blob: 9c7e6909dd29089f308902778164812d2c1ad216 [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;
Vlastimil Babka6bace092014-12-10 15:43:31 -080044 unsigned long high_pfn = 0;
Mel Gorman748446b2010-05-24 14:32:27 -070045
46 list_for_each_entry_safe(page, next, freelist, lru) {
Vlastimil Babka6bace092014-12-10 15:43:31 -080047 unsigned long pfn = page_to_pfn(page);
Mel Gorman748446b2010-05-24 14:32:27 -070048 list_del(&page->lru);
49 __free_page(page);
Vlastimil Babka6bace092014-12-10 15:43:31 -080050 if (pfn > high_pfn)
51 high_pfn = pfn;
Mel Gorman748446b2010-05-24 14:32:27 -070052 }
53
Vlastimil Babka6bace092014-12-10 15:43:31 -080054 return high_pfn;
Mel Gorman748446b2010-05-24 14:32:27 -070055}
56
Michal Nazarewiczff9543f2011-12-29 13:09:50 +010057static void map_pages(struct list_head *list)
58{
59 struct page *page;
60
61 list_for_each_entry(page, list, lru) {
62 arch_alloc_page(page, 0);
63 kernel_map_pages(page, 1, 1);
64 }
65}
66
Michal Nazarewicz47118af2011-12-29 13:09:50 +010067static inline bool migrate_async_suitable(int migratetype)
68{
69 return is_migrate_cma(migratetype) || migratetype == MIGRATE_MOVABLE;
70}
71
Vlastimil Babka7d49d882014-10-09 15:27:11 -070072/*
73 * Check that the whole (or subset of) a pageblock given by the interval of
74 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
75 * with the migration of free compaction scanner. The scanners then need to
76 * use only pfn_valid_within() check for arches that allow holes within
77 * pageblocks.
78 *
79 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
80 *
81 * It's possible on some configurations to have a setup like node0 node1 node0
82 * i.e. it's possible that all pages within a zones range of pages do not
83 * belong to a single zone. We assume that a border between node0 and node1
84 * can occur within a single pageblock, but not a node0 node1 node0
85 * interleaving within a single pageblock. It is therefore sufficient to check
86 * the first and last page of a pageblock and avoid checking each individual
87 * page in a pageblock.
88 */
89static struct page *pageblock_pfn_to_page(unsigned long start_pfn,
90 unsigned long end_pfn, struct zone *zone)
91{
92 struct page *start_page;
93 struct page *end_page;
94
95 /* end_pfn is one past the range we are checking */
96 end_pfn--;
97
98 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
99 return NULL;
100
101 start_page = pfn_to_page(start_pfn);
102
103 if (page_zone(start_page) != zone)
104 return NULL;
105
106 end_page = pfn_to_page(end_pfn);
107
108 /* This gives a shorter code than deriving page_zone(end_page) */
109 if (page_zone_id(start_page) != page_zone_id(end_page))
110 return NULL;
111
112 return start_page;
113}
114
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700115#ifdef CONFIG_COMPACTION
116/* Returns true if the pageblock should be scanned for pages to isolate. */
117static inline bool isolation_suitable(struct compact_control *cc,
118 struct page *page)
119{
120 if (cc->ignore_skip_hint)
121 return true;
122
123 return !get_pageblock_skip(page);
124}
125
126/*
127 * This function is called to clear all cached information on pageblocks that
128 * should be skipped for page isolation when the migrate and free page scanner
129 * meet.
130 */
Mel Gorman62997022012-10-08 16:32:47 -0700131static void __reset_isolation_suitable(struct zone *zone)
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700132{
133 unsigned long start_pfn = zone->zone_start_pfn;
Cody P Schafer108bcc92013-02-22 16:35:23 -0800134 unsigned long end_pfn = zone_end_pfn(zone);
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700135 unsigned long pfn;
136
David Rientjes35979ef2014-06-04 16:08:27 -0700137 zone->compact_cached_migrate_pfn[0] = start_pfn;
138 zone->compact_cached_migrate_pfn[1] = start_pfn;
Mel Gormanc89511a2012-10-08 16:32:45 -0700139 zone->compact_cached_free_pfn = end_pfn;
Mel Gorman62997022012-10-08 16:32:47 -0700140 zone->compact_blockskip_flush = false;
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700141
142 /* Walk the zone and mark every pageblock as suitable for isolation */
143 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
144 struct page *page;
145
146 cond_resched();
147
148 if (!pfn_valid(pfn))
149 continue;
150
151 page = pfn_to_page(pfn);
152 if (zone != page_zone(page))
153 continue;
154
155 clear_pageblock_skip(page);
156 }
157}
158
Mel Gorman62997022012-10-08 16:32:47 -0700159void reset_isolation_suitable(pg_data_t *pgdat)
160{
161 int zoneid;
162
163 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
164 struct zone *zone = &pgdat->node_zones[zoneid];
165 if (!populated_zone(zone))
166 continue;
167
168 /* Only flush if a full compaction finished recently */
169 if (zone->compact_blockskip_flush)
170 __reset_isolation_suitable(zone);
171 }
172}
173
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700174/*
175 * If no pages were isolated then mark this pageblock to be skipped in the
Mel Gorman62997022012-10-08 16:32:47 -0700176 * future. The information is later cleared by __reset_isolation_suitable().
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700177 */
Mel Gormanc89511a2012-10-08 16:32:45 -0700178static void update_pageblock_skip(struct compact_control *cc,
179 struct page *page, unsigned long nr_isolated,
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700180 bool migrate_scanner)
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700181{
Mel Gormanc89511a2012-10-08 16:32:45 -0700182 struct zone *zone = cc->zone;
David Rientjes35979ef2014-06-04 16:08:27 -0700183 unsigned long pfn;
Joonsoo Kim6815bf32013-12-18 17:08:52 -0800184
185 if (cc->ignore_skip_hint)
186 return;
187
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700188 if (!page)
189 return;
190
David Rientjes35979ef2014-06-04 16:08:27 -0700191 if (nr_isolated)
192 return;
193
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700194 set_pageblock_skip(page);
Mel Gormanc89511a2012-10-08 16:32:45 -0700195
David Rientjes35979ef2014-06-04 16:08:27 -0700196 pfn = page_to_pfn(page);
197
198 /* Update where async and sync compaction should restart */
199 if (migrate_scanner) {
David Rientjes35979ef2014-06-04 16:08:27 -0700200 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 {
David Rientjes35979ef2014-06-04 16:08:27 -0700206 if (pfn < zone->compact_cached_free_pfn)
207 zone->compact_cached_free_pfn = pfn;
Mel Gormanc89511a2012-10-08 16:32:45 -0700208 }
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700209}
210#else
211static inline bool isolation_suitable(struct compact_control *cc,
212 struct page *page)
213{
214 return true;
215}
216
Mel Gormanc89511a2012-10-08 16:32:45 -0700217static void update_pageblock_skip(struct compact_control *cc,
218 struct page *page, unsigned long nr_isolated,
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700219 bool migrate_scanner)
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700220{
221}
222#endif /* CONFIG_COMPACTION */
223
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700224/*
225 * Compaction requires the taking of some coarse locks that are potentially
226 * very heavily contended. For async compaction, back out if the lock cannot
227 * be taken immediately. For sync compaction, spin on the lock if needed.
228 *
229 * Returns true if the lock is held
230 * Returns false if the lock is not held and compaction should abort
231 */
232static bool compact_trylock_irqsave(spinlock_t *lock, unsigned long *flags,
233 struct compact_control *cc)
Mel Gorman2a1402a2012-10-08 16:32:33 -0700234{
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700235 if (cc->mode == MIGRATE_ASYNC) {
236 if (!spin_trylock_irqsave(lock, *flags)) {
237 cc->contended = COMPACT_CONTENDED_LOCK;
238 return false;
239 }
240 } else {
241 spin_lock_irqsave(lock, *flags);
242 }
Vlastimil Babka1f9efde2014-10-09 15:27:14 -0700243
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700244 return true;
Mel Gorman2a1402a2012-10-08 16:32:33 -0700245}
246
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100247/*
Mel Gormanc67fe372012-08-21 16:16:17 -0700248 * Compaction requires the taking of some coarse locks that are potentially
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700249 * very heavily contended. The lock should be periodically unlocked to avoid
250 * having disabled IRQs for a long time, even when there is nobody waiting on
251 * the lock. It might also be that allowing the IRQs will result in
252 * need_resched() becoming true. If scheduling is needed, async compaction
253 * aborts. Sync compaction schedules.
254 * Either compaction type will also abort if a fatal signal is pending.
255 * In either case if the lock was locked, it is dropped and not regained.
Mel Gormanc67fe372012-08-21 16:16:17 -0700256 *
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700257 * Returns true if compaction should abort due to fatal signal pending, or
258 * async compaction due to need_resched()
259 * Returns false when compaction can continue (sync compaction might have
260 * scheduled)
Mel Gormanc67fe372012-08-21 16:16:17 -0700261 */
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700262static bool compact_unlock_should_abort(spinlock_t *lock,
263 unsigned long flags, bool *locked, struct compact_control *cc)
Mel Gormanc67fe372012-08-21 16:16:17 -0700264{
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700265 if (*locked) {
266 spin_unlock_irqrestore(lock, flags);
267 *locked = false;
268 }
Vlastimil Babka1f9efde2014-10-09 15:27:14 -0700269
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700270 if (fatal_signal_pending(current)) {
271 cc->contended = COMPACT_CONTENDED_SCHED;
272 return true;
273 }
Mel Gormanc67fe372012-08-21 16:16:17 -0700274
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700275 if (need_resched()) {
David Rientjese0b9dae2014-06-04 16:08:28 -0700276 if (cc->mode == MIGRATE_ASYNC) {
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700277 cc->contended = COMPACT_CONTENDED_SCHED;
278 return true;
Mel Gormanc67fe372012-08-21 16:16:17 -0700279 }
Mel Gormanc67fe372012-08-21 16:16:17 -0700280 cond_resched();
Mel Gormanc67fe372012-08-21 16:16:17 -0700281 }
282
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700283 return false;
Mel Gormanc67fe372012-08-21 16:16:17 -0700284}
285
Vlastimil Babkabe976572014-06-04 16:10:41 -0700286/*
287 * Aside from avoiding lock contention, compaction also periodically checks
288 * need_resched() and either schedules in sync compaction or aborts async
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700289 * compaction. This is similar to what compact_unlock_should_abort() does, but
Vlastimil Babkabe976572014-06-04 16:10:41 -0700290 * is used where no lock is concerned.
