Loading...
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
29static inline void count_compact_event(enum vm_event_item item)
30{
31 count_vm_event(item);
32}
33
34static inline void count_compact_events(enum vm_event_item item, long delta)
35{
36 count_vm_events(item, delta);
37}
38#else
39#define count_compact_event(item) do { } while (0)
40#define count_compact_events(item, delta) do { } while (0)
41#endif
42
43#if defined CONFIG_COMPACTION || defined CONFIG_CMA
44
45#define CREATE_TRACE_POINTS
46#include <trace/events/compaction.h>
47
48#define block_start_pfn(pfn, order) round_down(pfn, 1UL << (order))
49#define block_end_pfn(pfn, order) ALIGN((pfn) + 1, 1UL << (order))
50#define pageblock_start_pfn(pfn) block_start_pfn(pfn, pageblock_order)
51#define pageblock_end_pfn(pfn) block_end_pfn(pfn, pageblock_order)
52
53/*
54 * Fragmentation score check interval for proactive compaction purposes.
55 */
56static const unsigned int HPAGE_FRAG_CHECK_INTERVAL_MSEC = 500;
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
71static 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
87static 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
114int 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}
128EXPORT_SYMBOL(PageMovable);
129
130void __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}
136EXPORT_SYMBOL(__SetPageMovable);
137
138void __ClearPageMovable(struct page *page)
139{
140 VM_BUG_ON_PAGE(!PageLocked(page), page);
141 VM_BUG_ON_PAGE(!PageMovable(page), page);
142 /*
143 * Clear registered address_space val with keeping PAGE_MAPPING_MOVABLE
144 * flag so that VM can catch up released page by driver after isolation.
145 * With it, VM migration doesn't try to put it back.
146 */
147 page->mapping = (void *)((unsigned long)page->mapping &
148 PAGE_MAPPING_MOVABLE);
149}
150EXPORT_SYMBOL(__ClearPageMovable);
151
152/* Do not skip compaction more than 64 times */
153#define COMPACT_MAX_DEFER_SHIFT 6
154
155/*
156 * Compaction is deferred when compaction fails to result in a page
157 * allocation success. 1 << compact_defer_shift, compactions are skipped up
158 * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
159 */
160void defer_compaction(struct zone *zone, int order)
161{
162 zone->compact_considered = 0;
163 zone->compact_defer_shift++;
164
165 if (order < zone->compact_order_failed)
166 zone->compact_order_failed = order;
167
168 if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
169 zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
170
171 trace_mm_compaction_defer_compaction(zone, order);
172}
173
174/* Returns true if compaction should be skipped this time */
175bool compaction_deferred(struct zone *zone, int order)
176{
177 unsigned long defer_limit = 1UL << zone->compact_defer_shift;
178
179 if (order < zone->compact_order_failed)
180 return false;
181
182 /* Avoid possible overflow */
183 if (++zone->compact_considered > defer_limit)
184 zone->compact_considered = defer_limit;
185
186 if (zone->compact_considered >= defer_limit)
187 return false;
188
189 trace_mm_compaction_deferred(zone, order);
190
191 return true;
192}
193
194/*
195 * Update defer tracking counters after successful compaction of given order,
196 * which means an allocation either succeeded (alloc_success == true) or is
197 * expected to succeed.
198 */
199void compaction_defer_reset(struct zone *zone, int order,
200 bool alloc_success)
201{
202 if (alloc_success) {
203 zone->compact_considered = 0;
204 zone->compact_defer_shift = 0;
205 }
206 if (order >= zone->compact_order_failed)
207 zone->compact_order_failed = order + 1;
208
209 trace_mm_compaction_defer_reset(zone, order);
210}
211
212/* Returns true if restarting compaction after many failures */
213bool compaction_restarting(struct zone *zone, int order)
214{
215 if (order < zone->compact_order_failed)
216 return false;
217
218 return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
219 zone->compact_considered >= 1UL << zone->compact_defer_shift;
220}
221
222/* Returns true if the pageblock should be scanned for pages to isolate. */
223static inline bool isolation_suitable(struct compact_control *cc,
224 struct page *page)
225{
226 if (cc->ignore_skip_hint)
227 return true;
228
229 return !get_pageblock_skip(page);
230}
231
232static void reset_cached_positions(struct zone *zone)
233{
234 zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
235 zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
236 zone->compact_cached_free_pfn =
237 pageblock_start_pfn(zone_end_pfn(zone) - 1);
238}
239
240/*
241 * Compound pages of >= pageblock_order should consistenly be skipped until
242 * released. It is always pointless to compact pages of such order (if they are
243 * migratable), and the pageblocks they occupy cannot contain any free pages.
244 */
245static bool pageblock_skip_persistent(struct page *page)
246{
247 if (!PageCompound(page))
248 return false;
249
250 page = compound_head(page);
251
252 if (compound_order(page) >= pageblock_order)
253 return true;
254
255 return false;
256}
257
258static bool
259__reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source,
260 bool check_target)
261{
262 struct page *page = pfn_to_online_page(pfn);
263 struct page *block_page;
264 struct page *end_page;
265 unsigned long block_pfn;
266
267 if (!page)
268 return false;
269 if (zone != page_zone(page))
270 return false;
271 if (pageblock_skip_persistent(page))
272 return false;
273
274 /*
275 * If skip is already cleared do no further checking once the
276 * restart points have been set.
277 */
278 if (check_source && check_target && !get_pageblock_skip(page))
279 return true;
280
281 /*
282 * If clearing skip for the target scanner, do not select a
283 * non-movable pageblock as the starting point.
284 */
285 if (!check_source && check_target &&
286 get_pageblock_migratetype(page) != MIGRATE_MOVABLE)
287 return false;
288
289 /* Ensure the start of the pageblock or zone is online and valid */
290 block_pfn = pageblock_start_pfn(pfn);
291 block_pfn = max(block_pfn, zone->zone_start_pfn);
292 block_page = pfn_to_online_page(block_pfn);
293 if (block_page) {
294 page = block_page;
295 pfn = block_pfn;
296 }
297
298 /* Ensure the end of the pageblock or zone is online and valid */
299 block_pfn = pageblock_end_pfn(pfn) - 1;
300 block_pfn = min(block_pfn, zone_end_pfn(zone) - 1);
301 end_page = pfn_to_online_page(block_pfn);
302 if (!end_page)
303 return false;
304
305 /*
306 * Only clear the hint if a sample indicates there is either a
307 * free page or an LRU page in the block. One or other condition
308 * is necessary for the block to be a migration source/target.
309 */
310 do {
311 if (pfn_valid_within(pfn)) {
312 if (check_source && PageLRU(page)) {
313 clear_pageblock_skip(page);
314 return true;
315 }
316
317 if (check_target && PageBuddy(page)) {
318 clear_pageblock_skip(page);
319 return true;
320 }
321 }
322
323 page += (1 << PAGE_ALLOC_COSTLY_ORDER);
324 pfn += (1 << PAGE_ALLOC_COSTLY_ORDER);
325 } while (page <= end_page);
326
327 return false;
328}
329
330/*
331 * This function is called to clear all cached information on pageblocks that
332 * should be skipped for page isolation when the migrate and free page scanner
333 * meet.
334 */
335static void __reset_isolation_suitable(struct zone *zone)
336{
337 unsigned long migrate_pfn = zone->zone_start_pfn;
338 unsigned long free_pfn = zone_end_pfn(zone) - 1;
339 unsigned long reset_migrate = free_pfn;
340 unsigned long reset_free = migrate_pfn;
341 bool source_set = false;
342 bool free_set = false;
343
344 if (!zone->compact_blockskip_flush)
345 return;
346
347 zone->compact_blockskip_flush = false;
348
349 /*
350 * Walk the zone and update pageblock skip information. Source looks
351 * for PageLRU while target looks for PageBuddy. When the scanner
352 * is found, both PageBuddy and PageLRU are checked as the pageblock
353 * is suitable as both source and target.
354 */
355 for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages,
356 free_pfn -= pageblock_nr_pages) {
357 cond_resched();
358
359 /* Update the migrate PFN */
360 if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) &&
361 migrate_pfn < reset_migrate) {
362 source_set = true;
363 reset_migrate = migrate_pfn;
364 zone->compact_init_migrate_pfn = reset_migrate;
365 zone->compact_cached_migrate_pfn[0] = reset_migrate;
366 zone->compact_cached_migrate_pfn[1] = reset_migrate;
367 }
368
369 /* Update the free PFN */
370 if (__reset_isolation_pfn(zone, free_pfn, free_set, true) &&
371 free_pfn > reset_free) {
372 free_set = true;
373 reset_free = free_pfn;
374 zone->compact_init_free_pfn = reset_free;
375 zone->compact_cached_free_pfn = reset_free;
376 }
377 }
378
379 /* Leave no distance if no suitable block was reset */
380 if (reset_migrate >= reset_free) {
381 zone->compact_cached_migrate_pfn[0] = migrate_pfn;
382 zone->compact_cached_migrate_pfn[1] = migrate_pfn;
383 zone->compact_cached_free_pfn = free_pfn;
384 }
385}
386
387void reset_isolation_suitable(pg_data_t *pgdat)
388{
389 int zoneid;
390
391 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
392 struct zone *zone = &pgdat->node_zones[zoneid];
393 if (!populated_zone(zone))
394 continue;
395
396 /* Only flush if a full compaction finished recently */
397 if (zone->compact_blockskip_flush)
398 __reset_isolation_suitable(zone);
399 }
400}
401
402/*
403 * Sets the pageblock skip bit if it was clear. Note that this is a hint as
404 * locks are not required for read/writers. Returns true if it was already set.
405 */
406static bool test_and_set_skip(struct compact_control *cc, struct page *page,
407 unsigned long pfn)
408{
409 bool skip;
410
411 /* Do no update if skip hint is being ignored */
412 if (cc->ignore_skip_hint)
413 return false;
414
415 if (!IS_ALIGNED(pfn, pageblock_nr_pages))
416 return false;
417
418 skip = get_pageblock_skip(page);
419 if (!skip && !cc->no_set_skip_hint)
420 set_pageblock_skip(page);
421
422 return skip;
423}
424
425static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
426{
427 struct zone *zone = cc->zone;
428
429 pfn = pageblock_end_pfn(pfn);
430
431 /* Set for isolation rather than compaction */
432 if (cc->no_set_skip_hint)
433 return;
434
435 if (pfn > zone->compact_cached_migrate_pfn[0])
436 zone->compact_cached_migrate_pfn[0] = pfn;
437 if (cc->mode != MIGRATE_ASYNC &&
438 pfn > zone->compact_cached_migrate_pfn[1])
439 zone->compact_cached_migrate_pfn[1] = pfn;
440}
441
442/*
443 * If no pages were isolated then mark this pageblock to be skipped in the
444 * future. The information is later cleared by __reset_isolation_suitable().
445 */
446static void update_pageblock_skip(struct compact_control *cc,
447 struct page *page, unsigned long pfn)
448{
449 struct zone *zone = cc->zone;
450
451 if (cc->no_set_skip_hint)
452 return;
453
454 if (!page)
455 return;
456
457 set_pageblock_skip(page);
458
459 /* Update where async and sync compaction should restart */
460 if (pfn < zone->compact_cached_free_pfn)
461 zone->compact_cached_free_pfn = pfn;
462}
463#else
464static inline bool isolation_suitable(struct compact_control *cc,
465 struct page *page)
466{
467 return true;
468}
469
470static inline bool pageblock_skip_persistent(struct page *page)
471{
472 return false;
473}
474
475static inline void update_pageblock_skip(struct compact_control *cc,
476 struct page *page, unsigned long pfn)
477{
478}
479
480static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
481{
482}
483
484static bool test_and_set_skip(struct compact_control *cc, struct page *page,
485 unsigned long pfn)
486{
487 return false;
488}
489#endif /* CONFIG_COMPACTION */
490
491/*
492 * Compaction requires the taking of some coarse locks that are potentially
493 * very heavily contended. For async compaction, trylock and record if the
494 * lock is contended. The lock will still be acquired but compaction will
495 * abort when the current block is finished regardless of success rate.
496 * Sync compaction acquires the lock.
497 *
498 * Always returns true which makes it easier to track lock state in callers.
499 */
500static bool compact_lock_irqsave(spinlock_t *lock, unsigned long *flags,
501 struct compact_control *cc)
502 __acquires(lock)
503{
504 /* Track if the lock is contended in async mode */
505 if (cc->mode == MIGRATE_ASYNC && !cc->contended) {
506 if (spin_trylock_irqsave(lock, *flags))
507 return true;
508
509 cc->contended = true;
510 }
511
512 spin_lock_irqsave(lock, *flags);
513 return true;
514}
515
516/*
517 * Compaction requires the taking of some coarse locks that are potentially
518 * very heavily contended. The lock should be periodically unlocked to avoid
519 * having disabled IRQs for a long time, even when there is nobody waiting on
520 * the lock. It might also be that allowing the IRQs will result in
521 * need_resched() becoming true. If scheduling is needed, async compaction
522 * aborts. Sync compaction schedules.
523 * Either compaction type will also abort if a fatal signal is pending.
524 * In either case if the lock was locked, it is dropped and not regained.
525 *
526 * Returns true if compaction should abort due to fatal signal pending, or
527 * async compaction due to need_resched()
528 * Returns false when compaction can continue (sync compaction might have
529 * scheduled)
530 */
531static bool compact_unlock_should_abort(spinlock_t *lock,
532 unsigned long flags, bool *locked, struct compact_control *cc)
533{
534 if (*locked) {
535 spin_unlock_irqrestore(lock, flags);
536 *locked = false;
537 }
538
539 if (fatal_signal_pending(current)) {
540 cc->contended = true;
541 return true;
542 }
543
544 cond_resched();
545
546 return false;
547}
548
549/*
550 * Isolate free pages onto a private freelist. If @strict is true, will abort
551 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
552 * (even though it may still end up isolating some pages).
553 */
554static unsigned long isolate_freepages_block(struct compact_control *cc,
555 unsigned long *start_pfn,
556 unsigned long end_pfn,
557 struct list_head *freelist,
558 unsigned int stride,
559 bool strict)
560{
561 int nr_scanned = 0, total_isolated = 0;
562 struct page *cursor;
563 unsigned long flags = 0;
564 bool locked = false;
565 unsigned long blockpfn = *start_pfn;
566 unsigned int order;
567
568 /* Strict mode is for isolation, speed is secondary */
569 if (strict)
570 stride = 1;
571
572 cursor = pfn_to_page(blockpfn);
573
574 /* Isolate free pages. */
575 for (; blockpfn < end_pfn; blockpfn += stride, cursor += stride) {
576 int isolated;
577 struct page *page = cursor;
578
579 /*
580 * Periodically drop the lock (if held) regardless of its
581 * contention, to give chance to IRQs. Abort if fatal signal
582 * pending or async compaction detects need_resched()
583 */
584 if (!(blockpfn % SWAP_CLUSTER_MAX)
585 && compact_unlock_should_abort(&cc->zone->lock, flags,
586 &locked, cc))
587 break;
588
589 nr_scanned++;
590 if (!pfn_valid_within(blockpfn))
591 goto isolate_fail;
592
593 /*
594 * For compound pages such as THP and hugetlbfs, we can save
595 * potentially a lot of iterations if we skip them at once.
596 * The check is racy, but we can consider only valid values
597 * and the only danger is skipping too much.
