Merge tag 'usb-3.18-rc3' of git://git.kernel.org/pub/scm/linux/kernel/git/gregkh/usb
[sfrench/cifs-2.6.git] / mm / compaction.c
1 /*
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>
15 #include <linux/sysctl.h>
16 #include <linux/sysfs.h>
17 #include <linux/balloon_compaction.h>
18 #include <linux/page-isolation.h>
19 #include "internal.h"
20
21 #ifdef CONFIG_COMPACTION
22 static inline void count_compact_event(enum vm_event_item item)
23 {
24         count_vm_event(item);
25 }
26
27 static 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
36 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
37
38 #define CREATE_TRACE_POINTS
39 #include <trace/events/compaction.h>
40
41 static 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
55 static 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
65 static inline bool migrate_async_suitable(int migratetype)
66 {
67         return is_migrate_cma(migratetype) || migratetype == MIGRATE_MOVABLE;
68 }
69
70 /*
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  */
87 static 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
113 #ifdef CONFIG_COMPACTION
114 /* Returns true if the pageblock should be scanned for pages to isolate. */
115 static 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  */
129 static void __reset_isolation_suitable(struct zone *zone)
130 {
131         unsigned long start_pfn = zone->zone_start_pfn;
132         unsigned long end_pfn = zone_end_pfn(zone);
133         unsigned long pfn;
134
135         zone->compact_cached_migrate_pfn[0] = start_pfn;
136         zone->compact_cached_migrate_pfn[1] = start_pfn;
137         zone->compact_cached_free_pfn = end_pfn;
138         zone->compact_blockskip_flush = false;
139
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
157 void 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
172 /*
173  * If no pages were isolated then mark this pageblock to be skipped in the
174  * future. The information is later cleared by __reset_isolation_suitable().
175  */
176 static void update_pageblock_skip(struct compact_control *cc,
177                         struct page *page, unsigned long nr_isolated,
178                         bool migrate_scanner)
179 {
180         struct zone *zone = cc->zone;
181         unsigned long pfn;
182
183         if (cc->ignore_skip_hint)
184                 return;
185
186         if (!page)
187                 return;
188
189         if (nr_isolated)
190                 return;
191
192         set_pageblock_skip(page);
193
194         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;
202                 if (cc->mode != MIGRATE_ASYNC &&
203                     pfn > zone->compact_cached_migrate_pfn[1])
204                         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;
210         }
211 }
212 #else
213 static inline bool isolation_suitable(struct compact_control *cc,
214                                         struct page *page)
215 {
216         return true;
217 }
218
219 static void update_pageblock_skip(struct compact_control *cc,
220                         struct page *page, unsigned long nr_isolated,
221                         bool migrate_scanner)
222 {
223 }
224 #endif /* CONFIG_COMPACTION */
225
226 /*
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  */
234 static bool compact_trylock_irqsave(spinlock_t *lock, unsigned long *flags,
235                                                 struct compact_control *cc)
236 {
237         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         }
245
246         return true;
247 }
248
249 /*
250  * Compaction requires the taking of some coarse locks that are potentially
251  * 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.
258  *
259  * 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)
263  */
264 static bool compact_unlock_should_abort(spinlock_t *lock,
265                 unsigned long flags, bool *locked, struct compact_control *cc)
266 {
267         if (*locked) {
268                 spin_unlock_irqrestore(lock, flags);
269                 *locked = false;
270         }
271
272         if (fatal_signal_pending(current)) {
273                 cc->contended = COMPACT_CONTENDED_SCHED;
274                 return true;
275         }
276
277         if (need_resched()) {
278                 if (cc->mode == MIGRATE_ASYNC) {
279                         cc->contended = COMPACT_CONTENDED_SCHED;
280                         return true;
281                 }
282                 cond_resched();
283         }
284
285         return false;
286 }
287
288 /*
289  * Aside from avoiding lock contention, compaction also periodically checks
290  * need_resched() and either schedules in sync compaction or aborts async
291  * compaction. This is similar to what compact_unlock_should_abort() does, but
292  * 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  */
297 static 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) {
302                         cc->contended = COMPACT_CONTENDED_SCHED;
303                         return true;
304                 }
305
306                 cond_resched();
307         }
308
309         return false;
310 }
311
312 /* Returns true if the page is within a block suitable for migration to */
313 static bool suitable_migration_target(struct page *page)
314 {
315         /* If the page is a large free page, then disallow migration */
316         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         }
325
326         /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
327         if (migrate_async_suitable(get_pageblock_migratetype(page)))
328                 return true;
329
330         /* Otherwise skip the block */
331         return false;
332 }
333
334 /*
335  * 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).
338  */
339 static unsigned long isolate_freepages_block(struct compact_control *cc,
340                                 unsigned long *start_pfn,
341                                 unsigned long end_pfn,
342                                 struct list_head *freelist,
343                                 bool strict)
344 {
345         int nr_scanned = 0, total_isolated = 0;
346         struct page *cursor, *valid_page = NULL;
347         unsigned long flags = 0;
348         bool locked = false;
349         unsigned long blockpfn = *start_pfn;
350
351         cursor = pfn_to_page(blockpfn);
352
353         /* Isolate free pages. */
354         for (; blockpfn < end_pfn; blockpfn++, cursor++) {
355                 int isolated, i;
356                 struct page *page = cursor;
357
358                 /*
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
368                 nr_scanned++;
369                 if (!pfn_valid_within(blockpfn))
370                         goto isolate_fail;
371
372                 if (!valid_page)
373                         valid_page = page;
374                 if (!PageBuddy(page))
375                         goto isolate_fail;
376
377                 /*
378                  * 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.