291 *
292 * Returns false when no scheduling was needed, or sync compaction scheduled.
293 * Returns true when async compaction should abort.
294 */
295static inline bool compact_should_abort(struct compact_control *cc)
296{
297 /* async compaction aborts if contended */
298 if (need_resched()) {
299 if (cc->mode == MIGRATE_ASYNC) {
Vlastimil Babka1f9efde2014-10-09 15:27:14 -0700300 cc->contended = COMPACT_CONTENDED_SCHED;
Vlastimil Babkabe976572014-06-04 16:10:41 -0700301 return true;
302 }
303
304 cond_resched();
305 }
306
307 return false;
308}
309
Mel Gormanf40d1e42012-10-08 16:32:36 -0700310/* Returns true if the page is within a block suitable for migration to */
311static bool suitable_migration_target(struct page *page)
312{
Joonsoo Kim7d348b92014-04-07 15:37:03 -0700313 /* If the page is a large free page, then disallow migration */
Vlastimil Babka99c0fd52014-10-09 15:27:23 -0700314 if (PageBuddy(page)) {
315 /*
316 * We are checking page_order without zone->lock taken. But
317 * the only small danger is that we skip a potentially suitable
318 * pageblock, so it's not worth to check order for valid range.
319 */
320 if (page_order_unsafe(page) >= pageblock_order)
321 return false;
322 }
Mel Gormanf40d1e42012-10-08 16:32:36 -0700323
324 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
Joonsoo Kim7d348b92014-04-07 15:37:03 -0700325 if (migrate_async_suitable(get_pageblock_migratetype(page)))
Mel Gormanf40d1e42012-10-08 16:32:36 -0700326 return true;
327
328 /* Otherwise skip the block */
329 return false;
330}
331
Mel Gormanc67fe372012-08-21 16:16:17 -0700332/*
Jerome Marchand9e4be472013-11-12 15:07:12 -0800333 * Isolate free pages onto a private freelist. If @strict is true, will abort
334 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
335 * (even though it may still end up isolating some pages).
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100336 */
Mel Gormanf40d1e42012-10-08 16:32:36 -0700337static unsigned long isolate_freepages_block(struct compact_control *cc,
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700338 unsigned long *start_pfn,
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100339 unsigned long end_pfn,
340 struct list_head *freelist,
341 bool strict)
Mel Gorman748446b2010-05-24 14:32:27 -0700342{
Mel Gormanb7aba692011-01-13 15:45:54 -0800343 int nr_scanned = 0, total_isolated = 0;
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700344 struct page *cursor, *valid_page = NULL;
Xiubo Lib8b2d822014-10-09 15:28:21 -0700345 unsigned long flags = 0;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700346 bool locked = false;
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700347 unsigned long blockpfn = *start_pfn;
Mel Gorman748446b2010-05-24 14:32:27 -0700348
Mel Gorman748446b2010-05-24 14:32:27 -0700349 cursor = pfn_to_page(blockpfn);
350
Mel Gormanf40d1e42012-10-08 16:32:36 -0700351 /* Isolate free pages. */
Mel Gorman748446b2010-05-24 14:32:27 -0700352 for (; blockpfn < end_pfn; blockpfn++, cursor++) {
353 int isolated, i;
354 struct page *page = cursor;
355
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700356 /*
357 * Periodically drop the lock (if held) regardless of its
358 * contention, to give chance to IRQs. Abort if fatal signal
359 * pending or async compaction detects need_resched()
360 */
361 if (!(blockpfn % SWAP_CLUSTER_MAX)
362 && compact_unlock_should_abort(&cc->zone->lock, flags,
363 &locked, cc))
364 break;
365
Mel Gormanb7aba692011-01-13 15:45:54 -0800366 nr_scanned++;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700367 if (!pfn_valid_within(blockpfn))
Laura Abbott2af120b2014-03-10 15:49:44 -0700368 goto isolate_fail;
369
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700370 if (!valid_page)
371 valid_page = page;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700372 if (!PageBuddy(page))
Laura Abbott2af120b2014-03-10 15:49:44 -0700373 goto isolate_fail;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700374
375 /*
Vlastimil Babka69b71892014-10-09 15:27:18 -0700376 * If we already hold the lock, we can skip some rechecking.
377 * Note that if we hold the lock now, checked_pageblock was
378 * already set in some previous iteration (or strict is true),
379 * so it is correct to skip the suitable migration target
380 * recheck as well.
Mel Gormanf40d1e42012-10-08 16:32:36 -0700381 */
Vlastimil Babka69b71892014-10-09 15:27:18 -0700382 if (!locked) {
383 /*
384 * The zone lock must be held to isolate freepages.
385 * Unfortunately this is a very coarse lock and can be
386 * heavily contended if there are parallel allocations
387 * or parallel compactions. For async compaction do not
388 * spin on the lock and we acquire the lock as late as
389 * possible.
390 */
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700391 locked = compact_trylock_irqsave(&cc->zone->lock,
392 &flags, cc);
Vlastimil Babka69b71892014-10-09 15:27:18 -0700393 if (!locked)
394 break;
Mel Gormanf40d1e42012-10-08 16:32:36 -0700395
Vlastimil Babka69b71892014-10-09 15:27:18 -0700396 /* Recheck this is a buddy page under lock */
397 if (!PageBuddy(page))
398 goto isolate_fail;
399 }
Mel Gorman748446b2010-05-24 14:32:27 -0700400
401 /* Found a free page, break it into order-0 pages */
402 isolated = split_free_page(page);
403 total_isolated += isolated;
404 for (i = 0; i < isolated; i++) {
405 list_add(&page->lru, freelist);
406 page++;
407 }
408
409 /* If a page was split, advance to the end of it */
410 if (isolated) {
411 blockpfn += isolated - 1;
412 cursor += isolated - 1;
Laura Abbott2af120b2014-03-10 15:49:44 -0700413 continue;
Mel Gorman748446b2010-05-24 14:32:27 -0700414 }
Laura Abbott2af120b2014-03-10 15:49:44 -0700415
416isolate_fail:
417 if (strict)
418 break;
419 else
420 continue;
421
Mel Gorman748446b2010-05-24 14:32:27 -0700422 }
423
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700424 /* Record how far we have got within the block */
425 *start_pfn = blockpfn;
426
Mel Gormanb7aba692011-01-13 15:45:54 -0800427 trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
Mel Gormanf40d1e42012-10-08 16:32:36 -0700428
429 /*
430 * If strict isolation is requested by CMA then check that all the
431 * pages requested were isolated. If there were any failures, 0 is
432 * returned and CMA will fail.
433 */
Laura Abbott2af120b2014-03-10 15:49:44 -0700434 if (strict && blockpfn < end_pfn)
Mel Gormanf40d1e42012-10-08 16:32:36 -0700435 total_isolated = 0;
436
437 if (locked)
438 spin_unlock_irqrestore(&cc->zone->lock, flags);
439
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700440 /* Update the pageblock-skip if the whole pageblock was scanned */
441 if (blockpfn == end_pfn)
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700442 update_pageblock_skip(cc, valid_page, total_isolated, false);
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700443
Minchan Kim010fc292012-12-20 15:05:06 -0800444 count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
Mel Gorman397487d2012-10-19 12:00:10 +0100445 if (total_isolated)
Minchan Kim010fc292012-12-20 15:05:06 -0800446 count_compact_events(COMPACTISOLATED, total_isolated);
Mel Gorman748446b2010-05-24 14:32:27 -0700447 return total_isolated;
448}
449
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100450/**
451 * isolate_freepages_range() - isolate free pages.
452 * @start_pfn: The first PFN to start isolating.
453 * @end_pfn: The one-past-last PFN.
454 *
455 * Non-free pages, invalid PFNs, or zone boundaries within the
456 * [start_pfn, end_pfn) range are considered errors, cause function to
457 * undo its actions and return zero.
458 *
459 * Otherwise, function returns one-past-the-last PFN of isolated page
460 * (which may be greater then end_pfn if end fell in a middle of
461 * a free page).
462 */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100463unsigned long
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700464isolate_freepages_range(struct compact_control *cc,
465 unsigned long start_pfn, unsigned long end_pfn)
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100466{
Mel Gormanf40d1e42012-10-08 16:32:36 -0700467 unsigned long isolated, pfn, block_end_pfn;
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100468 LIST_HEAD(freelist);
469
Vlastimil Babka7d49d882014-10-09 15:27:11 -0700470 pfn = start_pfn;
471 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100472
Vlastimil Babka7d49d882014-10-09 15:27:11 -0700473 for (; pfn < end_pfn; pfn += isolated,
474 block_end_pfn += pageblock_nr_pages) {
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700475 /* Protect pfn from changing by isolate_freepages_block */
476 unsigned long isolate_start_pfn = pfn;
Vlastimil Babka7d49d882014-10-09 15:27:11 -0700477
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100478 block_end_pfn = min(block_end_pfn, end_pfn);
479
Joonsoo Kim58420012014-11-13 15:19:07 -0800480 /*
481 * pfn could pass the block_end_pfn if isolated freepage
482 * is more than pageblock order. In this case, we adjust
483 * scanning range to right one.