598 */
599 if (PageCompound(page)) {
600 const unsigned int order = compound_order(page);
601
602 if (likely(order < MAX_ORDER)) {
603 blockpfn += (1UL << order) - 1;
604 cursor += (1UL << order) - 1;
605 }
606 goto isolate_fail;
607 }
608
609 if (!PageBuddy(page))
610 goto isolate_fail;
611
612 /*
613 * If we already hold the lock, we can skip some rechecking.
614 * Note that if we hold the lock now, checked_pageblock was
615 * already set in some previous iteration (or strict is true),
616 * so it is correct to skip the suitable migration target
617 * recheck as well.
618 */
619 if (!locked) {
620 locked = compact_lock_irqsave(&cc->zone->lock,
621 &flags, cc);
622
623 /* Recheck this is a buddy page under lock */
624 if (!PageBuddy(page))
625 goto isolate_fail;
626 }
627
628 /* Found a free page, will break it into order-0 pages */
629 order = page_order(page);
630 isolated = __isolate_free_page(page, order);
631 if (!isolated)
632 break;
633 set_page_private(page, order);
634
635 total_isolated += isolated;
636 cc->nr_freepages += isolated;
637 list_add_tail(&page->lru, freelist);
638
639 if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
640 blockpfn += isolated;
641 break;
642 }
643 /* Advance to the end of split page */
644 blockpfn += isolated - 1;
645 cursor += isolated - 1;
646 continue;
647
648isolate_fail:
649 if (strict)
650 break;
651 else
652 continue;
653
654 }
655
656 if (locked)
657 spin_unlock_irqrestore(&cc->zone->lock, flags);
658
659 /*
660 * There is a tiny chance that we have read bogus compound_order(),
661 * so be careful to not go outside of the pageblock.
662 */
663 if (unlikely(blockpfn > end_pfn))
664 blockpfn = end_pfn;
665
666 trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
667 nr_scanned, total_isolated);
668
669 /* Record how far we have got within the block */
670 *start_pfn = blockpfn;
671
672 /*
673 * If strict isolation is requested by CMA then check that all the
674 * pages requested were isolated. If there were any failures, 0 is
675 * returned and CMA will fail.
676 */
677 if (strict && blockpfn < end_pfn)
678 total_isolated = 0;
679
680 cc->total_free_scanned += nr_scanned;
681 if (total_isolated)
682 count_compact_events(COMPACTISOLATED, total_isolated);
683 return total_isolated;
684}
685
686/**
687 * isolate_freepages_range() - isolate free pages.
688 * @cc: Compaction control structure.
689 * @start_pfn: The first PFN to start isolating.
690 * @end_pfn: The one-past-last PFN.
691 *
692 * Non-free pages, invalid PFNs, or zone boundaries within the
693 * [start_pfn, end_pfn) range are considered errors, cause function to
694 * undo its actions and return zero.
695 *
696 * Otherwise, function returns one-past-the-last PFN of isolated page
697 * (which may be greater then end_pfn if end fell in a middle of
698 * a free page).
699 */
700unsigned long
701isolate_freepages_range(struct compact_control *cc,
702 unsigned long start_pfn, unsigned long end_pfn)
703{
704 unsigned long isolated, pfn, block_start_pfn, block_end_pfn;
705 LIST_HEAD(freelist);
706
707 pfn = start_pfn;
708 block_start_pfn = pageblock_start_pfn(pfn);
709 if (block_start_pfn < cc->zone->zone_start_pfn)
710 block_start_pfn = cc->zone->zone_start_pfn;
711 block_end_pfn = pageblock_end_pfn(pfn);
712
713 for (; pfn < end_pfn; pfn += isolated,
714 block_start_pfn = block_end_pfn,
715 block_end_pfn += pageblock_nr_pages) {
716 /* Protect pfn from changing by isolate_freepages_block */
717 unsigned long isolate_start_pfn = pfn;
718
719 block_end_pfn = min(block_end_pfn, end_pfn);
720
721 /*
722 * pfn could pass the block_end_pfn if isolated freepage
723 * is more than pageblock order. In this case, we adjust
724 * scanning range to right one.
725 */
726 if (pfn >= block_end_pfn) {
727 block_start_pfn = pageblock_start_pfn(pfn);
728 block_end_pfn = pageblock_end_pfn(pfn);
729 block_end_pfn = min(block_end_pfn, end_pfn);
730 }
731
732 if (!pageblock_pfn_to_page(block_start_pfn,
733 block_end_pfn, cc->zone))
734 break;
735
736 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
737 block_end_pfn, &freelist, 0, true);
738
739 /*
740 * In strict mode, isolate_freepages_block() returns 0 if
741 * there are any holes in the block (ie. invalid PFNs or
742 * non-free pages).
743 */
744 if (!isolated)
745 break;
746
747 /*
748 * If we managed to isolate pages, it is always (1 << n) *
749 * pageblock_nr_pages for some non-negative n. (Max order
750 * page may span two pageblocks).
751 */
752 }
753
754 /* __isolate_free_page() does not map the pages */
755 split_map_pages(&freelist);
756
757 if (pfn < end_pfn) {
758 /* Loop terminated early, cleanup. */
759 release_freepages(&freelist);
760 return 0;
761 }
762
763 /* We don't use freelists for anything. */
764 return pfn;
765}
766
767/* Similar to reclaim, but different enough that they don't share logic */
768static bool too_many_isolated(pg_data_t *pgdat)
769{
770 unsigned long active, inactive, isolated;
771
772 inactive = node_page_state(pgdat, NR_INACTIVE_FILE) +
773 node_page_state(pgdat, NR_INACTIVE_ANON);
774 active = node_page_state(pgdat, NR_ACTIVE_FILE) +
775 node_page_state(pgdat, NR_ACTIVE_ANON);
776 isolated = node_page_state(pgdat, NR_ISOLATED_FILE) +
777 node_page_state(pgdat, NR_ISOLATED_ANON);
778
779 return isolated > (inactive + active) / 2;
780}
781
782/**
783 * isolate_migratepages_block() - isolate all migrate-able pages within
784 * a single pageblock
785 * @cc: Compaction control structure.
786 * @low_pfn: The first PFN to isolate
787 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
788 * @isolate_mode: Isolation mode to be used.
789 *
790 * Isolate all pages that can be migrated from the range specified by
791 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
792 * Returns zero if there is a fatal signal pending, otherwise PFN of the
793 * first page that was not scanned (which may be both less, equal to or more
794 * than end_pfn).
795 *
796 * The pages are isolated on cc->migratepages list (not required to be empty),
797 * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
798 * is neither read nor updated.
799 */
800static unsigned long
801isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
802 unsigned long end_pfn, isolate_mode_t isolate_mode)
803{
804 pg_data_t *pgdat = cc->zone->zone_pgdat;
805 unsigned long nr_scanned = 0, nr_isolated = 0;
806 struct lruvec *lruvec;
807 unsigned long flags = 0;
808 bool locked = false;
809 struct page *page = NULL, *valid_page = NULL;
810 unsigned long start_pfn = low_pfn;
811 bool skip_on_failure = false;
812 unsigned long next_skip_pfn = 0;
813 bool skip_updated = false;
814
815 /*
816 * Ensure that there are not too many pages isolated from the LRU
817 * list by either parallel reclaimers or compaction. If there are,
818 * delay for some time until fewer pages are isolated
819 */
820 while (unlikely(too_many_isolated(pgdat))) {
821 /* async migration should just abort */
822 if (cc->mode == MIGRATE_ASYNC)
823 return 0;
824
825 congestion_wait(BLK_RW_ASYNC, HZ/10);
826
827 if (fatal_signal_pending(current))
828 return 0;
829 }
830
831 cond_resched();
832
833 if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
834 skip_on_failure = true;
835 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
836 }
837
838 /* Time to isolate some pages for migration */
839 for (; low_pfn < end_pfn; low_pfn++) {
840
841 if (skip_on_failure && low_pfn >= next_skip_pfn) {
842 /*
843 * We have isolated all migration candidates in the
844 * previous order-aligned block, and did not skip it due
845 * to failure. We should migrate the pages now and
846 * hopefully succeed compaction.
847 */
848 if (nr_isolated)
849 break;
850
851 /*
852 * We failed to isolate in the previous order-aligned
853 * block. Set the new boundary to the end of the
854 * current block. Note we can't simply increase
855 * next_skip_pfn by 1 << order, as low_pfn might have
856 * been incremented by a higher number due to skipping
857 * a compound or a high-order buddy page in the
858 * previous loop iteration.
859 */
860 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
861 }
862
863 /*
864 * Periodically drop the lock (if held) regardless of its
865 * contention, to give chance to IRQs. Abort completely if
866 * a fatal signal is pending.
867 */
868 if (!(low_pfn % SWAP_CLUSTER_MAX)
869 && compact_unlock_should_abort(&pgdat->lru_lock,
870 flags, &locked, cc)) {
871 low_pfn = 0;
872 goto fatal_pending;
873 }
874
875 if (!pfn_valid_within(low_pfn))
876 goto isolate_fail;
877 nr_scanned++;
878
879 page = pfn_to_page(low_pfn);
880
881 /*
882 * Check if the pageblock has already been marked skipped.
883 * Only the aligned PFN is checked as the caller isolates
884 * COMPACT_CLUSTER_MAX at a time so the second call must
885 * not falsely conclude that the block should be skipped.
886 */
887 if (!valid_page && IS_ALIGNED(low_pfn, pageblock_nr_pages)) {
888 if (!cc->ignore_skip_hint && get_pageblock_skip(page)) {
889 low_pfn = end_pfn;
890 goto isolate_abort;
891 }
892 valid_page = page;
893 }
894
895 /*
896 * Skip if free. We read page order here without zone lock
897 * which is generally unsafe, but the race window is small and
898 * the worst thing that can happen is that we skip some
899 * potential isolation targets.
900 */
901 if (PageBuddy(page)) {
902 unsigned long freepage_order = page_order_unsafe(page);
903
904 /*
905 * Without lock, we cannot be sure that what we got is
906 * a valid page order. Consider only values in the
907 * valid order range to prevent low_pfn overflow.
908 */
909 if (freepage_order > 0 && freepage_order < MAX_ORDER)
910 low_pfn += (1UL << freepage_order) - 1;
911 continue;
912 }
913
914 /*
915 * Regardless of being on LRU, compound pages such as THP and
916 * hugetlbfs are not to be compacted unless we are attempting
917 * an allocation much larger than the huge page size (eg CMA).
918 * We can potentially save a lot of iterations if we skip them
919 * at once. The check is racy, but we can consider only valid
920 * values and the only danger is skipping too much.
921 */
922 if (PageCompound(page) && !cc->alloc_contig) {
923 const unsigned int order = compound_order(page);
924
925 if (likely(order < MAX_ORDER))
926 low_pfn += (1UL << order) - 1;
927 goto isolate_fail;
928 }
929
930 /*
931 * Check may be lockless but that's ok as we recheck later.
932 * It's possible to migrate LRU and non-lru movable pages.
933 * Skip any other type of page
934 */
935 if (!PageLRU(page)) {
936 /*
937 * __PageMovable can return false positive so we need
938 * to verify it under page_lock.
939 */
940 if (unlikely(__PageMovable(page)) &&
941 !PageIsolated(page)) {
942 if (locked) {
943 spin_unlock_irqrestore(&pgdat->lru_lock,
944 flags);
945 locked = false;
946 }
947
948 if (!isolate_movable_page(page, isolate_mode))
949 goto isolate_success;
950 }
951
952 goto isolate_fail;
953 }
954
955 /*
956 * Migration will fail if an anonymous page is pinned in memory,
957 * so avoid taking lru_lock and isolating it unnecessarily in an
958 * admittedly racy check.
959 */
960 if (!page_mapping(page) &&
961 page_count(page) > page_mapcount(page))
962 goto isolate_fail;
963
964 /*
965 * Only allow to migrate anonymous pages in GFP_NOFS context
966 * because those do not depend on fs locks.
967 */
968 if (!(cc->gfp_mask & __GFP_FS) && page_mapping(page))
969 goto isolate_fail;
970
971 /* If we already hold the lock, we can skip some rechecking */
972 if (!locked) {
973 locked = compact_lock_irqsave(&pgdat->lru_lock,
974 &flags, cc);
975
976 /* Try get exclusive access under lock */
977 if (!skip_updated) {
978 skip_updated = true;
979 if (test_and_set_skip(cc, page, low_pfn))
980 goto isolate_abort;
981 }
982
983 /* Recheck PageLRU and PageCompound under lock */
984 if (!PageLRU(page))
985 goto isolate_fail;
986
987 /*
988 * Page become compound since the non-locked check,
989 * and it's on LRU. It can only be a THP so the order
990 * is safe to read and it's 0 for tail pages.
991 */
992 if (unlikely(PageCompound(page) && !cc->alloc_contig)) {
993 low_pfn += compound_nr(page) - 1;
994 goto isolate_fail;
995 }
996 }
997
998 lruvec = mem_cgroup_page_lruvec(page, pgdat);
999
1000 /* Try isolate the page */
1001 if (__isolate_lru_page(page, isolate_mode) != 0)
1002 goto isolate_fail;
1003
1004 /* The whole page is taken off the LRU; skip the tail pages. */
1005 if (PageCompound(page))
1006 low_pfn += compound_nr(page) - 1;
1007
1008 /* Successfully isolated */
1009 del_page_from_lru_list(page, lruvec, page_lru(page));
1010 mod_node_page_state(page_pgdat(page),
1011 NR_ISOLATED_ANON + page_is_file_lru(page),
1012 thp_nr_pages(page));
1013
1014isolate_success:
1015 list_add(&page->lru, &cc->migratepages);
1016 cc->nr_migratepages++;
1017 nr_isolated++;
1018
1019 /*
1020 * Avoid isolating too much unless this block is being
1021 * rescanned (e.g. dirty/writeback pages, parallel allocation)
1022 * or a lock is contended. For contention, isolate quickly to
1023 * potentially remove one source of contention.
1024 */
1025 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX &&
1026 !cc->rescan && !cc->contended) {
1027 ++low_pfn;
1028 break;
1029 }
1030
1031 continue;
1032isolate_fail:
1033 if (!skip_on_failure)
1034 continue;
1035
1036 /*
1037 * We have isolated some pages, but then failed. Release them
1038 * instead of migrating, as we cannot form the cc->order buddy
1039 * page anyway.
1040 */
1041 if (nr_isolated) {
1042 if (locked) {
1043 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
1044 locked = false;
1045 }
1046 putback_movable_pages(&cc->migratepages);
1047 cc->nr_migratepages = 0;
1048 nr_isolated = 0;
1049 }
1050
1051 if (low_pfn < next_skip_pfn) {
1052 low_pfn = next_skip_pfn - 1;
1053 /*
1054 * The check near the loop beginning would have updated
1055 * next_skip_pfn too, but this is a bit simpler.
1056 */
1057 next_skip_pfn += 1UL << cc->order;
1058 }
1059 }
1060
1061 /*
1062 * The PageBuddy() check could have potentially brought us outside
1063 * the range to be scanned.
1064 */
1065 if (unlikely(low_pfn > end_pfn))
1066 low_pfn = end_pfn;
1067
1068isolate_abort:
1069 if (locked)
1070 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
1071
1072 /*
1073 * Updated the cached scanner pfn once the pageblock has been scanned
1074 * Pages will either be migrated in which case there is no point
1075 * scanning in the near future or migration failed in which case the
1076 * failure reason may persist. The block is marked for skipping if
1077 * there were no pages isolated in the block or if the block is
1078 * rescanned twice in a row.
