xref: /openbmc/linux/mm/compaction.c (revision 930beb5a)
1 /*
2  * linux/mm/compaction.c
3  *
4  * Memory compaction for the reduction of external fragmentation. Note that
5  * this heavily depends upon page migration to do all the real heavy
6  * lifting
7  *
8  * Copyright IBM Corp. 2007-2010 Mel Gorman <mel@csn.ul.ie>
9  */
10 #include <linux/swap.h>
11 #include <linux/migrate.h>
12 #include <linux/compaction.h>
13 #include <linux/mm_inline.h>
14 #include <linux/backing-dev.h>
15 #include <linux/sysctl.h>
16 #include <linux/sysfs.h>
17 #include <linux/balloon_compaction.h>
18 #include <linux/page-isolation.h>
19 #include "internal.h"
20 
21 #ifdef CONFIG_COMPACTION
22 static inline void count_compact_event(enum vm_event_item item)
23 {
24 	count_vm_event(item);
25 }
26 
27 static inline void count_compact_events(enum vm_event_item item, long delta)
28 {
29 	count_vm_events(item, delta);
30 }
31 #else
32 #define count_compact_event(item) do { } while (0)
33 #define count_compact_events(item, delta) do { } while (0)
34 #endif
35 
36 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
37 
38 #define CREATE_TRACE_POINTS
39 #include <trace/events/compaction.h>
40 
41 static unsigned long release_freepages(struct list_head *freelist)
42 {
43 	struct page *page, *next;
44 	unsigned long count = 0;
45 
46 	list_for_each_entry_safe(page, next, freelist, lru) {
47 		list_del(&page->lru);
48 		__free_page(page);
49 		count++;
50 	}
51 
52 	return count;
53 }
54 
55 static void map_pages(struct list_head *list)
56 {
57 	struct page *page;
58 
59 	list_for_each_entry(page, list, lru) {
60 		arch_alloc_page(page, 0);
61 		kernel_map_pages(page, 1, 1);
62 	}
63 }
64 
65 static inline bool migrate_async_suitable(int migratetype)
66 {
67 	return is_migrate_cma(migratetype) || migratetype == MIGRATE_MOVABLE;
68 }
69 
70 #ifdef CONFIG_COMPACTION
71 /* Returns true if the pageblock should be scanned for pages to isolate. */
72 static inline bool isolation_suitable(struct compact_control *cc,
73 					struct page *page)
74 {
75 	if (cc->ignore_skip_hint)
76 		return true;
77 
78 	return !get_pageblock_skip(page);
79 }
80 
81 /*
82  * This function is called to clear all cached information on pageblocks that
83  * should be skipped for page isolation when the migrate and free page scanner
84  * meet.
85  */
86 static void __reset_isolation_suitable(struct zone *zone)
87 {
88 	unsigned long start_pfn = zone->zone_start_pfn;
89 	unsigned long end_pfn = zone_end_pfn(zone);
90 	unsigned long pfn;
91 
92 	zone->compact_cached_migrate_pfn = start_pfn;
93 	zone->compact_cached_free_pfn = end_pfn;
94 	zone->compact_blockskip_flush = false;
95 
96 	/* Walk the zone and mark every pageblock as suitable for isolation */
97 	for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
98 		struct page *page;
99 
100 		cond_resched();
101 
102 		if (!pfn_valid(pfn))
103 			continue;
104 
105 		page = pfn_to_page(pfn);
106 		if (zone != page_zone(page))
107 			continue;
108 
109 		clear_pageblock_skip(page);
110 	}
111 }
112 
113 void reset_isolation_suitable(pg_data_t *pgdat)
114 {
115 	int zoneid;
116 
117 	for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
118 		struct zone *zone = &pgdat->node_zones[zoneid];
119 		if (!populated_zone(zone))
120 			continue;
121 
122 		/* Only flush if a full compaction finished recently */
123 		if (zone->compact_blockskip_flush)
124 			__reset_isolation_suitable(zone);
125 	}
126 }
127 
128 /*
129  * If no pages were isolated then mark this pageblock to be skipped in the
130  * future. The information is later cleared by __reset_isolation_suitable().
131  */
132 static void update_pageblock_skip(struct compact_control *cc,
133 			struct page *page, unsigned long nr_isolated,
134 			bool migrate_scanner)
135 {
136 	struct zone *zone = cc->zone;
137 
138 	if (cc->ignore_skip_hint)
139 		return;
140 
141 	if (!page)
142 		return;
143 
144 	if (!nr_isolated) {
145 		unsigned long pfn = page_to_pfn(page);
146 		set_pageblock_skip(page);
147 
148 		/* Update where compaction should restart */
149 		if (migrate_scanner) {
150 			if (!cc->finished_update_migrate &&
151 			    pfn > zone->compact_cached_migrate_pfn)
152 				zone->compact_cached_migrate_pfn = pfn;
153 		} else {
154 			if (!cc->finished_update_free &&
155 			    pfn < zone->compact_cached_free_pfn)
156 				zone->compact_cached_free_pfn = pfn;
157 		}
158 	}
159 }
160 #else
161 static inline bool isolation_suitable(struct compact_control *cc,
162 					struct page *page)
163 {
164 	return true;
165 }
166 
167 static void update_pageblock_skip(struct compact_control *cc,
168 			struct page *page, unsigned long nr_isolated,
169 			bool migrate_scanner)
170 {
171 }
172 #endif /* CONFIG_COMPACTION */
173 
174 static inline bool should_release_lock(spinlock_t *lock)
175 {
176 	return need_resched() || spin_is_contended(lock);
177 }
178 
179 /*
180  * Compaction requires the taking of some coarse locks that are potentially
181  * very heavily contended. Check if the process needs to be scheduled or
182  * if the lock is contended. For async compaction, back out in the event
183  * if contention is severe. For sync compaction, schedule.