484 */
485 if (pfn >= block_end_pfn) {
486 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
487 block_end_pfn = min(block_end_pfn, end_pfn);
488 }
489
Vlastimil Babka7d49d882014-10-09 15:27:11 -0700490 if (!pageblock_pfn_to_page(pfn, block_end_pfn, cc->zone))
491 break;
492
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700493 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
494 block_end_pfn, &freelist, true);
Michal Nazarewicz85aa1252012-01-30 13:24:03 +0100495
496 /*
497 * In strict mode, isolate_freepages_block() returns 0 if
498 * there are any holes in the block (ie. invalid PFNs or
499 * non-free pages).
500 */
501 if (!isolated)
502 break;
503
504 /*
505 * If we managed to isolate pages, it is always (1 << n) *
506 * pageblock_nr_pages for some non-negative n. (Max order
507 * page may span two pageblocks).
508 */
509 }
510
511 /* split_free_page does not map the pages */
512 map_pages(&freelist);
513
514 if (pfn < end_pfn) {
515 /* Loop terminated early, cleanup. */
516 release_freepages(&freelist);
517 return 0;
518 }
519
520 /* We don't use freelists for anything. */
521 return pfn;
522}
523
Mel Gorman748446b2010-05-24 14:32:27 -0700524/* Update the number of anon and file isolated pages in the zone */
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700525static void acct_isolated(struct zone *zone, struct compact_control *cc)
Mel Gorman748446b2010-05-24 14:32:27 -0700526{
527 struct page *page;
Minchan Kimb9e84ac2011-10-31 17:06:44 -0700528 unsigned int count[2] = { 0, };
Mel Gorman748446b2010-05-24 14:32:27 -0700529
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700530 if (list_empty(&cc->migratepages))
531 return;
532
Minchan Kimb9e84ac2011-10-31 17:06:44 -0700533 list_for_each_entry(page, &cc->migratepages, lru)
534 count[!!page_is_file_cache(page)]++;
Mel Gorman748446b2010-05-24 14:32:27 -0700535
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700536 mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
537 mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
Mel Gorman748446b2010-05-24 14:32:27 -0700538}
539
540/* Similar to reclaim, but different enough that they don't share logic */
541static bool too_many_isolated(struct zone *zone)
542{
Minchan Kimbc693042010-09-09 16:38:00 -0700543 unsigned long active, inactive, isolated;
Mel Gorman748446b2010-05-24 14:32:27 -0700544
545 inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
546 zone_page_state(zone, NR_INACTIVE_ANON);
Minchan Kimbc693042010-09-09 16:38:00 -0700547 active = zone_page_state(zone, NR_ACTIVE_FILE) +
548 zone_page_state(zone, NR_ACTIVE_ANON);
Mel Gorman748446b2010-05-24 14:32:27 -0700549 isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
550 zone_page_state(zone, NR_ISOLATED_ANON);
551
Minchan Kimbc693042010-09-09 16:38:00 -0700552 return isolated > (inactive + active) / 2;
Mel Gorman748446b2010-05-24 14:32:27 -0700553}
554
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100555/**
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700556 * isolate_migratepages_block() - isolate all migrate-able pages within
557 * a single pageblock
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100558 * @cc: Compaction control structure.
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700559 * @low_pfn: The first PFN to isolate
560 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
561 * @isolate_mode: Isolation mode to be used.
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100562 *
563 * Isolate all pages that can be migrated from the range specified by
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700564 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
565 * Returns zero if there is a fatal signal pending, otherwise PFN of the
566 * first page that was not scanned (which may be both less, equal to or more
567 * than end_pfn).
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100568 *
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700569 * The pages are isolated on cc->migratepages list (not required to be empty),
570 * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
571 * is neither read nor updated.
Mel Gorman748446b2010-05-24 14:32:27 -0700572 */
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700573static unsigned long
574isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
575 unsigned long end_pfn, isolate_mode_t isolate_mode)
Mel Gorman748446b2010-05-24 14:32:27 -0700576{
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700577 struct zone *zone = cc->zone;
Mel Gormanb7aba692011-01-13 15:45:54 -0800578 unsigned long nr_scanned = 0, nr_isolated = 0;
Mel Gorman748446b2010-05-24 14:32:27 -0700579 struct list_head *migratelist = &cc->migratepages;
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700580 struct lruvec *lruvec;
Xiubo Lib8b2d822014-10-09 15:28:21 -0700581 unsigned long flags = 0;
Mel Gorman2a1402a2012-10-08 16:32:33 -0700582 bool locked = false;
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700583 struct page *page = NULL, *valid_page = NULL;
Mel Gorman748446b2010-05-24 14:32:27 -0700584
Mel Gorman748446b2010-05-24 14:32:27 -0700585 /*
586 * Ensure that there are not too many pages isolated from the LRU
587 * list by either parallel reclaimers or compaction. If there are,
588 * delay for some time until fewer pages are isolated
589 */
590 while (unlikely(too_many_isolated(zone))) {
Mel Gormanf9e35b32011-06-15 15:08:52 -0700591 /* async migration should just abort */
David Rientjese0b9dae2014-06-04 16:08:28 -0700592 if (cc->mode == MIGRATE_ASYNC)
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100593 return 0;
Mel Gormanf9e35b32011-06-15 15:08:52 -0700594
Mel Gorman748446b2010-05-24 14:32:27 -0700595 congestion_wait(BLK_RW_ASYNC, HZ/10);
596
597 if (fatal_signal_pending(current))
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100598 return 0;
Mel Gorman748446b2010-05-24 14:32:27 -0700599 }
600
Vlastimil Babkabe976572014-06-04 16:10:41 -0700601 if (compact_should_abort(cc))
602 return 0;
David Rientjesaeef4b82014-06-04 16:08:31 -0700603
Mel Gorman748446b2010-05-24 14:32:27 -0700604 /* Time to isolate some pages for migration */
Mel Gorman748446b2010-05-24 14:32:27 -0700605 for (; low_pfn < end_pfn; low_pfn++) {
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700606 /*
607 * Periodically drop the lock (if held) regardless of its
608 * contention, to give chance to IRQs. Abort async compaction
609 * if contended.
610 */
611 if (!(low_pfn % SWAP_CLUSTER_MAX)
612 && compact_unlock_should_abort(&zone->lru_lock, flags,
613 &locked, cc))
614 break;
Mel Gormanc67fe372012-08-21 16:16:17 -0700615
Mel Gorman748446b2010-05-24 14:32:27 -0700616 if (!pfn_valid_within(low_pfn))
617 continue;
Mel Gormanb7aba692011-01-13 15:45:54 -0800618 nr_scanned++;
Mel Gorman748446b2010-05-24 14:32:27 -0700619
Mel Gorman748446b2010-05-24 14:32:27 -0700620 page = pfn_to_page(low_pfn);
Mel Gormandc908602012-02-08 17:13:38 -0800621
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700622 if (!valid_page)
623 valid_page = page;
624
Mel Gorman6c144662014-01-23 15:53:38 -0800625 /*
Vlastimil Babka99c0fd52014-10-09 15:27:23 -0700626 * Skip if free. We read page order here without zone lock
627 * which is generally unsafe, but the race window is small and
628 * the worst thing that can happen is that we skip some
629 * potential isolation targets.
Mel Gorman6c144662014-01-23 15:53:38 -0800630 */
Vlastimil Babka99c0fd52014-10-09 15:27:23 -0700631 if (PageBuddy(page)) {
632 unsigned long freepage_order = page_order_unsafe(page);
633
634 /*
635 * Without lock, we cannot be sure that what we got is
636 * a valid page order. Consider only values in the
637 * valid order range to prevent low_pfn overflow.
638 */
639 if (freepage_order > 0 && freepage_order < MAX_ORDER)
640 low_pfn += (1UL << freepage_order) - 1;
Mel Gorman748446b2010-05-24 14:32:27 -0700641 continue;
Vlastimil Babka99c0fd52014-10-09 15:27:23 -0700642 }
Mel Gorman748446b2010-05-24 14:32:27 -0700643
Mel Gorman9927af742011-01-13 15:45:59 -0800644 /*
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800645 * Check may be lockless but that's ok as we recheck later.
646 * It's possible to migrate LRU pages and balloon pages
647 * Skip any other type of page
648 */
649 if (!PageLRU(page)) {
650 if (unlikely(balloon_page_movable(page))) {
Konstantin Khlebnikovd6d86c02014-10-09 15:29:27 -0700651 if (balloon_page_isolate(page)) {
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800652 /* Successfully isolated */
Joonsoo Kimb6c75012014-04-07 15:37:07 -0700653 goto isolate_success;
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800654 }
655 }
Andrea Arcangelibc835012011-01-13 15:47:08 -0800656 continue;
Rafael Aquinibf6bddf2012-12-11 16:02:42 -0800657 }
Andrea Arcangelibc835012011-01-13 15:47:08 -0800658
659 /*
Mel Gorman2a1402a2012-10-08 16:32:33 -0700660 * PageLRU is set. lru_lock normally excludes isolation
661 * splitting and collapsing (collapsing has already happened
662 * if PageLRU is set) but the lock is not necessarily taken
663 * here and it is wasteful to take it just to check transhuge.
