]> git.karo-electronics.de Git - karo-tx-linux.git/blob - mm/compaction.c
Merge tag 'armsoc-drivers' of git://git.kernel.org/pub/scm/linux/kernel/git/arm/arm-soc
[karo-tx-linux.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 <linux/kasan.h>
20 #include "internal.h"
21
22 #ifdef CONFIG_COMPACTION
23 static inline void count_compact_event(enum vm_event_item item)
24 {
25         count_vm_event(item);
26 }
27
28 static inline void count_compact_events(enum vm_event_item item, long delta)
29 {
30         count_vm_events(item, delta);
31 }
32 #else
33 #define count_compact_event(item) do { } while (0)
34 #define count_compact_events(item, delta) do { } while (0)
35 #endif
36
37 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
38
39 #define CREATE_TRACE_POINTS
40 #include <trace/events/compaction.h>
41
42 static unsigned long release_freepages(struct list_head *freelist)
43 {
44         struct page *page, *next;
45         unsigned long high_pfn = 0;
46
47         list_for_each_entry_safe(page, next, freelist, lru) {
48                 unsigned long pfn = page_to_pfn(page);
49                 list_del(&page->lru);
50                 __free_page(page);
51                 if (pfn > high_pfn)
52                         high_pfn = pfn;
53         }
54
55         return high_pfn;
56 }
57
58 static void map_pages(struct list_head *list)
59 {
60         struct page *page;
61
62         list_for_each_entry(page, list, lru) {
63                 arch_alloc_page(page, 0);
64                 kernel_map_pages(page, 1, 1);
65                 kasan_alloc_pages(page, 0);
66         }
67 }
68
69 static inline bool migrate_async_suitable(int migratetype)
70 {
71         return is_migrate_cma(migratetype) || migratetype == MIGRATE_MOVABLE;
72 }
73
74 /*
75  * Check that the whole (or subset of) a pageblock given by the interval of
76  * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
77  * with the migration of free compaction scanner. The scanners then need to
78  * use only pfn_valid_within() check for arches that allow holes within
79  * pageblocks.
80  *
81  * Return struct page pointer of start_pfn, or NULL if checks were not passed.
82  *
83  * It's possible on some configurations to have a setup like node0 node1 node0
84  * i.e. it's possible that all pages within a zones range of pages do not
85  * belong to a single zone. We assume that a border between node0 and node1
86  * can occur within a single pageblock, but not a node0 node1 node0
87  * interleaving within a single pageblock. It is therefore sufficient to check
88  * the first and last page of a pageblock and avoid checking each individual
89  * page in a pageblock.
90  */
91 static struct page *pageblock_pfn_to_page(unsigned long start_pfn,
92                                 unsigned long end_pfn, struct zone *zone)
93 {
94         struct page *start_page;
95         struct page *end_page;
96
97         /* end_pfn is one past the range we are checking */
98         end_pfn--;
99
100         if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
101                 return NULL;
102
103         start_page = pfn_to_page(start_pfn);
104
105         if (page_zone(start_page) != zone)
106                 return NULL;
107
108         end_page = pfn_to_page(end_pfn);
109
110         /* This gives a shorter code than deriving page_zone(end_page) */
111         if (page_zone_id(start_page) != page_zone_id(end_page))
112                 return NULL;
113
114         return start_page;
115 }
116
117 #ifdef CONFIG_COMPACTION
118
119 /* Do not skip compaction more than 64 times */
120 #define COMPACT_MAX_DEFER_SHIFT 6
121
122 /*
123  * Compaction is deferred when compaction fails to result in a page
124  * allocation success. 1 << compact_defer_limit compactions are skipped up
125  * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
126  */
127 void defer_compaction(struct zone *zone, int order)
128 {
129         zone->compact_considered = 0;
130         zone->compact_defer_shift++;
131
132         if (order < zone->compact_order_failed)
133                 zone->compact_order_failed = order;
134
135         if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
136                 zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
137
138         trace_mm_compaction_defer_compaction(zone, order);
139 }
140
141 /* Returns true if compaction should be skipped this time */
142 bool compaction_deferred(struct zone *zone, int order)
143 {
144         unsigned long defer_limit = 1UL << zone->compact_defer_shift;
145
146         if (order < zone->compact_order_failed)
147                 return false;
148
149         /* Avoid possible overflow */
150         if (++zone->compact_considered > defer_limit)
151                 zone->compact_considered = defer_limit;
152
153         if (zone->compact_considered >= defer_limit)
154                 return false;
155
156         trace_mm_compaction_deferred(zone, order);
157
158         return true;
159 }
160
161 /*
162  * Update defer tracking counters after successful compaction of given order,
163  * which means an allocation either succeeded (alloc_success == true) or is
164  * expected to succeed.
165  */
166 void compaction_defer_reset(struct zone *zone, int order,
167                 bool alloc_success)
168 {
169         if (alloc_success) {
170                 zone->compact_considered = 0;
171                 zone->compact_defer_shift = 0;
172         }
173         if (order >= zone->compact_order_failed)
174                 zone->compact_order_failed = order + 1;
175
176         trace_mm_compaction_defer_reset(zone, order);
177 }
178
179 /* Returns true if restarting compaction after many failures */
180 bool compaction_restarting(struct zone *zone, int order)
181 {
182         if (order < zone->compact_order_failed)
183                 return false;
184
185         return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
186                 zone->compact_considered >= 1UL << zone->compact_defer_shift;
187 }
188
189 /* Returns true if the pageblock should be scanned for pages to isolate. */
190 static inline bool isolation_suitable(struct compact_control *cc,
191                                         struct page *page)
192 {
193         if (cc->ignore_skip_hint)
194                 return true;
195
196         return !get_pageblock_skip(page);
197 }
198
199 static void reset_cached_positions(struct zone *zone)
200 {
201         zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
202         zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
203         zone->compact_cached_free_pfn = zone_end_pfn(zone);
204 }
205
206 /*
207  * This function is called to clear all cached information on pageblocks that
208  * should be skipped for page isolation when the migrate and free page scanner
209  * meet.
210  */
211 static void __reset_isolation_suitable(struct zone *zone)
212 {
213         unsigned long start_pfn = zone->zone_start_pfn;
214         unsigned long end_pfn = zone_end_pfn(zone);
215         unsigned long pfn;
216
217         zone->compact_blockskip_flush = false;
218
219         /* Walk the zone and mark every pageblock as suitable for isolation */
220         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
221                 struct page *page;
222
223                 cond_resched();
224
225                 if (!pfn_valid(pfn))
226                         continue;
227
228                 page = pfn_to_page(pfn);
229                 if (zone != page_zone(page))
230                         continue;
231
232                 clear_pageblock_skip(page);
233         }
234
235         reset_cached_positions(zone);
236 }
237
238 void reset_isolation_suitable(pg_data_t *pgdat)
239 {
240         int zoneid;
241
242         for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
243                 struct zone *zone = &pgdat->node_zones[zoneid];
244                 if (!populated_zone(zone))
245                         continue;
246
247                 /* Only flush if a full compaction finished recently */
248                 if (zone->compact_blockskip_flush)
249                         __reset_isolation_suitable(zone);
250         }
251 }
252
253 /*
254  * If no pages were isolated then mark this pageblock to be skipped in the
255  * future. The information is later cleared by __reset_isolation_suitable().