184  *
185  * Returns true if the lock is held.
186  * Returns false if the lock is released and compaction should abort
187  */
188 static bool compact_checklock_irqsave(spinlock_t *lock, unsigned long *flags,
189 				      bool locked, struct compact_control *cc)
190 {
191 	if (should_release_lock(lock)) {
192 		if (locked) {
193 			spin_unlock_irqrestore(lock, *flags);
194 			locked = false;
195 		}
196 
197 		/* async aborts if taking too long or contended */
198 		if (!cc->sync) {
199 			cc->contended = true;
200 			return false;
201 		}
202 
203 		cond_resched();
204 	}
205 
206 	if (!locked)
207 		spin_lock_irqsave(lock, *flags);
208 	return true;
209 }
210 
211 static inline bool compact_trylock_irqsave(spinlock_t *lock,
212 			unsigned long *flags, struct compact_control *cc)
213 {
214 	return compact_checklock_irqsave(lock, flags, false, cc);
215 }
216 
217 /* Returns true if the page is within a block suitable for migration to */
218 static bool suitable_migration_target(struct page *page)
219 {
220 	int migratetype = get_pageblock_migratetype(page);
221 
222 	/* Don't interfere with memory hot-remove or the min_free_kbytes blocks */
223 	if (migratetype == MIGRATE_RESERVE)
224 		return false;
225 
226 	if (is_migrate_isolate(migratetype))
227 		return false;
228 
229 	/* If the page is a large free page, then allow migration */
230 	if (PageBuddy(page) && page_order(page) >= pageblock_order)
231 		return true;
232 
233 	/* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
234 	if (migrate_async_suitable(migratetype))
235 		return true;
236 
237 	/* Otherwise skip the block */
238 	return false;
239 }
240 
241 /*
242  * Isolate free pages onto a private freelist. If @strict is true, will abort
243  * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
244  * (even though it may still end up isolating some pages).
245  */
246 static unsigned long isolate_freepages_block(struct compact_control *cc,
247 				unsigned long blockpfn,
248 				unsigned long end_pfn,
249 				struct list_head *freelist,
250 				bool strict)
251 {
252 	int nr_scanned = 0, total_isolated = 0;
253 	struct page *cursor, *valid_page = NULL;
254 	unsigned long nr_strict_required = end_pfn - blockpfn;
255 	unsigned long flags;
256 	bool locked = false;
257 
258 	cursor = pfn_to_page(blockpfn);
259 
260 	/* Isolate free pages. */
261 	for (; blockpfn < end_pfn; blockpfn++, cursor++) {
262 		int isolated, i;
263 		struct page *page = cursor;
264 
265 		nr_scanned++;
266 		if (!pfn_valid_within(blockpfn))
267 			continue;
268 		if (!valid_page)
269 			valid_page = page;
270 		if (!PageBuddy(page))
271 			continue;
272 
273 		/*
274 		 * The zone lock must be held to isolate freepages.
275 		 * Unfortunately this is a very coarse lock and can be
276 		 * heavily contended if there are parallel allocations
277 		 * or parallel compactions. For async compaction do not
278 		 * spin on the lock and we acquire the lock as late as
279 		 * possible.
280 		 */
281 		locked = compact_checklock_irqsave(&cc->zone->lock, &flags,
282 								locked, cc);
283 		if (!locked)
284 			break;
285 
286 		/* Recheck this is a suitable migration target under lock */
287 		if (!strict && !suitable_migration_target(page))
288 			break;
289 
290 		/* Recheck this is a buddy page under lock */
291 		if (!PageBuddy(page))
292 			continue;
293 
294 		/* Found a free page, break it into order-0 pages */
295 		isolated = split_free_page(page);
296 		if (!isolated && strict)
297 			break;
298 		total_isolated += isolated;
299 		for (i = 0; i < isolated; i++) {
300 			list_add(&page->lru, freelist);
301 			page++;
302 		}
303 
304 		/* If a page was split, advance to the end of it */
305 		if (isolated) {
306 			blockpfn += isolated - 1;
307 			cursor += isolated - 1;
308 		}
309 	}
310 
311 	trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
312 
313 	/*
314 	 * If strict isolation is requested by CMA then check that all the
315 	 * pages requested were isolated. If there were any failures, 0 is
316 	 * returned and CMA will fail.
317 	 */
318 	if (strict && nr_strict_required > total_isolated)
319 		total_isolated = 0;
320 
321 	if (locked)
322 		spin_unlock_irqrestore(&cc->zone->lock, flags);
323 
324 	/* Update the pageblock-skip if the whole pageblock was scanned */
325 	if (blockpfn == end_pfn)
326 		update_pageblock_skip(cc, valid_page, total_isolated, false);
327 
328 	count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
329 	if (total_isolated)
330 		count_compact_events(COMPACTISOLATED, total_isolated);
331 	return total_isolated;
332 }
333 
334 /**
335  * isolate_freepages_range() - isolate free pages.