664 * Check TransHuge without lock and skip the whole pageblock if
665 * it's either a transhuge or hugetlbfs page, as calling
666 * compound_order() without preventing THP from splitting the
667 * page underneath us may return surprising results.
Andrea Arcangelibc835012011-01-13 15:47:08 -0800668 */
669 if (PageTransHuge(page)) {
Mel Gorman2a1402a2012-10-08 16:32:33 -0700670 if (!locked)
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700671 low_pfn = ALIGN(low_pfn + 1,
672 pageblock_nr_pages) - 1;
673 else
674 low_pfn += (1 << compound_order(page)) - 1;
675
Mel Gorman2a1402a2012-10-08 16:32:33 -0700676 continue;
677 }
678
David Rientjes119d6d52014-04-03 14:48:00 -0700679 /*
680 * Migration will fail if an anonymous page is pinned in memory,
681 * so avoid taking lru_lock and isolating it unnecessarily in an
682 * admittedly racy check.
683 */
684 if (!page_mapping(page) &&
685 page_count(page) > page_mapcount(page))
686 continue;
687
Vlastimil Babka69b71892014-10-09 15:27:18 -0700688 /* If we already hold the lock, we can skip some rechecking */
689 if (!locked) {
Vlastimil Babka8b44d272014-10-09 15:27:16 -0700690 locked = compact_trylock_irqsave(&zone->lru_lock,
691 &flags, cc);
Vlastimil Babka69b71892014-10-09 15:27:18 -0700692 if (!locked)
693 break;
Mel Gorman2a1402a2012-10-08 16:32:33 -0700694
Vlastimil Babka69b71892014-10-09 15:27:18 -0700695 /* Recheck PageLRU and PageTransHuge under lock */
696 if (!PageLRU(page))
697 continue;
698 if (PageTransHuge(page)) {
699 low_pfn += (1 << compound_order(page)) - 1;
700 continue;
701 }
Andrea Arcangelibc835012011-01-13 15:47:08 -0800702 }
703
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700704 lruvec = mem_cgroup_page_lruvec(page, zone);
705
Mel Gorman748446b2010-05-24 14:32:27 -0700706 /* Try isolate the page */
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700707 if (__isolate_lru_page(page, isolate_mode) != 0)
Mel Gorman748446b2010-05-24 14:32:27 -0700708 continue;
709
Sasha Levin309381fea2014-01-23 15:52:54 -0800710 VM_BUG_ON_PAGE(PageTransCompound(page), page);
Andrea Arcangelibc835012011-01-13 15:47:08 -0800711
Mel Gorman748446b2010-05-24 14:32:27 -0700712 /* Successfully isolated */
Hugh Dickinsfa9add62012-05-29 15:07:09 -0700713 del_page_from_lru_list(page, lruvec, page_lru(page));
Joonsoo Kimb6c75012014-04-07 15:37:07 -0700714
715isolate_success:
Mel Gorman748446b2010-05-24 14:32:27 -0700716 list_add(&page->lru, migratelist);
Mel Gorman748446b2010-05-24 14:32:27 -0700717 cc->nr_migratepages++;
Mel Gormanb7aba692011-01-13 15:45:54 -0800718 nr_isolated++;
Mel Gorman748446b2010-05-24 14:32:27 -0700719
720 /* Avoid isolating too much */
Hillf Danton31b83842012-01-10 15:07:59 -0800721 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
722 ++low_pfn;
Mel Gorman748446b2010-05-24 14:32:27 -0700723 break;
Hillf Danton31b83842012-01-10 15:07:59 -0800724 }
Mel Gorman748446b2010-05-24 14:32:27 -0700725 }
726
Vlastimil Babka99c0fd52014-10-09 15:27:23 -0700727 /*
728 * The PageBuddy() check could have potentially brought us outside
729 * the range to be scanned.
730 */
731 if (unlikely(low_pfn > end_pfn))
732 low_pfn = end_pfn;
733
Mel Gormanc67fe372012-08-21 16:16:17 -0700734 if (locked)
735 spin_unlock_irqrestore(&zone->lru_lock, flags);
Mel Gorman748446b2010-05-24 14:32:27 -0700736
Vlastimil Babka50b5b092014-01-21 15:51:10 -0800737 /*
738 * Update the pageblock-skip information and cached scanner pfn,
739 * if the whole pageblock was scanned without isolating any page.
Vlastimil Babka50b5b092014-01-21 15:51:10 -0800740 */
David Rientjes35979ef2014-06-04 16:08:27 -0700741 if (low_pfn == end_pfn)
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700742 update_pageblock_skip(cc, valid_page, nr_isolated, true);
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700743
Mel Gormanb7aba692011-01-13 15:45:54 -0800744 trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
745
Minchan Kim010fc292012-12-20 15:05:06 -0800746 count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
Mel Gorman397487d2012-10-19 12:00:10 +0100747 if (nr_isolated)
Minchan Kim010fc292012-12-20 15:05:06 -0800748 count_compact_events(COMPACTISOLATED, nr_isolated);
Mel Gorman397487d2012-10-19 12:00:10 +0100749
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100750 return low_pfn;
751}
752
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700753/**
754 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
755 * @cc: Compaction control structure.
756 * @start_pfn: The first PFN to start isolating.
757 * @end_pfn: The one-past-last PFN.
758 *
759 * Returns zero if isolation fails fatally due to e.g. pending signal.
760 * Otherwise, function returns one-past-the-last PFN of isolated page
761 * (which may be greater than end_pfn if end fell in a middle of a THP page).
762 */
763unsigned long
764isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
765 unsigned long end_pfn)
766{
767 unsigned long pfn, block_end_pfn;
768
769 /* Scan block by block. First and last block may be incomplete */
770 pfn = start_pfn;
771 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
772
773 for (; pfn < end_pfn; pfn = block_end_pfn,
774 block_end_pfn += pageblock_nr_pages) {
775
776 block_end_pfn = min(block_end_pfn, end_pfn);
777
Vlastimil Babka7d49d882014-10-09 15:27:11 -0700778 if (!pageblock_pfn_to_page(pfn, block_end_pfn, cc->zone))
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700779 continue;
780
781 pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
782 ISOLATE_UNEVICTABLE);
783
784 /*
785 * In case of fatal failure, release everything that might
786 * have been isolated in the previous iteration, and signal
787 * the failure back to caller.
788 */
789 if (!pfn) {
790 putback_movable_pages(&cc->migratepages);
791 cc->nr_migratepages = 0;
792 break;
793 }
Joonsoo Kim6ea41c02014-10-29 14:50:20 -0700794
795 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
796 break;
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700797 }
798 acct_isolated(cc->zone, cc);
799
800 return pfn;
801}
802
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100803#endif /* CONFIG_COMPACTION || CONFIG_CMA */
804#ifdef CONFIG_COMPACTION
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100805/*
806 * Based on information in the current compact_control, find blocks
807 * suitable for isolating free pages from and then isolate them.
808 */
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700809static void isolate_freepages(struct compact_control *cc)
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100810{
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700811 struct zone *zone = cc->zone;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100812 struct page *page;
Vlastimil Babkac96b9e52014-06-04 16:07:26 -0700813 unsigned long block_start_pfn; /* start of current pageblock */
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700814 unsigned long isolate_start_pfn; /* exact pfn we start at */
Vlastimil Babkac96b9e52014-06-04 16:07:26 -0700815 unsigned long block_end_pfn; /* end of current pageblock */
816 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100817 int nr_freepages = cc->nr_freepages;
818 struct list_head *freelist = &cc->freepages;
819
820 /*
821 * Initialise the free scanner. The starting point is where we last
Vlastimil Babka49e068f2014-05-06 12:50:03 -0700822 * successfully isolated from, zone-cached value, or the end of the
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700823 * zone when isolating for the first time. For looping we also need
824 * this pfn aligned down to the pageblock boundary, because we do
Vlastimil Babkac96b9e52014-06-04 16:07:26 -0700825 * block_start_pfn -= pageblock_nr_pages in the for loop.
826 * For ending point, take care when isolating in last pageblock of a
827 * a zone which ends in the middle of a pageblock.
Vlastimil Babka49e068f2014-05-06 12:50:03 -0700828 * The low boundary is the end of the pageblock the migration scanner
829 * is using.
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100830 */
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700831 isolate_start_pfn = cc->free_pfn;
Vlastimil Babkac96b9e52014-06-04 16:07:26 -0700832 block_start_pfn = cc->free_pfn & ~(pageblock_nr_pages-1);
833 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
834 zone_end_pfn(zone));
Vlastimil Babka7ed695e2014-01-21 15:51:09 -0800835 low_pfn = ALIGN(cc->migrate_pfn + 1, pageblock_nr_pages);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100836
837 /*
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100838 * Isolate free pages until enough are available to migrate the
839 * pages on cc->migratepages. We stop searching if the migrate
840 * and free page scanners meet or enough free pages are isolated.
841 */
Vlastimil Babkac96b9e52014-06-04 16:07:26 -0700842 for (; block_start_pfn >= low_pfn && cc->nr_migratepages > nr_freepages;
843 block_end_pfn = block_start_pfn,
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700844 block_start_pfn -= pageblock_nr_pages,
845 isolate_start_pfn = block_start_pfn) {
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100846 unsigned long isolated;
847
David Rientjesf6ea3ad2013-09-30 13:45:03 -0700848 /*
849 * This can iterate a massively long zone without finding any
850 * suitable migration targets, so periodically check if we need
Vlastimil Babkabe976572014-06-04 16:10:41 -0700851 * to schedule, or even abort async compaction.
