blob: 64df5fe052db656dcfd38999abdf561942dfd4c5 [file] [log] [blame]
Mel Gorman748446b2010-05-24 14:32:27 -07001/*
2 * linux/mm/compaction.c
3 *
4 * Memory compaction for the reduction of external fragmentation. Note that
5 * this heavily depends upon page migration to do all the real heavy
6 * lifting
7 *
8 * Copyright IBM Corp. 2007-2010 Mel Gorman <mel@csn.ul.ie>
9 */
Vlastimil Babka698b1b32016-03-17 14:18:08 -070010#include <linux/cpu.h>
Mel Gorman748446b2010-05-24 14:32:27 -070011#include <linux/swap.h>
12#include <linux/migrate.h>
13#include <linux/compaction.h>
14#include <linux/mm_inline.h>
15#include <linux/backing-dev.h>
Mel Gorman76ab0f52010-05-24 14:32:28 -070016#include <linux/sysctl.h>
Mel Gormaned4a6d72010-05-24 14:32:29 -070017#include <linux/sysfs.h>
Minchan Kim194159f2013-02-22 16:33:58 -080018#include <linux/page-isolation.h>
Andrey Ryabininb8c73fc2015-02-13 14:39:28 -080019#include <linux/kasan.h>
Vlastimil Babka698b1b32016-03-17 14:18:08 -070020#include <linux/kthread.h>
21#include <linux/freezer.h>
Joonsoo Kim83358ec2016-07-26 15:23:43 -070022#include <linux/page_owner.h>
Mel Gorman748446b2010-05-24 14:32:27 -070023#include "internal.h"
24
Minchan Kim010fc292012-12-20 15:05:06 -080025#ifdef CONFIG_COMPACTION
26static inline void count_compact_event(enum vm_event_item item)
27{
28 count_vm_event(item);
29}
30
31static inline void count_compact_events(enum vm_event_item item, long delta)
32{
33 count_vm_events(item, delta);
34}
35#else
36#define count_compact_event(item) do { } while (0)
37#define count_compact_events(item, delta) do { } while (0)
38#endif
39
Michal Nazarewiczff9543f2011-12-29 13:09:50 +010040#if defined CONFIG_COMPACTION || defined CONFIG_CMA
41
Mel Gormanb7aba692011-01-13 15:45:54 -080042#define CREATE_TRACE_POINTS
43#include <trace/events/compaction.h>
44
Vlastimil Babka06b66402016-05-19 17:11:48 -070045#define block_start_pfn(pfn, order) round_down(pfn, 1UL << (order))
46#define block_end_pfn(pfn, order) ALIGN((pfn) + 1, 1UL << (order))
47#define pageblock_start_pfn(pfn) block_start_pfn(pfn, pageblock_order)
48#define pageblock_end_pfn(pfn) block_end_pfn(pfn, pageblock_order)
49
Mel Gorman748446b2010-05-24 14:32:27 -070050static unsigned long release_freepages(struct list_head *freelist)
51{
52 struct page *page, *next;
Vlastimil Babka6bace092014-12-10 15:43:31 -080053 unsigned long high_pfn = 0;
Mel Gorman748446b2010-05-24 14:32:27 -070054
55 list_for_each_entry_safe(page, next, freelist, lru) {
Vlastimil Babka6bace092014-12-10 15:43:31 -080056 unsigned long pfn = page_to_pfn(page);
Mel Gorman748446b2010-05-24 14:32:27 -070057 list_del(&page->lru);
58 __free_page(page);
Vlastimil Babka6bace092014-12-10 15:43:31 -080059 if (pfn > high_pfn)
60 high_pfn = pfn;
Mel Gorman748446b2010-05-24 14:32:27 -070061 }
62
Vlastimil Babka6bace092014-12-10 15:43:31 -080063 return high_pfn;
Mel Gorman748446b2010-05-24 14:32:27 -070064}
65
Michal Nazarewiczff9543f2011-12-29 13:09:50 +010066static void map_pages(struct list_head *list)
67{
Joonsoo Kim66c64222016-07-26 15:23:40 -070068 unsigned int i, order, nr_pages;
69 struct page *page, *next;
70 LIST_HEAD(tmp_list);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +010071
Joonsoo Kim66c64222016-07-26 15:23:40 -070072 list_for_each_entry_safe(page, next, list, lru) {
73 list_del(&page->lru);
74
75 order = page_private(page);
76 nr_pages = 1 << order;
Joonsoo Kim66c64222016-07-26 15:23:40 -070077
Joonsoo Kim46f24fd2016-07-26 15:23:58 -070078 post_alloc_hook(page, order, __GFP_MOVABLE);
Joonsoo Kim66c64222016-07-26 15:23:40 -070079 if (order)
80 split_page(page, order);
81
82 for (i = 0; i < nr_pages; i++) {
83 list_add(&page->lru, &tmp_list);
84 page++;
85 }
Michal Nazarewiczff9543f2011-12-29 13:09:50 +010086 }
Joonsoo Kim66c64222016-07-26 15:23:40 -070087
88 list_splice(&tmp_list, list);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +010089}
90
Michal Nazarewicz47118af2011-12-29 13:09:50 +010091static inline bool migrate_async_suitable(int migratetype)
92{
93 return is_migrate_cma(migratetype) || migratetype == MIGRATE_MOVABLE;
94}
95
Mel Gormanbb13ffe2012-10-08 16:32:41 -070096#ifdef CONFIG_COMPACTION
Joonsoo Kim24e27162015-02-11 15:27:09 -080097
Minchan Kimbda807d2016-07-26 15:23:05 -070098int PageMovable(struct page *page)
99{
100 struct address_space *mapping;
101
102 VM_BUG_ON_PAGE(!PageLocked(page), page);
103 if (!__PageMovable(page))
104 return 0;
105
106 mapping = page_mapping(page);
107 if (mapping && mapping->a_ops && mapping->a_ops->isolate_page)
108 return 1;
109
110 return 0;
111}
112EXPORT_SYMBOL(PageMovable);
113
114void __SetPageMovable(struct page *page, struct address_space *mapping)
115{
116 VM_BUG_ON_PAGE(!PageLocked(page), page);
117 VM_BUG_ON_PAGE((unsigned long)mapping & PAGE_MAPPING_MOVABLE, page);
118 page->mapping = (void *)((unsigned long)mapping | PAGE_MAPPING_MOVABLE);
119}
120EXPORT_SYMBOL(__SetPageMovable);
121
122void __ClearPageMovable(struct page *page)
123{
124 VM_BUG_ON_PAGE(!PageLocked(page), page);
125 VM_BUG_ON_PAGE(!PageMovable(page), page);
126 /*
127 * Clear registered address_space val with keeping PAGE_MAPPING_MOVABLE
128 * flag so that VM can catch up released page by driver after isolation.
129 * With it, VM migration doesn't try to put it back.
130 */
131 page->mapping = (void *)((unsigned long)page->mapping &
132 PAGE_MAPPING_MOVABLE);
133}
134EXPORT_SYMBOL(__ClearPageMovable);
135
Joonsoo Kim24e27162015-02-11 15:27:09 -0800136/* Do not skip compaction more than 64 times */
137#define COMPACT_MAX_DEFER_SHIFT 6
138
139/*
140 * Compaction is deferred when compaction fails to result in a page
141 * allocation success. 1 << compact_defer_limit compactions are skipped up
142 * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
143 */
144void defer_compaction(struct zone *zone, int order)
145{
146 zone->compact_considered = 0;
147 zone->compact_defer_shift++;
148
149 if (order < zone->compact_order_failed)
150 zone->compact_order_failed = order;
151
152 if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
153 zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
154
155 trace_mm_compaction_defer_compaction(zone, order);
156}
157
158/* Returns true if compaction should be skipped this time */
159bool compaction_deferred(struct zone *zone, int order)
160{
161 unsigned long defer_limit = 1UL << zone->compact_defer_shift;
162
163 if (order < zone->compact_order_failed)
164 return false;
165
166 /* Avoid possible overflow */
167 if (++zone->compact_considered > defer_limit)
168 zone->compact_considered = defer_limit;
169
170 if (zone->compact_considered >= defer_limit)
171 return false;
172
173 trace_mm_compaction_deferred(zone, order);
174
175 return true;
176}
177
178/*
179 * Update defer tracking counters after successful compaction of given order,
180 * which means an allocation either succeeded (alloc_success == true) or is
181 * expected to succeed.
182 */
183void compaction_defer_reset(struct zone *zone, int order,
184 bool alloc_success)
185{
186 if (alloc_success) {
187 zone->compact_considered = 0;
188 zone->compact_defer_shift = 0;
189 }
190 if (order >= zone->compact_order_failed)
191 zone->compact_order_failed = order + 1;
192
193 trace_mm_compaction_defer_reset(zone, order);
194}
195
196/* Returns true if restarting compaction after many failures */
197bool compaction_restarting(struct zone *zone, int order)
198{
199 if (order < zone->compact_order_failed)
200 return false;
201
202 return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
203 zone->compact_considered >= 1UL << zone->compact_defer_shift;
204}
205
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700206/* Returns true if the pageblock should be scanned for pages to isolate. */
207static inline bool isolation_suitable(struct compact_control *cc,
208 struct page *page)
209{
210 if (cc->ignore_skip_hint)
211 return true;
212
213 return !get_pageblock_skip(page);
214}
215
Vlastimil Babka02333642015-09-08 15:02:42 -0700216static void reset_cached_positions(struct zone *zone)
217{
218 zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
219 zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
Joonsoo Kim623446e2016-03-15 14:57:45 -0700220 zone->compact_cached_free_pfn =
Vlastimil Babka06b66402016-05-19 17:11:48 -0700221 pageblock_start_pfn(zone_end_pfn(zone) - 1);
Vlastimil Babka02333642015-09-08 15:02:42 -0700222}
223
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700224/*
225 * This function is called to clear all cached information on pageblocks that
226 * should be skipped for page isolation when the migrate and free page scanner
227 * meet.
228 */
Mel Gorman62997022012-10-08 16:32:47 -0700229static void __reset_isolation_suitable(struct zone *zone)
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700230{
231 unsigned long start_pfn = zone->zone_start_pfn;
Cody P Schafer108bcc92013-02-22 16:35:23 -0800232 unsigned long end_pfn = zone_end_pfn(zone);
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700233 unsigned long pfn;
234
Mel Gorman62997022012-10-08 16:32:47 -0700235 zone->compact_blockskip_flush = false;
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700236
237 /* Walk the zone and mark every pageblock as suitable for isolation */
238 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
239 struct page *page;
240
241 cond_resched();
242
243 if (!pfn_valid(pfn))
244 continue;
245
246 page = pfn_to_page(pfn);
247 if (zone != page_zone(page))
248 continue;
249
250 clear_pageblock_skip(page);
251 }
Vlastimil Babka02333642015-09-08 15:02:42 -0700252
253 reset_cached_positions(zone);
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700254}
255
Mel Gorman62997022012-10-08 16:32:47 -0700256void reset_isolation_suitable(pg_data_t *pgdat)
257{
258 int zoneid;
259
260 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
261 struct zone *zone = &pgdat->node_zones[zoneid];
262 if (!populated_zone(zone))
263 continue;
264
265 /* Only flush if a full compaction finished recently */
266 if (zone->compact_blockskip_flush)
267 __reset_isolation_suitable(zone);
268 }
269}
270
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700271/*
272 * If no pages were isolated then mark this pageblock to be skipped in the
Mel Gorman62997022012-10-08 16:32:47 -0700273 * future. The information is later cleared by __reset_isolation_suitable().
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700274 */
Mel Gormanc89511a2012-10-08 16:32:45 -0700275static void update_pageblock_skip(struct compact_control *cc,
276 struct page *page, unsigned long nr_isolated,
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700277 bool migrate_scanner)
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700278{
Mel Gormanc89511a2012-10-08 16:32:45 -0700279 struct zone *zone = cc->zone;
David Rientjes35979ef2014-06-04 16:08:27 -0700280 unsigned long pfn;
Joonsoo Kim6815bf32013-12-18 17:08:52 -0800281
282 if (cc->ignore_skip_hint)
283 return;
284
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700285 if (!page)
286 return;
287
David Rientjes35979ef2014-06-04 16:08:27 -0700288 if (nr_isolated)
289 return;
290
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700291 set_pageblock_skip(page);
Mel Gormanc89511a2012-10-08 16:32:45 -0700292
David Rientjes35979ef2014-06-04 16:08:27 -0700293 pfn = page_to_pfn(page);
294
295 /* Update where async and sync compaction should restart */
296 if (migrate_scanner) {
David Rientjes35979ef2014-06-04 16:08:27 -0700297 if (pfn > zone->compact_cached_migrate_pfn[0])
298 zone->compact_cached_migrate_pfn[0] = pfn;
David Rientjese0b9dae2014-06-04 16:08:28 -0700299 if (cc->mode != MIGRATE_ASYNC &&
300 pfn > zone->compact_cached_migrate_pfn[1])
David Rientjes35979ef2014-06-04 16:08:27 -0700301 zone->compact_cached_migrate_pfn[1] = pfn;
302 } else {
David Rientjes35979ef2014-06-04 16:08:27 -0700303 if (pfn < zone->compact_cached_free_pfn)
304 zone->compact_cached_free_pfn = pfn;
Mel Gormanc89511a2012-10-08 16:32:45 -0700305 }
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700306}
307#else
308static inline bool isolation_suitable(struct compact_control *cc,
309 struct page *page)
310{
311 return true;
312}
313
Mel Gormanc89511a2012-10-08 16:32:45 -0700314static void update_pageblock_skip(struct compact_control *cc,
315 struct page *page, unsigned long nr_isolated,
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700316 bool migrate_scanner)
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700317{
318}
319#endif /* CONFIG_COMPACTION */
320
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700321/*
322 * Compaction requires the taking of some coarse locks that are potentially
323 * very heavily contended. For async compaction, back out if the lock cannot
324 * be taken immediately. For sync compaction, spin on the lock if needed.
