1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * linux/mm/compaction.c
4  *
5  * Memory compaction for the reduction of external fragmentation. Note that
6  * this heavily depends upon page migration to do all the real heavy
7  * lifting
8  *
9  * Copyright IBM Corp. 2007-2010 Mel Gorman <mel@csn.ul.ie>
10  */
11 #include <linux/cpu.h>
12 #include <linux/swap.h>
13 #include <linux/migrate.h>
14 #include <linux/compaction.h>
15 #include <linux/mm_inline.h>
16 #include <linux/sched/signal.h>
17 #include <linux/backing-dev.h>
18 #include <linux/sysctl.h>
19 #include <linux/sysfs.h>
20 #include <linux/page-isolation.h>
21 #include <linux/kasan.h>
22 #include <linux/kthread.h>
23 #include <linux/freezer.h>
24 #include <linux/page_owner.h>
25 #include <linux/psi.h>
26 #include "internal.h"
27 
28 #ifdef CONFIG_COMPACTION
29 /*
30  * Fragmentation score check interval for proactive compaction purposes.
31  */
32 #define HPAGE_FRAG_CHECK_INTERVAL_MSEC	(500)
33 
count_compact_event(enum vm_event_item item)34 static inline void count_compact_event(enum vm_event_item item)
35 {
36 	count_vm_event(item);
37 }
38 
count_compact_events(enum vm_event_item item,long delta)39 static inline void count_compact_events(enum vm_event_item item, long delta)
40 {
41 	count_vm_events(item, delta);
42 }
43 #else
44 #define count_compact_event(item) do { } while (0)
45 #define count_compact_events(item, delta) do { } while (0)
46 #endif
47 
48 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
49 
50 #define CREATE_TRACE_POINTS
51 #include <trace/events/compaction.h>
52 
53 #define block_start_pfn(pfn, order)	round_down(pfn, 1UL << (order))
54 #define block_end_pfn(pfn, order)	ALIGN((pfn) + 1, 1UL << (order))
55 #define pageblock_start_pfn(pfn)	block_start_pfn(pfn, pageblock_order)
56 #define pageblock_end_pfn(pfn)		block_end_pfn(pfn, pageblock_order)
57 
58 /*
59  * Page order with-respect-to which proactive compaction
60  * calculates external fragmentation, which is used as
61  * the "fragmentation score" of a node/zone.
62  */
63 #if defined CONFIG_TRANSPARENT_HUGEPAGE
64 #define COMPACTION_HPAGE_ORDER	HPAGE_PMD_ORDER
65 #elif defined CONFIG_HUGETLBFS
66 #define COMPACTION_HPAGE_ORDER	HUGETLB_PAGE_ORDER
67 #else
68 #define COMPACTION_HPAGE_ORDER	(PMD_SHIFT - PAGE_SHIFT)
69 #endif
70 
release_freepages(struct list_head * freelist)71 static unsigned long release_freepages(struct list_head *freelist)
72 {
73 	struct page *page, *next;
74 	unsigned long high_pfn = 0;
75 
76 	list_for_each_entry_safe(page, next, freelist, lru) {
77 		unsigned long pfn = page_to_pfn(page);
78 		list_del(&page->lru);
79 		__free_page(page);
80 		if (pfn > high_pfn)
81 			high_pfn = pfn;
82 	}
83 
84 	return high_pfn;
85 }
86 
split_map_pages(struct list_head * list)87 static void split_map_pages(struct list_head *list)
88 {
89 	unsigned int i, order, nr_pages;
90 	struct page *page, *next;
91 	LIST_HEAD(tmp_list);
92 
93 	list_for_each_entry_safe(page, next, list, lru) {
94 		list_del(&page->lru);
95 
96 		order = page_private(page);
97 		nr_pages = 1 << order;
98 
99 		post_alloc_hook(page, order, __GFP_MOVABLE);
100 		if (order)
101 			split_page(page, order);
102 
103 		for (i = 0; i < nr_pages; i++) {
104 			list_add(&page->lru, &tmp_list);
105 			page++;
106 		}
107 	}
108 
109 	list_splice(&tmp_list, list);
110 }
111 
112 #ifdef CONFIG_COMPACTION
113 
PageMovable(struct page * page)114 int PageMovable(struct page *page)
115 {
116 	struct address_space *mapping;
117 
118 	VM_BUG_ON_PAGE(!PageLocked(page), page);
119 	if (!__PageMovable(page))
120 		return 0;
121 
122 	mapping = page_mapping(page);
123 	if (mapping && mapping->a_ops && mapping->a_ops->isolate_page)
124 		return 1;
125 
126 	return 0;
127 }
128 EXPORT_SYMBOL(PageMovable);
129 
__SetPageMovable(struct page * page,struct address_space * mapping)130 void __SetPageMovable(struct page *page, struct address_space *mapping)
131 {
132 	VM_BUG_ON_PAGE(!PageLocked(page), page);
133 	VM_BUG_ON_PAGE((unsigned long)mapping & PAGE_MAPPING_MOVABLE, page);
134 	page->mapping = (void *)((unsigned long)mapping | PAGE_MAPPING_MOVABLE);
135 }
136 EXPORT_SYMBOL(__SetPageMovable);
137 
__ClearPageMovable(struct page * page)138 void __ClearPageMovable(struct page *page)
139 {
140 	VM_BUG_ON_PAGE(!PageMovable(page), page);
141 	/*
142 	 * Clear registered address_space val with keeping PAGE_MAPPING_MOVABLE
143 	 * flag so that VM can catch up released page by driver after isolation.
144 	 * With it, VM migration doesn't try to put it back.
145 	 */
146 	page->mapping = (void *)((unsigned long)page->mapping &
147 				PAGE_MAPPING_MOVABLE);
148 }
149 EXPORT_SYMBOL(__ClearPageMovable);
150 
151 /* Do not skip compaction more than 64 times */
152 #define COMPACT_MAX_DEFER_SHIFT 6
153 
154 /*
155  * Compaction is deferred when compaction fails to result in a page
156  * allocation success. 1 << compact_defer_shift, compactions are skipped up
157  * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
158  */
defer_compaction(struct zone * zone,int order)159 static void defer_compaction(struct zone *zone, int order)
160 {
161 	zone->compact_considered = 0;
162 	zone->compact_defer_shift++;
163 
164 	if (order < zone->compact_order_failed)
165 		zone->compact_order_failed = order;
166 
167 	if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
168 		zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
169 
170 	trace_mm_compaction_defer_compaction(zone, order);
171 }
172 
173 /* Returns true if compaction should be skipped this time */
compaction_deferred(struct zone * zone,int order)174 static bool compaction_deferred(struct zone *zone, int order)
175 {
176 	unsigned long defer_limit = 1UL << zone->compact_defer_shift;
177 
178 	if (order < zone->compact_order_failed)
179 		return false;
180 
181 	/* Avoid possible overflow */
182 	if (++zone->compact_considered >= defer_limit) {
183 		zone->compact_considered = defer_limit;
184 		return false;
185 	}
186 
187 	trace_mm_compaction_deferred(zone, order);
188 
189 	return true;
190 }
191 
192 /*
193  * Update defer tracking counters after successful compaction of given order,
194  * which means an allocation either succeeded (alloc_success == true) or is
195  * expected to succeed.
196  */
compaction_defer_reset(struct zone * zone,int order,bool alloc_success)197 void compaction_defer_reset(struct zone *zone, int order,
198 		bool alloc_success)
199 {
200 	if (alloc_success) {
201 		zone->compact_considered = 0;
202 		zone->compact_defer_shift = 0;
203 	}
204 	if (order >= zone->compact_order_failed)
205 		zone->compact_order_failed = order + 1;
206 
207 	trace_mm_compaction_defer_reset(zone, order);
208 }
209 
210 /* Returns true if restarting compaction after many failures */
compaction_restarting(struct zone * zone,int order)211 static bool compaction_restarting(struct zone *zone, int order)
212 {
213 	if (order < zone->compact_order_failed)
214 		return false;
215 
216 	return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
217 		zone->compact_considered >= 1UL << zone->compact_defer_shift;
218 }
219 
220 /* Returns true if the pageblock should be scanned for pages to isolate. */
isolation_suitable(struct compact_control * cc,struct page * page)221 static inline bool isolation_suitable(struct compact_control *cc,
222 					struct page *page)
223 {
224 	if (cc->ignore_skip_hint)
225 		return true;
226 
227 	return !get_pageblock_skip(page);
228 }
229 
reset_cached_positions(struct zone * zone)230 static void reset_cached_positions(struct zone *zone)
231 {
232 	zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
233 	zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
234 	zone->compact_cached_free_pfn =
235 				pageblock_start_pfn(zone_end_pfn(zone) - 1);
236 }
237 
238 /*
239  * Compound pages of >= pageblock_order should consistently be skipped until
240  * released. It is always pointless to compact pages of such order (if they are
241  * migratable), and the pageblocks they occupy cannot contain any free pages.
242  */
pageblock_skip_persistent(struct page * page)243 static bool pageblock_skip_persistent(struct page *page)
244 {
245 	if (!PageCompound(page))
246 		return false;
247 
248 	page = compound_head(page);
249 
250 	if (compound_order(page) >= pageblock_order)
251 		return true;
252 
253 	return false;
254 }
255 
256 static bool
__reset_isolation_pfn(struct zone * zone,unsigned long pfn,bool check_source,bool check_target)257 __reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source,
258 							bool check_target)
259 {
260 	struct page *page = pfn_to_online_page(pfn);
261 	struct page *block_page;
262 	struct page *end_page;
263 	unsigned long block_pfn;
264 
265 	if (!page)
266 		return false;
267 	if (zone != page_zone(page))
268 		return false;
269 	if (pageblock_skip_persistent(page))
270 		return false;
271 
272 	/*
273 	 * If skip is already cleared do no further checking once the
274 	 * restart points have been set.
275 	 */
276 	if (check_source && check_target && !get_pageblock_skip(page))
277 		return true;
278 
279 	/*
280 	 * If clearing skip for the target scanner, do not select a
281 	 * non-movable pageblock as the starting point.
282 	 */
283 	if (!check_source && check_target &&
284 	    get_pageblock_migratetype(page) != MIGRATE_MOVABLE)
285 		return false;
286 
287 	/* Ensure the start of the pageblock or zone is online and valid */
288 	block_pfn = pageblock_start_pfn(pfn);
289 	block_pfn = max(block_pfn, zone->zone_start_pfn);
290 	block_page = pfn_to_online_page(block_pfn);
291 	if (block_page) {
292 		page = block_page;
293 		pfn = block_pfn;
294 	}
295 
296 	/* Ensure the end of the pageblock or zone is online and valid */
297 	block_pfn = pageblock_end_pfn(pfn) - 1;
298 	block_pfn = min(block_pfn, zone_end_pfn(zone) - 1);
299 	end_page = pfn_to_online_page(block_pfn);
300 	if (!end_page)
301 		return false;
302 
303 	/*
304 	 * Only clear the hint if a sample indicates there is either a
305 	 * free page or an LRU page in the block. One or other condition
306 	 * is necessary for the block to be a migration source/target.
307 	 */
308 	do {
309 		if (check_source && PageLRU(page)) {
310 			clear_pageblock_skip(page);
311 			return true;
312 		}
313 
314 		if (check_target && PageBuddy(page)) {
315 			clear_pageblock_skip(page);
316 			return true;
317 		}
318 
319 		page += (1 << PAGE_ALLOC_COSTLY_ORDER);
320 	} while (page <= end_page);
321 
322 	return false;
323 }
324 
325 /*
326  * This function is called to clear all cached information on pageblocks that
327  * should be skipped for page isolation when the migrate and free page scanner
328  * meet.
329  */
__reset_isolation_suitable(struct zone * zone)330 static void __reset_isolation_suitable(struct zone *zone)
331 {
332 	unsigned long migrate_pfn = zone->zone_start_pfn;
333 	unsigned long free_pfn = zone_end_pfn(zone) - 1;
334 	unsigned long reset_migrate = free_pfn;
335 	unsigned long reset_free = migrate_pfn;
336 	bool source_set = false;
337 	bool free_set = false;
338 
339 	if (!zone->compact_blockskip_flush)
340 		return;
341 
342 	zone->compact_blockskip_flush = false;
343 
344 	/*
345 	 * Walk the zone and update pageblock skip information. Source looks
346 	 * for PageLRU while target looks for PageBuddy. When the scanner
347 	 * is found, both PageBuddy and PageLRU are checked as the pageblock
348 	 * is suitable as both source and target.
349 	 */
350 	for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages,
351 					free_pfn -= pageblock_nr_pages) {
352 		cond_resched();
353 
354 		/* Update the migrate PFN */
355 		if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) &&
356 		    migrate_pfn < reset_migrate) {
357 			source_set = true;
358 			reset_migrate = migrate_pfn;
359 			zone->compact_init_migrate_pfn = reset_migrate;
360 			zone->compact_cached_migrate_pfn[0] = reset_migrate;
361 			zone->compact_cached_migrate_pfn[1] = reset_migrate;
362 		}
363 
364 		/* Update the free PFN */
365 		if (__reset_isolation_pfn(zone, free_pfn, free_set, true) &&
366 		    free_pfn > reset_free) {
367 			free_set = true;
368 			reset_free = free_pfn;
369 			zone->compact_init_free_pfn = reset_free;
370 			zone->compact_cached_free_pfn = reset_free;
371 		}
372 	}
373 
374 	/* Leave no distance if no suitable block was reset */
375 	if (reset_migrate >= reset_free) {
376 		zone->compact_cached_migrate_pfn[0] = migrate_pfn;
377 		zone->compact_cached_migrate_pfn[1] = migrate_pfn;
378 		zone->compact_cached_free_pfn = free_pfn;
379 	}
380 }
381 
reset_isolation_suitable(pg_data_t * pgdat)382 void reset_isolation_suitable(pg_data_t *pgdat)
383 {
384 	int zoneid;
385 
386 	for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
387 		struct zone *zone = &pgdat->node_zones[zoneid];
388 		if (!populated_zone(zone))
389 			continue;
390 
391 		/* Only flush if a full compaction finished recently */
392 		if (zone->compact_blockskip_flush)
393 			__reset_isolation_suitable(zone);
394 	}
395 }
396 
397 /*
398  * Sets the pageblock skip bit if it was clear. Note that this is a hint as
399  * locks are not required for read/writers. Returns true if it was already set.
