blob: f9792ba3537ccc830594e7954715ca66eb2e9654 [file] [log] [blame]
Mel Gorman748446b2010-05-24 14:32:27 -07001/*
2 * linux/mm/compaction.c
3 *
4 * Memory compaction for the reduction of external fragmentation. Note that
5 * this heavily depends upon page migration to do all the real heavy
6 * lifting
7 *
8 * Copyright IBM Corp. 2007-2010 Mel Gorman <mel@csn.ul.ie>
9 */
10#include <linux/swap.h>
11#include <linux/migrate.h>
12#include <linux/compaction.h>
13#include <linux/mm_inline.h>
14#include <linux/backing-dev.h>
Mel Gorman76ab0f52010-05-24 14:32:28 -070015#include <linux/sysctl.h>
Mel Gormaned4a6d72010-05-24 14:32:29 -070016#include <linux/sysfs.h>
Rafael Aquinibf6bddf2012-12-11 16:02:42 -080017#include <linux/balloon_compaction.h>
Minchan Kim194159f2013-02-22 16:33:58 -080018#include <linux/page-isolation.h>
Mel Gorman748446b2010-05-24 14:32:27 -070019#include "internal.h"
20
Minchan Kim010fc292012-12-20 15:05:06 -080021#ifdef CONFIG_COMPACTION
22static inline void count_compact_event(enum vm_event_item item)
23{
24 count_vm_event(item);
25}
26
27static inline void count_compact_events(enum vm_event_item item, long delta)
28{
29 count_vm_events(item, delta);
30}
31#else
32#define count_compact_event(item) do { } while (0)
33#define count_compact_events(item, delta) do { } while (0)
34#endif
35
Michal Nazarewiczff9543f2011-12-29 13:09:50 +010036#if defined CONFIG_COMPACTION || defined CONFIG_CMA
37
Mel Gormanb7aba692011-01-13 15:45:54 -080038#define CREATE_TRACE_POINTS
39#include <trace/events/compaction.h>
40
Mel Gorman748446b2010-05-24 14:32:27 -070041static unsigned long release_freepages(struct list_head *freelist)
42{
43 struct page *page, *next;
44 unsigned long count = 0;
45
46 list_for_each_entry_safe(page, next, freelist, lru) {
47 list_del(&page->lru);
48 __free_page(page);
49 count++;
50 }
51
52 return count;
53}
54
Michal Nazarewiczff9543f2011-12-29 13:09:50 +010055static void map_pages(struct list_head *list)
56{
57 struct page *page;
58
59 list_for_each_entry(page, list, lru) {
60 arch_alloc_page(page, 0);
61 kernel_map_pages(page, 1, 1);
62 }
63}
64
Michal Nazarewicz47118af2011-12-29 13:09:50 +010065static inline bool migrate_async_suitable(int migratetype)
66{
67 return is_migrate_cma(migratetype) || migratetype == MIGRATE_MOVABLE;
68}
69
Vlastimil Babka7d49d882014-10-09 15:27:11 -070070/*
71 * Check that the whole (or subset of) a pageblock given by the interval of
72 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
73 * with the migration of free compaction scanner. The scanners then need to
74 * use only pfn_valid_within() check for arches that allow holes within
75 * pageblocks.
76 *
77 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
78 *
79 * It's possible on some configurations to have a setup like node0 node1 node0
80 * i.e. it's possible that all pages within a zones range of pages do not
81 * belong to a single zone. We assume that a border between node0 and node1
82 * can occur within a single pageblock, but not a node0 node1 node0
83 * interleaving within a single pageblock. It is therefore sufficient to check
84 * the first and last page of a pageblock and avoid checking each individual
85 * page in a pageblock.
86 */
87static struct page *pageblock_pfn_to_page(unsigned long start_pfn,
88 unsigned long end_pfn, struct zone *zone)
89{
90 struct page *start_page;
91 struct page *end_page;
92
93 /* end_pfn is one past the range we are checking */
94 end_pfn--;
95
96 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
97 return NULL;
98
99 start_page = pfn_to_page(start_pfn);
100
101 if (page_zone(start_page) != zone)
102 return NULL;
103
104 end_page = pfn_to_page(end_pfn);
105
106 /* This gives a shorter code than deriving page_zone(end_page) */
107 if (page_zone_id(start_page) != page_zone_id(end_page))
108 return NULL;
109
110 return start_page;
111}
112
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700113#ifdef CONFIG_COMPACTION
114/* Returns true if the pageblock should be scanned for pages to isolate. */
115static inline bool isolation_suitable(struct compact_control *cc,
116 struct page *page)
117{
118 if (cc->ignore_skip_hint)
119 return true;
120
121 return !get_pageblock_skip(page);
122}
123
124/*
125 * This function is called to clear all cached information on pageblocks that
126 * should be skipped for page isolation when the migrate and free page scanner
127 * meet.
128 */
Mel Gorman62997022012-10-08 16:32:47 -0700129static void __reset_isolation_suitable(struct zone *zone)
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700130{
131 unsigned long start_pfn = zone->zone_start_pfn;
Cody P Schafer108bcc92013-02-22 16:35:23 -0800132 unsigned long end_pfn = zone_end_pfn(zone);
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700133 unsigned long pfn;
134
David Rientjes35979ef2014-06-04 16:08:27 -0700135 zone->compact_cached_migrate_pfn[0] = start_pfn;
136 zone->compact_cached_migrate_pfn[1] = start_pfn;
Mel Gormanc89511a2012-10-08 16:32:45 -0700137 zone->compact_cached_free_pfn = end_pfn;
Mel Gorman62997022012-10-08 16:32:47 -0700138 zone->compact_blockskip_flush = false;
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700139
140 /* Walk the zone and mark every pageblock as suitable for isolation */
141 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
142 struct page *page;
143
144 cond_resched();
145
146 if (!pfn_valid(pfn))
147 continue;
148
149 page = pfn_to_page(pfn);
150 if (zone != page_zone(page))
151 continue;
152
153 clear_pageblock_skip(page);
154 }
155}
156
Mel Gorman62997022012-10-08 16:32:47 -0700157void reset_isolation_suitable(pg_data_t *pgdat)
158{
159 int zoneid;
160
161 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
162 struct zone *zone = &pgdat->node_zones[zoneid];
163 if (!populated_zone(zone))
164 continue;
165
166 /* Only flush if a full compaction finished recently */
167 if (zone->compact_blockskip_flush)
168 __reset_isolation_suitable(zone);
169 }
170}
171
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700172/*
173 * If no pages were isolated then mark this pageblock to be skipped in the
Mel Gorman62997022012-10-08 16:32:47 -0700174 * future. The information is later cleared by __reset_isolation_suitable().
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700175 */
Mel Gormanc89511a2012-10-08 16:32:45 -0700176static void update_pageblock_skip(struct compact_control *cc,
177 struct page *page, unsigned long nr_isolated,
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700178 bool migrate_scanner)
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700179{
Mel Gormanc89511a2012-10-08 16:32:45 -0700180 struct zone *zone = cc->zone;
David Rientjes35979ef2014-06-04 16:08:27 -0700181 unsigned long pfn;
Joonsoo Kim6815bf32013-12-18 17:08:52 -0800182
183 if (cc->ignore_skip_hint)
184 return;
185
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700186 if (!page)
187 return;
188
David Rientjes35979ef2014-06-04 16:08:27 -0700189 if (nr_isolated)
190 return;
191
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700192 set_pageblock_skip(page);
Mel Gormanc89511a2012-10-08 16:32:45 -0700193
David Rientjes35979ef2014-06-04 16:08:27 -0700194 pfn = page_to_pfn(page);
195
196 /* Update where async and sync compaction should restart */
197 if (migrate_scanner) {
198 if (cc->finished_update_migrate)
199 return;
200 if (pfn > zone->compact_cached_migrate_pfn[0])
201 zone->compact_cached_migrate_pfn[0] = pfn;
David Rientjese0b9dae2014-06-04 16:08:28 -0700202 if (cc->mode != MIGRATE_ASYNC &&
203 pfn > zone->compact_cached_migrate_pfn[1])
David Rientjes35979ef2014-06-04 16:08:27 -0700204 zone->compact_cached_migrate_pfn[1] = pfn;
205 } else {
206 if (cc->finished_update_free)
207 return;
208 if (pfn < zone->compact_cached_free_pfn)
209 zone->compact_cached_free_pfn = pfn;
Mel Gormanc89511a2012-10-08 16:32:45 -0700210 }
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700211}
212#else
213static inline bool isolation_suitable(struct compact_control *cc,
214 struct page *page)
215{
216 return true;
217}
218
Mel Gormanc89511a2012-10-08 16:32:45 -0700219static void update_pageblock_skip(struct compact_control *cc,
220 struct page *page, unsigned long nr_isolated,
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700221 bool migrate_scanner)
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700222{
223}
224#endif /* CONFIG_COMPACTION */
225
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700226/*
227 * Compaction requires the taking of some coarse locks that are potentially
228 * very heavily contended. For async compaction, back out if the lock cannot
229 * be taken immediately. For sync compaction, spin on the lock if needed.