383                  */
384                 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                          */
393                         locked = compact_trylock_irqsave(&cc->zone->lock,
394                                                                 &flags, cc);
395                         if (!locked)
396                                 break;
397
398                         /* Recheck this is a buddy page under lock */
399                         if (!PageBuddy(page))
400                                 goto isolate_fail;
401                 }
402
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;
415                         continue;
416                 }
417
418 isolate_fail:
419                 if (strict)
420                         break;
421                 else
422                         continue;
423
424         }
425
426         /* Record how far we have got within the block */
427         *start_pfn = blockpfn;
428
429         trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
430
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          */
436         if (strict && blockpfn < end_pfn)
437                 total_isolated = 0;
438
439         if (locked)
440                 spin_unlock_irqrestore(&cc->zone->lock, flags);
441
442         /* Update the pageblock-skip if the whole pageblock was scanned */
443         if (blockpfn == end_pfn)
444                 update_pageblock_skip(cc, valid_page, total_isolated, false);
445
446         count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
447         if (total_isolated)
448                 count_compact_events(COMPACTISOLATED, total_isolated);
449         return total_isolated;
450 }
451
452 /**
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  */
465 unsigned long
466 isolate_freepages_range(struct compact_control *cc,
467                         unsigned long start_pfn, unsigned long end_pfn)
468 {
469         unsigned long isolated, pfn, block_end_pfn;
470         LIST_HEAD(freelist);
471
472         pfn = start_pfn;
473         block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
474
475         for (; pfn < end_pfn; pfn += isolated,
476                                 block_end_pfn += pageblock_nr_pages) {
477                 /* Protect pfn from changing by isolate_freepages_block */
478                 unsigned long isolate_start_pfn = pfn;
479
480                 block_end_pfn = min(block_end_pfn, end_pfn);
481
482                 if (!pageblock_pfn_to_page(pfn, block_end_pfn, cc->zone))
483                         break;
484
485                 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
486                                                 block_end_pfn, &freelist, true);
487
488                 /*
489                  * In strict mode, isolate_freepages_block() returns 0 if
490                  * there are any holes in the block (ie. invalid PFNs or
491                  * non-free pages).
492                  */
493                 if (!isolated)
494                         break;
495
496                 /*
497                  * If we managed to isolate pages, it is always (1 << n) *
498                  * pageblock_nr_pages for some non-negative n.  (Max order
499                  * page may span two pageblocks).
500                  */
501         }
502
503         /* split_free_page does not map the pages */
504         map_pages(&freelist);
505
506         if (pfn < end_pfn) {
507                 /* Loop terminated early, cleanup. */
508                 release_freepages(&freelist);
509                 return 0;
510         }
511
512         /* We don't use freelists for anything. */
513         return pfn;
514 }
515
516 /* Update the number of anon and file isolated pages in the zone */
517 static void acct_isolated(struct zone *zone, struct compact_control *cc)
518 {
519         struct page *page;
520         unsigned int count[2] = { 0, };
521
522         if (list_empty(&cc->migratepages))
523                 return;
524
525         list_for_each_entry(page, &cc->migratepages, lru)
526                 count[!!page_is_file_cache(page)]++;
527
528         mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
529         mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
530 }
531
532 /* Similar to reclaim, but different enough that they don't share logic */
533 static bool too_many_isolated(struct zone *zone)
534 {
535         unsigned long active, inactive, isolated;
536
537         inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
538                                         zone_page_state(zone, NR_INACTIVE_ANON);
539         active = zone_page_state(zone, NR_ACTIVE_FILE) +
540                                         zone_page_state(zone, NR_ACTIVE_ANON);
541         isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
542                                         zone_page_state(zone, NR_ISOLATED_ANON);
543
544         return isolated > (inactive + active) / 2;
545 }
546
547 /**
548  * isolate_migratepages_block() - isolate all migrate-able pages within
549  *                                a single pageblock
550  * @cc:         Compaction control structure.
551  * @low_pfn:    The first PFN to isolate
552  * @end_pfn:    The one-past-the-last PFN to isolate, within same pageblock
553  * @isolate_mode: Isolation mode to be used.
554  *
555  * Isolate all pages that can be migrated from the range specified by
556  * [low_pfn, end_pfn). The range is expected to be within same pageblock.
557  * Returns zero if there is a fatal signal pending, otherwise PFN of the
558  * first page that was not scanned (which may be both less, equal to or more
559  * than end_pfn).
560  *
561  * The pages are isolated on cc->migratepages list (not required to be empty),
562  * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
563  * is neither read nor updated.
