301 /*
1302 * Either we isolated something and proceed with migration. Or
1303 * we failed and compact_zone should decide if we should
1304 * continue or not.
1305 */
1306 break;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001307 }
1308
Vlastimil Babkaf2849aa2015-09-08 15:02:36 -07001309 /* Record where migration scanner will be restarted. */
1310 cc->migrate_pfn = low_pfn;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001311
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001312 return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001313}
1314
Yaowei Bai21c527a2015-11-05 18:47:20 -08001315/*
1316 * order == -1 is expected when compacting via
1317 * /proc/sys/vm/compact_memory
1318 */
1319static inline bool is_via_compact_memory(int order)
1320{
1321 return order == -1;
1322}
1323
Vlastimil Babkad39773a2017-05-08 15:54:46 -07001324static enum compact_result __compact_finished(struct zone *zone,
1325 struct compact_control *cc)
Mel Gorman748446b2010-05-24 14:32:27 -07001326{
Mel Gorman8fb74b92013-01-11 14:32:16 -08001327 unsigned int order;
Vlastimil Babkad39773a2017-05-08 15:54:46 -07001328 const int migratetype = cc->migratetype;
Mel Gorman56de7262010-05-24 14:32:30 -07001329
Vlastimil Babkabe976572014-06-04 16:10:41 -07001330 if (cc->contended || fatal_signal_pending(current))
Vlastimil Babka2d1e1042015-11-05 18:48:02 -08001331 return COMPACT_CONTENDED;
Mel Gorman748446b2010-05-24 14:32:27 -07001332
Mel Gorman753341a2012-10-08 16:32:40 -07001333 /* Compaction run completes if the migrate and free scanner meet */
Vlastimil Babkaf2849aa2015-09-08 15:02:36 -07001334 if (compact_scanners_met(cc)) {
Vlastimil Babka55b7c4c2014-01-21 15:51:11 -08001335 /* Let the next compaction start anew. */
Vlastimil Babka023336412015-09-08 15:02:42 -07001336 reset_cached_positions(zone);
Vlastimil Babka55b7c4c2014-01-21 15:51:11 -08001337
Mel Gorman62997022012-10-08 16:32:47 -07001338 /*
1339 * Mark that the PG_migrate_skip information should be cleared
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07001340 * by kswapd when it goes to sleep. kcompactd does not set the
Mel Gorman62997022012-10-08 16:32:47 -07001341 * flag itself as the decision to be clear should be directly
1342 * based on an allocation request.
1343 */
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07001344 if (cc->direct_compaction)
Mel Gorman62997022012-10-08 16:32:47 -07001345 zone->compact_blockskip_flush = true;
1346
Michal Hockoc8f7de02016-05-20 16:56:47 -07001347 if (cc->whole_zone)
1348 return COMPACT_COMPLETE;
1349 else
1350 return COMPACT_PARTIAL_SKIPPED;
Mel Gormanbb13ffe2012-10-08 16:32:41 -07001351 }
Mel Gorman748446b2010-05-24 14:32:27 -07001352
Yaowei Bai21c527a2015-11-05 18:47:20 -08001353 if (is_via_compact_memory(cc->order))
Mel Gorman56de7262010-05-24 14:32:30 -07001354 return COMPACT_CONTINUE;
1355
Vlastimil Babkabaf6a9a2017-05-08 15:54:52 -07001356 if (cc->finishing_block) {
1357 /*
1358 * We have finished the pageblock, but better check again that
1359 * we really succeeded.
1360 */
1361 if (IS_ALIGNED(cc->migrate_pfn, pageblock_nr_pages))
1362 cc->finishing_block = false;
1363 else
1364 return COMPACT_CONTINUE;
1365 }
1366
Mel Gorman56de7262010-05-24 14:32:30 -07001367 /* Direct compactor: Is a suitable page free? */
Mel Gorman8fb74b92013-01-11 14:32:16 -08001368 for (order = cc->order; order < MAX_ORDER; order++) {
1369 struct free_area *area = &zone->free_area[order];
Joonsoo Kim2149cda2015-04-14 15:45:21 -07001370 bool can_steal;
Mel Gorman56de7262010-05-24 14:32:30 -07001371
Mel Gorman8fb74b92013-01-11 14:32:16 -08001372 /* Job done if page is free of the right migratetype */
David Rientjes6d7ce552014-10-09 15:27:27 -07001373 if (!list_empty(&area->free_list[migratetype]))
Vlastimil Babkacf378312016-10-07 16:57:41 -07001374 return COMPACT_SUCCESS;
Mel Gorman8fb74b92013-01-11 14:32:16 -08001375
Joonsoo Kim2149cda2015-04-14 15:45:21 -07001376#ifdef CONFIG_CMA
1377 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
1378 if (migratetype == MIGRATE_MOVABLE &&
1379 !list_empty(&area->free_list[MIGRATE_CMA]))
Vlastimil Babkacf378312016-10-07 16:57:41 -07001380 return COMPACT_SUCCESS;
Joonsoo Kim2149cda2015-04-14 15:45:21 -07001381#endif
1382 /*
1383 * Job done if allocation would steal freepages from
1384 * other migratetype buddy lists.
1385 */
1386 if (find_suitable_fallback(area, order, migratetype,
Vlastimil Babkabaf6a9a2017-05-08 15:54:52 -07001387 true, &can_steal) != -1) {
1388
1389 /* movable pages are OK in any pageblock */
1390 if (migratetype == MIGRATE_MOVABLE)
1391 return COMPACT_SUCCESS;
1392
1393 /*
1394 * We are stealing for a non-movable allocation. Make
1395 * sure we finish compacting the current pageblock
1396 * first so it is as free as possible and we won't
1397 * have to steal another one soon. This only applies
1398 * to sync compaction, as async compaction operates
1399 * on pageblocks of the same migratetype.