230 *
231 * Returns true if the lock is held
232 * Returns false if the lock is not held and compaction should abort
233 */
234static bool compact_trylock_irqsave(spinlock_t *lock, unsigned long *flags,
235 struct compact_control *cc)
Mel Gorman2a1402a2012-10-08 16:32:33 -0700236{
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700237 if (cc->mode == MIGRATE_ASYNC) {
238 if (!spin_trylock_irqsave(lock, *flags)) {
239 cc->contended = COMPACT_CONTENDED_LOCK;
240 return false;
241 }
242 } else {
243 spin_lock_irqsave(lock, *flags);
244 }
Vlastimil Babka1f9efde2014-10-09 15:27:14 -0700245
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700246 return true;
Mel Gorman2a1402a2012-10-08 16:32:33 -0700247}
248
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100249/*
Mel Gormanc67fe372012-08-21 16:16:17 -0700250 * Compaction requires the taking of some coarse locks that are potentially
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700251 * very heavily contended. The lock should be periodically unlocked to avoid
252 * having disabled IRQs for a long time, even when there is nobody waiting on
253 * the lock. It might also be that allowing the IRQs will result in
254 * need_resched() becoming true. If scheduling is needed, async compaction
255 * aborts. Sync compaction schedules.
256 * Either compaction type will also abort if a fatal signal is pending.
257 * In either case if the lock was locked, it is dropped and not regained.
Mel Gormanc67fe372012-08-21 16:16:17 -0700258 *
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700259 * Returns true if compaction should abort due to fatal signal pending, or
260 * async compaction due to need_resched()
261 * Returns false when compaction can continue (sync compaction might have
262 * scheduled)
Mel Gormanc67fe372012-08-21 16:16:17 -0700263 */
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700264static bool compact_unlock_should_abort(spinlock_t *lock,
265 unsigned long flags, bool *locked, struct compact_control *cc)
Mel Gormanc67fe372012-08-21 16:16:17 -0700266{
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700267 if (*locked) {
268 spin_unlock_irqrestore(lock, flags);
269 *locked = false;
270 }
Vlastimil Babka1f9efde2014-10-09 15:27:14 -0700271
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700272 if (fatal_signal_pending(current)) {
273 cc->contended = COMPACT_CONTENDED_SCHED;
274 return true;
275 }
Mel Gormanc67fe372012-08-21 16:16:17 -0700276
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700277 if (need_resched()) {
David Rientjese0b9dae2014-06-04 16:08:28 -0700278 if (cc->mode == MIGRATE_ASYNC) {
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700279 cc->contended = COMPACT_CONTENDED_SCHED;
280 return true;
Mel Gormanc67fe372012-08-21 16:16:17 -0700281 }
Mel Gormanc67fe372012-08-21 16:16:17 -0700282 cond_resched();
Mel Gormanc67fe372012-08-21 16:16:17 -0700283 }
284
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700285 return false;
Mel Gormanc67fe372012-08-21 16:16:17 -0700286}
287
Vlastimil Babkabe976572014-06-04 16:10:41 -0700288/*
289 * Aside from avoiding lock contention, compaction also periodically checks
290 * need_resched() and either schedules in sync compaction or aborts async
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700291 * compaction. This is similar to what compact_unlock_should_abort() does, but
Vlastimil Babkabe976572014-06-04 16:10:41 -0700292 * is used where no lock is concerned.
293 *
294 * Returns false when no scheduling was needed, or sync compaction scheduled.
295 * Returns true when async compaction should abort.
296 */
297static inline bool compact_should_abort(struct compact_control *cc)
298{
299 /* async compaction aborts if contended */
300 if (need_resched()) {
301 if (cc->mode == MIGRATE_ASYNC) {
Vlastimil Babka1f9efde2014-10-09 15:27:14 -0700302 cc->contended = COMPACT_CONTENDED_SCHED;
Vlastimil Babkabe976572014-06-04 16:10:41 -0700303 return true;
304 }
305
306 cond_resched();
307 }
308
309 return false;
310}
311
Mel Gormanf40d1e42012-10-08 16:32:36 -0700312/* Returns true if the page is within a block suitable for migration to */
313static bool suitable_migration_target(struct page *page)
314{
Joonsoo Kim7d348b92014-04-07 15:37:03 -0700315 /* If the page is a large free page, then disallow migration */
Vlastimil Babka99c0fd52014-10-09 15:27:23 -0700316 if (PageBuddy(page)) {
317 /*
318 * We are checking page_order without zone->lock taken. But
319 * the only small danger is that we skip a potentially suitable
320 * pageblock, so it's not worth to check order for valid range.
321 */
322 if (page_order_unsafe(page) >= pageblock_order)
323 return false;
324 }
Mel Gormanf40d1e42012-10-08 16:32:36 -0700325
326 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
Joonsoo Kim7d348b92014-04-07 15:37:03 -0700327 if (migrate_async_suitable(get_pageblock_migratetype(page)))
Mel Gormanf40d1e42012-10-08 16:32:36 -0700328 return true;
329
330 /* Otherwise skip the block */
331 return false;
332}
333
Mel Gormanc67fe372012-08-21 16:16:17 -0700334/*
Jerome Marchand9e4be472013-11-12 15:07:12 -0800335 * Isolate free pages onto a private freelist. If @strict is true, will abort
336 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
337 * (even though it may still end up isolating some pages).
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100338 */
Mel Gormanf40d1e42012-10-08 16:32:36 -0700339static unsigned long isolate_freepages_block(struct compact_control *cc,
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700340 unsigned long *start_pfn,
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100341 unsigned long end_pfn,
342 struct list_head *freelist,
343 bool strict)
Mel Gorman748446b2010-05-24 14:32:27 -0700344{
Mel Gormanb7aba692011-01-13 15:45:54 -0800345 int nr_scanned = 0, total_isolated = 0;
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700346 struct page *cursor, *valid_page = NULL;
Xiubo Lib8b2d822014-10-09 15:28:21 -0700347 unsigned long flags = 0;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700348 bool locked = false;
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700349 unsigned long blockpfn = *start_pfn;
Mel Gorman748446b2010-05-24 14:32:27 -0700350
Mel Gorman748446b2010-05-24 14:32:27 -0700351 cursor = pfn_to_page(blockpfn);
352
Mel Gormanf40d1e42012-10-08 16:32:36 -0700353 /* Isolate free pages. */
Mel Gorman748446b2010-05-24 14:32:27 -0700354 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
355 int isolated, i;
356 struct page *page = cursor;
357
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700358 /*
359 * Periodically drop the lock (if held) regardless of its
360 * contention, to give chance to IRQs. Abort if fatal signal
361 * pending or async compaction detects need_resched()
362 */
363 if (!(blockpfn % SWAP_CLUSTER_MAX)
364 && compact_unlock_should_abort(&cc->zone->lock, flags,
365 &locked, cc))
366 break;
367
Mel Gormanb7aba692011-01-13 15:45:54 -0800368 nr_scanned++;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700369 if (!pfn_valid_within(blockpfn))
Laura Abbott2af120b2014-03-10 15:49:44 -0700370 goto isolate_fail;
371
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700372 if (!valid_page)
373 valid_page = page;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700374 if (!PageBuddy(page))
Laura Abbott2af120b2014-03-10 15:49:44 -0700375 goto isolate_fail;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700376
377 /*
Vlastimil Babka69b71892014-10-09 15:27:18 -0700378 * If we already hold the lock, we can skip some rechecking.
379 * Note that if we hold the lock now, checked_pageblock was
380 * already set in some previous iteration (or strict is true),
381 * so it is correct to skip the suitable migration target
382 * recheck as well.
Mel Gormanf40d1e42012-10-08 16:32:36 -0700383 */
Vlastimil Babka69b71892014-10-09 15:27:18 -0700384 if (!locked) {
385 /*
386 * The zone lock must be held to isolate freepages.
387 * Unfortunately this is a very coarse lock and can be
388 * heavily contended if there are parallel allocations
389 * or parallel compactions. For async compaction do not
390 * spin on the lock and we acquire the lock as late as
391 * possible.
392 */
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700393 locked = compact_trylock_irqsave(&cc->zone->lock,
394 &flags, cc);
Vlastimil Babka69b71892014-10-09 15:27:18 -0700395 if (!locked)
396 break;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700397
Vlastimil Babka69b71892014-10-09 15:27:18 -0700398 /* Recheck this is a buddy page under lock */
399 if (!PageBuddy(page))
400 goto isolate_fail;
401 }
Mel Gorman748446b2010-05-24 14:32:27 -0700402
403 /* Found a free page, break it into order-0 pages */
404 isolated = split_free_page(page);
405 total_isolated += isolated;
406 for (i = 0; i < isolated; i++) {
407 list_add(&page->lru, freelist);
408 page++;
409 }
410
411 /* If a page was split, advance to the end of it */
412 if (isolated) {
413 blockpfn += isolated - 1;
414 cursor += isolated - 1;
Laura Abbott2af120b2014-03-10 15:49:44 -0700415 continue;
Mel Gorman748446b2010-05-24 14:32:27 -0700416 }
Laura Abbott2af120b2014-03-10 15:49:44 -0700417
418isolate_fail:
419 if (strict)
420 break;
421 else
422 continue;
423
Mel Gorman748446b2010-05-24 14:32:27 -0700424 }
425
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700426 /* Record how far we have got within the block */
427 *start_pfn = blockpfn;
428
Mel Gormanb7aba692011-01-13 15:45:54 -0800429 trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
Mel Gormanf40d1e42012-10-08 16:32:36 -0700430
431 /*
432 * If strict isolation is requested by CMA then check that all the
433 * pages requested were isolated. If there were any failures, 0 is
434 * returned and CMA will fail.