David Rientjesf6ea3ad2013-09-30 13:45:03 -0700852 */
Vlastimil Babkabe976572014-06-04 16:10:41 -0700853 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
854 && compact_should_abort(cc))
855 break;
David Rientjesf6ea3ad2013-09-30 13:45:03 -0700856
Vlastimil Babka7d49d882014-10-09 15:27:11 -0700857 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
858 zone);
859 if (!page)
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100860 continue;
861
862 /* Check the block is suitable for migration */
Linus Torvalds68e3e922012-06-03 20:05:57 -0700863 if (!suitable_migration_target(page))
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100864 continue;
Linus Torvalds68e3e922012-06-03 20:05:57 -0700865
Mel Gormanbb13ffe2012-10-08 16:32:41 -0700866 /* If isolation recently failed, do not retry */
867 if (!isolation_suitable(cc, page))
868 continue;
869
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700870 /* Found a block suitable for isolating free pages from. */
871 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
Vlastimil Babkac96b9e52014-06-04 16:07:26 -0700872 block_end_pfn, freelist, false);
Mel Gormanf40d1e42012-10-08 16:32:36 -0700873 nr_freepages += isolated;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100874
875 /*
Vlastimil Babkae14c7202014-10-09 15:27:20 -0700876 * Remember where the free scanner should restart next time,
877 * which is where isolate_freepages_block() left off.
878 * But if it scanned the whole pageblock, isolate_start_pfn
879 * now points at block_end_pfn, which is the start of the next
880 * pageblock.
881 * In that case we will however want to restart at the start
882 * of the previous pageblock.
883 */
884 cc->free_pfn = (isolate_start_pfn < block_end_pfn) ?
885 isolate_start_pfn :
886 block_start_pfn - pageblock_nr_pages;
887
888 /*
Vlastimil Babkabe976572014-06-04 16:10:41 -0700889 * isolate_freepages_block() might have aborted due to async
890 * compaction being contended
891 */
892 if (cc->contended)
893 break;
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100894 }
895
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100896 /* split_free_page does not map the pages */
897 map_pages(freelist);
Michal Nazarewicz2fe86e02012-01-30 13:16:26 +0100898
Vlastimil Babka7ed695e2014-01-21 15:51:09 -0800899 /*
900 * If we crossed the migrate scanner, we want to keep it that way
901 * so that compact_finished() may detect this
902 */
Vlastimil Babkac96b9e52014-06-04 16:07:26 -0700903 if (block_start_pfn < low_pfn)
Vlastimil Babkae9ade562014-06-04 16:08:34 -0700904 cc->free_pfn = cc->migrate_pfn;
Vlastimil Babkac96b9e52014-06-04 16:07:26 -0700905
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100906 cc->nr_freepages = nr_freepages;
Mel Gorman748446b2010-05-24 14:32:27 -0700907}
908
909/*
910 * This is a migrate-callback that "allocates" freepages by taking pages
911 * from the isolated freelists in the block we are migrating to.
912 */
913static struct page *compaction_alloc(struct page *migratepage,
914 unsigned long data,
915 int **result)
916{
917 struct compact_control *cc = (struct compact_control *)data;
918 struct page *freepage;
919
Vlastimil Babkabe976572014-06-04 16:10:41 -0700920 /*
921 * Isolate free pages if necessary, and if we are not aborting due to
922 * contention.
923 */
Mel Gorman748446b2010-05-24 14:32:27 -0700924 if (list_empty(&cc->freepages)) {
Vlastimil Babkabe976572014-06-04 16:10:41 -0700925 if (!cc->contended)
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700926 isolate_freepages(cc);
Mel Gorman748446b2010-05-24 14:32:27 -0700927
928 if (list_empty(&cc->freepages))
929 return NULL;
930 }
931
932 freepage = list_entry(cc->freepages.next, struct page, lru);
933 list_del(&freepage->lru);
934 cc->nr_freepages--;
935
936 return freepage;
937}
938
939/*
David Rientjesd53aea32014-06-04 16:08:26 -0700940 * This is a migrate-callback that "frees" freepages back to the isolated
941 * freelist. All pages on the freelist are from the same zone, so there is no
942 * special handling needed for NUMA.
943 */
944static void compaction_free(struct page *page, unsigned long data)
945{
946 struct compact_control *cc = (struct compact_control *)data;
947
948 list_add(&page->lru, &cc->freepages);
949 cc->nr_freepages++;
950}
951
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100952/* possible outcome of isolate_migratepages */
953typedef enum {
954 ISOLATE_ABORT, /* Abort compaction now */
955 ISOLATE_NONE, /* No pages isolated, continue scanning */
956 ISOLATE_SUCCESS, /* Pages isolated, migrate */
957} isolate_migrate_t;
958
959/*
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700960 * Isolate all pages that can be migrated from the first suitable block,
961 * starting at the block pointed to by the migrate scanner pfn within
962 * compact_control.
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100963 */
964static isolate_migrate_t isolate_migratepages(struct zone *zone,
965 struct compact_control *cc)
966{
967 unsigned long low_pfn, end_pfn;
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700968 struct page *page;
969 const isolate_mode_t isolate_mode =
970 (cc->mode == MIGRATE_ASYNC ? ISOLATE_ASYNC_MIGRATE : 0);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100971
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700972 /*
973 * Start at where we last stopped, or beginning of the zone as
974 * initialized by compact_zone()
975 */
976 low_pfn = cc->migrate_pfn;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100977
978 /* Only scan within a pageblock boundary */
Mel Gormana9aacbc2013-02-22 16:32:25 -0800979 end_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages);
Michal Nazarewiczff9543f2011-12-29 13:09:50 +0100980
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700981 /*
982 * Iterate over whole pageblocks until we find the first suitable.
983 * Do not cross the free scanner.
984 */
985 for (; end_pfn <= cc->free_pfn;
986 low_pfn = end_pfn, end_pfn += pageblock_nr_pages) {
987
988 /*
989 * This can potentially iterate a massively long zone with
990 * many pageblocks unsuitable, so periodically check if we
991 * need to schedule, or even abort async compaction.
992 */
993 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
994 && compact_should_abort(cc))
995 break;
996
Vlastimil Babka7d49d882014-10-09 15:27:11 -0700997 page = pageblock_pfn_to_page(low_pfn, end_pfn, zone);
998 if (!page)
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -0700999 continue;
1000
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001001 /* If isolation recently failed, do not retry */
1002 if (!isolation_suitable(cc, page))
1003 continue;
1004
1005 /*
1006 * For async compaction, also only scan in MOVABLE blocks.
1007 * Async compaction is optimistic to see if the minimum amount
1008 * of work satisfies the allocation.
1009 */
1010 if (cc->mode == MIGRATE_ASYNC &&
1011 !migrate_async_suitable(get_pageblock_migratetype(page)))
1012 continue;
1013
1014 /* Perform the isolation */
1015 low_pfn = isolate_migratepages_block(cc, low_pfn, end_pfn,
1016 isolate_mode);
1017
1018 if (!low_pfn || cc->contended)
1019 return ISOLATE_ABORT;
1020
1021 /*
1022 * Either we isolated something and proceed with migration. Or
1023 * we failed and compact_zone should decide if we should
1024 * continue or not.
1025 */
1026 break;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001027 }
1028
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001029 acct_isolated(zone, cc);
Vlastimil Babka1d5bfe12014-11-13 15:19:30 -08001030 /*
1031 * Record where migration scanner will be restarted. If we end up in
1032 * the same pageblock as the free scanner, make the scanners fully
1033 * meet so that compact_finished() terminates compaction.
1034 */
1035 cc->migrate_pfn = (end_pfn <= cc->free_pfn) ? low_pfn : cc->free_pfn;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001036
Vlastimil Babkaedc2ca62014-10-09 15:27:09 -07001037 return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001038}
1039
David Rientjes6d7ce552014-10-09 15:27:27 -07001040static int compact_finished(struct zone *zone, struct compact_control *cc,
1041 const int migratetype)
Mel Gorman748446b2010-05-24 14:32:27 -07001042{
Mel Gorman8fb74b92013-01-11 14:32:16 -08001043 unsigned int order;
Andrea Arcangeli5a03b052011-01-13 15:47:11 -08001044 unsigned long watermark;
Mel Gorman56de7262010-05-24 14:32:30 -07001045
Vlastimil Babkabe976572014-06-04 16:10:41 -07001046 if (cc->contended || fatal_signal_pending(current))
Mel Gorman748446b2010-05-24 14:32:27 -07001047 return COMPACT_PARTIAL;
1048
Mel Gorman753341a2012-10-08 16:32:40 -07001049 /* Compaction run completes if the migrate and free scanner meet */
Mel Gormanbb13ffe2012-10-08 16:32:41 -07001050 if (cc->free_pfn <= cc->migrate_pfn) {
Vlastimil Babka55b7c4c2014-01-21 15:51:11 -08001051 /* Let the next compaction start anew. */
David Rientjes35979ef2014-06-04 16:08:27 -07001052 zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
1053 zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
Vlastimil Babka55b7c4c2014-01-21 15:51:11 -08001054 zone->compact_cached_free_pfn = zone_end_pfn(zone);
1055
Mel Gorman62997022012-10-08 16:32:47 -07001056 /*
1057 * Mark that the PG_migrate_skip information should be cleared
1058 * by kswapd when it goes to sleep. kswapd does not set the
1059 * flag itself as the decision to be clear should be directly
1060 * based on an allocation request.