1400 */
1401 if (cc->mode == MIGRATE_ASYNC ||
1402 IS_ALIGNED(cc->migrate_pfn,
1403 pageblock_nr_pages)) {
1404 return COMPACT_SUCCESS;
1405 }
1406
1407 cc->finishing_block = true;
1408 return COMPACT_CONTINUE;
1409 }
Mel Gorman56de7262010-05-24 14:32:30 -07001410 }
1411
Joonsoo Kim837d0262015-02-11 15:27:06 -08001412 return COMPACT_NO_SUITABLE_PAGE;
1413}
1414
Michal Hockoea7ab982016-05-20 16:56:38 -07001415static enum compact_result compact_finished(struct zone *zone,
Vlastimil Babkad39773a2017-05-08 15:54:46 -07001416 struct compact_control *cc)
Joonsoo Kim837d0262015-02-11 15:27:06 -08001417{
1418 int ret;
1419
Vlastimil Babkad39773a2017-05-08 15:54:46 -07001420 ret = __compact_finished(zone, cc);
Joonsoo Kim837d0262015-02-11 15:27:06 -08001421 trace_mm_compaction_finished(zone, cc->order, ret);
1422 if (ret == COMPACT_NO_SUITABLE_PAGE)
1423 ret = COMPACT_CONTINUE;
1424
1425 return ret;
Mel Gorman748446b2010-05-24 14:32:27 -07001426}
1427
Mel Gorman3e7d3442011-01-13 15:45:56 -08001428/*
1429 * compaction_suitable: Is this suitable to run compaction on this zone now?
1430 * Returns
1431 * COMPACT_SKIPPED - If there are too few free pages for compaction
Vlastimil Babkacf378312016-10-07 16:57:41 -07001432 * COMPACT_SUCCESS - If the allocation would succeed without compaction
Mel Gorman3e7d3442011-01-13 15:45:56 -08001433 * COMPACT_CONTINUE - If compaction should run now
1434 */
Michal Hockoea7ab982016-05-20 16:56:38 -07001435static enum compact_result __compaction_suitable(struct zone *zone, int order,
Mel Gormanc6038442016-05-19 17:13:38 -07001436 unsigned int alloc_flags,
Michal Hocko86a294a2016-05-20 16:57:12 -07001437 int classzone_idx,
1438 unsigned long wmark_target)
Mel Gorman3e7d3442011-01-13 15:45:56 -08001439{
Mel Gorman3e7d3442011-01-13 15:45:56 -08001440 unsigned long watermark;
1441
Yaowei Bai21c527a2015-11-05 18:47:20 -08001442 if (is_via_compact_memory(order))
Michal Hocko3957c772011-06-15 15:08:25 -07001443 return COMPACT_CONTINUE;
1444
Vlastimil Babkaf2b82282016-10-07 16:57:50 -07001445 watermark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
Vlastimil Babkaebff3982014-12-10 15:43:22 -08001446 /*
1447 * If watermarks for high-order allocation are already met, there
1448 * should be no need for compaction at all.
1449 */
1450 if (zone_watermark_ok(zone, order, watermark, classzone_idx,
1451 alloc_flags))
Vlastimil Babkacf378312016-10-07 16:57:41 -07001452 return COMPACT_SUCCESS;
Vlastimil Babkaebff3982014-12-10 15:43:22 -08001453
Michal Hocko3957c772011-06-15 15:08:25 -07001454 /*
Vlastimil Babka9861a622016-10-07 16:57:53 -07001455 * Watermarks for order-0 must be met for compaction to be able to
Vlastimil Babka984fdba2016-10-07 16:57:57 -07001456 * isolate free pages for migration targets. This means that the
1457 * watermark and alloc_flags have to match, or be more pessimistic than
1458 * the check in __isolate_free_page(). We don't use the direct
1459 * compactor's alloc_flags, as they are not relevant for freepage
1460 * isolation. We however do use the direct compactor's classzone_idx to
1461 * skip over zones where lowmem reserves would prevent allocation even
1462 * if compaction succeeds.
Vlastimil Babka8348faf2016-10-07 16:58:00 -07001463 * For costly orders, we require low watermark instead of min for
1464 * compaction to proceed to increase its chances.
Vlastimil Babka984fdba2016-10-07 16:57:57 -07001465 * ALLOC_CMA is used, as pages in CMA pageblocks are considered
1466 * suitable migration targets
Mel Gorman3e7d3442011-01-13 15:45:56 -08001467 */
Vlastimil Babka8348faf2016-10-07 16:58:00 -07001468 watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
1469 low_wmark_pages(zone) : min_wmark_pages(zone);
1470 watermark += compact_gap(order);
Michal Hocko86a294a2016-05-20 16:57:12 -07001471 if (!__zone_watermark_ok(zone, 0, watermark, classzone_idx,
Vlastimil Babka984fdba2016-10-07 16:57:57 -07001472 ALLOC_CMA, wmark_target))
Mel Gorman3e7d3442011-01-13 15:45:56 -08001473 return COMPACT_SKIPPED;
1474
Vlastimil Babkacc5c9f02016-10-07 17:00:43 -07001475 return COMPACT_CONTINUE;
1476}
1477
1478enum compact_result compaction_suitable(struct zone *zone, int order,
1479 unsigned int alloc_flags,
1480 int classzone_idx)
1481{
1482 enum compact_result ret;
1483 int fragindex;
1484
1485 ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx,
1486 zone_page_state(zone, NR_FREE_PAGES));
Mel Gorman3e7d3442011-01-13 15:45:56 -08001487 /*
1488 * fragmentation index determines if allocation failures are due to
1489 * low memory or external fragmentation
1490 *
Vlastimil Babkaebff3982014-12-10 15:43:22 -08001491 * index of -1000 would imply allocations might succeed depending on
1492 * watermarks, but we already failed the high-order watermark check
Mel Gorman3e7d3442011-01-13 15:45:56 -08001493 * index towards 0 implies failure is due to lack of memory
1494 * index towards 1000 implies failure is due to fragmentation
1495 *
Vlastimil Babka20311422016-10-07 17:00:46 -07001496 * Only compact if a failure would be due to fragmentation. Also
1497 * ignore fragindex for non-costly orders where the alternative to
1498 * a successful reclaim/compaction is OOM. Fragindex and the
1499 * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
1500 * excessive compaction for costly orders, but it should not be at the
1501 * expense of system stability.