325 *
326 * Returns true if the lock is held
327 * Returns false if the lock is not held and compaction should abort
328 */
329static bool compact_trylock_irqsave(spinlock_t *lock, unsigned long *flags,
330 struct compact_control *cc)
Mel Gorman2a1402a2012-10-08 16:32:33 -0700331{
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700332 if (cc->mode == MIGRATE_ASYNC) {
333 if (!spin_trylock_irqsave(lock, *flags)) {
334 cc->contended = COMPACT_CONTENDED_LOCK;
335 return false;
336 }
337 } else {
338 spin_lock_irqsave(lock, *flags);
339 }
Vlastimil Babka1f9efde2014-10-09 15:27:14 -0700340
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700341 return true;
Mel Gorman2a1402a2012-10-08 16:32:33 -0700342}
343
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100344/*
Mel Gormanc67fe372012-08-21 16:16:17 -0700345 * Compaction requires the taking of some coarse locks that are potentially
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700346 * very heavily contended. The lock should be periodically unlocked to avoid
347 * having disabled IRQs for a long time, even when there is nobody waiting on
348 * the lock. It might also be that allowing the IRQs will result in
349 * need_resched() becoming true. If scheduling is needed, async compaction
350 * aborts. Sync compaction schedules.
351 * Either compaction type will also abort if a fatal signal is pending.
352 * In either case if the lock was locked, it is dropped and not regained.
Mel Gormanc67fe372012-08-21 16:16:17 -0700353 *
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700354 * Returns true if compaction should abort due to fatal signal pending, or
355 * async compaction due to need_resched()
356 * Returns false when compaction can continue (sync compaction might have
357 * scheduled)
Mel Gormanc67fe372012-08-21 16:16:17 -0700358 */
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700359static bool compact_unlock_should_abort(spinlock_t *lock,
360 unsigned long flags, bool *locked, struct compact_control *cc)
Mel Gormanc67fe372012-08-21 16:16:17 -0700361{
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700362 if (*locked) {
363 spin_unlock_irqrestore(lock, flags);
364 *locked = false;
365 }
Vlastimil Babka1f9efde2014-10-09 15:27:14 -0700366
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700367 if (fatal_signal_pending(current)) {
368 cc->contended = COMPACT_CONTENDED_SCHED;
369 return true;
370 }
Mel Gormanc67fe372012-08-21 16:16:17 -0700371
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700372 if (need_resched()) {
David Rientjese0b9dae2014-06-04 16:08:28 -0700373 if (cc->mode == MIGRATE_ASYNC) {
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700374 cc->contended = COMPACT_CONTENDED_SCHED;
375 return true;
Mel Gormanc67fe372012-08-21 16:16:17 -0700376 }
Mel Gormanc67fe372012-08-21 16:16:17 -0700377 cond_resched();
Mel Gormanc67fe372012-08-21 16:16:17 -0700378 }
379
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700380 return false;
Mel Gormanc67fe372012-08-21 16:16:17 -0700381}
382
Vlastimil Babkabe976572014-06-04 16:10:41 -0700383/*
384 * Aside from avoiding lock contention, compaction also periodically checks
385 * need_resched() and either schedules in sync compaction or aborts async
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700386 * compaction. This is similar to what compact_unlock_should_abort() does, but
Vlastimil Babkabe976572014-06-04 16:10:41 -0700387 * is used where no lock is concerned.
388 *
389 * Returns false when no scheduling was needed, or sync compaction scheduled.
390 * Returns true when async compaction should abort.
391 */
392static inline bool compact_should_abort(struct compact_control *cc)
393{
394 /* async compaction aborts if contended */
395 if (need_resched()) {
396 if (cc->mode == MIGRATE_ASYNC) {
Vlastimil Babka1f9efde2014-10-09 15:27:14 -0700397 cc->contended = COMPACT_CONTENDED_SCHED;
Vlastimil Babkabe976572014-06-04 16:10:41 -0700398 return true;
399 }
400
401 cond_resched();
402 }
403
404 return false;
405}
406
Mel Gormanc67fe372012-08-21 16:16:17 -0700407/*
Jerome Marchand9e4be472013-11-12 15:07:12 -0800408 * Isolate free pages onto a private freelist. If @strict is true, will abort
409 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
410 * (even though it may still end up isolating some pages).
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100411 */
Mel Gormanf40d1e42012-10-08 16:32:36 -0700412static unsigned long isolate_freepages_block(struct compact_control *cc,
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700413 unsigned long *start_pfn,
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100414 unsigned long end_pfn,
415 struct list_head *freelist,
416 bool strict)
Mel Gorman748446b2010-05-24 14:32:27 -0700417{
Mel Gormanb7aba692011-01-13 15:45:54 -0800418 int nr_scanned = 0, total_isolated = 0;
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700419 struct page *cursor, *valid_page = NULL;
Xiubo Lib8b2d822014-10-09 15:28:21 -0700420 unsigned long flags = 0;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700421 bool locked = false;
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700422 unsigned long blockpfn = *start_pfn;
Joonsoo Kim66c64222016-07-26 15:23:40 -0700423 unsigned int order;
Mel Gorman748446b2010-05-24 14:32:27 -0700424
Mel Gorman748446b2010-05-24 14:32:27 -0700425 cursor = pfn_to_page(blockpfn);
426
Mel Gormanf40d1e42012-10-08 16:32:36 -0700427 /* Isolate free pages. */
Mel Gorman748446b2010-05-24 14:32:27 -0700428 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
Joonsoo Kim66c64222016-07-26 15:23:40 -0700429 int isolated;
Mel Gorman748446b2010-05-24 14:32:27 -0700430 struct page *page = cursor;
431
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700432 /*
433 * Periodically drop the lock (if held) regardless of its
434 * contention, to give chance to IRQs. Abort if fatal signal
435 * pending or async compaction detects need_resched()
436 */
437 if (!(blockpfn % SWAP_CLUSTER_MAX)
438 && compact_unlock_should_abort(&cc->zone->lock, flags,
439 &locked, cc))
440 break;
441
Mel Gormanb7aba692011-01-13 15:45:54 -0800442 nr_scanned++;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700443 if (!pfn_valid_within(blockpfn))
Laura Abbott2af120b2014-03-10 15:49:44 -0700444 goto isolate_fail;
445
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700446 if (!valid_page)
447 valid_page = page;
Vlastimil Babka9fcd6d22015-09-08 15:02:49 -0700448
449 /*
450 * For compound pages such as THP and hugetlbfs, we can save
451 * potentially a lot of iterations if we skip them at once.
452 * The check is racy, but we can consider only valid values
453 * and the only danger is skipping too much.
454 */
455 if (PageCompound(page)) {
456 unsigned int comp_order = compound_order(page);
457
458 if (likely(comp_order < MAX_ORDER)) {
459 blockpfn += (1UL << comp_order) - 1;
460 cursor += (1UL << comp_order) - 1;
461 }
462
463 goto isolate_fail;
464 }
465
Mel Gormanf40d1e42012-10-08 16:32:36 -0700466 if (!PageBuddy(page))
Laura Abbott2af120b2014-03-10 15:49:44 -0700467 goto isolate_fail;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700468
469 /*
Vlastimil Babka69b71892014-10-09 15:27:18 -0700470 * If we already hold the lock, we can skip some rechecking.
471 * Note that if we hold the lock now, checked_pageblock was
472 * already set in some previous iteration (or strict is true),
473 * so it is correct to skip the suitable migration target
474 * recheck as well.
Mel Gormanf40d1e42012-10-08 16:32:36 -0700475 */
Vlastimil Babka69b71892014-10-09 15:27:18 -0700476 if (!locked) {
477 /*
478 * The zone lock must be held to isolate freepages.
479 * Unfortunately this is a very coarse lock and can be
480 * heavily contended if there are parallel allocations
481 * or parallel compactions. For async compaction do not
482 * spin on the lock and we acquire the lock as late as
483 * possible.
484 */
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700485 locked = compact_trylock_irqsave(&cc->zone->lock,
486 &flags, cc);
Vlastimil Babka69b71892014-10-09 15:27:18 -0700487 if (!locked)
488 break;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700489
Vlastimil Babka69b71892014-10-09 15:27:18 -0700490 /* Recheck this is a buddy page under lock */
491 if (!PageBuddy(page))
492 goto isolate_fail;
493 }
Mel Gorman748446b2010-05-24 14:32:27 -0700494
Joonsoo Kim66c64222016-07-26 15:23:40 -0700495 /* Found a free page, will break it into order-0 pages */
496 order = page_order(page);
497 isolated = __isolate_free_page(page, order);
David Rientjesa4f04f22016-06-24 14:50:10 -0700498 if (!isolated)
499 break;
Joonsoo Kim66c64222016-07-26 15:23:40 -0700500 set_page_private(page, order);
David Rientjesa4f04f22016-06-24 14:50:10 -0700501
Mel Gorman748446b2010-05-24 14:32:27 -0700502 total_isolated += isolated;
David Rientjesa4f04f22016-06-24 14:50:10 -0700503 cc->nr_freepages += isolated;
Joonsoo Kim66c64222016-07-26 15:23:40 -0700504 list_add_tail(&page->lru, freelist);
505
David Rientjesa4f04f22016-06-24 14:50:10 -0700506 if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
507 blockpfn += isolated;
508 break;
Mel Gorman748446b2010-05-24 14:32:27 -0700509 }
David Rientjesa4f04f22016-06-24 14:50:10 -0700510 /* Advance to the end of split page */
511 blockpfn += isolated - 1;
512 cursor += isolated - 1;
513 continue;
Laura Abbott2af120b2014-03-10 15:49:44 -0700514
515isolate_fail:
516 if (strict)
517 break;
518 else
519 continue;
520
Mel Gorman748446b2010-05-24 14:32:27 -0700521 }
522
David Rientjesa4f04f22016-06-24 14:50:10 -0700523 if (locked)
524 spin_unlock_irqrestore(&cc->zone->lock, flags);
525
Vlastimil Babka9fcd6d22015-09-08 15:02:49 -0700526 /*
527 * There is a tiny chance that we have read bogus compound_order(),
528 * so be careful to not go outside of the pageblock.
529 */
530 if (unlikely(blockpfn > end_pfn))
531 blockpfn = end_pfn;
532
Joonsoo Kime34d85f2015-02-11 15:27:04 -0800533 trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
534 nr_scanned, total_isolated);
535
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700536 /* Record how far we have got within the block */
537 *start_pfn = blockpfn;
538
Mel Gormanf40d1e42012-10-08 16:32:36 -0700539 /*
540 * If strict isolation is requested by CMA then check that all the
541 * pages requested were isolated. If there were any failures, 0 is
542 * returned and CMA will fail.
543 */
Laura Abbott2af120b2014-03-10 15:49:44 -0700544 if (strict && blockpfn < end_pfn)
Mel Gormanf40d1e42012-10-08 16:32:36 -0700545 total_isolated = 0;
546
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700547 /* Update the pageblock-skip if the whole pageblock was scanned */
548 if (blockpfn == end_pfn)
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700549 update_pageblock_skip(cc, valid_page, total_isolated, false);
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700550
Minchan Kim010fc292012-12-20 15:05:06 -0800551 count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
Mel Gorman397487d2012-10-19 12:00:10 +0100552 if (total_isolated)
Minchan Kim010fc292012-12-20 15:05:06 -0800553 count_compact_events(COMPACTISOLATED, total_isolated);
Mel Gorman748446b2010-05-24 14:32:27 -0700554 return total_isolated;
555}
556
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100557/**
558 * isolate_freepages_range() - isolate free pages.
559 * @start_pfn: The first PFN to start isolating.
560 * @end_pfn: The one-past-last PFN.
561 *
562 * Non-free pages, invalid PFNs, or zone boundaries within the
563 * [start_pfn, end_pfn) range are considered errors, cause function to
564 * undo its actions and return zero.
565 *
566 * Otherwise, function returns one-past-the-last PFN of isolated page
567 * (which may be greater then end_pfn if end fell in a middle of
568 * a free page).
