xref: /openbmc/linux/mm/compaction.c (revision ae3473231e77a3f1909d48cd144cebe5e1d049b3)
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/cpu.h>
11 #include <linux/swap.h>
12 #include <linux/migrate.h>
13 #include <linux/compaction.h>
14 #include <linux/mm_inline.h>
15 #include <linux/backing-dev.h>
16 #include <linux/sysctl.h>
17 #include <linux/sysfs.h>
18 #include <linux/page-isolation.h>
19 #include <linux/kasan.h>
20 #include <linux/kthread.h>
21 #include <linux/freezer.h>
22 #include <linux/page_owner.h>
23 #include "internal.h"
24 
25 #ifdef CONFIG_COMPACTION
26 static inline void count_compact_event(enum vm_event_item item)
27 {
28 	count_vm_event(item);
29 }
30 
31 static inline void count_compact_events(enum vm_event_item item, long delta)
32 {
33 	count_vm_events(item, delta);
34 }
35 #else
36 #define count_compact_event(item) do { } while (0)
37 #define count_compact_events(item, delta) do { } while (0)
38 #endif
39 
40 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
41 
42 #define CREATE_TRACE_POINTS
43 #include <trace/events/compaction.h>
44 
45 #define block_start_pfn(pfn, order)	round_down(pfn, 1UL << (order))
46 #define block_end_pfn(pfn, order)	ALIGN((pfn) + 1, 1UL << (order))
47 #define pageblock_start_pfn(pfn)	block_start_pfn(pfn, pageblock_order)
48 #define pageblock_end_pfn(pfn)		block_end_pfn(pfn, pageblock_order)
49 
50 static unsigned long release_freepages(struct list_head *freelist)
51 {
52 	struct page *page, *next;
53 	unsigned long high_pfn = 0;
54 
55 	list_for_each_entry_safe(page, next, freelist, lru) {
56 		unsigned long pfn = page_to_pfn(page);
57 		list_del(&page->lru);
58 		__free_page(page);
59 		if (pfn > high_pfn)
60 			high_pfn = pfn;
61 	}
62 
63 	return high_pfn;
64 }
65 
66 static void map_pages(struct list_head *list)
67 {
68 	unsigned int i, order, nr_pages;
69 	struct page *page, *next;
70 	LIST_HEAD(tmp_list);
71 
72 	list_for_each_entry_safe(page, next, list, lru) {
73 		list_del(&page->lru);
74 
75 		order = page_private(page);
76 		nr_pages = 1 << order;
77 
78 		post_alloc_hook(page, order, __GFP_MOVABLE);
79 		if (order)
80 			split_page(page, order);
81 
82 		for (i = 0; i < nr_pages; i++) {
83 			list_add(&page->lru, &tmp_list);
84 			page++;
85 		}
86 	}
87 
88 	list_splice(&tmp_list, list);
89 }
90 
91 static inline bool migrate_async_suitable(int migratetype)
92 {
93 	return is_migrate_cma(migratetype) || migratetype == MIGRATE_MOVABLE;
94 }
95 
96 #ifdef CONFIG_COMPACTION
97 
98 int PageMovable(struct page *page)
99 {
100 	struct address_space *mapping;
101 
102 	VM_BUG_ON_PAGE(!PageLocked(page), page);
103 	if (!__PageMovable(page))
104 		return 0;
105 
106 	mapping = page_mapping(page);
107 	if (mapping && mapping->a_ops && mapping->a_ops->isolate_page)
108 		return 1;
109 
110 	return 0;
111 }
112 EXPORT_SYMBOL(PageMovable);
113 
114 void __SetPageMovable(struct page *page, struct address_space *mapping)
115 {
116 	VM_BUG_ON_PAGE(!PageLocked(page), page);
117 	VM_BUG_ON_PAGE((unsigned long)mapping & PAGE_MAPPING_MOVABLE, page);
118 	page->mapping = (void *)((unsigned long)mapping | PAGE_MAPPING_MOVABLE);
119 }
120 EXPORT_SYMBOL(__SetPageMovable);
121 
122 void __ClearPageMovable(struct page *page)
123 {
124 	VM_BUG_ON_PAGE(!PageLocked(page), page);
125 	VM_BUG_ON_PAGE(!PageMovable(page), page);
126 	/*
127 	 * Clear registered address_space val with keeping PAGE_MAPPING_MOVABLE
128 	 * flag so that VM can catch up released page by driver after isolation.
129 	 * With it, VM migration doesn't try to put it back.
130 	 */
131 	page->mapping = (void *)((unsigned long)page->mapping &
132 				PAGE_MAPPING_MOVABLE);
133 }
134 EXPORT_SYMBOL(__ClearPageMovable);
135 
136 /* Do not skip compaction more than 64 times */
137 #define COMPACT_MAX_DEFER_SHIFT 6
138 
139 /*
140  * Compaction is deferred when compaction fails to result in a page
141  * allocation success. 1 << compact_defer_limit compactions are skipped up
142  * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
143  */
144 void defer_compaction(struct zone *zone, int order)
145 {
146 	zone->compact_considered = 0;
147 	zone->compact_defer_shift++;
148 
149 	if (order < zone->compact_order_failed)
150 		zone->compact_order_failed = order;
151 
152 	if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
153 		zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
154 
155 	trace_mm_compaction_defer_compaction(zone, order);
156 }
157 
158 /* Returns true if compaction should be skipped this time */
159 bool compaction_deferred(struct zone *zone, int order)
160 {
161 	unsigned long defer_limit = 1UL << zone->compact_defer_shift;
162 
163 	if (order < zone->compact_order_failed)
164 		return false;
165 
166 	/* Avoid possible overflow */
167 	if (++zone->compact_considered > defer_limit)
168 		zone->compact_considered = defer_limit;
169 
170 	if (zone->compact_considered >= defer_limit)
171 		return false;
172 
173 	trace_mm_compaction_deferred(zone, order);
174 
175 	return true;
176 }
177 
178 /*
179  * Update defer tracking counters after successful compaction of given order,
180  * which means an allocation either succeeded (alloc_success == true) or is
181  * expected to succeed.
182  */
183 void compaction_defer_reset(struct zone *zone, int order,
184 		bool alloc_success)
185 {
186 	if (alloc_success) {
187 		zone->compact_considered = 0;
188 		zone->compact_defer_shift = 0;
189 	}
190 	if (order >= zone->compact_order_failed)
191 		zone->compact_order_failed = order + 1;
192 
193 	trace_mm_compaction_defer_reset(zone, order);
194 }
195 
196 /* Returns true if restarting compaction after many failures */
197 bool compaction_restarting(struct zone *zone, int order)
198 {
199 	if (order < zone->compact_order_failed)
200 		return false;
201 
202 	return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
203 		zone->compact_considered >= 1UL << zone->compact_defer_shift;
204 }
205 
206 /* Returns true if the pageblock should be scanned for pages to isolate. */
207 static inline bool isolation_suitable(struct compact_control *cc,
208 					struct page *page)
209 {
210 	if (cc->ignore_skip_hint)
211 		return true;
212 
213 	return !get_pageblock_skip(page);
214 }
215 
216 static void reset_cached_positions(struct zone *zone)
217 {
218 	zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
219 	zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
220 	zone->compact_cached_free_pfn =
221 				pageblock_start_pfn(zone_end_pfn(zone) - 1);
222 }
223 
224 /*
225  * This function is called to clear all cached information on pageblocks that
226  * should be skipped for page isolation when the migrate and free page scanner
227  * meet.
228  */
229 static void __reset_isolation_suitable(struct zone *zone)
230 {
231 	unsigned long start_pfn = zone->zone_start_pfn;
232 	unsigned long end_pfn = zone_end_pfn(zone);
233 	unsigned long pfn;
234 
235 	zone->compact_blockskip_flush = false;
236 
237 	/* Walk the zone and mark every pageblock as suitable for isolation */
238 	for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
239 		struct page *page;
240 
241 		cond_resched();
242 
243 		if (!pfn_valid(pfn))
244 			continue;
245 
246 		page = pfn_to_page(pfn);
247 		if (zone != page_zone(page))
248 			continue;
249 
250 		clear_pageblock_skip(page);
251 	}
252 
253 	reset_cached_positions(zone);
254 }
255 
256 void reset_isolation_suitable(pg_data_t *pgdat)
257 {
258 	int zoneid;
259 
260 	for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
261 		struct zone *zone = &pgdat->node_zones[zoneid];
262 		if (!populated_zone(zone))
263 			continue;
264 
265 		/* Only flush if a full compaction finished recently */
266 		if (zone->compact_blockskip_flush)
267 			__reset_isolation_suitable(zone);
268 	}
269 }
270 
271 /*
272  * If no pages were isolated then mark this pageblock to be skipped in the
273  * future. The information is later cleared by __reset_isolation_suitable().