400  */
test_and_set_skip(struct compact_control * cc,struct page * page,unsigned long pfn)401 static bool test_and_set_skip(struct compact_control *cc, struct page *page,
402 							unsigned long pfn)
403 {
404 	bool skip;
405 
406 	/* Do no update if skip hint is being ignored */
407 	if (cc->ignore_skip_hint)
408 		return false;
409 
410 	if (!IS_ALIGNED(pfn, pageblock_nr_pages))
411 		return false;
412 
413 	skip = get_pageblock_skip(page);
414 	if (!skip && !cc->no_set_skip_hint)
415 		set_pageblock_skip(page);
416 
417 	return skip;
418 }
419 
update_cached_migrate(struct compact_control * cc,unsigned long pfn)420 static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
421 {
422 	struct zone *zone = cc->zone;
423 
424 	pfn = pageblock_end_pfn(pfn);
425 
426 	/* Set for isolation rather than compaction */
427 	if (cc->no_set_skip_hint)
428 		return;
429 
430 	if (pfn > zone->compact_cached_migrate_pfn[0])
431 		zone->compact_cached_migrate_pfn[0] = pfn;
432 	if (cc->mode != MIGRATE_ASYNC &&
433 	    pfn > zone->compact_cached_migrate_pfn[1])
434 		zone->compact_cached_migrate_pfn[1] = pfn;
435 }
436 
437 /*
438  * If no pages were isolated then mark this pageblock to be skipped in the
439  * future. The information is later cleared by __reset_isolation_suitable().
440  */
update_pageblock_skip(struct compact_control * cc,struct page * page,unsigned long pfn)441 static void update_pageblock_skip(struct compact_control *cc,
442 			struct page *page, unsigned long pfn)
443 {
444 	struct zone *zone = cc->zone;
445 
446 	if (cc->no_set_skip_hint)
447 		return;
448 
449 	if (!page)
450 		return;
451 
452 	set_pageblock_skip(page);
453 
454 	/* Update where async and sync compaction should restart */
455 	if (pfn < zone->compact_cached_free_pfn)
456 		zone->compact_cached_free_pfn = pfn;
457 }
458 #else
isolation_suitable(struct compact_control * cc,struct page * page)459 static inline bool isolation_suitable(struct compact_control *cc,
460 					struct page *page)
461 {
462 	return true;
463 }
464 
pageblock_skip_persistent(struct page * page)465 static inline bool pageblock_skip_persistent(struct page *page)
466 {
467 	return false;
468 }
469 
update_pageblock_skip(struct compact_control * cc,struct page * page,unsigned long pfn)470 static inline void update_pageblock_skip(struct compact_control *cc,
471 			struct page *page, unsigned long pfn)
472 {
473 }
474 
update_cached_migrate(struct compact_control * cc,unsigned long pfn)475 static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
476 {
477 }
478 
test_and_set_skip(struct compact_control * cc,struct page * page,unsigned long pfn)479 static bool test_and_set_skip(struct compact_control *cc, struct page *page,
480 							unsigned long pfn)
481 {
482 	return false;
483 }
484 #endif /* CONFIG_COMPACTION */
485 
486 /*
487  * Compaction requires the taking of some coarse locks that are potentially
488  * very heavily contended. For async compaction, trylock and record if the
489  * lock is contended. The lock will still be acquired but compaction will
490  * abort when the current block is finished regardless of success rate.
491  * Sync compaction acquires the lock.
492  *
493  * Always returns true which makes it easier to track lock state in callers.
494  */
compact_lock_irqsave(spinlock_t * lock,unsigned long * flags,struct compact_control * cc)495 static bool compact_lock_irqsave(spinlock_t *lock, unsigned long *flags,
496 						struct compact_control *cc)
497 	__acquires(lock)
498 {
499 	/* Track if the lock is contended in async mode */
500 	if (cc->mode == MIGRATE_ASYNC && !cc->contended) {
501 		if (spin_trylock_irqsave(lock, *flags))
502 			return true;
503 
504 		cc->contended = true;
505 	}
506 
507 	spin_lock_irqsave(lock, *flags);
508 	return true;
509 }
510 
511 /*
512  * Compaction requires the taking of some coarse locks that are potentially
513  * very heavily contended. The lock should be periodically unlocked to avoid
514  * having disabled IRQs for a long time, even when there is nobody waiting on
515  * the lock. It might also be that allowing the IRQs will result in
516  * need_resched() becoming true. If scheduling is needed, compaction schedules.
517  * Either compaction type will also abort if a fatal signal is pending.
518  * In either case if the lock was locked, it is dropped and not regained.
519  *
520  * Returns true if compaction should abort due to fatal signal pending.
521  * Returns false when compaction can continue.
522  */
compact_unlock_should_abort(spinlock_t * lock,unsigned long flags,bool * locked,struct compact_control * cc)523 static bool compact_unlock_should_abort(spinlock_t *lock,
524 		unsigned long flags, bool *locked, struct compact_control *cc)
525 {
526 	if (*locked) {
527 		spin_unlock_irqrestore(lock, flags);
528 		*locked = false;
529 	}
530 
531 	if (fatal_signal_pending(current)) {
532 		cc->contended = true;
533 		return true;
534 	}
535 
536 	cond_resched();
537 
538 	return false;
539 }
540 
541 /*
542  * Isolate free pages onto a private freelist. If @strict is true, will abort
543  * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
544  * (even though it may still end up isolating some pages).
545  */
isolate_freepages_block(struct compact_control * cc,unsigned long * start_pfn,unsigned long end_pfn,struct list_head * freelist,unsigned int stride,bool strict)546 static unsigned long isolate_freepages_block(struct compact_control *cc,
547 				unsigned long *start_pfn,
548 				unsigned long end_pfn,
549 				struct list_head *freelist,
550 				unsigned int stride,
551 				bool strict)
552 {
553 	int nr_scanned = 0, total_isolated = 0;
554 	struct page *cursor;
555 	unsigned long flags = 0;
556 	bool locked = false;
557 	unsigned long blockpfn = *start_pfn;
558 	unsigned int order;
559 
560 	/* Strict mode is for isolation, speed is secondary */
561 	if (strict)
562 		stride = 1;
563 
564 	cursor = pfn_to_page(blockpfn);
565 
566 	/* Isolate free pages. */
567 	for (; blockpfn < end_pfn; blockpfn += stride, cursor += stride) {
568 		int isolated;
569 		struct page *page = cursor;
570 
571 		/*
572 		 * Periodically drop the lock (if held) regardless of its
573 		 * contention, to give chance to IRQs. Abort if fatal signal
574 		 * pending.
575 		 */
576 		if (!(blockpfn % COMPACT_CLUSTER_MAX)
577 		    && compact_unlock_should_abort(&cc->zone->lock, flags,
578 								&locked, cc))
579 			break;
580 
581 		nr_scanned++;
582 
583 		/*
584 		 * For compound pages such as THP and hugetlbfs, we can save
585 		 * potentially a lot of iterations if we skip them at once.
586 		 * The check is racy, but we can consider only valid values
587 		 * and the only danger is skipping too much.
588 		 */
589 		if (PageCompound(page)) {
590 			const unsigned int order = compound_order(page);
591 
592 			if (likely(order < MAX_ORDER)) {
593 				blockpfn += (1UL << order) - 1;
594 				cursor += (1UL << order) - 1;
595 			}
596 			goto isolate_fail;
597 		}
598 
599 		if (!PageBuddy(page))
600 			goto isolate_fail;
601 
602 		/* If we already hold the lock, we can skip some rechecking. */
603 		if (!locked) {
604 			locked = compact_lock_irqsave(&cc->zone->lock,
605 								&flags, cc);
606 
607 			/* Recheck this is a buddy page under lock */
608 			if (!PageBuddy(page))
609 				goto isolate_fail;
610 		}
611 
612 		/* Found a free page, will break it into order-0 pages */
613 		order = buddy_order(page);
614 		isolated = __isolate_free_page(page, order);
615 		if (!isolated)
616 			break;
617 		set_page_private(page, order);
618 
619 		total_isolated += isolated;
620 		cc->nr_freepages += isolated;
621 		list_add_tail(&page->lru, freelist);
622 
623 		if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
624 			blockpfn += isolated;
625 			break;
626 		}
627 		/* Advance to the end of split page */
628 		blockpfn += isolated - 1;
629 		cursor += isolated - 1;
630 		continue;
631 
632 isolate_fail:
633 		if (strict)
634 			break;
635 		else
636 			continue;
637 
638 	}
639 
640 	if (locked)
641 		spin_unlock_irqrestore(&cc->zone->lock, flags);
642 
643 	/*
644 	 * There is a tiny chance that we have read bogus compound_order(),
645 	 * so be careful to not go outside of the pageblock.
646 	 */
647 	if (unlikely(blockpfn > end_pfn))
648 		blockpfn = end_pfn;
649 
650 	trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
651 					nr_scanned, total_isolated);
652 
653 	/* Record how far we have got within the block */
654 	*start_pfn = blockpfn;
655 
656 	/*
657 	 * If strict isolation is requested by CMA then check that all the
658 	 * pages requested were isolated. If there were any failures, 0 is
659 	 * returned and CMA will fail.
660 	 */
661 	if (strict && blockpfn < end_pfn)
662 		total_isolated = 0;
663 
664 	cc->total_free_scanned += nr_scanned;
665 	if (total_isolated)
666 		count_compact_events(COMPACTISOLATED, total_isolated);
667 	return total_isolated;
668 }
669 
670 /**
671  * isolate_freepages_range() - isolate free pages.
672  * @cc:        Compaction control structure.
673  * @start_pfn: The first PFN to start isolating.
674  * @end_pfn:   The one-past-last PFN.
675  *
676  * Non-free pages, invalid PFNs, or zone boundaries within the
677  * [start_pfn, end_pfn) range are considered errors, cause function to
678  * undo its actions and return zero.
679  *
680  * Otherwise, function returns one-past-the-last PFN of isolated page
681  * (which may be greater then end_pfn if end fell in a middle of
682  * a free page).
683  */
684 unsigned long
isolate_freepages_range(struct compact_control * cc,unsigned long start_pfn,unsigned long end_pfn)685 isolate_freepages_range(struct compact_control *cc,
686 			unsigned long start_pfn, unsigned long end_pfn)
687 {
688 	unsigned long isolated, pfn, block_start_pfn, block_end_pfn;
689 	LIST_HEAD(freelist);
690 
691 	pfn = start_pfn;
692 	block_start_pfn = pageblock_start_pfn(pfn);
693 	if (block_start_pfn < cc->zone->zone_start_pfn)
694 		block_start_pfn = cc->zone->zone_start_pfn;
695 	block_end_pfn = pageblock_end_pfn(pfn);
696 
697 	for (; pfn < end_pfn; pfn += isolated,
698 				block_start_pfn = block_end_pfn,
699 				block_end_pfn += pageblock_nr_pages) {
700 		/* Protect pfn from changing by isolate_freepages_block */
701 		unsigned long isolate_start_pfn = pfn;
702 
703 		block_end_pfn = min(block_end_pfn, end_pfn);
704 
705 		/*
706 		 * pfn could pass the block_end_pfn if isolated freepage
707 		 * is more than pageblock order. In this case, we adjust
708 		 * scanning range to right one.
709 		 */
710 		if (pfn >= block_end_pfn) {
711 			block_start_pfn = pageblock_start_pfn(pfn);
712 			block_end_pfn = pageblock_end_pfn(pfn);
713 			block_end_pfn = min(block_end_pfn, end_pfn);
714 		}
715 
716 		if (!pageblock_pfn_to_page(block_start_pfn,
717 					block_end_pfn, cc->zone))
718 			break;
719 
720 		isolated = isolate_freepages_block(cc, &isolate_start_pfn,
721 					block_end_pfn, &freelist, 0, true);
722 
723 		/*
724 		 * In strict mode, isolate_freepages_block() returns 0 if
725 		 * there are any holes in the block (ie. invalid PFNs or
726 		 * non-free pages).
727 		 */
728 		if (!isolated)
729 			break;
730 
731 		/*
732 		 * If we managed to isolate pages, it is always (1 << n) *
733 		 * pageblock_nr_pages for some non-negative n.  (Max order
734 		 * page may span two pageblocks).
735 		 */
736 	}
737 
738 	/* __isolate_free_page() does not map the pages */
739 	split_map_pages(&freelist);
740 
741 	if (pfn < end_pfn) {
742 		/* Loop terminated early, cleanup. */
743 		release_freepages(&freelist);
744 		return 0;
745 	}
746 
747 	/* We don't use freelists for anything. */
748 	return pfn;
749 }
750 
751 /* Similar to reclaim, but different enough that they don't share logic */
too_many_isolated(pg_data_t * pgdat)752 static bool too_many_isolated(pg_data_t *pgdat)
753 {
754 	bool too_many;
755 
756 	unsigned long active, inactive, isolated;
757 
758 	inactive = node_page_state(pgdat, NR_INACTIVE_FILE) +
759 			node_page_state(pgdat, NR_INACTIVE_ANON);
760 	active = node_page_state(pgdat, NR_ACTIVE_FILE) +
761 			node_page_state(pgdat, NR_ACTIVE_ANON);
762 	isolated = node_page_state(pgdat, NR_ISOLATED_FILE) +
763 			node_page_state(pgdat, NR_ISOLATED_ANON);
764 
765 	too_many = isolated > (inactive + active) / 2;
766 	if (!too_many)
767 		wake_throttle_isolated(pgdat);
768 
769 	return too_many;
770 }
771 
772 /**
773  * isolate_migratepages_block() - isolate all migrate-able pages within
774  *				  a single pageblock
775  * @cc:		Compaction control structure.
776  * @low_pfn:	The first PFN to isolate
777  * @end_pfn:	The one-past-the-last PFN to isolate, within same pageblock
778  * @mode:	Isolation mode to be used.
779  *
780  * Isolate all pages that can be migrated from the range specified by
781  * [low_pfn, end_pfn). The range is expected to be within same pageblock.
782  * Returns errno, like -EAGAIN or -EINTR in case e.g signal pending or congestion,
783  * -ENOMEM in case we could not allocate a page, or 0.
784  * cc->migrate_pfn will contain the next pfn to scan.
785  *
786  * The pages are isolated on cc->migratepages list (not required to be empty),
787  * and cc->nr_migratepages is updated accordingly.