256  */
257 static void update_pageblock_skip(struct compact_control *cc,
258                         struct page *page, unsigned long nr_isolated,
259                         bool migrate_scanner)
260 {
261         struct zone *zone = cc->zone;
262         unsigned long pfn;
263
264         if (cc->ignore_skip_hint)
265                 return;
266
267         if (!page)
268                 return;
269
270         if (nr_isolated)
271                 return;
272
273         set_pageblock_skip(page);
274
275         pfn = page_to_pfn(page);
276
277         /* Update where async and sync compaction should restart */
278         if (migrate_scanner) {
279                 if (pfn > zone->compact_cached_migrate_pfn[0])
280                         zone->compact_cached_migrate_pfn[0] = pfn;
281                 if (cc->mode != MIGRATE_ASYNC &&
282                     pfn > zone->compact_cached_migrate_pfn[1])
283                         zone->compact_cached_migrate_pfn[1] = pfn;
284         } else {
285                 if (pfn < zone->compact_cached_free_pfn)
286                         zone->compact_cached_free_pfn = pfn;
287         }
288 }
289 #else
290 static inline bool isolation_suitable(struct compact_control *cc,
291                                         struct page *page)
292 {
293         return true;
294 }
295
296 static void update_pageblock_skip(struct compact_control *cc,
297                         struct page *page, unsigned long nr_isolated,
298                         bool migrate_scanner)
299 {
300 }
301 #endif /* CONFIG_COMPACTION */
302
303 /*
304  * Compaction requires the taking of some coarse locks that are potentially
305  * very heavily contended. For async compaction, back out if the lock cannot
306  * be taken immediately. For sync compaction, spin on the lock if needed.
307  *
308  * Returns true if the lock is held
309  * Returns false if the lock is not held and compaction should abort
310  */
311 static bool compact_trylock_irqsave(spinlock_t *lock, unsigned long *flags,
312                                                 struct compact_control *cc)
313 {
314         if (cc->mode == MIGRATE_ASYNC) {
315                 if (!spin_trylock_irqsave(lock, *flags)) {
316                         cc->contended = COMPACT_CONTENDED_LOCK;
317                         return false;
318                 }
319         } else {
320                 spin_lock_irqsave(lock, *flags);
321         }
322
323         return true;
324 }
325
326 /*
327  * Compaction requires the taking of some coarse locks that are potentially
328  * very heavily contended. The lock should be periodically unlocked to avoid
329  * having disabled IRQs for a long time, even when there is nobody waiting on
330  * the lock. It might also be that allowing the IRQs will result in
331  * need_resched() becoming true. If scheduling is needed, async compaction
332  * aborts. Sync compaction schedules.
333  * Either compaction type will also abort if a fatal signal is pending.
334  * In either case if the lock was locked, it is dropped and not regained.
335  *
336  * Returns true if compaction should abort due to fatal signal pending, or
337  *              async compaction due to need_resched()
338  * Returns false when compaction can continue (sync compaction might have
339  *              scheduled)
340  */
341 static bool compact_unlock_should_abort(spinlock_t *lock,
342                 unsigned long flags, bool *locked, struct compact_control *cc)
343 {
344         if (*locked) {
345                 spin_unlock_irqrestore(lock, flags);
346                 *locked = false;
347         }
348
349         if (fatal_signal_pending(current)) {
350                 cc->contended = COMPACT_CONTENDED_SCHED;
351                 return true;
352         }
353
354         if (need_resched()) {
355                 if (cc->mode == MIGRATE_ASYNC) {
356                         cc->contended = COMPACT_CONTENDED_SCHED;
357                         return true;
358                 }
359                 cond_resched();
360         }
361
362         return false;
363 }
364
365 /*
366  * Aside from avoiding lock contention, compaction also periodically checks
367  * need_resched() and either schedules in sync compaction or aborts async
368  * compaction. This is similar to what compact_unlock_should_abort() does, but
369  * is used where no lock is concerned.
370  *
371  * Returns false when no scheduling was needed, or sync compaction scheduled.
372  * Returns true when async compaction should abort.
373  */
374 static inline bool compact_should_abort(struct compact_control *cc)
375 {
376         /* async compaction aborts if contended */
377         if (need_resched()) {
378                 if (cc->mode == MIGRATE_ASYNC) {
379                         cc->contended = COMPACT_CONTENDED_SCHED;
380                         return true;
381                 }
382
383                 cond_resched();
384         }
385
386         return false;
387 }
388
389 /*
390  * Isolate free pages onto a private freelist. If @strict is true, will abort
391  * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
392  * (even though it may still end up isolating some pages).
393  */
394 static unsigned long isolate_freepages_block(struct compact_control *cc,
395                                 unsigned long *start_pfn,
396                                 unsigned long end_pfn,
397                                 struct list_head *freelist,
398                                 bool strict)
399 {
400         int nr_scanned = 0, total_isolated = 0;
401         struct page *cursor, *valid_page = NULL;
402         unsigned long flags = 0;
403         bool locked = false;
404         unsigned long blockpfn = *start_pfn;
405
406         cursor = pfn_to_page(blockpfn);
407
408         /* Isolate free pages. */
409         for (; blockpfn < end_pfn; blockpfn++, cursor++) {
410                 int isolated, i;
411                 struct page *page = cursor;
412
413                 /*
414                  * Periodically drop the lock (if held) regardless of its
415                  * contention, to give chance to IRQs. Abort if fatal signal
416                  * pending or async compaction detects need_resched()
417                  */
418                 if (!(blockpfn % SWAP_CLUSTER_MAX)
419                     && compact_unlock_should_abort(&cc->zone->lock, flags,
420                                                                 &locked, cc))
421                         break;
422
423                 nr_scanned++;
424                 if (!pfn_valid_within(blockpfn))
425                         goto isolate_fail;
426
427                 if (!valid_page)
428                         valid_page = page;
429
430                 /*
431                  * For compound pages such as THP and hugetlbfs, we can save
432                  * potentially a lot of iterations if we skip them at once.
433                  * The check is racy, but we can consider only valid values
434                  * and the only danger is skipping too much.
435                  */
436                 if (PageCompound(page)) {
437                         unsigned int comp_order = compound_order(page);
438
439                         if (likely(comp_order < MAX_ORDER)) {
440                                 blockpfn += (1UL << comp_order) - 1;
441                                 cursor += (1UL << comp_order) - 1;
442                         }
443
444                         goto isolate_fail;
445                 }
446
447                 if (!PageBuddy(page))
448                         goto isolate_fail;
449
450                 /*
451                  * If we already hold the lock, we can skip some rechecking.
452                  * Note that if we hold the lock now, checked_pageblock was
453                  * already set in some previous iteration (or strict is true),
454                  * so it is correct to skip the suitable migration target
455                  * recheck as well.
456                  */
457                 if (!locked) {
458                         /*
459                          * The zone lock must be held to isolate freepages.
460                          * Unfortunately this is a very coarse lock and can be
461                          * heavily contended if there are parallel allocations
462                          * or parallel compactions. For async compaction do not
463                          * spin on the lock and we acquire the lock as late as
464                          * possible.