336  * @start_pfn: The first PFN to start isolating.
337  * @end_pfn:   The one-past-last PFN.
338  *
339  * Non-free pages, invalid PFNs, or zone boundaries within the
340  * [start_pfn, end_pfn) range are considered errors, cause function to
341  * undo its actions and return zero.
342  *
343  * Otherwise, function returns one-past-the-last PFN of isolated page
344  * (which may be greater then end_pfn if end fell in a middle of
345  * a free page).
346  */
347 unsigned long
348 isolate_freepages_range(struct compact_control *cc,
349 			unsigned long start_pfn, unsigned long end_pfn)
350 {
351 	unsigned long isolated, pfn, block_end_pfn;
352 	LIST_HEAD(freelist);
353 
354 	for (pfn = start_pfn; pfn < end_pfn; pfn += isolated) {
355 		if (!pfn_valid(pfn) || cc->zone != page_zone(pfn_to_page(pfn)))
356 			break;
357 
358 		/*
359 		 * On subsequent iterations ALIGN() is actually not needed,
360 		 * but we keep it that we not to complicate the code.
361 		 */
362 		block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
363 		block_end_pfn = min(block_end_pfn, end_pfn);
364 
365 		isolated = isolate_freepages_block(cc, pfn, block_end_pfn,
366 						   &freelist, true);
367 
368 		/*
369 		 * In strict mode, isolate_freepages_block() returns 0 if
370 		 * there are any holes in the block (ie. invalid PFNs or
371 		 * non-free pages).
372 		 */
373 		if (!isolated)
374 			break;
375 
376 		/*
377 		 * If we managed to isolate pages, it is always (1 << n) *
378 		 * pageblock_nr_pages for some non-negative n.  (Max order
379 		 * page may span two pageblocks).
380 		 */
381 	}
382 
383 	/* split_free_page does not map the pages */
384 	map_pages(&freelist);
385 
386 	if (pfn < end_pfn) {
387 		/* Loop terminated early, cleanup. */
388 		release_freepages(&freelist);
389 		return 0;
390 	}
391 
392 	/* We don't use freelists for anything. */
393 	return pfn;
394 }
395 
396 /* Update the number of anon and file isolated pages in the zone */
397 static void acct_isolated(struct zone *zone, bool locked, struct compact_control *cc)
398 {
399 	struct page *page;
400 	unsigned int count[2] = { 0, };
401 
402 	list_for_each_entry(page, &cc->migratepages, lru)
403 		count[!!page_is_file_cache(page)]++;
404 
405 	/* If locked we can use the interrupt unsafe versions */
406 	if (locked) {
407 		__mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
408 		__mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
409 	} else {
410 		mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
411 		mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
412 	}
413 }
414 
415 /* Similar to reclaim, but different enough that they don't share logic */
416 static bool too_many_isolated(struct zone *zone)
417 {
418 	unsigned long active, inactive, isolated;
419 
420 	inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
421 					zone_page_state(zone, NR_INACTIVE_ANON);
422 	active = zone_page_state(zone, NR_ACTIVE_FILE) +
423 					zone_page_state(zone, NR_ACTIVE_ANON);
424 	isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
425 					zone_page_state(zone, NR_ISOLATED_ANON);
426 
427 	return isolated > (inactive + active) / 2;
428 }
429 
430 /**
431  * isolate_migratepages_range() - isolate all migrate-able pages in range.
432  * @zone:	Zone pages are in.
433  * @cc:		Compaction control structure.
434  * @low_pfn:	The first PFN of the range.
435  * @end_pfn:	The one-past-the-last PFN of the range.
436  * @unevictable: true if it allows to isolate unevictable pages
437  *
438  * Isolate all pages that can be migrated from the range specified by
439  * [low_pfn, end_pfn).  Returns zero if there is a fatal signal
440  * pending), otherwise PFN of the first page that was not scanned
441  * (which may be both less, equal to or more then end_pfn).
442  *
443  * Assumes that cc->migratepages is empty and cc->nr_migratepages is
444  * zero.
445  *
446  * Apart from cc->migratepages and cc->nr_migratetypes this function
447  * does not modify any cc's fields, in particular it does not modify
448  * (or read for that matter) cc->migrate_pfn.