274  */
275 static void update_pageblock_skip(struct compact_control *cc,
276 			struct page *page, unsigned long nr_isolated,
277 			bool migrate_scanner)
278 {
279 	struct zone *zone = cc->zone;
280 	unsigned long pfn;
281 
282 	if (cc->ignore_skip_hint)
283 		return;
284 
285 	if (!page)
286 		return;
287 
288 	if (nr_isolated)
289 		return;
290 
291 	set_pageblock_skip(page);
292 
293 	pfn = page_to_pfn(page);
294 
295 	/* Update where async and sync compaction should restart */
296 	if (migrate_scanner) {
297 		if (pfn > zone->compact_cached_migrate_pfn[0])
298 			zone->compact_cached_migrate_pfn[0] = pfn;
299 		if (cc->mode != MIGRATE_ASYNC &&
300 		    pfn > zone->compact_cached_migrate_pfn[1])
301 			zone->compact_cached_migrate_pfn[1] = pfn;
302 	} else {
303 		if (pfn < zone->compact_cached_free_pfn)
304 			zone->compact_cached_free_pfn = pfn;
305 	}
306 }
307 #else
308 static inline bool isolation_suitable(struct compact_control *cc,
309 					struct page *page)
310 {
311 	return true;
312 }
313 
314 static void update_pageblock_skip(struct compact_control *cc,
315 			struct page *page, unsigned long nr_isolated,
316 			bool migrate_scanner)
317 {
318 }
319 #endif /* CONFIG_COMPACTION */
320 
321 /*
322  * Compaction requires the taking of some coarse locks that are potentially
323  * very heavily contended. For async compaction, back out if the lock cannot
324  * be taken immediately. For sync compaction, spin on the lock if needed.
325  *
326  * Returns true if the lock is held
327  * Returns false if the lock is not held and compaction should abort
328  */
329 static bool compact_trylock_irqsave(spinlock_t *lock, unsigned long *flags,
330 						struct compact_control *cc)
331 {
332 	if (cc->mode == MIGRATE_ASYNC) {
333 		if (!spin_trylock_irqsave(lock, *flags)) {
334 			cc->contended = true;
335 			return false;
336 		}
337 	} else {
338 		spin_lock_irqsave(lock, *flags);
339 	}
340 
341 	return true;
342 }
343 
344 /*
345  * Compaction requires the taking of some coarse locks that are potentially
346  * very heavily contended. The lock should be periodically unlocked to avoid
347  * having disabled IRQs for a long time, even when there is nobody waiting on
348  * the lock. It might also be that allowing the IRQs will result in
349  * need_resched() becoming true. If scheduling is needed, async compaction
350  * aborts. Sync compaction schedules.
351  * Either compaction type will also abort if a fatal signal is pending.
352  * In either case if the lock was locked, it is dropped and not regained.
353  *
354  * Returns true if compaction should abort due to fatal signal pending, or
355  *		async compaction due to need_resched()
356  * Returns false when compaction can continue (sync compaction might have
357  *		scheduled)
358  */
359 static bool compact_unlock_should_abort(spinlock_t *lock,
360 		unsigned long flags, bool *locked, struct compact_control *cc)
361 {
362 	if (*locked) {
363 		spin_unlock_irqrestore(lock, flags);
364 		*locked = false;
365 	}
366 
367 	if (fatal_signal_pending(current)) {
368 		cc->contended = true;
369 		return true;
370 	}
371 
372 	if (need_resched()) {
373 		if (cc->mode == MIGRATE_ASYNC) {
374 			cc->contended = true;
375 			return true;
376 		}
377 		cond_resched();
378 	}
379 
380 	return false;
381 }
382 
383 /*
384  * Aside from avoiding lock contention, compaction also periodically checks
385  * need_resched() and either schedules in sync compaction or aborts async
386  * compaction. This is similar to what compact_unlock_should_abort() does, but
387  * is used where no lock is concerned.
388  *
389  * Returns false when no scheduling was needed, or sync compaction scheduled.
390  * Returns true when async compaction should abort.
391  */
392 static inline bool compact_should_abort(struct compact_control *cc)
393 {
394 	/* async compaction aborts if contended */
395 	if (need_resched()) {
396 		if (cc->mode == MIGRATE_ASYNC) {
397 			cc->contended = true;
398 			return true;
399 		}
400 
401 		cond_resched();
402 	}
403 
404 	return false;
405 }
406 
407 /*
408  * Isolate free pages onto a private freelist. If @strict is true, will abort
409  * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
410  * (even though it may still end up isolating some pages).
411  */
412 static unsigned long isolate_freepages_block(struct compact_control *cc,
413 				unsigned long *start_pfn,
414 				unsigned long end_pfn,
415 				struct list_head *freelist,
416 				bool strict)
417 {
418 	int nr_scanned = 0, total_isolated = 0;
419 	struct page *cursor, *valid_page = NULL;
420 	unsigned long flags = 0;
421 	bool locked = false;
422 	unsigned long blockpfn = *start_pfn;
423 	unsigned int order;
424 
425 	cursor = pfn_to_page(blockpfn);
426 
427 	/* Isolate free pages. */
428 	for (; blockpfn < end_pfn; blockpfn++, cursor++) {
429 		int isolated;
430 		struct page *page = cursor;
431 
432 		/*
433 		 * Periodically drop the lock (if held) regardless of its
434 		 * contention, to give chance to IRQs. Abort if fatal signal
435 		 * pending or async compaction detects need_resched()
436 		 */
437 		if (!(blockpfn % SWAP_CLUSTER_MAX)
438 		    && compact_unlock_should_abort(&cc->zone->lock, flags,
439 								&locked, cc))
440 			break;
441 
442 		nr_scanned++;
443 		if (!pfn_valid_within(blockpfn))
444 			goto isolate_fail;
445 
446 		if (!valid_page)
447 			valid_page = page;
448 
449 		/*
450 		 * For compound pages such as THP and hugetlbfs, we can save
451 		 * potentially a lot of iterations if we skip them at once.
452 		 * The check is racy, but we can consider only valid values
453 		 * and the only danger is skipping too much.
454 		 */
455 		if (PageCompound(page)) {
456 			unsigned int comp_order = compound_order(page);
457 
458 			if (likely(comp_order < MAX_ORDER)) {
459 				blockpfn += (1UL << comp_order) - 1;
460 				cursor += (1UL << comp_order) - 1;
461 			}
462 
463 			goto isolate_fail;
464 		}
465 
466 		if (!PageBuddy(page))
467 			goto isolate_fail;
468 
469 		/*
470 		 * If we already hold the lock, we can skip some rechecking.
471 		 * Note that if we hold the lock now, checked_pageblock was
472 		 * already set in some previous iteration (or strict is true),
473 		 * so it is correct to skip the suitable migration target
474 		 * recheck as well.
475 		 */
476 		if (!locked) {
477 			/*
478 			 * The zone lock must be held to isolate freepages.
479 			 * Unfortunately this is a very coarse lock and can be
480 			 * heavily contended if there are parallel allocations
481 			 * or parallel compactions. For async compaction do not
482 			 * spin on the lock and we acquire the lock as late as
483 			 * possible.
484 			 */
485 			locked = compact_trylock_irqsave(&cc->zone->lock,
486 								&flags, cc);
487 			if (!locked)
488 				break;
489 
490 			/* Recheck this is a buddy page under lock */
491 			if (!PageBuddy(page))
492 				goto isolate_fail;
493 		}
494 
495 		/* Found a free page, will break it into order-0 pages */
496 		order = page_order(page);
497 		isolated = __isolate_free_page(page, order);
498 		if (!isolated)
499 			break;
500 		set_page_private(page, order);
501 
502 		total_isolated += isolated;
503 		cc->nr_freepages += isolated;
504 		list_add_tail(&page->lru, freelist);
505 
506 		if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
507 			blockpfn += isolated;
508 			break;
509 		}
510 		/* Advance to the end of split page */
511 		blockpfn += isolated - 1;
512 		cursor += isolated - 1;
513 		continue;
514 
515 isolate_fail:
516 		if (strict)
517 			break;
518 		else
519 			continue;
520 
521 	}
522 
523 	if (locked)
524 		spin_unlock_irqrestore(&cc->zone->lock, flags);
525 
526 	/*
527 	 * There is a tiny chance that we have read bogus compound_order(),
528 	 * so be careful to not go outside of the pageblock.
529 	 */
530 	if (unlikely(blockpfn > end_pfn))
531 		blockpfn = end_pfn;
532 
533 	trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
534 					nr_scanned, total_isolated);
535 
536 	/* Record how far we have got within the block */
537 	*start_pfn = blockpfn;
538 
539 	/*
540 	 * If strict isolation is requested by CMA then check that all the
541 	 * pages requested were isolated. If there were any failures, 0 is
542 	 * returned and CMA will fail.
543 	 */
544 	if (strict && blockpfn < end_pfn)
545 		total_isolated = 0;
546 
547 	/* Update the pageblock-skip if the whole pageblock was scanned */
548 	if (blockpfn == end_pfn)
549 		update_pageblock_skip(cc, valid_page, total_isolated, false);
550 
551 	count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
552 	if (total_isolated)
553 		count_compact_events(COMPACTISOLATED, total_isolated);
554 	return total_isolated;
555 }
556 
557 /**
558  * isolate_freepages_range() - isolate free pages.
559  * @start_pfn: The first PFN to start isolating.
560  * @end_pfn:   The one-past-last PFN.