1079 */
1080 if (low_pfn == end_pfn && (!nr_isolated || cc->rescan)) {
1081 if (valid_page && !skip_updated)
1082 set_pageblock_skip(valid_page);
1083 update_cached_migrate(cc, low_pfn);
1084 }
1085
1086 trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
1087 nr_scanned, nr_isolated);
1088
1089fatal_pending:
1090 cc->total_migrate_scanned += nr_scanned;
1091 if (nr_isolated)
1092 count_compact_events(COMPACTISOLATED, nr_isolated);
1093
1094 return low_pfn;
1095}
1096
1097/**
1098 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
1099 * @cc: Compaction control structure.
1100 * @start_pfn: The first PFN to start isolating.
1101 * @end_pfn: The one-past-last PFN.
1102 *
1103 * Returns zero if isolation fails fatally due to e.g. pending signal.
1104 * Otherwise, function returns one-past-the-last PFN of isolated page
1105 * (which may be greater than end_pfn if end fell in a middle of a THP page).
1106 */
1107unsigned long
1108isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
1109 unsigned long end_pfn)
1110{
1111 unsigned long pfn, block_start_pfn, block_end_pfn;
1112
1113 /* Scan block by block. First and last block may be incomplete */
1114 pfn = start_pfn;
1115 block_start_pfn = pageblock_start_pfn(pfn);
1116 if (block_start_pfn < cc->zone->zone_start_pfn)
1117 block_start_pfn = cc->zone->zone_start_pfn;
1118 block_end_pfn = pageblock_end_pfn(pfn);
1119
1120 for (; pfn < end_pfn; pfn = block_end_pfn,
1121 block_start_pfn = block_end_pfn,
1122 block_end_pfn += pageblock_nr_pages) {
1123
1124 block_end_pfn = min(block_end_pfn, end_pfn);
1125
1126 if (!pageblock_pfn_to_page(block_start_pfn,
1127 block_end_pfn, cc->zone))
1128 continue;
1129
1130 pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
1131 ISOLATE_UNEVICTABLE);
1132
1133 if (!pfn)
1134 break;
1135
1136 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
1137 break;
1138 }
1139
1140 return pfn;
1141}
1142
1143#endif /* CONFIG_COMPACTION || CONFIG_CMA */
1144#ifdef CONFIG_COMPACTION
1145
1146static bool suitable_migration_source(struct compact_control *cc,
1147 struct page *page)
1148{
1149 int block_mt;
1150
1151 if (pageblock_skip_persistent(page))
1152 return false;
1153
1154 if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction)
1155 return true;
1156
1157 block_mt = get_pageblock_migratetype(page);
1158
1159 if (cc->migratetype == MIGRATE_MOVABLE)
1160 return is_migrate_movable(block_mt);
1161 else
1162 return block_mt == cc->migratetype;
1163}
1164
1165/* Returns true if the page is within a block suitable for migration to */
1166static bool suitable_migration_target(struct compact_control *cc,
1167 struct page *page)
1168{
1169 /* If the page is a large free page, then disallow migration */
1170 if (PageBuddy(page)) {
1171 /*
1172 * We are checking page_order without zone->lock taken. But
1173 * the only small danger is that we skip a potentially suitable
1174 * pageblock, so it's not worth to check order for valid range.
1175 */
1176 if (page_order_unsafe(page) >= pageblock_order)
1177 return false;
1178 }
1179
1180 if (cc->ignore_block_suitable)
1181 return true;
1182
1183 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1184 if (is_migrate_movable(get_pageblock_migratetype(page)))
1185 return true;
1186
1187 /* Otherwise skip the block */
1188 return false;
1189}
1190
1191static inline unsigned int
1192freelist_scan_limit(struct compact_control *cc)
1193{
1194 unsigned short shift = BITS_PER_LONG - 1;
1195
1196 return (COMPACT_CLUSTER_MAX >> min(shift, cc->fast_search_fail)) + 1;
1197}
1198
1199/*
1200 * Test whether the free scanner has reached the same or lower pageblock than
1201 * the migration scanner, and compaction should thus terminate.
1202 */
1203static inline bool compact_scanners_met(struct compact_control *cc)
1204{
1205 return (cc->free_pfn >> pageblock_order)
1206 <= (cc->migrate_pfn >> pageblock_order);
1207}
1208
1209/*
1210 * Used when scanning for a suitable migration target which scans freelists
1211 * in reverse. Reorders the list such as the unscanned pages are scanned
1212 * first on the next iteration of the free scanner
1213 */
1214static void
1215move_freelist_head(struct list_head *freelist, struct page *freepage)
1216{
1217 LIST_HEAD(sublist);
1218
1219 if (!list_is_last(freelist, &freepage->lru)) {
1220 list_cut_before(&sublist, freelist, &freepage->lru);
1221 if (!list_empty(&sublist))
1222 list_splice_tail(&sublist, freelist);
1223 }
1224}
1225
1226/*
1227 * Similar to move_freelist_head except used by the migration scanner
1228 * when scanning forward. It's possible for these list operations to
1229 * move against each other if they search the free list exactly in
1230 * lockstep.
1231 */
1232static void
1233move_freelist_tail(struct list_head *freelist, struct page *freepage)
1234{
1235 LIST_HEAD(sublist);
1236
1237 if (!list_is_first(freelist, &freepage->lru)) {
1238 list_cut_position(&sublist, freelist, &freepage->lru);
1239 if (!list_empty(&sublist))
1240 list_splice_tail(&sublist, freelist);
1241 }
1242}
1243
1244static void
1245fast_isolate_around(struct compact_control *cc, unsigned long pfn, unsigned long nr_isolated)
1246{
1247 unsigned long start_pfn, end_pfn;
1248 struct page *page = pfn_to_page(pfn);
1249
1250 /* Do not search around if there are enough pages already */
1251 if (cc->nr_freepages >= cc->nr_migratepages)
1252 return;
1253
1254 /* Minimise scanning during async compaction */
1255 if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC)
1256 return;
1257
1258 /* Pageblock boundaries */
1259 start_pfn = pageblock_start_pfn(pfn);
1260 end_pfn = min(pageblock_end_pfn(pfn), zone_end_pfn(cc->zone)) - 1;
1261
1262 /* Scan before */
1263 if (start_pfn != pfn) {
1264 isolate_freepages_block(cc, &start_pfn, pfn, &cc->freepages, 1, false);
1265 if (cc->nr_freepages >= cc->nr_migratepages)
1266 return;
1267 }
1268
1269 /* Scan after */
1270 start_pfn = pfn + nr_isolated;
1271 if (start_pfn < end_pfn)
1272 isolate_freepages_block(cc, &start_pfn, end_pfn, &cc->freepages, 1, false);
1273
1274 /* Skip this pageblock in the future as it's full or nearly full */
1275 if (cc->nr_freepages < cc->nr_migratepages)
1276 set_pageblock_skip(page);
1277}
1278
1279/* Search orders in round-robin fashion */
1280static int next_search_order(struct compact_control *cc, int order)
1281{
1282 order--;
1283 if (order < 0)
1284 order = cc->order - 1;
1285
1286 /* Search wrapped around? */
1287 if (order == cc->search_order) {
1288 cc->search_order--;
1289 if (cc->search_order < 0)
1290 cc->search_order = cc->order - 1;
1291 return -1;
1292 }
1293
1294 return order;
1295}
1296
1297static unsigned long
1298fast_isolate_freepages(struct compact_control *cc)
1299{
1300 unsigned int limit = min(1U, freelist_scan_limit(cc) >> 1);
1301 unsigned int nr_scanned = 0;
1302 unsigned long low_pfn, min_pfn, high_pfn = 0, highest = 0;
1303 unsigned long nr_isolated = 0;
1304 unsigned long distance;
1305 struct page *page = NULL;
1306 bool scan_start = false;
1307 int order;
1308
1309 /* Full compaction passes in a negative order */
1310 if (cc->order <= 0)
1311 return cc->free_pfn;
1312
1313 /*
1314 * If starting the scan, use a deeper search and use the highest
1315 * PFN found if a suitable one is not found.
1316 */
1317 if (cc->free_pfn >= cc->zone->compact_init_free_pfn) {
1318 limit = pageblock_nr_pages >> 1;
1319 scan_start = true;
1320 }
1321
1322 /*
1323 * Preferred point is in the top quarter of the scan space but take
1324 * a pfn from the top half if the search is problematic.
1325 */
1326 distance = (cc->free_pfn - cc->migrate_pfn);
1327 low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2));
1328 min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1));
1329
1330 if (WARN_ON_ONCE(min_pfn > low_pfn))
1331 low_pfn = min_pfn;
1332
1333 /*
1334 * Search starts from the last successful isolation order or the next
1335 * order to search after a previous failure
1336 */
1337 cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order);
1338
1339 for (order = cc->search_order;
1340 !page && order >= 0;
1341 order = next_search_order(cc, order)) {
1342 struct free_area *area = &cc->zone->free_area[order];
1343 struct list_head *freelist;
1344 struct page *freepage;
1345 unsigned long flags;
1346 unsigned int order_scanned = 0;
1347
1348 if (!area->nr_free)
1349 continue;
1350
1351 spin_lock_irqsave(&cc->zone->lock, flags);
1352 freelist = &area->free_list[MIGRATE_MOVABLE];
1353 list_for_each_entry_reverse(freepage, freelist, lru) {
1354 unsigned long pfn;
1355
1356 order_scanned++;
1357 nr_scanned++;
1358 pfn = page_to_pfn(freepage);
1359
1360 if (pfn >= highest)
1361 highest = pageblock_start_pfn(pfn);
1362
1363 if (pfn >= low_pfn) {
1364 cc->fast_search_fail = 0;
1365 cc->search_order = order;
1366 page = freepage;
1367 break;
1368 }
1369
1370 if (pfn >= min_pfn && pfn > high_pfn) {
1371 high_pfn = pfn;
1372
1373 /* Shorten the scan if a candidate is found */
1374 limit >>= 1;
1375 }
1376
1377 if (order_scanned >= limit)
1378 break;
1379 }
1380
1381 /* Use a minimum pfn if a preferred one was not found */
1382 if (!page && high_pfn) {
1383 page = pfn_to_page(high_pfn);
1384
1385 /* Update freepage for the list reorder below */
1386 freepage = page;
1387 }
1388
1389 /* Reorder to so a future search skips recent pages */
1390 move_freelist_head(freelist, freepage);
1391
1392 /* Isolate the page if available */
1393 if (page) {
1394 if (__isolate_free_page(page, order)) {
1395 set_page_private(page, order);
1396 nr_isolated = 1 << order;
1397 cc->nr_freepages += nr_isolated;
1398 list_add_tail(&page->lru, &cc->freepages);
1399 count_compact_events(COMPACTISOLATED, nr_isolated);
1400 } else {
1401 /* If isolation fails, abort the search */
1402 order = cc->search_order + 1;
1403 page = NULL;
1404 }
1405 }
1406
1407 spin_unlock_irqrestore(&cc->zone->lock, flags);
1408
1409 /*
1410 * Smaller scan on next order so the total scan ig related
1411 * to freelist_scan_limit.
1412 */
1413 if (order_scanned >= limit)
1414 limit = min(1U, limit >> 1);
1415 }
1416
1417 if (!page) {
1418 cc->fast_search_fail++;
1419 if (scan_start) {
1420 /*
1421 * Use the highest PFN found above min. If one was
1422 * not found, be pessimistic for direct compaction
1423 * and use the min mark.
1424 */
1425 if (highest) {
1426 page = pfn_to_page(highest);
1427 cc->free_pfn = highest;
1428 } else {
1429 if (cc->direct_compaction && pfn_valid(min_pfn)) {
1430 page = pageblock_pfn_to_page(min_pfn,
1431 pageblock_end_pfn(min_pfn),
1432 cc->zone);
1433 cc->free_pfn = min_pfn;
1434 }
1435 }
1436 }
1437 }
1438
1439 if (highest && highest >= cc->zone->compact_cached_free_pfn) {
1440 highest -= pageblock_nr_pages;
1441 cc->zone->compact_cached_free_pfn = highest;
1442 }
1443
1444 cc->total_free_scanned += nr_scanned;
1445 if (!page)
1446 return cc->free_pfn;
1447
1448 low_pfn = page_to_pfn(page);
1449 fast_isolate_around(cc, low_pfn, nr_isolated);
1450 return low_pfn;
1451}
1452
1453/*
1454 * Based on information in the current compact_control, find blocks
1455 * suitable for isolating free pages from and then isolate them.
1456 */
1457static void isolate_freepages(struct compact_control *cc)
1458{
1459 struct zone *zone = cc->zone;
1460 struct page *page;
1461 unsigned long block_start_pfn; /* start of current pageblock */
1462 unsigned long isolate_start_pfn; /* exact pfn we start at */
1463 unsigned long block_end_pfn; /* end of current pageblock */
1464 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
1465 struct list_head *freelist = &cc->freepages;
1466 unsigned int stride;
1467
1468 /* Try a small search of the free lists for a candidate */
1469 isolate_start_pfn = fast_isolate_freepages(cc);
1470 if (cc->nr_freepages)
1471 goto splitmap;
1472
1473 /*
1474 * Initialise the free scanner. The starting point is where we last
1475 * successfully isolated from, zone-cached value, or the end of the
1476 * zone when isolating for the first time. For looping we also need
1477 * this pfn aligned down to the pageblock boundary, because we do
1478 * block_start_pfn -= pageblock_nr_pages in the for loop.
1479 * For ending point, take care when isolating in last pageblock of a
1480 * zone which ends in the middle of a pageblock.
1481 * The low boundary is the end of the pageblock the migration scanner
1482 * is using.
1483 */
1484 isolate_start_pfn = cc->free_pfn;
1485 block_start_pfn = pageblock_start_pfn(isolate_start_pfn);
1486 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1487 zone_end_pfn(zone));
1488 low_pfn = pageblock_end_pfn(cc->migrate_pfn);
1489 stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1;
1490
1491 /*
1492 * Isolate free pages until enough are available to migrate the
1493 * pages on cc->migratepages. We stop searching if the migrate
1494 * and free page scanners meet or enough free pages are isolated.
1495 */
1496 for (; block_start_pfn >= low_pfn;
1497 block_end_pfn = block_start_pfn,
1498 block_start_pfn -= pageblock_nr_pages,
1499 isolate_start_pfn = block_start_pfn) {
1500 unsigned long nr_isolated;
1501
1502 /*
1503 * This can iterate a massively long zone without finding any
1504 * suitable migration targets, so periodically check resched.
1505 */
1506 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)))
1507 cond_resched();
1508
1509 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1510 zone);
1511 if (!page)
1512 continue;
1513
1514 /* Check the block is suitable for migration */
1515 if (!suitable_migration_target(cc, page))
1516 continue;
1517
1518 /* If isolation recently failed, do not retry */
1519 if (!isolation_suitable(cc, page))
1520 continue;
1521
1522 /* Found a block suitable for isolating free pages from. */
1523 nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn,
1524 block_end_pfn, freelist, stride, false);
1525
1526 /* Update the skip hint if the full pageblock was scanned */
1527 if (isolate_start_pfn == block_end_pfn)
1528 update_pageblock_skip(cc, page, block_start_pfn);
1529
1530 /* Are enough freepages isolated? */
1531 if (cc->nr_freepages >= cc->nr_migratepages) {
1532 if (isolate_start_pfn >= block_end_pfn) {
1533 /*
1534 * Restart at previous pageblock if more
1535 * freepages can be isolated next time.