435 */
Laura Abbott2af120b2014-03-10 15:49:44 -0700436 if (strict && blockpfn < end_pfn)
Mel Gormanf40d1e42012-10-08 16:32:36 -0700437 total_isolated = 0;
438
439 if (locked)
440 spin_unlock_irqrestore(&cc->zone->lock, flags);
441
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700442 /* Update the pageblock-skip if the whole pageblock was scanned */
443 if (blockpfn == end_pfn)
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700444 update_pageblock_skip(cc, valid_page, total_isolated, false);
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700445
Minchan Kim010fc292012-12-20 15:05:06 -0800446 count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
Mel Gorman397487d2012-10-19 12:00:10 +0100447 if (total_isolated)
Minchan Kim010fc292012-12-20 15:05:06 -0800448 count_compact_events(COMPACTISOLATED, total_isolated);
Mel Gorman748446b2010-05-24 14:32:27 -0700449 return total_isolated;
450}
451
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100452/**
453 * isolate_freepages_range() - isolate free pages.
454 * @start_pfn: The first PFN to start isolating.
455 * @end_pfn: The one-past-last PFN.
456 *
457 * Non-free pages, invalid PFNs, or zone boundaries within the
458 * [start_pfn, end_pfn) range are considered errors, cause function to
459 * undo its actions and return zero.
460 *
461 * Otherwise, function returns one-past-the-last PFN of isolated page
462 * (which may be greater then end_pfn if end fell in a middle of
463 * a free page).
464 */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100465unsigned long
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700466isolate_freepages_range(struct compact_control *cc,
467 unsigned long start_pfn, unsigned long end_pfn)
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100468{
Mel Gormanf40d1e42012-10-08 16:32:36 -0700469 unsigned long isolated, pfn, block_end_pfn;
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100470 LIST_HEAD(freelist);
471
Vlastimil Babka7d49d882014-10-09 15:27:11 -0700472 pfn = start_pfn;
473 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100474
Vlastimil Babka7d49d882014-10-09 15:27:11 -0700475 for (; pfn < end_pfn; pfn += isolated,
476 block_end_pfn += pageblock_nr_pages) {
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700477 /* Protect pfn from changing by isolate_freepages_block */
478 unsigned long isolate_start_pfn = pfn;
Vlastimil Babka7d49d882014-10-09 15:27:11 -0700479
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100480 block_end_pfn = min(block_end_pfn, end_pfn);
481
Joonsoo Kim58420012014-11-13 15:19:07 -0800482 /*
483 * pfn could pass the block_end_pfn if isolated freepage
484 * is more than pageblock order. In this case, we adjust
485 * scanning range to right one.
486 */
487 if (pfn >= block_end_pfn) {
488 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
489 block_end_pfn = min(block_end_pfn, end_pfn);
490 }
491
Vlastimil Babka7d49d882014-10-09 15:27:11 -0700492 if (!pageblock_pfn_to_page(pfn, block_end_pfn, cc->zone))
493 break;
494
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700495 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
496 block_end_pfn, &freelist, true);
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100497
498 /*
499 * In strict mode, isolate_freepages_block() returns 0 if
500 * there are any holes in the block (ie. invalid PFNs or
501 * non-free pages).
502 */
503 if (!isolated)
504 break;
505
506 /*
507 * If we managed to isolate pages, it is always (1 << n) *
508 * pageblock_nr_pages for some non-negative n. (Max order
509 * page may span two pageblocks).
510 */
511 }
512
513 /* split_free_page does not map the pages */
514 map_pages(&freelist);
515
516 if (pfn < end_pfn) {
517 /* Loop terminated early, cleanup. */
518 release_freepages(&freelist);
519 return 0;
520 }
521
522 /* We don't use freelists for anything. */
523 return pfn;
524}
525
Mel Gorman748446b2010-05-24 14:32:27 -0700526/* Update the number of anon and file isolated pages in the zone */
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700527static void acct_isolated(struct zone *zone, struct compact_control *cc)
Mel Gorman748446b2010-05-24 14:32:27 -0700528{
529 struct page *page;
Minchan Kimb9e84ac2011-10-31 17:06:44 -0700530 unsigned int count[2] = { 0, };
Mel Gorman748446b2010-05-24 14:32:27 -0700531
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700532 if (list_empty(&cc->migratepages))
533 return;
534
Minchan Kimb9e84ac2011-10-31 17:06:44 -0700535 list_for_each_entry(page, &cc->migratepages, lru)
536 count[!!page_is_file_cache(page)]++;
Mel Gorman748446b2010-05-24 14:32:27 -0700537
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700538 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
539 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
Mel Gorman748446b2010-05-24 14:32:27 -0700540}
541
542/* Similar to reclaim, but different enough that they don't share logic */
543static bool too_many_isolated(struct zone *zone)
544{
Minchan Kimbc693042010-09-09 16:38:00 -0700545 unsigned long active, inactive, isolated;
Mel Gorman748446b2010-05-24 14:32:27 -0700546
547 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
548 zone_page_state(zone, NR_INACTIVE_ANON);
Minchan Kimbc693042010-09-09 16:38:00 -0700549 active = zone_page_state(zone, NR_ACTIVE_FILE) +
550 zone_page_state(zone, NR_ACTIVE_ANON);
Mel Gorman748446b2010-05-24 14:32:27 -0700551 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
552 zone_page_state(zone, NR_ISOLATED_ANON);
553
Minchan Kimbc693042010-09-09 16:38:00 -0700554 return isolated > (inactive + active) / 2;
Mel Gorman748446b2010-05-24 14:32:27 -0700555}
556
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100557/**
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700558 * isolate_migratepages_block() - isolate all migrate-able pages within
559 * a single pageblock
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100560 * @cc: Compaction control structure.
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700561 * @low_pfn: The first PFN to isolate
562 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
563 * @isolate_mode: Isolation mode to be used.
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100564 *
565 * Isolate all pages that can be migrated from the range specified by
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700566 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
567 * Returns zero if there is a fatal signal pending, otherwise PFN of the
568 * first page that was not scanned (which may be both less, equal to or more
569 * than end_pfn).
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100570 *
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700571 * The pages are isolated on cc->migratepages list (not required to be empty),
572 * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
573 * is neither read nor updated.
Mel Gorman748446b2010-05-24 14:32:27 -0700574 */
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700575static unsigned long
576isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
577 unsigned long end_pfn, isolate_mode_t isolate_mode)
Mel Gorman748446b2010-05-24 14:32:27 -0700578{
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700579 struct zone *zone = cc->zone;
Mel Gormanb7aba692011-01-13 15:45:54 -0800580 unsigned long nr_scanned = 0, nr_isolated = 0;
Mel Gorman748446b2010-05-24 14:32:27 -0700581 struct list_head *migratelist = &cc->migratepages;
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700582 struct lruvec *lruvec;
Xiubo Lib8b2d822014-10-09 15:28:21 -0700583 unsigned long flags = 0;
Mel Gorman2a1402a2012-10-08 16:32:33 -0700584 bool locked = false;
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700585 struct page *page = NULL, *valid_page = NULL;
Mel Gorman748446b2010-05-24 14:32:27 -0700586
Mel Gorman748446b2010-05-24 14:32:27 -0700587 /*
588 * Ensure that there are not too many pages isolated from the LRU
589 * list by either parallel reclaimers or compaction. If there are,
590 * delay for some time until fewer pages are isolated
591 */
592 while (unlikely(too_many_isolated(zone))) {
Mel Gormanf9e35b32011-06-15 15:08:52 -0700593 /* async migration should just abort */
David Rientjese0b9dae2014-06-04 16:08:28 -0700594 if (cc->mode == MIGRATE_ASYNC)
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100595 return 0;
Mel Gormanf9e35b32011-06-15 15:08:52 -0700596
Mel Gorman748446b2010-05-24 14:32:27 -0700597 congestion_wait(BLK_RW_ASYNC, HZ/10);
598
599 if (fatal_signal_pending(current))
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100600 return 0;
Mel Gorman748446b2010-05-24 14:32:27 -0700601 }
602
Vlastimil Babkabe976572014-06-04 16:10:41 -0700603 if (compact_should_abort(cc))
604 return 0;
David Rientjesaeef4b82014-06-04 16:08:31 -0700605
Mel Gorman748446b2010-05-24 14:32:27 -0700606 /* Time to isolate some pages for migration */
Mel Gorman748446b2010-05-24 14:32:27 -0700607 for (; low_pfn < end_pfn; low_pfn++) {
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700608 /*
609 * Periodically drop the lock (if held) regardless of its
610 * contention, to give chance to IRQs. Abort async compaction
611 * if contended.
612 */
613 if (!(low_pfn % SWAP_CLUSTER_MAX)
614 && compact_unlock_should_abort(&zone->lru_lock, flags,
615 &locked, cc))
616 break;
Mel Gormanc67fe372012-08-21 16:16:17 -0700617
Mel Gorman748446b2010-05-24 14:32:27 -0700618 if (!pfn_valid_within(low_pfn))
619 continue;
Mel Gormanb7aba692011-01-13 15:45:54 -0800620 nr_scanned++;
Mel Gorman748446b2010-05-24 14:32:27 -0700621
Mel Gorman748446b2010-05-24 14:32:27 -0700622 page = pfn_to_page(low_pfn);
Mel Gormandc908602012-02-08 17:13:38 -0800623
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700624 if (!valid_page)
625 valid_page = page;
626
Mel Gorman6c144662014-01-23 15:53:38 -0800627 /*
Vlastimil Babka99c0fd52014-10-09 15:27:23 -0700628 * Skip if free. We read page order here without zone lock
629 * which is generally unsafe, but the race window is small and
630 * the worst thing that can happen is that we skip some
631 * potential isolation targets.
