Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/ebiederm...
[firefly-linux-kernel-4.4.55.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 #ifdef CONFIG_TRACEPOINTS
39 static const char *const compaction_status_string[] = {
40         "deferred",
41         "skipped",
42         "continue",
43         "partial",
44         "complete",
45         "no_suitable_page",
46         "not_suitable_zone",
47 };
48 #endif
49
50 #define CREATE_TRACE_POINTS
51 #include <trace/events/compaction.h>
52
53 static unsigned long release_freepages(struct list_head *freelist)
54 {
55         struct page *page, *next;
56         unsigned long high_pfn = 0;
57
58         list_for_each_entry_safe(page, next, freelist, lru) {
59                 unsigned long pfn = page_to_pfn(page);
60                 list_del(&page->lru);
61                 __free_page(page);
62                 if (pfn > high_pfn)
63                         high_pfn = pfn;
64         }
65
66         return high_pfn;
67 }
68
69 static void map_pages(struct list_head *list)
70 {
71         struct page *page;
72
73         list_for_each_entry(page, list, lru) {
74                 arch_alloc_page(page, 0);
75                 kernel_map_pages(page, 1, 1);
76                 kasan_alloc_pages(page, 0);
77         }
78 }
79
80 static inline bool migrate_async_suitable(int migratetype)
81 {
82         return is_migrate_cma(migratetype) || migratetype == MIGRATE_MOVABLE;
83 }
84
85 /*
86  * Check that the whole (or subset of) a pageblock given by the interval of
87  * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
88  * with the migration of free compaction scanner. The scanners then need to
89  * use only pfn_valid_within() check for arches that allow holes within
90  * pageblocks.
91  *
92  * Return struct page pointer of start_pfn, or NULL if checks were not passed.
93  *
94  * It's possible on some configurations to have a setup like node0 node1 node0
95  * i.e. it's possible that all pages within a zones range of pages do not
96  * belong to a single zone. We assume that a border between node0 and node1
97  * can occur within a single pageblock, but not a node0 node1 node0
98  * interleaving within a single pageblock. It is therefore sufficient to check
99  * the first and last page of a pageblock and avoid checking each individual
100  * page in a pageblock.
101  */
102 static struct page *pageblock_pfn_to_page(unsigned long start_pfn,
103                                 unsigned long end_pfn, struct zone *zone)
104 {
105         struct page *start_page;
106         struct page *end_page;
107
108         /* end_pfn is one past the range we are checking */
109         end_pfn--;
110
111         if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
112                 return NULL;
113
114         start_page = pfn_to_page(start_pfn);
115
116         if (page_zone(start_page) != zone)
117                 return NULL;
118
119         end_page = pfn_to_page(end_pfn);
120
121         /* This gives a shorter code than deriving page_zone(end_page) */
122         if (page_zone_id(start_page) != page_zone_id(end_page))
123                 return NULL;
124
125         return start_page;
126 }
127
128 #ifdef CONFIG_COMPACTION
129
130 /* Do not skip compaction more than 64 times */
131 #define COMPACT_MAX_DEFER_SHIFT 6
132
133 /*
134  * Compaction is deferred when compaction fails to result in a page
135  * allocation success. 1 << compact_defer_limit compactions are skipped up
136  * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
137  */
138 void defer_compaction(struct zone *zone, int order)
139 {
140         zone->compact_considered = 0;
141         zone->compact_defer_shift++;
142
143         if (order < zone->compact_order_failed)
144                 zone->compact_order_failed = order;
145
146         if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
147                 zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
148
149         trace_mm_compaction_defer_compaction(zone, order);
150 }
151
152 /* Returns true if compaction should be skipped this time */
153 bool compaction_deferred(struct zone *zone, int order)
154 {
155         unsigned long defer_limit = 1UL << zone->compact_defer_shift;
156
157         if (order < zone->compact_order_failed)
158                 return false;
159
160         /* Avoid possible overflow */
161         if (++zone->compact_considered > defer_limit)
162                 zone->compact_considered = defer_limit;
163
164         if (zone->compact_considered >= defer_limit)
165                 return false;
166
167         trace_mm_compaction_deferred(zone, order);
168
169         return true;
170 }
171
172 /*
173  * Update defer tracking counters after successful compaction of given order,
174  * which means an allocation either succeeded (alloc_success == true) or is
175  * expected to succeed.
176  */
177 void compaction_defer_reset(struct zone *zone, int order,
178                 bool alloc_success)
179 {
180         if (alloc_success) {
181                 zone->compact_considered = 0;
182                 zone->compact_defer_shift = 0;
183         }
184         if (order >= zone->compact_order_failed)
185                 zone->compact_order_failed = order + 1;
186
187         trace_mm_compaction_defer_reset(zone, order);
188 }
189
190 /* Returns true if restarting compaction after many failures */
191 bool compaction_restarting(struct zone *zone, int order)
192 {
193         if (order < zone->compact_order_failed)
194                 return false;
195
196         return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
197                 zone->compact_considered >= 1UL << zone->compact_defer_shift;
198 }
199
200 /* Returns true if the pageblock should be scanned for pages to isolate. */
201 static inline bool isolation_suitable(struct compact_control *cc,
202                                         struct page *page)
203 {
204         if (cc->ignore_skip_hint)
205                 return true;
206
207         return !get_pageblock_skip(page);
208 }
209
210 /*
211  * This function is called to clear all cached information on pageblocks that
212  * should be skipped for page isolation when the migrate and free page scanner
213  * meet.
214  */
215 static void __reset_isolation_suitable(struct zone *zone)
216 {
217         unsigned long start_pfn = zone->zone_start_pfn;
218         unsigned long end_pfn = zone_end_pfn(zone);
219         unsigned long pfn;
220
221         zone->compact_cached_migrate_pfn[0] = start_pfn;
222         zone->compact_cached_migrate_pfn[1] = start_pfn;
223         zone->compact_cached_free_pfn = end_pfn;
224         zone->compact_blockskip_flush = false;
225
226         /* Walk the zone and mark every pageblock as suitable for isolation */
227         for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
228                 struct page *page;
229
230                 cond_resched();
231
232                 if (!pfn_valid(pfn))
233                         continue;
234
235                 page = pfn_to_page(pfn);
236                 if (zone != page_zone(page))
237                         continue;
238
239                 clear_pageblock_skip(page);
240         }
241 }
242
243 void reset_isolation_suitable(pg_data_t *pgdat)
244 {
245         int zoneid;
246
247         for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
248                 struct zone *zone = &pgdat->node_zones[zoneid];
249                 if (!populated_zone(zone))
250                         continue;
251
252                 /* Only flush if a full compaction finished recently */
253                 if (zone->compact_blockskip_flush)
254                         __reset_isolation_suitable(zone);
255         }
256 }
257
258 /*
259  * If no pages were isolated then mark this pageblock to be skipped in the
260  * future. The information is later cleared by __reset_isolation_suitable().