1536 */
1537 isolate_start_pfn =
1538 block_start_pfn - pageblock_nr_pages;
1539 }
1540 break;
1541 } else if (isolate_start_pfn < block_end_pfn) {
1542 /*
1543 * If isolation failed early, do not continue
1544 * needlessly.
1545 */
1546 break;
1547 }
1548
1549 /* Adjust stride depending on isolation */
1550 if (nr_isolated) {
1551 stride = 1;
1552 continue;
1553 }
1554 stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1);
1555 }
1556
1557 /*
1558 * Record where the free scanner will restart next time. Either we
1559 * broke from the loop and set isolate_start_pfn based on the last
1560 * call to isolate_freepages_block(), or we met the migration scanner
1561 * and the loop terminated due to isolate_start_pfn < low_pfn
1562 */
1563 cc->free_pfn = isolate_start_pfn;
1564
1565splitmap:
1566 /* __isolate_free_page() does not map the pages */
1567 split_map_pages(freelist);
1568}
1569
1570/*
1571 * This is a migrate-callback that "allocates" freepages by taking pages
1572 * from the isolated freelists in the block we are migrating to.
1573 */
1574static struct page *compaction_alloc(struct page *migratepage,
1575 unsigned long data)
1576{
1577 struct compact_control *cc = (struct compact_control *)data;
1578 struct page *freepage;
1579
1580 if (list_empty(&cc->freepages)) {
1581 isolate_freepages(cc);
1582
1583 if (list_empty(&cc->freepages))
1584 return NULL;
1585 }
1586
1587 freepage = list_entry(cc->freepages.next, struct page, lru);
1588 list_del(&freepage->lru);
1589 cc->nr_freepages--;
1590
1591 return freepage;
1592}
1593
1594/*
1595 * This is a migrate-callback that "frees" freepages back to the isolated
1596 * freelist. All pages on the freelist are from the same zone, so there is no
1597 * special handling needed for NUMA.
1598 */
1599static void compaction_free(struct page *page, unsigned long data)
1600{
1601 struct compact_control *cc = (struct compact_control *)data;
1602
1603 list_add(&page->lru, &cc->freepages);
1604 cc->nr_freepages++;
1605}
1606
1607/* possible outcome of isolate_migratepages */
1608typedef enum {
1609 ISOLATE_ABORT, /* Abort compaction now */
1610 ISOLATE_NONE, /* No pages isolated, continue scanning */
1611 ISOLATE_SUCCESS, /* Pages isolated, migrate */
1612} isolate_migrate_t;
1613
1614/*
1615 * Allow userspace to control policy on scanning the unevictable LRU for
1616 * compactable pages.
1617 */
1618#ifdef CONFIG_PREEMPT_RT
1619int sysctl_compact_unevictable_allowed __read_mostly = 0;
1620#else
1621int sysctl_compact_unevictable_allowed __read_mostly = 1;
1622#endif
1623
1624static inline void
1625update_fast_start_pfn(struct compact_control *cc, unsigned long pfn)
1626{
1627 if (cc->fast_start_pfn == ULONG_MAX)
1628 return;
1629
1630 if (!cc->fast_start_pfn)
1631 cc->fast_start_pfn = pfn;
1632
1633 cc->fast_start_pfn = min(cc->fast_start_pfn, pfn);
1634}
1635
1636static inline unsigned long
1637reinit_migrate_pfn(struct compact_control *cc)
1638{
1639 if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX)
1640 return cc->migrate_pfn;
1641
1642 cc->migrate_pfn = cc->fast_start_pfn;
1643 cc->fast_start_pfn = ULONG_MAX;
1644
1645 return cc->migrate_pfn;
1646}
1647
1648/*
1649 * Briefly search the free lists for a migration source that already has
1650 * some free pages to reduce the number of pages that need migration
1651 * before a pageblock is free.
1652 */
1653static unsigned long fast_find_migrateblock(struct compact_control *cc)
1654{
1655 unsigned int limit = freelist_scan_limit(cc);
1656 unsigned int nr_scanned = 0;
1657 unsigned long distance;
1658 unsigned long pfn = cc->migrate_pfn;
1659 unsigned long high_pfn;
1660 int order;
1661
1662 /* Skip hints are relied on to avoid repeats on the fast search */
1663 if (cc->ignore_skip_hint)
1664 return pfn;
1665
1666 /*
1667 * If the migrate_pfn is not at the start of a zone or the start
1668 * of a pageblock then assume this is a continuation of a previous
1669 * scan restarted due to COMPACT_CLUSTER_MAX.
1670 */
1671 if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn))
1672 return pfn;
1673
1674 /*
1675 * For smaller orders, just linearly scan as the number of pages
1676 * to migrate should be relatively small and does not necessarily
1677 * justify freeing up a large block for a small allocation.
1678 */
1679 if (cc->order <= PAGE_ALLOC_COSTLY_ORDER)
1680 return pfn;
1681
1682 /*
1683 * Only allow kcompactd and direct requests for movable pages to
1684 * quickly clear out a MOVABLE pageblock for allocation. This
1685 * reduces the risk that a large movable pageblock is freed for
1686 * an unmovable/reclaimable small allocation.
1687 */
1688 if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE)
1689 return pfn;
1690
1691 /*
1692 * When starting the migration scanner, pick any pageblock within the
1693 * first half of the search space. Otherwise try and pick a pageblock
1694 * within the first eighth to reduce the chances that a migration
1695 * target later becomes a source.
1696 */
1697 distance = (cc->free_pfn - cc->migrate_pfn) >> 1;
1698 if (cc->migrate_pfn != cc->zone->zone_start_pfn)
1699 distance >>= 2;
1700 high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance);
1701
1702 for (order = cc->order - 1;
1703 order >= PAGE_ALLOC_COSTLY_ORDER && pfn == cc->migrate_pfn && nr_scanned < limit;
1704 order--) {
1705 struct free_area *area = &cc->zone->free_area[order];
1706 struct list_head *freelist;
1707 unsigned long flags;
1708 struct page *freepage;
1709
1710 if (!area->nr_free)
1711 continue;
1712
1713 spin_lock_irqsave(&cc->zone->lock, flags);
1714 freelist = &area->free_list[MIGRATE_MOVABLE];
1715 list_for_each_entry(freepage, freelist, lru) {
1716 unsigned long free_pfn;
1717
1718 nr_scanned++;
1719 free_pfn = page_to_pfn(freepage);
1720 if (free_pfn < high_pfn) {
1721 /*
1722 * Avoid if skipped recently. Ideally it would
1723 * move to the tail but even safe iteration of
1724 * the list assumes an entry is deleted, not
1725 * reordered.
1726 */
1727 if (get_pageblock_skip(freepage)) {
1728 if (list_is_last(freelist, &freepage->lru))
1729 break;
1730
1731 continue;
1732 }
1733
1734 /* Reorder to so a future search skips recent pages */
1735 move_freelist_tail(freelist, freepage);
1736
1737 update_fast_start_pfn(cc, free_pfn);
1738 pfn = pageblock_start_pfn(free_pfn);
1739 cc->fast_search_fail = 0;
1740 set_pageblock_skip(freepage);
1741 break;
1742 }
1743
1744 if (nr_scanned >= limit) {
1745 cc->fast_search_fail++;
1746 move_freelist_tail(freelist, freepage);
1747 break;
1748 }
1749 }
1750 spin_unlock_irqrestore(&cc->zone->lock, flags);
1751 }
1752
1753 cc->total_migrate_scanned += nr_scanned;
1754
1755 /*
1756 * If fast scanning failed then use a cached entry for a page block
1757 * that had free pages as the basis for starting a linear scan.
1758 */
1759 if (pfn == cc->migrate_pfn)
1760 pfn = reinit_migrate_pfn(cc);
1761
1762 return pfn;
1763}
1764
1765/*
1766 * Isolate all pages that can be migrated from the first suitable block,
1767 * starting at the block pointed to by the migrate scanner pfn within
1768 * compact_control.
1769 */
1770static isolate_migrate_t isolate_migratepages(struct compact_control *cc)
1771{
1772 unsigned long block_start_pfn;
1773 unsigned long block_end_pfn;
1774 unsigned long low_pfn;
1775 struct page *page;
1776 const isolate_mode_t isolate_mode =
1777 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1778 (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1779 bool fast_find_block;
1780
1781 /*
1782 * Start at where we last stopped, or beginning of the zone as
1783 * initialized by compact_zone(). The first failure will use
1784 * the lowest PFN as the starting point for linear scanning.
1785 */
1786 low_pfn = fast_find_migrateblock(cc);
1787 block_start_pfn = pageblock_start_pfn(low_pfn);
1788 if (block_start_pfn < cc->zone->zone_start_pfn)
1789 block_start_pfn = cc->zone->zone_start_pfn;
1790
1791 /*
1792 * fast_find_migrateblock marks a pageblock skipped so to avoid
1793 * the isolation_suitable check below, check whether the fast
1794 * search was successful.
1795 */
1796 fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail;
1797
1798 /* Only scan within a pageblock boundary */
1799 block_end_pfn = pageblock_end_pfn(low_pfn);
1800
1801 /*
1802 * Iterate over whole pageblocks until we find the first suitable.
1803 * Do not cross the free scanner.
1804 */
1805 for (; block_end_pfn <= cc->free_pfn;
1806 fast_find_block = false,
1807 low_pfn = block_end_pfn,
1808 block_start_pfn = block_end_pfn,
1809 block_end_pfn += pageblock_nr_pages) {
1810
1811 /*
1812 * This can potentially iterate a massively long zone with
1813 * many pageblocks unsuitable, so periodically check if we
1814 * need to schedule.
1815 */
1816 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)))
1817 cond_resched();
1818
1819 page = pageblock_pfn_to_page(block_start_pfn,
1820 block_end_pfn, cc->zone);
1821 if (!page)
1822 continue;
1823
1824 /*
1825 * If isolation recently failed, do not retry. Only check the
1826 * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock
1827 * to be visited multiple times. Assume skip was checked
1828 * before making it "skip" so other compaction instances do
1829 * not scan the same block.
1830 */
1831 if (IS_ALIGNED(low_pfn, pageblock_nr_pages) &&
1832 !fast_find_block && !isolation_suitable(cc, page))
1833 continue;
1834
1835 /*
1836 * For async compaction, also only scan in MOVABLE blocks
1837 * without huge pages. Async compaction is optimistic to see
1838 * if the minimum amount of work satisfies the allocation.
1839 * The cached PFN is updated as it's possible that all
1840 * remaining blocks between source and target are unsuitable
1841 * and the compaction scanners fail to meet.
1842 */
1843 if (!suitable_migration_source(cc, page)) {
1844 update_cached_migrate(cc, block_end_pfn);
1845 continue;
1846 }
1847
1848 /* Perform the isolation */
1849 low_pfn = isolate_migratepages_block(cc, low_pfn,
1850 block_end_pfn, isolate_mode);
1851
1852 if (!low_pfn)
1853 return ISOLATE_ABORT;
1854
1855 /*
1856 * Either we isolated something and proceed with migration. Or
1857 * we failed and compact_zone should decide if we should
1858 * continue or not.
1859 */
1860 break;
1861 }
1862
1863 /* Record where migration scanner will be restarted. */
1864 cc->migrate_pfn = low_pfn;
1865
1866 return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
1867}
1868
1869/*
1870 * order == -1 is expected when compacting via
1871 * /proc/sys/vm/compact_memory
1872 */
1873static inline bool is_via_compact_memory(int order)
1874{
1875 return order == -1;
1876}
1877
1878static bool kswapd_is_running(pg_data_t *pgdat)
1879{
1880 return pgdat->kswapd && (pgdat->kswapd->state == TASK_RUNNING);
1881}
1882
1883/*
1884 * A zone's fragmentation score is the external fragmentation wrt to the
1885 * COMPACTION_HPAGE_ORDER scaled by the zone's size. It returns a value
1886 * in the range [0, 100].
1887 *
1888 * The scaling factor ensures that proactive compaction focuses on larger
1889 * zones like ZONE_NORMAL, rather than smaller, specialized zones like
1890 * ZONE_DMA32. For smaller zones, the score value remains close to zero,
1891 * and thus never exceeds the high threshold for proactive compaction.
1892 */
1893static unsigned int fragmentation_score_zone(struct zone *zone)
1894{
1895 unsigned long score;
1896
1897 score = zone->present_pages *
1898 extfrag_for_order(zone, COMPACTION_HPAGE_ORDER);
1899 return div64_ul(score, zone->zone_pgdat->node_present_pages + 1);
1900}
1901
1902/*
1903 * The per-node proactive (background) compaction process is started by its
1904 * corresponding kcompactd thread when the node's fragmentation score
1905 * exceeds the high threshold. The compaction process remains active till
1906 * the node's score falls below the low threshold, or one of the back-off
1907 * conditions is met.
1908 */
1909static unsigned int fragmentation_score_node(pg_data_t *pgdat)
1910{
1911 unsigned int score = 0;
1912 int zoneid;
1913
1914 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
1915 struct zone *zone;
1916
1917 zone = &pgdat->node_zones[zoneid];
1918 score += fragmentation_score_zone(zone);
1919 }
1920
1921 return score;
1922}
1923
1924static unsigned int fragmentation_score_wmark(pg_data_t *pgdat, bool low)
1925{
1926 unsigned int wmark_low;
1927
1928 /*
1929 * Cap the low watermak to avoid excessive compaction
1930 * activity in case a user sets the proactivess tunable
1931 * close to 100 (maximum).
1932 */
1933 wmark_low = max(100U - sysctl_compaction_proactiveness, 5U);
1934 return low ? wmark_low : min(wmark_low + 10, 100U);
1935}
1936
1937static bool should_proactive_compact_node(pg_data_t *pgdat)
1938{
1939 int wmark_high;
1940
1941 if (!sysctl_compaction_proactiveness || kswapd_is_running(pgdat))
1942 return false;
1943
1944 wmark_high = fragmentation_score_wmark(pgdat, false);
1945 return fragmentation_score_node(pgdat) > wmark_high;
1946}
1947
1948static enum compact_result __compact_finished(struct compact_control *cc)
1949{
1950 unsigned int order;
1951 const int migratetype = cc->migratetype;
1952 int ret;
1953
1954 /* Compaction run completes if the migrate and free scanner meet */
1955 if (compact_scanners_met(cc)) {
1956 /* Let the next compaction start anew. */
1957 reset_cached_positions(cc->zone);
1958
1959 /*
1960 * Mark that the PG_migrate_skip information should be cleared
1961 * by kswapd when it goes to sleep. kcompactd does not set the
1962 * flag itself as the decision to be clear should be directly
1963 * based on an allocation request.