788  */
789 static int
isolate_migratepages_block(struct compact_control * cc,unsigned long low_pfn,unsigned long end_pfn,isolate_mode_t mode)790 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
791 			unsigned long end_pfn, isolate_mode_t mode)
792 {
793 	pg_data_t *pgdat = cc->zone->zone_pgdat;
794 	unsigned long nr_scanned = 0, nr_isolated = 0;
795 	struct lruvec *lruvec;
796 	unsigned long flags = 0;
797 	struct lruvec *locked = NULL;
798 	struct page *page = NULL, *valid_page = NULL;
799 	struct address_space *mapping;
800 	unsigned long start_pfn = low_pfn;
801 	bool skip_on_failure = false;
802 	unsigned long next_skip_pfn = 0;
803 	bool skip_updated = false;
804 	int ret = 0;
805 
806 	cc->migrate_pfn = low_pfn;
807 
808 	/*
809 	 * Ensure that there are not too many pages isolated from the LRU
810 	 * list by either parallel reclaimers or compaction. If there are,
811 	 * delay for some time until fewer pages are isolated
812 	 */
813 	while (unlikely(too_many_isolated(pgdat))) {
814 		/* stop isolation if there are still pages not migrated */
815 		if (cc->nr_migratepages)
816 			return -EAGAIN;
817 
818 		/* async migration should just abort */
819 		if (cc->mode == MIGRATE_ASYNC)
820 			return -EAGAIN;
821 
822 		reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED);
823 
824 		if (fatal_signal_pending(current))
825 			return -EINTR;
826 	}
827 
828 	cond_resched();
829 
830 	if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
831 		skip_on_failure = true;
832 		next_skip_pfn = block_end_pfn(low_pfn, cc->order);
833 	}
834 
835 	/* Time to isolate some pages for migration */
836 	for (; low_pfn < end_pfn; low_pfn++) {
837 
838 		if (skip_on_failure && low_pfn >= next_skip_pfn) {
839 			/*
840 			 * We have isolated all migration candidates in the
841 			 * previous order-aligned block, and did not skip it due
842 			 * to failure. We should migrate the pages now and
843 			 * hopefully succeed compaction.
844 			 */
845 			if (nr_isolated)
846 				break;
847 
848 			/*
849 			 * We failed to isolate in the previous order-aligned
850 			 * block. Set the new boundary to the end of the
851 			 * current block. Note we can't simply increase
852 			 * next_skip_pfn by 1 << order, as low_pfn might have
853 			 * been incremented by a higher number due to skipping
854 			 * a compound or a high-order buddy page in the
855 			 * previous loop iteration.
856 			 */
857 			next_skip_pfn = block_end_pfn(low_pfn, cc->order);
858 		}
859 
860 		/*
861 		 * Periodically drop the lock (if held) regardless of its
862 		 * contention, to give chance to IRQs. Abort completely if
863 		 * a fatal signal is pending.
864 		 */
865 		if (!(low_pfn % COMPACT_CLUSTER_MAX)) {
866 			if (locked) {
867 				unlock_page_lruvec_irqrestore(locked, flags);
868 				locked = NULL;
869 			}
870 
871 			if (fatal_signal_pending(current)) {
872 				cc->contended = true;
873 				ret = -EINTR;
874 
875 				goto fatal_pending;
876 			}
877 
878 			cond_resched();
879 		}
880 
881 		nr_scanned++;
882 
883 		page = pfn_to_page(low_pfn);
884 
885 		/*
886 		 * Check if the pageblock has already been marked skipped.
887 		 * Only the aligned PFN is checked as the caller isolates
888 		 * COMPACT_CLUSTER_MAX at a time so the second call must
889 		 * not falsely conclude that the block should be skipped.
890 		 */
891 		if (!valid_page && IS_ALIGNED(low_pfn, pageblock_nr_pages)) {
892 			if (!isolation_suitable(cc, page)) {
893 				low_pfn = end_pfn;
894 				page = NULL;
895 				goto isolate_abort;
896 			}
897 			valid_page = page;
898 		}
899 
900 		if (PageHuge(page) && cc->alloc_contig) {
901 			ret = isolate_or_dissolve_huge_page(page, &cc->migratepages);
902 
903 			/*
904 			 * Fail isolation in case isolate_or_dissolve_huge_page()
905 			 * reports an error. In case of -ENOMEM, abort right away.
906 			 */
907 			if (ret < 0) {
908 				 /* Do not report -EBUSY down the chain */
909 				if (ret == -EBUSY)
910 					ret = 0;
911 				low_pfn += compound_nr(page) - 1;
912 				goto isolate_fail;
913 			}
914 
915 			if (PageHuge(page)) {
916 				/*
917 				 * Hugepage was successfully isolated and placed
918 				 * on the cc->migratepages list.
919 				 */
920 				low_pfn += compound_nr(page) - 1;
921 				goto isolate_success_no_list;
922 			}
923 
924 			/*
925 			 * Ok, the hugepage was dissolved. Now these pages are
926 			 * Buddy and cannot be re-allocated because they are
927 			 * isolated. Fall-through as the check below handles
928 			 * Buddy pages.
929 			 */
930 		}
931 
932 		/*
933 		 * Skip if free. We read page order here without zone lock
934 		 * which is generally unsafe, but the race window is small and
935 		 * the worst thing that can happen is that we skip some
936 		 * potential isolation targets.
937 		 */
938 		if (PageBuddy(page)) {
939 			unsigned long freepage_order = buddy_order_unsafe(page);
940 
941 			/*
942 			 * Without lock, we cannot be sure that what we got is
943 			 * a valid page order. Consider only values in the
944 			 * valid order range to prevent low_pfn overflow.
945 			 */
946 			if (freepage_order > 0 && freepage_order < MAX_ORDER)
947 				low_pfn += (1UL << freepage_order) - 1;
948 			continue;
949 		}
950 
951 		/*
952 		 * Regardless of being on LRU, compound pages such as THP and
953 		 * hugetlbfs are not to be compacted unless we are attempting
954 		 * an allocation much larger than the huge page size (eg CMA).
955 		 * We can potentially save a lot of iterations if we skip them
956 		 * at once. The check is racy, but we can consider only valid
957 		 * values and the only danger is skipping too much.
958 		 */
959 		if (PageCompound(page) && !cc->alloc_contig) {
960 			const unsigned int order = compound_order(page);
961 
962 			if (likely(order < MAX_ORDER))
963 				low_pfn += (1UL << order) - 1;
964 			goto isolate_fail;
965 		}
966 
967 		/*
968 		 * Check may be lockless but that's ok as we recheck later.
969 		 * It's possible to migrate LRU and non-lru movable pages.
970 		 * Skip any other type of page
971 		 */
972 		if (!PageLRU(page)) {
973 			/*
974 			 * __PageMovable can return false positive so we need
975 			 * to verify it under page_lock.
976 			 */
977 			if (unlikely(__PageMovable(page)) &&
978 					!PageIsolated(page)) {
979 				if (locked) {
980 					unlock_page_lruvec_irqrestore(locked, flags);
981 					locked = NULL;
982 				}
983 
984 				if (!isolate_movable_page(page, mode))
985 					goto isolate_success;
986 			}
987 
988 			goto isolate_fail;
989 		}
990 
991 		/*
992 		 * Migration will fail if an anonymous page is pinned in memory,
993 		 * so avoid taking lru_lock and isolating it unnecessarily in an
994 		 * admittedly racy check.
995 		 */
996 		mapping = page_mapping(page);
997 		if (!mapping && page_count(page) > page_mapcount(page))
998 			goto isolate_fail;
999 
1000 		/*
1001 		 * Only allow to migrate anonymous pages in GFP_NOFS context
1002 		 * because those do not depend on fs locks.
1003 		 */
1004 		if (!(cc->gfp_mask & __GFP_FS) && mapping)
1005 			goto isolate_fail;
1006 
1007 		/*
1008 		 * Be careful not to clear PageLRU until after we're
1009 		 * sure the page is not being freed elsewhere -- the
1010 		 * page release code relies on it.
1011 		 */
1012 		if (unlikely(!get_page_unless_zero(page)))
1013 			goto isolate_fail;
1014 
1015 		/* Only take pages on LRU: a check now makes later tests safe */
1016 		if (!PageLRU(page))
1017 			goto isolate_fail_put;
1018 
1019 		/* Compaction might skip unevictable pages but CMA takes them */
1020 		if (!(mode & ISOLATE_UNEVICTABLE) && PageUnevictable(page))
1021 			goto isolate_fail_put;
1022 
1023 		/*
1024 		 * To minimise LRU disruption, the caller can indicate with
1025 		 * ISOLATE_ASYNC_MIGRATE that it only wants to isolate pages
1026 		 * it will be able to migrate without blocking - clean pages
1027 		 * for the most part.  PageWriteback would require blocking.
1028 		 */
1029 		if ((mode & ISOLATE_ASYNC_MIGRATE) && PageWriteback(page))
1030 			goto isolate_fail_put;
1031 
1032 		if ((mode & ISOLATE_ASYNC_MIGRATE) && PageDirty(page)) {
1033 			bool migrate_dirty;
1034 
1035 			/*
1036 			 * Only pages without mappings or that have a
1037 			 * ->migratepage callback are possible to migrate
1038 			 * without blocking. However, we can be racing with
1039 			 * truncation so it's necessary to lock the page
1040 			 * to stabilise the mapping as truncation holds
1041 			 * the page lock until after the page is removed
1042 			 * from the page cache.
1043 			 */
1044 			if (!trylock_page(page))
1045 				goto isolate_fail_put;
1046 
1047 			mapping = page_mapping(page);
1048 			migrate_dirty = !mapping || mapping->a_ops->migratepage;
1049 			unlock_page(page);
1050 			if (!migrate_dirty)
1051 				goto isolate_fail_put;
1052 		}
1053 
1054 		/* Try isolate the page */
1055 		if (!TestClearPageLRU(page))
1056 			goto isolate_fail_put;
1057 
1058 		lruvec = folio_lruvec(page_folio(page));
1059 
1060 		/* If we already hold the lock, we can skip some rechecking */
1061 		if (lruvec != locked) {
1062 			if (locked)
1063 				unlock_page_lruvec_irqrestore(locked, flags);
1064 
1065 			compact_lock_irqsave(&lruvec->lru_lock, &flags, cc);
1066 			locked = lruvec;
1067 
1068 			lruvec_memcg_debug(lruvec, page_folio(page));
1069 
1070 			/* Try get exclusive access under lock */
1071 			if (!skip_updated) {
1072 				skip_updated = true;
1073 				if (test_and_set_skip(cc, page, low_pfn))
1074 					goto isolate_abort;
1075 			}
1076 
1077 			/*
1078 			 * Page become compound since the non-locked check,
1079 			 * and it's on LRU. It can only be a THP so the order
1080 			 * is safe to read and it's 0 for tail pages.
1081 			 */
1082 			if (unlikely(PageCompound(page) && !cc->alloc_contig)) {
1083 				low_pfn += compound_nr(page) - 1;
1084 				SetPageLRU(page);
1085 				goto isolate_fail_put;
1086 			}
1087 		}
1088 
1089 		/* The whole page is taken off the LRU; skip the tail pages. */
1090 		if (PageCompound(page))
1091 			low_pfn += compound_nr(page) - 1;
1092 
1093 		/* Successfully isolated */
1094 		del_page_from_lru_list(page, lruvec);
1095 		mod_node_page_state(page_pgdat(page),
1096 				NR_ISOLATED_ANON + page_is_file_lru(page),
1097 				thp_nr_pages(page));
1098 
1099 isolate_success:
1100 		list_add(&page->lru, &cc->migratepages);
1101 isolate_success_no_list:
1102 		cc->nr_migratepages += compound_nr(page);
1103 		nr_isolated += compound_nr(page);
1104 
1105 		/*
1106 		 * Avoid isolating too much unless this block is being
1107 		 * rescanned (e.g. dirty/writeback pages, parallel allocation)
1108 		 * or a lock is contended. For contention, isolate quickly to
1109 		 * potentially remove one source of contention.
1110 		 */
1111 		if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX &&
1112 		    !cc->rescan && !cc->contended) {
1113 			++low_pfn;
1114 			break;
1115 		}
1116 
1117 		continue;
1118 
1119 isolate_fail_put:
1120 		/* Avoid potential deadlock in freeing page under lru_lock */
1121 		if (locked) {
1122 			unlock_page_lruvec_irqrestore(locked, flags);
1123 			locked = NULL;
1124 		}
1125 		put_page(page);
1126 
1127 isolate_fail:
1128 		if (!skip_on_failure && ret != -ENOMEM)
1129 			continue;
1130 
1131 		/*
1132 		 * We have isolated some pages, but then failed. Release them
1133 		 * instead of migrating, as we cannot form the cc->order buddy
1134 		 * page anyway.
1135 		 */
1136 		if (nr_isolated) {
1137 			if (locked) {
1138 				unlock_page_lruvec_irqrestore(locked, flags);
1139 				locked = NULL;
1140 			}
1141 			putback_movable_pages(&cc->migratepages);
1142 			cc->nr_migratepages = 0;
1143 			nr_isolated = 0;
1144 		}
1145 
1146 		if (low_pfn < next_skip_pfn) {
1147 			low_pfn = next_skip_pfn - 1;
1148 			/*
1149 			 * The check near the loop beginning would have updated
1150 			 * next_skip_pfn too, but this is a bit simpler.
1151 			 */
1152 			next_skip_pfn += 1UL << cc->order;
1153 		}
1154 
1155 		if (ret == -ENOMEM)
1156 			break;
1157 	}
1158 
1159 	/*
1160 	 * The PageBuddy() check could have potentially brought us outside
1161 	 * the range to be scanned.
1162 	 */
1163 	if (unlikely(low_pfn > end_pfn))
1164 		low_pfn = end_pfn;
1165 
1166 	page = NULL;
1167 
1168 isolate_abort:
1169 	if (locked)
1170 		unlock_page_lruvec_irqrestore(locked, flags);
1171 	if (page) {
1172 		SetPageLRU(page);
1173 		put_page(page);
1174 	}
1175 
1176 	/*
1177 	 * Updated the cached scanner pfn once the pageblock has been scanned
1178 	 * Pages will either be migrated in which case there is no point
1179 	 * scanning in the near future or migration failed in which case the
1180 	 * failure reason may persist. The block is marked for skipping if
1181 	 * there were no pages isolated in the block or if the block is
1182 	 * rescanned twice in a row.