569 */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100570unsigned long
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700571isolate_freepages_range(struct compact_control *cc,
572 unsigned long start_pfn, unsigned long end_pfn)
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100573{
Joonsoo Kime1409c32016-03-15 14:57:48 -0700574 unsigned long isolated, pfn, block_start_pfn, block_end_pfn;
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100575 LIST_HEAD(freelist);
576
Vlastimil Babka7d49d882014-10-09 15:27:11 -0700577 pfn = start_pfn;
Vlastimil Babka06b66402016-05-19 17:11:48 -0700578 block_start_pfn = pageblock_start_pfn(pfn);
Joonsoo Kime1409c32016-03-15 14:57:48 -0700579 if (block_start_pfn < cc->zone->zone_start_pfn)
580 block_start_pfn = cc->zone->zone_start_pfn;
Vlastimil Babka06b66402016-05-19 17:11:48 -0700581 block_end_pfn = pageblock_end_pfn(pfn);
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100582
Vlastimil Babka7d49d882014-10-09 15:27:11 -0700583 for (; pfn < end_pfn; pfn += isolated,
Joonsoo Kime1409c32016-03-15 14:57:48 -0700584 block_start_pfn = block_end_pfn,
Vlastimil Babka7d49d882014-10-09 15:27:11 -0700585 block_end_pfn += pageblock_nr_pages) {
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700586 /* Protect pfn from changing by isolate_freepages_block */
587 unsigned long isolate_start_pfn = pfn;
Vlastimil Babka7d49d882014-10-09 15:27:11 -0700588
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100589 block_end_pfn = min(block_end_pfn, end_pfn);
590
Joonsoo Kim58420012014-11-13 15:19:07 -0800591 /*
592 * pfn could pass the block_end_pfn if isolated freepage
593 * is more than pageblock order. In this case, we adjust
594 * scanning range to right one.
595 */
596 if (pfn >= block_end_pfn) {
Vlastimil Babka06b66402016-05-19 17:11:48 -0700597 block_start_pfn = pageblock_start_pfn(pfn);
598 block_end_pfn = pageblock_end_pfn(pfn);
Joonsoo Kim58420012014-11-13 15:19:07 -0800599 block_end_pfn = min(block_end_pfn, end_pfn);
600 }
601
Joonsoo Kime1409c32016-03-15 14:57:48 -0700602 if (!pageblock_pfn_to_page(block_start_pfn,
603 block_end_pfn, cc->zone))
Vlastimil Babka7d49d882014-10-09 15:27:11 -0700604 break;
605
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700606 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
607 block_end_pfn, &freelist, true);
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100608
609 /*
610 * In strict mode, isolate_freepages_block() returns 0 if
611 * there are any holes in the block (ie. invalid PFNs or
612 * non-free pages).
613 */
614 if (!isolated)
615 break;
616
617 /*
618 * If we managed to isolate pages, it is always (1 << n) *
619 * pageblock_nr_pages for some non-negative n. (Max order
620 * page may span two pageblocks).
621 */
622 }
623
Joonsoo Kim66c64222016-07-26 15:23:40 -0700624 /* __isolate_free_page() does not map the pages */
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100625 map_pages(&freelist);
626
627 if (pfn < end_pfn) {
628 /* Loop terminated early, cleanup. */
629 release_freepages(&freelist);
630 return 0;
631 }
632
633 /* We don't use freelists for anything. */
634 return pfn;
635}
636
Mel Gorman748446b2010-05-24 14:32:27 -0700637/* Update the number of anon and file isolated pages in the zone */
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700638static void acct_isolated(struct zone *zone, struct compact_control *cc)
Mel Gorman748446b2010-05-24 14:32:27 -0700639{
640 struct page *page;
Minchan Kimb9e84ac2011-10-31 17:06:44 -0700641 unsigned int count[2] = { 0, };
Mel Gorman748446b2010-05-24 14:32:27 -0700642
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700643 if (list_empty(&cc->migratepages))
644 return;
645
Minchan Kimb9e84ac2011-10-31 17:06:44 -0700646 list_for_each_entry(page, &cc->migratepages, lru)
647 count[!!page_is_file_cache(page)]++;
Mel Gorman748446b2010-05-24 14:32:27 -0700648
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700649 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
650 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
Mel Gorman748446b2010-05-24 14:32:27 -0700651}
652
653/* Similar to reclaim, but different enough that they don't share logic */
654static bool too_many_isolated(struct zone *zone)
655{
Minchan Kimbc693042010-09-09 16:38:00 -0700656 unsigned long active, inactive, isolated;
Mel Gorman748446b2010-05-24 14:32:27 -0700657
658 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
659 zone_page_state(zone, NR_INACTIVE_ANON);
Minchan Kimbc693042010-09-09 16:38:00 -0700660 active = zone_page_state(zone, NR_ACTIVE_FILE) +
661 zone_page_state(zone, NR_ACTIVE_ANON);
Mel Gorman748446b2010-05-24 14:32:27 -0700662 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
663 zone_page_state(zone, NR_ISOLATED_ANON);
664
Minchan Kimbc693042010-09-09 16:38:00 -0700665 return isolated > (inactive + active) / 2;
Mel Gorman748446b2010-05-24 14:32:27 -0700666}
667
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100668/**
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700669 * isolate_migratepages_block() - isolate all migrate-able pages within
670 * a single pageblock
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100671 * @cc: Compaction control structure.
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700672 * @low_pfn: The first PFN to isolate
673 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
674 * @isolate_mode: Isolation mode to be used.
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100675 *
676 * Isolate all pages that can be migrated from the range specified by
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700677 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
678 * Returns zero if there is a fatal signal pending, otherwise PFN of the
679 * first page that was not scanned (which may be both less, equal to or more
680 * than end_pfn).
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100681 *
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700682 * The pages are isolated on cc->migratepages list (not required to be empty),
683 * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
684 * is neither read nor updated.
Mel Gorman748446b2010-05-24 14:32:27 -0700685 */
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700686static unsigned long
687isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
688 unsigned long end_pfn, isolate_mode_t isolate_mode)
Mel Gorman748446b2010-05-24 14:32:27 -0700689{
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700690 struct zone *zone = cc->zone;
Mel Gormanb7aba692011-01-13 15:45:54 -0800691 unsigned long nr_scanned = 0, nr_isolated = 0;
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700692 struct lruvec *lruvec;
Xiubo Lib8b2d822014-10-09 15:28:21 -0700693 unsigned long flags = 0;
Mel Gorman2a1402a2012-10-08 16:32:33 -0700694 bool locked = false;
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700695 struct page *page = NULL, *valid_page = NULL;
Joonsoo Kime34d85f2015-02-11 15:27:04 -0800696 unsigned long start_pfn = low_pfn;
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700697 bool skip_on_failure = false;
698 unsigned long next_skip_pfn = 0;
Mel Gorman748446b2010-05-24 14:32:27 -0700699
Mel Gorman748446b2010-05-24 14:32:27 -0700700 /*
701 * Ensure that there are not too many pages isolated from the LRU
702 * list by either parallel reclaimers or compaction. If there are,
703 * delay for some time until fewer pages are isolated
704 */
705 while (unlikely(too_many_isolated(zone))) {
Mel Gormanf9e35b32011-06-15 15:08:52 -0700706 /* async migration should just abort */
David Rientjese0b9dae2014-06-04 16:08:28 -0700707 if (cc->mode == MIGRATE_ASYNC)
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100708 return 0;
Mel Gormanf9e35b32011-06-15 15:08:52 -0700709
Mel Gorman748446b2010-05-24 14:32:27 -0700710 congestion_wait(BLK_RW_ASYNC, HZ/10);
711
712 if (fatal_signal_pending(current))
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100713 return 0;
Mel Gorman748446b2010-05-24 14:32:27 -0700714 }
715
Vlastimil Babkabe976572014-06-04 16:10:41 -0700716 if (compact_should_abort(cc))
717 return 0;
David Rientjesaeef4b82014-06-04 16:08:31 -0700718
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700719 if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
720 skip_on_failure = true;
721 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
722 }
723
Mel Gorman748446b2010-05-24 14:32:27 -0700724 /* Time to isolate some pages for migration */
Mel Gorman748446b2010-05-24 14:32:27 -0700725 for (; low_pfn < end_pfn; low_pfn++) {
Vlastimil Babka29c0dde2015-09-08 15:02:46 -0700726
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700727 if (skip_on_failure && low_pfn >= next_skip_pfn) {
728 /*
729 * We have isolated all migration candidates in the
730 * previous order-aligned block, and did not skip it due
731 * to failure. We should migrate the pages now and
732 * hopefully succeed compaction.
733 */
734 if (nr_isolated)
735 break;
736
737 /*
738 * We failed to isolate in the previous order-aligned
739 * block. Set the new boundary to the end of the
740 * current block. Note we can't simply increase
741 * next_skip_pfn by 1 << order, as low_pfn might have
742 * been incremented by a higher number due to skipping
743 * a compound or a high-order buddy page in the
744 * previous loop iteration.
745 */
746 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
747 }
748
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700749 /*
750 * Periodically drop the lock (if held) regardless of its
751 * contention, to give chance to IRQs. Abort async compaction
752 * if contended.
753 */
754 if (!(low_pfn % SWAP_CLUSTER_MAX)
755 && compact_unlock_should_abort(&zone->lru_lock, flags,
756 &locked, cc))
757 break;
Mel Gormanc67fe372012-08-21 16:16:17 -0700758
Mel Gorman748446b2010-05-24 14:32:27 -0700759 if (!pfn_valid_within(low_pfn))
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700760 goto isolate_fail;
Mel Gormanb7aba692011-01-13 15:45:54 -0800761 nr_scanned++;
Mel Gorman748446b2010-05-24 14:32:27 -0700762
Mel Gorman748446b2010-05-24 14:32:27 -0700763 page = pfn_to_page(low_pfn);
Mel Gormandc908602012-02-08 17:13:38 -0800764
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700765 if (!valid_page)
766 valid_page = page;
767
Mel Gorman6c144662014-01-23 15:53:38 -0800768 /*
Vlastimil Babka99c0fd52014-10-09 15:27:23 -0700769 * Skip if free. We read page order here without zone lock
770 * which is generally unsafe, but the race window is small and
771 * the worst thing that can happen is that we skip some
772 * potential isolation targets.
Mel Gorman6c144662014-01-23 15:53:38 -0800773 */
Vlastimil Babka99c0fd52014-10-09 15:27:23 -0700774 if (PageBuddy(page)) {
775 unsigned long freepage_order = page_order_unsafe(page);
776
777 /*
778 * Without lock, we cannot be sure that what we got is
779 * a valid page order. Consider only values in the
780 * valid order range to prevent low_pfn overflow.
781 */
782 if (freepage_order > 0 && freepage_order < MAX_ORDER)
783 low_pfn += (1UL << freepage_order) - 1;
Mel Gorman748446b2010-05-24 14:32:27 -0700784 continue;
Vlastimil Babka99c0fd52014-10-09 15:27:23 -0700785 }
Mel Gorman748446b2010-05-24 14:32:27 -0700786
Mel Gorman9927af742011-01-13 15:45:59 -0800787 /*
Vlastimil Babka29c0dde2015-09-08 15:02:46 -0700788 * Regardless of being on LRU, compound pages such as THP and
789 * hugetlbfs are not to be compacted. We can potentially save
790 * a lot of iterations if we skip them at once. The check is
791 * racy, but we can consider only valid values and the only
792 * danger is skipping too much.
Andrea Arcangelibc835012011-01-13 15:47:08 -0800793 */
Vlastimil Babka29c0dde2015-09-08 15:02:46 -0700794 if (PageCompound(page)) {
795 unsigned int comp_order = compound_order(page);
796
797 if (likely(comp_order < MAX_ORDER))
798 low_pfn += (1UL << comp_order) - 1;
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700799
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700800 goto isolate_fail;
Mel Gorman2a1402a2012-10-08 16:32:33 -0700801 }
802
Minchan Kimbda807d2016-07-26 15:23:05 -0700803 /*
804 * Check may be lockless but that's ok as we recheck later.
805 * It's possible to migrate LRU and non-lru movable pages.
806 * Skip any other type of page
807 */
808 if (!PageLRU(page)) {
Minchan Kimbda807d2016-07-26 15:23:05 -0700809 /*
810 * __PageMovable can return false positive so we need
811 * to verify it under page_lock.
812 */
813 if (unlikely(__PageMovable(page)) &&
814 !PageIsolated(page)) {
815 if (locked) {
816 spin_unlock_irqrestore(&zone->lru_lock,
817 flags);
818 locked = false;
819 }
820
821 if (isolate_movable_page(page, isolate_mode))
822 goto isolate_success;
823 }
824
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700825 goto isolate_fail;
Minchan Kimbda807d2016-07-26 15:23:05 -0700826 }
Vlastimil Babka29c0dde2015-09-08 15:02:46 -0700827
David Rientjes119d6d52014-04-03 14:48:00 -0700828 /*
829 * Migration will fail if an anonymous page is pinned in memory,
830 * so avoid taking lru_lock and isolating it unnecessarily in an
831 * admittedly racy check.
832 */
833 if (!page_mapping(page) &&
834 page_count(page) > page_mapcount(page))
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700835 goto isolate_fail;
David Rientjes119d6d52014-04-03 14:48:00 -0700836
Vlastimil Babka69b71892014-10-09 15:27:18 -0700837 /* If we already hold the lock, we can skip some rechecking */
838 if (!locked) {
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700839 locked = compact_trylock_irqsave(&zone->lru_lock,
840 &flags, cc);
Vlastimil Babka69b71892014-10-09 15:27:18 -0700841 if (!locked)
842 break;
Mel Gorman2a1402a2012-10-08 16:32:33 -0700843
Vlastimil Babka29c0dde2015-09-08 15:02:46 -0700844 /* Recheck PageLRU and PageCompound under lock */
Vlastimil Babka69b71892014-10-09 15:27:18 -0700845 if (!PageLRU(page))
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700846 goto isolate_fail;
Vlastimil Babka29c0dde2015-09-08 15:02:46 -0700847
848 /*
849 * Page become compound since the non-locked check,
850 * and it's on LRU. It can only be a THP so the order
851 * is safe to read and it's 0 for tail pages.