Mel Gorman3e7d3442011-01-13 15:45:56 -08001502 */
Vlastimil Babka20311422016-10-07 17:00:46 -07001503 if (ret == COMPACT_CONTINUE && (order > PAGE_ALLOC_COSTLY_ORDER)) {
Vlastimil Babkacc5c9f02016-10-07 17:00:43 -07001504 fragindex = fragmentation_index(zone, order);
1505 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
1506 ret = COMPACT_NOT_SUITABLE_ZONE;
1507 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08001508
Joonsoo Kim837d0262015-02-11 15:27:06 -08001509 trace_mm_compaction_suitable(zone, order, ret);
1510 if (ret == COMPACT_NOT_SUITABLE_ZONE)
1511 ret = COMPACT_SKIPPED;
1512
1513 return ret;
1514}
1515
Michal Hocko86a294a2016-05-20 16:57:12 -07001516bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
1517 int alloc_flags)
1518{
1519 struct zone *zone;
1520 struct zoneref *z;
1521
1522 /*
1523 * Make sure at least one zone would pass __compaction_suitable if we continue
1524 * retrying the reclaim.
1525 */
1526 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1527 ac->nodemask) {
1528 unsigned long available;
1529 enum compact_result compact_result;
1530
1531 /*
1532 * Do not consider all the reclaimable memory because we do not
1533 * want to trash just for a single high order allocation which
1534 * is even not guaranteed to appear even if __compaction_suitable
1535 * is happy about the watermark check.
1536 */
Mel Gorman5a1c84b2016-07-28 15:47:31 -07001537 available = zone_reclaimable_pages(zone) / order;
Michal Hocko86a294a2016-05-20 16:57:12 -07001538 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
1539 compact_result = __compaction_suitable(zone, order, alloc_flags,
1540 ac_classzone_idx(ac), available);
Vlastimil Babkacc5c9f02016-10-07 17:00:43 -07001541 if (compact_result != COMPACT_SKIPPED)
Michal Hocko86a294a2016-05-20 16:57:12 -07001542 return true;
1543 }
1544
1545 return false;
1546}
1547
Michal Hockoea7ab982016-05-20 16:56:38 -07001548static enum compact_result compact_zone(struct zone *zone, struct compact_control *cc)
Mel Gorman748446b2010-05-24 14:32:27 -07001549{
Michal Hockoea7ab982016-05-20 16:56:38 -07001550 enum compact_result ret;
Mel Gormanc89511a2012-10-08 16:32:45 -07001551 unsigned long start_pfn = zone->zone_start_pfn;
Cody P Schafer108bcc92013-02-22 16:35:23 -08001552 unsigned long end_pfn = zone_end_pfn(zone);
David Rientjese0b9dae2014-06-04 16:08:28 -07001553 const bool sync = cc->mode != MIGRATE_ASYNC;
Mel Gorman748446b2010-05-24 14:32:27 -07001554
Vlastimil Babkad39773a2017-05-08 15:54:46 -07001555 cc->migratetype = gfpflags_to_migratetype(cc->gfp_mask);
Vlastimil Babkaebff3982014-12-10 15:43:22 -08001556 ret = compaction_suitable(zone, cc->order, cc->alloc_flags,
1557 cc->classzone_idx);
Michal Hockoc46649d2016-05-20 16:56:41 -07001558 /* Compaction is likely to fail */
Vlastimil Babkacf378312016-10-07 16:57:41 -07001559 if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED)
Mel Gorman3e7d3442011-01-13 15:45:56 -08001560 return ret;
Michal Hockoc46649d2016-05-20 16:56:41 -07001561
1562 /* huh, compaction_suitable is returning something unexpected */
1563 VM_BUG_ON(ret != COMPACT_CONTINUE);
Mel Gorman3e7d3442011-01-13 15:45:56 -08001564
Mel Gormanc89511a2012-10-08 16:32:45 -07001565 /*
Vlastimil Babkad3132e42014-01-21 15:51:08 -08001566 * Clear pageblock skip if there were failures recently and compaction
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07001567 * is about to be retried after being deferred.
Vlastimil Babkad3132e42014-01-21 15:51:08 -08001568 */
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07001569 if (compaction_restarting(zone, cc->order))
Vlastimil Babkad3132e42014-01-21 15:51:08 -08001570 __reset_isolation_suitable(zone);
1571
1572 /*
Mel Gormanc89511a2012-10-08 16:32:45 -07001573 * Setup to move all movable pages to the end of the zone. Used cached
Vlastimil Babka06ed2992016-10-07 16:57:35 -07001574 * information on where the scanners should start (unless we explicitly
1575 * want to compact the whole zone), but check that it is initialised
1576 * by ensuring the values are within zone boundaries.
Mel Gormanc89511a2012-10-08 16:32:45 -07001577 */
Vlastimil Babka06ed2992016-10-07 16:57:35 -07001578 if (cc->whole_zone) {
Mel Gormanc89511a2012-10-08 16:32:45 -07001579 cc->migrate_pfn = start_pfn;
Vlastimil Babka06ed2992016-10-07 16:57:35 -07001580 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
1581 } else {
1582 cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync];
1583 cc->free_pfn = zone->compact_cached_free_pfn;
1584 if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
1585 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
1586 zone->compact_cached_free_pfn = cc->free_pfn;
1587 }
1588 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
1589 cc->migrate_pfn = start_pfn;
1590 zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
1591 zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
1592 }
Michal Hockoc8f7de02016-05-20 16:56:47 -07001593
Vlastimil Babka06ed2992016-10-07 16:57:35 -07001594 if (cc->migrate_pfn == start_pfn)
1595 cc->whole_zone = true;
1596 }
Michal Hockoc8f7de02016-05-20 16:56:47 -07001597
Joonsoo Kim1a167182015-09-08 15:03:59 -07001598 cc->last_migrated_pfn = 0;
Mel Gorman748446b2010-05-24 14:32:27 -07001599
Joonsoo Kim16c4a092015-02-11 15:27:01 -08001600 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
1601 cc->free_pfn, end_pfn, sync);
Mel Gorman0eb927c2014-01-21 15:51:05 -08001602
Mel Gorman748446b2010-05-24 14:32:27 -07001603 migrate_prep_local();
1604
Vlastimil Babkad39773a2017-05-08 15:54:46 -07001605 while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
Minchan Kim9d502c12011-03-22 16:30:39 -07001606 int err;
Mel Gorman748446b2010-05-24 14:32:27 -07001607
Mel Gormanf9e35b32011-06-15 15:08:52 -07001608 switch (isolate_migratepages(zone, cc)) {
1609 case ISOLATE_ABORT:
Vlastimil Babka2d1e1042015-11-05 18:48:02 -08001610 ret = COMPACT_CONTENDED;
Rafael Aquini5733c7d2012-12-11 16:02:47 -08001611 putback_movable_pages(&cc->migratepages);
Shaohua Lie64c5232012-10-08 16:32:27 -07001612 cc->nr_migratepages = 0;
Mel Gormanf9e35b32011-06-15 15:08:52 -07001613 goto out;
1614 case ISOLATE_NONE:
Vlastimil Babkafdaf7f52014-12-10 15:43:34 -08001615 /*
1616 * We haven't isolated and migrated anything, but
1617 * there might still be unflushed migrations from
1618 * previous cc->order aligned block.