1183 	 */
1184 	if (low_pfn == end_pfn && (!nr_isolated || cc->rescan)) {
1185 		if (valid_page && !skip_updated)
1186 			set_pageblock_skip(valid_page);
1187 		update_cached_migrate(cc, low_pfn);
1188 	}
1189 
1190 	trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
1191 						nr_scanned, nr_isolated);
1192 
1193 fatal_pending:
1194 	cc->total_migrate_scanned += nr_scanned;
1195 	if (nr_isolated)
1196 		count_compact_events(COMPACTISOLATED, nr_isolated);
1197 
1198 	cc->migrate_pfn = low_pfn;
1199 
1200 	return ret;
1201 }
1202 
1203 /**
1204  * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
1205  * @cc:        Compaction control structure.
1206  * @start_pfn: The first PFN to start isolating.
1207  * @end_pfn:   The one-past-last PFN.
1208  *
1209  * Returns -EAGAIN when contented, -EINTR in case of a signal pending, -ENOMEM
1210  * in case we could not allocate a page, or 0.
1211  */
1212 int
isolate_migratepages_range(struct compact_control * cc,unsigned long start_pfn,unsigned long end_pfn)1213 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
1214 							unsigned long end_pfn)
1215 {
1216 	unsigned long pfn, block_start_pfn, block_end_pfn;
1217 	int ret = 0;
1218 
1219 	/* Scan block by block. First and last block may be incomplete */
1220 	pfn = start_pfn;
1221 	block_start_pfn = pageblock_start_pfn(pfn);
1222 	if (block_start_pfn < cc->zone->zone_start_pfn)
1223 		block_start_pfn = cc->zone->zone_start_pfn;
1224 	block_end_pfn = pageblock_end_pfn(pfn);
1225 
1226 	for (; pfn < end_pfn; pfn = block_end_pfn,
1227 				block_start_pfn = block_end_pfn,
1228 				block_end_pfn += pageblock_nr_pages) {
1229 
1230 		block_end_pfn = min(block_end_pfn, end_pfn);
1231 
1232 		if (!pageblock_pfn_to_page(block_start_pfn,
1233 					block_end_pfn, cc->zone))
1234 			continue;
1235 
1236 		ret = isolate_migratepages_block(cc, pfn, block_end_pfn,
1237 						 ISOLATE_UNEVICTABLE);
1238 
1239 		if (ret)
1240 			break;
1241 
1242 		if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX)
1243 			break;
1244 	}
1245 
1246 	return ret;
1247 }
1248 
1249 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
1250 #ifdef CONFIG_COMPACTION
1251 
suitable_migration_source(struct compact_control * cc,struct page * page)1252 static bool suitable_migration_source(struct compact_control *cc,
1253 							struct page *page)
1254 {
1255 	int block_mt;
1256 
1257 	if (pageblock_skip_persistent(page))
1258 		return false;
1259 
1260 	if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction)
1261 		return true;
1262 
1263 	block_mt = get_pageblock_migratetype(page);
1264 
1265 	if (cc->migratetype == MIGRATE_MOVABLE)
1266 		return is_migrate_movable(block_mt);
1267 	else
1268 		return block_mt == cc->migratetype;
1269 }
1270 
1271 /* Returns true if the page is within a block suitable for migration to */
suitable_migration_target(struct compact_control * cc,struct page * page)1272 static bool suitable_migration_target(struct compact_control *cc,
1273 							struct page *page)
1274 {
1275 	/* If the page is a large free page, then disallow migration */
1276 	if (PageBuddy(page)) {
1277 		/*
1278 		 * We are checking page_order without zone->lock taken. But
1279 		 * the only small danger is that we skip a potentially suitable
1280 		 * pageblock, so it's not worth to check order for valid range.
1281 		 */
1282 		if (buddy_order_unsafe(page) >= pageblock_order)
1283 			return false;
1284 	}
1285 
1286 	if (cc->ignore_block_suitable)
1287 		return true;
1288 
1289 	/* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1290 	if (is_migrate_movable(get_pageblock_migratetype(page)))
1291 		return true;
1292 
1293 	/* Otherwise skip the block */
1294 	return false;
1295 }
1296 
1297 static inline unsigned int
freelist_scan_limit(struct compact_control * cc)1298 freelist_scan_limit(struct compact_control *cc)
1299 {
1300 	unsigned short shift = BITS_PER_LONG - 1;
1301 
1302 	return (COMPACT_CLUSTER_MAX >> min(shift, cc->fast_search_fail)) + 1;
1303 }
1304 
1305 /*
1306  * Test whether the free scanner has reached the same or lower pageblock than
1307  * the migration scanner, and compaction should thus terminate.
1308  */
compact_scanners_met(struct compact_control * cc)1309 static inline bool compact_scanners_met(struct compact_control *cc)
1310 {
1311 	return (cc->free_pfn >> pageblock_order)
1312 		<= (cc->migrate_pfn >> pageblock_order);
1313 }
1314 
1315 /*
1316  * Used when scanning for a suitable migration target which scans freelists
1317  * in reverse. Reorders the list such as the unscanned pages are scanned
1318  * first on the next iteration of the free scanner
1319  */
1320 static void
move_freelist_head(struct list_head * freelist,struct page * freepage)1321 move_freelist_head(struct list_head *freelist, struct page *freepage)
1322 {
1323 	LIST_HEAD(sublist);
1324 
1325 	if (!list_is_last(freelist, &freepage->lru)) {
1326 		list_cut_before(&sublist, freelist, &freepage->lru);
1327 		list_splice_tail(&sublist, freelist);
1328 	}
1329 }
1330 
1331 /*
1332  * Similar to move_freelist_head except used by the migration scanner
1333  * when scanning forward. It's possible for these list operations to
1334  * move against each other if they search the free list exactly in
1335  * lockstep.
1336  */
1337 static void
move_freelist_tail(struct list_head * freelist,struct page * freepage)1338 move_freelist_tail(struct list_head *freelist, struct page *freepage)
1339 {
1340 	LIST_HEAD(sublist);
1341 
1342 	if (!list_is_first(freelist, &freepage->lru)) {
1343 		list_cut_position(&sublist, freelist, &freepage->lru);
1344 		list_splice_tail(&sublist, freelist);
1345 	}
1346 }
1347 
1348 static void
fast_isolate_around(struct compact_control * cc,unsigned long pfn,unsigned long nr_isolated)1349 fast_isolate_around(struct compact_control *cc, unsigned long pfn, unsigned long nr_isolated)
1350 {
1351 	unsigned long start_pfn, end_pfn;
1352 	struct page *page;
1353 
1354 	/* Do not search around if there are enough pages already */
1355 	if (cc->nr_freepages >= cc->nr_migratepages)
1356 		return;
1357 
1358 	/* Minimise scanning during async compaction */
1359 	if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC)
1360 		return;
1361 
1362 	/* Pageblock boundaries */
1363 	start_pfn = max(pageblock_start_pfn(pfn), cc->zone->zone_start_pfn);
1364 	end_pfn = min(pageblock_end_pfn(pfn), zone_end_pfn(cc->zone));
1365 
1366 	page = pageblock_pfn_to_page(start_pfn, end_pfn, cc->zone);
1367 	if (!page)
1368 		return;
1369 
1370 	/* Scan before */
1371 	if (start_pfn != pfn) {
1372 		isolate_freepages_block(cc, &start_pfn, pfn, &cc->freepages, 1, false);
1373 		if (cc->nr_freepages >= cc->nr_migratepages)
1374 			return;
1375 	}
1376 
1377 	/* Scan after */
1378 	start_pfn = pfn + nr_isolated;
1379 	if (start_pfn < end_pfn)
1380 		isolate_freepages_block(cc, &start_pfn, end_pfn, &cc->freepages, 1, false);
1381 
1382 	/* Skip this pageblock in the future as it's full or nearly full */
1383 	if (cc->nr_freepages < cc->nr_migratepages)
1384 		set_pageblock_skip(page);
1385 }
1386 
1387 /* Search orders in round-robin fashion */
next_search_order(struct compact_control * cc,int order)1388 static int next_search_order(struct compact_control *cc, int order)
1389 {
1390 	order--;
1391 	if (order < 0)
1392 		order = cc->order - 1;
1393 
1394 	/* Search wrapped around? */
1395 	if (order == cc->search_order) {
1396 		cc->search_order--;
1397 		if (cc->search_order < 0)
1398 			cc->search_order = cc->order - 1;
1399 		return -1;
1400 	}
1401 
1402 	return order;
1403 }
1404 
1405 static unsigned long
fast_isolate_freepages(struct compact_control * cc)1406 fast_isolate_freepages(struct compact_control *cc)
1407 {
1408 	unsigned int limit = max(1U, freelist_scan_limit(cc) >> 1);
1409 	unsigned int nr_scanned = 0;
1410 	unsigned long low_pfn, min_pfn, highest = 0;
1411 	unsigned long nr_isolated = 0;
1412 	unsigned long distance;
1413 	struct page *page = NULL;
1414 	bool scan_start = false;
1415 	int order;
1416 
1417 	/* Full compaction passes in a negative order */
1418 	if (cc->order <= 0)
1419 		return cc->free_pfn;
1420 
1421 	/*
1422 	 * If starting the scan, use a deeper search and use the highest
1423 	 * PFN found if a suitable one is not found.
1424 	 */
1425 	if (cc->free_pfn >= cc->zone->compact_init_free_pfn) {
1426 		limit = pageblock_nr_pages >> 1;
1427 		scan_start = true;
1428 	}
1429 
1430 	/*
1431 	 * Preferred point is in the top quarter of the scan space but take
1432 	 * a pfn from the top half if the search is problematic.
1433 	 */
1434 	distance = (cc->free_pfn - cc->migrate_pfn);
1435 	low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2));
1436 	min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1));
1437 
1438 	if (WARN_ON_ONCE(min_pfn > low_pfn))
1439 		low_pfn = min_pfn;
1440 
1441 	/*
1442 	 * Search starts from the last successful isolation order or the next
1443 	 * order to search after a previous failure
1444 	 */
1445 	cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order);
1446 
1447 	for (order = cc->search_order;
1448 	     !page && order >= 0;
1449 	     order = next_search_order(cc, order)) {
1450 		struct free_area *area = &cc->zone->free_area[order];
1451 		struct list_head *freelist;
1452 		struct page *freepage;
1453 		unsigned long flags;
1454 		unsigned int order_scanned = 0;
1455 		unsigned long high_pfn = 0;
1456 
1457 		if (!area->nr_free)
1458 			continue;
1459 
1460 		spin_lock_irqsave(&cc->zone->lock, flags);
1461 		freelist = &area->free_list[MIGRATE_MOVABLE];
1462 		list_for_each_entry_reverse(freepage, freelist, lru) {
1463 			unsigned long pfn;
1464 
1465 			order_scanned++;
1466 			nr_scanned++;
1467 			pfn = page_to_pfn(freepage);
1468 
1469 			if (pfn >= highest)
1470 				highest = max(pageblock_start_pfn(pfn),
1471 					      cc->zone->zone_start_pfn);
1472 
1473 			if (pfn >= low_pfn) {
1474 				cc->fast_search_fail = 0;
1475 				cc->search_order = order;
1476 				page = freepage;
1477 				break;
1478 			}
1479 
1480 			if (pfn >= min_pfn && pfn > high_pfn) {
1481 				high_pfn = pfn;
1482 
1483 				/* Shorten the scan if a candidate is found */
1484 				limit >>= 1;
1485 			}
1486 
1487 			if (order_scanned >= limit)
1488 				break;
1489 		}
1490 
1491 		/* Use a minimum pfn if a preferred one was not found */
1492 		if (!page && high_pfn) {
1493 			page = pfn_to_page(high_pfn);
1494 
1495 			/* Update freepage for the list reorder below */
1496 			freepage = page;
1497 		}
1498 
1499 		/* Reorder to so a future search skips recent pages */
1500 		move_freelist_head(freelist, freepage);
1501 
1502 		/* Isolate the page if available */
1503 		if (page) {
1504 			if (__isolate_free_page(page, order)) {
1505 				set_page_private(page, order);
1506 				nr_isolated = 1 << order;
1507 				cc->nr_freepages += nr_isolated;
1508 				list_add_tail(&page->lru, &cc->freepages);
1509 				count_compact_events(COMPACTISOLATED, nr_isolated);
1510 			} else {
1511 				/* If isolation fails, abort the search */
1512 				order = cc->search_order + 1;
1513 				page = NULL;
1514 			}
1515 		}
1516 
1517 		spin_unlock_irqrestore(&cc->zone->lock, flags);
1518 
1519 		/*
1520 		 * Smaller scan on next order so the total scan is related
1521 		 * to freelist_scan_limit.
1522 		 */
1523 		if (order_scanned >= limit)
1524 			limit = max(1U, limit >> 1);
1525 	}
1526 
1527 	if (!page) {
1528 		cc->fast_search_fail++;
1529 		if (scan_start) {
1530 			/*
1531 			 * Use the highest PFN found above min. If one was
1532 			 * not found, be pessimistic for direct compaction
1533 			 * and use the min mark.
1534 			 */
1535 			if (highest >= min_pfn) {
1536 				page = pfn_to_page(highest);
1537 				cc->free_pfn = highest;
1538 			} else {
1539 				if (cc->direct_compaction && pfn_valid(min_pfn)) {
1540 					page = pageblock_pfn_to_page(min_pfn,
1541 						min(pageblock_end_pfn(min_pfn),
1542 						    zone_end_pfn(cc->zone)),
1543 						cc->zone);
1544 					cc->free_pfn = min_pfn;
1545 				}
1546 			}
1547 		}
1548 	}
1549 
1550 	if (highest && highest >= cc->zone->compact_cached_free_pfn) {
1551 		highest -= pageblock_nr_pages;
1552 		cc->zone->compact_cached_free_pfn = highest;
1553 	}
1554 
1555 	cc->total_free_scanned += nr_scanned;
1556 	if (!page)
1557 		return cc->free_pfn;
1558 
1559 	low_pfn = page_to_pfn(page);
1560 	fast_isolate_around(cc, low_pfn, nr_isolated);
1561 	return low_pfn;
1562 }
1563 
1564 /*
1565  * Based on information in the current compact_control, find blocks
1566  * suitable for isolating free pages from and then isolate them.