1964 */
1965 if (cc->direct_compaction)
1966 cc->zone->compact_blockskip_flush = true;
1967
1968 if (cc->whole_zone)
1969 return COMPACT_COMPLETE;
1970 else
1971 return COMPACT_PARTIAL_SKIPPED;
1972 }
1973
1974 if (cc->proactive_compaction) {
1975 int score, wmark_low;
1976 pg_data_t *pgdat;
1977
1978 pgdat = cc->zone->zone_pgdat;
1979 if (kswapd_is_running(pgdat))
1980 return COMPACT_PARTIAL_SKIPPED;
1981
1982 score = fragmentation_score_zone(cc->zone);
1983 wmark_low = fragmentation_score_wmark(pgdat, true);
1984
1985 if (score > wmark_low)
1986 ret = COMPACT_CONTINUE;
1987 else
1988 ret = COMPACT_SUCCESS;
1989
1990 goto out;
1991 }
1992
1993 if (is_via_compact_memory(cc->order))
1994 return COMPACT_CONTINUE;
1995
1996 /*
1997 * Always finish scanning a pageblock to reduce the possibility of
1998 * fallbacks in the future. This is particularly important when
1999 * migration source is unmovable/reclaimable but it's not worth
2000 * special casing.
2001 */
2002 if (!IS_ALIGNED(cc->migrate_pfn, pageblock_nr_pages))
2003 return COMPACT_CONTINUE;
2004
2005 /* Direct compactor: Is a suitable page free? */
2006 ret = COMPACT_NO_SUITABLE_PAGE;
2007 for (order = cc->order; order < MAX_ORDER; order++) {
2008 struct free_area *area = &cc->zone->free_area[order];
2009 bool can_steal;
2010
2011 /* Job done if page is free of the right migratetype */
2012 if (!free_area_empty(area, migratetype))
2013 return COMPACT_SUCCESS;
2014
2015#ifdef CONFIG_CMA
2016 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
2017 if (migratetype == MIGRATE_MOVABLE &&
2018 !free_area_empty(area, MIGRATE_CMA))
2019 return COMPACT_SUCCESS;
2020#endif
2021 /*
2022 * Job done if allocation would steal freepages from
2023 * other migratetype buddy lists.
2024 */
2025 if (find_suitable_fallback(area, order, migratetype,
2026 true, &can_steal) != -1) {
2027
2028 /* movable pages are OK in any pageblock */
2029 if (migratetype == MIGRATE_MOVABLE)
2030 return COMPACT_SUCCESS;
2031
2032 /*
2033 * We are stealing for a non-movable allocation. Make
2034 * sure we finish compacting the current pageblock
2035 * first so it is as free as possible and we won't
2036 * have to steal another one soon. This only applies
2037 * to sync compaction, as async compaction operates
2038 * on pageblocks of the same migratetype.
2039 */
2040 if (cc->mode == MIGRATE_ASYNC ||
2041 IS_ALIGNED(cc->migrate_pfn,
2042 pageblock_nr_pages)) {
2043 return COMPACT_SUCCESS;
2044 }
2045
2046 ret = COMPACT_CONTINUE;
2047 break;
2048 }
2049 }
2050
2051out:
2052 if (cc->contended || fatal_signal_pending(current))
2053 ret = COMPACT_CONTENDED;
2054
2055 return ret;
2056}
2057
2058static enum compact_result compact_finished(struct compact_control *cc)
2059{
2060 int ret;
2061
2062 ret = __compact_finished(cc);
2063 trace_mm_compaction_finished(cc->zone, cc->order, ret);
2064 if (ret == COMPACT_NO_SUITABLE_PAGE)
2065 ret = COMPACT_CONTINUE;
2066
2067 return ret;
2068}
2069
2070/*
2071 * compaction_suitable: Is this suitable to run compaction on this zone now?
2072 * Returns
2073 * COMPACT_SKIPPED - If there are too few free pages for compaction
2074 * COMPACT_SUCCESS - If the allocation would succeed without compaction
2075 * COMPACT_CONTINUE - If compaction should run now
2076 */
2077static enum compact_result __compaction_suitable(struct zone *zone, int order,
2078 unsigned int alloc_flags,
2079 int highest_zoneidx,
2080 unsigned long wmark_target)
2081{
2082 unsigned long watermark;
2083
2084 if (is_via_compact_memory(order))
2085 return COMPACT_CONTINUE;
2086
2087 watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
2088 /*
2089 * If watermarks for high-order allocation are already met, there
2090 * should be no need for compaction at all.
2091 */
2092 if (zone_watermark_ok(zone, order, watermark, highest_zoneidx,
2093 alloc_flags))
2094 return COMPACT_SUCCESS;
2095
2096 /*
2097 * Watermarks for order-0 must be met for compaction to be able to
2098 * isolate free pages for migration targets. This means that the
2099 * watermark and alloc_flags have to match, or be more pessimistic than
2100 * the check in __isolate_free_page(). We don't use the direct
2101 * compactor's alloc_flags, as they are not relevant for freepage
2102 * isolation. We however do use the direct compactor's highest_zoneidx
2103 * to skip over zones where lowmem reserves would prevent allocation
2104 * even if compaction succeeds.
2105 * For costly orders, we require low watermark instead of min for
2106 * compaction to proceed to increase its chances.
2107 * ALLOC_CMA is used, as pages in CMA pageblocks are considered
2108 * suitable migration targets
2109 */
2110 watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
2111 low_wmark_pages(zone) : min_wmark_pages(zone);
2112 watermark += compact_gap(order);
2113 if (!__zone_watermark_ok(zone, 0, watermark, highest_zoneidx,
2114 ALLOC_CMA, wmark_target))
2115 return COMPACT_SKIPPED;
2116
2117 return COMPACT_CONTINUE;
2118}
2119
2120enum compact_result compaction_suitable(struct zone *zone, int order,
2121 unsigned int alloc_flags,
2122 int highest_zoneidx)
2123{
2124 enum compact_result ret;
2125 int fragindex;
2126
2127 ret = __compaction_suitable(zone, order, alloc_flags, highest_zoneidx,
2128 zone_page_state(zone, NR_FREE_PAGES));
2129 /*
2130 * fragmentation index determines if allocation failures are due to
2131 * low memory or external fragmentation
2132 *
2133 * index of -1000 would imply allocations might succeed depending on
2134 * watermarks, but we already failed the high-order watermark check
2135 * index towards 0 implies failure is due to lack of memory
2136 * index towards 1000 implies failure is due to fragmentation
2137 *
2138 * Only compact if a failure would be due to fragmentation. Also
2139 * ignore fragindex for non-costly orders where the alternative to
2140 * a successful reclaim/compaction is OOM. Fragindex and the
2141 * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
2142 * excessive compaction for costly orders, but it should not be at the
2143 * expense of system stability.
2144 */
2145 if (ret == COMPACT_CONTINUE && (order > PAGE_ALLOC_COSTLY_ORDER)) {
2146 fragindex = fragmentation_index(zone, order);
2147 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
2148 ret = COMPACT_NOT_SUITABLE_ZONE;
2149 }
2150
2151 trace_mm_compaction_suitable(zone, order, ret);
2152 if (ret == COMPACT_NOT_SUITABLE_ZONE)
2153 ret = COMPACT_SKIPPED;
2154
2155 return ret;
2156}
2157
2158bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
2159 int alloc_flags)
2160{
2161 struct zone *zone;
2162 struct zoneref *z;
2163
2164 /*
2165 * Make sure at least one zone would pass __compaction_suitable if we continue
2166 * retrying the reclaim.
2167 */
2168 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2169 ac->highest_zoneidx, ac->nodemask) {
2170 unsigned long available;
2171 enum compact_result compact_result;
2172
2173 /*
2174 * Do not consider all the reclaimable memory because we do not
2175 * want to trash just for a single high order allocation which
2176 * is even not guaranteed to appear even if __compaction_suitable
2177 * is happy about the watermark check.
2178 */
2179 available = zone_reclaimable_pages(zone) / order;
2180 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
2181 compact_result = __compaction_suitable(zone, order, alloc_flags,
2182 ac->highest_zoneidx, available);
2183 if (compact_result != COMPACT_SKIPPED)
2184 return true;
2185 }
2186
2187 return false;
2188}
2189
2190static enum compact_result
2191compact_zone(struct compact_control *cc, struct capture_control *capc)
2192{
2193 enum compact_result ret;
2194 unsigned long start_pfn = cc->zone->zone_start_pfn;
2195 unsigned long end_pfn = zone_end_pfn(cc->zone);
2196 unsigned long last_migrated_pfn;
2197 const bool sync = cc->mode != MIGRATE_ASYNC;
2198 bool update_cached;
2199
2200 /*
2201 * These counters track activities during zone compaction. Initialize
2202 * them before compacting a new zone.
2203 */
2204 cc->total_migrate_scanned = 0;
2205 cc->total_free_scanned = 0;
2206 cc->nr_migratepages = 0;
2207 cc->nr_freepages = 0;
2208 INIT_LIST_HEAD(&cc->freepages);
2209 INIT_LIST_HEAD(&cc->migratepages);
2210
2211 cc->migratetype = gfp_migratetype(cc->gfp_mask);
2212 ret = compaction_suitable(cc->zone, cc->order, cc->alloc_flags,
2213 cc->highest_zoneidx);
2214 /* Compaction is likely to fail */
2215 if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED)
2216 return ret;
2217
2218 /* huh, compaction_suitable is returning something unexpected */
2219 VM_BUG_ON(ret != COMPACT_CONTINUE);
2220
2221 /*
2222 * Clear pageblock skip if there were failures recently and compaction
2223 * is about to be retried after being deferred.
2224 */
2225 if (compaction_restarting(cc->zone, cc->order))
2226 __reset_isolation_suitable(cc->zone);
2227
2228 /*
2229 * Setup to move all movable pages to the end of the zone. Used cached
2230 * information on where the scanners should start (unless we explicitly
2231 * want to compact the whole zone), but check that it is initialised
2232 * by ensuring the values are within zone boundaries.
2233 */
2234 cc->fast_start_pfn = 0;
2235 if (cc->whole_zone) {
2236 cc->migrate_pfn = start_pfn;
2237 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2238 } else {
2239 cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync];
2240 cc->free_pfn = cc->zone->compact_cached_free_pfn;
2241 if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
2242 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2243 cc->zone->compact_cached_free_pfn = cc->free_pfn;
2244 }
2245 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
2246 cc->migrate_pfn = start_pfn;
2247 cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
2248 cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
2249 }
2250
2251 if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn)
2252 cc->whole_zone = true;
2253 }
2254
2255 last_migrated_pfn = 0;
2256
2257 /*
2258 * Migrate has separate cached PFNs for ASYNC and SYNC* migration on
2259 * the basis that some migrations will fail in ASYNC mode. However,
2260 * if the cached PFNs match and pageblocks are skipped due to having
2261 * no isolation candidates, then the sync state does not matter.
2262 * Until a pageblock with isolation candidates is found, keep the
2263 * cached PFNs in sync to avoid revisiting the same blocks.
2264 */
2265 update_cached = !sync &&
2266 cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1];
2267
2268 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
2269 cc->free_pfn, end_pfn, sync);
2270
2271 migrate_prep_local();
2272
2273 while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) {
2274 int err;
2275 unsigned long start_pfn = cc->migrate_pfn;
2276
2277 /*
2278 * Avoid multiple rescans which can happen if a page cannot be
2279 * isolated (dirty/writeback in async mode) or if the migrated
2280 * pages are being allocated before the pageblock is cleared.
2281 * The first rescan will capture the entire pageblock for
2282 * migration. If it fails, it'll be marked skip and scanning
2283 * will proceed as normal.
2284 */
2285 cc->rescan = false;
2286 if (pageblock_start_pfn(last_migrated_pfn) ==
2287 pageblock_start_pfn(start_pfn)) {
2288 cc->rescan = true;
2289 }
2290
2291 switch (isolate_migratepages(cc)) {
2292 case ISOLATE_ABORT:
2293 ret = COMPACT_CONTENDED;
2294 putback_movable_pages(&cc->migratepages);
2295 cc->nr_migratepages = 0;
2296 goto out;
2297 case ISOLATE_NONE:
2298 if (update_cached) {
2299 cc->zone->compact_cached_migrate_pfn[1] =
2300 cc->zone->compact_cached_migrate_pfn[0];
2301 }
2302
2303 /*
2304 * We haven't isolated and migrated anything, but
2305 * there might still be unflushed migrations from
2306 * previous cc->order aligned block.
2307 */
2308 goto check_drain;
2309 case ISOLATE_SUCCESS:
2310 update_cached = false;
2311 last_migrated_pfn = start_pfn;
2312 ;
2313 }
2314
2315 err = migrate_pages(&cc->migratepages, compaction_alloc,
2316 compaction_free, (unsigned long)cc, cc->mode,
2317 MR_COMPACTION);
2318
2319 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
2320 &cc->migratepages);
2321
2322 /* All pages were either migrated or will be released */
2323 cc->nr_migratepages = 0;
2324 if (err) {
2325 putback_movable_pages(&cc->migratepages);
2326 /*
2327 * migrate_pages() may return -ENOMEM when scanners meet
2328 * and we want compact_finished() to detect it
2329 */
2330 if (err == -ENOMEM && !compact_scanners_met(cc)) {
2331 ret = COMPACT_CONTENDED;
2332 goto out;
2333 }
2334 /*
2335 * We failed to migrate at least one page in the current
2336 * order-aligned block, so skip the rest of it.
2337 */
2338 if (cc->direct_compaction &&
2339 (cc->mode == MIGRATE_ASYNC)) {
2340 cc->migrate_pfn = block_end_pfn(
2341 cc->migrate_pfn - 1, cc->order);
2342 /* Draining pcplists is useless in this case */
2343 last_migrated_pfn = 0;
2344 }
2345 }
2346
2347check_drain:
2348 /*
2349 * Has the migration scanner moved away from the previous
2350 * cc->order aligned block where we migrated from? If yes,
2351 * flush the pages that were freed, so that they can merge and
2352 * compact_finished() can detect immediately if allocation
2353 * would succeed.
2354 */
2355 if (cc->order > 0 && last_migrated_pfn) {
2356 unsigned long current_block_start =
2357 block_start_pfn(cc->migrate_pfn, cc->order);
2358
2359 if (last_migrated_pfn < current_block_start) {
2360 lru_add_drain_cpu_zone(cc->zone);
2361 /* No more flushing until we migrate again */
2362 last_migrated_pfn = 0;
2363 }
2364 }
2365
2366 /* Stop if a page has been captured */
2367 if (capc && capc->page) {
2368 ret = COMPACT_SUCCESS;
2369 break;
2370 }
2371 }
2372
2373out:
2374 /*
2375 * Release free pages and update where the free scanner should restart,
2376 * so we don't leave any returned pages behind in the next attempt.
2377 */
2378 if (cc->nr_freepages > 0) {
2379 unsigned long free_pfn = release_freepages(&cc->freepages);
2380
2381 cc->nr_freepages = 0;
2382 VM_BUG_ON(free_pfn == 0);
2383 /* The cached pfn is always the first in a pageblock */
2384 free_pfn = pageblock_start_pfn(free_pfn);
2385 /*
2386 * Only go back, not forward. The cached pfn might have been
2387 * already reset to zone end in compact_finished()
2388 */
2389 if (free_pfn > cc->zone->compact_cached_free_pfn)
2390 cc->zone->compact_cached_free_pfn = free_pfn;
2391 }
2392
2393 count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned);
2394 count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned);
2395
2396 trace_mm_compaction_end(start_pfn, cc->migrate_pfn,
2397 cc->free_pfn, end_pfn, sync, ret);
2398
2399 return ret;
2400}
2401
2402static enum compact_result compact_zone_order(struct zone *zone, int order,
2403 gfp_t gfp_mask, enum compact_priority prio,
2404 unsigned int alloc_flags, int highest_zoneidx,
2405 struct page **capture)
2406{
2407 enum compact_result ret;
2408 struct compact_control cc = {
2409 .order = order,
2410 .search_order = order,
2411 .gfp_mask = gfp_mask,
2412 .zone = zone,
2413 .mode = (prio == COMPACT_PRIO_ASYNC) ?