852 */
853 if (unlikely(PageCompound(page))) {
854 low_pfn += (1UL << compound_order(page)) - 1;
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700855 goto isolate_fail;
Vlastimil Babka69b71892014-10-09 15:27:18 -0700856 }
Andrea Arcangelibc835012011-01-13 15:47:08 -0800857 }
858
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700859 lruvec = mem_cgroup_page_lruvec(page, zone);
860
Mel Gorman748446b2010-05-24 14:32:27 -0700861 /* Try isolate the page */
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700862 if (__isolate_lru_page(page, isolate_mode) != 0)
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700863 goto isolate_fail;
Mel Gorman748446b2010-05-24 14:32:27 -0700864
Vlastimil Babka29c0dde2015-09-08 15:02:46 -0700865 VM_BUG_ON_PAGE(PageCompound(page), page);
Andrea Arcangelibc835012011-01-13 15:47:08 -0800866
Mel Gorman748446b2010-05-24 14:32:27 -0700867 /* Successfully isolated */
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700868 del_page_from_lru_list(page, lruvec, page_lru(page));
Joonsoo Kimb6c75012014-04-07 15:37:07 -0700869
870isolate_success:
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700871 list_add(&page->lru, &cc->migratepages);
Mel Gorman748446b2010-05-24 14:32:27 -0700872 cc->nr_migratepages++;
Mel Gormanb7aba692011-01-13 15:45:54 -0800873 nr_isolated++;
Mel Gorman748446b2010-05-24 14:32:27 -0700874
Vlastimil Babkaa34753d2016-05-19 17:11:51 -0700875 /*
876 * Record where we could have freed pages by migration and not
877 * yet flushed them to buddy allocator.
878 * - this is the lowest page that was isolated and likely be
879 * then freed by migration.
880 */
881 if (!cc->last_migrated_pfn)
882 cc->last_migrated_pfn = low_pfn;
883
Mel Gorman748446b2010-05-24 14:32:27 -0700884 /* Avoid isolating too much */
Hillf Danton31b83842012-01-10 15:07:59 -0800885 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
886 ++low_pfn;
Mel Gorman748446b2010-05-24 14:32:27 -0700887 break;
Hillf Danton31b83842012-01-10 15:07:59 -0800888 }
Vlastimil Babkafdd048e2016-05-19 17:11:55 -0700889
890 continue;
891isolate_fail:
892 if (!skip_on_failure)
893 continue;
894
895 /*
896 * We have isolated some pages, but then failed. Release them
897 * instead of migrating, as we cannot form the cc->order buddy
898 * page anyway.
899 */
900 if (nr_isolated) {
901 if (locked) {
902 spin_unlock_irqrestore(&zone->lru_lock, flags);
903 locked = false;
904 }
905 acct_isolated(zone, cc);
906 putback_movable_pages(&cc->migratepages);
907 cc->nr_migratepages = 0;
908 cc->last_migrated_pfn = 0;
909 nr_isolated = 0;
910 }
911
912 if (low_pfn < next_skip_pfn) {
913 low_pfn = next_skip_pfn - 1;
914 /*
915 * The check near the loop beginning would have updated
916 * next_skip_pfn too, but this is a bit simpler.
917 */
918 next_skip_pfn += 1UL << cc->order;
919 }
Mel Gorman748446b2010-05-24 14:32:27 -0700920 }
921
Vlastimil Babka99c0fd52014-10-09 15:27:23 -0700922 /*
923 * The PageBuddy() check could have potentially brought us outside
924 * the range to be scanned.
925 */
926 if (unlikely(low_pfn > end_pfn))
927 low_pfn = end_pfn;
928
Mel Gormanc67fe372012-08-21 16:16:17 -0700929 if (locked)
930 spin_unlock_irqrestore(&zone->lru_lock, flags);
Mel Gorman748446b2010-05-24 14:32:27 -0700931
Vlastimil Babka50b5b092014-01-21 15:51:10 -0800932 /*
933 * Update the pageblock-skip information and cached scanner pfn,
934 * if the whole pageblock was scanned without isolating any page.
Vlastimil Babka50b5b092014-01-21 15:51:10 -0800935 */
David Rientjes35979ef2014-06-04 16:08:27 -0700936 if (low_pfn == end_pfn)
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700937 update_pageblock_skip(cc, valid_page, nr_isolated, true);
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700938
Joonsoo Kime34d85f2015-02-11 15:27:04 -0800939 trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
940 nr_scanned, nr_isolated);
Mel Gormanb7aba692011-01-13 15:45:54 -0800941
Minchan Kim010fc292012-12-20 15:05:06 -0800942 count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
Mel Gorman397487d2012-10-19 12:00:10 +0100943 if (nr_isolated)
Minchan Kim010fc292012-12-20 15:05:06 -0800944 count_compact_events(COMPACTISOLATED, nr_isolated);
Mel Gorman397487d2012-10-19 12:00:10 +0100945
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100946 return low_pfn;
947}
948
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700949/**
950 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
951 * @cc: Compaction control structure.
952 * @start_pfn: The first PFN to start isolating.
953 * @end_pfn: The one-past-last PFN.
954 *
955 * Returns zero if isolation fails fatally due to e.g. pending signal.
956 * Otherwise, function returns one-past-the-last PFN of isolated page
957 * (which may be greater than end_pfn if end fell in a middle of a THP page).
958 */
959unsigned long
960isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
961 unsigned long end_pfn)
962{
Joonsoo Kime1409c32016-03-15 14:57:48 -0700963 unsigned long pfn, block_start_pfn, block_end_pfn;
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700964
965 /* Scan block by block. First and last block may be incomplete */
966 pfn = start_pfn;
Vlastimil Babka06b66402016-05-19 17:11:48 -0700967 block_start_pfn = pageblock_start_pfn(pfn);
Joonsoo Kime1409c32016-03-15 14:57:48 -0700968 if (block_start_pfn < cc->zone->zone_start_pfn)
969 block_start_pfn = cc->zone->zone_start_pfn;
Vlastimil Babka06b66402016-05-19 17:11:48 -0700970 block_end_pfn = pageblock_end_pfn(pfn);
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700971
972 for (; pfn < end_pfn; pfn = block_end_pfn,
Joonsoo Kime1409c32016-03-15 14:57:48 -0700973 block_start_pfn = block_end_pfn,
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700974 block_end_pfn += pageblock_nr_pages) {
975
976 block_end_pfn = min(block_end_pfn, end_pfn);
977
Joonsoo Kime1409c32016-03-15 14:57:48 -0700978 if (!pageblock_pfn_to_page(block_start_pfn,
979 block_end_pfn, cc->zone))
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700980 continue;
981
982 pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
983 ISOLATE_UNEVICTABLE);
984
Hugh Dickins14af4a52016-05-05 16:22:15 -0700985 if (!pfn)
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700986 break;
Joonsoo Kim6ea41c02014-10-29 14:50:20 -0700987
988 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
989 break;
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700990 }
991 acct_isolated(cc->zone, cc);
992
993 return pfn;
994}
995
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100996#endif /* CONFIG_COMPACTION || CONFIG_CMA */
997#ifdef CONFIG_COMPACTION
Andrew Morton018e9a42015-04-15 16:15:20 -0700998
999/* Returns true if the page is within a block suitable for migration to */
1000static bool suitable_migration_target(struct page *page)
1001{
1002 /* If the page is a large free page, then disallow migration */
1003 if (PageBuddy(page)) {
1004 /*
1005 * We are checking page_order without zone->lock taken. But
1006 * the only small danger is that we skip a potentially suitable
1007 * pageblock, so it's not worth to check order for valid range.
1008 */
1009 if (page_order_unsafe(page) >= pageblock_order)
1010 return false;
1011 }
1012
1013 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1014 if (migrate_async_suitable(get_pageblock_migratetype(page)))
1015 return true;
1016
1017 /* Otherwise skip the block */
1018 return false;
1019}
1020
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001021/*
Vlastimil Babkaf2849aa2015-09-08 15:02:36 -07001022 * Test whether the free scanner has reached the same or lower pageblock than
1023 * the migration scanner, and compaction should thus terminate.
1024 */
1025static inline bool compact_scanners_met(struct compact_control *cc)
1026{
1027 return (cc->free_pfn >> pageblock_order)
1028 <= (cc->migrate_pfn >> pageblock_order);
1029}
1030
1031/*
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001032 * Based on information in the current compact_control, find blocks
1033 * suitable for isolating free pages from and then isolate them.
1034 */
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001035static void isolate_freepages(struct compact_control *cc)
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001036{
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001037 struct zone *zone = cc->zone;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001038 struct page *page;
Vlastimil Babkac96b9e52014-06-04 16:07:26 -07001039 unsigned long block_start_pfn; /* start of current pageblock */
Vlastimil Babkae14c7202014-10-09 15:27:20 -07001040 unsigned long isolate_start_pfn; /* exact pfn we start at */
Vlastimil Babkac96b9e52014-06-04 16:07:26 -07001041 unsigned long block_end_pfn; /* end of current pageblock */
1042 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001043 struct list_head *freelist = &cc->freepages;
1044
1045 /*
1046 * Initialise the free scanner. The starting point is where we last
Vlastimil Babka49e068f2014-05-06 12:50:03 -07001047 * successfully isolated from, zone-cached value, or the end of the
Vlastimil Babkae14c7202014-10-09 15:27:20 -07001048 * zone when isolating for the first time. For looping we also need
1049 * this pfn aligned down to the pageblock boundary, because we do
Vlastimil Babkac96b9e52014-06-04 16:07:26 -07001050 * block_start_pfn -= pageblock_nr_pages in the for loop.
1051 * For ending point, take care when isolating in last pageblock of a
1052 * a zone which ends in the middle of a pageblock.
Vlastimil Babka49e068f2014-05-06 12:50:03 -07001053 * The low boundary is the end of the pageblock the migration scanner
1054 * is using.
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001055 */
Vlastimil Babkae14c7202014-10-09 15:27:20 -07001056 isolate_start_pfn = cc->free_pfn;
Vlastimil Babka06b66402016-05-19 17:11:48 -07001057 block_start_pfn = pageblock_start_pfn(cc->free_pfn);
Vlastimil Babkac96b9e52014-06-04 16:07:26 -07001058 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1059 zone_end_pfn(zone));
Vlastimil Babka06b66402016-05-19 17:11:48 -07001060 low_pfn = pageblock_end_pfn(cc->migrate_pfn);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001061
1062 /*
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001063 * Isolate free pages until enough are available to migrate the
1064 * pages on cc->migratepages. We stop searching if the migrate
1065 * and free page scanners meet or enough free pages are isolated.
1066 */
Vlastimil Babkaf5f61a32015-09-08 15:02:39 -07001067 for (; block_start_pfn >= low_pfn;
Vlastimil Babkac96b9e52014-06-04 16:07:26 -07001068 block_end_pfn = block_start_pfn,
Vlastimil Babkae14c7202014-10-09 15:27:20 -07001069 block_start_pfn -= pageblock_nr_pages,
1070 isolate_start_pfn = block_start_pfn) {
David Rientjesf6ea3ad2013-09-30 13:45:03 -07001071 /*
1072 * This can iterate a massively long zone without finding any
1073 * suitable migration targets, so periodically check if we need
Vlastimil Babkabe976572014-06-04 16:10:41 -07001074 * to schedule, or even abort async compaction.
David Rientjesf6ea3ad2013-09-30 13:45:03 -07001075 */
Vlastimil Babkabe976572014-06-04 16:10:41 -07001076 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1077 && compact_should_abort(cc))
1078 break;
David Rientjesf6ea3ad2013-09-30 13:45:03 -07001079
Vlastimil Babka7d49d882014-10-09 15:27:11 -07001080 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1081 zone);
1082 if (!page)
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001083 continue;
1084
1085 /* Check the block is suitable for migration */
Linus Torvalds68e3e922012-06-03 20:05:57 -07001086 if (!suitable_migration_target(page))
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001087 continue;
Linus Torvalds68e3e922012-06-03 20:05:57 -07001088
Mel Gormanbb13ffe2012-10-08 16:32:41 -07001089 /* If isolation recently failed, do not retry */
1090 if (!isolation_suitable(cc, page))
1091 continue;
1092
Vlastimil Babkae14c7202014-10-09 15:27:20 -07001093 /* Found a block suitable for isolating free pages from. */
David Rientjesa46cbf32016-07-14 12:06:50 -07001094 isolate_freepages_block(cc, &isolate_start_pfn, block_end_pfn,
1095 freelist, false);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001096
1097 /*
David Rientjesa46cbf32016-07-14 12:06:50 -07001098 * If we isolated enough freepages, or aborted due to lock
1099 * contention, terminate.
Vlastimil Babkae14c7202014-10-09 15:27:20 -07001100 */
Vlastimil Babkaf5f61a32015-09-08 15:02:39 -07001101 if ((cc->nr_freepages >= cc->nr_migratepages)
1102 || cc->contended) {
David Rientjesa46cbf32016-07-14 12:06:50 -07001103 if (isolate_start_pfn >= block_end_pfn) {
1104 /*
1105 * Restart at previous pageblock if more
1106 * freepages can be isolated next time.