564  */
565 static unsigned long
566 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
567                         unsigned long end_pfn, isolate_mode_t isolate_mode)
568 {
569         struct zone *zone = cc->zone;
570         unsigned long nr_scanned = 0, nr_isolated = 0;
571         struct list_head *migratelist = &cc->migratepages;
572         struct lruvec *lruvec;
573         unsigned long flags = 0;
574         bool locked = false;
575         struct page *page = NULL, *valid_page = NULL;
576
577         /*
578          * Ensure that there are not too many pages isolated from the LRU
579          * list by either parallel reclaimers or compaction. If there are,
580          * delay for some time until fewer pages are isolated
581          */
582         while (unlikely(too_many_isolated(zone))) {
583                 /* async migration should just abort */
584                 if (cc->mode == MIGRATE_ASYNC)
585                         return 0;
586
587                 congestion_wait(BLK_RW_ASYNC, HZ/10);
588
589                 if (fatal_signal_pending(current))
590                         return 0;
591         }
592
593         if (compact_should_abort(cc))
594                 return 0;
595
596         /* Time to isolate some pages for migration */
597         for (; low_pfn < end_pfn; low_pfn++) {
598                 /*
599                  * Periodically drop the lock (if held) regardless of its
600                  * contention, to give chance to IRQs. Abort async compaction
601                  * if contended.
602                  */
603                 if (!(low_pfn % SWAP_CLUSTER_MAX)
604                     && compact_unlock_should_abort(&zone->lru_lock, flags,
605                                                                 &locked, cc))
606                         break;
607
608                 if (!pfn_valid_within(low_pfn))
609                         continue;
610                 nr_scanned++;
611
612                 page = pfn_to_page(low_pfn);
613
614                 if (!valid_page)
615                         valid_page = page;
616
617                 /*
618                  * Skip if free. We read page order here without zone lock
619                  * which is generally unsafe, but the race window is small and
620                  * the worst thing that can happen is that we skip some
621                  * potential isolation targets.
622                  */
623                 if (PageBuddy(page)) {
624                         unsigned long freepage_order = page_order_unsafe(page);
625
626                         /*
627                          * Without lock, we cannot be sure that what we got is
628                          * a valid page order. Consider only values in the
629                          * valid order range to prevent low_pfn overflow.
630                          */
631                         if (freepage_order > 0 && freepage_order < MAX_ORDER)
632                                 low_pfn += (1UL << freepage_order) - 1;
633                         continue;
634                 }
635
636                 /*
637                  * Check may be lockless but that's ok as we recheck later.
638                  * It's possible to migrate LRU pages and balloon pages
639                  * Skip any other type of page
640                  */
641                 if (!PageLRU(page)) {
642                         if (unlikely(balloon_page_movable(page))) {
643                                 if (balloon_page_isolate(page)) {
644                                         /* Successfully isolated */
645                                         goto isolate_success;
646                                 }
647                         }
648                         continue;
649                 }
650
651                 /*
652                  * PageLRU is set. lru_lock normally excludes isolation
653                  * splitting and collapsing (collapsing has already happened
654                  * if PageLRU is set) but the lock is not necessarily taken
655                  * here and it is wasteful to take it just to check transhuge.
656                  * Check TransHuge without lock and skip the whole pageblock if
657                  * it's either a transhuge or hugetlbfs page, as calling
658                  * compound_order() without preventing THP from splitting the
659                  * page underneath us may return surprising results.
660                  */
661                 if (PageTransHuge(page)) {
662                         if (!locked)
663                                 low_pfn = ALIGN(low_pfn + 1,
664                                                 pageblock_nr_pages) - 1;
665                         else
666                                 low_pfn += (1 << compound_order(page)) - 1;
667
668                         continue;
669                 }
670
671                 /*
672                  * Migration will fail if an anonymous page is pinned in memory,
673                  * so avoid taking lru_lock and isolating it unnecessarily in an
674                  * admittedly racy check.
675                  */
676                 if (!page_mapping(page) &&
677                     page_count(page) > page_mapcount(page))
678                         continue;
679
680                 /* If we already hold the lock, we can skip some rechecking */
681                 if (!locked) {
682                         locked = compact_trylock_irqsave(&zone->lru_lock,
683                                                                 &flags, cc);
684                         if (!locked)
685                                 break;
686
687                         /* Recheck PageLRU and PageTransHuge under lock */
688                         if (!PageLRU(page))
689                                 continue;
690                         if (PageTransHuge(page)) {
691                                 low_pfn += (1 << compound_order(page)) - 1;
692                                 continue;
693                         }
694                 }
695
696                 lruvec = mem_cgroup_page_lruvec(page, zone);
697
698                 /* Try isolate the page */
699                 if (__isolate_lru_page(page, isolate_mode) != 0)
700                         continue;
701
702                 VM_BUG_ON_PAGE(PageTransCompound(page), page);
703
704                 /* Successfully isolated */
705                 del_page_from_lru_list(page, lruvec, page_lru(page));
706
707 isolate_success:
708                 cc->finished_update_migrate = true;
709                 list_add(&page->lru, migratelist);
710                 cc->nr_migratepages++;
711                 nr_isolated++;
712
713                 /* Avoid isolating too much */
714                 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
715                         ++low_pfn;
716                         break;
717                 }
718         }
719
720         /*
721          * The PageBuddy() check could have potentially brought us outside
722          * the range to be scanned.
723          */
724         if (unlikely(low_pfn > end_pfn))
725                 low_pfn = end_pfn;
726
727         if (locked)
728                 spin_unlock_irqrestore(&zone->lru_lock, flags);
729
730         /*
731          * Update the pageblock-skip information and cached scanner pfn,
732          * if the whole pageblock was scanned without isolating any page.