465                          */
466                         locked = compact_trylock_irqsave(&cc->zone->lock,
467                                                                 &flags, cc);
468                         if (!locked)
469                                 break;
470
471                         /* Recheck this is a buddy page under lock */
472                         if (!PageBuddy(page))
473                                 goto isolate_fail;
474                 }
475
476                 /* Found a free page, break it into order-0 pages */
477                 isolated = split_free_page(page);
478                 total_isolated += isolated;
479                 for (i = 0; i < isolated; i++) {
480                         list_add(&page->lru, freelist);
481                         page++;
482                 }
483
484                 /* If a page was split, advance to the end of it */
485                 if (isolated) {
486                         cc->nr_freepages += isolated;
487                         if (!strict &&
488                                 cc->nr_migratepages <= cc->nr_freepages) {
489                                 blockpfn += isolated;
490                                 break;
491                         }
492
493                         blockpfn += isolated - 1;
494                         cursor += isolated - 1;
495                         continue;
496                 }
497
498 isolate_fail:
499                 if (strict)
500                         break;
501                 else
502                         continue;
503
504         }
505
506         /*
507          * There is a tiny chance that we have read bogus compound_order(),
508          * so be careful to not go outside of the pageblock.
509          */
510         if (unlikely(blockpfn > end_pfn))
511                 blockpfn = end_pfn;
512
513         trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
514                                         nr_scanned, total_isolated);
515
516         /* Record how far we have got within the block */
517         *start_pfn = blockpfn;
518
519         /*
520          * If strict isolation is requested by CMA then check that all the
521          * pages requested were isolated. If there were any failures, 0 is
522          * returned and CMA will fail.
523          */
524         if (strict && blockpfn < end_pfn)
525                 total_isolated = 0;
526
527         if (locked)
528                 spin_unlock_irqrestore(&cc->zone->lock, flags);
529
530         /* Update the pageblock-skip if the whole pageblock was scanned */
531         if (blockpfn == end_pfn)
532                 update_pageblock_skip(cc, valid_page, total_isolated, false);
533
534         count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
535         if (total_isolated)
536                 count_compact_events(COMPACTISOLATED, total_isolated);
537         return total_isolated;
538 }
539
540 /**
541  * isolate_freepages_range() - isolate free pages.
542  * @start_pfn: The first PFN to start isolating.
543  * @end_pfn:   The one-past-last PFN.
544  *
545  * Non-free pages, invalid PFNs, or zone boundaries within the
546  * [start_pfn, end_pfn) range are considered errors, cause function to
547  * undo its actions and return zero.
548  *
549  * Otherwise, function returns one-past-the-last PFN of isolated page
550  * (which may be greater then end_pfn if end fell in a middle of
551  * a free page).
552  */
553 unsigned long
554 isolate_freepages_range(struct compact_control *cc,
555                         unsigned long start_pfn, unsigned long end_pfn)
556 {
557         unsigned long isolated, pfn, block_end_pfn;
558         LIST_HEAD(freelist);
559
560         pfn = start_pfn;
561         block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
562
563         for (; pfn < end_pfn; pfn += isolated,
564                                 block_end_pfn += pageblock_nr_pages) {
565                 /* Protect pfn from changing by isolate_freepages_block */
566                 unsigned long isolate_start_pfn = pfn;
567
568                 block_end_pfn = min(block_end_pfn, end_pfn);
569
570                 /*
571                  * pfn could pass the block_end_pfn if isolated freepage
572                  * is more than pageblock order. In this case, we adjust
573                  * scanning range to right one.
574                  */
575                 if (pfn >= block_end_pfn) {
576                         block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
577                         block_end_pfn = min(block_end_pfn, end_pfn);
578                 }
579
580                 if (!pageblock_pfn_to_page(pfn, block_end_pfn, cc->zone))
581                         break;
582
583                 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
584                                                 block_end_pfn, &freelist, true);
585
586                 /*
587                  * In strict mode, isolate_freepages_block() returns 0 if
588                  * there are any holes in the block (ie. invalid PFNs or
589                  * non-free pages).
590                  */
591                 if (!isolated)
592                         break;
593
594                 /*
595                  * If we managed to isolate pages, it is always (1 << n) *
596                  * pageblock_nr_pages for some non-negative n.  (Max order
597                  * page may span two pageblocks).
598                  */
599         }
600
601         /* split_free_page does not map the pages */
602         map_pages(&freelist);
603
604         if (pfn < end_pfn) {
605                 /* Loop terminated early, cleanup. */
606                 release_freepages(&freelist);
607                 return 0;
608         }
609
610         /* We don't use freelists for anything. */
611         return pfn;
612 }
613
614 /* Update the number of anon and file isolated pages in the zone */
615 static void acct_isolated(struct zone *zone, struct compact_control *cc)
616 {
617         struct page *page;
618         unsigned int count[2] = { 0, };
619
620         if (list_empty(&cc->migratepages))
621                 return;
622
623         list_for_each_entry(page, &cc->migratepages, lru)
624                 count[!!page_is_file_cache(page)]++;
625
626         mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
627         mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
628 }
629
630 /* Similar to reclaim, but different enough that they don't share logic */
631 static bool too_many_isolated(struct zone *zone)
632 {
633         unsigned long active, inactive, isolated;
634
635         inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
636                                         zone_page_state(zone, NR_INACTIVE_ANON);
637         active = zone_page_state(zone, NR_ACTIVE_FILE) +
638                                         zone_page_state(zone, NR_ACTIVE_ANON);
639         isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
640                                         zone_page_state(zone, NR_ISOLATED_ANON);
641
642         return isolated > (inactive + active) / 2;
643 }
644
645 /**
646  * isolate_migratepages_block() - isolate all migrate-able pages within
647  *                                a single pageblock
648  * @cc:         Compaction control structure.
649  * @low_pfn:    The first PFN to isolate
650  * @end_pfn:    The one-past-the-last PFN to isolate, within same pageblock
651  * @isolate_mode: Isolation mode to be used.
652  *
653  * Isolate all pages that can be migrated from the range specified by
654  * [low_pfn, end_pfn). The range is expected to be within same pageblock.
655  * Returns zero if there is a fatal signal pending, otherwise PFN of the
656  * first page that was not scanned (which may be both less, equal to or more
657  * than end_pfn).
658  *
659  * The pages are isolated on cc->migratepages list (not required to be empty),
660  * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
661  * is neither read nor updated.
662  */
663 static unsigned long
664 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
665                         unsigned long end_pfn, isolate_mode_t isolate_mode)
666 {
667         struct zone *zone = cc->zone;
668         unsigned long nr_scanned = 0, nr_isolated = 0;
669         struct list_head *migratelist = &cc->migratepages;
670         struct lruvec *lruvec;
671         unsigned long flags = 0;
672         bool locked = false;
673         struct page *page = NULL, *valid_page = NULL;
674         unsigned long start_pfn = low_pfn;
675
676         /*
677          * Ensure that there are not too many pages isolated from the LRU
678          * list by either parallel reclaimers or compaction. If there are,
679          * delay for some time until fewer pages are isolated
680          */
681         while (unlikely(too_many_isolated(zone))) {
682                 /* async migration should just abort */
683                 if (cc->mode == MIGRATE_ASYNC)
684                         return 0;
685
686                 congestion_wait(BLK_RW_ASYNC, HZ/10);
687
688                 if (fatal_signal_pending(current))
689                         return 0;
690         }
691
692         if (compact_should_abort(cc))
693                 return 0;
694
695         /* Time to isolate some pages for migration */
696         for (; low_pfn < end_pfn; low_pfn++) {
697                 bool is_lru;
698
699                 /*
700                  * Periodically drop the lock (if held) regardless of its
701                  * contention, to give chance to IRQs. Abort async compaction
702                  * if contended.