449  */
450 unsigned long
451 isolate_migratepages_range(struct zone *zone, struct compact_control *cc,
452 		unsigned long low_pfn, unsigned long end_pfn, bool unevictable)
453 {
454 	unsigned long last_pageblock_nr = 0, pageblock_nr;
455 	unsigned long nr_scanned = 0, nr_isolated = 0;
456 	struct list_head *migratelist = &cc->migratepages;
457 	isolate_mode_t mode = 0;
458 	struct lruvec *lruvec;
459 	unsigned long flags;
460 	bool locked = false;
461 	struct page *page = NULL, *valid_page = NULL;
462 
463 	/*
464 	 * Ensure that there are not too many pages isolated from the LRU
465 	 * list by either parallel reclaimers or compaction. If there are,
466 	 * delay for some time until fewer pages are isolated
467 	 */
468 	while (unlikely(too_many_isolated(zone))) {
469 		/* async migration should just abort */
470 		if (!cc->sync)
471 			return 0;
472 
473 		congestion_wait(BLK_RW_ASYNC, HZ/10);
474 
475 		if (fatal_signal_pending(current))
476 			return 0;
477 	}
478 
479 	/* Time to isolate some pages for migration */
480 	cond_resched();
481 	for (; low_pfn < end_pfn; low_pfn++) {
482 		/* give a chance to irqs before checking need_resched() */
483 		if (locked && !((low_pfn+1) % SWAP_CLUSTER_MAX)) {
484 			if (should_release_lock(&zone->lru_lock)) {
485 				spin_unlock_irqrestore(&zone->lru_lock, flags);
486 				locked = false;
487 			}
488 		}
489 
490 		/*
491 		 * migrate_pfn does not necessarily start aligned to a
492 		 * pageblock. Ensure that pfn_valid is called when moving
493 		 * into a new MAX_ORDER_NR_PAGES range in case of large
494 		 * memory holes within the zone
495 		 */
496 		if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
497 			if (!pfn_valid(low_pfn)) {
498 				low_pfn += MAX_ORDER_NR_PAGES - 1;
499 				continue;
500 			}
501 		}
502 
503 		if (!pfn_valid_within(low_pfn))
504 			continue;
505 		nr_scanned++;
506 
507 		/*
508 		 * Get the page and ensure the page is within the same zone.
509 		 * See the comment in isolate_freepages about overlapping
510 		 * nodes. It is deliberate that the new zone lock is not taken
511 		 * as memory compaction should not move pages between nodes.
512 		 */
513 		page = pfn_to_page(low_pfn);
514 		if (page_zone(page) != zone)
515 			continue;
516 
517 		if (!valid_page)
518 			valid_page = page;
519 
520 		/* If isolation recently failed, do not retry */
521 		pageblock_nr = low_pfn >> pageblock_order;
522 		if (!isolation_suitable(cc, page))
523 			goto next_pageblock;
524 
525 		/* Skip if free */
526 		if (PageBuddy(page))
527 			continue;
528 
529 		/*
530 		 * For async migration, also only scan in MOVABLE blocks. Async
531 		 * migration is optimistic to see if the minimum amount of work
532 		 * satisfies the allocation
533 		 */
534 		if (!cc->sync && last_pageblock_nr != pageblock_nr &&
535 		    !migrate_async_suitable(get_pageblock_migratetype(page))) {
536 			cc->finished_update_migrate = true;
537 			goto next_pageblock;
538 		}
539 
540 		/*
541 		 * Check may be lockless but that's ok as we recheck later.
542 		 * It's possible to migrate LRU pages and balloon pages
543 		 * Skip any other type of page
544 		 */
545 		if (!PageLRU(page)) {
546 			if (unlikely(balloon_page_movable(page))) {
547 				if (locked && balloon_page_isolate(page)) {
548 					/* Successfully isolated */
549 					cc->finished_update_migrate = true;
550 					list_add(&page->lru, migratelist);
551 					cc->nr_migratepages++;
552 					nr_isolated++;
553 					goto check_compact_cluster;
554 				}
555 			}
556 			continue;
557 		}
558 
559 		/*
560 		 * PageLRU is set. lru_lock normally excludes isolation
561 		 * splitting and collapsing (collapsing has already happened
562 		 * if PageLRU is set) but the lock is not necessarily taken
563 		 * here and it is wasteful to take it just to check transhuge.
564 		 * Check TransHuge without lock and skip the whole pageblock if
565 		 * it's either a transhuge or hugetlbfs page, as calling
566 		 * compound_order() without preventing THP from splitting the
567 		 * page underneath us may return surprising results.
568 		 */
569 		if (PageTransHuge(page)) {
570 			if (!locked)
571 				goto next_pageblock;
572 			low_pfn += (1 << compound_order(page)) - 1;
573 			continue;
574 		}
575 
576 		/* Check if it is ok to still hold the lock */
577 		locked = compact_checklock_irqsave(&zone->lru_lock, &flags,
578 								locked, cc);
579 		if (!locked || fatal_signal_pending(current))
580 			break;
581 
582 		/* Recheck PageLRU and PageTransHuge under lock */
583 		if (!PageLRU(page))
584 			continue;
585 		if (PageTransHuge(page)) {
586 			low_pfn += (1 << compound_order(page)) - 1;
587 			continue;
588 		}
589 
590 		if (!cc->sync)
591 			mode |= ISOLATE_ASYNC_MIGRATE;
592 
593 		if (unevictable)
594 			mode |= ISOLATE_UNEVICTABLE;
595 
596 		lruvec = mem_cgroup_page_lruvec(page, zone);
597 
598 		/* Try isolate the page */
599 		if (__isolate_lru_page(page, mode) != 0)
600 			continue;
601 
602 		VM_BUG_ON(PageTransCompound(page));
603 
604 		/* Successfully isolated */
605 		cc->finished_update_migrate = true;
606 		del_page_from_lru_list(page, lruvec, page_lru(page));
607 		list_add(&page->lru, migratelist);
608 		cc->nr_migratepages++;
609 		nr_isolated++;
610 
611 check_compact_cluster:
612 		/* Avoid isolating too much */
613 		if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
614 			++low_pfn;
615 			break;
616 		}
617 
618 		continue;
619 
620 next_pageblock:
621 		low_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages) - 1;
622 		last_pageblock_nr = pageblock_nr;
623 	}
624 
625 	acct_isolated(zone, locked, cc);
626 
627 	if (locked)
628 		spin_unlock_irqrestore(&zone->lru_lock, flags);
629 
630 	/* Update the pageblock-skip if the whole pageblock was scanned */
631 	if (low_pfn == end_pfn)
632 		update_pageblock_skip(cc, valid_page, nr_isolated, true);
633 
634 	trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
635 
636 	count_compact_events(COMPACTMIGRATE_SCANNED, nr_scanned);
637 	if (nr_isolated)
638 		count_compact_events(COMPACTISOLATED, nr_isolated);
639 
640 	return low_pfn;
641 }
642 
643 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
644 #ifdef CONFIG_COMPACTION
645 /*
646  * Based on information in the current compact_control, find blocks
647  * suitable for isolating free pages from and then isolate them.