733          */
734         if (low_pfn == end_pfn)
735                 update_pageblock_skip(cc, valid_page, nr_isolated, true);
736
737         trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
738
739         count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
740         if (nr_isolated)
741                 count_compact_events(COMPACTISOLATED, nr_isolated);
742
743         return low_pfn;
744 }
745
746 /**
747  * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
748  * @cc:        Compaction control structure.
749  * @start_pfn: The first PFN to start isolating.
750  * @end_pfn:   The one-past-last PFN.
751  *
752  * Returns zero if isolation fails fatally due to e.g. pending signal.
753  * Otherwise, function returns one-past-the-last PFN of isolated page
754  * (which may be greater than end_pfn if end fell in a middle of a THP page).
755  */
756 unsigned long
757 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
758                                                         unsigned long end_pfn)
759 {
760         unsigned long pfn, block_end_pfn;
761
762         /* Scan block by block. First and last block may be incomplete */
763         pfn = start_pfn;
764         block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
765
766         for (; pfn < end_pfn; pfn = block_end_pfn,
767                                 block_end_pfn += pageblock_nr_pages) {
768
769                 block_end_pfn = min(block_end_pfn, end_pfn);
770
771                 if (!pageblock_pfn_to_page(pfn, block_end_pfn, cc->zone))
772                         continue;
773
774                 pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
775                                                         ISOLATE_UNEVICTABLE);
776
777                 /*
778                  * In case of fatal failure, release everything that might
779                  * have been isolated in the previous iteration, and signal
780                  * the failure back to caller.
781                  */
782                 if (!pfn) {
783                         putback_movable_pages(&cc->migratepages);
784                         cc->nr_migratepages = 0;
785                         break;
786                 }
787
788                 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
789                         break;
790         }
791         acct_isolated(cc->zone, cc);
792
793         return pfn;
794 }
795
796 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
797 #ifdef CONFIG_COMPACTION
798 /*
799  * Based on information in the current compact_control, find blocks
800  * suitable for isolating free pages from and then isolate them.
801  */
802 static void isolate_freepages(struct compact_control *cc)
803 {
804         struct zone *zone = cc->zone;
805         struct page *page;
806         unsigned long block_start_pfn;  /* start of current pageblock */
807         unsigned long isolate_start_pfn; /* exact pfn we start at */
808         unsigned long block_end_pfn;    /* end of current pageblock */
809         unsigned long low_pfn;       /* lowest pfn scanner is able to scan */
810         int nr_freepages = cc->nr_freepages;
811         struct list_head *freelist = &cc->freepages;
812
813         /*
814          * Initialise the free scanner. The starting point is where we last
815          * successfully isolated from, zone-cached value, or the end of the
816          * zone when isolating for the first time. For looping we also need
817          * this pfn aligned down to the pageblock boundary, because we do
818          * block_start_pfn -= pageblock_nr_pages in the for loop.
819          * For ending point, take care when isolating in last pageblock of a
820          * a zone which ends in the middle of a pageblock.
821          * The low boundary is the end of the pageblock the migration scanner
822          * is using.
823          */
824         isolate_start_pfn = cc->free_pfn;
825         block_start_pfn = cc->free_pfn & ~(pageblock_nr_pages-1);
826         block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
827                                                 zone_end_pfn(zone));
828         low_pfn = ALIGN(cc->migrate_pfn + 1, pageblock_nr_pages);
829
830         /*
831          * Isolate free pages until enough are available to migrate the
832          * pages on cc->migratepages. We stop searching if the migrate
833          * and free page scanners meet or enough free pages are isolated.
834          */
835         for (; block_start_pfn >= low_pfn && cc->nr_migratepages > nr_freepages;
836                                 block_end_pfn = block_start_pfn,
837                                 block_start_pfn -= pageblock_nr_pages,
838                                 isolate_start_pfn = block_start_pfn) {
839                 unsigned long isolated;
840
841                 /*
842                  * This can iterate a massively long zone without finding any
843                  * suitable migration targets, so periodically check if we need
844                  * to schedule, or even abort async compaction.
845                  */
846                 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
847                                                 && compact_should_abort(cc))
848                         break;
849
850                 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
851                                                                         zone);
852                 if (!page)
853                         continue;
854
855                 /* Check the block is suitable for migration */
856                 if (!suitable_migration_target(page))
857                         continue;
858
859                 /* If isolation recently failed, do not retry */
860                 if (!isolation_suitable(cc, page))
861                         continue;
862
863                 /* Found a block suitable for isolating free pages from. */
864                 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
865                                         block_end_pfn, freelist, false);
866                 nr_freepages += isolated;
867
868                 /*
869                  * Remember where the free scanner should restart next time,
870                  * which is where isolate_freepages_block() left off.
871                  * But if it scanned the whole pageblock, isolate_start_pfn
872                  * now points at block_end_pfn, which is the start of the next
873                  * pageblock.
874                  * In that case we will however want to restart at the start
875                  * of the previous pageblock.
876                  */
877                 cc->free_pfn = (isolate_start_pfn < block_end_pfn) ?
878                                 isolate_start_pfn :
879                                 block_start_pfn - pageblock_nr_pages;
880
881                 /*
882                  * Set a flag that we successfully isolated in this pageblock.
883                  * In the next loop iteration, zone->compact_cached_free_pfn
884                  * will not be updated and thus it will effectively contain the
885                  * highest pageblock we isolated pages from.