1107 */
Vlastimil Babkaf5f61a32015-09-08 15:02:39 -07001108 isolate_start_pfn =
1109 block_start_pfn - pageblock_nr_pages;
David Rientjesa46cbf32016-07-14 12:06:50 -07001110 }
Vlastimil Babkabe976572014-06-04 16:10:41 -07001111 break;
David Rientjesa46cbf32016-07-14 12:06:50 -07001112 } else if (isolate_start_pfn < block_end_pfn) {
Vlastimil Babkaf5f61a32015-09-08 15:02:39 -07001113 /*
David Rientjesa46cbf32016-07-14 12:06:50 -07001114 * If isolation failed early, do not continue
1115 * needlessly.
Vlastimil Babkaf5f61a32015-09-08 15:02:39 -07001116 */
David Rientjesa46cbf32016-07-14 12:06:50 -07001117 break;
Vlastimil Babkaf5f61a32015-09-08 15:02:39 -07001118 }
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +01001119 }
1120
Joonsoo Kim66c64222016-07-26 15:23:40 -07001121 /* __isolate_free_page() does not map the pages */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001122 map_pages(freelist);
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +01001123
Vlastimil Babka7ed695e2014-01-21 15:51:09 -08001124 /*
Vlastimil Babkaf5f61a32015-09-08 15:02:39 -07001125 * Record where the free scanner will restart next time. Either we
1126 * broke from the loop and set isolate_start_pfn based on the last
1127 * call to isolate_freepages_block(), or we met the migration scanner
1128 * and the loop terminated due to isolate_start_pfn < low_pfn
Vlastimil Babka7ed695e2014-01-21 15:51:09 -08001129 */
Vlastimil Babkaf5f61a32015-09-08 15:02:39 -07001130 cc->free_pfn = isolate_start_pfn;
Mel Gorman748446b2010-05-24 14:32:27 -07001131}
1132
1133/*
1134 * This is a migrate-callback that "allocates" freepages by taking pages
1135 * from the isolated freelists in the block we are migrating to.
1136 */
1137static struct page *compaction_alloc(struct page *migratepage,
1138 unsigned long data,
1139 int **result)
1140{
1141 struct compact_control *cc = (struct compact_control *)data;
1142 struct page *freepage;
1143
Vlastimil Babkabe976572014-06-04 16:10:41 -07001144 /*
1145 * Isolate free pages if necessary, and if we are not aborting due to
1146 * contention.
1147 */
Mel Gorman748446b2010-05-24 14:32:27 -07001148 if (list_empty(&cc->freepages)) {
Vlastimil Babkabe976572014-06-04 16:10:41 -07001149 if (!cc->contended)
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001150 isolate_freepages(cc);
Mel Gorman748446b2010-05-24 14:32:27 -07001151
1152 if (list_empty(&cc->freepages))
1153 return NULL;
1154 }
1155
1156 freepage = list_entry(cc->freepages.next, struct page, lru);
1157 list_del(&freepage->lru);
1158 cc->nr_freepages--;
1159
1160 return freepage;
1161}
1162
1163/*
David Rientjesd53aea32014-06-04 16:08:26 -07001164 * This is a migrate-callback that "frees" freepages back to the isolated
1165 * freelist. All pages on the freelist are from the same zone, so there is no
1166 * special handling needed for NUMA.
1167 */
1168static void compaction_free(struct page *page, unsigned long data)
1169{
1170 struct compact_control *cc = (struct compact_control *)data;
1171
1172 list_add(&page->lru, &cc->freepages);
1173 cc->nr_freepages++;
1174}
1175
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001176/* possible outcome of isolate_migratepages */
1177typedef enum {
1178 ISOLATE_ABORT, /* Abort compaction now */
1179 ISOLATE_NONE, /* No pages isolated, continue scanning */
1180 ISOLATE_SUCCESS, /* Pages isolated, migrate */
1181} isolate_migrate_t;
1182
1183/*
Eric B Munson5bbe3542015-04-15 16:13:20 -07001184 * Allow userspace to control policy on scanning the unevictable LRU for
1185 * compactable pages.
1186 */
1187int sysctl_compact_unevictable_allowed __read_mostly = 1;
1188
1189/*
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001190 * Isolate all pages that can be migrated from the first suitable block,
1191 * starting at the block pointed to by the migrate scanner pfn within
1192 * compact_control.
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001193 */
1194static isolate_migrate_t isolate_migratepages(struct zone *zone,
1195 struct compact_control *cc)
1196{
Joonsoo Kime1409c32016-03-15 14:57:48 -07001197 unsigned long block_start_pfn;
1198 unsigned long block_end_pfn;
1199 unsigned long low_pfn;
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001200 struct page *page;
1201 const isolate_mode_t isolate_mode =
Eric B Munson5bbe3542015-04-15 16:13:20 -07001202 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001203 (cc->mode == MIGRATE_ASYNC ? ISOLATE_ASYNC_MIGRATE : 0);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001204
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001205 /*
1206 * Start at where we last stopped, or beginning of the zone as
1207 * initialized by compact_zone()
1208 */
1209 low_pfn = cc->migrate_pfn;
Vlastimil Babka06b66402016-05-19 17:11:48 -07001210 block_start_pfn = pageblock_start_pfn(low_pfn);
Joonsoo Kime1409c32016-03-15 14:57:48 -07001211 if (block_start_pfn < zone->zone_start_pfn)
1212 block_start_pfn = zone->zone_start_pfn;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001213
1214 /* Only scan within a pageblock boundary */
Vlastimil Babka06b66402016-05-19 17:11:48 -07001215 block_end_pfn = pageblock_end_pfn(low_pfn);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001216
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001217 /*
1218 * Iterate over whole pageblocks until we find the first suitable.
1219 * Do not cross the free scanner.
1220 */
Joonsoo Kime1409c32016-03-15 14:57:48 -07001221 for (; block_end_pfn <= cc->free_pfn;
1222 low_pfn = block_end_pfn,
1223 block_start_pfn = block_end_pfn,
1224 block_end_pfn += pageblock_nr_pages) {
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001225
1226 /*
1227 * This can potentially iterate a massively long zone with
1228 * many pageblocks unsuitable, so periodically check if we
1229 * need to schedule, or even abort async compaction.
1230 */
1231 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1232 && compact_should_abort(cc))
1233 break;
1234
Joonsoo Kime1409c32016-03-15 14:57:48 -07001235 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1236 zone);
Vlastimil Babka7d49d882014-10-09 15:27:11 -07001237 if (!page)
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001238 continue;
1239
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001240 /* If isolation recently failed, do not retry */
1241 if (!isolation_suitable(cc, page))
1242 continue;
1243
1244 /*
1245 * For async compaction, also only scan in MOVABLE blocks.
1246 * Async compaction is optimistic to see if the minimum amount
1247 * of work satisfies the allocation.
1248 */
1249 if (cc->mode == MIGRATE_ASYNC &&
1250 !migrate_async_suitable(get_pageblock_migratetype(page)))
1251 continue;
1252
1253 /* Perform the isolation */
Joonsoo Kime1409c32016-03-15 14:57:48 -07001254 low_pfn = isolate_migratepages_block(cc, low_pfn,
1255 block_end_pfn, isolate_mode);
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001256
Hugh Dickinsff599092015-02-12 15:00:28 -08001257 if (!low_pfn || cc->contended) {
1258 acct_isolated(zone, cc);
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001259 return ISOLATE_ABORT;
Hugh Dickinsff599092015-02-12 15:00:28 -08001260 }
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001261
1262 /*
1263 * Either we isolated something and proceed with migration. Or
1264 * we failed and compact_zone should decide if we should
1265 * continue or not.
1266 */
1267 break;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001268 }
1269
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001270 acct_isolated(zone, cc);
Vlastimil Babkaf2849aa2015-09-08 15:02:36 -07001271 /* Record where migration scanner will be restarted. */
1272 cc->migrate_pfn = low_pfn;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001273
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001274 return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001275}
1276
Yaowei Bai21c527a2015-11-05 18:47:20 -08001277/*
1278 * order == -1 is expected when compacting via
1279 * /proc/sys/vm/compact_memory
1280 */
1281static inline bool is_via_compact_memory(int order)
1282{
1283 return order == -1;
1284}
1285
Michal Hockoea7ab982016-05-20 16:56:38 -07001286static enum compact_result __compact_finished(struct zone *zone, struct compact_control *cc,
David Rientjes6d7ce552014-10-09 15:27:27 -07001287 const int migratetype)
Mel Gorman748446b2010-05-24 14:32:27 -07001288{
Mel Gorman8fb74b92013-01-11 14:32:16 -08001289 unsigned int order;
Andrea Arcangeli5a03b052011-01-13 15:47:11 -08001290 unsigned long watermark;
Mel Gorman56de7262010-05-24 14:32:30 -07001291
Vlastimil Babkabe976572014-06-04 16:10:41 -07001292 if (cc->contended || fatal_signal_pending(current))
Vlastimil Babka2d1e1042015-11-05 18:48:02 -08001293 return COMPACT_CONTENDED;
Mel Gorman748446b2010-05-24 14:32:27 -07001294
Mel Gorman753341a2012-10-08 16:32:40 -07001295 /* Compaction run completes if the migrate and free scanner meet */
Vlastimil Babkaf2849aa2015-09-08 15:02:36 -07001296 if (compact_scanners_met(cc)) {
Vlastimil Babka55b7c4c2014-01-21 15:51:11 -08001297 /* Let the next compaction start anew. */
Vlastimil Babka02333642015-09-08 15:02:42 -07001298 reset_cached_positions(zone);
Vlastimil Babka55b7c4c2014-01-21 15:51:11 -08001299
Mel Gorman62997022012-10-08 16:32:47 -07001300 /*
1301 * Mark that the PG_migrate_skip information should be cleared
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07001302 * by kswapd when it goes to sleep. kcompactd does not set the
Mel Gorman62997022012-10-08 16:32:47 -07001303 * flag itself as the decision to be clear should be directly
1304 * based on an allocation request.
1305 */
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07001306 if (cc->direct_compaction)
Mel Gorman62997022012-10-08 16:32:47 -07001307 zone->compact_blockskip_flush = true;
1308
Michal Hockoc8f7de02016-05-20 16:56:47 -07001309 if (cc->whole_zone)
1310 return COMPACT_COMPLETE;
1311 else
1312 return COMPACT_PARTIAL_SKIPPED;
Mel Gormanbb13ffe2012-10-08 16:32:41 -07001313 }
Mel Gorman748446b2010-05-24 14:32:27 -07001314
Yaowei Bai21c527a2015-11-05 18:47:20 -08001315 if (is_via_compact_memory(cc->order))
Mel Gorman56de7262010-05-24 14:32:30 -07001316 return COMPACT_CONTINUE;
1317
Michal Hocko3957c772011-06-15 15:08:25 -07001318 /* Compaction run is not finished if the watermark is not met */
1319 watermark = low_wmark_pages(zone);
Michal Hocko3957c772011-06-15 15:08:25 -07001320
Vlastimil Babkaebff3982014-12-10 15:43:22 -08001321 if (!zone_watermark_ok(zone, cc->order, watermark, cc->classzone_idx,
1322 cc->alloc_flags))
Michal Hocko3957c772011-06-15 15:08:25 -07001323 return COMPACT_CONTINUE;
1324
Mel Gorman56de7262010-05-24 14:32:30 -07001325 /* Direct compactor: Is a suitable page free? */
Mel Gorman8fb74b92013-01-11 14:32:16 -08001326 for (order = cc->order; order < MAX_ORDER; order++) {
1327 struct free_area *area = &zone->free_area[order];
Joonsoo Kim2149cda2015-04-14 15:45:21 -07001328 bool can_steal;
Mel Gorman56de7262010-05-24 14:32:30 -07001329
Mel Gorman8fb74b92013-01-11 14:32:16 -08001330 /* Job done if page is free of the right migratetype */
David Rientjes6d7ce552014-10-09 15:27:27 -07001331 if (!list_empty(&area->free_list[migratetype]))
Mel Gorman8fb74b92013-01-11 14:32:16 -08001332 return COMPACT_PARTIAL;
1333
Joonsoo Kim2149cda2015-04-14 15:45:21 -07001334#ifdef CONFIG_CMA
1335 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
1336 if (migratetype == MIGRATE_MOVABLE &&
1337 !list_empty(&area->free_list[MIGRATE_CMA]))
1338 return COMPACT_PARTIAL;
1339#endif
1340 /*
1341 * Job done if allocation would steal freepages from
1342 * other migratetype buddy lists.
1343 */
1344 if (find_suitable_fallback(area, order, migratetype,
1345 true, &can_steal) != -1)
Mel Gorman8fb74b92013-01-11 14:32:16 -08001346 return COMPACT_PARTIAL;
Mel Gorman56de7262010-05-24 14:32:30 -07001347 }
1348
Joonsoo Kim837d0262015-02-11 15:27:06 -08001349 return COMPACT_NO_SUITABLE_PAGE;
1350}
1351
Michal Hockoea7ab982016-05-20 16:56:38 -07001352static enum compact_result compact_finished(struct zone *zone,
1353 struct compact_control *cc,
1354 const int migratetype)
Joonsoo Kim837d0262015-02-11 15:27:06 -08001355{
1356 int ret;
1357
1358 ret = __compact_finished(zone, cc, migratetype);
1359 trace_mm_compaction_finished(zone, cc->order, ret);
1360 if (ret == COMPACT_NO_SUITABLE_PAGE)
1361 ret = COMPACT_CONTINUE;
1362
1363 return ret;
Mel Gorman748446b2010-05-24 14:32:27 -07001364}
1365
Mel Gorman3e7d3442011-01-13 15:45:56 -08001366/*
1367 * compaction_suitable: Is this suitable to run compaction on this zone now?