561  *
562  * Non-free pages, invalid PFNs, or zone boundaries within the
563  * [start_pfn, end_pfn) range are considered errors, cause function to
564  * undo its actions and return zero.
565  *
566  * Otherwise, function returns one-past-the-last PFN of isolated page
567  * (which may be greater then end_pfn if end fell in a middle of
568  * a free page).
569  */
570 unsigned long
571 isolate_freepages_range(struct compact_control *cc,
572 			unsigned long start_pfn, unsigned long end_pfn)
573 {
574 	unsigned long isolated, pfn, block_start_pfn, block_end_pfn;
575 	LIST_HEAD(freelist);
576 
577 	pfn = start_pfn;
578 	block_start_pfn = pageblock_start_pfn(pfn);
579 	if (block_start_pfn < cc->zone->zone_start_pfn)
580 		block_start_pfn = cc->zone->zone_start_pfn;
581 	block_end_pfn = pageblock_end_pfn(pfn);
582 
583 	for (; pfn < end_pfn; pfn += isolated,
584 				block_start_pfn = block_end_pfn,
585 				block_end_pfn += pageblock_nr_pages) {
586 		/* Protect pfn from changing by isolate_freepages_block */
587 		unsigned long isolate_start_pfn = pfn;
588 
589 		block_end_pfn = min(block_end_pfn, end_pfn);
590 
591 		/*
592 		 * pfn could pass the block_end_pfn if isolated freepage
593 		 * is more than pageblock order. In this case, we adjust
594 		 * scanning range to right one.
595 		 */
596 		if (pfn >= block_end_pfn) {
597 			block_start_pfn = pageblock_start_pfn(pfn);
598 			block_end_pfn = pageblock_end_pfn(pfn);
599 			block_end_pfn = min(block_end_pfn, end_pfn);
600 		}
601 
602 		if (!pageblock_pfn_to_page(block_start_pfn,
603 					block_end_pfn, cc->zone))
604 			break;
605 
606 		isolated = isolate_freepages_block(cc, &isolate_start_pfn,
607 						block_end_pfn, &freelist, true);
608 
609 		/*
610 		 * In strict mode, isolate_freepages_block() returns 0 if
611 		 * there are any holes in the block (ie. invalid PFNs or
612 		 * non-free pages).
613 		 */
614 		if (!isolated)
615 			break;
616 
617 		/*
618 		 * If we managed to isolate pages, it is always (1 << n) *
619 		 * pageblock_nr_pages for some non-negative n.  (Max order
620 		 * page may span two pageblocks).
621 		 */
622 	}
623 
624 	/* __isolate_free_page() does not map the pages */
625 	map_pages(&freelist);
626 
627 	if (pfn < end_pfn) {
628 		/* Loop terminated early, cleanup. */
629 		release_freepages(&freelist);
630 		return 0;
631 	}
632 
633 	/* We don't use freelists for anything. */
634 	return pfn;
635 }
636 
637 /* Similar to reclaim, but different enough that they don't share logic */
638 static bool too_many_isolated(struct zone *zone)
639 {
640 	unsigned long active, inactive, isolated;
641 
642 	inactive = node_page_state(zone->zone_pgdat, NR_INACTIVE_FILE) +
643 			node_page_state(zone->zone_pgdat, NR_INACTIVE_ANON);
644 	active = node_page_state(zone->zone_pgdat, NR_ACTIVE_FILE) +
645 			node_page_state(zone->zone_pgdat, NR_ACTIVE_ANON);
646 	isolated = node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE) +
647 			node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON);
648 
649 	return isolated > (inactive + active) / 2;
650 }
651 
652 /**
653  * isolate_migratepages_block() - isolate all migrate-able pages within
654  *				  a single pageblock
655  * @cc:		Compaction control structure.
656  * @low_pfn:	The first PFN to isolate
657  * @end_pfn:	The one-past-the-last PFN to isolate, within same pageblock
658  * @isolate_mode: Isolation mode to be used.
659  *
660  * Isolate all pages that can be migrated from the range specified by
661  * [low_pfn, end_pfn). The range is expected to be within same pageblock.
662  * Returns zero if there is a fatal signal pending, otherwise PFN of the
663  * first page that was not scanned (which may be both less, equal to or more
664  * than end_pfn).
665  *
666  * The pages are isolated on cc->migratepages list (not required to be empty),
667  * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
668  * is neither read nor updated.
669  */
670 static unsigned long
671 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
672 			unsigned long end_pfn, isolate_mode_t isolate_mode)
673 {
674 	struct zone *zone = cc->zone;
675 	unsigned long nr_scanned = 0, nr_isolated = 0;
676 	struct lruvec *lruvec;
677 	unsigned long flags = 0;
678 	bool locked = false;
679 	struct page *page = NULL, *valid_page = NULL;
680 	unsigned long start_pfn = low_pfn;
681 	bool skip_on_failure = false;
682 	unsigned long next_skip_pfn = 0;
683 
684 	/*
685 	 * Ensure that there are not too many pages isolated from the LRU
686 	 * list by either parallel reclaimers or compaction. If there are,
687 	 * delay for some time until fewer pages are isolated
688 	 */
689 	while (unlikely(too_many_isolated(zone))) {
690 		/* async migration should just abort */
691 		if (cc->mode == MIGRATE_ASYNC)
692 			return 0;
693 
694 		congestion_wait(BLK_RW_ASYNC, HZ/10);
695 
696 		if (fatal_signal_pending(current))
697 			return 0;
698 	}
699 
700 	if (compact_should_abort(cc))
701 		return 0;
702 
703 	if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
704 		skip_on_failure = true;
705 		next_skip_pfn = block_end_pfn(low_pfn, cc->order);
706 	}
707 
708 	/* Time to isolate some pages for migration */
709 	for (; low_pfn < end_pfn; low_pfn++) {
710 
711 		if (skip_on_failure && low_pfn >= next_skip_pfn) {
712 			/*
713 			 * We have isolated all migration candidates in the
714 			 * previous order-aligned block, and did not skip it due
715 			 * to failure. We should migrate the pages now and
716 			 * hopefully succeed compaction.
717 			 */
718 			if (nr_isolated)
719 				break;
720 
721 			/*
722 			 * We failed to isolate in the previous order-aligned
723 			 * block. Set the new boundary to the end of the
724 			 * current block. Note we can't simply increase
725 			 * next_skip_pfn by 1 << order, as low_pfn might have
726 			 * been incremented by a higher number due to skipping
727 			 * a compound or a high-order buddy page in the
728 			 * previous loop iteration.
729 			 */
730 			next_skip_pfn = block_end_pfn(low_pfn, cc->order);
731 		}
732 
733 		/*
734 		 * Periodically drop the lock (if held) regardless of its
735 		 * contention, to give chance to IRQs. Abort async compaction
736 		 * if contended.
737 		 */
738 		if (!(low_pfn % SWAP_CLUSTER_MAX)
739 		    && compact_unlock_should_abort(zone_lru_lock(zone), flags,
740 								&locked, cc))
741 			break;
742 
743 		if (!pfn_valid_within(low_pfn))
744 			goto isolate_fail;
745 		nr_scanned++;
746 
747 		page = pfn_to_page(low_pfn);
748 
749 		if (!valid_page)
750 			valid_page = page;
751 
752 		/*
753 		 * Skip if free. We read page order here without zone lock
754 		 * which is generally unsafe, but the race window is small and
755 		 * the worst thing that can happen is that we skip some
756 		 * potential isolation targets.
757 		 */
758 		if (PageBuddy(page)) {
759 			unsigned long freepage_order = page_order_unsafe(page);
760 
761 			/*
762 			 * Without lock, we cannot be sure that what we got is
763 			 * a valid page order. Consider only values in the
764 			 * valid order range to prevent low_pfn overflow.
765 			 */
766 			if (freepage_order > 0 && freepage_order < MAX_ORDER)
767 				low_pfn += (1UL << freepage_order) - 1;
768 			continue;
769 		}
770 
771 		/*
772 		 * Regardless of being on LRU, compound pages such as THP and
773 		 * hugetlbfs are not to be compacted. We can potentially save
774 		 * a lot of iterations if we skip them at once. The check is
775 		 * racy, but we can consider only valid values and the only
776 		 * danger is skipping too much.
777 		 */
778 		if (PageCompound(page)) {
779 			unsigned int comp_order = compound_order(page);
780 
781 			if (likely(comp_order < MAX_ORDER))
782 				low_pfn += (1UL << comp_order) - 1;
783 
784 			goto isolate_fail;
785 		}
786 
787 		/*
788 		 * Check may be lockless but that's ok as we recheck later.
789 		 * It's possible to migrate LRU and non-lru movable pages.
790 		 * Skip any other type of page
791 		 */
792 		if (!PageLRU(page)) {
793 			/*
794 			 * __PageMovable can return false positive so we need
795 			 * to verify it under page_lock.
796 			 */
797 			if (unlikely(__PageMovable(page)) &&
798 					!PageIsolated(page)) {
799 				if (locked) {
800 					spin_unlock_irqrestore(zone_lru_lock(zone),
801 									flags);
802 					locked = false;
803 				}
804 
805 				if (isolate_movable_page(page, isolate_mode))
806 					goto isolate_success;
807 			}
808 
809 			goto isolate_fail;
810 		}
811 
812 		/*
813 		 * Migration will fail if an anonymous page is pinned in memory,
814 		 * so avoid taking lru_lock and isolating it unnecessarily in an
815 		 * admittedly racy check.