886                  */
887                 if (isolated)
888                         cc->finished_update_free = true;
889
890                 /*
891                  * isolate_freepages_block() might have aborted due to async
892                  * compaction being contended
893                  */
894                 if (cc->contended)
895                         break;
896         }
897
898         /* split_free_page does not map the pages */
899         map_pages(freelist);
900
901         /*
902          * If we crossed the migrate scanner, we want to keep it that way
903          * so that compact_finished() may detect this
904          */
905         if (block_start_pfn < low_pfn)
906                 cc->free_pfn = cc->migrate_pfn;
907
908         cc->nr_freepages = nr_freepages;
909 }
910
911 /*
912  * This is a migrate-callback that "allocates" freepages by taking pages
913  * from the isolated freelists in the block we are migrating to.
914  */
915 static struct page *compaction_alloc(struct page *migratepage,
916                                         unsigned long data,
917                                         int **result)
918 {
919         struct compact_control *cc = (struct compact_control *)data;
920         struct page *freepage;
921
922         /*
923          * Isolate free pages if necessary, and if we are not aborting due to
924          * contention.
925          */
926         if (list_empty(&cc->freepages)) {
927                 if (!cc->contended)
928                         isolate_freepages(cc);
929
930                 if (list_empty(&cc->freepages))
931                         return NULL;
932         }
933
934         freepage = list_entry(cc->freepages.next, struct page, lru);
935         list_del(&freepage->lru);
936         cc->nr_freepages--;
937
938         return freepage;
939 }
940
941 /*
942  * This is a migrate-callback that "frees" freepages back to the isolated
943  * freelist.  All pages on the freelist are from the same zone, so there is no
944  * special handling needed for NUMA.
945  */
946 static void compaction_free(struct page *page, unsigned long data)
947 {
948         struct compact_control *cc = (struct compact_control *)data;
949
950         list_add(&page->lru, &cc->freepages);
951         cc->nr_freepages++;
952 }
953
954 /* possible outcome of isolate_migratepages */
955 typedef enum {
956         ISOLATE_ABORT,          /* Abort compaction now */
957         ISOLATE_NONE,           /* No pages isolated, continue scanning */
958         ISOLATE_SUCCESS,        /* Pages isolated, migrate */
959 } isolate_migrate_t;
960
961 /*
962  * Isolate all pages that can be migrated from the first suitable block,
963  * starting at the block pointed to by the migrate scanner pfn within
964  * compact_control.
965  */
966 static isolate_migrate_t isolate_migratepages(struct zone *zone,
967                                         struct compact_control *cc)
968 {
969         unsigned long low_pfn, end_pfn;
970         struct page *page;
971         const isolate_mode_t isolate_mode =
972                 (cc->mode == MIGRATE_ASYNC ? ISOLATE_ASYNC_MIGRATE : 0);
973
974         /*
975          * Start at where we last stopped, or beginning of the zone as
976          * initialized by compact_zone()
977          */
978         low_pfn = cc->migrate_pfn;
979
980         /* Only scan within a pageblock boundary */
981         end_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages);
982
983         /*
984          * Iterate over whole pageblocks until we find the first suitable.
985          * Do not cross the free scanner.
986          */
987         for (; end_pfn <= cc->free_pfn;
988                         low_pfn = end_pfn, end_pfn += pageblock_nr_pages) {
989
990                 /*
991                  * This can potentially iterate a massively long zone with
992                  * many pageblocks unsuitable, so periodically check if we
993                  * need to schedule, or even abort async compaction.
994                  */
995                 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
996                                                 && compact_should_abort(cc))
997                         break;
998
999                 page = pageblock_pfn_to_page(low_pfn, end_pfn, zone);
1000                 if (!page)
1001                         continue;
1002
1003                 /* If isolation recently failed, do not retry */
1004                 if (!isolation_suitable(cc, page))
1005                         continue;
1006
1007                 /*
1008                  * For async compaction, also only scan in MOVABLE blocks.
1009                  * Async compaction is optimistic to see if the minimum amount
1010                  * of work satisfies the allocation.
1011                  */
1012                 if (cc->mode == MIGRATE_ASYNC &&
1013                     !migrate_async_suitable(get_pageblock_migratetype(page)))
1014                         continue;
1015
1016                 /* Perform the isolation */
1017                 low_pfn = isolate_migratepages_block(cc, low_pfn, end_pfn,
1018                                                                 isolate_mode);
1019
1020                 if (!low_pfn || cc->contended)
1021                         return ISOLATE_ABORT;
1022
1023                 /*
1024                  * Either we isolated something and proceed with migration. Or
1025                  * we failed and compact_zone should decide if we should
1026                  * continue or not.
1027                  */
1028                 break;
1029         }
1030
1031         acct_isolated(zone, cc);
1032         /* Record where migration scanner will be restarted */
1033         cc->migrate_pfn = low_pfn;
1034
1035         return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
1036 }
1037
1038 static int compact_finished(struct zone *zone, struct compact_control *cc,
1039                             const int migratetype)
1040 {
1041         unsigned int order;
1042         unsigned long watermark;
1043
1044         if (cc->contended || fatal_signal_pending(current))
1045                 return COMPACT_PARTIAL;
1046
1047         /* Compaction run completes if the migrate and free scanner meet */
1048         if (cc->free_pfn <= cc->migrate_pfn) {
1049                 /* Let the next compaction start anew. */
1050                 zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
1051                 zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
1052                 zone->compact_cached_free_pfn = zone_end_pfn(zone);
1053
1054                 /*
1055                  * Mark that the PG_migrate_skip information should be cleared
1056                  * by kswapd when it goes to sleep. kswapd does not set the
1057                  * flag itself as the decision to be clear should be directly
1058                  * based on an allocation request.