648  */
649 static void isolate_freepages(struct zone *zone,
650 				struct compact_control *cc)
651 {
652 	struct page *page;
653 	unsigned long high_pfn, low_pfn, pfn, z_end_pfn, end_pfn;
654 	int nr_freepages = cc->nr_freepages;
655 	struct list_head *freelist = &cc->freepages;
656 
657 	/*
658 	 * Initialise the free scanner. The starting point is where we last
659 	 * scanned from (or the end of the zone if starting). The low point
660 	 * is the end of the pageblock the migration scanner is using.
661 	 */
662 	pfn = cc->free_pfn;
663 	low_pfn = cc->migrate_pfn + pageblock_nr_pages;
664 
665 	/*
666 	 * Take care that if the migration scanner is at the end of the zone
667 	 * that the free scanner does not accidentally move to the next zone
668 	 * in the next isolation cycle.
669 	 */
670 	high_pfn = min(low_pfn, pfn);
671 
672 	z_end_pfn = zone_end_pfn(zone);
673 
674 	/*
675 	 * Isolate free pages until enough are available to migrate the
676 	 * pages on cc->migratepages. We stop searching if the migrate
677 	 * and free page scanners meet or enough free pages are isolated.
678 	 */
679 	for (; pfn > low_pfn && cc->nr_migratepages > nr_freepages;
680 					pfn -= pageblock_nr_pages) {
681 		unsigned long isolated;
682 
683 		/*
684 		 * This can iterate a massively long zone without finding any
685 		 * suitable migration targets, so periodically check if we need
686 		 * to schedule.
687 		 */
688 		cond_resched();
689 
690 		if (!pfn_valid(pfn))
691 			continue;
692 
693 		/*
694 		 * Check for overlapping nodes/zones. It's possible on some
695 		 * configurations to have a setup like
696 		 * node0 node1 node0
697 		 * i.e. it's possible that all pages within a zones range of
698 		 * pages do not belong to a single zone.
699 		 */
700 		page = pfn_to_page(pfn);
701 		if (page_zone(page) != zone)
702 			continue;
703 
704 		/* Check the block is suitable for migration */
705 		if (!suitable_migration_target(page))
706 			continue;
707 
708 		/* If isolation recently failed, do not retry */
709 		if (!isolation_suitable(cc, page))
710 			continue;
711 
712 		/* Found a block suitable for isolating free pages from */
713 		isolated = 0;
714 
715 		/*
716 		 * As pfn may not start aligned, pfn+pageblock_nr_page
717 		 * may cross a MAX_ORDER_NR_PAGES boundary and miss
718 		 * a pfn_valid check. Ensure isolate_freepages_block()
719 		 * only scans within a pageblock
720 		 */
721 		end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
722 		end_pfn = min(end_pfn, z_end_pfn);
723 		isolated = isolate_freepages_block(cc, pfn, end_pfn,
724 						   freelist, false);
725 		nr_freepages += isolated;
726 
727 		/*
728 		 * Record the highest PFN we isolated pages from. When next
729 		 * looking for free pages, the search will restart here as
730 		 * page migration may have returned some pages to the allocator
731 		 */
732 		if (isolated) {
733 			cc->finished_update_free = true;
734 			high_pfn = max(high_pfn, pfn);
735 		}
736 	}
737 
738 	/* split_free_page does not map the pages */
739 	map_pages(freelist);
740 
741 	cc->free_pfn = high_pfn;
742 	cc->nr_freepages = nr_freepages;
743 }
744 
745 /*
746  * This is a migrate-callback that "allocates" freepages by taking pages
747  * from the isolated freelists in the block we are migrating to.
748  */
749 static struct page *compaction_alloc(struct page *migratepage,
750 					unsigned long data,
751 					int **result)
752 {
753 	struct compact_control *cc = (struct compact_control *)data;
754 	struct page *freepage;
755 
756 	/* Isolate free pages if necessary */
757 	if (list_empty(&cc->freepages)) {
758 		isolate_freepages(cc->zone, cc);
759 
760 		if (list_empty(&cc->freepages))
761 			return NULL;
762 	}
763 
764 	freepage = list_entry(cc->freepages.next, struct page, lru);
765 	list_del(&freepage->lru);
766 	cc->nr_freepages--;
767 
768 	return freepage;
769 }
770 
771 /*
772  * We cannot control nr_migratepages and nr_freepages fully when migration is
773  * running as migrate_pages() has no knowledge of compact_control. When
774  * migration is complete, we count the number of pages on the lists by hand.