261  */
262 static void update_pageblock_skip(struct compact_control *cc,
263                         struct page *page, unsigned long nr_isolated,
264                         bool migrate_scanner)
265 {
266         struct zone *zone = cc->zone;
267         unsigned long pfn;
268
269         if (cc->ignore_skip_hint)
270                 return;
271
272         if (!page)
273                 return;
274
275         if (nr_isolated)
276                 return;
277
278         set_pageblock_skip(page);
279
280         pfn = page_to_pfn(page);
281
282         /* Update where async and sync compaction should restart */
283         if (migrate_scanner) {
284                 if (pfn > zone->compact_cached_migrate_pfn[0])
285                         zone->compact_cached_migrate_pfn[0] = pfn;
286                 if (cc->mode != MIGRATE_ASYNC &&
287                     pfn > zone->compact_cached_migrate_pfn[1])
288                         zone->compact_cached_migrate_pfn[1] = pfn;
289         } else {
290                 if (pfn < zone->compact_cached_free_pfn)
291                         zone->compact_cached_free_pfn = pfn;
292         }
293 }
294 #else
295 static inline bool isolation_suitable(struct compact_control *cc,
296                                         struct page *page)
297 {
298         return true;
299 }
300
301 static void update_pageblock_skip(struct compact_control *cc,
302                         struct page *page, unsigned long nr_isolated,
303                         bool migrate_scanner)
304 {
305 }
306 #endif /* CONFIG_COMPACTION */
307
308 /*
309  * Compaction requires the taking of some coarse locks that are potentially
310  * very heavily contended. For async compaction, back out if the lock cannot
311  * be taken immediately. For sync compaction, spin on the lock if needed.
312  *
313  * Returns true if the lock is held
314  * Returns false if the lock is not held and compaction should abort
315  */
316 static bool compact_trylock_irqsave(spinlock_t *lock, unsigned long *flags,
317                                                 struct compact_control *cc)
318 {
319         if (cc->mode == MIGRATE_ASYNC) {
320                 if (!spin_trylock_irqsave(lock, *flags)) {
321                         cc->contended = COMPACT_CONTENDED_LOCK;
322                         return false;
323                 }
324         } else {
325                 spin_lock_irqsave(lock, *flags);
326         }
327
328         return true;
329 }
330
331 /*
332  * Compaction requires the taking of some coarse locks that are potentially
333  * very heavily contended. The lock should be periodically unlocked to avoid
334  * having disabled IRQs for a long time, even when there is nobody waiting on
335  * the lock. It might also be that allowing the IRQs will result in
336  * need_resched() becoming true. If scheduling is needed, async compaction
337  * aborts. Sync compaction schedules.
338  * Either compaction type will also abort if a fatal signal is pending.
339  * In either case if the lock was locked, it is dropped and not regained.
340  *
341  * Returns true if compaction should abort due to fatal signal pending, or
342  *              async compaction due to need_resched()
343  * Returns false when compaction can continue (sync compaction might have
344  *              scheduled)
345  */
346 static bool compact_unlock_should_abort(spinlock_t *lock,
347                 unsigned long flags, bool *locked, struct compact_control *cc)
348 {
349         if (*locked) {
350                 spin_unlock_irqrestore(lock, flags);
351                 *locked = false;
352         }
353
354         if (fatal_signal_pending(current)) {
355                 cc->contended = COMPACT_CONTENDED_SCHED;
356                 return true;
357         }
358
359         if (need_resched()) {
360                 if (cc->mode == MIGRATE_ASYNC) {
361                         cc->contended = COMPACT_CONTENDED_SCHED;
362                         return true;
363                 }
364                 cond_resched();
365         }
366
367         return false;
368 }
369
370 /*
371  * Aside from avoiding lock contention, compaction also periodically checks
372  * need_resched() and either schedules in sync compaction or aborts async
373  * compaction. This is similar to what compact_unlock_should_abort() does, but
374  * is used where no lock is concerned.
375  *
376  * Returns false when no scheduling was needed, or sync compaction scheduled.
377  * Returns true when async compaction should abort.
378  */
379 static inline bool compact_should_abort(struct compact_control *cc)
380 {
381         /* async compaction aborts if contended */
382         if (need_resched()) {
383                 if (cc->mode == MIGRATE_ASYNC) {
384                         cc->contended = COMPACT_CONTENDED_SCHED;
385                         return true;
386                 }
387
388                 cond_resched();
389         }
390
391         return false;
392 }
393
394 /*
395  * Isolate free pages onto a private freelist. If @strict is true, will abort
396  * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
397  * (even though it may still end up isolating some pages).
398  */
399 static unsigned long isolate_freepages_block(struct compact_control *cc,
400                                 unsigned long *start_pfn,
401                                 unsigned long end_pfn,
402                                 struct list_head *freelist,
403                                 bool strict)
404 {
405         int nr_scanned = 0, total_isolated = 0;
406         struct page *cursor, *valid_page = NULL;
407         unsigned long flags = 0;
408         bool locked = false;
409         unsigned long blockpfn = *start_pfn;
410
411         cursor = pfn_to_page(blockpfn);
412
413         /* Isolate free pages. */
414         for (; blockpfn < end_pfn; blockpfn++, cursor++) {
415                 int isolated, i;
416                 struct page *page = cursor;
417
418                 /*
419                  * Periodically drop the lock (if held) regardless of its
420                  * contention, to give chance to IRQs. Abort if fatal signal
421                  * pending or async compaction detects need_resched()
422                  */
423                 if (!(blockpfn % SWAP_CLUSTER_MAX)
424                     && compact_unlock_should_abort(&cc->zone->lock, flags,
425                                                                 &locked, cc))
426                         break;
427
428                 nr_scanned++;
429                 if (!pfn_valid_within(blockpfn))
430                         goto isolate_fail;
431
432                 if (!valid_page)
433                         valid_page = page;
434                 if (!PageBuddy(page))
435                         goto isolate_fail;
436
437                 /*
438                  * If we already hold the lock, we can skip some rechecking.
439                  * Note that if we hold the lock now, checked_pageblock was
440                  * already set in some previous iteration (or strict is true),
441                  * so it is correct to skip the suitable migration target
442                  * recheck as well.
443                  */
444                 if (!locked) {
445                         /*
446                          * The zone lock must be held to isolate freepages.
447                          * Unfortunately this is a very coarse lock and can be
448                          * heavily contended if there are parallel allocations
449                          * or parallel compactions. For async compaction do not
450                          * spin on the lock and we acquire the lock as late as
451                          * possible.
452                          */
453                         locked = compact_trylock_irqsave(&cc->zone->lock,
454                                                                 &flags, cc);
455                         if (!locked)
456                                 break;
457
458                         /* Recheck this is a buddy page under lock */
459                         if (!PageBuddy(page))
460                                 goto isolate_fail;
461                 }
462
463                 /* Found a free page, break it into order-0 pages */
464                 isolated = split_free_page(page);
465                 total_isolated += isolated;
466                 for (i = 0; i < isolated; i++) {
467                         list_add(&page->lru, freelist);
468                         page++;
469                 }
470
471                 /* If a page was split, advance to the end of it */
472                 if (isolated) {
473                         cc->nr_freepages += isolated;
474                         if (!strict &&
475                                 cc->nr_migratepages <= cc->nr_freepages) {
476                                 blockpfn += isolated;
477                                 break;
478                         }
479
480                         blockpfn += isolated - 1;
481                         cursor += isolated - 1;
482                         continue;
483                 }
484
485 isolate_fail:
486                 if (strict)
487                         break;
488                 else
489                         continue;
490
491         }
492
493         trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
494                                         nr_scanned, total_isolated);
495
496         /* Record how far we have got within the block */
497         *start_pfn = blockpfn;
498
499         /*
500          * If strict isolation is requested by CMA then check that all the
501          * pages requested were isolated. If there were any failures, 0 is
502          * returned and CMA will fail.