703                  */
704                 if (!(low_pfn % SWAP_CLUSTER_MAX)
705                     && compact_unlock_should_abort(&zone->lru_lock, flags,
706                                                                 &locked, cc))
707                         break;
708
709                 if (!pfn_valid_within(low_pfn))
710                         continue;
711                 nr_scanned++;
712
713                 page = pfn_to_page(low_pfn);
714
715                 if (!valid_page)
716                         valid_page = page;
717
718                 /*
719                  * Skip if free. We read page order here without zone lock
720                  * which is generally unsafe, but the race window is small and
721                  * the worst thing that can happen is that we skip some
722                  * potential isolation targets.
723                  */
724                 if (PageBuddy(page)) {
725                         unsigned long freepage_order = page_order_unsafe(page);
726
727                         /*
728                          * Without lock, we cannot be sure that what we got is
729                          * a valid page order. Consider only values in the
730                          * valid order range to prevent low_pfn overflow.
731                          */
732                         if (freepage_order > 0 && freepage_order < MAX_ORDER)
733                                 low_pfn += (1UL << freepage_order) - 1;
734                         continue;
735                 }
736
737                 /*
738                  * Check may be lockless but that's ok as we recheck later.
739                  * It's possible to migrate LRU pages and balloon pages
740                  * Skip any other type of page
741                  */
742                 is_lru = PageLRU(page);
743                 if (!is_lru) {
744                         if (unlikely(balloon_page_movable(page))) {
745                                 if (balloon_page_isolate(page)) {
746                                         /* Successfully isolated */
747                                         goto isolate_success;
748                                 }
749                         }
750                 }
751
752                 /*
753                  * Regardless of being on LRU, compound pages such as THP and
754                  * hugetlbfs are not to be compacted. We can potentially save
755                  * a lot of iterations if we skip them at once. The check is
756                  * racy, but we can consider only valid values and the only
757                  * danger is skipping too much.
758                  */
759                 if (PageCompound(page)) {
760                         unsigned int comp_order = compound_order(page);
761
762                         if (likely(comp_order < MAX_ORDER))
763                                 low_pfn += (1UL << comp_order) - 1;
764
765                         continue;
766                 }
767
768                 if (!is_lru)
769                         continue;
770
771                 /*
772                  * Migration will fail if an anonymous page is pinned in memory,
773                  * so avoid taking lru_lock and isolating it unnecessarily in an
774                  * admittedly racy check.
775                  */
776                 if (!page_mapping(page) &&
777                     page_count(page) > page_mapcount(page))
778                         continue;
779
780                 /* If we already hold the lock, we can skip some rechecking */
781                 if (!locked) {
782                         locked = compact_trylock_irqsave(&zone->lru_lock,
783                                                                 &flags, cc);
784                         if (!locked)
785                                 break;
786
787                         /* Recheck PageLRU and PageCompound under lock */
788                         if (!PageLRU(page))
789                                 continue;
790
791                         /*
792                          * Page become compound since the non-locked check,
793                          * and it's on LRU. It can only be a THP so the order
794                          * is safe to read and it's 0 for tail pages.
795                          */
796                         if (unlikely(PageCompound(page))) {
797                                 low_pfn += (1UL << compound_order(page)) - 1;
798                                 continue;
799                         }
800                 }
801
802                 lruvec = mem_cgroup_page_lruvec(page, zone);
803
804                 /* Try isolate the page */
805                 if (__isolate_lru_page(page, isolate_mode) != 0)
806                         continue;
807
808                 VM_BUG_ON_PAGE(PageCompound(page), page);
809
810                 /* Successfully isolated */
811                 del_page_from_lru_list(page, lruvec, page_lru(page));
812
813 isolate_success:
814                 list_add(&page->lru, migratelist);
815                 cc->nr_migratepages++;
816                 nr_isolated++;
817
818                 /* Avoid isolating too much */
819                 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
820                         ++low_pfn;
821                         break;
822                 }
823         }
824
825         /*
826          * The PageBuddy() check could have potentially brought us outside
827          * the range to be scanned.
828          */
829         if (unlikely(low_pfn > end_pfn))
830                 low_pfn = end_pfn;
831
832         if (locked)
833                 spin_unlock_irqrestore(&zone->lru_lock, flags);
834
835         /*
836          * Update the pageblock-skip information and cached scanner pfn,
837          * if the whole pageblock was scanned without isolating any page.
838          */
839         if (low_pfn == end_pfn)
840                 update_pageblock_skip(cc, valid_page, nr_isolated, true);
841
842         trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
843                                                 nr_scanned, nr_isolated);
844
845         count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
846         if (nr_isolated)
847                 count_compact_events(COMPACTISOLATED, nr_isolated);
848
849         return low_pfn;
850 }
851
852 /**
853  * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
854  * @cc:        Compaction control structure.
855  * @start_pfn: The first PFN to start isolating.
856  * @end_pfn:   The one-past-last PFN.
857  *
858  * Returns zero if isolation fails fatally due to e.g. pending signal.
859  * Otherwise, function returns one-past-the-last PFN of isolated page
860  * (which may be greater than end_pfn if end fell in a middle of a THP page).
861  */
862 unsigned long
863 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
864                                                         unsigned long end_pfn)
865 {
866         unsigned long pfn, block_end_pfn;
867
868         /* Scan block by block. First and last block may be incomplete */
869         pfn = start_pfn;
870         block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
871
872         for (; pfn < end_pfn; pfn = block_end_pfn,
873                                 block_end_pfn += pageblock_nr_pages) {
874
875                 block_end_pfn = min(block_end_pfn, end_pfn);
876
877                 if (!pageblock_pfn_to_page(pfn, block_end_pfn, cc->zone))
878                         continue;
879
880                 pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
881                                                         ISOLATE_UNEVICTABLE);
882
883                 /*
884                  * In case of fatal failure, release everything that might
885                  * have been isolated in the previous iteration, and signal
886                  * the failure back to caller.
887                  */
888                 if (!pfn) {
889                         putback_movable_pages(&cc->migratepages);
890                         cc->nr_migratepages = 0;
891                         break;
892                 }
893
894                 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
895                         break;
896         }
897         acct_isolated(cc->zone, cc);
898
899         return pfn;
900 }
901
902 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
903 #ifdef CONFIG_COMPACTION
904
905 /* Returns true if the page is within a block suitable for migration to */
906 static bool suitable_migration_target(struct page *page)
907 {
908         /* If the page is a large free page, then disallow migration */
909         if (PageBuddy(page)) {
910                 /*
911                  * We are checking page_order without zone->lock taken. But
912                  * the only small danger is that we skip a potentially suitable
913                  * pageblock, so it's not worth to check order for valid range.
914                  */
915                 if (page_order_unsafe(page) >= pageblock_order)
916                         return false;
917         }
918
919         /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
920         if (migrate_async_suitable(get_pageblock_migratetype(page)))
921                 return true;
922
923         /* Otherwise skip the block */
924         return false;
925 }
926
927 /*
928  * Test whether the free scanner has reached the same or lower pageblock than
929  * the migration scanner, and compaction should thus terminate.
930  */
931 static inline bool compact_scanners_met(struct compact_control *cc)
932 {
933         return (cc->free_pfn >> pageblock_order)
934                 <= (cc->migrate_pfn >> pageblock_order);
935 }
936
937 /*
938  * Based on information in the current compact_control, find blocks
939  * suitable for isolating free pages from and then isolate them.