Mel Gorman6c144662014-01-23 15:53:38 -0800632 */
Vlastimil Babka99c0fd52014-10-09 15:27:23 -0700633 if (PageBuddy(page)) {
634 unsigned long freepage_order = page_order_unsafe(page);
635
636 /*
637 * Without lock, we cannot be sure that what we got is
638 * a valid page order. Consider only values in the
639 * valid order range to prevent low_pfn overflow.
640 */
641 if (freepage_order > 0 && freepage_order < MAX_ORDER)
642 low_pfn += (1UL << freepage_order) - 1;
Mel Gorman748446b2010-05-24 14:32:27 -0700643 continue;
Vlastimil Babka99c0fd52014-10-09 15:27:23 -0700644 }
Mel Gorman748446b2010-05-24 14:32:27 -0700645
Mel Gorman9927af742011-01-13 15:45:59 -0800646 /*
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800647 * Check may be lockless but that's ok as we recheck later.
648 * It's possible to migrate LRU pages and balloon pages
649 * Skip any other type of page
650 */
651 if (!PageLRU(page)) {
652 if (unlikely(balloon_page_movable(page))) {
Konstantin Khlebnikovd6d86c02014-10-09 15:29:27 -0700653 if (balloon_page_isolate(page)) {
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800654 /* Successfully isolated */
Joonsoo Kimb6c75012014-04-07 15:37:07 -0700655 goto isolate_success;
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800656 }
657 }
Andrea Arcangelibc835012011-01-13 15:47:08 -0800658 continue;
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800659 }
Andrea Arcangelibc835012011-01-13 15:47:08 -0800660
661 /*
Mel Gorman2a1402a2012-10-08 16:32:33 -0700662 * PageLRU is set. lru_lock normally excludes isolation
663 * splitting and collapsing (collapsing has already happened
664 * if PageLRU is set) but the lock is not necessarily taken
665 * here and it is wasteful to take it just to check transhuge.
666 * Check TransHuge without lock and skip the whole pageblock if
667 * it's either a transhuge or hugetlbfs page, as calling
668 * compound_order() without preventing THP from splitting the
669 * page underneath us may return surprising results.
Andrea Arcangelibc835012011-01-13 15:47:08 -0800670 */
671 if (PageTransHuge(page)) {
Mel Gorman2a1402a2012-10-08 16:32:33 -0700672 if (!locked)
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700673 low_pfn = ALIGN(low_pfn + 1,
674 pageblock_nr_pages) - 1;
675 else
676 low_pfn += (1 << compound_order(page)) - 1;
677
Mel Gorman2a1402a2012-10-08 16:32:33 -0700678 continue;
679 }
680
David Rientjes119d6d52014-04-03 14:48:00 -0700681 /*
682 * Migration will fail if an anonymous page is pinned in memory,
683 * so avoid taking lru_lock and isolating it unnecessarily in an
684 * admittedly racy check.
685 */
686 if (!page_mapping(page) &&
687 page_count(page) > page_mapcount(page))
688 continue;
689
Vlastimil Babka69b71892014-10-09 15:27:18 -0700690 /* If we already hold the lock, we can skip some rechecking */
691 if (!locked) {
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700692 locked = compact_trylock_irqsave(&zone->lru_lock,
693 &flags, cc);
Vlastimil Babka69b71892014-10-09 15:27:18 -0700694 if (!locked)
695 break;
Mel Gorman2a1402a2012-10-08 16:32:33 -0700696
Vlastimil Babka69b71892014-10-09 15:27:18 -0700697 /* Recheck PageLRU and PageTransHuge under lock */
698 if (!PageLRU(page))
699 continue;
700 if (PageTransHuge(page)) {
701 low_pfn += (1 << compound_order(page)) - 1;
702 continue;
703 }
Andrea Arcangelibc835012011-01-13 15:47:08 -0800704 }
705
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700706 lruvec = mem_cgroup_page_lruvec(page, zone);
707
Mel Gorman748446b2010-05-24 14:32:27 -0700708 /* Try isolate the page */
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700709 if (__isolate_lru_page(page, isolate_mode) != 0)
Mel Gorman748446b2010-05-24 14:32:27 -0700710 continue;
711
Sasha Levin309381fea2014-01-23 15:52:54 -0800712 VM_BUG_ON_PAGE(PageTransCompound(page), page);
Andrea Arcangelibc835012011-01-13 15:47:08 -0800713
Mel Gorman748446b2010-05-24 14:32:27 -0700714 /* Successfully isolated */
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700715 del_page_from_lru_list(page, lruvec, page_lru(page));
Joonsoo Kimb6c75012014-04-07 15:37:07 -0700716
717isolate_success:
718 cc->finished_update_migrate = true;
Mel Gorman748446b2010-05-24 14:32:27 -0700719 list_add(&page->lru, migratelist);
Mel Gorman748446b2010-05-24 14:32:27 -0700720 cc->nr_migratepages++;
Mel Gormanb7aba692011-01-13 15:45:54 -0800721 nr_isolated++;
Mel Gorman748446b2010-05-24 14:32:27 -0700722
723 /* Avoid isolating too much */
Hillf Danton31b83842012-01-10 15:07:59 -0800724 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
725 ++low_pfn;
Mel Gorman748446b2010-05-24 14:32:27 -0700726 break;
Hillf Danton31b83842012-01-10 15:07:59 -0800727 }
Mel Gorman748446b2010-05-24 14:32:27 -0700728 }
729
Vlastimil Babka99c0fd52014-10-09 15:27:23 -0700730 /*
731 * The PageBuddy() check could have potentially brought us outside
732 * the range to be scanned.
733 */
734 if (unlikely(low_pfn > end_pfn))
735 low_pfn = end_pfn;
736
Mel Gormanc67fe372012-08-21 16:16:17 -0700737 if (locked)
738 spin_unlock_irqrestore(&zone->lru_lock, flags);
Mel Gorman748446b2010-05-24 14:32:27 -0700739
Vlastimil Babka50b5b092014-01-21 15:51:10 -0800740 /*
741 * Update the pageblock-skip information and cached scanner pfn,
742 * if the whole pageblock was scanned without isolating any page.
Vlastimil Babka50b5b092014-01-21 15:51:10 -0800743 */
David Rientjes35979ef2014-06-04 16:08:27 -0700744 if (low_pfn == end_pfn)
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700745 update_pageblock_skip(cc, valid_page, nr_isolated, true);
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700746
Mel Gormanb7aba692011-01-13 15:45:54 -0800747 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
748
Minchan Kim010fc292012-12-20 15:05:06 -0800749 count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
Mel Gorman397487d2012-10-19 12:00:10 +0100750 if (nr_isolated)
Minchan Kim010fc292012-12-20 15:05:06 -0800751 count_compact_events(COMPACTISOLATED, nr_isolated);
Mel Gorman397487d2012-10-19 12:00:10 +0100752
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100753 return low_pfn;
754}
755
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700756/**
757 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
758 * @cc: Compaction control structure.
759 * @start_pfn: The first PFN to start isolating.
760 * @end_pfn: The one-past-last PFN.
761 *
762 * Returns zero if isolation fails fatally due to e.g. pending signal.
763 * Otherwise, function returns one-past-the-last PFN of isolated page
764 * (which may be greater than end_pfn if end fell in a middle of a THP page).
765 */
766unsigned long
767isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
768 unsigned long end_pfn)
769{
770 unsigned long pfn, block_end_pfn;
771
772 /* Scan block by block. First and last block may be incomplete */
773 pfn = start_pfn;
774 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
775
776 for (; pfn < end_pfn; pfn = block_end_pfn,
777 block_end_pfn += pageblock_nr_pages) {
778
779 block_end_pfn = min(block_end_pfn, end_pfn);
780
Vlastimil Babka7d49d882014-10-09 15:27:11 -0700781 if (!pageblock_pfn_to_page(pfn, block_end_pfn, cc->zone))
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700782 continue;
783
784 pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
785 ISOLATE_UNEVICTABLE);
786
787 /*
788 * In case of fatal failure, release everything that might
789 * have been isolated in the previous iteration, and signal
790 * the failure back to caller.
791 */
792 if (!pfn) {
793 putback_movable_pages(&cc->migratepages);
794 cc->nr_migratepages = 0;
795 break;
796 }
Joonsoo Kim6ea41c02014-10-29 14:50:20 -0700797
798 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
799 break;
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700800 }
801 acct_isolated(cc->zone, cc);
802
803 return pfn;
804}
805
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100806#endif /* CONFIG_COMPACTION || CONFIG_CMA */
807#ifdef CONFIG_COMPACTION
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100808/*
809 * Based on information in the current compact_control, find blocks
810 * suitable for isolating free pages from and then isolate them.