503          */
504         if (strict && blockpfn < end_pfn)
505                 total_isolated = 0;
506
507         if (locked)
508                 spin_unlock_irqrestore(&cc->zone->lock, flags);
509
510         /* Update the pageblock-skip if the whole pageblock was scanned */
511         if (blockpfn == end_pfn)
512                 update_pageblock_skip(cc, valid_page, total_isolated, false);
513
514         count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
515         if (total_isolated)
516                 count_compact_events(COMPACTISOLATED, total_isolated);
517         return total_isolated;
518 }
519
520 /**
521  * isolate_freepages_range() - isolate free pages.
522  * @start_pfn: The first PFN to start isolating.
523  * @end_pfn:   The one-past-last PFN.
524  *
525  * Non-free pages, invalid PFNs, or zone boundaries within the
526  * [start_pfn, end_pfn) range are considered errors, cause function to
527  * undo its actions and return zero.
528  *
529  * Otherwise, function returns one-past-the-last PFN of isolated page
530  * (which may be greater then end_pfn if end fell in a middle of
531  * a free page).
532  */
533 unsigned long
534 isolate_freepages_range(struct compact_control *cc,
535                         unsigned long start_pfn, unsigned long end_pfn)
536 {
537         unsigned long isolated, pfn, block_end_pfn;
538         LIST_HEAD(freelist);
539
540         pfn = start_pfn;
541         block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
542
543         for (; pfn < end_pfn; pfn += isolated,
544                                 block_end_pfn += pageblock_nr_pages) {
545                 /* Protect pfn from changing by isolate_freepages_block */
546                 unsigned long isolate_start_pfn = pfn;
547
548                 block_end_pfn = min(block_end_pfn, end_pfn);
549
550                 /*
551                  * pfn could pass the block_end_pfn if isolated freepage
552                  * is more than pageblock order. In this case, we adjust
553                  * scanning range to right one.
554                  */
555                 if (pfn >= block_end_pfn) {
556                         block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
557                         block_end_pfn = min(block_end_pfn, end_pfn);
558                 }
559
560                 if (!pageblock_pfn_to_page(pfn, block_end_pfn, cc->zone))
561                         break;
562
563                 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
564                                                 block_end_pfn, &freelist, true);
565
566                 /*
567                  * In strict mode, isolate_freepages_block() returns 0 if
568                  * there are any holes in the block (ie. invalid PFNs or
569                  * non-free pages).
570                  */
571                 if (!isolated)
572                         break;
573
574                 /*
575                  * If we managed to isolate pages, it is always (1 << n) *
576                  * pageblock_nr_pages for some non-negative n.  (Max order
577                  * page may span two pageblocks).
578                  */
579         }
580
581         /* split_free_page does not map the pages */
582         map_pages(&freelist);
583
584         if (pfn < end_pfn) {
585                 /* Loop terminated early, cleanup. */
586                 release_freepages(&freelist);
587                 return 0;
588         }
589
590         /* We don't use freelists for anything. */
591         return pfn;
592 }
593
594 /* Update the number of anon and file isolated pages in the zone */
595 static void acct_isolated(struct zone *zone, struct compact_control *cc)
596 {
597         struct page *page;
598         unsigned int count[2] = { 0, };
599
600         if (list_empty(&cc->migratepages))
601                 return;
602
603         list_for_each_entry(page, &cc->migratepages, lru)
604                 count[!!page_is_file_cache(page)]++;
605
606         mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
607         mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
608 }
609
610 /* Similar to reclaim, but different enough that they don't share logic */
611 static bool too_many_isolated(struct zone *zone)
612 {
613         unsigned long active, inactive, isolated;
614
615         inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
616                                         zone_page_state(zone, NR_INACTIVE_ANON);
617         active = zone_page_state(zone, NR_ACTIVE_FILE) +
618                                         zone_page_state(zone, NR_ACTIVE_ANON);
619         isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
620                                         zone_page_state(zone, NR_ISOLATED_ANON);
621
622         return isolated > (inactive + active) / 2;
623 }
624
625 /**
626  * isolate_migratepages_block() - isolate all migrate-able pages within
627  *                                a single pageblock
628  * @cc:         Compaction control structure.
629  * @low_pfn:    The first PFN to isolate
630  * @end_pfn:    The one-past-the-last PFN to isolate, within same pageblock
631  * @isolate_mode: Isolation mode to be used.
632  *
633  * Isolate all pages that can be migrated from the range specified by
634  * [low_pfn, end_pfn). The range is expected to be within same pageblock.
635  * Returns zero if there is a fatal signal pending, otherwise PFN of the
636  * first page that was not scanned (which may be both less, equal to or more
637  * than end_pfn).
638  *
639  * The pages are isolated on cc->migratepages list (not required to be empty),
640  * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
641  * is neither read nor updated.
642  */
643 static unsigned long
644 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
645                         unsigned long end_pfn, isolate_mode_t isolate_mode)
646 {
647         struct zone *zone = cc->zone;
648         unsigned long nr_scanned = 0, nr_isolated = 0;
649         struct list_head *migratelist = &cc->migratepages;
650         struct lruvec *lruvec;
651         unsigned long flags = 0;
652         bool locked = false;
653         struct page *page = NULL, *valid_page = NULL;
654         unsigned long start_pfn = low_pfn;
655
656         /*
657          * Ensure that there are not too many pages isolated from the LRU
658          * list by either parallel reclaimers or compaction. If there are,
659          * delay for some time until fewer pages are isolated
660          */
661         while (unlikely(too_many_isolated(zone))) {
662                 /* async migration should just abort */
663                 if (cc->mode == MIGRATE_ASYNC)
664                         return 0;
665
666                 congestion_wait(BLK_RW_ASYNC, HZ/10);
667
668                 if (fatal_signal_pending(current))
669                         return 0;
670         }
671
672         if (compact_should_abort(cc))
673                 return 0;
674
675         /* Time to isolate some pages for migration */
676         for (; low_pfn < end_pfn; low_pfn++) {
677                 /*
678                  * Periodically drop the lock (if held) regardless of its
679                  * contention, to give chance to IRQs. Abort async compaction
680                  * if contended.
681                  */
682                 if (!(low_pfn % SWAP_CLUSTER_MAX)
683                     && compact_unlock_should_abort(&zone->lru_lock, flags,
684                                                                 &locked, cc))
685                         break;
686
687                 if (!pfn_valid_within(low_pfn))
688                         continue;
689                 nr_scanned++;
690
691                 page = pfn_to_page(low_pfn);
692
693                 if (!valid_page)
694                         valid_page = page;
695
696                 /*
697                  * Skip if free. We read page order here without zone lock
698                  * which is generally unsafe, but the race window is small and
699                  * the worst thing that can happen is that we skip some
700                  * potential isolation targets.