816 		 */
817 		if (!page_mapping(page) &&
818 		    page_count(page) > page_mapcount(page))
819 			goto isolate_fail;
820 
821 		/*
822 		 * Only allow to migrate anonymous pages in GFP_NOFS context
823 		 * because those do not depend on fs locks.
824 		 */
825 		if (!(cc->gfp_mask & __GFP_FS) && page_mapping(page))
826 			goto isolate_fail;
827 
828 		/* If we already hold the lock, we can skip some rechecking */
829 		if (!locked) {
830 			locked = compact_trylock_irqsave(zone_lru_lock(zone),
831 								&flags, cc);
832 			if (!locked)
833 				break;
834 
835 			/* Recheck PageLRU and PageCompound under lock */
836 			if (!PageLRU(page))
837 				goto isolate_fail;
838 
839 			/*
840 			 * Page become compound since the non-locked check,
841 			 * and it's on LRU. It can only be a THP so the order
842 			 * is safe to read and it's 0 for tail pages.
843 			 */
844 			if (unlikely(PageCompound(page))) {
845 				low_pfn += (1UL << compound_order(page)) - 1;
846 				goto isolate_fail;
847 			}
848 		}
849 
850 		lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
851 
852 		/* Try isolate the page */
853 		if (__isolate_lru_page(page, isolate_mode) != 0)
854 			goto isolate_fail;
855 
856 		VM_BUG_ON_PAGE(PageCompound(page), page);
857 
858 		/* Successfully isolated */
859 		del_page_from_lru_list(page, lruvec, page_lru(page));
860 		inc_node_page_state(page,
861 				NR_ISOLATED_ANON + page_is_file_cache(page));
862 
863 isolate_success:
864 		list_add(&page->lru, &cc->migratepages);
865 		cc->nr_migratepages++;
866 		nr_isolated++;
867 
868 		/*
869 		 * Record where we could have freed pages by migration and not
870 		 * yet flushed them to buddy allocator.
871 		 * - this is the lowest page that was isolated and likely be
872 		 * then freed by migration.
873 		 */
874 		if (!cc->last_migrated_pfn)
875 			cc->last_migrated_pfn = low_pfn;
876 
877 		/* Avoid isolating too much */
878 		if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
879 			++low_pfn;
880 			break;
881 		}
882 
883 		continue;
884 isolate_fail:
885 		if (!skip_on_failure)
886 			continue;
887 
888 		/*
889 		 * We have isolated some pages, but then failed. Release them
890 		 * instead of migrating, as we cannot form the cc->order buddy
891 		 * page anyway.
892 		 */
893 		if (nr_isolated) {
894 			if (locked) {
895 				spin_unlock_irqrestore(zone_lru_lock(zone), flags);
896 				locked = false;
897 			}
898 			putback_movable_pages(&cc->migratepages);
899 			cc->nr_migratepages = 0;
900 			cc->last_migrated_pfn = 0;
901 			nr_isolated = 0;
902 		}
903 
904 		if (low_pfn < next_skip_pfn) {
905 			low_pfn = next_skip_pfn - 1;
906 			/*
907 			 * The check near the loop beginning would have updated
908 			 * next_skip_pfn too, but this is a bit simpler.
909 			 */
910 			next_skip_pfn += 1UL << cc->order;
911 		}
912 	}
913 
914 	/*
915 	 * The PageBuddy() check could have potentially brought us outside
916 	 * the range to be scanned.
917 	 */
918 	if (unlikely(low_pfn > end_pfn))
919 		low_pfn = end_pfn;
920 
921 	if (locked)
922 		spin_unlock_irqrestore(zone_lru_lock(zone), flags);
923 
924 	/*
925 	 * Update the pageblock-skip information and cached scanner pfn,
926 	 * if the whole pageblock was scanned without isolating any page.
927 	 */
928 	if (low_pfn == end_pfn)
929 		update_pageblock_skip(cc, valid_page, nr_isolated, true);
930 
931 	trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
932 						nr_scanned, nr_isolated);
933 
934 	count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
935 	if (nr_isolated)
936 		count_compact_events(COMPACTISOLATED, nr_isolated);
937 
938 	return low_pfn;
939 }
940 
941 /**
942  * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
943  * @cc:        Compaction control structure.
944  * @start_pfn: The first PFN to start isolating.
945  * @end_pfn:   The one-past-last PFN.
946  *
947  * Returns zero if isolation fails fatally due to e.g. pending signal.
948  * Otherwise, function returns one-past-the-last PFN of isolated page
949  * (which may be greater than end_pfn if end fell in a middle of a THP page).
950  */
951 unsigned long
952 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
953 							unsigned long end_pfn)
954 {
955 	unsigned long pfn, block_start_pfn, block_end_pfn;
956 
957 	/* Scan block by block. First and last block may be incomplete */
958 	pfn = start_pfn;
959 	block_start_pfn = pageblock_start_pfn(pfn);
960 	if (block_start_pfn < cc->zone->zone_start_pfn)
961 		block_start_pfn = cc->zone->zone_start_pfn;
962 	block_end_pfn = pageblock_end_pfn(pfn);
963 
964 	for (; pfn < end_pfn; pfn = block_end_pfn,
965 				block_start_pfn = block_end_pfn,
966 				block_end_pfn += pageblock_nr_pages) {
967 
968 		block_end_pfn = min(block_end_pfn, end_pfn);
969 
970 		if (!pageblock_pfn_to_page(block_start_pfn,
971 					block_end_pfn, cc->zone))
972 			continue;
973 
974 		pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
975 							ISOLATE_UNEVICTABLE);
976 
977 		if (!pfn)
978 			break;
979 
980 		if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
981 			break;
982 	}
983 
984 	return pfn;
985 }
986 
987 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
988 #ifdef CONFIG_COMPACTION
989 
990 /* Returns true if the page is within a block suitable for migration to */
991 static bool suitable_migration_target(struct compact_control *cc,
992 							struct page *page)
993 {
994 	if (cc->ignore_block_suitable)
995 		return true;
996 
997 	/* If the page is a large free page, then disallow migration */
998 	if (PageBuddy(page)) {
999 		/*
1000 		 * We are checking page_order without zone->lock taken. But
1001 		 * the only small danger is that we skip a potentially suitable
1002 		 * pageblock, so it's not worth to check order for valid range.
1003 		 */
1004 		if (page_order_unsafe(page) >= pageblock_order)
1005 			return false;
1006 	}
1007 
1008 	/* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1009 	if (migrate_async_suitable(get_pageblock_migratetype(page)))
1010 		return true;
1011 
1012 	/* Otherwise skip the block */
1013 	return false;
1014 }
1015 
1016 /*
1017  * Test whether the free scanner has reached the same or lower pageblock than
1018  * the migration scanner, and compaction should thus terminate.
1019  */
1020 static inline bool compact_scanners_met(struct compact_control *cc)
1021 {
1022 	return (cc->free_pfn >> pageblock_order)
1023 		<= (cc->migrate_pfn >> pageblock_order);
1024 }
1025 
1026 /*
1027  * Based on information in the current compact_control, find blocks
1028  * suitable for isolating free pages from and then isolate them.
1029  */
1030 static void isolate_freepages(struct compact_control *cc)
1031 {
1032 	struct zone *zone = cc->zone;
1033 	struct page *page;
1034 	unsigned long block_start_pfn;	/* start of current pageblock */
1035 	unsigned long isolate_start_pfn; /* exact pfn we start at */
1036 	unsigned long block_end_pfn;	/* end of current pageblock */
1037 	unsigned long low_pfn;	     /* lowest pfn scanner is able to scan */
1038 	struct list_head *freelist = &cc->freepages;
1039 
1040 	/*
1041 	 * Initialise the free scanner. The starting point is where we last
1042 	 * successfully isolated from, zone-cached value, or the end of the
1043 	 * zone when isolating for the first time. For looping we also need
1044 	 * this pfn aligned down to the pageblock boundary, because we do
1045 	 * block_start_pfn -= pageblock_nr_pages in the for loop.
1046 	 * For ending point, take care when isolating in last pageblock of a
1047 	 * a zone which ends in the middle of a pageblock.
1048 	 * The low boundary is the end of the pageblock the migration scanner
1049 	 * is using.
1050 	 */
1051 	isolate_start_pfn = cc->free_pfn;
1052 	block_start_pfn = pageblock_start_pfn(cc->free_pfn);
1053 	block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1054 						zone_end_pfn(zone));
1055 	low_pfn = pageblock_end_pfn(cc->migrate_pfn);
1056 
1057 	/*
1058 	 * Isolate free pages until enough are available to migrate the
1059 	 * pages on cc->migratepages. We stop searching if the migrate
1060 	 * and free page scanners meet or enough free pages are isolated.