1567  */
isolate_freepages(struct compact_control * cc)1568 static void isolate_freepages(struct compact_control *cc)
1569 {
1570 	struct zone *zone = cc->zone;
1571 	struct page *page;
1572 	unsigned long block_start_pfn;	/* start of current pageblock */
1573 	unsigned long isolate_start_pfn; /* exact pfn we start at */
1574 	unsigned long block_end_pfn;	/* end of current pageblock */
1575 	unsigned long low_pfn;	     /* lowest pfn scanner is able to scan */
1576 	struct list_head *freelist = &cc->freepages;
1577 	unsigned int stride;
1578 
1579 	/* Try a small search of the free lists for a candidate */
1580 	fast_isolate_freepages(cc);
1581 	if (cc->nr_freepages)
1582 		goto splitmap;
1583 
1584 	/*
1585 	 * Initialise the free scanner. The starting point is where we last
1586 	 * successfully isolated from, zone-cached value, or the end of the
1587 	 * zone when isolating for the first time. For looping we also need
1588 	 * this pfn aligned down to the pageblock boundary, because we do
1589 	 * block_start_pfn -= pageblock_nr_pages in the for loop.
1590 	 * For ending point, take care when isolating in last pageblock of a
1591 	 * zone which ends in the middle of a pageblock.
1592 	 * The low boundary is the end of the pageblock the migration scanner
1593 	 * is using.
1594 	 */
1595 	isolate_start_pfn = cc->free_pfn;
1596 	block_start_pfn = pageblock_start_pfn(isolate_start_pfn);
1597 	block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1598 						zone_end_pfn(zone));
1599 	low_pfn = pageblock_end_pfn(cc->migrate_pfn);
1600 	stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1;
1601 
1602 	/*
1603 	 * Isolate free pages until enough are available to migrate the
1604 	 * pages on cc->migratepages. We stop searching if the migrate
1605 	 * and free page scanners meet or enough free pages are isolated.
1606 	 */
1607 	for (; block_start_pfn >= low_pfn;
1608 				block_end_pfn = block_start_pfn,
1609 				block_start_pfn -= pageblock_nr_pages,
1610 				isolate_start_pfn = block_start_pfn) {
1611 		unsigned long nr_isolated;
1612 
1613 		/*
1614 		 * This can iterate a massively long zone without finding any
1615 		 * suitable migration targets, so periodically check resched.
1616 		 */
1617 		if (!(block_start_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages)))
1618 			cond_resched();
1619 
1620 		page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1621 									zone);
1622 		if (!page)
1623 			continue;
1624 
1625 		/* Check the block is suitable for migration */
1626 		if (!suitable_migration_target(cc, page))
1627 			continue;
1628 
1629 		/* If isolation recently failed, do not retry */
1630 		if (!isolation_suitable(cc, page))
1631 			continue;
1632 
1633 		/* Found a block suitable for isolating free pages from. */
1634 		nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn,
1635 					block_end_pfn, freelist, stride, false);
1636 
1637 		/* Update the skip hint if the full pageblock was scanned */
1638 		if (isolate_start_pfn == block_end_pfn)
1639 			update_pageblock_skip(cc, page, block_start_pfn);
1640 
1641 		/* Are enough freepages isolated? */
1642 		if (cc->nr_freepages >= cc->nr_migratepages) {
1643 			if (isolate_start_pfn >= block_end_pfn) {
1644 				/*
1645 				 * Restart at previous pageblock if more
1646 				 * freepages can be isolated next time.
1647 				 */
1648 				isolate_start_pfn =
1649 					block_start_pfn - pageblock_nr_pages;
1650 			}
1651 			break;
1652 		} else if (isolate_start_pfn < block_end_pfn) {
1653 			/*
1654 			 * If isolation failed early, do not continue
1655 			 * needlessly.
1656 			 */
1657 			break;
1658 		}
1659 
1660 		/* Adjust stride depending on isolation */
1661 		if (nr_isolated) {
1662 			stride = 1;
1663 			continue;
1664 		}
1665 		stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1);
1666 	}
1667 
1668 	/*
1669 	 * Record where the free scanner will restart next time. Either we
1670 	 * broke from the loop and set isolate_start_pfn based on the last
1671 	 * call to isolate_freepages_block(), or we met the migration scanner
1672 	 * and the loop terminated due to isolate_start_pfn < low_pfn
1673 	 */
1674 	cc->free_pfn = isolate_start_pfn;
1675 
1676 splitmap:
1677 	/* __isolate_free_page() does not map the pages */
1678 	split_map_pages(freelist);
1679 }
1680 
1681 /*
1682  * This is a migrate-callback that "allocates" freepages by taking pages
1683  * from the isolated freelists in the block we are migrating to.
1684  */
compaction_alloc(struct page * migratepage,unsigned long data)1685 static struct page *compaction_alloc(struct page *migratepage,
1686 					unsigned long data)
1687 {
1688 	struct compact_control *cc = (struct compact_control *)data;
1689 	struct page *freepage;
1690 
1691 	if (list_empty(&cc->freepages)) {
1692 		isolate_freepages(cc);
1693 
1694 		if (list_empty(&cc->freepages))
1695 			return NULL;
1696 	}
1697 
1698 	freepage = list_entry(cc->freepages.next, struct page, lru);
1699 	list_del(&freepage->lru);
1700 	cc->nr_freepages--;
1701 
1702 	return freepage;
1703 }
1704 
1705 /*
1706  * This is a migrate-callback that "frees" freepages back to the isolated
1707  * freelist.  All pages on the freelist are from the same zone, so there is no
1708  * special handling needed for NUMA.
1709  */
compaction_free(struct page * page,unsigned long data)1710 static void compaction_free(struct page *page, unsigned long data)
1711 {
1712 	struct compact_control *cc = (struct compact_control *)data;
1713 
1714 	list_add(&page->lru, &cc->freepages);
1715 	cc->nr_freepages++;
1716 }
1717 
1718 /* possible outcome of isolate_migratepages */
1719 typedef enum {
1720 	ISOLATE_ABORT,		/* Abort compaction now */
1721 	ISOLATE_NONE,		/* No pages isolated, continue scanning */
1722 	ISOLATE_SUCCESS,	/* Pages isolated, migrate */
1723 } isolate_migrate_t;
1724 
1725 /*
1726  * Allow userspace to control policy on scanning the unevictable LRU for
1727  * compactable pages.
1728  */
1729 #ifdef CONFIG_PREEMPT_RT
1730 int sysctl_compact_unevictable_allowed __read_mostly = 0;
1731 #else
1732 int sysctl_compact_unevictable_allowed __read_mostly = 1;
1733 #endif
1734 
1735 static inline void
update_fast_start_pfn(struct compact_control * cc,unsigned long pfn)1736 update_fast_start_pfn(struct compact_control *cc, unsigned long pfn)
1737 {
1738 	if (cc->fast_start_pfn == ULONG_MAX)
1739 		return;
1740 
1741 	if (!cc->fast_start_pfn)
1742 		cc->fast_start_pfn = pfn;
1743 
1744 	cc->fast_start_pfn = min(cc->fast_start_pfn, pfn);
1745 }
1746 
1747 static inline unsigned long
reinit_migrate_pfn(struct compact_control * cc)1748 reinit_migrate_pfn(struct compact_control *cc)
1749 {
1750 	if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX)
1751 		return cc->migrate_pfn;
1752 
1753 	cc->migrate_pfn = cc->fast_start_pfn;
1754 	cc->fast_start_pfn = ULONG_MAX;
1755 
1756 	return cc->migrate_pfn;
1757 }
1758 
1759 /*
1760  * Briefly search the free lists for a migration source that already has
1761  * some free pages to reduce the number of pages that need migration
1762  * before a pageblock is free.
1763  */
fast_find_migrateblock(struct compact_control * cc)1764 static unsigned long fast_find_migrateblock(struct compact_control *cc)
1765 {
1766 	unsigned int limit = freelist_scan_limit(cc);
1767 	unsigned int nr_scanned = 0;
1768 	unsigned long distance;
1769 	unsigned long pfn = cc->migrate_pfn;
1770 	unsigned long high_pfn;
1771 	int order;
1772 	bool found_block = false;
1773 
1774 	/* Skip hints are relied on to avoid repeats on the fast search */
1775 	if (cc->ignore_skip_hint)
1776 		return pfn;
1777 
1778 	/*
1779 	 * If the migrate_pfn is not at the start of a zone or the start
1780 	 * of a pageblock then assume this is a continuation of a previous
1781 	 * scan restarted due to COMPACT_CLUSTER_MAX.
1782 	 */
1783 	if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn))
1784 		return pfn;
1785 
1786 	/*
1787 	 * For smaller orders, just linearly scan as the number of pages
1788 	 * to migrate should be relatively small and does not necessarily
1789 	 * justify freeing up a large block for a small allocation.
1790 	 */
1791 	if (cc->order <= PAGE_ALLOC_COSTLY_ORDER)
1792 		return pfn;
1793 
1794 	/*
1795 	 * Only allow kcompactd and direct requests for movable pages to
1796 	 * quickly clear out a MOVABLE pageblock for allocation. This
1797 	 * reduces the risk that a large movable pageblock is freed for
1798 	 * an unmovable/reclaimable small allocation.
1799 	 */
1800 	if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE)
1801 		return pfn;
1802 
1803 	/*
1804 	 * When starting the migration scanner, pick any pageblock within the
1805 	 * first half of the search space. Otherwise try and pick a pageblock
1806 	 * within the first eighth to reduce the chances that a migration
1807 	 * target later becomes a source.
1808 	 */
1809 	distance = (cc->free_pfn - cc->migrate_pfn) >> 1;
1810 	if (cc->migrate_pfn != cc->zone->zone_start_pfn)
1811 		distance >>= 2;
1812 	high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance);
1813 
1814 	for (order = cc->order - 1;
1815 	     order >= PAGE_ALLOC_COSTLY_ORDER && !found_block && nr_scanned < limit;
1816 	     order--) {
1817 		struct free_area *area = &cc->zone->free_area[order];
1818 		struct list_head *freelist;
1819 		unsigned long flags;
1820 		struct page *freepage;
1821 
1822 		if (!area->nr_free)
1823 			continue;
1824 
1825 		spin_lock_irqsave(&cc->zone->lock, flags);
1826 		freelist = &area->free_list[MIGRATE_MOVABLE];
1827 		list_for_each_entry(freepage, freelist, lru) {
1828 			unsigned long free_pfn;
1829 
1830 			if (nr_scanned++ >= limit) {
1831 				move_freelist_tail(freelist, freepage);
1832 				break;
1833 			}
1834 
1835 			free_pfn = page_to_pfn(freepage);
1836 			if (free_pfn < high_pfn) {
1837 				/*
1838 				 * Avoid if skipped recently. Ideally it would
1839 				 * move to the tail but even safe iteration of
1840 				 * the list assumes an entry is deleted, not
1841 				 * reordered.
1842 				 */
1843 				if (get_pageblock_skip(freepage))
1844 					continue;
1845 
1846 				/* Reorder to so a future search skips recent pages */
1847 				move_freelist_tail(freelist, freepage);
1848 
1849 				update_fast_start_pfn(cc, free_pfn);
1850 				pfn = pageblock_start_pfn(free_pfn);
1851 				if (pfn < cc->zone->zone_start_pfn)
1852 					pfn = cc->zone->zone_start_pfn;
1853 				cc->fast_search_fail = 0;
1854 				found_block = true;
1855 				set_pageblock_skip(freepage);
1856 				break;
1857 			}
1858 		}
1859 		spin_unlock_irqrestore(&cc->zone->lock, flags);
1860 	}
1861 
1862 	cc->total_migrate_scanned += nr_scanned;
1863 
1864 	/*
1865 	 * If fast scanning failed then use a cached entry for a page block
1866 	 * that had free pages as the basis for starting a linear scan.
1867 	 */
1868 	if (!found_block) {
1869 		cc->fast_search_fail++;
1870 		pfn = reinit_migrate_pfn(cc);
1871 	}
1872 	return pfn;
1873 }
1874 
1875 /*
1876  * Isolate all pages that can be migrated from the first suitable block,
1877  * starting at the block pointed to by the migrate scanner pfn within
1878  * compact_control.
1879  */
isolate_migratepages(struct compact_control * cc)1880 static isolate_migrate_t isolate_migratepages(struct compact_control *cc)
1881 {
1882 	unsigned long block_start_pfn;
1883 	unsigned long block_end_pfn;
1884 	unsigned long low_pfn;
1885 	struct page *page;
1886 	const isolate_mode_t isolate_mode =
1887 		(sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1888 		(cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1889 	bool fast_find_block;
1890 
1891 	/*
1892 	 * Start at where we last stopped, or beginning of the zone as
1893 	 * initialized by compact_zone(). The first failure will use
1894 	 * the lowest PFN as the starting point for linear scanning.
1895 	 */
1896 	low_pfn = fast_find_migrateblock(cc);
1897 	block_start_pfn = pageblock_start_pfn(low_pfn);
1898 	if (block_start_pfn < cc->zone->zone_start_pfn)
1899 		block_start_pfn = cc->zone->zone_start_pfn;
1900 
1901 	/*
1902 	 * fast_find_migrateblock marks a pageblock skipped so to avoid
1903 	 * the isolation_suitable check below, check whether the fast
1904 	 * search was successful.
1905 	 */
1906 	fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail;
1907 
1908 	/* Only scan within a pageblock boundary */
1909 	block_end_pfn = pageblock_end_pfn(low_pfn);
1910 
1911 	/*
1912 	 * Iterate over whole pageblocks until we find the first suitable.
1913 	 * Do not cross the free scanner.
1914 	 */
1915 	for (; block_end_pfn <= cc->free_pfn;
1916 			fast_find_block = false,
1917 			cc->migrate_pfn = low_pfn = block_end_pfn,
1918 			block_start_pfn = block_end_pfn,
1919 			block_end_pfn += pageblock_nr_pages) {
1920 
1921 		/*
1922 		 * This can potentially iterate a massively long zone with
1923 		 * many pageblocks unsuitable, so periodically check if we
1924 		 * need to schedule.
1925 		 */
1926 		if (!(low_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages)))
1927 			cond_resched();
1928 
1929 		page = pageblock_pfn_to_page(block_start_pfn,
1930 						block_end_pfn, cc->zone);
1931 		if (!page)
1932 			continue;
1933 
1934 		/*
1935 		 * If isolation recently failed, do not retry. Only check the
1936 		 * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock
1937 		 * to be visited multiple times. Assume skip was checked
1938 		 * before making it "skip" so other compaction instances do
1939 		 * not scan the same block.
1940 		 */
1941 		if (IS_ALIGNED(low_pfn, pageblock_nr_pages) &&
1942 		    !fast_find_block && !isolation_suitable(cc, page))
1943 			continue;
1944 
1945 		/*
1946 		 * For async direct compaction, only scan the pageblocks of the
1947 		 * same migratetype without huge pages. Async direct compaction
1948 		 * is optimistic to see if the minimum amount of work satisfies
1949 		 * the allocation. The cached PFN is updated as it's possible
1950 		 * that all remaining blocks between source and target are
1951 		 * unsuitable and the compaction scanners fail to meet.