1619 */
1620 goto check_drain;
Mel Gormanf9e35b32011-06-15 15:08:52 -07001621 case ISOLATE_SUCCESS:
1622 ;
1623 }
Mel Gorman748446b2010-05-24 14:32:27 -07001624
David Rientjesd53aea32014-06-04 16:08:26 -07001625 err = migrate_pages(&cc->migratepages, compaction_alloc,
David Rientjese0b9dae2014-06-04 16:08:28 -07001626 compaction_free, (unsigned long)cc, cc->mode,
Mel Gorman7b2a2d42012-10-19 14:07:31 +01001627 MR_COMPACTION);
Mel Gorman748446b2010-05-24 14:32:27 -07001628
Vlastimil Babkaf8c93012014-06-04 16:08:32 -07001629 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
1630 &cc->migratepages);
Mel Gorman748446b2010-05-24 14:32:27 -07001631
Vlastimil Babkaf8c93012014-06-04 16:08:32 -07001632 /* All pages were either migrated or will be released */
1633 cc->nr_migratepages = 0;
Minchan Kim9d502c12011-03-22 16:30:39 -07001634 if (err) {
Rafael Aquini5733c7d2012-12-11 16:02:47 -08001635 putback_movable_pages(&cc->migratepages);
Vlastimil Babka7ed695e2014-01-21 15:51:09 -08001636 /*
1637 * migrate_pages() may return -ENOMEM when scanners meet
1638 * and we want compact_finished() to detect it
1639 */
Vlastimil Babkaf2849aa2015-09-08 15:02:36 -07001640 if (err == -ENOMEM && !compact_scanners_met(cc)) {
Vlastimil Babka2d1e1042015-11-05 18:48:02 -08001641 ret = COMPACT_CONTENDED;
David Rientjes4bf2bba2012-07-11 14:02:13 -07001642 goto out;
1643 }
Vlastimil Babkafdd048e2016-05-19 17:11:55 -07001644 /*
1645 * We failed to migrate at least one page in the current
1646 * order-aligned block, so skip the rest of it.
1647 */
1648 if (cc->direct_compaction &&
1649 (cc->mode == MIGRATE_ASYNC)) {
1650 cc->migrate_pfn = block_end_pfn(
1651 cc->migrate_pfn - 1, cc->order);
1652 /* Draining pcplists is useless in this case */
1653 cc->last_migrated_pfn = 0;
1654
1655 }
Mel Gorman748446b2010-05-24 14:32:27 -07001656 }
Vlastimil Babkafdaf7f52014-12-10 15:43:34 -08001657
Vlastimil Babkafdaf7f52014-12-10 15:43:34 -08001658check_drain:
1659 /*
1660 * Has the migration scanner moved away from the previous
1661 * cc->order aligned block where we migrated from? If yes,
1662 * flush the pages that were freed, so that they can merge and
1663 * compact_finished() can detect immediately if allocation
1664 * would succeed.
1665 */
Joonsoo Kim1a167182015-09-08 15:03:59 -07001666 if (cc->order > 0 && cc->last_migrated_pfn) {
Vlastimil Babkafdaf7f52014-12-10 15:43:34 -08001667 int cpu;
1668 unsigned long current_block_start =
Vlastimil Babka06b66402016-05-19 17:11:48 -07001669 block_start_pfn(cc->migrate_pfn, cc->order);
Vlastimil Babkafdaf7f52014-12-10 15:43:34 -08001670
Joonsoo Kim1a167182015-09-08 15:03:59 -07001671 if (cc->last_migrated_pfn < current_block_start) {
Vlastimil Babkafdaf7f52014-12-10 15:43:34 -08001672 cpu = get_cpu();
1673 lru_add_drain_cpu(cpu);
1674 drain_local_pages(zone);
1675 put_cpu();
1676 /* No more flushing until we migrate again */
Joonsoo Kim1a167182015-09-08 15:03:59 -07001677 cc->last_migrated_pfn = 0;
Vlastimil Babkafdaf7f52014-12-10 15:43:34 -08001678 }
1679 }
1680
Mel Gorman748446b2010-05-24 14:32:27 -07001681 }
1682
Mel Gormanf9e35b32011-06-15 15:08:52 -07001683out:
Vlastimil Babka6bace092014-12-10 15:43:31 -08001684 /*
1685 * Release free pages and update where the free scanner should restart,
1686 * so we don't leave any returned pages behind in the next attempt.