940  */
941 static void isolate_freepages(struct compact_control *cc)
942 {
943         struct zone *zone = cc->zone;
944         struct page *page;
945         unsigned long block_start_pfn;  /* start of current pageblock */
946         unsigned long isolate_start_pfn; /* exact pfn we start at */
947         unsigned long block_end_pfn;    /* end of current pageblock */
948         unsigned long low_pfn;       /* lowest pfn scanner is able to scan */
949         struct list_head *freelist = &cc->freepages;
950
951         /*
952          * Initialise the free scanner. The starting point is where we last
953          * successfully isolated from, zone-cached value, or the end of the
954          * zone when isolating for the first time. For looping we also need
955          * this pfn aligned down to the pageblock boundary, because we do
956          * block_start_pfn -= pageblock_nr_pages in the for loop.
957          * For ending point, take care when isolating in last pageblock of a
958          * a zone which ends in the middle of a pageblock.
959          * The low boundary is the end of the pageblock the migration scanner
960          * is using.
961          */
962         isolate_start_pfn = cc->free_pfn;
963         block_start_pfn = cc->free_pfn & ~(pageblock_nr_pages-1);
964         block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
965                                                 zone_end_pfn(zone));
966         low_pfn = ALIGN(cc->migrate_pfn + 1, pageblock_nr_pages);
967
968         /*
969          * Isolate free pages until enough are available to migrate the
970          * pages on cc->migratepages. We stop searching if the migrate
971          * and free page scanners meet or enough free pages are isolated.
972          */
973         for (; block_start_pfn >= low_pfn;
974                                 block_end_pfn = block_start_pfn,
975                                 block_start_pfn -= pageblock_nr_pages,
976                                 isolate_start_pfn = block_start_pfn) {
977
978                 /*
979                  * This can iterate a massively long zone without finding any
980                  * suitable migration targets, so periodically check if we need
981                  * to schedule, or even abort async compaction.
982                  */
983                 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
984                                                 && compact_should_abort(cc))
985                         break;
986
987                 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
988                                                                         zone);
989                 if (!page)
990                         continue;
991
992                 /* Check the block is suitable for migration */
993                 if (!suitable_migration_target(page))
994                         continue;
995
996                 /* If isolation recently failed, do not retry */
997                 if (!isolation_suitable(cc, page))
998                         continue;
999
1000                 /* Found a block suitable for isolating free pages from. */
1001                 isolate_freepages_block(cc, &isolate_start_pfn,
1002                                         block_end_pfn, freelist, false);
1003
1004                 /*
1005                  * If we isolated enough freepages, or aborted due to async
1006                  * compaction being contended, terminate the loop.
1007                  * Remember where the free scanner should restart next time,
1008                  * which is where isolate_freepages_block() left off.
1009                  * But if it scanned the whole pageblock, isolate_start_pfn
1010                  * now points at block_end_pfn, which is the start of the next
1011                  * pageblock.
1012                  * In that case we will however want to restart at the start
1013                  * of the previous pageblock.
1014                  */
1015                 if ((cc->nr_freepages >= cc->nr_migratepages)
1016                                                         || cc->contended) {
1017                         if (isolate_start_pfn >= block_end_pfn)
1018                                 isolate_start_pfn =
1019                                         block_start_pfn - pageblock_nr_pages;
1020                         break;
1021                 } else {
1022                         /*
1023                          * isolate_freepages_block() should not terminate
1024                          * prematurely unless contended, or isolated enough
1025                          */
1026                         VM_BUG_ON(isolate_start_pfn < block_end_pfn);
1027                 }
1028         }
1029
1030         /* split_free_page does not map the pages */
1031         map_pages(freelist);
1032
1033         /*
1034          * Record where the free scanner will restart next time. Either we
1035          * broke from the loop and set isolate_start_pfn based on the last
1036          * call to isolate_freepages_block(), or we met the migration scanner
1037          * and the loop terminated due to isolate_start_pfn < low_pfn
1038          */
1039         cc->free_pfn = isolate_start_pfn;
1040 }
1041
1042 /*
1043  * This is a migrate-callback that "allocates" freepages by taking pages
1044  * from the isolated freelists in the block we are migrating to.
1045  */
1046 static struct page *compaction_alloc(struct page *migratepage,
1047                                         unsigned long data,
1048                                         int **result)
1049 {
1050         struct compact_control *cc = (struct compact_control *)data;
1051         struct page *freepage;
1052
1053         /*
1054          * Isolate free pages if necessary, and if we are not aborting due to
1055          * contention.
1056          */
1057         if (list_empty(&cc->freepages)) {
1058                 if (!cc->contended)
1059                         isolate_freepages(cc);
1060
1061                 if (list_empty(&cc->freepages))
1062                         return NULL;
1063         }
1064
1065         freepage = list_entry(cc->freepages.next, struct page, lru);
1066         list_del(&freepage->lru);
1067         cc->nr_freepages--;
1068
1069         return freepage;
1070 }
1071
1072 /*
1073  * This is a migrate-callback that "frees" freepages back to the isolated
1074  * freelist.  All pages on the freelist are from the same zone, so there is no
1075  * special handling needed for NUMA.
1076  */
1077 static void compaction_free(struct page *page, unsigned long data)
1078 {
1079         struct compact_control *cc = (struct compact_control *)data;
1080
1081         list_add(&page->lru, &cc->freepages);
1082         cc->nr_freepages++;
1083 }
1084
1085 /* possible outcome of isolate_migratepages */
1086 typedef enum {
1087         ISOLATE_ABORT,          /* Abort compaction now */
1088         ISOLATE_NONE,           /* No pages isolated, continue scanning */
1089         ISOLATE_SUCCESS,        /* Pages isolated, migrate */
1090 } isolate_migrate_t;
1091
1092 /*
1093  * Allow userspace to control policy on scanning the unevictable LRU for
1094  * compactable pages.
1095  */
1096 int sysctl_compact_unevictable_allowed __read_mostly = 1;
1097
1098 /*
1099  * Isolate all pages that can be migrated from the first suitable block,
1100  * starting at the block pointed to by the migrate scanner pfn within
1101  * compact_control.
1102  */
1103 static isolate_migrate_t isolate_migratepages(struct zone *zone,
1104                                         struct compact_control *cc)
1105 {
1106         unsigned long low_pfn, end_pfn;
1107         unsigned long isolate_start_pfn;
1108         struct page *page;
1109         const isolate_mode_t isolate_mode =
1110                 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1111                 (cc->mode == MIGRATE_ASYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1112
1113         /*
1114          * Start at where we last stopped, or beginning of the zone as
1115          * initialized by compact_zone()
1116          */
1117         low_pfn = cc->migrate_pfn;
1118
1119         /* Only scan within a pageblock boundary */
1120         end_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages);
1121
1122         /*
1123          * Iterate over whole pageblocks until we find the first suitable.
1124          * Do not cross the free scanner.
1125          */
1126         for (; end_pfn <= cc->free_pfn;
1127                         low_pfn = end_pfn, end_pfn += pageblock_nr_pages) {
1128
1129                 /*
1130                  * This can potentially iterate a massively long zone with
1131                  * many pageblocks unsuitable, so periodically check if we
1132                  * need to schedule, or even abort async compaction.
1133                  */
1134                 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1135                                                 && compact_should_abort(cc))
1136                         break;
1137
1138                 page = pageblock_pfn_to_page(low_pfn, end_pfn, zone);
1139                 if (!page)
1140                         continue;
1141
1142                 /* If isolation recently failed, do not retry */
1143                 if (!isolation_suitable(cc, page))
1144                         continue;
1145
1146                 /*
1147                  * For async compaction, also only scan in MOVABLE blocks.