1061 	 */
1062 	for (; block_start_pfn >= low_pfn;
1063 				block_end_pfn = block_start_pfn,
1064 				block_start_pfn -= pageblock_nr_pages,
1065 				isolate_start_pfn = block_start_pfn) {
1066 		/*
1067 		 * This can iterate a massively long zone without finding any
1068 		 * suitable migration targets, so periodically check if we need
1069 		 * to schedule, or even abort async compaction.
1070 		 */
1071 		if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1072 						&& compact_should_abort(cc))
1073 			break;
1074 
1075 		page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1076 									zone);
1077 		if (!page)
1078 			continue;
1079 
1080 		/* Check the block is suitable for migration */
1081 		if (!suitable_migration_target(cc, page))
1082 			continue;
1083 
1084 		/* If isolation recently failed, do not retry */
1085 		if (!isolation_suitable(cc, page))
1086 			continue;
1087 
1088 		/* Found a block suitable for isolating free pages from. */
1089 		isolate_freepages_block(cc, &isolate_start_pfn, block_end_pfn,
1090 					freelist, false);
1091 
1092 		/*
1093 		 * If we isolated enough freepages, or aborted due to lock
1094 		 * contention, terminate.
1095 		 */
1096 		if ((cc->nr_freepages >= cc->nr_migratepages)
1097 							|| cc->contended) {
1098 			if (isolate_start_pfn >= block_end_pfn) {
1099 				/*
1100 				 * Restart at previous pageblock if more
1101 				 * freepages can be isolated next time.
1102 				 */
1103 				isolate_start_pfn =
1104 					block_start_pfn - pageblock_nr_pages;
1105 			}
1106 			break;
1107 		} else if (isolate_start_pfn < block_end_pfn) {
1108 			/*
1109 			 * If isolation failed early, do not continue
1110 			 * needlessly.
1111 			 */
1112 			break;
1113 		}
1114 	}
1115 
1116 	/* __isolate_free_page() does not map the pages */
1117 	map_pages(freelist);
1118 
1119 	/*
1120 	 * Record where the free scanner will restart next time. Either we
1121 	 * broke from the loop and set isolate_start_pfn based on the last
1122 	 * call to isolate_freepages_block(), or we met the migration scanner
1123 	 * and the loop terminated due to isolate_start_pfn < low_pfn
1124 	 */
1125 	cc->free_pfn = isolate_start_pfn;
1126 }
1127 
1128 /*
1129  * This is a migrate-callback that "allocates" freepages by taking pages
1130  * from the isolated freelists in the block we are migrating to.
1131  */
1132 static struct page *compaction_alloc(struct page *migratepage,
1133 					unsigned long data,
1134 					int **result)
1135 {
1136 	struct compact_control *cc = (struct compact_control *)data;
1137 	struct page *freepage;
1138 
1139 	/*
1140 	 * Isolate free pages if necessary, and if we are not aborting due to
1141 	 * contention.
1142 	 */
1143 	if (list_empty(&cc->freepages)) {
1144 		if (!cc->contended)
1145 			isolate_freepages(cc);
1146 
1147 		if (list_empty(&cc->freepages))
1148 			return NULL;
1149 	}
1150 
1151 	freepage = list_entry(cc->freepages.next, struct page, lru);
1152 	list_del(&freepage->lru);
1153 	cc->nr_freepages--;
1154 
1155 	return freepage;
1156 }
1157 
1158 /*
1159  * This is a migrate-callback that "frees" freepages back to the isolated
1160  * freelist.  All pages on the freelist are from the same zone, so there is no
1161  * special handling needed for NUMA.
1162  */
1163 static void compaction_free(struct page *page, unsigned long data)
1164 {
1165 	struct compact_control *cc = (struct compact_control *)data;
1166 
1167 	list_add(&page->lru, &cc->freepages);
1168 	cc->nr_freepages++;
1169 }
1170 
1171 /* possible outcome of isolate_migratepages */
1172 typedef enum {
1173 	ISOLATE_ABORT,		/* Abort compaction now */
1174 	ISOLATE_NONE,		/* No pages isolated, continue scanning */
1175 	ISOLATE_SUCCESS,	/* Pages isolated, migrate */
1176 } isolate_migrate_t;
1177 
1178 /*
1179  * Allow userspace to control policy on scanning the unevictable LRU for
1180  * compactable pages.
1181  */
1182 int sysctl_compact_unevictable_allowed __read_mostly = 1;
1183 
1184 /*
1185  * Isolate all pages that can be migrated from the first suitable block,
1186  * starting at the block pointed to by the migrate scanner pfn within
1187  * compact_control.
1188  */
1189 static isolate_migrate_t isolate_migratepages(struct zone *zone,
1190 					struct compact_control *cc)
1191 {
1192 	unsigned long block_start_pfn;
1193 	unsigned long block_end_pfn;
1194 	unsigned long low_pfn;
1195 	struct page *page;
1196 	const isolate_mode_t isolate_mode =
1197 		(sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1198 		(cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1199 
1200 	/*
1201 	 * Start at where we last stopped, or beginning of the zone as
1202 	 * initialized by compact_zone()
1203 	 */
1204 	low_pfn = cc->migrate_pfn;
1205 	block_start_pfn = pageblock_start_pfn(low_pfn);
1206 	if (block_start_pfn < zone->zone_start_pfn)
1207 		block_start_pfn = zone->zone_start_pfn;
1208 
1209 	/* Only scan within a pageblock boundary */
1210 	block_end_pfn = pageblock_end_pfn(low_pfn);
1211 
1212 	/*
1213 	 * Iterate over whole pageblocks until we find the first suitable.
1214 	 * Do not cross the free scanner.
1215 	 */
1216 	for (; block_end_pfn <= cc->free_pfn;
1217 			low_pfn = block_end_pfn,
1218 			block_start_pfn = block_end_pfn,
1219 			block_end_pfn += pageblock_nr_pages) {
1220 
1221 		/*
1222 		 * This can potentially iterate a massively long zone with
1223 		 * many pageblocks unsuitable, so periodically check if we
1224 		 * need to schedule, or even abort async compaction.
1225 		 */
1226 		if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
1227 						&& compact_should_abort(cc))
1228 			break;
1229 
1230 		page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1231 									zone);
1232 		if (!page)
1233 			continue;
1234 
1235 		/* If isolation recently failed, do not retry */
1236 		if (!isolation_suitable(cc, page))
1237 			continue;
1238 
1239 		/*
1240 		 * For async compaction, also only scan in MOVABLE blocks.
1241 		 * Async compaction is optimistic to see if the minimum amount
1242 		 * of work satisfies the allocation.
1243 		 */
1244 		if (cc->mode == MIGRATE_ASYNC &&
1245 		    !migrate_async_suitable(get_pageblock_migratetype(page)))
1246 			continue;
1247 
1248 		/* Perform the isolation */
1249 		low_pfn = isolate_migratepages_block(cc, low_pfn,
1250 						block_end_pfn, isolate_mode);
1251 
1252 		if (!low_pfn || cc->contended)
1253 			return ISOLATE_ABORT;
1254 
1255 		/*
1256 		 * Either we isolated something and proceed with migration. Or
1257 		 * we failed and compact_zone should decide if we should
1258 		 * continue or not.
1259 		 */
1260 		break;
1261 	}
1262 
1263 	/* Record where migration scanner will be restarted. */
1264 	cc->migrate_pfn = low_pfn;
1265 
1266 	return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
1267 }
1268 
1269 /*
1270  * order == -1 is expected when compacting via
1271  * /proc/sys/vm/compact_memory
1272  */
1273 static inline bool is_via_compact_memory(int order)
1274 {
1275 	return order == -1;
1276 }
1277 
1278 static enum compact_result __compact_finished(struct zone *zone, struct compact_control *cc,
1279 			    const int migratetype)
1280 {
1281 	unsigned int order;
1282 	unsigned long watermark;
1283 
1284 	if (cc->contended || fatal_signal_pending(current))
1285 		return COMPACT_CONTENDED;
1286 
1287 	/* Compaction run completes if the migrate and free scanner meet */
1288 	if (compact_scanners_met(cc)) {
1289 		/* Let the next compaction start anew. */
1290 		reset_cached_positions(zone);
1291 
1292 		/*
1293 		 * Mark that the PG_migrate_skip information should be cleared
1294 		 * by kswapd when it goes to sleep. kcompactd does not set the
1295 		 * flag itself as the decision to be clear should be directly
1296 		 * based on an allocation request.
1297 		 */
1298 		if (cc->direct_compaction)
1299 			zone->compact_blockskip_flush = true;
1300 
1301 		if (cc->whole_zone)
1302 			return COMPACT_COMPLETE;
1303 		else
1304 			return COMPACT_PARTIAL_SKIPPED;
1305 	}
1306 
1307 	if (is_via_compact_memory(cc->order))
1308 		return COMPACT_CONTINUE;
1309 
1310 	/* Compaction run is not finished if the watermark is not met */
1311 	watermark = zone->watermark[cc->alloc_flags & ALLOC_WMARK_MASK];
1312 
1313 	if (!zone_watermark_ok(zone, cc->order, watermark, cc->classzone_idx,
1314 							cc->alloc_flags))
1315 		return COMPACT_CONTINUE;
1316 
1317 	/* Direct compactor: Is a suitable page free? */
1318 	for (order = cc->order; order < MAX_ORDER; order++) {
1319 		struct free_area *area = &zone->free_area[order];
1320 		bool can_steal;
1321 
1322 		/* Job done if page is free of the right migratetype */
1323 		if (!list_empty(&area->free_list[migratetype]))
1324 			return COMPACT_SUCCESS;
1325 
1326 #ifdef CONFIG_CMA
1327 		/* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
1328 		if (migratetype == MIGRATE_MOVABLE &&
1329 			!list_empty(&area->free_list[MIGRATE_CMA]))
1330 			return COMPACT_SUCCESS;
1331 #endif
1332 		/*
1333 		 * Job done if allocation would steal freepages from
1334 		 * other migratetype buddy lists.