1368 * Returns
1369 * COMPACT_SKIPPED - If there are too few free pages for compaction
1370 * COMPACT_PARTIAL - If the allocation would succeed without compaction
1371 * COMPACT_CONTINUE - If compaction should run now
1372 */
Michal Hockoea7ab982016-05-20 16:56:38 -07001373static enum compact_result __compaction_suitable(struct zone *zone, int order,
Mel Gormanc6038442016-05-19 17:13:38 -07001374 unsigned int alloc_flags,
Michal Hocko86a294a2016-05-20 16:57:12 -07001375 int classzone_idx,
1376 unsigned long wmark_target)
Mel Gorman3e7d3442011-01-13 15:45:56 -08001377{
1378 int fragindex;
1379 unsigned long watermark;
1380
Yaowei Bai21c527a2015-11-05 18:47:20 -08001381 if (is_via_compact_memory(order))
Michal Hocko3957c772011-06-15 15:08:25 -07001382 return COMPACT_CONTINUE;
1383
Vlastimil Babkaebff3982014-12-10 15:43:22 -08001384 watermark = low_wmark_pages(zone);
1385 /*
1386 * If watermarks for high-order allocation are already met, there
1387 * should be no need for compaction at all.
1388 */
1389 if (zone_watermark_ok(zone, order, watermark, classzone_idx,
1390 alloc_flags))
1391 return COMPACT_PARTIAL;
1392
Michal Hocko3957c772011-06-15 15:08:25 -07001393 /*
Mel Gorman3e7d3442011-01-13 15:45:56 -08001394 * Watermarks for order-0 must be met for compaction. Note the 2UL.
1395 * This is because during migration, copies of pages need to be
1396 * allocated and for a short time, the footprint is higher
1397 */
Vlastimil Babkaebff3982014-12-10 15:43:22 -08001398 watermark += (2UL << order);
Michal Hocko86a294a2016-05-20 16:57:12 -07001399 if (!__zone_watermark_ok(zone, 0, watermark, classzone_idx,
1400 alloc_flags, wmark_target))
Mel Gorman3e7d3442011-01-13 15:45:56 -08001401 return COMPACT_SKIPPED;
1402
1403 /*
1404 * fragmentation index determines if allocation failures are due to
1405 * low memory or external fragmentation
1406 *
Vlastimil Babkaebff3982014-12-10 15:43:22 -08001407 * index of -1000 would imply allocations might succeed depending on
1408 * watermarks, but we already failed the high-order watermark check
Mel Gorman3e7d3442011-01-13 15:45:56 -08001409 * index towards 0 implies failure is due to lack of memory
1410 * index towards 1000 implies failure is due to fragmentation
1411 *
1412 * Only compact if a failure would be due to fragmentation.
1413 */
1414 fragindex = fragmentation_index(zone, order);
1415 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
Joonsoo Kim837d0262015-02-11 15:27:06 -08001416 return COMPACT_NOT_SUITABLE_ZONE;
Mel Gorman3e7d3442011-01-13 15:45:56 -08001417
Mel Gorman3e7d3442011-01-13 15:45:56 -08001418 return COMPACT_CONTINUE;
1419}
1420
Michal Hockoea7ab982016-05-20 16:56:38 -07001421enum compact_result compaction_suitable(struct zone *zone, int order,
Mel Gormanc6038442016-05-19 17:13:38 -07001422 unsigned int alloc_flags,
1423 int classzone_idx)
Joonsoo Kim837d0262015-02-11 15:27:06 -08001424{
Michal Hockoea7ab982016-05-20 16:56:38 -07001425 enum compact_result ret;
Joonsoo Kim837d0262015-02-11 15:27:06 -08001426
Michal Hocko86a294a2016-05-20 16:57:12 -07001427 ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx,
1428 zone_page_state(zone, NR_FREE_PAGES));
Joonsoo Kim837d0262015-02-11 15:27:06 -08001429 trace_mm_compaction_suitable(zone, order, ret);
1430 if (ret == COMPACT_NOT_SUITABLE_ZONE)
1431 ret = COMPACT_SKIPPED;
1432
1433 return ret;
1434}
1435
Michal Hocko86a294a2016-05-20 16:57:12 -07001436bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
1437 int alloc_flags)
1438{
1439 struct zone *zone;
1440 struct zoneref *z;
1441
1442 /*
1443 * Make sure at least one zone would pass __compaction_suitable if we continue
1444 * retrying the reclaim.
1445 */
1446 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1447 ac->nodemask) {
1448 unsigned long available;
1449 enum compact_result compact_result;
1450
1451 /*
1452 * Do not consider all the reclaimable memory because we do not
1453 * want to trash just for a single high order allocation which
1454 * is even not guaranteed to appear even if __compaction_suitable
1455 * is happy about the watermark check.
1456 */
1457 available = zone_reclaimable_pages(zone) / order;
1458 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
1459 compact_result = __compaction_suitable(zone, order, alloc_flags,
1460 ac_classzone_idx(ac), available);
1461 if (compact_result != COMPACT_SKIPPED &&
1462 compact_result != COMPACT_NOT_SUITABLE_ZONE)
1463 return true;
1464 }
1465
1466 return false;
1467}
1468
Michal Hockoea7ab982016-05-20 16:56:38 -07001469static enum compact_result compact_zone(struct zone *zone, struct compact_control *cc)
Mel Gorman748446b2010-05-24 14:32:27 -07001470{
Michal Hockoea7ab982016-05-20 16:56:38 -07001471 enum compact_result ret;
Mel Gormanc89511a2012-10-08 16:32:45 -07001472 unsigned long start_pfn = zone->zone_start_pfn;
Cody P Schafer108bcc92013-02-22 16:35:23 -08001473 unsigned long end_pfn = zone_end_pfn(zone);
David Rientjes6d7ce552014-10-09 15:27:27 -07001474 const int migratetype = gfpflags_to_migratetype(cc->gfp_mask);
David Rientjese0b9dae2014-06-04 16:08:28 -07001475 const bool sync = cc->mode != MIGRATE_ASYNC;
Mel Gorman748446b2010-05-24 14:32:27 -07001476
Vlastimil Babkaebff3982014-12-10 15:43:22 -08001477 ret = compaction_suitable(zone, cc->order, cc->alloc_flags,
1478 cc->classzone_idx);
Michal Hockoc46649d2016-05-20 16:56:41 -07001479 /* Compaction is likely to fail */
1480 if (ret == COMPACT_PARTIAL || ret == COMPACT_SKIPPED)
Mel Gorman3e7d3442011-01-13 15:45:56 -08001481 return ret;
Michal Hockoc46649d2016-05-20 16:56:41 -07001482
1483 /* huh, compaction_suitable is returning something unexpected */
1484 VM_BUG_ON(ret != COMPACT_CONTINUE);
Mel Gorman3e7d3442011-01-13 15:45:56 -08001485
Mel Gormanc89511a2012-10-08 16:32:45 -07001486 /*
Vlastimil Babkad3132e42014-01-21 15:51:08 -08001487 * Clear pageblock skip if there were failures recently and compaction
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07001488 * is about to be retried after being deferred.
Vlastimil Babkad3132e42014-01-21 15:51:08 -08001489 */
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07001490 if (compaction_restarting(zone, cc->order))
Vlastimil Babkad3132e42014-01-21 15:51:08 -08001491 __reset_isolation_suitable(zone);
1492
1493 /*
Mel Gormanc89511a2012-10-08 16:32:45 -07001494 * Setup to move all movable pages to the end of the zone. Used cached
1495 * information on where the scanners should start but check that it
1496 * is initialised by ensuring the values are within zone boundaries.
1497 */
David Rientjese0b9dae2014-06-04 16:08:28 -07001498 cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync];
Mel Gormanc89511a2012-10-08 16:32:45 -07001499 cc->free_pfn = zone->compact_cached_free_pfn;
Joonsoo Kim623446e2016-03-15 14:57:45 -07001500 if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
Vlastimil Babka06b66402016-05-19 17:11:48 -07001501 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
Mel Gormanc89511a2012-10-08 16:32:45 -07001502 zone->compact_cached_free_pfn = cc->free_pfn;
1503 }
Joonsoo Kim623446e2016-03-15 14:57:45 -07001504 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
Mel Gormanc89511a2012-10-08 16:32:45 -07001505 cc->migrate_pfn = start_pfn;
David Rientjes35979ef2014-06-04 16:08:27 -07001506 zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
1507 zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
Mel Gormanc89511a2012-10-08 16:32:45 -07001508 }
Michal Hockoc8f7de02016-05-20 16:56:47 -07001509
1510 if (cc->migrate_pfn == start_pfn)
1511 cc->whole_zone = true;
1512
Joonsoo Kim1a167182015-09-08 15:03:59 -07001513 cc->last_migrated_pfn = 0;
Mel Gorman748446b2010-05-24 14:32:27 -07001514
Joonsoo Kim16c4a092015-02-11 15:27:01 -08001515 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
1516 cc->free_pfn, end_pfn, sync);
Mel Gorman0eb927c2014-01-21 15:51:05 -08001517
Mel Gorman748446b2010-05-24 14:32:27 -07001518 migrate_prep_local();
1519
David Rientjes6d7ce552014-10-09 15:27:27 -07001520 while ((ret = compact_finished(zone, cc, migratetype)) ==
1521 COMPACT_CONTINUE) {
Minchan Kim9d502c12011-03-22 16:30:39 -07001522 int err;
Mel Gorman748446b2010-05-24 14:32:27 -07001523
Mel Gormanf9e35b32011-06-15 15:08:52 -07001524 switch (isolate_migratepages(zone, cc)) {
1525 case ISOLATE_ABORT:
Vlastimil Babka2d1e1042015-11-05 18:48:02 -08001526 ret = COMPACT_CONTENDED;
Rafael Aquini5733c7d2012-12-11 16:02:47 -08001527 putback_movable_pages(&cc->migratepages);
Shaohua Lie64c5232012-10-08 16:32:27 -07001528 cc->nr_migratepages = 0;
Mel Gormanf9e35b32011-06-15 15:08:52 -07001529 goto out;
1530 case ISOLATE_NONE:
Vlastimil Babkafdaf7f52014-12-10 15:43:34 -08001531 /*
1532 * We haven't isolated and migrated anything, but
1533 * there might still be unflushed migrations from
1534 * previous cc->order aligned block.
1535 */
1536 goto check_drain;
Mel Gormanf9e35b32011-06-15 15:08:52 -07001537 case ISOLATE_SUCCESS:
1538 ;
1539 }
Mel Gorman748446b2010-05-24 14:32:27 -07001540
David Rientjesd53aea32014-06-04 16:08:26 -07001541 err = migrate_pages(&cc->migratepages, compaction_alloc,
David Rientjese0b9dae2014-06-04 16:08:28 -07001542 compaction_free, (unsigned long)cc, cc->mode,
Mel Gorman7b2a2d42012-10-19 14:07:31 +01001543 MR_COMPACTION);
Mel Gorman748446b2010-05-24 14:32:27 -07001544
Vlastimil Babkaf8c93012014-06-04 16:08:32 -07001545 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
1546 &cc->migratepages);
Mel Gorman748446b2010-05-24 14:32:27 -07001547
Vlastimil Babkaf8c93012014-06-04 16:08:32 -07001548 /* All pages were either migrated or will be released */
1549 cc->nr_migratepages = 0;
Minchan Kim9d502c12011-03-22 16:30:39 -07001550 if (err) {
Rafael Aquini5733c7d2012-12-11 16:02:47 -08001551 putback_movable_pages(&cc->migratepages);
Vlastimil Babka7ed695e2014-01-21 15:51:09 -08001552 /*
1553 * migrate_pages() may return -ENOMEM when scanners meet
1554 * and we want compact_finished() to detect it
1555 */
Vlastimil Babkaf2849aa2015-09-08 15:02:36 -07001556 if (err == -ENOMEM && !compact_scanners_met(cc)) {
Vlastimil Babka2d1e1042015-11-05 18:48:02 -08001557 ret = COMPACT_CONTENDED;
David Rientjes4bf2bba2012-07-11 14:02:13 -07001558 goto out;
1559 }
Vlastimil Babkafdd048e2016-05-19 17:11:55 -07001560 /*
1561 * We failed to migrate at least one page in the current
1562 * order-aligned block, so skip the rest of it.
1563 */
1564 if (cc->direct_compaction &&
1565 (cc->mode == MIGRATE_ASYNC)) {
1566 cc->migrate_pfn = block_end_pfn(
1567 cc->migrate_pfn - 1, cc->order);
1568 /* Draining pcplists is useless in this case */
1569 cc->last_migrated_pfn = 0;
1570
1571 }
Mel Gorman748446b2010-05-24 14:32:27 -07001572 }
Vlastimil Babkafdaf7f52014-12-10 15:43:34 -08001573
Vlastimil Babkafdaf7f52014-12-10 15:43:34 -08001574check_drain:
1575 /*
1576 * Has the migration scanner moved away from the previous
1577 * cc->order aligned block where we migrated from? If yes,
1578 * flush the pages that were freed, so that they can merge and
1579 * compact_finished() can detect immediately if allocation
1580 * would succeed.