775  */
776 static void update_nr_listpages(struct compact_control *cc)
777 {
778 	int nr_migratepages = 0;
779 	int nr_freepages = 0;
780 	struct page *page;
781 
782 	list_for_each_entry(page, &cc->migratepages, lru)
783 		nr_migratepages++;
784 	list_for_each_entry(page, &cc->freepages, lru)
785 		nr_freepages++;
786 
787 	cc->nr_migratepages = nr_migratepages;
788 	cc->nr_freepages = nr_freepages;
789 }
790 
791 /* possible outcome of isolate_migratepages */
792 typedef enum {
793 	ISOLATE_ABORT,		/* Abort compaction now */
794 	ISOLATE_NONE,		/* No pages isolated, continue scanning */
795 	ISOLATE_SUCCESS,	/* Pages isolated, migrate */
796 } isolate_migrate_t;
797 
798 /*
799  * Isolate all pages that can be migrated from the block pointed to by
800  * the migrate scanner within compact_control.
801  */
802 static isolate_migrate_t isolate_migratepages(struct zone *zone,
803 					struct compact_control *cc)
804 {
805 	unsigned long low_pfn, end_pfn;
806 
807 	/* Do not scan outside zone boundaries */
808 	low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
809 
810 	/* Only scan within a pageblock boundary */
811 	end_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages);
812 
813 	/* Do not cross the free scanner or scan within a memory hole */
814 	if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
815 		cc->migrate_pfn = end_pfn;
816 		return ISOLATE_NONE;
817 	}
818 
819 	/* Perform the isolation */
820 	low_pfn = isolate_migratepages_range(zone, cc, low_pfn, end_pfn, false);
821 	if (!low_pfn || cc->contended)
822 		return ISOLATE_ABORT;
823 
824 	cc->migrate_pfn = low_pfn;
825 
826 	return ISOLATE_SUCCESS;
827 }
828 
829 static int compact_finished(struct zone *zone,
830 			    struct compact_control *cc)
831 {
832 	unsigned int order;
833 	unsigned long watermark;
834 
835 	if (fatal_signal_pending(current))
836 		return COMPACT_PARTIAL;
837 
838 	/* Compaction run completes if the migrate and free scanner meet */
839 	if (cc->free_pfn <= cc->migrate_pfn) {
840 		/*
841 		 * Mark that the PG_migrate_skip information should be cleared
842 		 * by kswapd when it goes to sleep. kswapd does not set the
843 		 * flag itself as the decision to be clear should be directly
844 		 * based on an allocation request.
845 		 */
846 		if (!current_is_kswapd())
847 			zone->compact_blockskip_flush = true;
848 
849 		return COMPACT_COMPLETE;
850 	}
851 
852 	/*
853 	 * order == -1 is expected when compacting via
854 	 * /proc/sys/vm/compact_memory
855 	 */
856 	if (cc->order == -1)
857 		return COMPACT_CONTINUE;
858 
859 	/* Compaction run is not finished if the watermark is not met */
860 	watermark = low_wmark_pages(zone);
861 	watermark += (1 << cc->order);
862 
863 	if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
864 		return COMPACT_CONTINUE;
865 
866 	/* Direct compactor: Is a suitable page free? */
867 	for (order = cc->order; order < MAX_ORDER; order++) {
868 		struct free_area *area = &zone->free_area[order];
869 
870 		/* Job done if page is free of the right migratetype */
871 		if (!list_empty(&area->free_list[cc->migratetype]))
872 			return COMPACT_PARTIAL;
873 
874 		/* Job done if allocation would set block type */
875 		if (cc->order >= pageblock_order && area->nr_free)
876 			return COMPACT_PARTIAL;
877 	}
878 
879 	return COMPACT_CONTINUE;
880 }
881 
882 /*
883  * compaction_suitable: Is this suitable to run compaction on this zone now?
884  * Returns
885  *   COMPACT_SKIPPED  - If there are too few free pages for compaction
886  *   COMPACT_PARTIAL  - If the allocation would succeed without compaction
887  *   COMPACT_CONTINUE - If compaction should run now
888  */
889 unsigned long compaction_suitable(struct zone *zone, int order)
890 {
891 	int fragindex;
892 	unsigned long watermark;
893 
894 	/*
895 	 * order == -1 is expected when compacting via
896 	 * /proc/sys/vm/compact_memory
897 	 */
898 	if (order == -1)
899 		return COMPACT_CONTINUE;
900 
901 	/*
902 	 * Watermarks for order-0 must be met for compaction. Note the 2UL.
903 	 * This is because during migration, copies of pages need to be
904 	 * allocated and for a short time, the footprint is higher
905 	 */
906 	watermark = low_wmark_pages(zone) + (2UL << order);
907 	if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
908 		return COMPACT_SKIPPED;
909 
910 	/*
911 	 * fragmentation index determines if allocation failures are due to
912 	 * low memory or external fragmentation
913 	 *
914 	 * index of -1000 implies allocations might succeed depending on
915 	 * watermarks
916 	 * index towards 0 implies failure is due to lack of memory
917 	 * index towards 1000 implies failure is due to fragmentation
918 	 *
919 	 * Only compact if a failure would be due to fragmentation.