811 */
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700812static void isolate_freepages(struct compact_control *cc)
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100813{
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700814 struct zone *zone = cc->zone;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100815 struct page *page;
Vlastimil Babkac96b9e52014-06-04 16:07:26 -0700816 unsigned long block_start_pfn; /* start of current pageblock */
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700817 unsigned long isolate_start_pfn; /* exact pfn we start at */
Vlastimil Babkac96b9e52014-06-04 16:07:26 -0700818 unsigned long block_end_pfn; /* end of current pageblock */
819 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100820 int nr_freepages = cc->nr_freepages;
821 struct list_head *freelist = &cc->freepages;
822
823 /*
824 * Initialise the free scanner. The starting point is where we last
Vlastimil Babka49e068f2014-05-06 12:50:03 -0700825 * successfully isolated from, zone-cached value, or the end of the
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700826 * zone when isolating for the first time. For looping we also need
827 * this pfn aligned down to the pageblock boundary, because we do
Vlastimil Babkac96b9e52014-06-04 16:07:26 -0700828 * block_start_pfn -= pageblock_nr_pages in the for loop.
829 * For ending point, take care when isolating in last pageblock of a
830 * a zone which ends in the middle of a pageblock.
Vlastimil Babka49e068f2014-05-06 12:50:03 -0700831 * The low boundary is the end of the pageblock the migration scanner
832 * is using.
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100833 */
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700834 isolate_start_pfn = cc->free_pfn;
Vlastimil Babkac96b9e52014-06-04 16:07:26 -0700835 block_start_pfn = cc->free_pfn & ~(pageblock_nr_pages-1);
836 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
837 zone_end_pfn(zone));
Vlastimil Babka7ed695e2014-01-21 15:51:09 -0800838 low_pfn = ALIGN(cc->migrate_pfn + 1, pageblock_nr_pages);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100839
840 /*
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100841 * Isolate free pages until enough are available to migrate the
842 * pages on cc->migratepages. We stop searching if the migrate
843 * and free page scanners meet or enough free pages are isolated.
844 */
Vlastimil Babkac96b9e52014-06-04 16:07:26 -0700845 for (; block_start_pfn >= low_pfn && cc->nr_migratepages > nr_freepages;
846 block_end_pfn = block_start_pfn,
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700847 block_start_pfn -= pageblock_nr_pages,
848 isolate_start_pfn = block_start_pfn) {
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100849 unsigned long isolated;
850
David Rientjesf6ea3ad2013-09-30 13:45:03 -0700851 /*
852 * This can iterate a massively long zone without finding any
853 * suitable migration targets, so periodically check if we need
Vlastimil Babkabe976572014-06-04 16:10:41 -0700854 * to schedule, or even abort async compaction.
David Rientjesf6ea3ad2013-09-30 13:45:03 -0700855 */
Vlastimil Babkabe976572014-06-04 16:10:41 -0700856 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
857 && compact_should_abort(cc))
858 break;
David Rientjesf6ea3ad2013-09-30 13:45:03 -0700859
Vlastimil Babka7d49d882014-10-09 15:27:11 -0700860 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
861 zone);
862 if (!page)
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100863 continue;
864
865 /* Check the block is suitable for migration */
Linus Torvalds68e3e922012-06-03 20:05:57 -0700866 if (!suitable_migration_target(page))
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100867 continue;
Linus Torvalds68e3e922012-06-03 20:05:57 -0700868
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700869 /* If isolation recently failed, do not retry */
870 if (!isolation_suitable(cc, page))
871 continue;
872
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700873 /* Found a block suitable for isolating free pages from. */
874 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
Vlastimil Babkac96b9e52014-06-04 16:07:26 -0700875 block_end_pfn, freelist, false);
Mel Gormanf40d1e42012-10-08 16:32:36 -0700876 nr_freepages += isolated;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100877
878 /*
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700879 * Remember where the free scanner should restart next time,
880 * which is where isolate_freepages_block() left off.
881 * But if it scanned the whole pageblock, isolate_start_pfn
882 * now points at block_end_pfn, which is the start of the next
883 * pageblock.
884 * In that case we will however want to restart at the start
885 * of the previous pageblock.
886 */
887 cc->free_pfn = (isolate_start_pfn < block_end_pfn) ?
888 isolate_start_pfn :
889 block_start_pfn - pageblock_nr_pages;
890
891 /*
Vlastimil Babkae9ade562014-06-04 16:08:34 -0700892 * Set a flag that we successfully isolated in this pageblock.
893 * In the next loop iteration, zone->compact_cached_free_pfn
894 * will not be updated and thus it will effectively contain the
895 * highest pageblock we isolated pages from.
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100896 */
Vlastimil Babkae9ade562014-06-04 16:08:34 -0700897 if (isolated)
Mel Gormanc89511a2012-10-08 16:32:45 -0700898 cc->finished_update_free = true;
Vlastimil Babkabe976572014-06-04 16:10:41 -0700899
900 /*
901 * isolate_freepages_block() might have aborted due to async
902 * compaction being contended
903 */
904 if (cc->contended)
905 break;
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100906 }
907
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100908 /* split_free_page does not map the pages */
909 map_pages(freelist);
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100910
Vlastimil Babka7ed695e2014-01-21 15:51:09 -0800911 /*
912 * If we crossed the migrate scanner, we want to keep it that way
913 * so that compact_finished() may detect this
914 */
Vlastimil Babkac96b9e52014-06-04 16:07:26 -0700915 if (block_start_pfn < low_pfn)
Vlastimil Babkae9ade562014-06-04 16:08:34 -0700916 cc->free_pfn = cc->migrate_pfn;
Vlastimil Babkac96b9e52014-06-04 16:07:26 -0700917
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100918 cc->nr_freepages = nr_freepages;
Mel Gorman748446b2010-05-24 14:32:27 -0700919}
920
921/*
922 * This is a migrate-callback that "allocates" freepages by taking pages
923 * from the isolated freelists in the block we are migrating to.
924 */
925static struct page *compaction_alloc(struct page *migratepage,
926 unsigned long data,
927 int **result)
928{
929 struct compact_control *cc = (struct compact_control *)data;
930 struct page *freepage;
931
Vlastimil Babkabe976572014-06-04 16:10:41 -0700932 /*
933 * Isolate free pages if necessary, and if we are not aborting due to
934 * contention.
935 */
Mel Gorman748446b2010-05-24 14:32:27 -0700936 if (list_empty(&cc->freepages)) {
Vlastimil Babkabe976572014-06-04 16:10:41 -0700937 if (!cc->contended)
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700938 isolate_freepages(cc);
Mel Gorman748446b2010-05-24 14:32:27 -0700939
940 if (list_empty(&cc->freepages))
941 return NULL;
942 }
943
944 freepage = list_entry(cc->freepages.next, struct page, lru);
945 list_del(&freepage->lru);
946 cc->nr_freepages--;
947
948 return freepage;
949}
950
951/*
David Rientjesd53aea32014-06-04 16:08:26 -0700952 * This is a migrate-callback that "frees" freepages back to the isolated
953 * freelist. All pages on the freelist are from the same zone, so there is no
954 * special handling needed for NUMA.
955 */
956static void compaction_free(struct page *page, unsigned long data)
957{
958 struct compact_control *cc = (struct compact_control *)data;
959
960 list_add(&page->lru, &cc->freepages);
961 cc->nr_freepages++;
962}
963
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100964/* possible outcome of isolate_migratepages */
965typedef enum {
966 ISOLATE_ABORT, /* Abort compaction now */
967 ISOLATE_NONE, /* No pages isolated, continue scanning */
968 ISOLATE_SUCCESS, /* Pages isolated, migrate */
969} isolate_migrate_t;
970
971/*
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700972 * Isolate all pages that can be migrated from the first suitable block,
973 * starting at the block pointed to by the migrate scanner pfn within
974 * compact_control.
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100975 */
976static isolate_migrate_t isolate_migratepages(struct zone *zone,
977 struct compact_control *cc)
978{
979 unsigned long low_pfn, end_pfn;
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700980 struct page *page;
981 const isolate_mode_t isolate_mode =
982 (cc->mode == MIGRATE_ASYNC ? ISOLATE_ASYNC_MIGRATE : 0);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100983
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700984 /*
985 * Start at where we last stopped, or beginning of the zone as
986 * initialized by compact_zone()
987 */
988 low_pfn = cc->migrate_pfn;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100989
990 /* Only scan within a pageblock boundary */
Mel Gormana9aacbc2013-02-22 16:32:25 -0800991 end_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100992
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700993 /*
994 * Iterate over whole pageblocks until we find the first suitable.
995 * Do not cross the free scanner.
996 */
997 for (; end_pfn <= cc->free_pfn;
998 low_pfn = end_pfn, end_pfn += pageblock_nr_pages) {
999
1000 /*
1001 * This can potentially iterate a massively long zone with
1002 * many pageblocks unsuitable, so periodically check if we
1003 * need to schedule, or even abort async compaction.
1004 */
1005 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1006 && compact_should_abort(cc))
1007 break;
1008
Vlastimil Babka7d49d882014-10-09 15:27:11 -07001009 page = pageblock_pfn_to_page(low_pfn, end_pfn, zone);
1010 if (!page)
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001011 continue;
1012
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001013 /* If isolation recently failed, do not retry */
1014 if (!isolation_suitable(cc, page))
1015 continue;
1016
1017 /*
1018 * For async compaction, also only scan in MOVABLE blocks.
1019 * Async compaction is optimistic to see if the minimum amount
1020 * of work satisfies the allocation.