1059                  */
1060                 if (!current_is_kswapd())
1061                         zone->compact_blockskip_flush = true;
1062
1063                 return COMPACT_COMPLETE;
1064         }
1065
1066         /*
1067          * order == -1 is expected when compacting via
1068          * /proc/sys/vm/compact_memory
1069          */
1070         if (cc->order == -1)
1071                 return COMPACT_CONTINUE;
1072
1073         /* Compaction run is not finished if the watermark is not met */
1074         watermark = low_wmark_pages(zone);
1075         watermark += (1 << cc->order);
1076
1077         if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
1078                 return COMPACT_CONTINUE;
1079
1080         /* Direct compactor: Is a suitable page free? */
1081         for (order = cc->order; order < MAX_ORDER; order++) {
1082                 struct free_area *area = &zone->free_area[order];
1083
1084                 /* Job done if page is free of the right migratetype */
1085                 if (!list_empty(&area->free_list[migratetype]))
1086                         return COMPACT_PARTIAL;
1087
1088                 /* Job done if allocation would set block type */
1089                 if (cc->order >= pageblock_order && area->nr_free)
1090                         return COMPACT_PARTIAL;
1091         }
1092
1093         return COMPACT_CONTINUE;
1094 }
1095
1096 /*
1097  * compaction_suitable: Is this suitable to run compaction on this zone now?
1098  * Returns
1099  *   COMPACT_SKIPPED  - If there are too few free pages for compaction
1100  *   COMPACT_PARTIAL  - If the allocation would succeed without compaction
1101  *   COMPACT_CONTINUE - If compaction should run now
1102  */
1103 unsigned long compaction_suitable(struct zone *zone, int order)
1104 {
1105         int fragindex;
1106         unsigned long watermark;
1107
1108         /*
1109          * order == -1 is expected when compacting via
1110          * /proc/sys/vm/compact_memory
1111          */
1112         if (order == -1)
1113                 return COMPACT_CONTINUE;
1114
1115         /*
1116          * Watermarks for order-0 must be met for compaction. Note the 2UL.
1117          * This is because during migration, copies of pages need to be
1118          * allocated and for a short time, the footprint is higher
1119          */
1120         watermark = low_wmark_pages(zone) + (2UL << order);
1121         if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1122                 return COMPACT_SKIPPED;
1123
1124         /*
1125          * fragmentation index determines if allocation failures are due to
1126          * low memory or external fragmentation
1127          *
1128          * index of -1000 implies allocations might succeed depending on
1129          * watermarks
1130          * index towards 0 implies failure is due to lack of memory
1131          * index towards 1000 implies failure is due to fragmentation
1132          *
1133          * Only compact if a failure would be due to fragmentation.
1134          */
1135         fragindex = fragmentation_index(zone, order);
1136         if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
1137                 return COMPACT_SKIPPED;
1138
1139         if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
1140             0, 0))
1141                 return COMPACT_PARTIAL;
1142
1143         return COMPACT_CONTINUE;
1144 }
1145
1146 static int compact_zone(struct zone *zone, struct compact_control *cc)
1147 {
1148         int ret;
1149         unsigned long start_pfn = zone->zone_start_pfn;
1150         unsigned long end_pfn = zone_end_pfn(zone);
1151         const int migratetype = gfpflags_to_migratetype(cc->gfp_mask);
1152         const bool sync = cc->mode != MIGRATE_ASYNC;
1153
1154         ret = compaction_suitable(zone, cc->order);
1155         switch (ret) {
1156         case COMPACT_PARTIAL:
1157         case COMPACT_SKIPPED:
1158                 /* Compaction is likely to fail */
1159                 return ret;
1160         case COMPACT_CONTINUE:
1161                 /* Fall through to compaction */
1162                 ;
1163         }
1164
1165         /*
1166          * Clear pageblock skip if there were failures recently and compaction
1167          * is about to be retried after being deferred. kswapd does not do
1168          * this reset as it'll reset the cached information when going to sleep.
1169          */
1170         if (compaction_restarting(zone, cc->order) && !current_is_kswapd())
1171                 __reset_isolation_suitable(zone);
1172
1173         /*
1174          * Setup to move all movable pages to the end of the zone. Used cached
1175          * information on where the scanners should start but check that it
1176          * is initialised by ensuring the values are within zone boundaries.