1061 */
1062 if (!current_is_kswapd())
1063 zone->compact_blockskip_flush = true;
1064
Mel Gorman748446b2010-05-24 14:32:27 -07001065 return COMPACT_COMPLETE;
Mel Gormanbb13ffe2012-10-08 16:32:41 -07001066 }
Mel Gorman748446b2010-05-24 14:32:27 -07001067
Johannes Weiner82478fb2011-01-20 14:44:21 -08001068 /*
1069 * order == -1 is expected when compacting via
1070 * /proc/sys/vm/compact_memory
1071 */
Mel Gorman56de7262010-05-24 14:32:30 -07001072 if (cc->order == -1)
1073 return COMPACT_CONTINUE;
1074
Michal Hocko3957c772011-06-15 15:08:25 -07001075 /* Compaction run is not finished if the watermark is not met */
1076 watermark = low_wmark_pages(zone);
Michal Hocko3957c772011-06-15 15:08:25 -07001077
Vlastimil Babkaebff3982014-12-10 15:43:22 -08001078 if (!zone_watermark_ok(zone, cc->order, watermark, cc->classzone_idx,
1079 cc->alloc_flags))
Michal Hocko3957c772011-06-15 15:08:25 -07001080 return COMPACT_CONTINUE;
1081
Mel Gorman56de7262010-05-24 14:32:30 -07001082 /* Direct compactor: Is a suitable page free? */
Mel Gorman8fb74b92013-01-11 14:32:16 -08001083 for (order = cc->order; order < MAX_ORDER; order++) {
1084 struct free_area *area = &zone->free_area[order];
Mel Gorman56de7262010-05-24 14:32:30 -07001085
Mel Gorman8fb74b92013-01-11 14:32:16 -08001086 /* Job done if page is free of the right migratetype */
David Rientjes6d7ce552014-10-09 15:27:27 -07001087 if (!list_empty(&area->free_list[migratetype]))
Mel Gorman8fb74b92013-01-11 14:32:16 -08001088 return COMPACT_PARTIAL;
1089
1090 /* Job done if allocation would set block type */
1091 if (cc->order >= pageblock_order && area->nr_free)
1092 return COMPACT_PARTIAL;
Mel Gorman56de7262010-05-24 14:32:30 -07001093 }
1094
Mel Gorman748446b2010-05-24 14:32:27 -07001095 return COMPACT_CONTINUE;
1096}
1097
Mel Gorman3e7d3442011-01-13 15:45:56 -08001098/*
1099 * compaction_suitable: Is this suitable to run compaction on this zone now?
1100 * Returns
1101 * COMPACT_SKIPPED - If there are too few free pages for compaction
1102 * COMPACT_PARTIAL - If the allocation would succeed without compaction
1103 * COMPACT_CONTINUE - If compaction should run now
1104 */
Vlastimil Babkaebff3982014-12-10 15:43:22 -08001105unsigned long compaction_suitable(struct zone *zone, int order,
1106 int alloc_flags, int classzone_idx)
Mel Gorman3e7d3442011-01-13 15:45:56 -08001107{
1108 int fragindex;
1109 unsigned long watermark;
1110
1111 /*
Michal Hocko3957c772011-06-15 15:08:25 -07001112 * order == -1 is expected when compacting via
1113 * /proc/sys/vm/compact_memory
1114 */
1115 if (order == -1)
1116 return COMPACT_CONTINUE;
1117
Vlastimil Babkaebff3982014-12-10 15:43:22 -08001118 watermark = low_wmark_pages(zone);
1119 /*
1120 * If watermarks for high-order allocation are already met, there
1121 * should be no need for compaction at all.
1122 */
1123 if (zone_watermark_ok(zone, order, watermark, classzone_idx,
1124 alloc_flags))
1125 return COMPACT_PARTIAL;
1126
Michal Hocko3957c772011-06-15 15:08:25 -07001127 /*
Mel Gorman3e7d3442011-01-13 15:45:56 -08001128 * Watermarks for order-0 must be met for compaction. Note the 2UL.
1129 * This is because during migration, copies of pages need to be
1130 * allocated and for a short time, the footprint is higher
1131 */
Vlastimil Babkaebff3982014-12-10 15:43:22 -08001132 watermark += (2UL << order);
1133 if (!zone_watermark_ok(zone, 0, watermark, classzone_idx, alloc_flags))
Mel Gorman3e7d3442011-01-13 15:45:56 -08001134 return COMPACT_SKIPPED;
1135
1136 /*
1137 * fragmentation index determines if allocation failures are due to
1138 * low memory or external fragmentation
1139 *
Vlastimil Babkaebff3982014-12-10 15:43:22 -08001140 * index of -1000 would imply allocations might succeed depending on
1141 * watermarks, but we already failed the high-order watermark check
Mel Gorman3e7d3442011-01-13 15:45:56 -08001142 * index towards 0 implies failure is due to lack of memory
1143 * index towards 1000 implies failure is due to fragmentation
1144 *
1145 * Only compact if a failure would be due to fragmentation.
1146 */
1147 fragindex = fragmentation_index(zone, order);
1148 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
1149 return COMPACT_SKIPPED;
1150
Mel Gorman3e7d3442011-01-13 15:45:56 -08001151 return COMPACT_CONTINUE;
1152}
1153
Mel Gorman748446b2010-05-24 14:32:27 -07001154static int compact_zone(struct zone *zone, struct compact_control *cc)
1155{
1156 int ret;
Mel Gormanc89511a2012-10-08 16:32:45 -07001157 unsigned long start_pfn = zone->zone_start_pfn;
Cody P Schafer108bcc92013-02-22 16:35:23 -08001158 unsigned long end_pfn = zone_end_pfn(zone);
David Rientjes6d7ce552014-10-09 15:27:27 -07001159 const int migratetype = gfpflags_to_migratetype(cc->gfp_mask);
David Rientjese0b9dae2014-06-04 16:08:28 -07001160 const bool sync = cc->mode != MIGRATE_ASYNC;
Vlastimil Babkafdaf7f52014-12-10 15:43:34 -08001161 unsigned long last_migrated_pfn = 0;
Mel Gorman748446b2010-05-24 14:32:27 -07001162
Vlastimil Babkaebff3982014-12-10 15:43:22 -08001163 ret = compaction_suitable(zone, cc->order, cc->alloc_flags,
1164 cc->classzone_idx);
Mel Gorman3e7d3442011-01-13 15:45:56 -08001165 switch (ret) {
1166 case COMPACT_PARTIAL:
1167 case COMPACT_SKIPPED:
1168 /* Compaction is likely to fail */
1169 return ret;
1170 case COMPACT_CONTINUE:
1171 /* Fall through to compaction */
1172 ;
1173 }
1174
Mel Gormanc89511a2012-10-08 16:32:45 -07001175 /*
Vlastimil Babkad3132e42014-01-21 15:51:08 -08001176 * Clear pageblock skip if there were failures recently and compaction
1177 * is about to be retried after being deferred. kswapd does not do
1178 * this reset as it'll reset the cached information when going to sleep.
1179 */
1180 if (compaction_restarting(zone, cc->order) && !current_is_kswapd())
1181 __reset_isolation_suitable(zone);
1182
1183 /*
Mel Gormanc89511a2012-10-08 16:32:45 -07001184 * Setup to move all movable pages to the end of the zone. Used cached
1185 * information on where the scanners should start but check that it
1186 * is initialised by ensuring the values are within zone boundaries.
1187 */
David Rientjese0b9dae2014-06-04 16:08:28 -07001188 cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync];
Mel Gormanc89511a2012-10-08 16:32:45 -07001189 cc->free_pfn = zone->compact_cached_free_pfn;
1190 if (cc->free_pfn < start_pfn || cc->free_pfn > end_pfn) {
1191 cc->free_pfn = end_pfn & ~(pageblock_nr_pages-1);
1192 zone->compact_cached_free_pfn = cc->free_pfn;
1193 }
1194 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn > end_pfn) {
1195 cc->migrate_pfn = start_pfn;
David Rientjes35979ef2014-06-04 16:08:27 -07001196 zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
1197 zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
Mel Gormanc89511a2012-10-08 16:32:45 -07001198 }
Mel Gorman748446b2010-05-24 14:32:27 -07001199
Mel Gorman0eb927c2014-01-21 15:51:05 -08001200 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn, cc->free_pfn, end_pfn);
1201
Mel Gorman748446b2010-05-24 14:32:27 -07001202 migrate_prep_local();
1203
David Rientjes6d7ce552014-10-09 15:27:27 -07001204 while ((ret = compact_finished(zone, cc, migratetype)) ==
1205 COMPACT_CONTINUE) {
Minchan Kim9d502c12011-03-22 16:30:39 -07001206 int err;
Vlastimil Babkafdaf7f52014-12-10 15:43:34 -08001207 unsigned long isolate_start_pfn = cc->migrate_pfn;
Mel Gorman748446b2010-05-24 14:32:27 -07001208
Mel Gormanf9e35b32011-06-15 15:08:52 -07001209 switch (isolate_migratepages(zone, cc)) {
1210 case ISOLATE_ABORT:
1211 ret = COMPACT_PARTIAL;
Rafael Aquini5733c7d2012-12-11 16:02:47 -08001212 putback_movable_pages(&cc->migratepages);
Shaohua Lie64c5232012-10-08 16:32:27 -07001213 cc->nr_migratepages = 0;
Mel Gormanf9e35b32011-06-15 15:08:52 -07001214 goto out;
1215 case ISOLATE_NONE:
Vlastimil Babkafdaf7f52014-12-10 15:43:34 -08001216 /*
1217 * We haven't isolated and migrated anything, but
1218 * there might still be unflushed migrations from
1219 * previous cc->order aligned block.