1021 */
1022 if (cc->mode == MIGRATE_ASYNC &&
1023 !migrate_async_suitable(get_pageblock_migratetype(page)))
1024 continue;
1025
1026 /* Perform the isolation */
1027 low_pfn = isolate_migratepages_block(cc, low_pfn, end_pfn,
1028 isolate_mode);
1029
1030 if (!low_pfn || cc->contended)
1031 return ISOLATE_ABORT;
1032
1033 /*
1034 * Either we isolated something and proceed with migration. Or
1035 * we failed and compact_zone should decide if we should
1036 * continue or not.
1037 */
1038 break;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001039 }
1040
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001041 acct_isolated(zone, cc);
Vlastimil Babka1d5bfe12014-11-13 15:19:30 -08001042 /*
1043 * Record where migration scanner will be restarted. If we end up in
1044 * the same pageblock as the free scanner, make the scanners fully
1045 * meet so that compact_finished() terminates compaction.
1046 */
1047 cc->migrate_pfn = (end_pfn <= cc->free_pfn) ? low_pfn : cc->free_pfn;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001048
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001049 return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001050}
1051
David Rientjes6d7ce552014-10-09 15:27:27 -07001052static int compact_finished(struct zone *zone, struct compact_control *cc,
1053 const int migratetype)
Mel Gorman748446b2010-05-24 14:32:27 -07001054{
Mel Gorman8fb74b92013-01-11 14:32:16 -08001055 unsigned int order;
Andrea Arcangeli5a03b052011-01-13 15:47:11 -08001056 unsigned long watermark;
Mel Gorman56de7262010-05-24 14:32:30 -07001057
Vlastimil Babkabe976572014-06-04 16:10:41 -07001058 if (cc->contended || fatal_signal_pending(current))
Mel Gorman748446b2010-05-24 14:32:27 -07001059 return COMPACT_PARTIAL;
1060
Mel Gorman753341a2012-10-08 16:32:40 -07001061 /* Compaction run completes if the migrate and free scanner meet */
Mel Gormanbb13ffe2012-10-08 16:32:41 -07001062 if (cc->free_pfn <= cc->migrate_pfn) {
Vlastimil Babka55b7c4c2014-01-21 15:51:11 -08001063 /* Let the next compaction start anew. */
David Rientjes35979ef2014-06-04 16:08:27 -07001064 zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
1065 zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
Vlastimil Babka55b7c4c2014-01-21 15:51:11 -08001066 zone->compact_cached_free_pfn = zone_end_pfn(zone);
1067
Mel Gorman62997022012-10-08 16:32:47 -07001068 /*
1069 * Mark that the PG_migrate_skip information should be cleared
1070 * by kswapd when it goes to sleep. kswapd does not set the
1071 * flag itself as the decision to be clear should be directly
1072 * based on an allocation request.
1073 */
1074 if (!current_is_kswapd())
1075 zone->compact_blockskip_flush = true;
1076
Mel Gorman748446b2010-05-24 14:32:27 -07001077 return COMPACT_COMPLETE;
Mel Gormanbb13ffe2012-10-08 16:32:41 -07001078 }
Mel Gorman748446b2010-05-24 14:32:27 -07001079
Johannes Weiner82478fb2011-01-20 14:44:21 -08001080 /*
1081 * order == -1 is expected when compacting via
1082 * /proc/sys/vm/compact_memory
1083 */
Mel Gorman56de7262010-05-24 14:32:30 -07001084 if (cc->order == -1)
1085 return COMPACT_CONTINUE;
1086
Michal Hocko3957c772011-06-15 15:08:25 -07001087 /* Compaction run is not finished if the watermark is not met */
1088 watermark = low_wmark_pages(zone);
1089 watermark += (1 << cc->order);
1090
1091 if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
1092 return COMPACT_CONTINUE;
1093
Mel Gorman56de7262010-05-24 14:32:30 -07001094 /* Direct compactor: Is a suitable page free? */
Mel Gorman8fb74b92013-01-11 14:32:16 -08001095 for (order = cc->order; order < MAX_ORDER; order++) {
1096 struct free_area *area = &zone->free_area[order];
Mel Gorman56de7262010-05-24 14:32:30 -07001097
Mel Gorman8fb74b92013-01-11 14:32:16 -08001098 /* Job done if page is free of the right migratetype */
David Rientjes6d7ce552014-10-09 15:27:27 -07001099 if (!list_empty(&area->free_list[migratetype]))
Mel Gorman8fb74b92013-01-11 14:32:16 -08001100 return COMPACT_PARTIAL;
1101
1102 /* Job done if allocation would set block type */
1103 if (cc->order >= pageblock_order && area->nr_free)
1104 return COMPACT_PARTIAL;
Mel Gorman56de7262010-05-24 14:32:30 -07001105 }
1106
Mel Gorman748446b2010-05-24 14:32:27 -07001107 return COMPACT_CONTINUE;
1108}
1109
Mel Gorman3e7d3442011-01-13 15:45:56 -08001110/*
1111 * compaction_suitable: Is this suitable to run compaction on this zone now?
1112 * Returns
1113 * COMPACT_SKIPPED - If there are too few free pages for compaction
1114 * COMPACT_PARTIAL - If the allocation would succeed without compaction
1115 * COMPACT_CONTINUE - If compaction should run now
1116 */
1117unsigned long compaction_suitable(struct zone *zone, int order)
1118{
1119 int fragindex;
1120 unsigned long watermark;
1121
1122 /*
Michal Hocko3957c772011-06-15 15:08:25 -07001123 * order == -1 is expected when compacting via
1124 * /proc/sys/vm/compact_memory
1125 */
1126 if (order == -1)
1127 return COMPACT_CONTINUE;
1128
1129 /*
Mel Gorman3e7d3442011-01-13 15:45:56 -08001130 * Watermarks for order-0 must be met for compaction. Note the 2UL.
1131 * This is because during migration, copies of pages need to be
1132 * allocated and for a short time, the footprint is higher
1133 */
1134 watermark = low_wmark_pages(zone) + (2UL << order);
1135 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1136 return COMPACT_SKIPPED;
1137
1138 /*
1139 * fragmentation index determines if allocation failures are due to
1140 * low memory or external fragmentation
1141 *
Shaohua Lia582a732011-06-15 15:08:49 -07001142 * index of -1000 implies allocations might succeed depending on
1143 * watermarks
Mel Gorman3e7d3442011-01-13 15:45:56 -08001144 * index towards 0 implies failure is due to lack of memory
1145 * index towards 1000 implies failure is due to fragmentation
1146 *
1147 * Only compact if a failure would be due to fragmentation.
1148 */
1149 fragindex = fragmentation_index(zone, order);
1150 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
1151 return COMPACT_SKIPPED;
1152
Shaohua Lia582a732011-06-15 15:08:49 -07001153 if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
1154 0, 0))
Mel Gorman3e7d3442011-01-13 15:45:56 -08001155 return COMPACT_PARTIAL;
1156
1157 return COMPACT_CONTINUE;
1158}
1159
Mel Gorman748446b2010-05-24 14:32:27 -07001160static int compact_zone(struct zone *zone, struct compact_control *cc)
1161{
1162 int ret;
Mel Gormanc89511a2012-10-08 16:32:45 -07001163 unsigned long start_pfn = zone->zone_start_pfn;
Cody P Schafer108bcc92013-02-22 16:35:23 -08001164 unsigned long end_pfn = zone_end_pfn(zone);
David Rientjes6d7ce552014-10-09 15:27:27 -07001165 const int migratetype = gfpflags_to_migratetype(cc->gfp_mask);
David Rientjese0b9dae2014-06-04 16:08:28 -07001166 const bool sync = cc->mode != MIGRATE_ASYNC;
Mel Gorman748446b2010-05-24 14:32:27 -07001167
Mel Gorman3e7d3442011-01-13 15:45:56 -08001168 ret = compaction_suitable(zone, cc->order);
1169 switch (ret) {
1170 case COMPACT_PARTIAL:
1171 case COMPACT_SKIPPED:
1172 /* Compaction is likely to fail */
1173 return ret;
1174 case COMPACT_CONTINUE:
1175 /* Fall through to compaction */
1176 ;
1177 }
1178
Mel Gormanc89511a2012-10-08 16:32:45 -07001179 /*
Vlastimil Babkad3132e42014-01-21 15:51:08 -08001180 * Clear pageblock skip if there were failures recently and compaction
1181 * is about to be retried after being deferred. kswapd does not do
1182 * this reset as it'll reset the cached information when going to sleep.
1183 */
1184 if (compaction_restarting(zone, cc->order) && !current_is_kswapd())
1185 __reset_isolation_suitable(zone);
1186
1187 /*
Mel Gormanc89511a2012-10-08 16:32:45 -07001188 * Setup to move all movable pages to the end of the zone. Used cached
1189 * information on where the scanners should start but check that it
1190 * is initialised by ensuring the values are within zone boundaries.