1687 */
1688 if (cc->nr_freepages > 0) {
1689 unsigned long free_pfn = release_freepages(&cc->freepages);
1690
1691 cc->nr_freepages = 0;
1692 VM_BUG_ON(free_pfn == 0);
1693 /* The cached pfn is always the first in a pageblock */
Vlastimil Babka06b66402016-05-19 17:11:48 -07001694 free_pfn = pageblock_start_pfn(free_pfn);
Vlastimil Babka6bace092014-12-10 15:43:31 -08001695 /*
1696 * Only go back, not forward. The cached pfn might have been
1697 * already reset to zone end in compact_finished()
1698 */
1699 if (free_pfn > zone->compact_cached_free_pfn)
1700 zone->compact_cached_free_pfn = free_pfn;
1701 }
Mel Gorman748446b2010-05-24 14:32:27 -07001702
David Rientjes7f354a52017-02-22 15:44:50 -08001703 count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned);
1704 count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned);
1705
Joonsoo Kim16c4a092015-02-11 15:27:01 -08001706 trace_mm_compaction_end(start_pfn, cc->migrate_pfn,
1707 cc->free_pfn, end_pfn, sync, ret);
Mel Gorman0eb927c2014-01-21 15:51:05 -08001708
Mel Gorman748446b2010-05-24 14:32:27 -07001709 return ret;
1710}
Mel Gorman76ab0f52010-05-24 14:32:28 -07001711
Michal Hockoea7ab982016-05-20 16:56:38 -07001712static enum compact_result compact_zone_order(struct zone *zone, int order,
Vlastimil Babkac3486f52016-07-28 15:49:30 -07001713 gfp_t gfp_mask, enum compact_priority prio,
Mel Gormanc6038442016-05-19 17:13:38 -07001714 unsigned int alloc_flags, int classzone_idx)
Mel Gorman56de7262010-05-24 14:32:30 -07001715{
Michal Hockoea7ab982016-05-20 16:56:38 -07001716 enum compact_result ret;
Mel Gorman56de7262010-05-24 14:32:30 -07001717 struct compact_control cc = {
1718 .nr_freepages = 0,
1719 .nr_migratepages = 0,
David Rientjes7f354a52017-02-22 15:44:50 -08001720 .total_migrate_scanned = 0,
1721 .total_free_scanned = 0,
Mel Gorman56de7262010-05-24 14:32:30 -07001722 .order = order,
David Rientjes6d7ce552014-10-09 15:27:27 -07001723 .gfp_mask = gfp_mask,
Mel Gorman56de7262010-05-24 14:32:30 -07001724 .zone = zone,
Vlastimil Babkaa5508cd2016-07-28 15:49:28 -07001725 .mode = (prio == COMPACT_PRIO_ASYNC) ?
1726 MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT,
Vlastimil Babkaebff3982014-12-10 15:43:22 -08001727 .alloc_flags = alloc_flags,
1728 .classzone_idx = classzone_idx,
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07001729 .direct_compaction = true,
Vlastimil Babkaa8e025e2016-10-07 16:57:47 -07001730 .whole_zone = (prio == MIN_COMPACT_PRIORITY),
Vlastimil Babka9f7e3382016-10-07 17:00:37 -07001731 .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
1732 .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
Mel Gorman56de7262010-05-24 14:32:30 -07001733 };
1734 INIT_LIST_HEAD(&cc.freepages);
1735 INIT_LIST_HEAD(&cc.migratepages);
1736
Shaohua Lie64c5232012-10-08 16:32:27 -07001737 ret = compact_zone(zone, &cc);
1738
1739 VM_BUG_ON(!list_empty(&cc.freepages));
1740 VM_BUG_ON(!list_empty(&cc.migratepages));
1741
Shaohua Lie64c5232012-10-08 16:32:27 -07001742 return ret;
Mel Gorman56de7262010-05-24 14:32:30 -07001743}
1744
Mel Gorman5e771902010-05-24 14:32:31 -07001745int sysctl_extfrag_threshold = 500;
1746
Mel Gorman56de7262010-05-24 14:32:30 -07001747/**
1748 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
Mel Gorman56de7262010-05-24 14:32:30 -07001749 * @gfp_mask: The GFP mask of the current allocation
Vlastimil Babka1a6d53a2015-02-11 15:25:44 -08001750 * @order: The order of the current allocation
1751 * @alloc_flags: The allocation flags of the current allocation
1752 * @ac: The context of current allocation
David Rientjese0b9dae2014-06-04 16:08:28 -07001753 * @mode: The migration mode for async, sync light, or sync migration
Mel Gorman56de7262010-05-24 14:32:30 -07001754 *
1755 * This is the main entry point for direct page compaction.
1756 */
Michal Hockoea7ab982016-05-20 16:56:38 -07001757enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
Mel Gormanc6038442016-05-19 17:13:38 -07001758 unsigned int alloc_flags, const struct alloc_context *ac,
Vlastimil Babkac3486f52016-07-28 15:49:30 -07001759 enum compact_priority prio)
Mel Gorman56de7262010-05-24 14:32:30 -07001760{
Mel Gorman56de7262010-05-24 14:32:30 -07001761 int may_perform_io = gfp_mask & __GFP_IO;
Mel Gorman56de7262010-05-24 14:32:30 -07001762 struct zoneref *z;
1763 struct zone *zone;
Michal Hocko1d4746d2016-05-20 16:56:44 -07001764 enum compact_result rc = COMPACT_SKIPPED;
Mel Gorman56de7262010-05-24 14:32:30 -07001765
Michal Hocko73e64c52016-12-14 15:04:07 -08001766 /*
1767 * Check if the GFP flags allow compaction - GFP_NOIO is really
1768 * tricky context because the migration might require IO
1769 */
1770 if (!may_perform_io)
Vlastimil Babka53853e22014-10-09 15:27:02 -07001771 return COMPACT_SKIPPED;
Mel Gorman56de7262010-05-24 14:32:30 -07001772
Vlastimil Babkaa5508cd2016-07-28 15:49:28 -07001773 trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
Joonsoo Kim837d0262015-02-11 15:27:06 -08001774
Mel Gorman56de7262010-05-24 14:32:30 -07001775 /* Compact each zone in the list */
Vlastimil Babka1a6d53a2015-02-11 15:25:44 -08001776 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1777 ac->nodemask) {
Michal Hockoea7ab982016-05-20 16:56:38 -07001778 enum compact_result status;
Mel Gorman56de7262010-05-24 14:32:30 -07001779
Vlastimil Babkaa8e025e2016-10-07 16:57:47 -07001780 if (prio > MIN_COMPACT_PRIORITY
1781 && compaction_deferred(zone, order)) {
Michal Hocko1d4746d2016-05-20 16:56:44 -07001782 rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
Vlastimil Babka53853e22014-10-09 15:27:02 -07001783 continue;
Michal Hocko1d4746d2016-05-20 16:56:44 -07001784 }
Vlastimil Babka53853e22014-10-09 15:27:02 -07001785
Vlastimil Babkaa5508cd2016-07-28 15:49:28 -07001786 status = compact_zone_order(zone, order, gfp_mask, prio,
Vlastimil Babkac3486f52016-07-28 15:49:30 -07001787 alloc_flags, ac_classzone_idx(ac));
Mel Gorman56de7262010-05-24 14:32:30 -07001788 rc = max(status, rc);
1789
Vlastimil Babka7ceb0092016-10-07 16:57:44 -07001790 /* The allocation should succeed, stop compacting */
1791 if (status == COMPACT_SUCCESS) {
Vlastimil Babka53853e22014-10-09 15:27:02 -07001792 /*
1793 * We think the allocation will succeed in this zone,
1794 * but it is not certain, hence the false. The caller
1795 * will repeat this with true if allocation indeed
1796 * succeeds in this zone.