1220 */
1221 goto check_drain;
Mel Gormanf9e35b32011-06-15 15:08:52 -07001222 case ISOLATE_SUCCESS:
1223 ;
1224 }
Mel Gorman748446b2010-05-24 14:32:27 -07001225
David Rientjesd53aea32014-06-04 16:08:26 -07001226 err = migrate_pages(&cc->migratepages, compaction_alloc,
David Rientjese0b9dae2014-06-04 16:08:28 -07001227 compaction_free, (unsigned long)cc, cc->mode,
Mel Gorman7b2a2d42012-10-19 14:07:31 +01001228 MR_COMPACTION);
Mel Gorman748446b2010-05-24 14:32:27 -07001229
Vlastimil Babkaf8c93012014-06-04 16:08:32 -07001230 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
1231 &cc->migratepages);
Mel Gorman748446b2010-05-24 14:32:27 -07001232
Vlastimil Babkaf8c93012014-06-04 16:08:32 -07001233 /* All pages were either migrated or will be released */
1234 cc->nr_migratepages = 0;
Minchan Kim9d502c12011-03-22 16:30:39 -07001235 if (err) {
Rafael Aquini5733c7d2012-12-11 16:02:47 -08001236 putback_movable_pages(&cc->migratepages);
Vlastimil Babka7ed695e2014-01-21 15:51:09 -08001237 /*
1238 * migrate_pages() may return -ENOMEM when scanners meet
1239 * and we want compact_finished() to detect it
1240 */
1241 if (err == -ENOMEM && cc->free_pfn > cc->migrate_pfn) {
David Rientjes4bf2bba2012-07-11 14:02:13 -07001242 ret = COMPACT_PARTIAL;
1243 goto out;
1244 }
Mel Gorman748446b2010-05-24 14:32:27 -07001245 }
Vlastimil Babkafdaf7f52014-12-10 15:43:34 -08001246
1247 /*
1248 * Record where we could have freed pages by migration and not
1249 * yet flushed them to buddy allocator. We use the pfn that
1250 * isolate_migratepages() started from in this loop iteration
1251 * - this is the lowest page that could have been isolated and
1252 * then freed by migration.
1253 */
1254 if (!last_migrated_pfn)
1255 last_migrated_pfn = isolate_start_pfn;
1256
1257check_drain:
1258 /*
1259 * Has the migration scanner moved away from the previous
1260 * cc->order aligned block where we migrated from? If yes,
1261 * flush the pages that were freed, so that they can merge and
1262 * compact_finished() can detect immediately if allocation
1263 * would succeed.
1264 */
1265 if (cc->order > 0 && last_migrated_pfn) {
1266 int cpu;
1267 unsigned long current_block_start =
1268 cc->migrate_pfn & ~((1UL << cc->order) - 1);
1269
1270 if (last_migrated_pfn < current_block_start) {
1271 cpu = get_cpu();
1272 lru_add_drain_cpu(cpu);
1273 drain_local_pages(zone);
1274 put_cpu();
1275 /* No more flushing until we migrate again */
1276 last_migrated_pfn = 0;
1277 }
1278 }
1279
Mel Gorman748446b2010-05-24 14:32:27 -07001280 }
1281
Mel Gormanf9e35b32011-06-15 15:08:52 -07001282out:
Vlastimil Babka6bace092014-12-10 15:43:31 -08001283 /*
1284 * Release free pages and update where the free scanner should restart,
1285 * so we don't leave any returned pages behind in the next attempt.
1286 */
1287 if (cc->nr_freepages > 0) {
1288 unsigned long free_pfn = release_freepages(&cc->freepages);
1289
1290 cc->nr_freepages = 0;
1291 VM_BUG_ON(free_pfn == 0);
1292 /* The cached pfn is always the first in a pageblock */
1293 free_pfn &= ~(pageblock_nr_pages-1);
1294 /*
1295 * Only go back, not forward. The cached pfn might have been
1296 * already reset to zone end in compact_finished()
1297 */
1298 if (free_pfn > zone->compact_cached_free_pfn)
1299 zone->compact_cached_free_pfn = free_pfn;
1300 }
Mel Gorman748446b2010-05-24 14:32:27 -07001301
Mel Gorman0eb927c2014-01-21 15:51:05 -08001302 trace_mm_compaction_end(ret);
1303
Mel Gorman748446b2010-05-24 14:32:27 -07001304 return ret;
1305}
Mel Gorman76ab0f52010-05-24 14:32:28 -07001306
David Rientjese0b9dae2014-06-04 16:08:28 -07001307static unsigned long compact_zone_order(struct zone *zone, int order,
Vlastimil Babkaebff3982014-12-10 15:43:22 -08001308 gfp_t gfp_mask, enum migrate_mode mode, int *contended,
1309 int alloc_flags, int classzone_idx)
Mel Gorman56de7262010-05-24 14:32:30 -07001310{
Shaohua Lie64c5232012-10-08 16:32:27 -07001311 unsigned long ret;
Mel Gorman56de7262010-05-24 14:32:30 -07001312 struct compact_control cc = {
1313 .nr_freepages = 0,
1314 .nr_migratepages = 0,
1315 .order = order,
David Rientjes6d7ce552014-10-09 15:27:27 -07001316 .gfp_mask = gfp_mask,
Mel Gorman56de7262010-05-24 14:32:30 -07001317 .zone = zone,
David Rientjese0b9dae2014-06-04 16:08:28 -07001318 .mode = mode,
Vlastimil Babkaebff3982014-12-10 15:43:22 -08001319 .alloc_flags = alloc_flags,
1320 .classzone_idx = classzone_idx,
Mel Gorman56de7262010-05-24 14:32:30 -07001321 };
1322 INIT_LIST_HEAD(&cc.freepages);
1323 INIT_LIST_HEAD(&cc.migratepages);
1324
Shaohua Lie64c5232012-10-08 16:32:27 -07001325 ret = compact_zone(zone, &cc);
1326
1327 VM_BUG_ON(!list_empty(&cc.freepages));
1328 VM_BUG_ON(!list_empty(&cc.migratepages));
1329
1330 *contended = cc.contended;
1331 return ret;
Mel Gorman56de7262010-05-24 14:32:30 -07001332}
1333
Mel Gorman5e771902010-05-24 14:32:31 -07001334int sysctl_extfrag_threshold = 500;
1335
Mel Gorman56de7262010-05-24 14:32:30 -07001336/**
1337 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
Mel Gorman56de7262010-05-24 14:32:30 -07001338 * @gfp_mask: The GFP mask of the current allocation
Vlastimil Babka1a6d53a2015-02-11 15:25:44 -08001339 * @order: The order of the current allocation
1340 * @alloc_flags: The allocation flags of the current allocation
1341 * @ac: The context of current allocation
David Rientjese0b9dae2014-06-04 16:08:28 -07001342 * @mode: The migration mode for async, sync light, or sync migration
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001343 * @contended: Return value that determines if compaction was aborted due to
1344 * need_resched() or lock contention
Mel Gorman56de7262010-05-24 14:32:30 -07001345 *
1346 * This is the main entry point for direct page compaction.
1347 */
Vlastimil Babka1a6d53a2015-02-11 15:25:44 -08001348unsigned long try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
1349 int alloc_flags, const struct alloc_context *ac,
1350 enum migrate_mode mode, int *contended)
Mel Gorman56de7262010-05-24 14:32:30 -07001351{
Mel Gorman56de7262010-05-24 14:32:30 -07001352 int may_enter_fs = gfp_mask & __GFP_FS;
1353 int may_perform_io = gfp_mask & __GFP_IO;
Mel Gorman56de7262010-05-24 14:32:30 -07001354 struct zoneref *z;
1355 struct zone *zone;
Vlastimil Babka53853e22014-10-09 15:27:02 -07001356 int rc = COMPACT_DEFERRED;
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001357 int all_zones_contended = COMPACT_CONTENDED_LOCK; /* init for &= op */
1358
1359 *contended = COMPACT_CONTENDED_NONE;
Mel Gorman56de7262010-05-24 14:32:30 -07001360
Mel Gorman4ffb6332012-10-08 16:29:09 -07001361 /* Check if the GFP flags allow compaction */
Andrea Arcangelic5a73c32011-01-13 15:47:11 -08001362 if (!order || !may_enter_fs || !may_perform_io)
Vlastimil Babka53853e22014-10-09 15:27:02 -07001363 return COMPACT_SKIPPED;
Mel Gorman56de7262010-05-24 14:32:30 -07001364
1365 /* Compact each zone in the list */
Vlastimil Babka1a6d53a2015-02-11 15:25:44 -08001366 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1367 ac->nodemask) {
Mel Gorman56de7262010-05-24 14:32:30 -07001368 int status;
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001369 int zone_contended;
Mel Gorman56de7262010-05-24 14:32:30 -07001370
Vlastimil Babka53853e22014-10-09 15:27:02 -07001371 if (compaction_deferred(zone, order))
1372 continue;
1373
David Rientjese0b9dae2014-06-04 16:08:28 -07001374 status = compact_zone_order(zone, order, gfp_mask, mode,
Vlastimil Babka1a6d53a2015-02-11 15:25:44 -08001375 &zone_contended, alloc_flags,
1376 ac->classzone_idx);
Mel Gorman56de7262010-05-24 14:32:30 -07001377 rc = max(status, rc);
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001378 /*
1379 * It takes at least one zone that wasn't lock contended
1380 * to clear all_zones_contended.