1952 		 */
1953 		if (!suitable_migration_source(cc, page)) {
1954 			update_cached_migrate(cc, block_end_pfn);
1955 			continue;
1956 		}
1957 
1958 		/* Perform the isolation */
1959 		if (isolate_migratepages_block(cc, low_pfn, block_end_pfn,
1960 						isolate_mode))
1961 			return ISOLATE_ABORT;
1962 
1963 		/*
1964 		 * Either we isolated something and proceed with migration. Or
1965 		 * we failed and compact_zone should decide if we should
1966 		 * continue or not.
1967 		 */
1968 		break;
1969 	}
1970 
1971 	return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
1972 }
1973 
1974 /*
1975  * order == -1 is expected when compacting via
1976  * /proc/sys/vm/compact_memory
1977  */
is_via_compact_memory(int order)1978 static inline bool is_via_compact_memory(int order)
1979 {
1980 	return order == -1;
1981 }
1982 
kswapd_is_running(pg_data_t * pgdat)1983 static bool kswapd_is_running(pg_data_t *pgdat)
1984 {
1985 	return pgdat->kswapd && task_is_running(pgdat->kswapd);
1986 }
1987 
1988 /*
1989  * A zone's fragmentation score is the external fragmentation wrt to the
1990  * COMPACTION_HPAGE_ORDER. It returns a value in the range [0, 100].
1991  */
fragmentation_score_zone(struct zone * zone)1992 static unsigned int fragmentation_score_zone(struct zone *zone)
1993 {
1994 	return extfrag_for_order(zone, COMPACTION_HPAGE_ORDER);
1995 }
1996 
1997 /*
1998  * A weighted zone's fragmentation score is the external fragmentation
1999  * wrt to the COMPACTION_HPAGE_ORDER scaled by the zone's size. It
2000  * returns a value in the range [0, 100].
2001  *
2002  * The scaling factor ensures that proactive compaction focuses on larger
2003  * zones like ZONE_NORMAL, rather than smaller, specialized zones like
2004  * ZONE_DMA32. For smaller zones, the score value remains close to zero,
2005  * and thus never exceeds the high threshold for proactive compaction.
2006  */
fragmentation_score_zone_weighted(struct zone * zone)2007 static unsigned int fragmentation_score_zone_weighted(struct zone *zone)
2008 {
2009 	unsigned long score;
2010 
2011 	score = zone->present_pages * fragmentation_score_zone(zone);
2012 	return div64_ul(score, zone->zone_pgdat->node_present_pages + 1);
2013 }
2014 
2015 /*
2016  * The per-node proactive (background) compaction process is started by its
2017  * corresponding kcompactd thread when the node's fragmentation score
2018  * exceeds the high threshold. The compaction process remains active till
2019  * the node's score falls below the low threshold, or one of the back-off
2020  * conditions is met.
2021  */
fragmentation_score_node(pg_data_t * pgdat)2022 static unsigned int fragmentation_score_node(pg_data_t *pgdat)
2023 {
2024 	unsigned int score = 0;
2025 	int zoneid;
2026 
2027 	for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2028 		struct zone *zone;
2029 
2030 		zone = &pgdat->node_zones[zoneid];
2031 		score += fragmentation_score_zone_weighted(zone);
2032 	}
2033 
2034 	return score;
2035 }
2036 
fragmentation_score_wmark(pg_data_t * pgdat,bool low)2037 static unsigned int fragmentation_score_wmark(pg_data_t *pgdat, bool low)
2038 {
2039 	unsigned int wmark_low;
2040 
2041 	/*
2042 	 * Cap the low watermark to avoid excessive compaction
2043 	 * activity in case a user sets the proactiveness tunable
2044 	 * close to 100 (maximum).
2045 	 */
2046 	wmark_low = max(100U - sysctl_compaction_proactiveness, 5U);
2047 	return low ? wmark_low : min(wmark_low + 10, 100U);
2048 }
2049 
should_proactive_compact_node(pg_data_t * pgdat)2050 static bool should_proactive_compact_node(pg_data_t *pgdat)
2051 {
2052 	int wmark_high;
2053 
2054 	if (!sysctl_compaction_proactiveness || kswapd_is_running(pgdat))
2055 		return false;
2056 
2057 	wmark_high = fragmentation_score_wmark(pgdat, false);
2058 	return fragmentation_score_node(pgdat) > wmark_high;
2059 }
2060 
__compact_finished(struct compact_control * cc)2061 static enum compact_result __compact_finished(struct compact_control *cc)
2062 {
2063 	unsigned int order;
2064 	const int migratetype = cc->migratetype;
2065 	int ret;
2066 
2067 	/* Compaction run completes if the migrate and free scanner meet */
2068 	if (compact_scanners_met(cc)) {
2069 		/* Let the next compaction start anew. */
2070 		reset_cached_positions(cc->zone);
2071 
2072 		/*
2073 		 * Mark that the PG_migrate_skip information should be cleared
2074 		 * by kswapd when it goes to sleep. kcompactd does not set the
2075 		 * flag itself as the decision to be clear should be directly
2076 		 * based on an allocation request.
2077 		 */
2078 		if (cc->direct_compaction)
2079 			cc->zone->compact_blockskip_flush = true;
2080 
2081 		if (cc->whole_zone)
2082 			return COMPACT_COMPLETE;
2083 		else
2084 			return COMPACT_PARTIAL_SKIPPED;
2085 	}
2086 
2087 	if (cc->proactive_compaction) {
2088 		int score, wmark_low;
2089 		pg_data_t *pgdat;
2090 
2091 		pgdat = cc->zone->zone_pgdat;
2092 		if (kswapd_is_running(pgdat))
2093 			return COMPACT_PARTIAL_SKIPPED;
2094 
2095 		score = fragmentation_score_zone(cc->zone);
2096 		wmark_low = fragmentation_score_wmark(pgdat, true);
2097 
2098 		if (score > wmark_low)
2099 			ret = COMPACT_CONTINUE;
2100 		else
2101 			ret = COMPACT_SUCCESS;
2102 
2103 		goto out;
2104 	}
2105 
2106 	if (is_via_compact_memory(cc->order))
2107 		return COMPACT_CONTINUE;
2108 
2109 	/*
2110 	 * Always finish scanning a pageblock to reduce the possibility of
2111 	 * fallbacks in the future. This is particularly important when
2112 	 * migration source is unmovable/reclaimable but it's not worth
2113 	 * special casing.
2114 	 */
2115 	if (!IS_ALIGNED(cc->migrate_pfn, pageblock_nr_pages))
2116 		return COMPACT_CONTINUE;
2117 
2118 	/* Direct compactor: Is a suitable page free? */
2119 	ret = COMPACT_NO_SUITABLE_PAGE;
2120 	for (order = cc->order; order < MAX_ORDER; order++) {
2121 		struct free_area *area = &cc->zone->free_area[order];
2122 		bool can_steal;
2123 
2124 		/* Job done if page is free of the right migratetype */
2125 		if (!free_area_empty(area, migratetype))
2126 			return COMPACT_SUCCESS;
2127 
2128 #ifdef CONFIG_CMA
2129 		/* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
2130 		if (migratetype == MIGRATE_MOVABLE &&
2131 			!free_area_empty(area, MIGRATE_CMA))
2132 			return COMPACT_SUCCESS;
2133 #endif
2134 		/*
2135 		 * Job done if allocation would steal freepages from
2136 		 * other migratetype buddy lists.
2137 		 */
2138 		if (find_suitable_fallback(area, order, migratetype,
2139 						true, &can_steal) != -1)
2140 			/*
2141 			 * Movable pages are OK in any pageblock. If we are
2142 			 * stealing for a non-movable allocation, make sure
2143 			 * we finish compacting the current pageblock first
2144 			 * (which is assured by the above migrate_pfn align
2145 			 * check) so it is as free as possible and we won't
2146 			 * have to steal another one soon.
2147 			 */
2148 			return COMPACT_SUCCESS;
2149 	}
2150 
2151 out:
2152 	if (cc->contended || fatal_signal_pending(current))
2153 		ret = COMPACT_CONTENDED;
2154 
2155 	return ret;
2156 }
2157 
compact_finished(struct compact_control * cc)2158 static enum compact_result compact_finished(struct compact_control *cc)
2159 {
2160 	int ret;
2161 
2162 	ret = __compact_finished(cc);
2163 	trace_mm_compaction_finished(cc->zone, cc->order, ret);
2164 	if (ret == COMPACT_NO_SUITABLE_PAGE)
2165 		ret = COMPACT_CONTINUE;
2166 
2167 	return ret;
2168 }
2169 
__compaction_suitable(struct zone * zone,int order,unsigned int alloc_flags,int highest_zoneidx,unsigned long wmark_target)2170 static enum compact_result __compaction_suitable(struct zone *zone, int order,
2171 					unsigned int alloc_flags,
2172 					int highest_zoneidx,
2173 					unsigned long wmark_target)
2174 {
2175 	unsigned long watermark;
2176 
2177 	if (is_via_compact_memory(order))
2178 		return COMPACT_CONTINUE;
2179 
2180 	watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
2181 	/*
2182 	 * If watermarks for high-order allocation are already met, there
2183 	 * should be no need for compaction at all.
2184 	 */
2185 	if (zone_watermark_ok(zone, order, watermark, highest_zoneidx,
2186 								alloc_flags))
2187 		return COMPACT_SUCCESS;
2188 
2189 	/*
2190 	 * Watermarks for order-0 must be met for compaction to be able to
2191 	 * isolate free pages for migration targets. This means that the
2192 	 * watermark and alloc_flags have to match, or be more pessimistic than
2193 	 * the check in __isolate_free_page(). We don't use the direct
2194 	 * compactor's alloc_flags, as they are not relevant for freepage
2195 	 * isolation. We however do use the direct compactor's highest_zoneidx
2196 	 * to skip over zones where lowmem reserves would prevent allocation
2197 	 * even if compaction succeeds.
2198 	 * For costly orders, we require low watermark instead of min for
2199 	 * compaction to proceed to increase its chances.
2200 	 * ALLOC_CMA is used, as pages in CMA pageblocks are considered
2201 	 * suitable migration targets
2202 	 */
2203 	watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
2204 				low_wmark_pages(zone) : min_wmark_pages(zone);
2205 	watermark += compact_gap(order);
2206 	if (!__zone_watermark_ok(zone, 0, watermark, highest_zoneidx,
2207 						ALLOC_CMA, wmark_target))
2208 		return COMPACT_SKIPPED;
2209 
2210 	return COMPACT_CONTINUE;
2211 }
2212 
2213 /*
2214  * compaction_suitable: Is this suitable to run compaction on this zone now?
2215  * Returns
2216  *   COMPACT_SKIPPED  - If there are too few free pages for compaction
2217  *   COMPACT_SUCCESS  - If the allocation would succeed without compaction
2218  *   COMPACT_CONTINUE - If compaction should run now
2219  */
compaction_suitable(struct zone * zone,int order,unsigned int alloc_flags,int highest_zoneidx)2220 enum compact_result compaction_suitable(struct zone *zone, int order,
2221 					unsigned int alloc_flags,
2222 					int highest_zoneidx)
2223 {
2224 	enum compact_result ret;
2225 	int fragindex;
2226 
2227 	ret = __compaction_suitable(zone, order, alloc_flags, highest_zoneidx,
2228 				    zone_page_state(zone, NR_FREE_PAGES));
2229 	/*
2230 	 * fragmentation index determines if allocation failures are due to
2231 	 * low memory or external fragmentation
2232 	 *
2233 	 * index of -1000 would imply allocations might succeed depending on
2234 	 * watermarks, but we already failed the high-order watermark check
2235 	 * index towards 0 implies failure is due to lack of memory
2236 	 * index towards 1000 implies failure is due to fragmentation
2237 	 *
2238 	 * Only compact if a failure would be due to fragmentation. Also
2239 	 * ignore fragindex for non-costly orders where the alternative to
2240 	 * a successful reclaim/compaction is OOM. Fragindex and the
2241 	 * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
2242 	 * excessive compaction for costly orders, but it should not be at the
2243 	 * expense of system stability.
2244 	 */
2245 	if (ret == COMPACT_CONTINUE && (order > PAGE_ALLOC_COSTLY_ORDER)) {
2246 		fragindex = fragmentation_index(zone, order);
2247 		if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
2248 			ret = COMPACT_NOT_SUITABLE_ZONE;
2249 	}
2250 
2251 	trace_mm_compaction_suitable(zone, order, ret);
2252 	if (ret == COMPACT_NOT_SUITABLE_ZONE)
2253 		ret = COMPACT_SKIPPED;
2254 
2255 	return ret;
2256 }
2257 
compaction_zonelist_suitable(struct alloc_context * ac,int order,int alloc_flags)2258 bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
2259 		int alloc_flags)
2260 {
2261 	struct zone *zone;
2262 	struct zoneref *z;
2263 
2264 	/*
2265 	 * Make sure at least one zone would pass __compaction_suitable if we continue
2266 	 * retrying the reclaim.
2267 	 */
2268 	for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2269 				ac->highest_zoneidx, ac->nodemask) {
2270 		unsigned long available;
2271 		enum compact_result compact_result;
2272 
2273 		/*
2274 		 * Do not consider all the reclaimable memory because we do not
2275 		 * want to trash just for a single high order allocation which
2276 		 * is even not guaranteed to appear even if __compaction_suitable
2277 		 * is happy about the watermark check.
2278 		 */
2279 		available = zone_reclaimable_pages(zone) / order;
2280 		available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
2281 		compact_result = __compaction_suitable(zone, order, alloc_flags,
2282 				ac->highest_zoneidx, available);
2283 		if (compact_result == COMPACT_CONTINUE)
2284 			return true;
2285 	}
2286 
2287 	return false;
2288 }
2289 
2290 static enum compact_result
compact_zone(struct compact_control * cc,struct capture_control * capc)2291 compact_zone(struct compact_control *cc, struct capture_control *capc)
2292 {
2293 	enum compact_result ret;
2294 	unsigned long start_pfn = cc->zone->zone_start_pfn;
2295 	unsigned long end_pfn = zone_end_pfn(cc->zone);
2296 	unsigned long last_migrated_pfn;
2297 	const bool sync = cc->mode != MIGRATE_ASYNC;
2298 	bool update_cached;
2299 	unsigned int nr_succeeded = 0;
2300 
2301 	/*
2302 	 * These counters track activities during zone compaction.  Initialize
2303 	 * them before compacting a new zone.