701                  */
702                 if (PageBuddy(page)) {
703                         unsigned long freepage_order = page_order_unsafe(page);
704
705                         /*
706                          * Without lock, we cannot be sure that what we got is
707                          * a valid page order. Consider only values in the
708                          * valid order range to prevent low_pfn overflow.
709                          */
710                         if (freepage_order > 0 && freepage_order < MAX_ORDER)
711                                 low_pfn += (1UL << freepage_order) - 1;
712                         continue;
713                 }
714
715                 /*
716                  * Check may be lockless but that's ok as we recheck later.
717                  * It's possible to migrate LRU pages and balloon pages
718                  * Skip any other type of page
719                  */
720                 if (!PageLRU(page)) {
721                         if (unlikely(balloon_page_movable(page))) {
722                                 if (balloon_page_isolate(page)) {
723                                         /* Successfully isolated */
724                                         goto isolate_success;
725                                 }
726                         }
727                         continue;
728                 }
729
730                 /*
731                  * PageLRU is set. lru_lock normally excludes isolation
732                  * splitting and collapsing (collapsing has already happened
733                  * if PageLRU is set) but the lock is not necessarily taken
734                  * here and it is wasteful to take it just to check transhuge.
735                  * Check TransHuge without lock and skip the whole pageblock if
736                  * it's either a transhuge or hugetlbfs page, as calling
737                  * compound_order() without preventing THP from splitting the
738                  * page underneath us may return surprising results.
739                  */
740                 if (PageTransHuge(page)) {
741                         if (!locked)
742                                 low_pfn = ALIGN(low_pfn + 1,
743                                                 pageblock_nr_pages) - 1;
744                         else
745                                 low_pfn += (1 << compound_order(page)) - 1;
746
747                         continue;
748                 }
749
750                 /*
751                  * Migration will fail if an anonymous page is pinned in memory,
752                  * so avoid taking lru_lock and isolating it unnecessarily in an
753                  * admittedly racy check.
754                  */
755                 if (!page_mapping(page) &&
756                     page_count(page) > page_mapcount(page))
757                         continue;
758
759                 /* If we already hold the lock, we can skip some rechecking */
760                 if (!locked) {
761                         locked = compact_trylock_irqsave(&zone->lru_lock,
762                                                                 &flags, cc);
763                         if (!locked)
764                                 break;
765
766                         /* Recheck PageLRU and PageTransHuge under lock */
767                         if (!PageLRU(page))
768                                 continue;
769                         if (PageTransHuge(page)) {
770                                 low_pfn += (1 << compound_order(page)) - 1;
771                                 continue;
772                         }
773                 }
774
775                 lruvec = mem_cgroup_page_lruvec(page, zone);
776
777                 /* Try isolate the page */
778                 if (__isolate_lru_page(page, isolate_mode) != 0)
779                         continue;
780
781                 VM_BUG_ON_PAGE(PageTransCompound(page), page);
782
783                 /* Successfully isolated */
784                 del_page_from_lru_list(page, lruvec, page_lru(page));
785
786 isolate_success:
787                 list_add(&page->lru, migratelist);
788                 cc->nr_migratepages++;
789                 nr_isolated++;
790
791                 /* Avoid isolating too much */
792                 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
793                         ++low_pfn;
794                         break;
795                 }
796         }
797
798         /*
799          * The PageBuddy() check could have potentially brought us outside
800          * the range to be scanned.
801          */
802         if (unlikely(low_pfn > end_pfn))
803                 low_pfn = end_pfn;
804
805         if (locked)
806                 spin_unlock_irqrestore(&zone->lru_lock, flags);
807
808         /*
809          * Update the pageblock-skip information and cached scanner pfn,
810          * if the whole pageblock was scanned without isolating any page.
811          */
812         if (low_pfn == end_pfn)
813                 update_pageblock_skip(cc, valid_page, nr_isolated, true);
814
815         trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
816                                                 nr_scanned, nr_isolated);
817
818         count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
819         if (nr_isolated)
820                 count_compact_events(COMPACTISOLATED, nr_isolated);
821
822         return low_pfn;
823 }
824
825 /**
826  * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
827  * @cc:        Compaction control structure.
828  * @start_pfn: The first PFN to start isolating.
829  * @end_pfn:   The one-past-last PFN.
830  *
831  * Returns zero if isolation fails fatally due to e.g. pending signal.
832  * Otherwise, function returns one-past-the-last PFN of isolated page
833  * (which may be greater than end_pfn if end fell in a middle of a THP page).
834  */
835 unsigned long
836 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
837                                                         unsigned long end_pfn)
838 {
839         unsigned long pfn, block_end_pfn;
840
841         /* Scan block by block. First and last block may be incomplete */
842         pfn = start_pfn;
843         block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
844
845         for (; pfn < end_pfn; pfn = block_end_pfn,
846                                 block_end_pfn += pageblock_nr_pages) {
847
848                 block_end_pfn = min(block_end_pfn, end_pfn);
849
850                 if (!pageblock_pfn_to_page(pfn, block_end_pfn, cc->zone))
851                         continue;
852
853                 pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
854                                                         ISOLATE_UNEVICTABLE);
855
856                 /*
857                  * In case of fatal failure, release everything that might
858                  * have been isolated in the previous iteration, and signal
859                  * the failure back to caller.
860                  */
861                 if (!pfn) {
862                         putback_movable_pages(&cc->migratepages);
863                         cc->nr_migratepages = 0;
864                         break;
865                 }
866
867                 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
868                         break;
869         }
870         acct_isolated(cc->zone, cc);
871
872         return pfn;
873 }
874
875 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
876 #ifdef CONFIG_COMPACTION
877
878 /* Returns true if the page is within a block suitable for migration to */
879 static bool suitable_migration_target(struct page *page)
880 {
881         /* If the page is a large free page, then disallow migration */
882         if (PageBuddy(page)) {
883                 /*
884                  * We are checking page_order without zone->lock taken. But
885                  * the only small danger is that we skip a potentially suitable
886                  * pageblock, so it's not worth to check order for valid range.
887                  */
888                 if (page_order_unsafe(page) >= pageblock_order)
889                         return false;
890         }
891
892         /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
893         if (migrate_async_suitable(get_pageblock_migratetype(page)))
894                 return true;
895
896         /* Otherwise skip the block */
897         return false;
898 }
899
900 /*
901  * Based on information in the current compact_control, find blocks
902  * suitable for isolating free pages from and then isolate them.
903  */
904 static void isolate_freepages(struct compact_control *cc)
905 {
906         struct zone *zone = cc->zone;
907         struct page *page;
908         unsigned long block_start_pfn;  /* start of current pageblock */
909         unsigned long isolate_start_pfn; /* exact pfn we start at */
910         unsigned long block_end_pfn;    /* end of current pageblock */
911         unsigned long low_pfn;       /* lowest pfn scanner is able to scan */
912         struct list_head *freelist = &cc->freepages;
913
914         /*
915          * Initialise the free scanner. The starting point is where we last
916          * successfully isolated from, zone-cached value, or the end of the
917          * zone when isolating for the first time. For looping we also need
918          * this pfn aligned down to the pageblock boundary, because we do
919          * block_start_pfn -= pageblock_nr_pages in the for loop.