1797 */
1798 compaction_defer_reset(zone, order, false);
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001799
Vlastimil Babkac3486f52016-07-28 15:49:30 -07001800 break;
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001801 }
1802
Vlastimil Babkaa5508cd2016-07-28 15:49:28 -07001803 if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
Vlastimil Babkac3486f52016-07-28 15:49:30 -07001804 status == COMPACT_PARTIAL_SKIPPED))
Vlastimil Babka53853e22014-10-09 15:27:02 -07001805 /*
1806 * We think that allocation won't succeed in this zone
1807 * so we defer compaction there. If it ends up
1808 * succeeding after all, it will be reset.
1809 */
1810 defer_compaction(zone, order);
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001811
1812 /*
1813 * We might have stopped compacting due to need_resched() in
1814 * async compaction, or due to a fatal signal detected. In that
Vlastimil Babkac3486f52016-07-28 15:49:30 -07001815 * case do not try further zones
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001816 */
Vlastimil Babkac3486f52016-07-28 15:49:30 -07001817 if ((prio == COMPACT_PRIO_ASYNC && need_resched())
1818 || fatal_signal_pending(current))
1819 break;
Mel Gorman56de7262010-05-24 14:32:30 -07001820 }
1821
1822 return rc;
1823}
1824
1825
Mel Gorman76ab0f52010-05-24 14:32:28 -07001826/* Compact all zones within a node */
Andrew Morton7103f162013-02-22 16:32:33 -08001827static void compact_node(int nid)
Rik van Riel7be62de2012-03-21 16:33:52 -07001828{
Vlastimil Babka791cae92016-10-07 16:57:38 -07001829 pg_data_t *pgdat = NODE_DATA(nid);
1830 int zoneid;
1831 struct zone *zone;
Rik van Riel7be62de2012-03-21 16:33:52 -07001832 struct compact_control cc = {
1833 .order = -1,
David Rientjes7f354a52017-02-22 15:44:50 -08001834 .total_migrate_scanned = 0,
1835 .total_free_scanned = 0,
David Rientjese0b9dae2014-06-04 16:08:28 -07001836 .mode = MIGRATE_SYNC,
David Rientjes91ca9182014-04-03 14:47:23 -07001837 .ignore_skip_hint = true,
Vlastimil Babka06ed2992016-10-07 16:57:35 -07001838 .whole_zone = true,
Michal Hocko73e64c52016-12-14 15:04:07 -08001839 .gfp_mask = GFP_KERNEL,
Rik van Riel7be62de2012-03-21 16:33:52 -07001840 };
1841
Vlastimil Babka791cae92016-10-07 16:57:38 -07001842
1843 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1844
1845 zone = &pgdat->node_zones[zoneid];
1846 if (!populated_zone(zone))
1847 continue;
1848
1849 cc.nr_freepages = 0;
1850 cc.nr_migratepages = 0;
1851 cc.zone = zone;
1852 INIT_LIST_HEAD(&cc.freepages);
1853 INIT_LIST_HEAD(&cc.migratepages);
1854
1855 compact_zone(zone, &cc);
1856
1857 VM_BUG_ON(!list_empty(&cc.freepages));
1858 VM_BUG_ON(!list_empty(&cc.migratepages));
1859 }
Rik van Riel7be62de2012-03-21 16:33:52 -07001860}
1861
Mel Gorman76ab0f52010-05-24 14:32:28 -07001862/* Compact all nodes in the system */
Jason Liu7964c062013-01-11 14:31:47 -08001863static void compact_nodes(void)
Mel Gorman76ab0f52010-05-24 14:32:28 -07001864{
1865 int nid;
1866
Hugh Dickins8575ec22012-03-21 16:33:53 -07001867 /* Flush pending updates to the LRU lists */
1868 lru_add_drain_all();
1869
Mel Gorman76ab0f52010-05-24 14:32:28 -07001870 for_each_online_node(nid)
1871 compact_node(nid);
Mel Gorman76ab0f52010-05-24 14:32:28 -07001872}
1873
1874/* The written value is actually unused, all memory is compacted */
1875int sysctl_compact_memory;
1876
Yaowei Baifec4eb22016-01-14 15:20:09 -08001877/*
1878 * This is the entry point for compacting all nodes via
1879 * /proc/sys/vm/compact_memory
1880 */
Mel Gorman76ab0f52010-05-24 14:32:28 -07001881int sysctl_compaction_handler(struct ctl_table *table, int write,
1882 void __user *buffer, size_t *length, loff_t *ppos)
1883{
1884 if (write)
Jason Liu7964c062013-01-11 14:31:47 -08001885 compact_nodes();
Mel Gorman76ab0f52010-05-24 14:32:28 -07001886
1887 return 0;
1888}
Mel Gormaned4a6d72010-05-24 14:32:29 -07001889
Mel Gorman5e771902010-05-24 14:32:31 -07001890int sysctl_extfrag_handler(struct ctl_table *table, int write,
1891 void __user *buffer, size_t *length, loff_t *ppos)
1892{
1893 proc_dointvec_minmax(table, write, buffer, length, ppos);
1894
1895 return 0;
1896}
1897
Mel Gormaned4a6d72010-05-24 14:32:29 -07001898#if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
Rashika Kheria74e77fb2014-04-03 14:48:01 -07001899static ssize_t sysfs_compact_node(struct device *dev,
Kay Sievers10fbcf42011-12-21 14:48:43 -08001900 struct device_attribute *attr,