1191 */
David Rientjese0b9dae2014-06-04 16:08:28 -07001192 cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync];
Mel Gormanc89511a2012-10-08 16:32:45 -07001193 cc->free_pfn = zone->compact_cached_free_pfn;
1194 if (cc->free_pfn < start_pfn || cc->free_pfn > end_pfn) {
1195 cc->free_pfn = end_pfn & ~(pageblock_nr_pages-1);
1196 zone->compact_cached_free_pfn = cc->free_pfn;
1197 }
1198 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn > end_pfn) {
1199 cc->migrate_pfn = start_pfn;
David Rientjes35979ef2014-06-04 16:08:27 -07001200 zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
1201 zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
Mel Gormanc89511a2012-10-08 16:32:45 -07001202 }
Mel Gorman748446b2010-05-24 14:32:27 -07001203
Mel Gorman0eb927c2014-01-21 15:51:05 -08001204 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn, cc->free_pfn, end_pfn);
1205
Mel Gorman748446b2010-05-24 14:32:27 -07001206 migrate_prep_local();
1207
David Rientjes6d7ce552014-10-09 15:27:27 -07001208 while ((ret = compact_finished(zone, cc, migratetype)) ==
1209 COMPACT_CONTINUE) {
Minchan Kim9d502c12011-03-22 16:30:39 -07001210 int err;
Mel Gorman748446b2010-05-24 14:32:27 -07001211
Mel Gormanf9e35b32011-06-15 15:08:52 -07001212 switch (isolate_migratepages(zone, cc)) {
1213 case ISOLATE_ABORT:
1214 ret = COMPACT_PARTIAL;
Rafael Aquini5733c7d2012-12-11 16:02:47 -08001215 putback_movable_pages(&cc->migratepages);
Shaohua Lie64c5232012-10-08 16:32:27 -07001216 cc->nr_migratepages = 0;
Mel Gormanf9e35b32011-06-15 15:08:52 -07001217 goto out;
1218 case ISOLATE_NONE:
Mel Gorman748446b2010-05-24 14:32:27 -07001219 continue;
Mel Gormanf9e35b32011-06-15 15:08:52 -07001220 case ISOLATE_SUCCESS:
1221 ;
1222 }
Mel Gorman748446b2010-05-24 14:32:27 -07001223
David Rientjesd53aea32014-06-04 16:08:26 -07001224 err = migrate_pages(&cc->migratepages, compaction_alloc,
David Rientjese0b9dae2014-06-04 16:08:28 -07001225 compaction_free, (unsigned long)cc, cc->mode,
Mel Gorman7b2a2d42012-10-19 14:07:31 +01001226 MR_COMPACTION);
Mel Gorman748446b2010-05-24 14:32:27 -07001227
Vlastimil Babkaf8c93012014-06-04 16:08:32 -07001228 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
1229 &cc->migratepages);
Mel Gorman748446b2010-05-24 14:32:27 -07001230
Vlastimil Babkaf8c93012014-06-04 16:08:32 -07001231 /* All pages were either migrated or will be released */
1232 cc->nr_migratepages = 0;
Minchan Kim9d502c12011-03-22 16:30:39 -07001233 if (err) {
Rafael Aquini5733c7d2012-12-11 16:02:47 -08001234 putback_movable_pages(&cc->migratepages);
Vlastimil Babka7ed695e2014-01-21 15:51:09 -08001235 /*
1236 * migrate_pages() may return -ENOMEM when scanners meet
1237 * and we want compact_finished() to detect it
1238 */
1239 if (err == -ENOMEM && cc->free_pfn > cc->migrate_pfn) {
David Rientjes4bf2bba2012-07-11 14:02:13 -07001240 ret = COMPACT_PARTIAL;
1241 goto out;
1242 }
Mel Gorman748446b2010-05-24 14:32:27 -07001243 }
Mel Gorman748446b2010-05-24 14:32:27 -07001244 }
1245
Mel Gormanf9e35b32011-06-15 15:08:52 -07001246out:
Mel Gorman748446b2010-05-24 14:32:27 -07001247 /* Release free pages and check accounting */
1248 cc->nr_freepages -= release_freepages(&cc->freepages);
1249 VM_BUG_ON(cc->nr_freepages != 0);
1250
Mel Gorman0eb927c2014-01-21 15:51:05 -08001251 trace_mm_compaction_end(ret);
1252
Mel Gorman748446b2010-05-24 14:32:27 -07001253 return ret;
1254}
Mel Gorman76ab0f52010-05-24 14:32:28 -07001255
David Rientjese0b9dae2014-06-04 16:08:28 -07001256static unsigned long compact_zone_order(struct zone *zone, int order,
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001257 gfp_t gfp_mask, enum migrate_mode mode, int *contended)
Mel Gorman56de7262010-05-24 14:32:30 -07001258{
Shaohua Lie64c5232012-10-08 16:32:27 -07001259 unsigned long ret;
Mel Gorman56de7262010-05-24 14:32:30 -07001260 struct compact_control cc = {
1261 .nr_freepages = 0,
1262 .nr_migratepages = 0,
1263 .order = order,
David Rientjes6d7ce552014-10-09 15:27:27 -07001264 .gfp_mask = gfp_mask,
Mel Gorman56de7262010-05-24 14:32:30 -07001265 .zone = zone,
David Rientjese0b9dae2014-06-04 16:08:28 -07001266 .mode = mode,
Mel Gorman56de7262010-05-24 14:32:30 -07001267 };
1268 INIT_LIST_HEAD(&cc.freepages);
1269 INIT_LIST_HEAD(&cc.migratepages);
1270
Shaohua Lie64c5232012-10-08 16:32:27 -07001271 ret = compact_zone(zone, &cc);
1272
1273 VM_BUG_ON(!list_empty(&cc.freepages));
1274 VM_BUG_ON(!list_empty(&cc.migratepages));
1275
1276 *contended = cc.contended;
1277 return ret;
Mel Gorman56de7262010-05-24 14:32:30 -07001278}
1279
Mel Gorman5e771902010-05-24 14:32:31 -07001280int sysctl_extfrag_threshold = 500;
1281
Mel Gorman56de7262010-05-24 14:32:30 -07001282/**
1283 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1284 * @zonelist: The zonelist used for the current allocation
1285 * @order: The order of the current allocation
1286 * @gfp_mask: The GFP mask of the current allocation
1287 * @nodemask: The allowed nodes to allocate from
David Rientjese0b9dae2014-06-04 16:08:28 -07001288 * @mode: The migration mode for async, sync light, or sync migration
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001289 * @contended: Return value that determines if compaction was aborted due to
1290 * need_resched() or lock contention
Vlastimil Babka53853e22014-10-09 15:27:02 -07001291 * @candidate_zone: Return the zone where we think allocation should succeed
Mel Gorman56de7262010-05-24 14:32:30 -07001292 *
1293 * This is the main entry point for direct page compaction.
1294 */
1295unsigned long try_to_compact_pages(struct zonelist *zonelist,
Mel Gorman77f1fe62011-01-13 15:45:57 -08001296 int order, gfp_t gfp_mask, nodemask_t *nodemask,
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001297 enum migrate_mode mode, int *contended,
Vlastimil Babka53853e22014-10-09 15:27:02 -07001298 struct zone **candidate_zone)
Mel Gorman56de7262010-05-24 14:32:30 -07001299{
1300 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1301 int may_enter_fs = gfp_mask & __GFP_FS;
1302 int may_perform_io = gfp_mask & __GFP_IO;
Mel Gorman56de7262010-05-24 14:32:30 -07001303 struct zoneref *z;
1304 struct zone *zone;
Vlastimil Babka53853e22014-10-09 15:27:02 -07001305 int rc = COMPACT_DEFERRED;
Bartlomiej Zolnierkiewiczd95ea5d2012-10-08 16:32:05 -07001306 int alloc_flags = 0;
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001307 int all_zones_contended = COMPACT_CONTENDED_LOCK; /* init for &= op */
1308
1309 *contended = COMPACT_CONTENDED_NONE;
Mel Gorman56de7262010-05-24 14:32:30 -07001310
Mel Gorman4ffb6332012-10-08 16:29:09 -07001311 /* Check if the GFP flags allow compaction */
Andrea Arcangelic5a73c32011-01-13 15:47:11 -08001312 if (!order || !may_enter_fs || !may_perform_io)
Vlastimil Babka53853e22014-10-09 15:27:02 -07001313 return COMPACT_SKIPPED;
Mel Gorman56de7262010-05-24 14:32:30 -07001314
Bartlomiej Zolnierkiewiczd95ea5d2012-10-08 16:32:05 -07001315#ifdef CONFIG_CMA
David Rientjes43e7a342014-10-09 15:27:25 -07001316 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
Bartlomiej Zolnierkiewiczd95ea5d2012-10-08 16:32:05 -07001317 alloc_flags |= ALLOC_CMA;
1318#endif
Mel Gorman56de7262010-05-24 14:32:30 -07001319 /* Compact each zone in the list */
1320 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1321 nodemask) {
Mel Gorman56de7262010-05-24 14:32:30 -07001322 int status;
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001323 int zone_contended;
Mel Gorman56de7262010-05-24 14:32:30 -07001324
Vlastimil Babka53853e22014-10-09 15:27:02 -07001325 if (compaction_deferred(zone, order))
1326 continue;
1327
David Rientjese0b9dae2014-06-04 16:08:28 -07001328 status = compact_zone_order(zone, order, gfp_mask, mode,
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001329 &zone_contended);
Mel Gorman56de7262010-05-24 14:32:30 -07001330 rc = max(status, rc);
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001331 /*
1332 * It takes at least one zone that wasn't lock contended
1333 * to clear all_zones_contended.