1177          */
1178         cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync];
1179         cc->free_pfn = zone->compact_cached_free_pfn;
1180         if (cc->free_pfn < start_pfn || cc->free_pfn > end_pfn) {
1181                 cc->free_pfn = end_pfn & ~(pageblock_nr_pages-1);
1182                 zone->compact_cached_free_pfn = cc->free_pfn;
1183         }
1184         if (cc->migrate_pfn < start_pfn || cc->migrate_pfn > end_pfn) {
1185                 cc->migrate_pfn = start_pfn;
1186                 zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
1187                 zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
1188         }
1189
1190         trace_mm_compaction_begin(start_pfn, cc->migrate_pfn, cc->free_pfn, end_pfn);
1191
1192         migrate_prep_local();
1193
1194         while ((ret = compact_finished(zone, cc, migratetype)) ==
1195                                                 COMPACT_CONTINUE) {
1196                 int err;
1197
1198                 switch (isolate_migratepages(zone, cc)) {
1199                 case ISOLATE_ABORT:
1200                         ret = COMPACT_PARTIAL;
1201                         putback_movable_pages(&cc->migratepages);
1202                         cc->nr_migratepages = 0;
1203                         goto out;
1204                 case ISOLATE_NONE:
1205                         continue;
1206                 case ISOLATE_SUCCESS:
1207                         ;
1208                 }
1209
1210                 err = migrate_pages(&cc->migratepages, compaction_alloc,
1211                                 compaction_free, (unsigned long)cc, cc->mode,
1212                                 MR_COMPACTION);
1213
1214                 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
1215                                                         &cc->migratepages);
1216
1217                 /* All pages were either migrated or will be released */
1218                 cc->nr_migratepages = 0;
1219                 if (err) {
1220                         putback_movable_pages(&cc->migratepages);
1221                         /*
1222                          * migrate_pages() may return -ENOMEM when scanners meet
1223                          * and we want compact_finished() to detect it
1224                          */
1225                         if (err == -ENOMEM && cc->free_pfn > cc->migrate_pfn) {
1226                                 ret = COMPACT_PARTIAL;
1227                                 goto out;
1228                         }
1229                 }
1230         }
1231
1232 out:
1233         /* Release free pages and check accounting */
1234         cc->nr_freepages -= release_freepages(&cc->freepages);
1235         VM_BUG_ON(cc->nr_freepages != 0);
1236
1237         trace_mm_compaction_end(ret);
1238
1239         return ret;
1240 }
1241
1242 static unsigned long compact_zone_order(struct zone *zone, int order,
1243                 gfp_t gfp_mask, enum migrate_mode mode, int *contended)
1244 {
1245         unsigned long ret;
1246         struct compact_control cc = {
1247                 .nr_freepages = 0,
1248                 .nr_migratepages = 0,
1249                 .order = order,
1250                 .gfp_mask = gfp_mask,
1251                 .zone = zone,
1252                 .mode = mode,
1253         };
1254         INIT_LIST_HEAD(&cc.freepages);
1255         INIT_LIST_HEAD(&cc.migratepages);
1256
1257         ret = compact_zone(zone, &cc);
1258
1259         VM_BUG_ON(!list_empty(&cc.freepages));
1260         VM_BUG_ON(!list_empty(&cc.migratepages));
1261
1262         *contended = cc.contended;
1263         return ret;
1264 }
1265
1266 int sysctl_extfrag_threshold = 500;
1267
1268 /**
1269  * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1270  * @zonelist: The zonelist used for the current allocation
1271  * @order: The order of the current allocation
1272  * @gfp_mask: The GFP mask of the current allocation
1273  * @nodemask: The allowed nodes to allocate from
1274  * @mode: The migration mode for async, sync light, or sync migration
1275  * @contended: Return value that determines if compaction was aborted due to
1276  *             need_resched() or lock contention
1277  * @candidate_zone: Return the zone where we think allocation should succeed
1278  *
1279  * This is the main entry point for direct page compaction.
1280  */
1281 unsigned long try_to_compact_pages(struct zonelist *zonelist,
1282                         int order, gfp_t gfp_mask, nodemask_t *nodemask,
1283                         enum migrate_mode mode, int *contended,
1284                         struct zone **candidate_zone)
1285 {
1286         enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1287         int may_enter_fs = gfp_mask & __GFP_FS;
1288         int may_perform_io = gfp_mask & __GFP_IO;
1289         struct zoneref *z;
1290         struct zone *zone;
1291         int rc = COMPACT_DEFERRED;
1292         int alloc_flags = 0;
1293         int all_zones_contended = COMPACT_CONTENDED_LOCK; /* init for &= op */
1294
1295         *contended = COMPACT_CONTENDED_NONE;
1296
1297         /* Check if the GFP flags allow compaction */
1298         if (!order || !may_enter_fs || !may_perform_io)
1299                 return COMPACT_SKIPPED;
1300
1301 #ifdef CONFIG_CMA
1302         if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
1303                 alloc_flags |= ALLOC_CMA;
1304 #endif
1305         /* Compact each zone in the list */
1306         for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1307                                                                 nodemask) {
1308                 int status;
1309                 int zone_contended;
1310
1311                 if (compaction_deferred(zone, order))
1312                         continue;
1313
1314                 status = compact_zone_order(zone, order, gfp_mask, mode,
1315                                                         &zone_contended);
1316                 rc = max(status, rc);
1317                 /*
1318                  * It takes at least one zone that wasn't lock contended
1319                  * to clear all_zones_contended.
1320                  */
1321                 all_zones_contended &= zone_contended;
1322
1323                 /* If a normal allocation would succeed, stop compacting */
1324                 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0,
1325                                       alloc_flags)) {
1326                         *candidate_zone = zone;
1327                         /*
1328                          * We think the allocation will succeed in this zone,
1329                          * but it is not certain, hence the false. The caller
1330                          * will repeat this with true if allocation indeed
1331                          * succeeds in this zone.