1381 */
1382 all_zones_contended &= zone_contended;
Mel Gorman56de7262010-05-24 14:32:30 -07001383
Mel Gorman3e7d3442011-01-13 15:45:56 -08001384 /* If a normal allocation would succeed, stop compacting */
Vlastimil Babkaebff3982014-12-10 15:43:22 -08001385 if (zone_watermark_ok(zone, order, low_wmark_pages(zone),
Vlastimil Babka1a6d53a2015-02-11 15:25:44 -08001386 ac->classzone_idx, alloc_flags)) {
Vlastimil Babka53853e22014-10-09 15:27:02 -07001387 /*
1388 * We think the allocation will succeed in this zone,
1389 * but it is not certain, hence the false. The caller
1390 * will repeat this with true if allocation indeed
1391 * succeeds in this zone.
1392 */
1393 compaction_defer_reset(zone, order, false);
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001394 /*
1395 * It is possible that async compaction aborted due to
1396 * need_resched() and the watermarks were ok thanks to
1397 * somebody else freeing memory. The allocation can
1398 * however still fail so we better signal the
1399 * need_resched() contention anyway (this will not
1400 * prevent the allocation attempt).
1401 */
1402 if (zone_contended == COMPACT_CONTENDED_SCHED)
1403 *contended = COMPACT_CONTENDED_SCHED;
1404
1405 goto break_loop;
1406 }
1407
Vlastimil Babkaf8669792014-12-10 15:43:28 -08001408 if (mode != MIGRATE_ASYNC && status == COMPACT_COMPLETE) {
Vlastimil Babka53853e22014-10-09 15:27:02 -07001409 /*
1410 * We think that allocation won't succeed in this zone
1411 * so we defer compaction there. If it ends up
1412 * succeeding after all, it will be reset.
1413 */
1414 defer_compaction(zone, order);
1415 }
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001416
1417 /*
1418 * We might have stopped compacting due to need_resched() in
1419 * async compaction, or due to a fatal signal detected. In that
1420 * case do not try further zones and signal need_resched()
1421 * contention.
1422 */
1423 if ((zone_contended == COMPACT_CONTENDED_SCHED)
1424 || fatal_signal_pending(current)) {
1425 *contended = COMPACT_CONTENDED_SCHED;
1426 goto break_loop;
1427 }
1428
1429 continue;
1430break_loop:
1431 /*
1432 * We might not have tried all the zones, so be conservative
1433 * and assume they are not all lock contended.
1434 */
1435 all_zones_contended = 0;
1436 break;
Mel Gorman56de7262010-05-24 14:32:30 -07001437 }
1438
Vlastimil Babka1f9efde2014-10-09 15:27:14 -07001439 /*
1440 * If at least one zone wasn't deferred or skipped, we report if all
1441 * zones that were tried were lock contended.
1442 */
1443 if (rc > COMPACT_SKIPPED && all_zones_contended)
1444 *contended = COMPACT_CONTENDED_LOCK;
1445
Mel Gorman56de7262010-05-24 14:32:30 -07001446 return rc;
1447}
1448
1449
Mel Gorman76ab0f52010-05-24 14:32:28 -07001450/* Compact all zones within a node */
Andrew Morton7103f162013-02-22 16:32:33 -08001451static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
Mel Gorman76ab0f52010-05-24 14:32:28 -07001452{
1453 int zoneid;
Mel Gorman76ab0f52010-05-24 14:32:28 -07001454 struct zone *zone;
1455
Mel Gorman76ab0f52010-05-24 14:32:28 -07001456 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
Mel Gorman76ab0f52010-05-24 14:32:28 -07001457
1458 zone = &pgdat->node_zones[zoneid];
1459 if (!populated_zone(zone))
1460 continue;
1461
Rik van Riel7be62de2012-03-21 16:33:52 -07001462 cc->nr_freepages = 0;
1463 cc->nr_migratepages = 0;
1464 cc->zone = zone;
1465 INIT_LIST_HEAD(&cc->freepages);
1466 INIT_LIST_HEAD(&cc->migratepages);
Mel Gorman76ab0f52010-05-24 14:32:28 -07001467
Dan Carpenteraad6ec32012-03-21 16:33:54 -07001468 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
Rik van Riel7be62de2012-03-21 16:33:52 -07001469 compact_zone(zone, cc);
Mel Gorman76ab0f52010-05-24 14:32:28 -07001470
Rik van Rielaff62242012-03-21 16:33:52 -07001471 if (cc->order > 0) {
Vlastimil Babkade6c60a2014-01-21 15:51:07 -08001472 if (zone_watermark_ok(zone, cc->order,
1473 low_wmark_pages(zone), 0, 0))
1474 compaction_defer_reset(zone, cc->order, false);
Rik van Rielaff62242012-03-21 16:33:52 -07001475 }
1476
Rik van Riel7be62de2012-03-21 16:33:52 -07001477 VM_BUG_ON(!list_empty(&cc->freepages));
1478 VM_BUG_ON(!list_empty(&cc->migratepages));
Mel Gorman76ab0f52010-05-24 14:32:28 -07001479 }
Mel Gorman76ab0f52010-05-24 14:32:28 -07001480}
1481
Andrew Morton7103f162013-02-22 16:32:33 -08001482void compact_pgdat(pg_data_t *pgdat, int order)
Rik van Riel7be62de2012-03-21 16:33:52 -07001483{
1484 struct compact_control cc = {
1485 .order = order,
David Rientjese0b9dae2014-06-04 16:08:28 -07001486 .mode = MIGRATE_ASYNC,
Rik van Riel7be62de2012-03-21 16:33:52 -07001487 };
1488
Mel Gorman3a7200a2013-09-11 14:22:19 -07001489 if (!order)
1490 return;
1491
Andrew Morton7103f162013-02-22 16:32:33 -08001492 __compact_pgdat(pgdat, &cc);
Rik van Riel7be62de2012-03-21 16:33:52 -07001493}
1494
Andrew Morton7103f162013-02-22 16:32:33 -08001495static void compact_node(int nid)
Rik van Riel7be62de2012-03-21 16:33:52 -07001496{
Rik van Riel7be62de2012-03-21 16:33:52 -07001497 struct compact_control cc = {
1498 .order = -1,
David Rientjese0b9dae2014-06-04 16:08:28 -07001499 .mode = MIGRATE_SYNC,
David Rientjes91ca9182014-04-03 14:47:23 -07001500 .ignore_skip_hint = true,
Rik van Riel7be62de2012-03-21 16:33:52 -07001501 };
1502
Andrew Morton7103f162013-02-22 16:32:33 -08001503 __compact_pgdat(NODE_DATA(nid), &cc);
Rik van Riel7be62de2012-03-21 16:33:52 -07001504}
1505
Mel Gorman76ab0f52010-05-24 14:32:28 -07001506/* Compact all nodes in the system */
Jason Liu7964c062013-01-11 14:31:47 -08001507static void compact_nodes(void)
Mel Gorman76ab0f52010-05-24 14:32:28 -07001508{
1509 int nid;
1510
Hugh Dickins8575ec22012-03-21 16:33:53 -07001511 /* Flush pending updates to the LRU lists */
1512 lru_add_drain_all();
1513
Mel Gorman76ab0f52010-05-24 14:32:28 -07001514 for_each_online_node(nid)
1515 compact_node(nid);
Mel Gorman76ab0f52010-05-24 14:32:28 -07001516}
1517
1518/* The written value is actually unused, all memory is compacted */
1519int sysctl_compact_memory;
1520
1521/* This is the entry point for compacting all nodes via /proc/sys/vm */
1522int sysctl_compaction_handler(struct ctl_table *table, int write,
1523 void __user *buffer, size_t *length, loff_t *ppos)
1524{
1525 if (write)
Jason Liu7964c062013-01-11 14:31:47 -08001526 compact_nodes();
Mel Gorman76ab0f52010-05-24 14:32:28 -07001527
1528 return 0;
1529}
Mel Gormaned4a6d72010-05-24 14:32:29 -07001530
Mel Gorman5e771902010-05-24 14:32:31 -07001531int sysctl_extfrag_handler(struct ctl_table *table, int write,
1532 void __user *buffer, size_t *length, loff_t *ppos)
1533{
1534 proc_dointvec_minmax(table, write, buffer, length, ppos);
1535
1536 return 0;
1537}
1538
Mel Gormaned4a6d72010-05-24 14:32:29 -07001539#if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
Rashika Kheria74e77fb2014-04-03 14:48:01 -07001540static ssize_t sysfs_compact_node(struct device *dev,
Kay Sievers10fbcf42011-12-21 14:48:43 -08001541 struct device_attribute *attr,
Mel Gormaned4a6d72010-05-24 14:32:29 -07001542 const char *buf, size_t count)
1543{
Hugh Dickins8575ec22012-03-21 16:33:53 -07001544 int nid = dev->id;
1545
1546 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1547 /* Flush pending updates to the LRU lists */
1548 lru_add_drain_all();
1549
1550 compact_node(nid);
1551 }
Mel Gormaned4a6d72010-05-24 14:32:29 -07001552
1553 return count;
1554}
Kay Sievers10fbcf42011-12-21 14:48:43 -08001555static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
Mel Gormaned4a6d72010-05-24 14:32:29 -07001556
1557int compaction_register_node(struct node *node)
1558{
Kay Sievers10fbcf42011-12-21 14:48:43 -08001559 return device_create_file(&node->dev, &dev_attr_compact);
Mel Gormaned4a6d72010-05-24 14:32:29 -07001560}
1561
1562void compaction_unregister_node(struct node *node)
1563{
Kay Sievers10fbcf42011-12-21 14:48:43 -08001564 return device_remove_file(&node->dev, &dev_attr_compact);
Mel Gormaned4a6d72010-05-24 14:32:29 -07001565}
1566#endif /* CONFIG_SYSFS && CONFIG_NUMA */
Michal Nazarewiczff9543f2011-12-29 13:09:50 +01001567
1568#endif /* CONFIG_COMPACTION */