1335 		 */
1336 		if (find_suitable_fallback(area, order, migratetype,
1337 						true, &can_steal) != -1)
1338 			return COMPACT_SUCCESS;
1339 	}
1340 
1341 	return COMPACT_NO_SUITABLE_PAGE;
1342 }
1343 
1344 static enum compact_result compact_finished(struct zone *zone,
1345 			struct compact_control *cc,
1346 			const int migratetype)
1347 {
1348 	int ret;
1349 
1350 	ret = __compact_finished(zone, cc, migratetype);
1351 	trace_mm_compaction_finished(zone, cc->order, ret);
1352 	if (ret == COMPACT_NO_SUITABLE_PAGE)
1353 		ret = COMPACT_CONTINUE;
1354 
1355 	return ret;
1356 }
1357 
1358 /*
1359  * compaction_suitable: Is this suitable to run compaction on this zone now?
1360  * Returns
1361  *   COMPACT_SKIPPED  - If there are too few free pages for compaction
1362  *   COMPACT_SUCCESS  - If the allocation would succeed without compaction
1363  *   COMPACT_CONTINUE - If compaction should run now
1364  */
1365 static enum compact_result __compaction_suitable(struct zone *zone, int order,
1366 					unsigned int alloc_flags,
1367 					int classzone_idx,
1368 					unsigned long wmark_target)
1369 {
1370 	unsigned long watermark;
1371 
1372 	if (is_via_compact_memory(order))
1373 		return COMPACT_CONTINUE;
1374 
1375 	watermark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1376 	/*
1377 	 * If watermarks for high-order allocation are already met, there
1378 	 * should be no need for compaction at all.
1379 	 */
1380 	if (zone_watermark_ok(zone, order, watermark, classzone_idx,
1381 								alloc_flags))
1382 		return COMPACT_SUCCESS;
1383 
1384 	/*
1385 	 * Watermarks for order-0 must be met for compaction to be able to
1386 	 * isolate free pages for migration targets. This means that the
1387 	 * watermark and alloc_flags have to match, or be more pessimistic than
1388 	 * the check in __isolate_free_page(). We don't use the direct
1389 	 * compactor's alloc_flags, as they are not relevant for freepage
1390 	 * isolation. We however do use the direct compactor's classzone_idx to
1391 	 * skip over zones where lowmem reserves would prevent allocation even
1392 	 * if compaction succeeds.
1393 	 * For costly orders, we require low watermark instead of min for
1394 	 * compaction to proceed to increase its chances.
1395 	 * ALLOC_CMA is used, as pages in CMA pageblocks are considered
1396 	 * suitable migration targets
1397 	 */
1398 	watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
1399 				low_wmark_pages(zone) : min_wmark_pages(zone);
1400 	watermark += compact_gap(order);
1401 	if (!__zone_watermark_ok(zone, 0, watermark, classzone_idx,
1402 						ALLOC_CMA, wmark_target))
1403 		return COMPACT_SKIPPED;
1404 
1405 	return COMPACT_CONTINUE;
1406 }
1407 
1408 enum compact_result compaction_suitable(struct zone *zone, int order,
1409 					unsigned int alloc_flags,
1410 					int classzone_idx)
1411 {
1412 	enum compact_result ret;
1413 	int fragindex;
1414 
1415 	ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx,
1416 				    zone_page_state(zone, NR_FREE_PAGES));
1417 	/*
1418 	 * fragmentation index determines if allocation failures are due to
1419 	 * low memory or external fragmentation
1420 	 *
1421 	 * index of -1000 would imply allocations might succeed depending on
1422 	 * watermarks, but we already failed the high-order watermark check
1423 	 * index towards 0 implies failure is due to lack of memory
1424 	 * index towards 1000 implies failure is due to fragmentation
1425 	 *
1426 	 * Only compact if a failure would be due to fragmentation. Also
1427 	 * ignore fragindex for non-costly orders where the alternative to
1428 	 * a successful reclaim/compaction is OOM. Fragindex and the
1429 	 * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
1430 	 * excessive compaction for costly orders, but it should not be at the
1431 	 * expense of system stability.
1432 	 */
1433 	if (ret == COMPACT_CONTINUE && (order > PAGE_ALLOC_COSTLY_ORDER)) {
1434 		fragindex = fragmentation_index(zone, order);
1435 		if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
1436 			ret = COMPACT_NOT_SUITABLE_ZONE;
1437 	}
1438 
1439 	trace_mm_compaction_suitable(zone, order, ret);
1440 	if (ret == COMPACT_NOT_SUITABLE_ZONE)
1441 		ret = COMPACT_SKIPPED;
1442 
1443 	return ret;
1444 }
1445 
1446 bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
1447 		int alloc_flags)
1448 {
1449 	struct zone *zone;
1450 	struct zoneref *z;
1451 
1452 	/*
1453 	 * Make sure at least one zone would pass __compaction_suitable if we continue
1454 	 * retrying the reclaim.
1455 	 */
1456 	for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1457 					ac->nodemask) {
1458 		unsigned long available;
1459 		enum compact_result compact_result;
1460 
1461 		/*
1462 		 * Do not consider all the reclaimable memory because we do not
1463 		 * want to trash just for a single high order allocation which
1464 		 * is even not guaranteed to appear even if __compaction_suitable
1465 		 * is happy about the watermark check.
1466 		 */
1467 		available = zone_reclaimable_pages(zone) / order;
1468 		available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
1469 		compact_result = __compaction_suitable(zone, order, alloc_flags,
1470 				ac_classzone_idx(ac), available);
1471 		if (compact_result != COMPACT_SKIPPED)
1472 			return true;
1473 	}
1474 
1475 	return false;
1476 }
1477 
1478 static enum compact_result compact_zone(struct zone *zone, struct compact_control *cc)
1479 {
1480 	enum compact_result ret;
1481 	unsigned long start_pfn = zone->zone_start_pfn;
1482 	unsigned long end_pfn = zone_end_pfn(zone);
1483 	const int migratetype = gfpflags_to_migratetype(cc->gfp_mask);
1484 	const bool sync = cc->mode != MIGRATE_ASYNC;
1485 
1486 	ret = compaction_suitable(zone, cc->order, cc->alloc_flags,
1487 							cc->classzone_idx);
1488 	/* Compaction is likely to fail */
1489 	if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED)
1490 		return ret;
1491 
1492 	/* huh, compaction_suitable is returning something unexpected */
1493 	VM_BUG_ON(ret != COMPACT_CONTINUE);
1494 
1495 	/*
1496 	 * Clear pageblock skip if there were failures recently and compaction
1497 	 * is about to be retried after being deferred.
1498 	 */
1499 	if (compaction_restarting(zone, cc->order))
1500 		__reset_isolation_suitable(zone);
1501 
1502 	/*
1503 	 * Setup to move all movable pages to the end of the zone. Used cached
1504 	 * information on where the scanners should start (unless we explicitly
1505 	 * want to compact the whole zone), but check that it is initialised
1506 	 * by ensuring the values are within zone boundaries.
1507 	 */
1508 	if (cc->whole_zone) {
1509 		cc->migrate_pfn = start_pfn;
1510 		cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
1511 	} else {
1512 		cc->migrate_pfn = zone->compact_cached_migrate_pfn[sync];
1513 		cc->free_pfn = zone->compact_cached_free_pfn;
1514 		if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
1515 			cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
1516 			zone->compact_cached_free_pfn = cc->free_pfn;
1517 		}
1518 		if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
1519 			cc->migrate_pfn = start_pfn;
1520 			zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
1521 			zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
1522 		}
1523 
1524 		if (cc->migrate_pfn == start_pfn)
1525 			cc->whole_zone = true;
1526 	}
1527 
1528 	cc->last_migrated_pfn = 0;
1529 
1530 	trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
1531 				cc->free_pfn, end_pfn, sync);
1532 
1533 	migrate_prep_local();
1534 
1535 	while ((ret = compact_finished(zone, cc, migratetype)) ==
1536 						COMPACT_CONTINUE) {
1537 		int err;
1538 
1539 		switch (isolate_migratepages(zone, cc)) {
1540 		case ISOLATE_ABORT:
1541 			ret = COMPACT_CONTENDED;
1542 			putback_movable_pages(&cc->migratepages);
1543 			cc->nr_migratepages = 0;
1544 			goto out;
1545 		case ISOLATE_NONE:
1546 			/*
1547 			 * We haven't isolated and migrated anything, but
1548 			 * there might still be unflushed migrations from
1549 			 * previous cc->order aligned block.