1332                          */
1333                         compaction_defer_reset(zone, order, false);
1334                         /*
1335                          * It is possible that async compaction aborted due to
1336                          * need_resched() and the watermarks were ok thanks to
1337                          * somebody else freeing memory. The allocation can
1338                          * however still fail so we better signal the
1339                          * need_resched() contention anyway (this will not
1340                          * prevent the allocation attempt).
1341                          */
1342                         if (zone_contended == COMPACT_CONTENDED_SCHED)
1343                                 *contended = COMPACT_CONTENDED_SCHED;
1344
1345                         goto break_loop;
1346                 }
1347
1348                 if (mode != MIGRATE_ASYNC) {
1349                         /*
1350                          * We think that allocation won't succeed in this zone
1351                          * so we defer compaction there. If it ends up
1352                          * succeeding after all, it will be reset.
1353                          */
1354                         defer_compaction(zone, order);
1355                 }
1356
1357                 /*
1358                  * We might have stopped compacting due to need_resched() in
1359                  * async compaction, or due to a fatal signal detected. In that
1360                  * case do not try further zones and signal need_resched()
1361                  * contention.
1362                  */
1363                 if ((zone_contended == COMPACT_CONTENDED_SCHED)
1364                                         || fatal_signal_pending(current)) {
1365                         *contended = COMPACT_CONTENDED_SCHED;
1366                         goto break_loop;
1367                 }
1368
1369                 continue;
1370 break_loop:
1371                 /*
1372                  * We might not have tried all the zones, so  be conservative
1373                  * and assume they are not all lock contended.
1374                  */
1375                 all_zones_contended = 0;
1376                 break;
1377         }
1378
1379         /*
1380          * If at least one zone wasn't deferred or skipped, we report if all
1381          * zones that were tried were lock contended.
1382          */
1383         if (rc > COMPACT_SKIPPED && all_zones_contended)
1384                 *contended = COMPACT_CONTENDED_LOCK;
1385
1386         return rc;
1387 }
1388
1389
1390 /* Compact all zones within a node */
1391 static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
1392 {
1393         int zoneid;
1394         struct zone *zone;
1395
1396         for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1397
1398                 zone = &pgdat->node_zones[zoneid];
1399                 if (!populated_zone(zone))
1400                         continue;
1401
1402                 cc->nr_freepages = 0;
1403                 cc->nr_migratepages = 0;
1404                 cc->zone = zone;
1405                 INIT_LIST_HEAD(&cc->freepages);
1406                 INIT_LIST_HEAD(&cc->migratepages);
1407
1408                 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
1409                         compact_zone(zone, cc);
1410
1411                 if (cc->order > 0) {
1412                         if (zone_watermark_ok(zone, cc->order,
1413                                                 low_wmark_pages(zone), 0, 0))
1414                                 compaction_defer_reset(zone, cc->order, false);
1415                 }
1416
1417                 VM_BUG_ON(!list_empty(&cc->freepages));
1418                 VM_BUG_ON(!list_empty(&cc->migratepages));
1419         }
1420 }
1421
1422 void compact_pgdat(pg_data_t *pgdat, int order)
1423 {
1424         struct compact_control cc = {
1425                 .order = order,
1426                 .mode = MIGRATE_ASYNC,
1427         };
1428
1429         if (!order)
1430                 return;
1431
1432         __compact_pgdat(pgdat, &cc);
1433 }
1434
1435 static void compact_node(int nid)
1436 {
1437         struct compact_control cc = {
1438                 .order = -1,
1439                 .mode = MIGRATE_SYNC,
1440                 .ignore_skip_hint = true,
1441         };
1442
1443         __compact_pgdat(NODE_DATA(nid), &cc);
1444 }
1445
1446 /* Compact all nodes in the system */
1447 static void compact_nodes(void)
1448 {
1449         int nid;
1450
1451         /* Flush pending updates to the LRU lists */
1452         lru_add_drain_all();
1453
1454         for_each_online_node(nid)
1455                 compact_node(nid);
1456 }
1457
1458 /* The written value is actually unused, all memory is compacted */
1459 int sysctl_compact_memory;
1460
1461 /* This is the entry point for compacting all nodes via /proc/sys/vm */
1462 int sysctl_compaction_handler(struct ctl_table *table, int write,
1463                         void __user *buffer, size_t *length, loff_t *ppos)
1464 {
1465         if (write)
1466                 compact_nodes();
1467
1468         return 0;
1469 }
1470
1471 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1472                         void __user *buffer, size_t *length, loff_t *ppos)
1473 {
1474         proc_dointvec_minmax(table, write, buffer, length, ppos);
1475
1476         return 0;
1477 }
1478
1479 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1480 static ssize_t sysfs_compact_node(struct device *dev,
1481                         struct device_attribute *attr,
1482                         const char *buf, size_t count)
1483 {
1484         int nid = dev->id;
1485
1486         if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1487                 /* Flush pending updates to the LRU lists */
1488                 lru_add_drain_all();
1489
1490                 compact_node(nid);
1491         }
1492
1493         return count;
1494 }
1495 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1496
1497 int compaction_register_node(struct node *node)
1498 {
1499         return device_create_file(&node->dev, &dev_attr_compact);
1500 }
1501
1502 void compaction_unregister_node(struct node *node)
1503 {
1504         return device_remove_file(&node->dev, &dev_attr_compact);
1505 }
1506 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1507
1508 #endif /* CONFIG_COMPACTION */