2414 MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT,
2415 .alloc_flags = alloc_flags,
2416 .highest_zoneidx = highest_zoneidx,
2417 .direct_compaction = true,
2418 .whole_zone = (prio == MIN_COMPACT_PRIORITY),
2419 .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
2420 .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
2421 };
2422 struct capture_control capc = {
2423 .cc = &cc,
2424 .page = NULL,
2425 };
2426
2427 /*
2428 * Make sure the structs are really initialized before we expose the
2429 * capture control, in case we are interrupted and the interrupt handler
2430 * frees a page.
2431 */
2432 barrier();
2433 WRITE_ONCE(current->capture_control, &capc);
2434
2435 ret = compact_zone(&cc, &capc);
2436
2437 VM_BUG_ON(!list_empty(&cc.freepages));
2438 VM_BUG_ON(!list_empty(&cc.migratepages));
2439
2440 /*
2441 * Make sure we hide capture control first before we read the captured
2442 * page pointer, otherwise an interrupt could free and capture a page
2443 * and we would leak it.
2444 */
2445 WRITE_ONCE(current->capture_control, NULL);
2446 *capture = READ_ONCE(capc.page);
2447
2448 return ret;
2449}
2450
2451int sysctl_extfrag_threshold = 500;
2452
2453/**
2454 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
2455 * @gfp_mask: The GFP mask of the current allocation
2456 * @order: The order of the current allocation
2457 * @alloc_flags: The allocation flags of the current allocation
2458 * @ac: The context of current allocation
2459 * @prio: Determines how hard direct compaction should try to succeed
2460 * @capture: Pointer to free page created by compaction will be stored here
2461 *
2462 * This is the main entry point for direct page compaction.
2463 */
2464enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
2465 unsigned int alloc_flags, const struct alloc_context *ac,
2466 enum compact_priority prio, struct page **capture)
2467{
2468 int may_perform_io = gfp_mask & __GFP_IO;
2469 struct zoneref *z;
2470 struct zone *zone;
2471 enum compact_result rc = COMPACT_SKIPPED;
2472
2473 /*
2474 * Check if the GFP flags allow compaction - GFP_NOIO is really
2475 * tricky context because the migration might require IO
2476 */
2477 if (!may_perform_io)
2478 return COMPACT_SKIPPED;
2479
2480 trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
2481
2482 /* Compact each zone in the list */
2483 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2484 ac->highest_zoneidx, ac->nodemask) {
2485 enum compact_result status;
2486
2487 if (prio > MIN_COMPACT_PRIORITY
2488 && compaction_deferred(zone, order)) {
2489 rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
2490 continue;
2491 }
2492
2493 status = compact_zone_order(zone, order, gfp_mask, prio,
2494 alloc_flags, ac->highest_zoneidx, capture);
2495 rc = max(status, rc);
2496
2497 /* The allocation should succeed, stop compacting */
2498 if (status == COMPACT_SUCCESS) {
2499 /*
2500 * We think the allocation will succeed in this zone,
2501 * but it is not certain, hence the false. The caller
2502 * will repeat this with true if allocation indeed
2503 * succeeds in this zone.
2504 */
2505 compaction_defer_reset(zone, order, false);
2506
2507 break;
2508 }
2509
2510 if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
2511 status == COMPACT_PARTIAL_SKIPPED))
2512 /*
2513 * We think that allocation won't succeed in this zone
2514 * so we defer compaction there. If it ends up
2515 * succeeding after all, it will be reset.
2516 */
2517 defer_compaction(zone, order);
2518
2519 /*
2520 * We might have stopped compacting due to need_resched() in
2521 * async compaction, or due to a fatal signal detected. In that
2522 * case do not try further zones
2523 */
2524 if ((prio == COMPACT_PRIO_ASYNC && need_resched())
2525 || fatal_signal_pending(current))
2526 break;
2527 }
2528
2529 return rc;
2530}
2531
2532/*
2533 * Compact all zones within a node till each zone's fragmentation score
2534 * reaches within proactive compaction thresholds (as determined by the
2535 * proactiveness tunable).
2536 *
2537 * It is possible that the function returns before reaching score targets
2538 * due to various back-off conditions, such as, contention on per-node or
2539 * per-zone locks.
2540 */
2541static void proactive_compact_node(pg_data_t *pgdat)
2542{
2543 int zoneid;
2544 struct zone *zone;
2545 struct compact_control cc = {
2546 .order = -1,
2547 .mode = MIGRATE_SYNC_LIGHT,
2548 .ignore_skip_hint = true,
2549 .whole_zone = true,
2550 .gfp_mask = GFP_KERNEL,
2551 .proactive_compaction = true,
2552 };
2553
2554 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2555 zone = &pgdat->node_zones[zoneid];
2556 if (!populated_zone(zone))
2557 continue;
2558
2559 cc.zone = zone;
2560
2561 compact_zone(&cc, NULL);
2562
2563 VM_BUG_ON(!list_empty(&cc.freepages));
2564 VM_BUG_ON(!list_empty(&cc.migratepages));
2565 }
2566}
2567
2568/* Compact all zones within a node */
2569static void compact_node(int nid)
2570{
2571 pg_data_t *pgdat = NODE_DATA(nid);
2572 int zoneid;
2573 struct zone *zone;
2574 struct compact_control cc = {
2575 .order = -1,
2576 .mode = MIGRATE_SYNC,
2577 .ignore_skip_hint = true,
2578 .whole_zone = true,
2579 .gfp_mask = GFP_KERNEL,
2580 };
2581
2582
2583 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2584
2585 zone = &pgdat->node_zones[zoneid];
2586 if (!populated_zone(zone))
2587 continue;
2588
2589 cc.zone = zone;
2590
2591 compact_zone(&cc, NULL);
2592
2593 VM_BUG_ON(!list_empty(&cc.freepages));
2594 VM_BUG_ON(!list_empty(&cc.migratepages));
2595 }
2596}
2597
2598/* Compact all nodes in the system */
2599static void compact_nodes(void)
2600{
2601 int nid;
2602
2603 /* Flush pending updates to the LRU lists */
2604 lru_add_drain_all();
2605
2606 for_each_online_node(nid)
2607 compact_node(nid);
2608}
2609
2610/* The written value is actually unused, all memory is compacted */
2611int sysctl_compact_memory;
2612
2613/*
2614 * Tunable for proactive compaction. It determines how
2615 * aggressively the kernel should compact memory in the
2616 * background. It takes values in the range [0, 100].
2617 */
2618unsigned int __read_mostly sysctl_compaction_proactiveness = 20;
2619
2620/*
2621 * This is the entry point for compacting all nodes via
2622 * /proc/sys/vm/compact_memory
2623 */
2624int sysctl_compaction_handler(struct ctl_table *table, int write,
2625 void *buffer, size_t *length, loff_t *ppos)
2626{
2627 if (write)
2628 compact_nodes();
2629
2630 return 0;
2631}
2632
2633#if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
2634static ssize_t sysfs_compact_node(struct device *dev,
2635 struct device_attribute *attr,
2636 const char *buf, size_t count)
2637{
2638 int nid = dev->id;
2639
2640 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
2641 /* Flush pending updates to the LRU lists */
2642 lru_add_drain_all();
2643
2644 compact_node(nid);
2645 }
2646
2647 return count;
2648}
2649static DEVICE_ATTR(compact, 0200, NULL, sysfs_compact_node);
2650
2651int compaction_register_node(struct node *node)
2652{
2653 return device_create_file(&node->dev, &dev_attr_compact);
2654}
2655
2656void compaction_unregister_node(struct node *node)
2657{
2658 return device_remove_file(&node->dev, &dev_attr_compact);
2659}
2660#endif /* CONFIG_SYSFS && CONFIG_NUMA */
2661
2662static inline bool kcompactd_work_requested(pg_data_t *pgdat)
2663{
2664 return pgdat->kcompactd_max_order > 0 || kthread_should_stop();
2665}
2666
2667static bool kcompactd_node_suitable(pg_data_t *pgdat)
2668{
2669 int zoneid;
2670 struct zone *zone;
2671 enum zone_type highest_zoneidx = pgdat->kcompactd_highest_zoneidx;
2672
2673 for (zoneid = 0; zoneid <= highest_zoneidx; zoneid++) {
2674 zone = &pgdat->node_zones[zoneid];
2675
2676 if (!populated_zone(zone))
2677 continue;
2678
2679 if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0,
2680 highest_zoneidx) == COMPACT_CONTINUE)
2681 return true;
2682 }
2683
2684 return false;
2685}
2686
2687static void kcompactd_do_work(pg_data_t *pgdat)
2688{
2689 /*
2690 * With no special task, compact all zones so that a page of requested
2691 * order is allocatable.
2692 */
2693 int zoneid;
2694 struct zone *zone;
2695 struct compact_control cc = {
2696 .order = pgdat->kcompactd_max_order,
2697 .search_order = pgdat->kcompactd_max_order,
2698 .highest_zoneidx = pgdat->kcompactd_highest_zoneidx,
2699 .mode = MIGRATE_SYNC_LIGHT,
2700 .ignore_skip_hint = false,
2701 .gfp_mask = GFP_KERNEL,
2702 };
2703 trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
2704 cc.highest_zoneidx);
2705 count_compact_event(KCOMPACTD_WAKE);
2706
2707 for (zoneid = 0; zoneid <= cc.highest_zoneidx; zoneid++) {
2708 int status;
2709
2710 zone = &pgdat->node_zones[zoneid];
2711 if (!populated_zone(zone))
2712 continue;
2713
2714 if (compaction_deferred(zone, cc.order))
2715 continue;
2716
2717 if (compaction_suitable(zone, cc.order, 0, zoneid) !=
2718 COMPACT_CONTINUE)
2719 continue;
2720
2721 if (kthread_should_stop())
2722 return;
2723
2724 cc.zone = zone;
2725 status = compact_zone(&cc, NULL);
2726
2727 if (status == COMPACT_SUCCESS) {
2728 compaction_defer_reset(zone, cc.order, false);
2729 } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
2730 /*
2731 * Buddy pages may become stranded on pcps that could
2732 * otherwise coalesce on the zone's free area for
2733 * order >= cc.order. This is ratelimited by the
2734 * upcoming deferral.
2735 */
2736 drain_all_pages(zone);
2737
2738 /*
2739 * We use sync migration mode here, so we defer like
2740 * sync direct compaction does.
2741 */
2742 defer_compaction(zone, cc.order);
2743 }
2744
2745 count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
2746 cc.total_migrate_scanned);
2747 count_compact_events(KCOMPACTD_FREE_SCANNED,
2748 cc.total_free_scanned);
2749
2750 VM_BUG_ON(!list_empty(&cc.freepages));
2751 VM_BUG_ON(!list_empty(&cc.migratepages));
2752 }
2753
2754 /*
2755 * Regardless of success, we are done until woken up next. But remember
2756 * the requested order/highest_zoneidx in case it was higher/tighter
2757 * than our current ones
2758 */
2759 if (pgdat->kcompactd_max_order <= cc.order)
2760 pgdat->kcompactd_max_order = 0;
2761 if (pgdat->kcompactd_highest_zoneidx >= cc.highest_zoneidx)
2762 pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
2763}
2764
2765void wakeup_kcompactd(pg_data_t *pgdat, int order, int highest_zoneidx)
2766{
2767 if (!order)
2768 return;
2769
2770 if (pgdat->kcompactd_max_order < order)
2771 pgdat->kcompactd_max_order = order;
2772
2773 if (pgdat->kcompactd_highest_zoneidx > highest_zoneidx)
2774 pgdat->kcompactd_highest_zoneidx = highest_zoneidx;
2775
2776 /*
2777 * Pairs with implicit barrier in wait_event_freezable()
2778 * such that wakeups are not missed.
2779 */
2780 if (!wq_has_sleeper(&pgdat->kcompactd_wait))
2781 return;
2782
2783 if (!kcompactd_node_suitable(pgdat))
2784 return;
2785
2786 trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
2787 highest_zoneidx);
2788 wake_up_interruptible(&pgdat->kcompactd_wait);
2789}
2790
2791/*
2792 * The background compaction daemon, started as a kernel thread
2793 * from the init process.
2794 */
2795static int kcompactd(void *p)
2796{
2797 pg_data_t *pgdat = (pg_data_t*)p;
2798 struct task_struct *tsk = current;
2799 unsigned int proactive_defer = 0;
2800
2801 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2802
2803 if (!cpumask_empty(cpumask))
2804 set_cpus_allowed_ptr(tsk, cpumask);
2805
2806 set_freezable();
2807
2808 pgdat->kcompactd_max_order = 0;
2809 pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
2810
2811 while (!kthread_should_stop()) {
2812 unsigned long pflags;
2813
2814 trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
2815 if (wait_event_freezable_timeout(pgdat->kcompactd_wait,
2816 kcompactd_work_requested(pgdat),
2817 msecs_to_jiffies(HPAGE_FRAG_CHECK_INTERVAL_MSEC))) {
2818
2819 psi_memstall_enter(&pflags);
2820 kcompactd_do_work(pgdat);
2821 psi_memstall_leave(&pflags);
2822 continue;
2823 }
2824
2825 /* kcompactd wait timeout */
2826 if (should_proactive_compact_node(pgdat)) {
2827 unsigned int prev_score, score;
2828
2829 if (proactive_defer) {
2830 proactive_defer--;
2831 continue;
2832 }
2833 prev_score = fragmentation_score_node(pgdat);
2834 proactive_compact_node(pgdat);
2835 score = fragmentation_score_node(pgdat);
2836 /*
2837 * Defer proactive compaction if the fragmentation
2838 * score did not go down i.e. no progress made.
2839 */
2840 proactive_defer = score < prev_score ?
2841 0 : 1 << COMPACT_MAX_DEFER_SHIFT;
2842 }
2843 }
2844
2845 return 0;
2846}
2847
2848/*
2849 * This kcompactd start function will be called by init and node-hot-add.
2850 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
2851 */
2852int kcompactd_run(int nid)
2853{
2854 pg_data_t *pgdat = NODE_DATA(nid);
2855 int ret = 0;
2856
2857 if (pgdat->kcompactd)
2858 return 0;
2859
2860 pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
2861 if (IS_ERR(pgdat->kcompactd)) {
2862 pr_err("Failed to start kcompactd on node %d\n", nid);
2863 ret = PTR_ERR(pgdat->kcompactd);
2864 pgdat->kcompactd = NULL;
2865 }
2866 return ret;
2867}
2868
2869/*
2870 * Called by memory hotplug when all memory in a node is offlined. Caller must
2871 * hold mem_hotplug_begin/end().
2872 */
2873void kcompactd_stop(int nid)
2874{
2875 struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
2876
2877 if (kcompactd) {
2878 kthread_stop(kcompactd);
2879 NODE_DATA(nid)->kcompactd = NULL;
2880 }
2881}
2882
2883/*
2884 * It's optimal to keep kcompactd on the same CPUs as their memory, but
2885 * not required for correctness. So if the last cpu in a node goes
2886 * away, we get changed to run anywhere: as the first one comes back,
2887 * restore their cpu bindings.