1334 */
1335 all_zones_contended &= zone_contended;
Mel Gorman56de7262010-05-24 14:32:30 -07001336
Mel Gorman3e7d3442011-01-13 15:45:56 -08001337 /* If a normal allocation would succeed, stop compacting */
Bartlomiej Zolnierkiewiczd95ea5d2012-10-08 16:32:05 -07001338 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0,
Vlastimil Babka53853e22014-10-09 15:27:02 -07001339 alloc_flags)) {
1340 *candidate_zone = zone;
1341 /*
1342 * We think the allocation will succeed in this zone,
1343 * but it is not certain, hence the false. The caller
1344 * will repeat this with true if allocation indeed
1345 * succeeds in this zone.
1346 */
1347 compaction_defer_reset(zone, order, false);
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001348 /*
1349 * It is possible that async compaction aborted due to
1350 * need_resched() and the watermarks were ok thanks to
1351 * somebody else freeing memory. The allocation can
1352 * however still fail so we better signal the
1353 * need_resched() contention anyway (this will not
1354 * prevent the allocation attempt).
1355 */
1356 if (zone_contended == COMPACT_CONTENDED_SCHED)
1357 *contended = COMPACT_CONTENDED_SCHED;
1358
1359 goto break_loop;
1360 }
1361
1362 if (mode != MIGRATE_ASYNC) {
Vlastimil Babka53853e22014-10-09 15:27:02 -07001363 /*
1364 * We think that allocation won't succeed in this zone
1365 * so we defer compaction there. If it ends up
1366 * succeeding after all, it will be reset.
1367 */
1368 defer_compaction(zone, order);
1369 }
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001370
1371 /*
1372 * We might have stopped compacting due to need_resched() in
1373 * async compaction, or due to a fatal signal detected. In that
1374 * case do not try further zones and signal need_resched()
1375 * contention.
1376 */
1377 if ((zone_contended == COMPACT_CONTENDED_SCHED)
1378 || fatal_signal_pending(current)) {
1379 *contended = COMPACT_CONTENDED_SCHED;
1380 goto break_loop;
1381 }
1382
1383 continue;
1384break_loop:
1385 /*
1386 * We might not have tried all the zones, so be conservative
1387 * and assume they are not all lock contended.
1388 */
1389 all_zones_contended = 0;
1390 break;
Mel Gorman56de7262010-05-24 14:32:30 -07001391 }
1392
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001393 /*
1394 * If at least one zone wasn't deferred or skipped, we report if all
1395 * zones that were tried were lock contended.
1396 */
1397 if (rc > COMPACT_SKIPPED && all_zones_contended)
1398 *contended = COMPACT_CONTENDED_LOCK;
1399
Mel Gorman56de7262010-05-24 14:32:30 -07001400 return rc;
1401}
1402
1403
Mel Gorman76ab0f52010-05-24 14:32:28 -07001404/* Compact all zones within a node */
Andrew Morton7103f162013-02-22 16:32:33 -08001405static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
Mel Gorman76ab0f52010-05-24 14:32:28 -07001406{
1407 int zoneid;
Mel Gorman76ab0f52010-05-24 14:32:28 -07001408 struct zone *zone;
1409
Mel Gorman76ab0f52010-05-24 14:32:28 -07001410 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
Mel Gorman76ab0f52010-05-24 14:32:28 -07001411
1412 zone = &pgdat->node_zones[zoneid];
1413 if (!populated_zone(zone))
1414 continue;
1415
Rik van Riel7be62de2012-03-21 16:33:52 -07001416 cc->nr_freepages = 0;
1417 cc->nr_migratepages = 0;
1418 cc->zone = zone;
1419 INIT_LIST_HEAD(&cc->freepages);
1420 INIT_LIST_HEAD(&cc->migratepages);
Mel Gorman76ab0f52010-05-24 14:32:28 -07001421
Dan Carpenteraad6ec32012-03-21 16:33:54 -07001422 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
Rik van Riel7be62de2012-03-21 16:33:52 -07001423 compact_zone(zone, cc);
Mel Gorman76ab0f52010-05-24 14:32:28 -07001424
Rik van Rielaff62242012-03-21 16:33:52 -07001425 if (cc->order > 0) {
Vlastimil Babkade6c60a2014-01-21 15:51:07 -08001426 if (zone_watermark_ok(zone, cc->order,
1427 low_wmark_pages(zone), 0, 0))
1428 compaction_defer_reset(zone, cc->order, false);
Rik van Rielaff62242012-03-21 16:33:52 -07001429 }
1430
Rik van Riel7be62de2012-03-21 16:33:52 -07001431 VM_BUG_ON(!list_empty(&cc->freepages));
1432 VM_BUG_ON(!list_empty(&cc->migratepages));
Mel Gorman76ab0f52010-05-24 14:32:28 -07001433 }
Mel Gorman76ab0f52010-05-24 14:32:28 -07001434}
1435
Andrew Morton7103f162013-02-22 16:32:33 -08001436void compact_pgdat(pg_data_t *pgdat, int order)
Rik van Riel7be62de2012-03-21 16:33:52 -07001437{
1438 struct compact_control cc = {
1439 .order = order,
David Rientjese0b9dae2014-06-04 16:08:28 -07001440 .mode = MIGRATE_ASYNC,
Rik van Riel7be62de2012-03-21 16:33:52 -07001441 };
1442
Mel Gorman3a7200a2013-09-11 14:22:19 -07001443 if (!order)
1444 return;
1445
Andrew Morton7103f162013-02-22 16:32:33 -08001446 __compact_pgdat(pgdat, &cc);
Rik van Riel7be62de2012-03-21 16:33:52 -07001447}
1448
Andrew Morton7103f162013-02-22 16:32:33 -08001449static void compact_node(int nid)
Rik van Riel7be62de2012-03-21 16:33:52 -07001450{
Rik van Riel7be62de2012-03-21 16:33:52 -07001451 struct compact_control cc = {
1452 .order = -1,
David Rientjese0b9dae2014-06-04 16:08:28 -07001453 .mode = MIGRATE_SYNC,
David Rientjes91ca9182014-04-03 14:47:23 -07001454 .ignore_skip_hint = true,
Rik van Riel7be62de2012-03-21 16:33:52 -07001455 };
1456
Andrew Morton7103f162013-02-22 16:32:33 -08001457 __compact_pgdat(NODE_DATA(nid), &cc);
Rik van Riel7be62de2012-03-21 16:33:52 -07001458}
1459
Mel Gorman76ab0f52010-05-24 14:32:28 -07001460/* Compact all nodes in the system */
Jason Liu7964c062013-01-11 14:31:47 -08001461static void compact_nodes(void)
Mel Gorman76ab0f52010-05-24 14:32:28 -07001462{
1463 int nid;
1464
Hugh Dickins8575ec22012-03-21 16:33:53 -07001465 /* Flush pending updates to the LRU lists */
1466 lru_add_drain_all();
1467
Mel Gorman76ab0f52010-05-24 14:32:28 -07001468 for_each_online_node(nid)
1469 compact_node(nid);
Mel Gorman76ab0f52010-05-24 14:32:28 -07001470}
1471
1472/* The written value is actually unused, all memory is compacted */
1473int sysctl_compact_memory;
1474
1475/* This is the entry point for compacting all nodes via /proc/sys/vm */
1476int sysctl_compaction_handler(struct ctl_table *table, int write,
1477 void __user *buffer, size_t *length, loff_t *ppos)
1478{
1479 if (write)
Jason Liu7964c062013-01-11 14:31:47 -08001480 compact_nodes();
Mel Gorman76ab0f52010-05-24 14:32:28 -07001481
1482 return 0;
1483}
Mel Gormaned4a6d72010-05-24 14:32:29 -07001484
Mel Gorman5e771902010-05-24 14:32:31 -07001485int sysctl_extfrag_handler(struct ctl_table *table, int write,
1486 void __user *buffer, size_t *length, loff_t *ppos)
1487{
1488 proc_dointvec_minmax(table, write, buffer, length, ppos);
1489
1490 return 0;
1491}
1492
Mel Gormaned4a6d72010-05-24 14:32:29 -07001493#if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
Rashika Kheria74e77fb2014-04-03 14:48:01 -07001494static ssize_t sysfs_compact_node(struct device *dev,
Kay Sievers10fbcf42011-12-21 14:48:43 -08001495 struct device_attribute *attr,
Mel Gormaned4a6d72010-05-24 14:32:29 -07001496 const char *buf, size_t count)
1497{
Hugh Dickins8575ec22012-03-21 16:33:53 -07001498 int nid = dev->id;
1499
1500 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1501 /* Flush pending updates to the LRU lists */
1502 lru_add_drain_all();
1503
1504 compact_node(nid);
1505 }
Mel Gormaned4a6d72010-05-24 14:32:29 -07001506
1507 return count;
1508}
Kay Sievers10fbcf42011-12-21 14:48:43 -08001509static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
Mel Gormaned4a6d72010-05-24 14:32:29 -07001510
1511int compaction_register_node(struct node *node)
1512{
Kay Sievers10fbcf42011-12-21 14:48:43 -08001513 return device_create_file(&node->dev, &dev_attr_compact);
Mel Gormaned4a6d72010-05-24 14:32:29 -07001514}
1515
1516void compaction_unregister_node(struct node *node)
1517{
Kay Sievers10fbcf42011-12-21 14:48:43 -08001518 return device_remove_file(&node->dev, &dev_attr_compact);
Mel Gormaned4a6d72010-05-24 14:32:29 -07001519}
1520#endif /* CONFIG_SYSFS && CONFIG_NUMA */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001521
1522#endif /* CONFIG_COMPACTION */