920 	 */
921 	fragindex = fragmentation_index(zone, order);
922 	if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
923 		return COMPACT_SKIPPED;
924 
925 	if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
926 	    0, 0))
927 		return COMPACT_PARTIAL;
928 
929 	return COMPACT_CONTINUE;
930 }
931 
932 static int compact_zone(struct zone *zone, struct compact_control *cc)
933 {
934 	int ret;
935 	unsigned long start_pfn = zone->zone_start_pfn;
936 	unsigned long end_pfn = zone_end_pfn(zone);
937 
938 	ret = compaction_suitable(zone, cc->order);
939 	switch (ret) {
940 	case COMPACT_PARTIAL:
941 	case COMPACT_SKIPPED:
942 		/* Compaction is likely to fail */
943 		return ret;
944 	case COMPACT_CONTINUE:
945 		/* Fall through to compaction */
946 		;
947 	}
948 
949 	/*
950 	 * Setup to move all movable pages to the end of the zone. Used cached
951 	 * information on where the scanners should start but check that it
952 	 * is initialised by ensuring the values are within zone boundaries.
953 	 */
954 	cc->migrate_pfn = zone->compact_cached_migrate_pfn;
955 	cc->free_pfn = zone->compact_cached_free_pfn;
956 	if (cc->free_pfn < start_pfn || cc->free_pfn > end_pfn) {
957 		cc->free_pfn = end_pfn & ~(pageblock_nr_pages-1);
958 		zone->compact_cached_free_pfn = cc->free_pfn;
959 	}
960 	if (cc->migrate_pfn < start_pfn || cc->migrate_pfn > end_pfn) {
961 		cc->migrate_pfn = start_pfn;
962 		zone->compact_cached_migrate_pfn = cc->migrate_pfn;
963 	}
964 
965 	/*
966 	 * Clear pageblock skip if there were failures recently and compaction
967 	 * is about to be retried after being deferred. kswapd does not do
968 	 * this reset as it'll reset the cached information when going to sleep.
969 	 */
970 	if (compaction_restarting(zone, cc->order) && !current_is_kswapd())
971 		__reset_isolation_suitable(zone);
972 
973 	migrate_prep_local();
974 
975 	while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
976 		unsigned long nr_migrate, nr_remaining;
977 		int err;
978 
979 		switch (isolate_migratepages(zone, cc)) {
980 		case ISOLATE_ABORT:
981 			ret = COMPACT_PARTIAL;
982 			putback_movable_pages(&cc->migratepages);
983 			cc->nr_migratepages = 0;
984 			goto out;
985 		case ISOLATE_NONE:
986 			continue;
987 		case ISOLATE_SUCCESS:
988 			;
989 		}
990 
991 		nr_migrate = cc->nr_migratepages;
992 		err = migrate_pages(&cc->migratepages, compaction_alloc,
993 				(unsigned long)cc,
994 				cc->sync ? MIGRATE_SYNC_LIGHT : MIGRATE_ASYNC,
995 				MR_COMPACTION);
996 		update_nr_listpages(cc);
997 		nr_remaining = cc->nr_migratepages;
998 
999 		trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
1000 						nr_remaining);
1001 
1002 		/* Release isolated pages not migrated */
1003 		if (err) {
1004 			putback_movable_pages(&cc->migratepages);
1005 			cc->nr_migratepages = 0;
1006 			if (err == -ENOMEM) {
1007 				ret = COMPACT_PARTIAL;
1008 				goto out;
1009 			}
1010 		}
1011 	}
1012 
1013 out:
1014 	/* Release free pages and check accounting */
1015 	cc->nr_freepages -= release_freepages(&cc->freepages);
1016 	VM_BUG_ON(cc->nr_freepages != 0);
1017 
1018 	return ret;
1019 }
1020 
1021 static unsigned long compact_zone_order(struct zone *zone,
1022 				 int order, gfp_t gfp_mask,
1023 				 bool sync, bool *contended)
1024 {
1025 	unsigned long ret;
1026 	struct compact_control cc = {
1027 		.nr_freepages = 0,
1028 		.nr_migratepages = 0,
1029 		.order = order,
1030 		.migratetype = allocflags_to_migratetype(gfp_mask),
1031 		.zone = zone,
1032 		.sync = sync,
1033 	};
1034 	INIT_LIST_HEAD(&cc.freepages);
1035 	INIT_LIST_HEAD(&cc.migratepages);
1036 
1037 	ret = compact_zone(zone, &cc);
1038 
1039 	VM_BUG_ON(!list_empty(&cc.freepages));
1040 	VM_BUG_ON(!list_empty(&cc.migratepages));
1041 
1042 	*contended = cc.contended;
1043 	return ret;
1044 }
1045 
1046 int sysctl_extfrag_threshold = 500;
1047 
1048 /**
1049  * try_to_compact_pages - Direct compact to satisfy a high-order allocation
1050  * @zonelist: The zonelist used for the current allocation
1051  * @order: The order of the current allocation
1052  * @gfp_mask: The GFP mask of the current allocation
1053  * @nodemask: The allowed nodes to allocate from
1054  * @sync: Whether migration is synchronous or not
1055  * @contended: Return value that is true if compaction was aborted due to lock contention
1056  * @page: Optionally capture a free page of the requested order during compaction
1057  *
1058  * This is the main entry point for direct page compaction.