2888 */
2889static int kcompactd_cpu_online(unsigned int cpu)
2890{
2891 int nid;
2892
2893 for_each_node_state(nid, N_MEMORY) {
2894 pg_data_t *pgdat = NODE_DATA(nid);
2895 const struct cpumask *mask;
2896
2897 mask = cpumask_of_node(pgdat->node_id);
2898
2899 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2900 /* One of our CPUs online: restore mask */
2901 set_cpus_allowed_ptr(pgdat->kcompactd, mask);
2902 }
2903 return 0;
2904}
2905
2906static int __init kcompactd_init(void)
2907{
2908 int nid;
2909 int ret;
2910
2911 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
2912 "mm/compaction:online",
2913 kcompactd_cpu_online, NULL);
2914 if (ret < 0) {
2915 pr_err("kcompactd: failed to register hotplug callbacks.\n");
2916 return ret;
2917 }
2918
2919 for_each_node_state(nid, N_MEMORY)
2920 kcompactd_run(nid);
2921 return 0;
2922}
2923subsys_initcall(kcompactd_init)
2924
2925#endif /* CONFIG_COMPACTION */
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
29static inline void count_compact_event(enum vm_event_item item)
30{
31 count_vm_event(item);
32}
33
34static inline void count_compact_events(enum vm_event_item item, long delta)
35{
36 count_vm_events(item, delta);
37}
38#else
39#define count_compact_event(item) do { } while (0)
40#define count_compact_events(item, delta) do { } while (0)
41#endif
42
43#if defined CONFIG_COMPACTION || defined CONFIG_CMA
44
45#define CREATE_TRACE_POINTS
46#include <trace/events/compaction.h>
47
48#define block_start_pfn(pfn, order) round_down(pfn, 1UL << (order))
49#define block_end_pfn(pfn, order) ALIGN((pfn) + 1, 1UL << (order))
50#define pageblock_start_pfn(pfn) block_start_pfn(pfn, pageblock_order)
51#define pageblock_end_pfn(pfn) block_end_pfn(pfn, pageblock_order)
52
53static unsigned long release_freepages(struct list_head *freelist)
54{
55 struct page *page, *next;
56 unsigned long high_pfn = 0;
57
58 list_for_each_entry_safe(page, next, freelist, lru) {
59 unsigned long pfn = page_to_pfn(page);
60 list_del(&page->lru);
61 __free_page(page);
62 if (pfn > high_pfn)
63 high_pfn = pfn;
64 }
65
66 return high_pfn;
67}
68
69static void split_map_pages(struct list_head *list)
70{
71 unsigned int i, order, nr_pages;
72 struct page *page, *next;
73 LIST_HEAD(tmp_list);
74
75 list_for_each_entry_safe(page, next, list, lru) {
76 list_del(&page->lru);
77
78 order = page_private(page);
79 nr_pages = 1 << order;
80
81 post_alloc_hook(page, order, __GFP_MOVABLE);
82 if (order)
83 split_page(page, order);
84
85 for (i = 0; i < nr_pages; i++) {
86 list_add(&page->lru, &tmp_list);
87 page++;
88 }
89 }
90
91 list_splice(&tmp_list, list);
92}
93
94#ifdef CONFIG_COMPACTION
95
96int PageMovable(struct page *page)
97{
98 struct address_space *mapping;
99
100 VM_BUG_ON_PAGE(!PageLocked(page), page);
101 if (!__PageMovable(page))
102 return 0;
103
104 mapping = page_mapping(page);
105 if (mapping && mapping->a_ops && mapping->a_ops->isolate_page)
106 return 1;
107
108 return 0;
109}
110EXPORT_SYMBOL(PageMovable);
111
112void __SetPageMovable(struct page *page, struct address_space *mapping)
113{
114 VM_BUG_ON_PAGE(!PageLocked(page), page);
115 VM_BUG_ON_PAGE((unsigned long)mapping & PAGE_MAPPING_MOVABLE, page);
116 page->mapping = (void *)((unsigned long)mapping | PAGE_MAPPING_MOVABLE);
117}
118EXPORT_SYMBOL(__SetPageMovable);
119
120void __ClearPageMovable(struct page *page)
121{
122 VM_BUG_ON_PAGE(!PageLocked(page), page);
123 VM_BUG_ON_PAGE(!PageMovable(page), page);
124 /*
125 * Clear registered address_space val with keeping PAGE_MAPPING_MOVABLE
126 * flag so that VM can catch up released page by driver after isolation.
127 * With it, VM migration doesn't try to put it back.
128 */
129 page->mapping = (void *)((unsigned long)page->mapping &
130 PAGE_MAPPING_MOVABLE);
131}
132EXPORT_SYMBOL(__ClearPageMovable);
133
134/* Do not skip compaction more than 64 times */
135#define COMPACT_MAX_DEFER_SHIFT 6
136
137/*
138 * Compaction is deferred when compaction fails to result in a page
139 * allocation success. 1 << compact_defer_limit compactions are skipped up
140 * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
141 */
142void defer_compaction(struct zone *zone, int order)
143{
144 zone->compact_considered = 0;
145 zone->compact_defer_shift++;
146
147 if (order < zone->compact_order_failed)
148 zone->compact_order_failed = order;
149
150 if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
151 zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
152
153 trace_mm_compaction_defer_compaction(zone, order);
154}
155
156/* Returns true if compaction should be skipped this time */
157bool compaction_deferred(struct zone *zone, int order)
158{
159 unsigned long defer_limit = 1UL << zone->compact_defer_shift;
160
161 if (order < zone->compact_order_failed)
162 return false;
163
164 /* Avoid possible overflow */
165 if (++zone->compact_considered > defer_limit)
166 zone->compact_considered = defer_limit;
167
168 if (zone->compact_considered >= defer_limit)
169 return false;
170
171 trace_mm_compaction_deferred(zone, order);
172
173 return true;
174}
175
176/*
177 * Update defer tracking counters after successful compaction of given order,
178 * which means an allocation either succeeded (alloc_success == true) or is
179 * expected to succeed.
180 */
181void compaction_defer_reset(struct zone *zone, int order,
182 bool alloc_success)
183{
184 if (alloc_success) {
185 zone->compact_considered = 0;
186 zone->compact_defer_shift = 0;
187 }
188 if (order >= zone->compact_order_failed)
189 zone->compact_order_failed = order + 1;
190
191 trace_mm_compaction_defer_reset(zone, order);
192}
193
194/* Returns true if restarting compaction after many failures */
195bool compaction_restarting(struct zone *zone, int order)
196{
197 if (order < zone->compact_order_failed)
198 return false;
199
200 return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
201 zone->compact_considered >= 1UL << zone->compact_defer_shift;
202}
203
204/* Returns true if the pageblock should be scanned for pages to isolate. */
205static inline bool isolation_suitable(struct compact_control *cc,
206 struct page *page)
207{
208 if (cc->ignore_skip_hint)
209 return true;
210
211 return !get_pageblock_skip(page);
212}
213
214static void reset_cached_positions(struct zone *zone)
215{
216 zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
217 zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
218 zone->compact_cached_free_pfn =
219 pageblock_start_pfn(zone_end_pfn(zone) - 1);
220}
221
222/*
223 * Compound pages of >= pageblock_order should consistenly be skipped until
224 * released. It is always pointless to compact pages of such order (if they are
225 * migratable), and the pageblocks they occupy cannot contain any free pages.
226 */
227static bool pageblock_skip_persistent(struct page *page)
228{
229 if (!PageCompound(page))
230 return false;
231
232 page = compound_head(page);
233
234 if (compound_order(page) >= pageblock_order)
235 return true;
236
237 return false;
238}
239
240static bool
241__reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source,
242 bool check_target)
243{
244 struct page *page = pfn_to_online_page(pfn);
245 struct page *block_page;
246 struct page *end_page;
247 unsigned long block_pfn;
248
249 if (!page)
250 return false;
251 if (zone != page_zone(page))
252 return false;
253 if (pageblock_skip_persistent(page))
254 return false;
255
256 /*
257 * If skip is already cleared do no further checking once the
258 * restart points have been set.
259 */
260 if (check_source && check_target && !get_pageblock_skip(page))
261 return true;
262
263 /*
264 * If clearing skip for the target scanner, do not select a
265 * non-movable pageblock as the starting point.
266 */
267 if (!check_source && check_target &&
268 get_pageblock_migratetype(page) != MIGRATE_MOVABLE)
269 return false;
270
271 /* Ensure the start of the pageblock or zone is online and valid */
272 block_pfn = pageblock_start_pfn(pfn);
273 block_pfn = max(block_pfn, zone->zone_start_pfn);
274 block_page = pfn_to_online_page(block_pfn);
275 if (block_page) {
276 page = block_page;
277 pfn = block_pfn;
278 }
279
280 /* Ensure the end of the pageblock or zone is online and valid */
281 block_pfn = pageblock_end_pfn(pfn) - 1;
282 block_pfn = min(block_pfn, zone_end_pfn(zone) - 1);
283 end_page = pfn_to_online_page(block_pfn);
284 if (!end_page)
285 return false;
286
287 /*
288 * Only clear the hint if a sample indicates there is either a
289 * free page or an LRU page in the block. One or other condition
290 * is necessary for the block to be a migration source/target.
291 */
292 do {
293 if (pfn_valid_within(pfn)) {
294 if (check_source && PageLRU(page)) {
295 clear_pageblock_skip(page);
296 return true;
297 }
298
299 if (check_target && PageBuddy(page)) {
300 clear_pageblock_skip(page);
301 return true;
302 }
303 }
304
305 page += (1 << PAGE_ALLOC_COSTLY_ORDER);
306 pfn += (1 << PAGE_ALLOC_COSTLY_ORDER);
307 } while (page <= end_page);
308
309 return false;
310}
311
312/*
313 * This function is called to clear all cached information on pageblocks that
314 * should be skipped for page isolation when the migrate and free page scanner
315 * meet.
316 */
317static void __reset_isolation_suitable(struct zone *zone)
318{
319 unsigned long migrate_pfn = zone->zone_start_pfn;
320 unsigned long free_pfn = zone_end_pfn(zone) - 1;
321 unsigned long reset_migrate = free_pfn;
322 unsigned long reset_free = migrate_pfn;
323 bool source_set = false;
324 bool free_set = false;
325
326 if (!zone->compact_blockskip_flush)
327 return;
328
329 zone->compact_blockskip_flush = false;
330
331 /*
332 * Walk the zone and update pageblock skip information. Source looks
333 * for PageLRU while target looks for PageBuddy. When the scanner
334 * is found, both PageBuddy and PageLRU are checked as the pageblock
335 * is suitable as both source and target.
336 */
337 for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages,
338 free_pfn -= pageblock_nr_pages) {
339 cond_resched();
340
341 /* Update the migrate PFN */
342 if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) &&
343 migrate_pfn < reset_migrate) {
344 source_set = true;
345 reset_migrate = migrate_pfn;
346 zone->compact_init_migrate_pfn = reset_migrate;
347 zone->compact_cached_migrate_pfn[0] = reset_migrate;
348 zone->compact_cached_migrate_pfn[1] = reset_migrate;
349 }
350
351 /* Update the free PFN */
352 if (__reset_isolation_pfn(zone, free_pfn, free_set, true) &&
353 free_pfn > reset_free) {
354 free_set = true;
355 reset_free = free_pfn;
356 zone->compact_init_free_pfn = reset_free;
357 zone->compact_cached_free_pfn = reset_free;
358 }
359 }
360
361 /* Leave no distance if no suitable block was reset */
362 if (reset_migrate >= reset_free) {
363 zone->compact_cached_migrate_pfn[0] = migrate_pfn;
364 zone->compact_cached_migrate_pfn[1] = migrate_pfn;
365 zone->compact_cached_free_pfn = free_pfn;
366 }
367}
368
369void reset_isolation_suitable(pg_data_t *pgdat)
370{
371 int zoneid;
372
373 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
374 struct zone *zone = &pgdat->node_zones[zoneid];
375 if (!populated_zone(zone))
376 continue;
377
378 /* Only flush if a full compaction finished recently */
379 if (zone->compact_blockskip_flush)
380 __reset_isolation_suitable(zone);
381 }
382}
383
384/*
385 * Sets the pageblock skip bit if it was clear. Note that this is a hint as
386 * locks are not required for read/writers. Returns true if it was already set.
387 */
388static bool test_and_set_skip(struct compact_control *cc, struct page *page,
389 unsigned long pfn)
390{
391 bool skip;
392
393 /* Do no update if skip hint is being ignored */
394 if (cc->ignore_skip_hint)
395 return false;
396
397 if (!IS_ALIGNED(pfn, pageblock_nr_pages))
398 return false;
399
400 skip = get_pageblock_skip(page);
401 if (!skip && !cc->no_set_skip_hint)
402 set_pageblock_skip(page);
403
404 return skip;
405}
406
407static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
408{
409 struct zone *zone = cc->zone;
410
411 pfn = pageblock_end_pfn(pfn);
412
413 /* Set for isolation rather than compaction */
414 if (cc->no_set_skip_hint)
415 return;
416
417 if (pfn > zone->compact_cached_migrate_pfn[0])
418 zone->compact_cached_migrate_pfn[0] = pfn;
419 if (cc->mode != MIGRATE_ASYNC &&
420 pfn > zone->compact_cached_migrate_pfn[1])
421 zone->compact_cached_migrate_pfn[1] = pfn;
422}
423
424/*
425 * If no pages were isolated then mark this pageblock to be skipped in the
426 * future. The information is later cleared by __reset_isolation_suitable().
427 */
428static void update_pageblock_skip(struct compact_control *cc,
429 struct page *page, unsigned long pfn)
430{
431 struct zone *zone = cc->zone;
432
433 if (cc->no_set_skip_hint)
434 return;
435
436 if (!page)
437 return;
438
439 set_pageblock_skip(page);
440
441 /* Update where async and sync compaction should restart */
442 if (pfn < zone->compact_cached_free_pfn)
443 zone->compact_cached_free_pfn = pfn;
444}
445#else
446static inline bool isolation_suitable(struct compact_control *cc,
447 struct page *page)
448{
449 return true;
450}
451
452static inline bool pageblock_skip_persistent(struct page *page)
453{
454 return false;
455}
456
457static inline void update_pageblock_skip(struct compact_control *cc,
458 struct page *page, unsigned long pfn)
459{
460}
461
462static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
463{
464}
465
466static bool test_and_set_skip(struct compact_control *cc, struct page *page,
467 unsigned long pfn)
468{
469 return false;
470}
471#endif /* CONFIG_COMPACTION */
472
473/*
474 * Compaction requires the taking of some coarse locks that are potentially
475 * very heavily contended. For async compaction, trylock and record if the
476 * lock is contended. The lock will still be acquired but compaction will
477 * abort when the current block is finished regardless of success rate.
478 * Sync compaction acquires the lock.
479 *
480 * Always returns true which makes it easier to track lock state in callers.