920          * For ending point, take care when isolating in last pageblock of a
921          * a zone which ends in the middle of a pageblock.
922          * The low boundary is the end of the pageblock the migration scanner
923          * is using.
924          */
925         isolate_start_pfn = cc->free_pfn;
926         block_start_pfn = cc->free_pfn & ~(pageblock_nr_pages-1);
927         block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
928                                                 zone_end_pfn(zone));
929         low_pfn = ALIGN(cc->migrate_pfn + 1, pageblock_nr_pages);
930
931         /*
932          * Isolate free pages until enough are available to migrate the
933          * pages on cc->migratepages. We stop searching if the migrate
934          * and free page scanners meet or enough free pages are isolated.
935          */
936         for (; block_start_pfn >= low_pfn &&
937                         cc->nr_migratepages > cc->nr_freepages;
938                                 block_end_pfn = block_start_pfn,
939                                 block_start_pfn -= pageblock_nr_pages,
940                                 isolate_start_pfn = block_start_pfn) {
941
942                 /*
943                  * This can iterate a massively long zone without finding any
944                  * suitable migration targets, so periodically check if we need
945                  * to schedule, or even abort async compaction.
946                  */
947                 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
948                                                 && compact_should_abort(cc))
949                         break;
950
951                 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
952                                                                         zone);
953                 if (!page)
954                         continue;
955
956                 /* Check the block is suitable for migration */
957                 if (!suitable_migration_target(page))
958                         continue;
959
960                 /* If isolation recently failed, do not retry */
961                 if (!isolation_suitable(cc, page))
962                         continue;
963
964                 /* Found a block suitable for isolating free pages from. */
965                 isolate_freepages_block(cc, &isolate_start_pfn,
966                                         block_end_pfn, freelist, false);
967
968                 /*
969                  * Remember where the free scanner should restart next time,
970                  * which is where isolate_freepages_block() left off.
971                  * But if it scanned the whole pageblock, isolate_start_pfn
972                  * now points at block_end_pfn, which is the start of the next
973                  * pageblock.
974                  * In that case we will however want to restart at the start
975                  * of the previous pageblock.
976                  */
977                 cc->free_pfn = (isolate_start_pfn < block_end_pfn) ?
978                                 isolate_start_pfn :
979                                 block_start_pfn - pageblock_nr_pages;
980
981                 /*
982                  * isolate_freepages_block() might have aborted due to async
983                  * compaction being contended
984                  */
985                 if (cc->contended)
986                         break;
987         }
988
989         /* split_free_page does not map the pages */
990         map_pages(freelist);
991
992         /*
993          * If we crossed the migrate scanner, we want to keep it that way
994          * so that compact_finished() may detect this
995          */
996         if (block_start_pfn < low_pfn)
997                 cc->free_pfn = cc->migrate_pfn;
998 }
999
1000 /*
1001  * This is a migrate-callback that "allocates" freepages by taking pages
1002  * from the isolated freelists in the block we are migrating to.
1003  */
1004 static struct page *compaction_alloc(struct page *migratepage,
1005                                         unsigned long data,
1006                                         int **result)
1007 {
1008         struct compact_control *cc = (struct compact_control *)data;
1009         struct page *freepage;
1010
1011         /*
1012          * Isolate free pages if necessary, and if we are not aborting due to
1013          * contention.
1014          */
1015         if (list_empty(&cc->freepages)) {
1016                 if (!cc->contended)
1017                         isolate_freepages(cc);
1018
1019                 if (list_empty(&cc->freepages))
1020                         return NULL;
1021         }
1022
1023         freepage = list_entry(cc->freepages.next, struct page, lru);
1024         list_del(&freepage->lru);
1025         cc->nr_freepages--;
1026
1027         return freepage;
1028 }
1029
1030 /*
1031  * This is a migrate-callback that "frees" freepages back to the isolated
1032  * freelist.  All pages on the freelist are from the same zone, so there is no
1033  * special handling needed for NUMA.
1034  */
1035 static void compaction_free(struct page *page, unsigned long data)
1036 {
1037         struct compact_control *cc = (struct compact_control *)data;
1038
1039         list_add(&page->lru, &cc->freepages);
1040         cc->nr_freepages++;
1041 }
1042
1043 /* possible outcome of isolate_migratepages */
1044 typedef enum {
1045         ISOLATE_ABORT,          /* Abort compaction now */
1046         ISOLATE_NONE,           /* No pages isolated, continue scanning */
1047         ISOLATE_SUCCESS,        /* Pages isolated, migrate */
1048 } isolate_migrate_t;
1049
1050 /*
1051  * Allow userspace to control policy on scanning the unevictable LRU for
1052  * compactable pages.
1053  */
1054 int sysctl_compact_unevictable_allowed __read_mostly = 1;
1055
1056 /*
1057  * Isolate all pages that can be migrated from the first suitable block,
1058  * starting at the block pointed to by the migrate scanner pfn within
1059  * compact_control.
1060  */
1061 static isolate_migrate_t isolate_migratepages(struct zone *zone,
1062                                         struct compact_control *cc)
1063 {
1064         unsigned long low_pfn, end_pfn;
1065         struct page *page;
1066         const isolate_mode_t isolate_mode =
1067                 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1068                 (cc->mode == MIGRATE_ASYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1069
1070         /*
1071          * Start at where we last stopped, or beginning of the zone as
1072          * initialized by compact_zone()
1073          */
1074         low_pfn = cc->migrate_pfn;
1075
1076         /* Only scan within a pageblock boundary */
1077         end_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages);
1078
1079         /*
1080          * Iterate over whole pageblocks until we find the first suitable.
1081          * Do not cross the free scanner.
1082          */
1083         for (; end_pfn <= cc->free_pfn;
1084                         low_pfn = end_pfn, end_pfn += pageblock_nr_pages) {
1085
1086                 /*
1087                  * This can potentially iterate a massively long zone with
1088                  * many pageblocks unsuitable, so periodically check if we
1089                  * need to schedule, or even abort async compaction.
1090                  */
1091                 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1092                                                 && compact_should_abort(cc))
1093                         break;
1094
1095                 page = pageblock_pfn_to_page(low_pfn, end_pfn, zone);
1096                 if (!page)
1097                         continue;
1098
1099                 /* If isolation recently failed, do not retry */
1100                 if (!isolation_suitable(cc, page))
1101                         continue;
1102
1103                 /*
1104                  * For async compaction, also only scan in MOVABLE blocks.
1105                  * Async compaction is optimistic to see if the minimum amount
1106                  * of work satisfies the allocation.
1107                  */
1108                 if (cc->mode == MIGRATE_ASYNC &&
1109                     !migrate_async_suitable(get_pageblock_migratetype(page)))
1110                         continue;
1111
1112                 /* Perform the isolation */
1113                 low_pfn = isolate_migratepages_block(cc, low_pfn, end_pfn,
1114                                                                 isolate_mode);
1115
1116                 if (!low_pfn || cc->contended) {
1117                         acct_isolated(zone, cc);
1118                         return ISOLATE_ABORT;
1119                 }
1120
1121                 /*
1122                  * Either we isolated something and proceed with migration. Or
1123                  * we failed and compact_zone should decide if we should
1124                  * continue or not.