1148                  * Async compaction is optimistic to see if the minimum amount
1149                  * of work satisfies the allocation.
1150                  */
1151                 if (cc->mode == MIGRATE_ASYNC &&
1152                     !migrate_async_suitable(get_pageblock_migratetype(page)))
1153                         continue;
1154
1155                 /* Perform the isolation */
1156                 isolate_start_pfn = low_pfn;
1157                 low_pfn = isolate_migratepages_block(cc, low_pfn, end_pfn,
1158                                                                 isolate_mode);
1159
1160                 if (!low_pfn || cc->contended) {
1161                         acct_isolated(zone, cc);
1162                         return ISOLATE_ABORT;
1163                 }
1164
1165                 /*
1166                  * Record where we could have freed pages by migration and not
1167                  * yet flushed them to buddy allocator.
1168                  * - this is the lowest page that could have been isolated and
1169                  * then freed by migration.
1170                  */
1171                 if (cc->nr_migratepages && !cc->last_migrated_pfn)
1172                         cc->last_migrated_pfn = isolate_start_pfn;
1173
1174                 /*
1175                  * Either we isolated something and proceed with migration. Or
1176                  * we failed and compact_zone should decide if we should
1177                  * continue or not.
1178                  */
1179                 break;
1180         }
1181
1182         acct_isolated(zone, cc);
1183         /* Record where migration scanner will be restarted. */
1184         cc->migrate_pfn = low_pfn;
1185
1186         return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
1187 }
1188
1189 /*
1190  * order == -1 is expected when compacting via
1191  * /proc/sys/vm/compact_memory
1192  */
1193 static inline bool is_via_compact_memory(int order)
1194 {
1195         return order == -1;
1196 }
1197
1198 static int __compact_finished(struct zone *zone, struct compact_control *cc,
1199                             const int migratetype)
1200 {
1201         unsigned int order;
1202         unsigned long watermark;
1203
1204         if (cc->contended || fatal_signal_pending(current))
1205                 return COMPACT_CONTENDED;
1206
1207         /* Compaction run completes if the migrate and free scanner meet */
1208         if (compact_scanners_met(cc)) {
1209                 /* Let the next compaction start anew. */
1210                 reset_cached_positions(zone);
1211
1212                 /*
1213                  * Mark that the PG_migrate_skip information should be cleared
1214                  * by kswapd when it goes to sleep. kswapd does not set the
1215                  * flag itself as the decision to be clear should be directly
1216                  * based on an allocation request.
1217                  */
1218                 if (!current_is_kswapd())
1219                         zone->compact_blockskip_flush = true;
1220
1221                 return COMPACT_COMPLETE;
1222         }
1223
1224         if (is_via_compact_memory(cc->order))
1225                 return COMPACT_CONTINUE;
1226
1227         /* Compaction run is not finished if the watermark is not met */
1228         watermark = low_wmark_pages(zone);
1229
1230         if (!zone_watermark_ok(zone, cc->order, watermark, cc->classzone_idx,
1231                                                         cc->alloc_flags))
1232                 return COMPACT_CONTINUE;
1233
1234         /* Direct compactor: Is a suitable page free? */
1235         for (order = cc->order; order < MAX_ORDER; order++) {
1236                 struct free_area *area = &zone->free_area[order];
1237                 bool can_steal;
1238
1239                 /* Job done if page is free of the right migratetype */
1240                 if (!list_empty(&area->free_list[migratetype]))
1241                         return COMPACT_PARTIAL;
1242
1243 #ifdef CONFIG_CMA
1244                 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
1245                 if (migratetype == MIGRATE_MOVABLE &&
1246                         !list_empty(&area->free_list[MIGRATE_CMA]))
1247                         return COMPACT_PARTIAL;
1248 #endif
1249                 /*
1250                  * Job done if allocation would steal freepages from
1251                  * other migratetype buddy lists.
1252                  */
1253                 if (find_suitable_fallback(area, order, migratetype,
1254                                                 true, &can_steal) != -1)
1255                         return COMPACT_PARTIAL;
1256         }
1257
1258         return COMPACT_NO_SUITABLE_PAGE;
1259 }
1260
1261 static int compact_finished(struct zone *zone, struct compact_control *cc,
1262                             const int migratetype)
1263 {
1264         int ret;
1265
1266         ret = __compact_finished(zone, cc, migratetype);
1267         trace_mm_compaction_finished(zone, cc->order, ret);
1268         if (ret == COMPACT_NO_SUITABLE_PAGE)
1269                 ret = COMPACT_CONTINUE;
1270
1271         return ret;
1272 }
1273
1274 /*
1275  * compaction_suitable: Is this suitable to run compaction on this zone now?
1276  * Returns
1277  *   COMPACT_SKIPPED  - If there are too few free pages for compaction
1278  *   COMPACT_PARTIAL  - If the allocation would succeed without compaction
1279  *   COMPACT_CONTINUE - If compaction should run now
1280  */
1281 static unsigned long __compaction_suitable(struct zone *zone, int order,
1282                                         int alloc_flags, int classzone_idx)
1283 {
1284         int fragindex;
1285         unsigned long watermark;
1286
1287         if (is_via_compact_memory(order))
1288                 return COMPACT_CONTINUE;
1289
1290         watermark = low_wmark_pages(zone);
1291         /*
1292          * If watermarks for high-order allocation are already met, there
1293          * should be no need for compaction at all.
1294          */
1295         if (zone_watermark_ok(zone, order, watermark, classzone_idx,
1296                                                                 alloc_flags))
1297                 return COMPACT_PARTIAL;
1298
1299         /*
1300          * Watermarks for order-0 must be met for compaction. Note the 2UL.
1301          * This is because during migration, copies of pages need to be
1302          * allocated and for a short time, the footprint is higher
1303          */
1304         watermark += (2UL << order);
1305         if (!zone_watermark_ok(zone, 0, watermark, classzone_idx, alloc_flags))
1306                 return COMPACT_SKIPPED;
1307
1308         /*
1309          * fragmentation index determines if allocation failures are due to
1310          * low memory or external fragmentation
1311          *
1312          * index of -1000 would imply allocations might succeed depending on
1313          * watermarks, but we already failed the high-order watermark check
1314          * index towards 0 implies failure is due to lack of memory
1315          * index towards 1000 implies failure is due to fragmentation
1316          *
1317          * Only compact if a failure would be due to fragmentation.