Mel Gormaned4a6d72010-05-24 14:32:29 -07001901 const char *buf, size_t count)
1902{
Hugh Dickins8575ec22012-03-21 16:33:53 -07001903 int nid = dev->id;
1904
1905 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1906 /* Flush pending updates to the LRU lists */
1907 lru_add_drain_all();
1908
1909 compact_node(nid);
1910 }
Mel Gormaned4a6d72010-05-24 14:32:29 -07001911
1912 return count;
1913}
Kay Sievers10fbcf42011-12-21 14:48:43 -08001914static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
Mel Gormaned4a6d72010-05-24 14:32:29 -07001915
1916int compaction_register_node(struct node *node)
1917{
Kay Sievers10fbcf42011-12-21 14:48:43 -08001918 return device_create_file(&node->dev, &dev_attr_compact);
Mel Gormaned4a6d72010-05-24 14:32:29 -07001919}
1920
1921void compaction_unregister_node(struct node *node)
1922{
Kay Sievers10fbcf42011-12-21 14:48:43 -08001923 return device_remove_file(&node->dev, &dev_attr_compact);
Mel Gormaned4a6d72010-05-24 14:32:29 -07001924}
1925#endif /* CONFIG_SYSFS && CONFIG_NUMA */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001926
Vlastimil Babka698b1b32016-03-17 14:18:08 -07001927static inline bool kcompactd_work_requested(pg_data_t *pgdat)
1928{
Vlastimil Babka172400c2016-05-05 16:22:32 -07001929 return pgdat->kcompactd_max_order > 0 || kthread_should_stop();
Vlastimil Babka698b1b32016-03-17 14:18:08 -07001930}
1931
1932static bool kcompactd_node_suitable(pg_data_t *pgdat)
1933{
1934 int zoneid;
1935 struct zone *zone;
1936 enum zone_type classzone_idx = pgdat->kcompactd_classzone_idx;
1937
Chen Feng6cd9dc32016-05-20 16:59:02 -07001938 for (zoneid = 0; zoneid <= classzone_idx; zoneid++) {
Vlastimil Babka698b1b32016-03-17 14:18:08 -07001939 zone = &pgdat->node_zones[zoneid];
1940
1941 if (!populated_zone(zone))
1942 continue;
1943
1944 if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0,
1945 classzone_idx) == COMPACT_CONTINUE)
1946 return true;
1947 }
1948
1949 return false;
1950}
1951
1952static void kcompactd_do_work(pg_data_t *pgdat)
1953{
1954 /*
1955 * With no special task, compact all zones so that a page of requested
1956 * order is allocatable.
1957 */
1958 int zoneid;
1959 struct zone *zone;
1960 struct compact_control cc = {
1961 .order = pgdat->kcompactd_max_order,
David Rientjes7f354a52017-02-22 15:44:50 -08001962 .total_migrate_scanned = 0,
1963 .total_free_scanned = 0,
Vlastimil Babka698b1b32016-03-17 14:18:08 -07001964 .classzone_idx = pgdat->kcompactd_classzone_idx,
1965 .mode = MIGRATE_SYNC_LIGHT,
David Rientjesa0647dc2017-11-17 15:26:27 -08001966 .ignore_skip_hint = false,
Michal Hocko73e64c52016-12-14 15:04:07 -08001967 .gfp_mask = GFP_KERNEL,
Vlastimil Babka698b1b32016-03-17 14:18:08 -07001968 };
Vlastimil Babka698b1b32016-03-17 14:18:08 -07001969 trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
1970 cc.classzone_idx);
David Rientjes7f354a52017-02-22 15:44:50 -08001971 count_compact_event(KCOMPACTD_WAKE);
Vlastimil Babka698b1b32016-03-17 14:18:08 -07001972
Chen Feng6cd9dc32016-05-20 16:59:02 -07001973 for (zoneid = 0; zoneid <= cc.classzone_idx; zoneid++) {
Vlastimil Babka698b1b32016-03-17 14:18:08 -07001974 int status;
1975
1976 zone = &pgdat->node_zones[zoneid];
1977 if (!populated_zone(zone))
1978 continue;
1979
1980 if (compaction_deferred(zone, cc.order))
1981 continue;
1982
1983 if (compaction_suitable(zone, cc.order, 0, zoneid) !=
1984 COMPACT_CONTINUE)
1985 continue;
1986
1987 cc.nr_freepages = 0;
1988 cc.nr_migratepages = 0;
David Rientjes7f354a52017-02-22 15:44:50 -08001989 cc.total_migrate_scanned = 0;
1990 cc.total_free_scanned = 0;
Vlastimil Babka698b1b32016-03-17 14:18:08 -07001991 cc.zone = zone;
1992 INIT_LIST_HEAD(&cc.freepages);
1993 INIT_LIST_HEAD(&cc.migratepages);
1994
Vlastimil Babka172400c2016-05-05 16:22:32 -07001995 if (kthread_should_stop())
1996 return;
Vlastimil Babka698b1b32016-03-17 14:18:08 -07001997 status = compact_zone(zone, &cc);
1998
Vlastimil Babka7ceb0092016-10-07 16:57:44 -07001999 if (status == COMPACT_SUCCESS) {
Vlastimil Babka698b1b32016-03-17 14:18:08 -07002000 compaction_defer_reset(zone, cc.order, false);
Michal Hockoc8f7de02016-05-20 16:56:47 -07002001 } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
Vlastimil Babka698b1b32016-03-17 14:18:08 -07002002 /*
2003 * We use sync migration mode here, so we defer like
2004 * sync direct compaction does.