1581 */
Joonsoo Kim1a167182015-09-08 15:03:59 -07001582 if (cc->order > 0 && cc->last_migrated_pfn) {
Vlastimil Babkafdaf7f52014-12-10 15:43:34 -08001583 int cpu;
1584 unsigned long current_block_start =
Vlastimil Babka06b66402016-05-19 17:11:48 -07001585 block_start_pfn(cc->migrate_pfn, cc->order);
Vlastimil Babkafdaf7f52014-12-10 15:43:34 -08001586
Joonsoo Kim1a167182015-09-08 15:03:59 -07001587 if (cc->last_migrated_pfn < current_block_start) {
Vlastimil Babkafdaf7f52014-12-10 15:43:34 -08001588 cpu = get_cpu();
1589 lru_add_drain_cpu(cpu);
1590 drain_local_pages(zone);
1591 put_cpu();
1592 /* No more flushing until we migrate again */
Joonsoo Kim1a167182015-09-08 15:03:59 -07001593 cc->last_migrated_pfn = 0;
Vlastimil Babkafdaf7f52014-12-10 15:43:34 -08001594 }
1595 }
1596
Mel Gorman748446b2010-05-24 14:32:27 -07001597 }
1598
Mel Gormanf9e35b32011-06-15 15:08:52 -07001599out:
Vlastimil Babka6bace092014-12-10 15:43:31 -08001600 /*
1601 * Release free pages and update where the free scanner should restart,
1602 * so we don't leave any returned pages behind in the next attempt.
1603 */
1604 if (cc->nr_freepages > 0) {
1605 unsigned long free_pfn = release_freepages(&cc->freepages);
1606
1607 cc->nr_freepages = 0;
1608 VM_BUG_ON(free_pfn == 0);
1609 /* The cached pfn is always the first in a pageblock */
Vlastimil Babka06b66402016-05-19 17:11:48 -07001610 free_pfn = pageblock_start_pfn(free_pfn);
Vlastimil Babka6bace092014-12-10 15:43:31 -08001611 /*
1612 * Only go back, not forward. The cached pfn might have been
1613 * already reset to zone end in compact_finished()
1614 */
1615 if (free_pfn > zone->compact_cached_free_pfn)
1616 zone->compact_cached_free_pfn = free_pfn;
1617 }
Mel Gorman748446b2010-05-24 14:32:27 -07001618
Joonsoo Kim16c4a092015-02-11 15:27:01 -08001619 trace_mm_compaction_end(start_pfn, cc->migrate_pfn,
1620 cc->free_pfn, end_pfn, sync, ret);
Mel Gorman0eb927c2014-01-21 15:51:05 -08001621
Vlastimil Babka2d1e1042015-11-05 18:48:02 -08001622 if (ret == COMPACT_CONTENDED)
1623 ret = COMPACT_PARTIAL;
1624
Mel Gorman748446b2010-05-24 14:32:27 -07001625 return ret;
1626}
Mel Gorman76ab0f52010-05-24 14:32:28 -07001627
Michal Hockoea7ab982016-05-20 16:56:38 -07001628static enum compact_result compact_zone_order(struct zone *zone, int order,
Vlastimil Babkaebff3982014-12-10 15:43:22 -08001629 gfp_t gfp_mask, enum migrate_mode mode, int *contended,
Mel Gormanc6038442016-05-19 17:13:38 -07001630 unsigned int alloc_flags, int classzone_idx)
Mel Gorman56de7262010-05-24 14:32:30 -07001631{
Michal Hockoea7ab982016-05-20 16:56:38 -07001632 enum compact_result ret;
Mel Gorman56de7262010-05-24 14:32:30 -07001633 struct compact_control cc = {
1634 .nr_freepages = 0,
1635 .nr_migratepages = 0,
1636 .order = order,
David Rientjes6d7ce552014-10-09 15:27:27 -07001637 .gfp_mask = gfp_mask,
Mel Gorman56de7262010-05-24 14:32:30 -07001638 .zone = zone,
David Rientjese0b9dae2014-06-04 16:08:28 -07001639 .mode = mode,
Vlastimil Babkaebff3982014-12-10 15:43:22 -08001640 .alloc_flags = alloc_flags,
1641 .classzone_idx = classzone_idx,
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07001642 .direct_compaction = true,
Mel Gorman56de7262010-05-24 14:32:30 -07001643 };
1644 INIT_LIST_HEAD(&cc.freepages);
1645 INIT_LIST_HEAD(&cc.migratepages);
1646
Shaohua Lie64c5232012-10-08 16:32:27 -07001647 ret = compact_zone(zone, &cc);
1648
1649 VM_BUG_ON(!list_empty(&cc.freepages));
1650 VM_BUG_ON(!list_empty(&cc.migratepages));
1651
1652 *contended = cc.contended;
1653 return ret;
Mel Gorman56de7262010-05-24 14:32:30 -07001654}
1655
Mel Gorman5e771902010-05-24 14:32:31 -07001656int sysctl_extfrag_threshold = 500;
1657
Mel Gorman56de7262010-05-24 14:32:30 -07001658/**
1659 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
Mel Gorman56de7262010-05-24 14:32:30 -07001660 * @gfp_mask: The GFP mask of the current allocation
Vlastimil Babka1a6d53a2015-02-11 15:25:44 -08001661 * @order: The order of the current allocation
1662 * @alloc_flags: The allocation flags of the current allocation
1663 * @ac: The context of current allocation
David Rientjese0b9dae2014-06-04 16:08:28 -07001664 * @mode: The migration mode for async, sync light, or sync migration
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001665 * @contended: Return value that determines if compaction was aborted due to
1666 * need_resched() or lock contention
Mel Gorman56de7262010-05-24 14:32:30 -07001667 *
1668 * This is the main entry point for direct page compaction.
1669 */
Michal Hockoea7ab982016-05-20 16:56:38 -07001670enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
Mel Gormanc6038442016-05-19 17:13:38 -07001671 unsigned int alloc_flags, const struct alloc_context *ac,
1672 enum migrate_mode mode, int *contended)
Mel Gorman56de7262010-05-24 14:32:30 -07001673{
Mel Gorman56de7262010-05-24 14:32:30 -07001674 int may_enter_fs = gfp_mask & __GFP_FS;
1675 int may_perform_io = gfp_mask & __GFP_IO;
Mel Gorman56de7262010-05-24 14:32:30 -07001676 struct zoneref *z;
1677 struct zone *zone;
Michal Hocko1d4746d2016-05-20 16:56:44 -07001678 enum compact_result rc = COMPACT_SKIPPED;
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001679 int all_zones_contended = COMPACT_CONTENDED_LOCK; /* init for &= op */
1680
1681 *contended = COMPACT_CONTENDED_NONE;
Mel Gorman56de7262010-05-24 14:32:30 -07001682
Mel Gorman4ffb6332012-10-08 16:29:09 -07001683 /* Check if the GFP flags allow compaction */
Andrea Arcangelic5a73c32011-01-13 15:47:11 -08001684 if (!order || !may_enter_fs || !may_perform_io)
Vlastimil Babka53853e22014-10-09 15:27:02 -07001685 return COMPACT_SKIPPED;
Mel Gorman56de7262010-05-24 14:32:30 -07001686
Joonsoo Kim837d0262015-02-11 15:27:06 -08001687 trace_mm_compaction_try_to_compact_pages(order, gfp_mask, mode);
1688
Mel Gorman56de7262010-05-24 14:32:30 -07001689 /* Compact each zone in the list */
Vlastimil Babka1a6d53a2015-02-11 15:25:44 -08001690 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1691 ac->nodemask) {
Michal Hockoea7ab982016-05-20 16:56:38 -07001692 enum compact_result status;
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001693 int zone_contended;
Mel Gorman56de7262010-05-24 14:32:30 -07001694
Michal Hocko1d4746d2016-05-20 16:56:44 -07001695 if (compaction_deferred(zone, order)) {
1696 rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
Vlastimil Babka53853e22014-10-09 15:27:02 -07001697 continue;
Michal Hocko1d4746d2016-05-20 16:56:44 -07001698 }
Vlastimil Babka53853e22014-10-09 15:27:02 -07001699
David Rientjese0b9dae2014-06-04 16:08:28 -07001700 status = compact_zone_order(zone, order, gfp_mask, mode,
Vlastimil Babka1a6d53a2015-02-11 15:25:44 -08001701 &zone_contended, alloc_flags,
Mel Gorman93ea9962016-05-19 17:14:13 -07001702 ac_classzone_idx(ac));
Mel Gorman56de7262010-05-24 14:32:30 -07001703 rc = max(status, rc);
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001704 /*
1705 * It takes at least one zone that wasn't lock contended
1706 * to clear all_zones_contended.
1707 */
1708 all_zones_contended &= zone_contended;
Mel Gorman56de7262010-05-24 14:32:30 -07001709
Mel Gorman3e7d3442011-01-13 15:45:56 -08001710 /* If a normal allocation would succeed, stop compacting */
Vlastimil Babkaebff3982014-12-10 15:43:22 -08001711 if (zone_watermark_ok(zone, order, low_wmark_pages(zone),
Mel Gorman93ea9962016-05-19 17:14:13 -07001712 ac_classzone_idx(ac), alloc_flags)) {
Vlastimil Babka53853e22014-10-09 15:27:02 -07001713 /*
1714 * We think the allocation will succeed in this zone,
1715 * but it is not certain, hence the false. The caller
1716 * will repeat this with true if allocation indeed
1717 * succeeds in this zone.
1718 */
1719 compaction_defer_reset(zone, order, false);
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001720 /*
1721 * It is possible that async compaction aborted due to
1722 * need_resched() and the watermarks were ok thanks to
1723 * somebody else freeing memory. The allocation can
1724 * however still fail so we better signal the
1725 * need_resched() contention anyway (this will not
1726 * prevent the allocation attempt).
1727 */
1728 if (zone_contended == COMPACT_CONTENDED_SCHED)
1729 *contended = COMPACT_CONTENDED_SCHED;
1730
1731 goto break_loop;
1732 }
1733
Michal Hockoc8f7de02016-05-20 16:56:47 -07001734 if (mode != MIGRATE_ASYNC && (status == COMPACT_COMPLETE ||
1735 status == COMPACT_PARTIAL_SKIPPED)) {
Vlastimil Babka53853e22014-10-09 15:27:02 -07001736 /*
1737 * We think that allocation won't succeed in this zone
1738 * so we defer compaction there. If it ends up
1739 * succeeding after all, it will be reset.
1740 */
1741 defer_compaction(zone, order);
1742 }
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001743
1744 /*
1745 * We might have stopped compacting due to need_resched() in
1746 * async compaction, or due to a fatal signal detected. In that
1747 * case do not try further zones and signal need_resched()
1748 * contention.
1749 */
1750 if ((zone_contended == COMPACT_CONTENDED_SCHED)
1751 || fatal_signal_pending(current)) {
1752 *contended = COMPACT_CONTENDED_SCHED;
1753 goto break_loop;
1754 }
1755
1756 continue;
1757break_loop:
1758 /*
1759 * We might not have tried all the zones, so be conservative
1760 * and assume they are not all lock contended.
1761 */
1762 all_zones_contended = 0;
1763 break;
Mel Gorman56de7262010-05-24 14:32:30 -07001764 }
1765
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001766 /*
1767 * If at least one zone wasn't deferred or skipped, we report if all
1768 * zones that were tried were lock contended.
1769 */
Michal Hocko1d4746d2016-05-20 16:56:44 -07001770 if (rc > COMPACT_INACTIVE && all_zones_contended)
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001771 *contended = COMPACT_CONTENDED_LOCK;
1772
Mel Gorman56de7262010-05-24 14:32:30 -07001773 return rc;
1774}
1775
1776
Mel Gorman76ab0f52010-05-24 14:32:28 -07001777/* Compact all zones within a node */
Andrew Morton7103f162013-02-22 16:32:33 -08001778static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
Mel Gorman76ab0f52010-05-24 14:32:28 -07001779{
1780 int zoneid;
Mel Gorman76ab0f52010-05-24 14:32:28 -07001781 struct zone *zone;
1782
Mel Gorman76ab0f52010-05-24 14:32:28 -07001783 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
Mel Gorman76ab0f52010-05-24 14:32:28 -07001784
1785 zone = &pgdat->node_zones[zoneid];
1786 if (!populated_zone(zone))
1787 continue;
1788
Rik van Riel7be62de2012-03-21 16:33:52 -07001789 cc->nr_freepages = 0;
1790 cc->nr_migratepages = 0;
1791 cc->zone = zone;
1792 INIT_LIST_HEAD(&cc->freepages);
1793 INIT_LIST_HEAD(&cc->migratepages);
Mel Gorman76ab0f52010-05-24 14:32:28 -07001794
Gioh Kim195b0c62015-04-15 16:13:33 -07001795 /*
1796 * When called via /proc/sys/vm/compact_memory
1797 * this makes sure we compact the whole zone regardless of
1798 * cached scanner positions.