1318          */
1319         fragindex = fragmentation_index(zone, order);
1320         if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
1321                 return COMPACT_NOT_SUITABLE_ZONE;
1322
1323         return COMPACT_CONTINUE;
1324 }
1325
1326 unsigned long compaction_suitable(struct zone *zone, int order,
1327                                         int alloc_flags, int classzone_idx)
1328 {
1329         unsigned long ret;
1330
1331         ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx);
1332         trace_mm_compaction_suitable(zone, order, ret);
1333         if (ret == COMPACT_NOT_SUITABLE_ZONE)
1334                 ret = COMPACT_SKIPPED;
1335
1336         return ret;
1337 }
1338
1339 static int compact_zone(struct zone *zone, struct compact_control *cc)
1340 {
1341         int ret;
1342         unsigned long start_pfn = zone->zone_start_pfn;
1343         unsigned long end_pfn = zone_end_pfn(zone);
1344         const int migratetype = gfpflags_to_migratetype(cc->gfp_mask);
1345         const bool sync = cc->mode != MIGRATE_ASYNC;
1346
1347         ret = compaction_suitable(zone, cc->order, cc->alloc_flags,
1348                                                         cc->classzone_idx);
1349         switch (ret) {
1350         case COMPACT_PARTIAL:
1351         case COMPACT_SKIPPED:
1352                 /* Compaction is likely to fail */
1353                 return ret;
1354         case COMPACT_CONTINUE:
1355                 /* Fall through to compaction */
1356                 ;
1357         }
1358
1359         /*
1360          * Clear pageblock skip if there were failures recently and compaction
1361          * is about to be retried after being deferred. kswapd does not do
1362          * this reset as it'll reset the cached information when going to sleep.
1363          */
1364         if (compaction_restarting(zone, cc->order) && !current_is_kswapd())
1365                 __reset_isolation_suitable(zone);
1366
1367         /*
1368          * Setup to move all movable pages to the end of the zone. Used cached
1369          * information on where the scanners should start but check that it
1370          * is initialised by ensuring the values are within zone boundaries.
1371          */
1372         cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync];
1373         cc->free_pfn = zone->compact_cached_free_pfn;
1374         if (cc->free_pfn < start_pfn || cc->free_pfn > end_pfn) {
1375                 cc->free_pfn = end_pfn & ~(pageblock_nr_pages-1);
1376                 zone->compact_cached_free_pfn = cc->free_pfn;
1377         }
1378         if (cc->migrate_pfn < start_pfn || cc->migrate_pfn > end_pfn) {
1379                 cc->migrate_pfn = start_pfn;
1380                 zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
1381                 zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
1382         }
1383         cc->last_migrated_pfn = 0;
1384
1385         trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
1386                                 cc->free_pfn, end_pfn, sync);
1387
1388         migrate_prep_local();
1389
1390         while ((ret = compact_finished(zone, cc, migratetype)) ==
1391                                                 COMPACT_CONTINUE) {
1392                 int err;
1393
1394                 switch (isolate_migratepages(zone, cc)) {
1395                 case ISOLATE_ABORT:
1396                         ret = COMPACT_CONTENDED;
1397                         putback_movable_pages(&cc->migratepages);
1398                         cc->nr_migratepages = 0;
1399                         goto out;
1400                 case ISOLATE_NONE:
1401                         /*
1402                          * We haven't isolated and migrated anything, but
1403                          * there might still be unflushed migrations from
1404                          * previous cc->order aligned block.
1405                          */
1406                         goto check_drain;
1407                 case ISOLATE_SUCCESS:
1408                         ;
1409                 }
1410
1411                 err = migrate_pages(&cc->migratepages, compaction_alloc,
1412                                 compaction_free, (unsigned long)cc, cc->mode,
1413                                 MR_COMPACTION);
1414
1415                 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
1416                                                         &cc->migratepages);
1417
1418                 /* All pages were either migrated or will be released */
1419                 cc->nr_migratepages = 0;
1420                 if (err) {
1421                         putback_movable_pages(&cc->migratepages);
1422                         /*
1423                          * migrate_pages() may return -ENOMEM when scanners meet
1424                          * and we want compact_finished() to detect it
1425                          */
1426                         if (err == -ENOMEM && !compact_scanners_met(cc)) {
1427                                 ret = COMPACT_CONTENDED;
1428                                 goto out;
1429                         }
1430                 }
1431
1432 check_drain:
1433                 /*
1434                  * Has the migration scanner moved away from the previous
1435                  * cc->order aligned block where we migrated from? If yes,
1436                  * flush the pages that were freed, so that they can merge and
1437                  * compact_finished() can detect immediately if allocation
1438                  * would succeed.
1439                  */
1440                 if (cc->order > 0 && cc->last_migrated_pfn) {
1441                         int cpu;
1442                         unsigned long current_block_start =
1443                                 cc->migrate_pfn & ~((1UL << cc->order) - 1);
1444
1445                         if (cc->last_migrated_pfn < current_block_start) {
1446                                 cpu = get_cpu();
1447                                 lru_add_drain_cpu(cpu);
1448                                 drain_local_pages(zone);
1449                                 put_cpu();
1450                                 /* No more flushing until we migrate again */
1451                                 cc->last_migrated_pfn = 0;
1452                         }
1453                 }
1454
1455         }
1456
1457 out:
1458         /*
1459          * Release free pages and update where the free scanner should restart,
1460          * so we don't leave any returned pages behind in the next attempt.
1461          */
1462         if (cc->nr_freepages > 0) {
1463                 unsigned long free_pfn = release_freepages(&cc->freepages);
1464
1465                 cc->nr_freepages = 0;
1466                 VM_BUG_ON(free_pfn == 0);
1467                 /* The cached pfn is always the first in a pageblock */
1468                 free_pfn &= ~(pageblock_nr_pages-1);
1469                 /*
1470                  * Only go back, not forward. The cached pfn might have been
1471                  * already reset to zone end in compact_finished()
1472                  */
1473                 if (free_pfn > zone->compact_cached_free_pfn)
1474                         zone->compact_cached_free_pfn = free_pfn;
1475         }
1476
1477         trace_mm_compaction_end(start_pfn, cc->migrate_pfn,
1478                                 cc->free_pfn, end_pfn, sync, ret);
1479
1480         if (ret == COMPACT_CONTENDED)
1481                 ret = COMPACT_PARTIAL;
1482
1483         return ret;
1484 }
1485
1486 static unsigned long compact_zone_order(struct zone *zone, int order,
1487                 gfp_t gfp_mask, enum migrate_mode mode, int *contended,
1488                 int alloc_flags, int classzone_idx)
1489 {
1490         unsigned long ret;
1491         struct compact_control cc = {
1492                 .nr_freepages = 0,
1493                 .nr_migratepages = 0,
1494                 .order = order,
1495                 .gfp_mask = gfp_mask,
1496                 .zone = zone,
1497                 .mode = mode,
1498                 .alloc_flags = alloc_flags,
1499                 .classzone_idx = classzone_idx,
1500         };
1501         INIT_LIST_HEAD(&cc.freepages);
1502         INIT_LIST_HEAD(&cc.migratepages);
1503
1504         ret = compact_zone(zone, &cc);
1505
1506         VM_BUG_ON(!list_empty(&cc.freepages));
1507         VM_BUG_ON(!list_empty(&cc.migratepages));
1508
1509         *contended = cc.contended;
1510         return ret;
1511 }
1512
1513 int sysctl_extfrag_threshold = 500;
1514
1515 /**
1516  * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1517  * @gfp_mask: The GFP mask of the current allocation
1518  * @order: The order of the current allocation
1519  * @alloc_flags: The allocation flags of the current allocation
1520  * @ac: The context of current allocation
1521  * @mode: The migration mode for async, sync light, or sync migration
1522  * @contended: Return value that determines if compaction was aborted due to
1523  *             need_resched() or lock contention
1524  *
1525  * This is the main entry point for direct page compaction.