1125                  */
1126                 break;
1127         }
1128
1129         acct_isolated(zone, cc);
1130         /*
1131          * Record where migration scanner will be restarted. If we end up in
1132          * the same pageblock as the free scanner, make the scanners fully
1133          * meet so that compact_finished() terminates compaction.
1134          */
1135         cc->migrate_pfn = (end_pfn <= cc->free_pfn) ? low_pfn : cc->free_pfn;
1136
1137         return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
1138 }
1139
1140 static int __compact_finished(struct zone *zone, struct compact_control *cc,
1141                             const int migratetype)
1142 {
1143         unsigned int order;
1144         unsigned long watermark;
1145
1146         if (cc->contended || fatal_signal_pending(current))
1147                 return COMPACT_PARTIAL;
1148
1149         /* Compaction run completes if the migrate and free scanner meet */
1150         if (cc->free_pfn <= cc->migrate_pfn) {
1151                 /* Let the next compaction start anew. */
1152                 zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
1153                 zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
1154                 zone->compact_cached_free_pfn = zone_end_pfn(zone);
1155
1156                 /*
1157                  * Mark that the PG_migrate_skip information should be cleared
1158                  * by kswapd when it goes to sleep. kswapd does not set the
1159                  * flag itself as the decision to be clear should be directly
1160                  * based on an allocation request.
1161                  */
1162                 if (!current_is_kswapd())
1163                         zone->compact_blockskip_flush = true;
1164
1165                 return COMPACT_COMPLETE;
1166         }
1167
1168         /*
1169          * order == -1 is expected when compacting via
1170          * /proc/sys/vm/compact_memory
1171          */
1172         if (cc->order == -1)
1173                 return COMPACT_CONTINUE;
1174
1175         /* Compaction run is not finished if the watermark is not met */
1176         watermark = low_wmark_pages(zone);
1177
1178         if (!zone_watermark_ok(zone, cc->order, watermark, cc->classzone_idx,
1179                                                         cc->alloc_flags))
1180                 return COMPACT_CONTINUE;
1181
1182         /* Direct compactor: Is a suitable page free? */
1183         for (order = cc->order; order < MAX_ORDER; order++) {
1184                 struct free_area *area = &zone->free_area[order];
1185                 bool can_steal;
1186
1187                 /* Job done if page is free of the right migratetype */
1188                 if (!list_empty(&area->free_list[migratetype]))
1189                         return COMPACT_PARTIAL;
1190
1191 #ifdef CONFIG_CMA
1192                 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
1193                 if (migratetype == MIGRATE_MOVABLE &&
1194                         !list_empty(&area->free_list[MIGRATE_CMA]))
1195                         return COMPACT_PARTIAL;
1196 #endif
1197                 /*
1198                  * Job done if allocation would steal freepages from
1199                  * other migratetype buddy lists.
1200                  */
1201                 if (find_suitable_fallback(area, order, migratetype,
1202                                                 true, &can_steal) != -1)
1203                         return COMPACT_PARTIAL;
1204         }
1205
1206         return COMPACT_NO_SUITABLE_PAGE;
1207 }
1208
1209 static int compact_finished(struct zone *zone, struct compact_control *cc,
1210                             const int migratetype)
1211 {
1212         int ret;
1213
1214         ret = __compact_finished(zone, cc, migratetype);
1215         trace_mm_compaction_finished(zone, cc->order, ret);
1216         if (ret == COMPACT_NO_SUITABLE_PAGE)
1217                 ret = COMPACT_CONTINUE;
1218
1219         return ret;
1220 }
1221
1222 /*
1223  * compaction_suitable: Is this suitable to run compaction on this zone now?
1224  * Returns
1225  *   COMPACT_SKIPPED  - If there are too few free pages for compaction
1226  *   COMPACT_PARTIAL  - If the allocation would succeed without compaction
1227  *   COMPACT_CONTINUE - If compaction should run now
1228  */
1229 static unsigned long __compaction_suitable(struct zone *zone, int order,
1230                                         int alloc_flags, int classzone_idx)
1231 {
1232         int fragindex;
1233         unsigned long watermark;
1234
1235         /*
1236          * order == -1 is expected when compacting via
1237          * /proc/sys/vm/compact_memory
1238          */
1239         if (order == -1)
1240                 return COMPACT_CONTINUE;
1241
1242         watermark = low_wmark_pages(zone);
1243         /*
1244          * If watermarks for high-order allocation are already met, there
1245          * should be no need for compaction at all.
1246          */
1247         if (zone_watermark_ok(zone, order, watermark, classzone_idx,
1248                                                                 alloc_flags))
1249                 return COMPACT_PARTIAL;
1250
1251         /*
1252          * Watermarks for order-0 must be met for compaction. Note the 2UL.
1253          * This is because during migration, copies of pages need to be
1254          * allocated and for a short time, the footprint is higher
1255          */
1256         watermark += (2UL << order);
1257         if (!zone_watermark_ok(zone, 0, watermark, classzone_idx, alloc_flags))
1258                 return COMPACT_SKIPPED;
1259
1260         /*
1261          * fragmentation index determines if allocation failures are due to
1262          * low memory or external fragmentation
1263          *
1264          * index of -1000 would imply allocations might succeed depending on
1265          * watermarks, but we already failed the high-order watermark check
1266          * index towards 0 implies failure is due to lack of memory
1267          * index towards 1000 implies failure is due to fragmentation
1268          *
1269          * Only compact if a failure would be due to fragmentation.
1270          */
1271         fragindex = fragmentation_index(zone, order);
1272         if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
1273                 return COMPACT_NOT_SUITABLE_ZONE;
1274
1275         return COMPACT_CONTINUE;
1276 }
1277
1278 unsigned long compaction_suitable(struct zone *zone, int order,
1279                                         int alloc_flags, int classzone_idx)
1280 {
1281         unsigned long ret;
1282
1283         ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx);
1284         trace_mm_compaction_suitable(zone, order, ret);
1285         if (ret == COMPACT_NOT_SUITABLE_ZONE)
1286                 ret = COMPACT_SKIPPED;
1287
1288         return ret;
1289 }
1290
1291 static int compact_zone(struct zone *zone, struct compact_control *cc)
1292 {
1293         int ret;
1294         unsigned long start_pfn = zone->zone_start_pfn;
1295         unsigned long end_pfn = zone_end_pfn(zone);
1296         const int migratetype = gfpflags_to_migratetype(cc->gfp_mask);
1297         const bool sync = cc->mode != MIGRATE_ASYNC;
1298         unsigned long last_migrated_pfn = 0;
1299
1300         ret = compaction_suitable(zone, cc->order, cc->alloc_flags,
1301                                                         cc->classzone_idx);
1302         switch (ret) {
1303         case COMPACT_PARTIAL:
1304         case COMPACT_SKIPPED:
1305                 /* Compaction is likely to fail */
1306                 return ret;
1307         case COMPACT_CONTINUE:
1308                 /* Fall through to compaction */
1309                 ;
1310         }
1311
1312         /*
1313          * Clear pageblock skip if there were failures recently and compaction
1314          * is about to be retried after being deferred. kswapd does not do
1315          * this reset as it'll reset the cached information when going to sleep.