1550 			 */
1551 			goto check_drain;
1552 		case ISOLATE_SUCCESS:
1553 			;
1554 		}
1555 
1556 		err = migrate_pages(&cc->migratepages, compaction_alloc,
1557 				compaction_free, (unsigned long)cc, cc->mode,
1558 				MR_COMPACTION);
1559 
1560 		trace_mm_compaction_migratepages(cc->nr_migratepages, err,
1561 							&cc->migratepages);
1562 
1563 		/* All pages were either migrated or will be released */
1564 		cc->nr_migratepages = 0;
1565 		if (err) {
1566 			putback_movable_pages(&cc->migratepages);
1567 			/*
1568 			 * migrate_pages() may return -ENOMEM when scanners meet
1569 			 * and we want compact_finished() to detect it
1570 			 */
1571 			if (err == -ENOMEM && !compact_scanners_met(cc)) {
1572 				ret = COMPACT_CONTENDED;
1573 				goto out;
1574 			}
1575 			/*
1576 			 * We failed to migrate at least one page in the current
1577 			 * order-aligned block, so skip the rest of it.
1578 			 */
1579 			if (cc->direct_compaction &&
1580 						(cc->mode == MIGRATE_ASYNC)) {
1581 				cc->migrate_pfn = block_end_pfn(
1582 						cc->migrate_pfn - 1, cc->order);
1583 				/* Draining pcplists is useless in this case */
1584 				cc->last_migrated_pfn = 0;
1585 
1586 			}
1587 		}
1588 
1589 check_drain:
1590 		/*
1591 		 * Has the migration scanner moved away from the previous
1592 		 * cc->order aligned block where we migrated from? If yes,
1593 		 * flush the pages that were freed, so that they can merge and
1594 		 * compact_finished() can detect immediately if allocation
1595 		 * would succeed.
1596 		 */
1597 		if (cc->order > 0 && cc->last_migrated_pfn) {
1598 			int cpu;
1599 			unsigned long current_block_start =
1600 				block_start_pfn(cc->migrate_pfn, cc->order);
1601 
1602 			if (cc->last_migrated_pfn < current_block_start) {
1603 				cpu = get_cpu();
1604 				lru_add_drain_cpu(cpu);
1605 				drain_local_pages(zone);
1606 				put_cpu();
1607 				/* No more flushing until we migrate again */
1608 				cc->last_migrated_pfn = 0;
1609 			}
1610 		}
1611 
1612 	}
1613 
1614 out:
1615 	/*
1616 	 * Release free pages and update where the free scanner should restart,
1617 	 * so we don't leave any returned pages behind in the next attempt.
1618 	 */
1619 	if (cc->nr_freepages > 0) {
1620 		unsigned long free_pfn = release_freepages(&cc->freepages);
1621 
1622 		cc->nr_freepages = 0;
1623 		VM_BUG_ON(free_pfn == 0);
1624 		/* The cached pfn is always the first in a pageblock */
1625 		free_pfn = pageblock_start_pfn(free_pfn);
1626 		/*
1627 		 * Only go back, not forward. The cached pfn might have been
1628 		 * already reset to zone end in compact_finished()
1629 		 */
1630 		if (free_pfn > zone->compact_cached_free_pfn)
1631 			zone->compact_cached_free_pfn = free_pfn;
1632 	}
1633 
1634 	trace_mm_compaction_end(start_pfn, cc->migrate_pfn,
1635 				cc->free_pfn, end_pfn, sync, ret);
1636 
1637 	return ret;
1638 }
1639 
1640 static enum compact_result compact_zone_order(struct zone *zone, int order,
1641 		gfp_t gfp_mask, enum compact_priority prio,
1642 		unsigned int alloc_flags, int classzone_idx)
1643 {
1644 	enum compact_result ret;
1645 	struct compact_control cc = {
1646 		.nr_freepages = 0,
1647 		.nr_migratepages = 0,
1648 		.order = order,
1649 		.gfp_mask = gfp_mask,
1650 		.zone = zone,
1651 		.mode = (prio == COMPACT_PRIO_ASYNC) ?
1652 					MIGRATE_ASYNC :	MIGRATE_SYNC_LIGHT,
1653 		.alloc_flags = alloc_flags,
1654 		.classzone_idx = classzone_idx,
1655 		.direct_compaction = true,
1656 		.whole_zone = (prio == MIN_COMPACT_PRIORITY),
1657 		.ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
1658 		.ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
1659 	};
1660 	INIT_LIST_HEAD(&cc.freepages);
1661 	INIT_LIST_HEAD(&cc.migratepages);
1662 
1663 	ret = compact_zone(zone, &cc);
1664 
1665 	VM_BUG_ON(!list_empty(&cc.freepages));
1666 	VM_BUG_ON(!list_empty(&cc.migratepages));
1667 
1668 	return ret;
1669 }
1670 
1671 int sysctl_extfrag_threshold = 500;
1672 
1673 /**
1674  * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1675  * @gfp_mask: The GFP mask of the current allocation
1676  * @order: The order of the current allocation
1677  * @alloc_flags: The allocation flags of the current allocation
1678  * @ac: The context of current allocation
1679  * @mode: The migration mode for async, sync light, or sync migration
1680  *
1681  * This is the main entry point for direct page compaction.
1682  */
1683 enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
1684 		unsigned int alloc_flags, const struct alloc_context *ac,
1685 		enum compact_priority prio)
1686 {
1687 	int may_perform_io = gfp_mask & __GFP_IO;
1688 	struct zoneref *z;
1689 	struct zone *zone;
1690 	enum compact_result rc = COMPACT_SKIPPED;
1691 
1692 	/*
1693 	 * Check if the GFP flags allow compaction - GFP_NOIO is really
1694 	 * tricky context because the migration might require IO
1695 	 */
1696 	if (!may_perform_io)
1697 		return COMPACT_SKIPPED;
1698 
1699 	trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
1700 
1701 	/* Compact each zone in the list */
1702 	for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
1703 								ac->nodemask) {
1704 		enum compact_result status;
1705 
1706 		if (prio > MIN_COMPACT_PRIORITY
1707 					&& compaction_deferred(zone, order)) {
1708 			rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
1709 			continue;
1710 		}
1711 
1712 		status = compact_zone_order(zone, order, gfp_mask, prio,
1713 					alloc_flags, ac_classzone_idx(ac));
1714 		rc = max(status, rc);
1715 
1716 		/* The allocation should succeed, stop compacting */
1717 		if (status == COMPACT_SUCCESS) {
1718 			/*
1719 			 * We think the allocation will succeed in this zone,
1720 			 * but it is not certain, hence the false. The caller
1721 			 * will repeat this with true if allocation indeed
1722 			 * succeeds in this zone.
1723 			 */
1724 			compaction_defer_reset(zone, order, false);
1725 
1726 			break;
1727 		}
1728 
1729 		if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
1730 					status == COMPACT_PARTIAL_SKIPPED))
1731 			/*
1732 			 * We think that allocation won't succeed in this zone
1733 			 * so we defer compaction there. If it ends up
1734 			 * succeeding after all, it will be reset.