2005 */
2006 defer_compaction(zone, cc.order);
2007 }
2008
David Rientjes7f354a52017-02-22 15:44:50 -08002009 count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
2010 cc.total_migrate_scanned);
2011 count_compact_events(KCOMPACTD_FREE_SCANNED,
2012 cc.total_free_scanned);
2013
Vlastimil Babka698b1b32016-03-17 14:18:08 -07002014 VM_BUG_ON(!list_empty(&cc.freepages));
2015 VM_BUG_ON(!list_empty(&cc.migratepages));
2016 }
2017
2018 /*
2019 * Regardless of success, we are done until woken up next. But remember
2020 * the requested order/classzone_idx in case it was higher/tighter than
2021 * our current ones
2022 */
2023 if (pgdat->kcompactd_max_order <= cc.order)
2024 pgdat->kcompactd_max_order = 0;
2025 if (pgdat->kcompactd_classzone_idx >= cc.classzone_idx)
2026 pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
2027}
2028
2029void wakeup_kcompactd(pg_data_t *pgdat, int order, int classzone_idx)
2030{
2031 if (!order)
2032 return;
2033
2034 if (pgdat->kcompactd_max_order < order)
2035 pgdat->kcompactd_max_order = order;
2036
2037 if (pgdat->kcompactd_classzone_idx > classzone_idx)
2038 pgdat->kcompactd_classzone_idx = classzone_idx;
2039
Davidlohr Bueso68186002017-10-03 16:15:03 -07002040 /*
2041 * Pairs with implicit barrier in wait_event_freezable()
2042 * such that wakeups are not missed.
2043 */
2044 if (!wq_has_sleeper(&pgdat->kcompactd_wait))
Vlastimil Babka698b1b32016-03-17 14:18:08 -07002045 return;
2046
2047 if (!kcompactd_node_suitable(pgdat))
2048 return;
2049
2050 trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
2051 classzone_idx);
2052 wake_up_interruptible(&pgdat->kcompactd_wait);
2053}
2054
2055/*
2056 * The background compaction daemon, started as a kernel thread
2057 * from the init process.
2058 */
2059static int kcompactd(void *p)
2060{
2061 pg_data_t *pgdat = (pg_data_t*)p;
2062 struct task_struct *tsk = current;
2063
2064 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2065
2066 if (!cpumask_empty(cpumask))
2067 set_cpus_allowed_ptr(tsk, cpumask);
2068
2069 set_freezable();
2070
2071 pgdat->kcompactd_max_order = 0;
2072 pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
2073
2074 while (!kthread_should_stop()) {
2075 trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
2076 wait_event_freezable(pgdat->kcompactd_wait,
2077 kcompactd_work_requested(pgdat));
2078
2079 kcompactd_do_work(pgdat);
2080 }
2081
2082 return 0;
2083}
2084
2085/*
2086 * This kcompactd start function will be called by init and node-hot-add.
2087 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
2088 */
2089int kcompactd_run(int nid)
2090{
2091 pg_data_t *pgdat = NODE_DATA(nid);
2092 int ret = 0;
2093
2094 if (pgdat->kcompactd)
2095 return 0;
2096
2097 pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
2098 if (IS_ERR(pgdat->kcompactd)) {
2099 pr_err("Failed to start kcompactd on node %d\n", nid);
2100 ret = PTR_ERR(pgdat->kcompactd);
2101 pgdat->kcompactd = NULL;
2102 }
2103 return ret;
2104}
2105
2106/*
2107 * Called by memory hotplug when all memory in a node is offlined. Caller must
2108 * hold mem_hotplug_begin/end().
2109 */
2110void kcompactd_stop(int nid)
2111{
2112 struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
2113
2114 if (kcompactd) {
2115 kthread_stop(kcompactd);
2116 NODE_DATA(nid)->kcompactd = NULL;
2117 }
2118}
2119
2120/*
2121 * It's optimal to keep kcompactd on the same CPUs as their memory, but
2122 * not required for correctness. So if the last cpu in a node goes
2123 * away, we get changed to run anywhere: as the first one comes back,
2124 * restore their cpu bindings.
2125 */
Anna-Maria Gleixnere46b1db2016-11-27 00:13:42 +01002126static int kcompactd_cpu_online(unsigned int cpu)
Vlastimil Babka698b1b32016-03-17 14:18:08 -07002127{
2128 int nid;
2129
Anna-Maria Gleixnere46b1db2016-11-27 00:13:42 +01002130 for_each_node_state(nid, N_MEMORY) {
2131 pg_data_t *pgdat = NODE_DATA(nid);
2132 const struct cpumask *mask;
Vlastimil Babka698b1b32016-03-17 14:18:08 -07002133
Anna-Maria Gleixnere46b1db2016-11-27 00:13:42 +01002134 mask = cpumask_of_node(pgdat->node_id);
Vlastimil Babka698b1b32016-03-17 14:18:08 -07002135
Anna-Maria Gleixnere46b1db2016-11-27 00:13:42 +01002136 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2137 /* One of our CPUs online: restore mask */
2138 set_cpus_allowed_ptr(pgdat->kcompactd, mask);
Vlastimil Babka698b1b32016-03-17 14:18:08 -07002139 }
Anna-Maria Gleixnere46b1db2016-11-27 00:13:42 +01002140 return 0;
Vlastimil Babka698b1b32016-03-17 14:18:08 -07002141}
2142
2143static int __init kcompactd_init(void)
2144{
2145 int nid;
Anna-Maria Gleixnere46b1db2016-11-27 00:13:42 +01002146 int ret;
2147
2148 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
2149 "mm/compaction:online",
2150 kcompactd_cpu_online, NULL);
2151 if (ret < 0) {
2152 pr_err("kcompactd: failed to register hotplug callbacks.\n");
2153 return ret;
2154 }
Vlastimil Babka698b1b32016-03-17 14:18:08 -07002155
2156 for_each_node_state(nid, N_MEMORY)
2157 kcompactd_run(nid);
Vlastimil Babka698b1b32016-03-17 14:18:08 -07002158 return 0;
2159}
2160subsys_initcall(kcompactd_init)
2161
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01002162#endif /* CONFIG_COMPACTION */