1316          */
1317         if (compaction_restarting(zone, cc->order) && !current_is_kswapd())
1318                 __reset_isolation_suitable(zone);
1319
1320         /*
1321          * Setup to move all movable pages to the end of the zone. Used cached
1322          * information on where the scanners should start but check that it
1323          * is initialised by ensuring the values are within zone boundaries.
1324          */
1325         cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync];
1326         cc->free_pfn = zone->compact_cached_free_pfn;
1327         if (cc->free_pfn < start_pfn || cc->free_pfn > end_pfn) {
1328                 cc->free_pfn = end_pfn & ~(pageblock_nr_pages-1);
1329                 zone->compact_cached_free_pfn = cc->free_pfn;
1330         }
1331         if (cc->migrate_pfn < start_pfn || cc->migrate_pfn > end_pfn) {
1332                 cc->migrate_pfn = start_pfn;
1333                 zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
1334                 zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
1335         }
1336
1337         trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
1338                                 cc->free_pfn, end_pfn, sync);
1339
1340         migrate_prep_local();
1341
1342         while ((ret = compact_finished(zone, cc, migratetype)) ==
1343                                                 COMPACT_CONTINUE) {
1344                 int err;
1345                 unsigned long isolate_start_pfn = cc->migrate_pfn;
1346
1347                 switch (isolate_migratepages(zone, cc)) {
1348                 case ISOLATE_ABORT:
1349                         ret = COMPACT_PARTIAL;
1350                         putback_movable_pages(&cc->migratepages);
1351                         cc->nr_migratepages = 0;
1352                         goto out;
1353                 case ISOLATE_NONE:
1354                         /*
1355                          * We haven't isolated and migrated anything, but
1356                          * there might still be unflushed migrations from
1357                          * previous cc->order aligned block.
1358                          */
1359                         goto check_drain;
1360                 case ISOLATE_SUCCESS:
1361                         ;
1362                 }
1363
1364                 err = migrate_pages(&cc->migratepages, compaction_alloc,
1365                                 compaction_free, (unsigned long)cc, cc->mode,
1366                                 MR_COMPACTION);
1367
1368                 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
1369                                                         &cc->migratepages);
1370
1371                 /* All pages were either migrated or will be released */
1372                 cc->nr_migratepages = 0;
1373                 if (err) {
1374                         putback_movable_pages(&cc->migratepages);
1375                         /*
1376                          * migrate_pages() may return -ENOMEM when scanners meet
1377                          * and we want compact_finished() to detect it
1378                          */
1379                         if (err == -ENOMEM && cc->free_pfn > cc->migrate_pfn) {
1380                                 ret = COMPACT_PARTIAL;
1381                                 goto out;
1382                         }
1383                 }
1384
1385                 /*
1386                  * Record where we could have freed pages by migration and not
1387                  * yet flushed them to buddy allocator. We use the pfn that
1388                  * isolate_migratepages() started from in this loop iteration
1389                  * - this is the lowest page that could have been isolated and
1390                  * then freed by migration.
1391                  */
1392                 if (!last_migrated_pfn)
1393                         last_migrated_pfn = isolate_start_pfn;
1394
1395 check_drain:
1396                 /*
1397                  * Has the migration scanner moved away from the previous
1398                  * cc->order aligned block where we migrated from? If yes,
1399                  * flush the pages that were freed, so that they can merge and
1400                  * compact_finished() can detect immediately if allocation
1401                  * would succeed.
1402                  */
1403                 if (cc->order > 0 && last_migrated_pfn) {
1404                         int cpu;
1405                         unsigned long current_block_start =
1406                                 cc->migrate_pfn & ~((1UL << cc->order) - 1);
1407
1408                         if (last_migrated_pfn < current_block_start) {
1409                                 cpu = get_cpu();
1410                                 lru_add_drain_cpu(cpu);
1411                                 drain_local_pages(zone);
1412                                 put_cpu();
1413                                 /* No more flushing until we migrate again */
1414                                 last_migrated_pfn = 0;
1415                         }
1416                 }
1417
1418         }
1419
1420 out:
1421         /*
1422          * Release free pages and update where the free scanner should restart,
1423          * so we don't leave any returned pages behind in the next attempt.
1424          */
1425         if (cc->nr_freepages > 0) {
1426                 unsigned long free_pfn = release_freepages(&cc->freepages);
1427
1428                 cc->nr_freepages = 0;
1429                 VM_BUG_ON(free_pfn == 0);
1430                 /* The cached pfn is always the first in a pageblock */
1431                 free_pfn &= ~(pageblock_nr_pages-1);
1432                 /*
1433                  * Only go back, not forward. The cached pfn might have been
1434                  * already reset to zone end in compact_finished()
1435                  */
1436                 if (free_pfn > zone->compact_cached_free_pfn)
1437                         zone->compact_cached_free_pfn = free_pfn;
1438         }
1439
1440         trace_mm_compaction_end(start_pfn, cc->migrate_pfn,
1441                                 cc->free_pfn, end_pfn, sync, ret);
1442
1443         return ret;
1444 }
1445
1446 static unsigned long compact_zone_order(struct zone *zone, int order,
1447                 gfp_t gfp_mask, enum migrate_mode mode, int *contended,
1448                 int alloc_flags, int classzone_idx)
1449 {
1450         unsigned long ret;
1451         struct compact_control cc = {
1452                 .nr_freepages = 0,
1453                 .nr_migratepages = 0,
1454                 .order = order,
1455                 .gfp_mask = gfp_mask,
1456                 .zone = zone,
1457                 .mode = mode,
1458                 .alloc_flags = alloc_flags,
1459                 .classzone_idx = classzone_idx,
1460         };
1461         INIT_LIST_HEAD(&cc.freepages);
1462         INIT_LIST_HEAD(&cc.migratepages);
1463
1464         ret = compact_zone(zone, &cc);
1465
1466         VM_BUG_ON(!list_empty(&cc.freepages));
1467         VM_BUG_ON(!list_empty(&cc.migratepages));
1468
1469         *contended = cc.contended;
1470         return ret;
1471 }
1472
1473 int sysctl_extfrag_threshold = 500;
1474
1475 /**
1476  * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1477  * @gfp_mask: The GFP mask of the current allocation
1478  * @order: The order of the current allocation
1479  * @alloc_flags: The allocation flags of the current allocation
1480  * @ac: The context of current allocation
1481  * @mode: The migration mode for async, sync light, or sync migration
1482  * @contended: Return value that determines if compaction was aborted due to
1483  *             need_resched() or lock contention
1484  *
1485  * This is the main entry point for direct page compaction.