1735 			 */
1736 			defer_compaction(zone, order);
1737 
1738 		/*
1739 		 * We might have stopped compacting due to need_resched() in
1740 		 * async compaction, or due to a fatal signal detected. In that
1741 		 * case do not try further zones
1742 		 */
1743 		if ((prio == COMPACT_PRIO_ASYNC && need_resched())
1744 					|| fatal_signal_pending(current))
1745 			break;
1746 	}
1747 
1748 	return rc;
1749 }
1750 
1751 
1752 /* Compact all zones within a node */
1753 static void compact_node(int nid)
1754 {
1755 	pg_data_t *pgdat = NODE_DATA(nid);
1756 	int zoneid;
1757 	struct zone *zone;
1758 	struct compact_control cc = {
1759 		.order = -1,
1760 		.mode = MIGRATE_SYNC,
1761 		.ignore_skip_hint = true,
1762 		.whole_zone = true,
1763 		.gfp_mask = GFP_KERNEL,
1764 	};
1765 
1766 
1767 	for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1768 
1769 		zone = &pgdat->node_zones[zoneid];
1770 		if (!populated_zone(zone))
1771 			continue;
1772 
1773 		cc.nr_freepages = 0;
1774 		cc.nr_migratepages = 0;
1775 		cc.zone = zone;
1776 		INIT_LIST_HEAD(&cc.freepages);
1777 		INIT_LIST_HEAD(&cc.migratepages);
1778 
1779 		compact_zone(zone, &cc);
1780 
1781 		VM_BUG_ON(!list_empty(&cc.freepages));
1782 		VM_BUG_ON(!list_empty(&cc.migratepages));
1783 	}
1784 }
1785 
1786 /* Compact all nodes in the system */
1787 static void compact_nodes(void)
1788 {
1789 	int nid;
1790 
1791 	/* Flush pending updates to the LRU lists */
1792 	lru_add_drain_all();
1793 
1794 	for_each_online_node(nid)
1795 		compact_node(nid);
1796 }
1797 
1798 /* The written value is actually unused, all memory is compacted */
1799 int sysctl_compact_memory;
1800 
1801 /*
1802  * This is the entry point for compacting all nodes via
1803  * /proc/sys/vm/compact_memory
1804  */
1805 int sysctl_compaction_handler(struct ctl_table *table, int write,
1806 			void __user *buffer, size_t *length, loff_t *ppos)
1807 {
1808 	if (write)
1809 		compact_nodes();
1810 
1811 	return 0;
1812 }
1813 
1814 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1815 			void __user *buffer, size_t *length, loff_t *ppos)
1816 {
1817 	proc_dointvec_minmax(table, write, buffer, length, ppos);
1818 
1819 	return 0;
1820 }
1821 
1822 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1823 static ssize_t sysfs_compact_node(struct device *dev,
1824 			struct device_attribute *attr,
1825 			const char *buf, size_t count)
1826 {
1827 	int nid = dev->id;
1828 
1829 	if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1830 		/* Flush pending updates to the LRU lists */
1831 		lru_add_drain_all();
1832 
1833 		compact_node(nid);
1834 	}
1835 
1836 	return count;
1837 }
1838 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1839 
1840 int compaction_register_node(struct node *node)
1841 {
1842 	return device_create_file(&node->dev, &dev_attr_compact);
1843 }
1844 
1845 void compaction_unregister_node(struct node *node)
1846 {
1847 	return device_remove_file(&node->dev, &dev_attr_compact);
1848 }
1849 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1850 
1851 static inline bool kcompactd_work_requested(pg_data_t *pgdat)
1852 {
1853 	return pgdat->kcompactd_max_order > 0 || kthread_should_stop();
1854 }
1855 
1856 static bool kcompactd_node_suitable(pg_data_t *pgdat)
1857 {
1858 	int zoneid;
1859 	struct zone *zone;
1860 	enum zone_type classzone_idx = pgdat->kcompactd_classzone_idx;
1861 
1862 	for (zoneid = 0; zoneid <= classzone_idx; zoneid++) {
1863 		zone = &pgdat->node_zones[zoneid];
1864 
1865 		if (!populated_zone(zone))
1866 			continue;
1867 
1868 		if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0,
1869 					classzone_idx) == COMPACT_CONTINUE)
1870 			return true;
1871 	}
1872 
1873 	return false;
1874 }
1875 
1876 static void kcompactd_do_work(pg_data_t *pgdat)
1877 {
1878 	/*
1879 	 * With no special task, compact all zones so that a page of requested
1880 	 * order is allocatable.
1881 	 */
1882 	int zoneid;
1883 	struct zone *zone;
1884 	struct compact_control cc = {
1885 		.order = pgdat->kcompactd_max_order,
1886 		.classzone_idx = pgdat->kcompactd_classzone_idx,
1887 		.mode = MIGRATE_SYNC_LIGHT,
1888 		.ignore_skip_hint = true,
1889 		.gfp_mask = GFP_KERNEL,
1890 
1891 	};
1892 	trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
1893 							cc.classzone_idx);
1894 	count_vm_event(KCOMPACTD_WAKE);
1895 
1896 	for (zoneid = 0; zoneid <= cc.classzone_idx; zoneid++) {
1897 		int status;
1898 
1899 		zone = &pgdat->node_zones[zoneid];
1900 		if (!populated_zone(zone))
1901 			continue;
1902 
1903 		if (compaction_deferred(zone, cc.order))
1904 			continue;
1905 
1906 		if (compaction_suitable(zone, cc.order, 0, zoneid) !=
1907 							COMPACT_CONTINUE)
1908 			continue;
1909 
1910 		cc.nr_freepages = 0;
1911 		cc.nr_migratepages = 0;
1912 		cc.zone = zone;
1913 		INIT_LIST_HEAD(&cc.freepages);
1914 		INIT_LIST_HEAD(&cc.migratepages);
1915 
1916 		if (kthread_should_stop())
1917 			return;
1918 		status = compact_zone(zone, &cc);
1919 
1920 		if (status == COMPACT_SUCCESS) {
1921 			compaction_defer_reset(zone, cc.order, false);
1922 		} else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
1923 			/*
1924 			 * We use sync migration mode here, so we defer like
1925 			 * sync direct compaction does.
1926 			 */
1927 			defer_compaction(zone, cc.order);
1928 		}
1929 
1930 		VM_BUG_ON(!list_empty(&cc.freepages));
1931 		VM_BUG_ON(!list_empty(&cc.migratepages));
1932 	}
1933 
1934 	/*
1935 	 * Regardless of success, we are done until woken up next. But remember
1936 	 * the requested order/classzone_idx in case it was higher/tighter than
1937 	 * our current ones
1938 	 */
1939 	if (pgdat->kcompactd_max_order <= cc.order)
1940 		pgdat->kcompactd_max_order = 0;
1941 	if (pgdat->kcompactd_classzone_idx >= cc.classzone_idx)
1942 		pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
1943 }
1944 
1945 void wakeup_kcompactd(pg_data_t *pgdat, int order, int classzone_idx)
1946 {
1947 	if (!order)
1948 		return;
1949 
1950 	if (pgdat->kcompactd_max_order < order)
1951 		pgdat->kcompactd_max_order = order;
1952 
1953 	if (pgdat->kcompactd_classzone_idx > classzone_idx)
1954 		pgdat->kcompactd_classzone_idx = classzone_idx;
1955 
1956 	if (!waitqueue_active(&pgdat->kcompactd_wait))
1957 		return;
1958 
1959 	if (!kcompactd_node_suitable(pgdat))
1960 		return;
1961 
1962 	trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
1963 							classzone_idx);
1964 	wake_up_interruptible(&pgdat->kcompactd_wait);
1965 }
1966 
1967 /*
1968  * The background compaction daemon, started as a kernel thread
1969  * from the init process.
1970  */
1971 static int kcompactd(void *p)
1972 {
1973 	pg_data_t *pgdat = (pg_data_t*)p;
1974 	struct task_struct *tsk = current;
1975 
1976 	const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1977 
1978 	if (!cpumask_empty(cpumask))
1979 		set_cpus_allowed_ptr(tsk, cpumask);
1980 
1981 	set_freezable();
1982 
1983 	pgdat->kcompactd_max_order = 0;
1984 	pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
1985 
1986 	while (!kthread_should_stop()) {
1987 		trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
1988 		wait_event_freezable(pgdat->kcompactd_wait,
1989 				kcompactd_work_requested(pgdat));
1990 
1991 		kcompactd_do_work(pgdat);
1992 	}
1993 
1994 	return 0;
1995 }
1996 
1997 /*
1998  * This kcompactd start function will be called by init and node-hot-add.
1999  * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
2000  */
2001 int kcompactd_run(int nid)
2002 {
2003 	pg_data_t *pgdat = NODE_DATA(nid);
2004 	int ret = 0;
2005 
2006 	if (pgdat->kcompactd)
2007 		return 0;
2008 
2009 	pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
2010 	if (IS_ERR(pgdat->kcompactd)) {
2011 		pr_err("Failed to start kcompactd on node %d\n", nid);
2012 		ret = PTR_ERR(pgdat->kcompactd);
2013 		pgdat->kcompactd = NULL;
2014 	}
2015 	return ret;
2016 }
2017 
2018 /*
2019  * Called by memory hotplug when all memory in a node is offlined. Caller must
2020  * hold mem_hotplug_begin/end().
2021  */
2022 void kcompactd_stop(int nid)
2023 {
2024 	struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
2025 
2026 	if (kcompactd) {
2027 		kthread_stop(kcompactd);
2028 		NODE_DATA(nid)->kcompactd = NULL;
2029 	}
2030 }
2031 
2032 /*
2033  * It's optimal to keep kcompactd on the same CPUs as their memory, but
2034  * not required for correctness. So if the last cpu in a node goes
2035  * away, we get changed to run anywhere: as the first one comes back,
2036  * restore their cpu bindings.
2037  */
2038 static int kcompactd_cpu_online(unsigned int cpu)
2039 {
2040 	int nid;
2041 
2042 	for_each_node_state(nid, N_MEMORY) {
2043 		pg_data_t *pgdat = NODE_DATA(nid);
2044 		const struct cpumask *mask;
2045 
2046 		mask = cpumask_of_node(pgdat->node_id);
2047 
2048 		if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2049 			/* One of our CPUs online: restore mask */
2050 			set_cpus_allowed_ptr(pgdat->kcompactd, mask);
2051 	}
2052 	return 0;
2053 }
2054 
2055 static int __init kcompactd_init(void)
2056 {
2057 	int nid;
2058 	int ret;
2059 
2060 	ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
2061 					"mm/compaction:online",
2062 					kcompactd_cpu_online, NULL);
2063 	if (ret < 0) {
2064 		pr_err("kcompactd: failed to register hotplug callbacks.\n");
2065 		return ret;
2066 	}
2067 
2068 	for_each_node_state(nid, N_MEMORY)
2069 		kcompactd_run(nid);
2070 	return 0;
2071 }
2072 subsys_initcall(kcompactd_init)
2073 
2074 #endif /* CONFIG_COMPACTION */
2075