1526  */
1527 unsigned long try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
1528                         int alloc_flags, const struct alloc_context *ac,
1529                         enum migrate_mode mode, int *contended)
1530 {
1531         int may_enter_fs = gfp_mask & __GFP_FS;
1532         int may_perform_io = gfp_mask & __GFP_IO;
1533         struct zoneref *z;
1534         struct zone *zone;
1535         int rc = COMPACT_DEFERRED;
1536         int all_zones_contended = COMPACT_CONTENDED_LOCK; /* init for &= op */
1537
1538         *contended = COMPACT_CONTENDED_NONE;
1539
1540         /* Check if the GFP flags allow compaction */
1541         if (!order || !may_enter_fs || !may_perform_io)
1542                 return COMPACT_SKIPPED;
1543
1544         trace_mm_compaction_try_to_compact_pages(order, gfp_mask, mode);
1545
1546         /* Compact each zone in the list */
1547         for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1548                                                                 ac->nodemask) {
1549                 int status;
1550                 int zone_contended;
1551
1552                 if (compaction_deferred(zone, order))
1553                         continue;
1554
1555                 status = compact_zone_order(zone, order, gfp_mask, mode,
1556                                 &zone_contended, alloc_flags,
1557                                 ac->classzone_idx);
1558                 rc = max(status, rc);
1559                 /*
1560                  * It takes at least one zone that wasn't lock contended
1561                  * to clear all_zones_contended.
1562                  */
1563                 all_zones_contended &= zone_contended;
1564
1565                 /* If a normal allocation would succeed, stop compacting */
1566                 if (zone_watermark_ok(zone, order, low_wmark_pages(zone),
1567                                         ac->classzone_idx, alloc_flags)) {
1568                         /*
1569                          * We think the allocation will succeed in this zone,
1570                          * but it is not certain, hence the false. The caller
1571                          * will repeat this with true if allocation indeed
1572                          * succeeds in this zone.
1573                          */
1574                         compaction_defer_reset(zone, order, false);
1575                         /*
1576                          * It is possible that async compaction aborted due to
1577                          * need_resched() and the watermarks were ok thanks to
1578                          * somebody else freeing memory. The allocation can
1579                          * however still fail so we better signal the
1580                          * need_resched() contention anyway (this will not
1581                          * prevent the allocation attempt).
1582                          */
1583                         if (zone_contended == COMPACT_CONTENDED_SCHED)
1584                                 *contended = COMPACT_CONTENDED_SCHED;
1585
1586                         goto break_loop;
1587                 }
1588
1589                 if (mode != MIGRATE_ASYNC && status == COMPACT_COMPLETE) {
1590                         /*
1591                          * We think that allocation won't succeed in this zone
1592                          * so we defer compaction there. If it ends up
1593                          * succeeding after all, it will be reset.
1594                          */
1595                         defer_compaction(zone, order);
1596                 }
1597
1598                 /*
1599                  * We might have stopped compacting due to need_resched() in
1600                  * async compaction, or due to a fatal signal detected. In that
1601                  * case do not try further zones and signal need_resched()
1602                  * contention.
1603                  */
1604                 if ((zone_contended == COMPACT_CONTENDED_SCHED)
1605                                         || fatal_signal_pending(current)) {
1606                         *contended = COMPACT_CONTENDED_SCHED;
1607                         goto break_loop;
1608                 }
1609
1610                 continue;
1611 break_loop:
1612                 /*
1613                  * We might not have tried all the zones, so  be conservative
1614                  * and assume they are not all lock contended.
1615                  */
1616                 all_zones_contended = 0;
1617                 break;
1618         }
1619
1620         /*
1621          * If at least one zone wasn't deferred or skipped, we report if all
1622          * zones that were tried were lock contended.
1623          */
1624         if (rc > COMPACT_SKIPPED && all_zones_contended)
1625                 *contended = COMPACT_CONTENDED_LOCK;
1626
1627         return rc;
1628 }
1629
1630
1631 /* Compact all zones within a node */
1632 static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
1633 {
1634         int zoneid;
1635         struct zone *zone;
1636
1637         for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1638
1639                 zone = &pgdat->node_zones[zoneid];
1640                 if (!populated_zone(zone))
1641                         continue;
1642
1643                 cc->nr_freepages = 0;
1644                 cc->nr_migratepages = 0;
1645                 cc->zone = zone;
1646                 INIT_LIST_HEAD(&cc->freepages);
1647                 INIT_LIST_HEAD(&cc->migratepages);
1648
1649                 /*
1650                  * When called via /proc/sys/vm/compact_memory
1651                  * this makes sure we compact the whole zone regardless of
1652                  * cached scanner positions.
1653                  */
1654                 if (is_via_compact_memory(cc->order))
1655                         __reset_isolation_suitable(zone);
1656
1657                 if (is_via_compact_memory(cc->order) ||
1658                                 !compaction_deferred(zone, cc->order))
1659                         compact_zone(zone, cc);
1660
1661                 if (cc->order > 0) {
1662                         if (zone_watermark_ok(zone, cc->order,
1663                                                 low_wmark_pages(zone), 0, 0))
1664                                 compaction_defer_reset(zone, cc->order, false);
1665                 }
1666
1667                 VM_BUG_ON(!list_empty(&cc->freepages));
1668                 VM_BUG_ON(!list_empty(&cc->migratepages));
1669         }
1670 }
1671
1672 void compact_pgdat(pg_data_t *pgdat, int order)
1673 {
1674         struct compact_control cc = {
1675                 .order = order,
1676                 .mode = MIGRATE_ASYNC,
1677         };
1678
1679         if (!order)
1680                 return;
1681
1682         __compact_pgdat(pgdat, &cc);
1683 }
1684
1685 static void compact_node(int nid)
1686 {
1687         struct compact_control cc = {
1688                 .order = -1,
1689                 .mode = MIGRATE_SYNC,
1690                 .ignore_skip_hint = true,
1691         };
1692
1693         __compact_pgdat(NODE_DATA(nid), &cc);
1694 }
1695
1696 /* Compact all nodes in the system */
1697 static void compact_nodes(void)
1698 {
1699         int nid;
1700
1701         /* Flush pending updates to the LRU lists */
1702         lru_add_drain_all();
1703
1704         for_each_online_node(nid)
1705                 compact_node(nid);
1706 }
1707
1708 /* The written value is actually unused, all memory is compacted */
1709 int sysctl_compact_memory;
1710
1711 /* This is the entry point for compacting all nodes via /proc/sys/vm */
1712 int sysctl_compaction_handler(struct ctl_table *table, int write,
1713                         void __user *buffer, size_t *length, loff_t *ppos)
1714 {
1715         if (write)
1716                 compact_nodes();
1717
1718         return 0;
1719 }
1720
1721 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1722                         void __user *buffer, size_t *length, loff_t *ppos)
1723 {
1724         proc_dointvec_minmax(table, write, buffer, length, ppos);
1725
1726         return 0;
1727 }
1728
1729 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1730 static ssize_t sysfs_compact_node(struct device *dev,
1731                         struct device_attribute *attr,
1732                         const char *buf, size_t count)
1733 {
1734         int nid = dev->id;
1735
1736         if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1737                 /* Flush pending updates to the LRU lists */
1738                 lru_add_drain_all();
1739
1740                 compact_node(nid);
1741         }
1742
1743         return count;
1744 }
1745 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1746
1747 int compaction_register_node(struct node *node)
1748 {
1749         return device_create_file(&node->dev, &dev_attr_compact);
1750 }
1751
1752 void compaction_unregister_node(struct node *node)
1753 {
1754         return device_remove_file(&node->dev, &dev_attr_compact);
1755 }
1756 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1757
1758 #endif /* CONFIG_COMPACTION */