1486  */
1487 unsigned long try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
1488                         int alloc_flags, const struct alloc_context *ac,
1489                         enum migrate_mode mode, int *contended)
1490 {
1491         int may_enter_fs = gfp_mask & __GFP_FS;
1492         int may_perform_io = gfp_mask & __GFP_IO;
1493         struct zoneref *z;
1494         struct zone *zone;
1495         int rc = COMPACT_DEFERRED;
1496         int all_zones_contended = COMPACT_CONTENDED_LOCK; /* init for &= op */
1497
1498         *contended = COMPACT_CONTENDED_NONE;
1499
1500         /* Check if the GFP flags allow compaction */
1501         if (!order || !may_enter_fs || !may_perform_io)
1502                 return COMPACT_SKIPPED;
1503
1504         trace_mm_compaction_try_to_compact_pages(order, gfp_mask, mode);
1505
1506         /* Compact each zone in the list */
1507         for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1508                                                                 ac->nodemask) {
1509                 int status;
1510                 int zone_contended;
1511
1512                 if (compaction_deferred(zone, order))
1513                         continue;
1514
1515                 status = compact_zone_order(zone, order, gfp_mask, mode,
1516                                 &zone_contended, alloc_flags,
1517                                 ac->classzone_idx);
1518                 rc = max(status, rc);
1519                 /*
1520                  * It takes at least one zone that wasn't lock contended
1521                  * to clear all_zones_contended.
1522                  */
1523                 all_zones_contended &= zone_contended;
1524
1525                 /* If a normal allocation would succeed, stop compacting */
1526                 if (zone_watermark_ok(zone, order, low_wmark_pages(zone),
1527                                         ac->classzone_idx, alloc_flags)) {
1528                         /*
1529                          * We think the allocation will succeed in this zone,
1530                          * but it is not certain, hence the false. The caller
1531                          * will repeat this with true if allocation indeed
1532                          * succeeds in this zone.
1533                          */
1534                         compaction_defer_reset(zone, order, false);
1535                         /*
1536                          * It is possible that async compaction aborted due to
1537                          * need_resched() and the watermarks were ok thanks to
1538                          * somebody else freeing memory. The allocation can
1539                          * however still fail so we better signal the
1540                          * need_resched() contention anyway (this will not
1541                          * prevent the allocation attempt).
1542                          */
1543                         if (zone_contended == COMPACT_CONTENDED_SCHED)
1544                                 *contended = COMPACT_CONTENDED_SCHED;
1545
1546                         goto break_loop;
1547                 }
1548
1549                 if (mode != MIGRATE_ASYNC && status == COMPACT_COMPLETE) {
1550                         /*
1551                          * We think that allocation won't succeed in this zone
1552                          * so we defer compaction there. If it ends up
1553                          * succeeding after all, it will be reset.
1554                          */
1555                         defer_compaction(zone, order);
1556                 }
1557
1558                 /*
1559                  * We might have stopped compacting due to need_resched() in
1560                  * async compaction, or due to a fatal signal detected. In that
1561                  * case do not try further zones and signal need_resched()
1562                  * contention.
1563                  */
1564                 if ((zone_contended == COMPACT_CONTENDED_SCHED)
1565                                         || fatal_signal_pending(current)) {
1566                         *contended = COMPACT_CONTENDED_SCHED;
1567                         goto break_loop;
1568                 }
1569
1570                 continue;
1571 break_loop:
1572                 /*
1573                  * We might not have tried all the zones, so  be conservative
1574                  * and assume they are not all lock contended.
1575                  */
1576                 all_zones_contended = 0;
1577                 break;
1578         }
1579
1580         /*
1581          * If at least one zone wasn't deferred or skipped, we report if all
1582          * zones that were tried were lock contended.
1583          */
1584         if (rc > COMPACT_SKIPPED && all_zones_contended)
1585                 *contended = COMPACT_CONTENDED_LOCK;
1586
1587         return rc;
1588 }
1589
1590
1591 /* Compact all zones within a node */
1592 static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
1593 {
1594         int zoneid;
1595         struct zone *zone;
1596
1597         for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1598
1599                 zone = &pgdat->node_zones[zoneid];
1600                 if (!populated_zone(zone))
1601                         continue;
1602
1603                 cc->nr_freepages = 0;
1604                 cc->nr_migratepages = 0;
1605                 cc->zone = zone;
1606                 INIT_LIST_HEAD(&cc->freepages);
1607                 INIT_LIST_HEAD(&cc->migratepages);
1608
1609                 /*
1610                  * When called via /proc/sys/vm/compact_memory
1611                  * this makes sure we compact the whole zone regardless of
1612                  * cached scanner positions.
1613                  */
1614                 if (cc->order == -1)
1615                         __reset_isolation_suitable(zone);
1616
1617                 if (cc->order == -1 || !compaction_deferred(zone, cc->order))
1618                         compact_zone(zone, cc);
1619
1620                 if (cc->order > 0) {
1621                         if (zone_watermark_ok(zone, cc->order,
1622                                                 low_wmark_pages(zone), 0, 0))
1623                                 compaction_defer_reset(zone, cc->order, false);
1624                 }
1625
1626                 VM_BUG_ON(!list_empty(&cc->freepages));
1627                 VM_BUG_ON(!list_empty(&cc->migratepages));
1628         }
1629 }
1630
1631 void compact_pgdat(pg_data_t *pgdat, int order)
1632 {
1633         struct compact_control cc = {
1634                 .order = order,
1635                 .mode = MIGRATE_ASYNC,
1636         };
1637
1638         if (!order)
1639                 return;
1640
1641         __compact_pgdat(pgdat, &cc);
1642 }
1643
1644 static void compact_node(int nid)
1645 {
1646         struct compact_control cc = {
1647                 .order = -1,
1648                 .mode = MIGRATE_SYNC,
1649                 .ignore_skip_hint = true,
1650         };
1651
1652         __compact_pgdat(NODE_DATA(nid), &cc);
1653 }
1654
1655 /* Compact all nodes in the system */
1656 static void compact_nodes(void)
1657 {
1658         int nid;
1659
1660         /* Flush pending updates to the LRU lists */
1661         lru_add_drain_all();
1662
1663         for_each_online_node(nid)
1664                 compact_node(nid);
1665 }
1666
1667 /* The written value is actually unused, all memory is compacted */
1668 int sysctl_compact_memory;
1669
1670 /* This is the entry point for compacting all nodes via /proc/sys/vm */
1671 int sysctl_compaction_handler(struct ctl_table *table, int write,
1672                         void __user *buffer, size_t *length, loff_t *ppos)
1673 {
1674         if (write)
1675                 compact_nodes();
1676
1677         return 0;
1678 }
1679
1680 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1681                         void __user *buffer, size_t *length, loff_t *ppos)
1682 {
1683         proc_dointvec_minmax(table, write, buffer, length, ppos);
1684
1685         return 0;
1686 }
1687
1688 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1689 static ssize_t sysfs_compact_node(struct device *dev,
1690                         struct device_attribute *attr,
1691                         const char *buf, size_t count)
1692 {
1693         int nid = dev->id;
1694
1695         if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1696                 /* Flush pending updates to the LRU lists */
1697                 lru_add_drain_all();
1698
1699                 compact_node(nid);
1700         }
1701
1702         return count;
1703 }
1704 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1705
1706 int compaction_register_node(struct node *node)
1707 {
1708         return device_create_file(&node->dev, &dev_attr_compact);
1709 }
1710
1711 void compaction_unregister_node(struct node *node)
1712 {
1713         return device_remove_file(&node->dev, &dev_attr_compact);
1714 }
1715 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1716
1717 #endif /* CONFIG_COMPACTION */