1799 */
Yaowei Bai21c527a2015-11-05 18:47:20 -08001800 if (is_via_compact_memory(cc->order))
Gioh Kim195b0c62015-04-15 16:13:33 -07001801 __reset_isolation_suitable(zone);
1802
Yaowei Bai21c527a2015-11-05 18:47:20 -08001803 if (is_via_compact_memory(cc->order) ||
1804 !compaction_deferred(zone, cc->order))
Rik van Riel7be62de2012-03-21 16:33:52 -07001805 compact_zone(zone, cc);
Mel Gorman76ab0f52010-05-24 14:32:28 -07001806
Rik van Riel7be62de2012-03-21 16:33:52 -07001807 VM_BUG_ON(!list_empty(&cc->freepages));
1808 VM_BUG_ON(!list_empty(&cc->migratepages));
Joonsoo Kim75469342016-01-14 15:20:48 -08001809
1810 if (is_via_compact_memory(cc->order))
1811 continue;
1812
1813 if (zone_watermark_ok(zone, cc->order,
1814 low_wmark_pages(zone), 0, 0))
1815 compaction_defer_reset(zone, cc->order, false);
Mel Gorman76ab0f52010-05-24 14:32:28 -07001816 }
Mel Gorman76ab0f52010-05-24 14:32:28 -07001817}
1818
Andrew Morton7103f162013-02-22 16:32:33 -08001819void compact_pgdat(pg_data_t *pgdat, int order)
Rik van Riel7be62de2012-03-21 16:33:52 -07001820{
1821 struct compact_control cc = {
1822 .order = order,
David Rientjese0b9dae2014-06-04 16:08:28 -07001823 .mode = MIGRATE_ASYNC,
Rik van Riel7be62de2012-03-21 16:33:52 -07001824 };
1825
Mel Gorman3a7200a2013-09-11 14:22:19 -07001826 if (!order)
1827 return;
1828
Andrew Morton7103f162013-02-22 16:32:33 -08001829 __compact_pgdat(pgdat, &cc);
Rik van Riel7be62de2012-03-21 16:33:52 -07001830}
1831
Andrew Morton7103f162013-02-22 16:32:33 -08001832static void compact_node(int nid)
Rik van Riel7be62de2012-03-21 16:33:52 -07001833{
Rik van Riel7be62de2012-03-21 16:33:52 -07001834 struct compact_control cc = {
1835 .order = -1,
David Rientjese0b9dae2014-06-04 16:08:28 -07001836 .mode = MIGRATE_SYNC,
David Rientjes91ca9182014-04-03 14:47:23 -07001837 .ignore_skip_hint = true,
Rik van Riel7be62de2012-03-21 16:33:52 -07001838 };
1839
Andrew Morton7103f162013-02-22 16:32:33 -08001840 __compact_pgdat(NODE_DATA(nid), &cc);
Rik van Riel7be62de2012-03-21 16:33:52 -07001841}
1842
Mel Gorman76ab0f52010-05-24 14:32:28 -07001843/* Compact all nodes in the system */
Jason Liu7964c062013-01-11 14:31:47 -08001844static void compact_nodes(void)
Mel Gorman76ab0f52010-05-24 14:32:28 -07001845{
1846 int nid;
1847
Hugh Dickins8575ec22012-03-21 16:33:53 -07001848 /* Flush pending updates to the LRU lists */
1849 lru_add_drain_all();
1850
Mel Gorman76ab0f52010-05-24 14:32:28 -07001851 for_each_online_node(nid)
1852 compact_node(nid);
Mel Gorman76ab0f52010-05-24 14:32:28 -07001853}
1854
1855/* The written value is actually unused, all memory is compacted */
1856int sysctl_compact_memory;
1857
Yaowei Baifec4eb22016-01-14 15:20:09 -08001858/*
1859 * This is the entry point for compacting all nodes via
1860 * /proc/sys/vm/compact_memory
1861 */
Mel Gorman76ab0f52010-05-24 14:32:28 -07001862int sysctl_compaction_handler(struct ctl_table *table, int write,
1863 void __user *buffer, size_t *length, loff_t *ppos)
1864{
1865 if (write)
Jason Liu7964c062013-01-11 14:31:47 -08001866 compact_nodes();
Mel Gorman76ab0f52010-05-24 14:32:28 -07001867
1868 return 0;
1869}
Mel Gormaned4a6d72010-05-24 14:32:29 -07001870
Mel Gorman5e771902010-05-24 14:32:31 -07001871int sysctl_extfrag_handler(struct ctl_table *table, int write,
1872 void __user *buffer, size_t *length, loff_t *ppos)
1873{
1874 proc_dointvec_minmax(table, write, buffer, length, ppos);
1875
1876 return 0;
1877}
1878
Mel Gormaned4a6d72010-05-24 14:32:29 -07001879#if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
Rashika Kheria74e77fb2014-04-03 14:48:01 -07001880static ssize_t sysfs_compact_node(struct device *dev,
Kay Sievers10fbcf42011-12-21 14:48:43 -08001881 struct device_attribute *attr,
Mel Gormaned4a6d72010-05-24 14:32:29 -07001882 const char *buf, size_t count)
1883{
Hugh Dickins8575ec22012-03-21 16:33:53 -07001884 int nid = dev->id;
1885
1886 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1887 /* Flush pending updates to the LRU lists */
1888 lru_add_drain_all();
1889
1890 compact_node(nid);
1891 }
Mel Gormaned4a6d72010-05-24 14:32:29 -07001892
1893 return count;
1894}
Kay Sievers10fbcf42011-12-21 14:48:43 -08001895static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
Mel Gormaned4a6d72010-05-24 14:32:29 -07001896
1897int compaction_register_node(struct node *node)
1898{
Kay Sievers10fbcf42011-12-21 14:48:43 -08001899 return device_create_file(&node->dev, &dev_attr_compact);
Mel Gormaned4a6d72010-05-24 14:32:29 -07001900}
1901
1902void compaction_unregister_node(struct node *node)
1903{
Kay Sievers10fbcf42011-12-21 14:48:43 -08001904 return device_remove_file(&node->dev, &dev_attr_compact);
Mel Gormaned4a6d72010-05-24 14:32:29 -07001905}
1906#endif /* CONFIG_SYSFS && CONFIG_NUMA */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001907
Vlastimil Babka698b1b32016-03-17 14:18:08 -07001908static inline bool kcompactd_work_requested(pg_data_t *pgdat)
1909{
Vlastimil Babka172400c2016-05-05 16:22:32 -07001910 return pgdat->kcompactd_max_order > 0 || kthread_should_stop();
Vlastimil Babka698b1b32016-03-17 14:18:08 -07001911}
1912
1913static bool kcompactd_node_suitable(pg_data_t *pgdat)
1914{
1915 int zoneid;
1916 struct zone *zone;
1917 enum zone_type classzone_idx = pgdat->kcompactd_classzone_idx;
1918
Chen Feng6cd9dc32016-05-20 16:59:02 -07001919 for (zoneid = 0; zoneid <= classzone_idx; zoneid++) {
Vlastimil Babka698b1b32016-03-17 14:18:08 -07001920 zone = &pgdat->node_zones[zoneid];
1921
1922 if (!populated_zone(zone))
1923 continue;
1924
1925 if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0,
1926 classzone_idx) == COMPACT_CONTINUE)
1927 return true;
1928 }
1929
1930 return false;
1931}
1932
1933static void kcompactd_do_work(pg_data_t *pgdat)
1934{
1935 /*
1936 * With no special task, compact all zones so that a page of requested
1937 * order is allocatable.
1938 */
1939 int zoneid;
1940 struct zone *zone;
1941 struct compact_control cc = {
1942 .order = pgdat->kcompactd_max_order,
1943 .classzone_idx = pgdat->kcompactd_classzone_idx,
1944 .mode = MIGRATE_SYNC_LIGHT,
1945 .ignore_skip_hint = true,
1946
1947 };
1948 bool success = false;
1949
1950 trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
1951 cc.classzone_idx);
1952 count_vm_event(KCOMPACTD_WAKE);
1953
Chen Feng6cd9dc32016-05-20 16:59:02 -07001954 for (zoneid = 0; zoneid <= cc.classzone_idx; zoneid++) {
Vlastimil Babka698b1b32016-03-17 14:18:08 -07001955 int status;
1956
1957 zone = &pgdat->node_zones[zoneid];
1958 if (!populated_zone(zone))
1959 continue;
1960
1961 if (compaction_deferred(zone, cc.order))
1962 continue;
1963
1964 if (compaction_suitable(zone, cc.order, 0, zoneid) !=
1965 COMPACT_CONTINUE)
1966 continue;
1967
1968 cc.nr_freepages = 0;
1969 cc.nr_migratepages = 0;
1970 cc.zone = zone;
1971 INIT_LIST_HEAD(&cc.freepages);
1972 INIT_LIST_HEAD(&cc.migratepages);
1973
Vlastimil Babka172400c2016-05-05 16:22:32 -07001974 if (kthread_should_stop())
1975 return;
Vlastimil Babka698b1b32016-03-17 14:18:08 -07001976 status = compact_zone(zone, &cc);
1977
1978 if (zone_watermark_ok(zone, cc.order, low_wmark_pages(zone),
1979 cc.classzone_idx, 0)) {
1980 success = true;
1981 compaction_defer_reset(zone, cc.order, false);
Michal Hockoc8f7de02016-05-20 16:56:47 -07001982 } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
Vlastimil Babka698b1b32016-03-17 14:18:08 -07001983 /*
1984 * We use sync migration mode here, so we defer like
1985 * sync direct compaction does.
1986 */
1987 defer_compaction(zone, cc.order);
1988 }
1989
1990 VM_BUG_ON(!list_empty(&cc.freepages));
1991 VM_BUG_ON(!list_empty(&cc.migratepages));
1992 }
1993
1994 /*
1995 * Regardless of success, we are done until woken up next. But remember
1996 * the requested order/classzone_idx in case it was higher/tighter than
1997 * our current ones
1998 */
1999 if (pgdat->kcompactd_max_order <= cc.order)
2000 pgdat->kcompactd_max_order = 0;
2001 if (pgdat->kcompactd_classzone_idx >= cc.classzone_idx)
2002 pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
2003}
2004
2005void wakeup_kcompactd(pg_data_t *pgdat, int order, int classzone_idx)
2006{
2007 if (!order)
2008 return;
2009
2010 if (pgdat->kcompactd_max_order < order)
2011 pgdat->kcompactd_max_order = order;
2012
2013 if (pgdat->kcompactd_classzone_idx > classzone_idx)
2014 pgdat->kcompactd_classzone_idx = classzone_idx;
2015
2016 if (!waitqueue_active(&pgdat->kcompactd_wait))
2017 return;
2018
2019 if (!kcompactd_node_suitable(pgdat))
2020 return;
2021
2022 trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
2023 classzone_idx);
2024 wake_up_interruptible(&pgdat->kcompactd_wait);
2025}
2026
2027/*
2028 * The background compaction daemon, started as a kernel thread
2029 * from the init process.
2030 */
2031static int kcompactd(void *p)
2032{
2033 pg_data_t *pgdat = (pg_data_t*)p;
2034 struct task_struct *tsk = current;
2035
2036 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2037
2038 if (!cpumask_empty(cpumask))
2039 set_cpus_allowed_ptr(tsk, cpumask);
2040
2041 set_freezable();
2042
2043 pgdat->kcompactd_max_order = 0;
2044 pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
2045
2046 while (!kthread_should_stop()) {
2047 trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
2048 wait_event_freezable(pgdat->kcompactd_wait,
2049 kcompactd_work_requested(pgdat));
2050
2051 kcompactd_do_work(pgdat);
2052 }
2053
2054 return 0;
2055}
2056
2057/*
2058 * This kcompactd start function will be called by init and node-hot-add.
2059 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
2060 */
2061int kcompactd_run(int nid)
2062{
2063 pg_data_t *pgdat = NODE_DATA(nid);
2064 int ret = 0;
2065
2066 if (pgdat->kcompactd)
2067 return 0;
2068
2069 pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
2070 if (IS_ERR(pgdat->kcompactd)) {
2071 pr_err("Failed to start kcompactd on node %d\n", nid);
2072 ret = PTR_ERR(pgdat->kcompactd);
2073 pgdat->kcompactd = NULL;
2074 }
2075 return ret;
2076}
2077
2078/*
2079 * Called by memory hotplug when all memory in a node is offlined. Caller must
2080 * hold mem_hotplug_begin/end().
2081 */
2082void kcompactd_stop(int nid)
2083{
2084 struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
2085
2086 if (kcompactd) {
2087 kthread_stop(kcompactd);
2088 NODE_DATA(nid)->kcompactd = NULL;
2089 }
2090}
2091
2092/*
2093 * It's optimal to keep kcompactd on the same CPUs as their memory, but
2094 * not required for correctness. So if the last cpu in a node goes
2095 * away, we get changed to run anywhere: as the first one comes back,
2096 * restore their cpu bindings.
2097 */
2098static int cpu_callback(struct notifier_block *nfb, unsigned long action,
2099 void *hcpu)
2100{
2101 int nid;
2102
2103 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
2104 for_each_node_state(nid, N_MEMORY) {
2105 pg_data_t *pgdat = NODE_DATA(nid);
2106 const struct cpumask *mask;
2107
2108 mask = cpumask_of_node(pgdat->node_id);
2109
2110 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2111 /* One of our CPUs online: restore mask */
2112 set_cpus_allowed_ptr(pgdat->kcompactd, mask);
2113 }
2114 }
2115 return NOTIFY_OK;
2116}
2117
2118static int __init kcompactd_init(void)
2119{
2120 int nid;
2121
2122 for_each_node_state(nid, N_MEMORY)
2123 kcompactd_run(nid);
2124 hotcpu_notifier(cpu_callback, 0);
2125 return 0;
2126}
2127subsys_initcall(kcompactd_init)
2128
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01002129#endif /* CONFIG_COMPACTION */