1059  */
1060 unsigned long try_to_compact_pages(struct zonelist *zonelist,
1061 			int order, gfp_t gfp_mask, nodemask_t *nodemask,
1062 			bool sync, bool *contended)
1063 {
1064 	enum zone_type high_zoneidx = gfp_zone(gfp_mask);
1065 	int may_enter_fs = gfp_mask & __GFP_FS;
1066 	int may_perform_io = gfp_mask & __GFP_IO;
1067 	struct zoneref *z;
1068 	struct zone *zone;
1069 	int rc = COMPACT_SKIPPED;
1070 	int alloc_flags = 0;
1071 
1072 	/* Check if the GFP flags allow compaction */
1073 	if (!order || !may_enter_fs || !may_perform_io)
1074 		return rc;
1075 
1076 	count_compact_event(COMPACTSTALL);
1077 
1078 #ifdef CONFIG_CMA
1079 	if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
1080 		alloc_flags |= ALLOC_CMA;
1081 #endif
1082 	/* Compact each zone in the list */
1083 	for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1084 								nodemask) {
1085 		int status;
1086 
1087 		status = compact_zone_order(zone, order, gfp_mask, sync,
1088 						contended);
1089 		rc = max(status, rc);
1090 
1091 		/* If a normal allocation would succeed, stop compacting */
1092 		if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0,
1093 				      alloc_flags))
1094 			break;
1095 	}
1096 
1097 	return rc;
1098 }
1099 
1100 
1101 /* Compact all zones within a node */
1102 static void __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
1103 {
1104 	int zoneid;
1105 	struct zone *zone;
1106 
1107 	for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1108 
1109 		zone = &pgdat->node_zones[zoneid];
1110 		if (!populated_zone(zone))
1111 			continue;
1112 
1113 		cc->nr_freepages = 0;
1114 		cc->nr_migratepages = 0;
1115 		cc->zone = zone;
1116 		INIT_LIST_HEAD(&cc->freepages);
1117 		INIT_LIST_HEAD(&cc->migratepages);
1118 
1119 		if (cc->order == -1 || !compaction_deferred(zone, cc->order))
1120 			compact_zone(zone, cc);
1121 
1122 		if (cc->order > 0) {
1123 			int ok = zone_watermark_ok(zone, cc->order,
1124 						low_wmark_pages(zone), 0, 0);
1125 			if (ok && cc->order >= zone->compact_order_failed)
1126 				zone->compact_order_failed = cc->order + 1;
1127 			/* Currently async compaction is never deferred. */
1128 			else if (!ok && cc->sync)
1129 				defer_compaction(zone, cc->order);
1130 		}
1131 
1132 		VM_BUG_ON(!list_empty(&cc->freepages));
1133 		VM_BUG_ON(!list_empty(&cc->migratepages));
1134 	}
1135 }
1136 
1137 void compact_pgdat(pg_data_t *pgdat, int order)
1138 {
1139 	struct compact_control cc = {
1140 		.order = order,
1141 		.sync = false,
1142 	};
1143 
1144 	if (!order)
1145 		return;
1146 
1147 	__compact_pgdat(pgdat, &cc);
1148 }
1149 
1150 static void compact_node(int nid)
1151 {
1152 	struct compact_control cc = {
1153 		.order = -1,
1154 		.sync = true,
1155 	};
1156 
1157 	__compact_pgdat(NODE_DATA(nid), &cc);
1158 }
1159 
1160 /* Compact all nodes in the system */
1161 static void compact_nodes(void)
1162 {
1163 	int nid;
1164 
1165 	/* Flush pending updates to the LRU lists */
1166 	lru_add_drain_all();
1167 
1168 	for_each_online_node(nid)
1169 		compact_node(nid);
1170 }
1171 
1172 /* The written value is actually unused, all memory is compacted */
1173 int sysctl_compact_memory;
1174 
1175 /* This is the entry point for compacting all nodes via /proc/sys/vm */
1176 int sysctl_compaction_handler(struct ctl_table *table, int write,
1177 			void __user *buffer, size_t *length, loff_t *ppos)
1178 {
1179 	if (write)
1180 		compact_nodes();
1181 
1182 	return 0;
1183 }
1184 
1185 int sysctl_extfrag_handler(struct ctl_table *table, int write,
1186 			void __user *buffer, size_t *length, loff_t *ppos)
1187 {
1188 	proc_dointvec_minmax(table, write, buffer, length, ppos);
1189 
1190 	return 0;
1191 }
1192 
1193 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
1194 ssize_t sysfs_compact_node(struct device *dev,
1195 			struct device_attribute *attr,
1196 			const char *buf, size_t count)
1197 {
1198 	int nid = dev->id;
1199 
1200 	if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
1201 		/* Flush pending updates to the LRU lists */
1202 		lru_add_drain_all();
1203 
1204 		compact_node(nid);
1205 	}
1206 
1207 	return count;
1208 }
1209 static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
1210 
1211 int compaction_register_node(struct node *node)
1212 {
1213 	return device_create_file(&node->dev, &dev_attr_compact);
1214 }
1215 
1216 void compaction_unregister_node(struct node *node)
1217 {
1218 	return device_remove_file(&node->dev, &dev_attr_compact);
1219 }
1220 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
1221 
1222 #endif /* CONFIG_COMPACTION */
1223