2304 	 */
2305 	cc->total_migrate_scanned = 0;
2306 	cc->total_free_scanned = 0;
2307 	cc->nr_migratepages = 0;
2308 	cc->nr_freepages = 0;
2309 	INIT_LIST_HEAD(&cc->freepages);
2310 	INIT_LIST_HEAD(&cc->migratepages);
2311 
2312 	cc->migratetype = gfp_migratetype(cc->gfp_mask);
2313 	ret = compaction_suitable(cc->zone, cc->order, cc->alloc_flags,
2314 							cc->highest_zoneidx);
2315 	/* Compaction is likely to fail */
2316 	if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED)
2317 		return ret;
2318 
2319 	/* huh, compaction_suitable is returning something unexpected */
2320 	VM_BUG_ON(ret != COMPACT_CONTINUE);
2321 
2322 	/*
2323 	 * Clear pageblock skip if there were failures recently and compaction
2324 	 * is about to be retried after being deferred.
2325 	 */
2326 	if (compaction_restarting(cc->zone, cc->order))
2327 		__reset_isolation_suitable(cc->zone);
2328 
2329 	/*
2330 	 * Setup to move all movable pages to the end of the zone. Used cached
2331 	 * information on where the scanners should start (unless we explicitly
2332 	 * want to compact the whole zone), but check that it is initialised
2333 	 * by ensuring the values are within zone boundaries.
2334 	 */
2335 	cc->fast_start_pfn = 0;
2336 	if (cc->whole_zone) {
2337 		cc->migrate_pfn = start_pfn;
2338 		cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2339 	} else {
2340 		cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync];
2341 		cc->free_pfn = cc->zone->compact_cached_free_pfn;
2342 		if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
2343 			cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2344 			cc->zone->compact_cached_free_pfn = cc->free_pfn;
2345 		}
2346 		if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
2347 			cc->migrate_pfn = start_pfn;
2348 			cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
2349 			cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
2350 		}
2351 
2352 		if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn)
2353 			cc->whole_zone = true;
2354 	}
2355 
2356 	last_migrated_pfn = 0;
2357 
2358 	/*
2359 	 * Migrate has separate cached PFNs for ASYNC and SYNC* migration on
2360 	 * the basis that some migrations will fail in ASYNC mode. However,
2361 	 * if the cached PFNs match and pageblocks are skipped due to having
2362 	 * no isolation candidates, then the sync state does not matter.
2363 	 * Until a pageblock with isolation candidates is found, keep the
2364 	 * cached PFNs in sync to avoid revisiting the same blocks.
2365 	 */
2366 	update_cached = !sync &&
2367 		cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1];
2368 
2369 	trace_mm_compaction_begin(cc, start_pfn, end_pfn, sync);
2370 
2371 	/* lru_add_drain_all could be expensive with involving other CPUs */
2372 	lru_add_drain();
2373 
2374 	while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) {
2375 		int err;
2376 		unsigned long iteration_start_pfn = cc->migrate_pfn;
2377 
2378 		/*
2379 		 * Avoid multiple rescans which can happen if a page cannot be
2380 		 * isolated (dirty/writeback in async mode) or if the migrated
2381 		 * pages are being allocated before the pageblock is cleared.
2382 		 * The first rescan will capture the entire pageblock for
2383 		 * migration. If it fails, it'll be marked skip and scanning
2384 		 * will proceed as normal.
2385 		 */
2386 		cc->rescan = false;
2387 		if (pageblock_start_pfn(last_migrated_pfn) ==
2388 		    pageblock_start_pfn(iteration_start_pfn)) {
2389 			cc->rescan = true;
2390 		}
2391 
2392 		switch (isolate_migratepages(cc)) {
2393 		case ISOLATE_ABORT:
2394 			ret = COMPACT_CONTENDED;
2395 			putback_movable_pages(&cc->migratepages);
2396 			cc->nr_migratepages = 0;
2397 			goto out;
2398 		case ISOLATE_NONE:
2399 			if (update_cached) {
2400 				cc->zone->compact_cached_migrate_pfn[1] =
2401 					cc->zone->compact_cached_migrate_pfn[0];
2402 			}
2403 
2404 			/*
2405 			 * We haven't isolated and migrated anything, but
2406 			 * there might still be unflushed migrations from
2407 			 * previous cc->order aligned block.
2408 			 */
2409 			goto check_drain;
2410 		case ISOLATE_SUCCESS:
2411 			update_cached = false;
2412 			last_migrated_pfn = iteration_start_pfn;
2413 		}
2414 
2415 		err = migrate_pages(&cc->migratepages, compaction_alloc,
2416 				compaction_free, (unsigned long)cc, cc->mode,
2417 				MR_COMPACTION, &nr_succeeded);
2418 
2419 		trace_mm_compaction_migratepages(cc, nr_succeeded);
2420 
2421 		/* All pages were either migrated or will be released */
2422 		cc->nr_migratepages = 0;
2423 		if (err) {
2424 			putback_movable_pages(&cc->migratepages);
2425 			/*
2426 			 * migrate_pages() may return -ENOMEM when scanners meet
2427 			 * and we want compact_finished() to detect it
2428 			 */
2429 			if (err == -ENOMEM && !compact_scanners_met(cc)) {
2430 				ret = COMPACT_CONTENDED;
2431 				goto out;
2432 			}
2433 			/*
2434 			 * We failed to migrate at least one page in the current
2435 			 * order-aligned block, so skip the rest of it.
2436 			 */
2437 			if (cc->direct_compaction &&
2438 						(cc->mode == MIGRATE_ASYNC)) {
2439 				cc->migrate_pfn = block_end_pfn(
2440 						cc->migrate_pfn - 1, cc->order);
2441 				/* Draining pcplists is useless in this case */
2442 				last_migrated_pfn = 0;
2443 			}
2444 		}
2445 
2446 check_drain:
2447 		/*
2448 		 * Has the migration scanner moved away from the previous
2449 		 * cc->order aligned block where we migrated from? If yes,
2450 		 * flush the pages that were freed, so that they can merge and
2451 		 * compact_finished() can detect immediately if allocation
2452 		 * would succeed.
2453 		 */
2454 		if (cc->order > 0 && last_migrated_pfn) {
2455 			unsigned long current_block_start =
2456 				block_start_pfn(cc->migrate_pfn, cc->order);
2457 
2458 			if (last_migrated_pfn < current_block_start) {
2459 				lru_add_drain_cpu_zone(cc->zone);
2460 				/* No more flushing until we migrate again */
2461 				last_migrated_pfn = 0;
2462 			}
2463 		}
2464 
2465 		/* Stop if a page has been captured */
2466 		if (capc && capc->page) {
2467 			ret = COMPACT_SUCCESS;
2468 			break;
2469 		}
2470 	}
2471 
2472 out:
2473 	/*
2474 	 * Release free pages and update where the free scanner should restart,
2475 	 * so we don't leave any returned pages behind in the next attempt.
2476 	 */
2477 	if (cc->nr_freepages > 0) {
2478 		unsigned long free_pfn = release_freepages(&cc->freepages);
2479 
2480 		cc->nr_freepages = 0;
2481 		VM_BUG_ON(free_pfn == 0);
2482 		/* The cached pfn is always the first in a pageblock */
2483 		free_pfn = pageblock_start_pfn(free_pfn);
2484 		/*
2485 		 * Only go back, not forward. The cached pfn might have been
2486 		 * already reset to zone end in compact_finished()
2487 		 */
2488 		if (free_pfn > cc->zone->compact_cached_free_pfn)
2489 			cc->zone->compact_cached_free_pfn = free_pfn;
2490 	}
2491 
2492 	count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned);
2493 	count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned);
2494 
2495 	trace_mm_compaction_end(cc, start_pfn, end_pfn, sync, ret);
2496 
2497 	return ret;
2498 }
2499 
compact_zone_order(struct zone * zone,int order,gfp_t gfp_mask,enum compact_priority prio,unsigned int alloc_flags,int highest_zoneidx,struct page ** capture)2500 static enum compact_result compact_zone_order(struct zone *zone, int order,
2501 		gfp_t gfp_mask, enum compact_priority prio,
2502 		unsigned int alloc_flags, int highest_zoneidx,
2503 		struct page **capture)
2504 {
2505 	enum compact_result ret;
2506 	struct compact_control cc = {
2507 		.order = order,
2508 		.search_order = order,
2509 		.gfp_mask = gfp_mask,
2510 		.zone = zone,
2511 		.mode = (prio == COMPACT_PRIO_ASYNC) ?
2512 					MIGRATE_ASYNC :	MIGRATE_SYNC_LIGHT,
2513 		.alloc_flags = alloc_flags,
2514 		.highest_zoneidx = highest_zoneidx,
2515 		.direct_compaction = true,
2516 		.whole_zone = (prio == MIN_COMPACT_PRIORITY),
2517 		.ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
2518 		.ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
2519 	};
2520 	struct capture_control capc = {
2521 		.cc = &cc,
2522 		.page = NULL,
2523 	};
2524 
2525 	/*
2526 	 * Make sure the structs are really initialized before we expose the
2527 	 * capture control, in case we are interrupted and the interrupt handler
2528 	 * frees a page.
2529 	 */
2530 	barrier();
2531 	WRITE_ONCE(current->capture_control, &capc);
2532 
2533 	ret = compact_zone(&cc, &capc);
2534 
2535 	VM_BUG_ON(!list_empty(&cc.freepages));
2536 	VM_BUG_ON(!list_empty(&cc.migratepages));
2537 
2538 	/*
2539 	 * Make sure we hide capture control first before we read the captured
2540 	 * page pointer, otherwise an interrupt could free and capture a page
2541 	 * and we would leak it.
2542 	 */
2543 	WRITE_ONCE(current->capture_control, NULL);
2544 	*capture = READ_ONCE(capc.page);
2545 	/*
2546 	 * Technically, it is also possible that compaction is skipped but
2547 	 * the page is still captured out of luck(IRQ came and freed the page).
2548 	 * Returning COMPACT_SUCCESS in such cases helps in properly accounting
2549 	 * the COMPACT[STALL|FAIL] when compaction is skipped.
2550 	 */
2551 	if (*capture)
2552 		ret = COMPACT_SUCCESS;
2553 
2554 	return ret;
2555 }
2556 
2557 int sysctl_extfrag_threshold = 500;
2558 
2559 /**
2560  * try_to_compact_pages - Direct compact to satisfy a high-order allocation
2561  * @gfp_mask: The GFP mask of the current allocation
2562  * @order: The order of the current allocation
2563  * @alloc_flags: The allocation flags of the current allocation
2564  * @ac: The context of current allocation
2565  * @prio: Determines how hard direct compaction should try to succeed
2566  * @capture: Pointer to free page created by compaction will be stored here
2567  *
2568  * This is the main entry point for direct page compaction.
2569  */
try_to_compact_pages(gfp_t gfp_mask,unsigned int order,unsigned int alloc_flags,const struct alloc_context * ac,enum compact_priority prio,struct page ** capture)2570 enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
2571 		unsigned int alloc_flags, const struct alloc_context *ac,
2572 		enum compact_priority prio, struct page **capture)
2573 {
2574 	int may_perform_io = (__force int)(gfp_mask & __GFP_IO);
2575 	struct zoneref *z;
2576 	struct zone *zone;
2577 	enum compact_result rc = COMPACT_SKIPPED;
2578 
2579 	/*
2580 	 * Check if the GFP flags allow compaction - GFP_NOIO is really
2581 	 * tricky context because the migration might require IO
2582 	 */
2583 	if (!may_perform_io)
2584 		return COMPACT_SKIPPED;
2585 
2586 	trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
2587 
2588 	/* Compact each zone in the list */
2589 	for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2590 					ac->highest_zoneidx, ac->nodemask) {
2591 		enum compact_result status;
2592 
2593 		if (prio > MIN_COMPACT_PRIORITY
2594 					&& compaction_deferred(zone, order)) {
2595 			rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
2596 			continue;
2597 		}
2598 
2599 		status = compact_zone_order(zone, order, gfp_mask, prio,
2600 				alloc_flags, ac->highest_zoneidx, capture);
2601 		rc = max(status, rc);
2602 
2603 		/* The allocation should succeed, stop compacting */
2604 		if (status == COMPACT_SUCCESS) {
2605 			/*
2606 			 * We think the allocation will succeed in this zone,
2607 			 * but it is not certain, hence the false. The caller
2608 			 * will repeat this with true if allocation indeed
2609 			 * succeeds in this zone.
2610 			 */
2611 			compaction_defer_reset(zone, order, false);
2612 
2613 			break;
2614 		}
2615 
2616 		if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
2617 					status == COMPACT_PARTIAL_SKIPPED))
2618 			/*
2619 			 * We think that allocation won't succeed in this zone
2620 			 * so we defer compaction there. If it ends up
2621 			 * succeeding after all, it will be reset.
2622 			 */
2623 			defer_compaction(zone, order);
2624 
2625 		/*
2626 		 * We might have stopped compacting due to need_resched() in
2627 		 * async compaction, or due to a fatal signal detected. In that
2628 		 * case do not try further zones
2629 		 */
2630 		if ((prio == COMPACT_PRIO_ASYNC && need_resched())
2631 					|| fatal_signal_pending(current))
2632 			break;
2633 	}
2634 
2635 	return rc;
2636 }
2637 
2638 /*
2639  * Compact all zones within a node till each zone's fragmentation score
2640  * reaches within proactive compaction thresholds (as determined by the
2641  * proactiveness tunable).
2642  *
2643  * It is possible that the function returns before reaching score targets
2644  * due to various back-off conditions, such as, contention on per-node or
2645  * per-zone locks.
2646  */
proactive_compact_node(pg_data_t * pgdat)2647 static void proactive_compact_node(pg_data_t *pgdat)
2648 {
2649 	int zoneid;
2650 	struct zone *zone;
2651 	struct compact_control cc = {
2652 		.order = -1,
2653 		.mode = MIGRATE_SYNC_LIGHT,
2654 		.ignore_skip_hint = true,
2655 		.whole_zone = true,
2656 		.gfp_mask = GFP_KERNEL,
2657 		.proactive_compaction = true,
2658 	};
2659 
2660 	for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2661 		zone = &pgdat->node_zones[zoneid];
2662 		if (!populated_zone(zone))
2663 			continue;
2664 
2665 		cc.zone = zone;
2666 
2667 		compact_zone(&cc, NULL);
2668 
2669 		VM_BUG_ON(!list_empty(&cc.freepages));
2670 		VM_BUG_ON(!list_empty(&cc.migratepages));
2671 	}
2672 }
2673 
2674 /* Compact all zones within a node */
compact_node(int nid)2675 static void compact_node(int nid)
2676 {
2677 	pg_data_t *pgdat = NODE_DATA(nid);
2678 	int zoneid;
2679 	struct zone *zone;
2680 	struct compact_control cc = {
2681 		.order = -1,
2682 		.mode = MIGRATE_SYNC,
2683 		.ignore_skip_hint = true,
2684 		.whole_zone = true,
2685 		.gfp_mask = GFP_KERNEL,
2686 	};
2687 
2688 
2689 	for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2690 
2691 		zone = &pgdat->node_zones[zoneid];
2692 		if (!populated_zone(zone))
2693 			continue;
2694 
2695 		cc.zone = zone;
2696 
2697 		compact_zone(&cc, NULL);
2698 
2699 		VM_BUG_ON(!list_empty(&cc.freepages));
2700 		VM_BUG_ON(!list_empty(&cc.migratepages));
2701 	}
2702 }
2703 
2704 /* Compact all nodes in the system */
compact_nodes(void)2705 static void compact_nodes(void)
2706 {
2707 	int nid;
2708 
2709 	/* Flush pending updates to the LRU lists */
2710 	lru_add_drain_all();
2711 
2712 	for_each_online_node(nid)
2713 		compact_node(nid);
2714 }
2715 
2716 /*
2717  * Tunable for proactive compaction. It determines how
2718  * aggressively the kernel should compact memory in the
2719  * background. It takes values in the range [0, 100].
2720  */
2721 unsigned int __read_mostly sysctl_compaction_proactiveness = 20;
2722 
compaction_proactiveness_sysctl_handler(struct ctl_table * table,int write,void * buffer,size_t * length,loff_t * ppos)2723 int compaction_proactiveness_sysctl_handler(struct ctl_table *table, int write,
2724 		void *buffer, size_t *length, loff_t *ppos)
2725 {
2726 	int rc, nid;
2727 
2728 	rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
2729 	if (rc)
2730 		return rc;
2731 
2732 	if (write && sysctl_compaction_proactiveness) {
2733 		for_each_online_node(nid) {
2734 			pg_data_t *pgdat = NODE_DATA(nid);
2735 
2736 			if (pgdat->proactive_compact_trigger)
2737 				continue;
2738 
2739 			pgdat->proactive_compact_trigger = true;
2740 			wake_up_interruptible(&pgdat->kcompactd_wait);
2741 		}
2742 	}
2743 
2744 	return 0;
2745 }
2746 
2747 /*
2748  * This is the entry point for compacting all nodes via
2749  * /proc/sys/vm/compact_memory
2750  */
sysctl_compaction_handler(struct ctl_table * table,int write,void * buffer,size_t * length,loff_t * ppos)2751 int sysctl_compaction_handler(struct ctl_table *table, int write,
2752 			void *buffer, size_t *length, loff_t *ppos)
2753 {
2754 	if (write)
2755 		compact_nodes();
2756 
2757 	return 0;
2758 }
2759 
2760 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
compact_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)2761 static ssize_t compact_store(struct device *dev,
2762 			     struct device_attribute *attr,
2763 			     const char *buf, size_t count)
2764 {
2765 	int nid = dev->id;
2766 
2767 	if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
2768 		/* Flush pending updates to the LRU lists */
2769 		lru_add_drain_all();
2770 
2771 		compact_node(nid);
2772 	}
2773 
2774 	return count;
2775 }
2776 static DEVICE_ATTR_WO(compact);
2777 
compaction_register_node(struct node * node)2778 int compaction_register_node(struct node *node)
2779 {
2780 	return device_create_file(&node->dev, &dev_attr_compact);
2781 }
2782 
compaction_unregister_node(struct node * node)2783 void compaction_unregister_node(struct node *node)
2784 {
2785 	return device_remove_file(&node->dev, &dev_attr_compact);
2786 }
2787 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
2788 
kcompactd_work_requested(pg_data_t * pgdat)2789 static inline bool kcompactd_work_requested(pg_data_t *pgdat)
2790 {
2791 	return pgdat->kcompactd_max_order > 0 || kthread_should_stop() ||
2792 		pgdat->proactive_compact_trigger;
2793 }
2794 
kcompactd_node_suitable(pg_data_t * pgdat)2795 static bool kcompactd_node_suitable(pg_data_t *pgdat)
2796 {
2797 	int zoneid;
2798 	struct zone *zone;
2799 	enum zone_type highest_zoneidx = pgdat->kcompactd_highest_zoneidx;
2800 
2801 	for (zoneid = 0; zoneid <= highest_zoneidx; zoneid++) {
2802 		zone = &pgdat->node_zones[zoneid];
2803 
2804 		if (!populated_zone(zone))
2805 			continue;
2806 
2807 		if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0,
2808 					highest_zoneidx) == COMPACT_CONTINUE)
2809 			return true;
2810 	}
2811 
2812 	return false;
2813 }
2814 
kcompactd_do_work(pg_data_t * pgdat)2815 static void kcompactd_do_work(pg_data_t *pgdat)
2816 {
2817 	/*
2818 	 * With no special task, compact all zones so that a page of requested
2819 	 * order is allocatable.
2820 	 */
2821 	int zoneid;
2822 	struct zone *zone;
2823 	struct compact_control cc = {
2824 		.order = pgdat->kcompactd_max_order,
2825 		.search_order = pgdat->kcompactd_max_order,
2826 		.highest_zoneidx = pgdat->kcompactd_highest_zoneidx,
2827 		.mode = MIGRATE_SYNC_LIGHT,
2828 		.ignore_skip_hint = false,
2829 		.gfp_mask = GFP_KERNEL,
2830 	};
2831 	trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
2832 							cc.highest_zoneidx);
2833 	count_compact_event(KCOMPACTD_WAKE);
2834 
2835 	for (zoneid = 0; zoneid <= cc.highest_zoneidx; zoneid++) {
2836 		int status;
2837 
2838 		zone = &pgdat->node_zones[zoneid];
2839 		if (!populated_zone(zone))
2840 			continue;
2841 
2842 		if (compaction_deferred(zone, cc.order))
2843 			continue;
2844 
2845 		if (compaction_suitable(zone, cc.order, 0, zoneid) !=
2846 							COMPACT_CONTINUE)
2847 			continue;
2848 
2849 		if (kthread_should_stop())
2850 			return;
2851 
2852 		cc.zone = zone;
2853 		status = compact_zone(&cc, NULL);
2854 
2855 		if (status == COMPACT_SUCCESS) {
2856 			compaction_defer_reset(zone, cc.order, false);
2857 		} else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
2858 			/*
2859 			 * Buddy pages may become stranded on pcps that could
2860 			 * otherwise coalesce on the zone's free area for
2861 			 * order >= cc.order.  This is ratelimited by the
2862 			 * upcoming deferral.
2863 			 */
2864 			drain_all_pages(zone);
2865 
2866 			/*
2867 			 * We use sync migration mode here, so we defer like
2868 			 * sync direct compaction does.
2869 			 */
2870 			defer_compaction(zone, cc.order);
2871 		}
2872 
2873 		count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
2874 				     cc.total_migrate_scanned);
2875 		count_compact_events(KCOMPACTD_FREE_SCANNED,
2876 				     cc.total_free_scanned);
2877 
2878 		VM_BUG_ON(!list_empty(&cc.freepages));
2879 		VM_BUG_ON(!list_empty(&cc.migratepages));
2880 	}
2881 
2882 	/*
2883 	 * Regardless of success, we are done until woken up next. But remember
2884 	 * the requested order/highest_zoneidx in case it was higher/tighter
2885 	 * than our current ones
2886 	 */
2887 	if (pgdat->kcompactd_max_order <= cc.order)
2888 		pgdat->kcompactd_max_order = 0;
2889 	if (pgdat->kcompactd_highest_zoneidx >= cc.highest_zoneidx)
2890 		pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
2891 }
2892 
wakeup_kcompactd(pg_data_t * pgdat,int order,int highest_zoneidx)2893 void wakeup_kcompactd(pg_data_t *pgdat, int order, int highest_zoneidx)
2894 {
2895 	if (!order)
2896 		return;
2897 
2898 	if (pgdat->kcompactd_max_order < order)
2899 		pgdat->kcompactd_max_order = order;
2900 
2901 	if (pgdat->kcompactd_highest_zoneidx > highest_zoneidx)
2902 		pgdat->kcompactd_highest_zoneidx = highest_zoneidx;
2903 
2904 	/*
2905 	 * Pairs with implicit barrier in wait_event_freezable()
2906 	 * such that wakeups are not missed.
2907 	 */
2908 	if (!wq_has_sleeper(&pgdat->kcompactd_wait))
2909 		return;
2910 
2911 	if (!kcompactd_node_suitable(pgdat))
2912 		return;
2913 
2914 	trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
2915 							highest_zoneidx);
2916 	wake_up_interruptible(&pgdat->kcompactd_wait);
2917 }
2918 
2919 /*
2920  * The background compaction daemon, started as a kernel thread
2921  * from the init process.
2922  */
kcompactd(void * p)2923 static int kcompactd(void *p)
2924 {
2925 	pg_data_t *pgdat = (pg_data_t *)p;
2926 	struct task_struct *tsk = current;
2927 	long default_timeout = msecs_to_jiffies(HPAGE_FRAG_CHECK_INTERVAL_MSEC);
2928 	long timeout = default_timeout;
2929 
2930 	const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2931 
2932 	if (!cpumask_empty(cpumask))
2933 		set_cpus_allowed_ptr(tsk, cpumask);
2934 
2935 	set_freezable();
2936 
2937 	pgdat->kcompactd_max_order = 0;
2938 	pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
2939 
2940 	while (!kthread_should_stop()) {
2941 		unsigned long pflags;
2942 
2943 		/*
2944 		 * Avoid the unnecessary wakeup for proactive compaction
2945 		 * when it is disabled.
2946 		 */
2947 		if (!sysctl_compaction_proactiveness)
2948 			timeout = MAX_SCHEDULE_TIMEOUT;
2949 		trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
2950 		if (wait_event_freezable_timeout(pgdat->kcompactd_wait,
2951 			kcompactd_work_requested(pgdat), timeout) &&
2952 			!pgdat->proactive_compact_trigger) {
2953 
2954 			psi_memstall_enter(&pflags);
2955 			kcompactd_do_work(pgdat);
2956 			psi_memstall_leave(&pflags);
2957 			/*
2958 			 * Reset the timeout value. The defer timeout from
2959 			 * proactive compaction is lost here but that is fine
2960 			 * as the condition of the zone changing substantionally
2961 			 * then carrying on with the previous defer interval is
2962 			 * not useful.
2963 			 */
2964 			timeout = default_timeout;
2965 			continue;
2966 		}
2967 
2968 		/*
2969 		 * Start the proactive work with default timeout. Based
2970 		 * on the fragmentation score, this timeout is updated.
2971 		 */
2972 		timeout = default_timeout;
2973 		if (should_proactive_compact_node(pgdat)) {
2974 			unsigned int prev_score, score;
2975 
2976 			prev_score = fragmentation_score_node(pgdat);
2977 			proactive_compact_node(pgdat);
2978 			score = fragmentation_score_node(pgdat);
2979 			/*
2980 			 * Defer proactive compaction if the fragmentation
2981 			 * score did not go down i.e. no progress made.
2982 			 */
2983 			if (unlikely(score >= prev_score))
2984 				timeout =
2985 				   default_timeout << COMPACT_MAX_DEFER_SHIFT;
2986 		}
2987 		if (unlikely(pgdat->proactive_compact_trigger))
2988 			pgdat->proactive_compact_trigger = false;
2989 	}
2990 
2991 	return 0;
2992 }
2993 
2994 /*
2995  * This kcompactd start function will be called by init and node-hot-add.
2996  * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
2997  */
kcompactd_run(int nid)2998 void kcompactd_run(int nid)
2999 {
3000 	pg_data_t *pgdat = NODE_DATA(nid);
3001 
3002 	if (pgdat->kcompactd)
3003 		return;
3004 
3005 	pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
3006 	if (IS_ERR(pgdat->kcompactd)) {
3007 		pr_err("Failed to start kcompactd on node %d\n", nid);
3008 		pgdat->kcompactd = NULL;
3009 	}
3010 }
3011 
3012 /*
3013  * Called by memory hotplug when all memory in a node is offlined. Caller must
3014  * hold mem_hotplug_begin/end().
3015  */
kcompactd_stop(int nid)3016 void kcompactd_stop(int nid)
3017 {
3018 	struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
3019 
3020 	if (kcompactd) {
3021 		kthread_stop(kcompactd);
3022 		NODE_DATA(nid)->kcompactd = NULL;
3023 	}
3024 }
3025 
3026 /*
3027  * It's optimal to keep kcompactd on the same CPUs as their memory, but
3028  * not required for correctness. So if the last cpu in a node goes
3029  * away, we get changed to run anywhere: as the first one comes back,
3030  * restore their cpu bindings.
3031  */
kcompactd_cpu_online(unsigned int cpu)3032 static int kcompactd_cpu_online(unsigned int cpu)
3033 {
3034 	int nid;
3035 
3036 	for_each_node_state(nid, N_MEMORY) {
3037 		pg_data_t *pgdat = NODE_DATA(nid);
3038 		const struct cpumask *mask;
3039 
3040 		mask = cpumask_of_node(pgdat->node_id);
3041 
3042 		if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3043 			/* One of our CPUs online: restore mask */
3044 			if (pgdat->kcompactd)
3045 				set_cpus_allowed_ptr(pgdat->kcompactd, mask);
3046 	}
3047 	return 0;
3048 }
3049 
kcompactd_init(void)3050 static int __init kcompactd_init(void)
3051 {
3052 	int nid;
3053 	int ret;
3054 
3055 	ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3056 					"mm/compaction:online",
3057 					kcompactd_cpu_online, NULL);
3058 	if (ret < 0) {
3059 		pr_err("kcompactd: failed to register hotplug callbacks.\n");
3060 		return ret;
3061 	}
3062 
3063 	for_each_node_state(nid, N_MEMORY)
3064 		kcompactd_run(nid);
3065 	return 0;
3066 }
3067 subsys_initcall(kcompactd_init)
3068 
3069 #endif /* CONFIG_COMPACTION */
3070