481 */
482static bool compact_lock_irqsave(spinlock_t *lock, unsigned long *flags,
483 struct compact_control *cc)
484{
485 /* Track if the lock is contended in async mode */
486 if (cc->mode == MIGRATE_ASYNC && !cc->contended) {
487 if (spin_trylock_irqsave(lock, *flags))
488 return true;
489
490 cc->contended = true;
491 }
492
493 spin_lock_irqsave(lock, *flags);
494 return true;
495}
496
497/*
498 * Compaction requires the taking of some coarse locks that are potentially
499 * very heavily contended. The lock should be periodically unlocked to avoid
500 * having disabled IRQs for a long time, even when there is nobody waiting on
501 * the lock. It might also be that allowing the IRQs will result in
502 * need_resched() becoming true. If scheduling is needed, async compaction
503 * aborts. Sync compaction schedules.
504 * Either compaction type will also abort if a fatal signal is pending.
505 * In either case if the lock was locked, it is dropped and not regained.
506 *
507 * Returns true if compaction should abort due to fatal signal pending, or
508 * async compaction due to need_resched()
509 * Returns false when compaction can continue (sync compaction might have
510 * scheduled)
511 */
512static bool compact_unlock_should_abort(spinlock_t *lock,
513 unsigned long flags, bool *locked, struct compact_control *cc)
514{
515 if (*locked) {
516 spin_unlock_irqrestore(lock, flags);
517 *locked = false;
518 }
519
520 if (fatal_signal_pending(current)) {
521 cc->contended = true;
522 return true;
523 }
524
525 cond_resched();
526
527 return false;
528}
529
530/*
531 * Isolate free pages onto a private freelist. If @strict is true, will abort
532 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
533 * (even though it may still end up isolating some pages).
534 */
535static unsigned long isolate_freepages_block(struct compact_control *cc,
536 unsigned long *start_pfn,
537 unsigned long end_pfn,
538 struct list_head *freelist,
539 unsigned int stride,
540 bool strict)
541{
542 int nr_scanned = 0, total_isolated = 0;
543 struct page *cursor;
544 unsigned long flags = 0;
545 bool locked = false;
546 unsigned long blockpfn = *start_pfn;
547 unsigned int order;
548
549 /* Strict mode is for isolation, speed is secondary */
550 if (strict)
551 stride = 1;
552
553 cursor = pfn_to_page(blockpfn);
554
555 /* Isolate free pages. */
556 for (; blockpfn < end_pfn; blockpfn += stride, cursor += stride) {
557 int isolated;
558 struct page *page = cursor;
559
560 /*
561 * Periodically drop the lock (if held) regardless of its
562 * contention, to give chance to IRQs. Abort if fatal signal
563 * pending or async compaction detects need_resched()
564 */
565 if (!(blockpfn % SWAP_CLUSTER_MAX)
566 && compact_unlock_should_abort(&cc->zone->lock, flags,
567 &locked, cc))
568 break;
569
570 nr_scanned++;
571 if (!pfn_valid_within(blockpfn))
572 goto isolate_fail;
573
574 /*
575 * For compound pages such as THP and hugetlbfs, we can save
576 * potentially a lot of iterations if we skip them at once.
577 * The check is racy, but we can consider only valid values
578 * and the only danger is skipping too much.
579 */
580 if (PageCompound(page)) {
581 const unsigned int order = compound_order(page);
582
583 if (likely(order < MAX_ORDER)) {
584 blockpfn += (1UL << order) - 1;
585 cursor += (1UL << order) - 1;
586 }
587 goto isolate_fail;
588 }
589
590 if (!PageBuddy(page))
591 goto isolate_fail;
592
593 /*
594 * If we already hold the lock, we can skip some rechecking.
595 * Note that if we hold the lock now, checked_pageblock was
596 * already set in some previous iteration (or strict is true),
597 * so it is correct to skip the suitable migration target
598 * recheck as well.
599 */
600 if (!locked) {
601 locked = compact_lock_irqsave(&cc->zone->lock,
602 &flags, cc);
603
604 /* Recheck this is a buddy page under lock */
605 if (!PageBuddy(page))
606 goto isolate_fail;
607 }
608
609 /* Found a free page, will break it into order-0 pages */
610 order = page_order(page);
611 isolated = __isolate_free_page(page, order);
612 if (!isolated)
613 break;
614 set_page_private(page, order);
615
616 total_isolated += isolated;
617 cc->nr_freepages += isolated;
618 list_add_tail(&page->lru, freelist);
619
620 if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
621 blockpfn += isolated;
622 break;
623 }
624 /* Advance to the end of split page */
625 blockpfn += isolated - 1;
626 cursor += isolated - 1;
627 continue;
628
629isolate_fail:
630 if (strict)
631 break;
632 else
633 continue;
634
635 }
636
637 if (locked)
638 spin_unlock_irqrestore(&cc->zone->lock, flags);
639
640 /*
641 * There is a tiny chance that we have read bogus compound_order(),
642 * so be careful to not go outside of the pageblock.
643 */
644 if (unlikely(blockpfn > end_pfn))
645 blockpfn = end_pfn;
646
647 trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
648 nr_scanned, total_isolated);
649
650 /* Record how far we have got within the block */
651 *start_pfn = blockpfn;
652
653 /*
654 * If strict isolation is requested by CMA then check that all the
655 * pages requested were isolated. If there were any failures, 0 is
656 * returned and CMA will fail.
657 */
658 if (strict && blockpfn < end_pfn)
659 total_isolated = 0;
660
661 cc->total_free_scanned += nr_scanned;
662 if (total_isolated)
663 count_compact_events(COMPACTISOLATED, total_isolated);
664 return total_isolated;
665}
666
667/**
668 * isolate_freepages_range() - isolate free pages.
669 * @cc: Compaction control structure.
670 * @start_pfn: The first PFN to start isolating.
671 * @end_pfn: The one-past-last PFN.
672 *
673 * Non-free pages, invalid PFNs, or zone boundaries within the
674 * [start_pfn, end_pfn) range are considered errors, cause function to
675 * undo its actions and return zero.
676 *
677 * Otherwise, function returns one-past-the-last PFN of isolated page
678 * (which may be greater then end_pfn if end fell in a middle of
679 * a free page).
680 */
681unsigned long
682isolate_freepages_range(struct compact_control *cc,
683 unsigned long start_pfn, unsigned long end_pfn)
684{
685 unsigned long isolated, pfn, block_start_pfn, block_end_pfn;
686 LIST_HEAD(freelist);
687
688 pfn = start_pfn;
689 block_start_pfn = pageblock_start_pfn(pfn);
690 if (block_start_pfn < cc->zone->zone_start_pfn)
691 block_start_pfn = cc->zone->zone_start_pfn;
692 block_end_pfn = pageblock_end_pfn(pfn);
693
694 for (; pfn < end_pfn; pfn += isolated,
695 block_start_pfn = block_end_pfn,
696 block_end_pfn += pageblock_nr_pages) {
697 /* Protect pfn from changing by isolate_freepages_block */
698 unsigned long isolate_start_pfn = pfn;
699
700 block_end_pfn = min(block_end_pfn, end_pfn);
701
702 /*
703 * pfn could pass the block_end_pfn if isolated freepage
704 * is more than pageblock order. In this case, we adjust
705 * scanning range to right one.
706 */
707 if (pfn >= block_end_pfn) {
708 block_start_pfn = pageblock_start_pfn(pfn);
709 block_end_pfn = pageblock_end_pfn(pfn);
710 block_end_pfn = min(block_end_pfn, end_pfn);
711 }
712
713 if (!pageblock_pfn_to_page(block_start_pfn,
714 block_end_pfn, cc->zone))
715 break;
716
717 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
718 block_end_pfn, &freelist, 0, true);
719
720 /*
721 * In strict mode, isolate_freepages_block() returns 0 if
722 * there are any holes in the block (ie. invalid PFNs or
723 * non-free pages).
724 */
725 if (!isolated)
726 break;
727
728 /*
729 * If we managed to isolate pages, it is always (1 << n) *
730 * pageblock_nr_pages for some non-negative n. (Max order
731 * page may span two pageblocks).
732 */
733 }
734
735 /* __isolate_free_page() does not map the pages */
736 split_map_pages(&freelist);
737
738 if (pfn < end_pfn) {
739 /* Loop terminated early, cleanup. */
740 release_freepages(&freelist);
741 return 0;
742 }
743
744 /* We don't use freelists for anything. */
745 return pfn;
746}
747
748/* Similar to reclaim, but different enough that they don't share logic */
749static bool too_many_isolated(pg_data_t *pgdat)
750{
751 unsigned long active, inactive, isolated;
752
753 inactive = node_page_state(pgdat, NR_INACTIVE_FILE) +
754 node_page_state(pgdat, NR_INACTIVE_ANON);
755 active = node_page_state(pgdat, NR_ACTIVE_FILE) +
756 node_page_state(pgdat, NR_ACTIVE_ANON);
757 isolated = node_page_state(pgdat, NR_ISOLATED_FILE) +
758 node_page_state(pgdat, NR_ISOLATED_ANON);
759
760 return isolated > (inactive + active) / 2;
761}
762
763/**
764 * isolate_migratepages_block() - isolate all migrate-able pages within
765 * a single pageblock
766 * @cc: Compaction control structure.
767 * @low_pfn: The first PFN to isolate
768 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
769 * @isolate_mode: Isolation mode to be used.
770 *
771 * Isolate all pages that can be migrated from the range specified by
772 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
773 * Returns zero if there is a fatal signal pending, otherwise PFN of the
774 * first page that was not scanned (which may be both less, equal to or more
775 * than end_pfn).
776 *
777 * The pages are isolated on cc->migratepages list (not required to be empty),
778 * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
779 * is neither read nor updated.
780 */
781static unsigned long
782isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
783 unsigned long end_pfn, isolate_mode_t isolate_mode)
784{
785 pg_data_t *pgdat = cc->zone->zone_pgdat;
786 unsigned long nr_scanned = 0, nr_isolated = 0;
787 struct lruvec *lruvec;
788 unsigned long flags = 0;
789 bool locked = false;
790 struct page *page = NULL, *valid_page = NULL;
791 unsigned long start_pfn = low_pfn;
792 bool skip_on_failure = false;
793 unsigned long next_skip_pfn = 0;
794 bool skip_updated = false;
795
796 /*
797 * Ensure that there are not too many pages isolated from the LRU
798 * list by either parallel reclaimers or compaction. If there are,
799 * delay for some time until fewer pages are isolated
800 */
801 while (unlikely(too_many_isolated(pgdat))) {
802 /* async migration should just abort */
803 if (cc->mode == MIGRATE_ASYNC)
804 return 0;
805
806 congestion_wait(BLK_RW_ASYNC, HZ/10);
807
808 if (fatal_signal_pending(current))
809 return 0;
810 }
811
812 cond_resched();
813
814 if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
815 skip_on_failure = true;
816 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
817 }
818
819 /* Time to isolate some pages for migration */
820 for (; low_pfn < end_pfn; low_pfn++) {
821
822 if (skip_on_failure && low_pfn >= next_skip_pfn) {
823 /*
824 * We have isolated all migration candidates in the
825 * previous order-aligned block, and did not skip it due
826 * to failure. We should migrate the pages now and
827 * hopefully succeed compaction.
828 */
829 if (nr_isolated)
830 break;
831
832 /*
833 * We failed to isolate in the previous order-aligned
834 * block. Set the new boundary to the end of the
835 * current block. Note we can't simply increase
836 * next_skip_pfn by 1 << order, as low_pfn might have
837 * been incremented by a higher number due to skipping
838 * a compound or a high-order buddy page in the
839 * previous loop iteration.
840 */
841 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
842 }
843
844 /*
845 * Periodically drop the lock (if held) regardless of its
846 * contention, to give chance to IRQs. Abort completely if
847 * a fatal signal is pending.
848 */
849 if (!(low_pfn % SWAP_CLUSTER_MAX)
850 && compact_unlock_should_abort(&pgdat->lru_lock,
851 flags, &locked, cc)) {
852 low_pfn = 0;
853 goto fatal_pending;
854 }
855
856 if (!pfn_valid_within(low_pfn))
857 goto isolate_fail;
858 nr_scanned++;
859
860 page = pfn_to_page(low_pfn);
861
862 /*
863 * Check if the pageblock has already been marked skipped.
864 * Only the aligned PFN is checked as the caller isolates
865 * COMPACT_CLUSTER_MAX at a time so the second call must
866 * not falsely conclude that the block should be skipped.
867 */
868 if (!valid_page && IS_ALIGNED(low_pfn, pageblock_nr_pages)) {
869 if (!cc->ignore_skip_hint && get_pageblock_skip(page)) {
870 low_pfn = end_pfn;
871 goto isolate_abort;
872 }
873 valid_page = page;
874 }
875
876 /*
877 * Skip if free. We read page order here without zone lock
878 * which is generally unsafe, but the race window is small and
879 * the worst thing that can happen is that we skip some
880 * potential isolation targets.
881 */
882 if (PageBuddy(page)) {
883 unsigned long freepage_order = page_order_unsafe(page);
884
885 /*
886 * Without lock, we cannot be sure that what we got is
887 * a valid page order. Consider only values in the
888 * valid order range to prevent low_pfn overflow.
889 */
890 if (freepage_order > 0 && freepage_order < MAX_ORDER)
891 low_pfn += (1UL << freepage_order) - 1;
892 continue;
893 }
894
895 /*
896 * Regardless of being on LRU, compound pages such as THP and
897 * hugetlbfs are not to be compacted. We can potentially save
898 * a lot of iterations if we skip them at once. The check is
899 * racy, but we can consider only valid values and the only
900 * danger is skipping too much.
901 */
902 if (PageCompound(page)) {
903 const unsigned int order = compound_order(page);
904
905 if (likely(order < MAX_ORDER))
906 low_pfn += (1UL << order) - 1;
907 goto isolate_fail;
908 }
909
910 /*
911 * Check may be lockless but that's ok as we recheck later.
912 * It's possible to migrate LRU and non-lru movable pages.
913 * Skip any other type of page
914 */
915 if (!PageLRU(page)) {
916 /*
917 * __PageMovable can return false positive so we need
918 * to verify it under page_lock.
919 */
920 if (unlikely(__PageMovable(page)) &&
921 !PageIsolated(page)) {
922 if (locked) {
923 spin_unlock_irqrestore(&pgdat->lru_lock,
924 flags);
925 locked = false;
926 }
927
928 if (!isolate_movable_page(page, isolate_mode))
929 goto isolate_success;
930 }
931
932 goto isolate_fail;
933 }
934
935 /*
936 * Migration will fail if an anonymous page is pinned in memory,
937 * so avoid taking lru_lock and isolating it unnecessarily in an
938 * admittedly racy check.
939 */
940 if (!page_mapping(page) &&
941 page_count(page) > page_mapcount(page))
942 goto isolate_fail;
943
944 /*
945 * Only allow to migrate anonymous pages in GFP_NOFS context
946 * because those do not depend on fs locks.
947 */
948 if (!(cc->gfp_mask & __GFP_FS) && page_mapping(page))
949 goto isolate_fail;
950
951 /* If we already hold the lock, we can skip some rechecking */
952 if (!locked) {
953 locked = compact_lock_irqsave(&pgdat->lru_lock,
954 &flags, cc);
955
956 /* Try get exclusive access under lock */
957 if (!skip_updated) {
958 skip_updated = true;
959 if (test_and_set_skip(cc, page, low_pfn))
960 goto isolate_abort;
961 }
962
963 /* Recheck PageLRU and PageCompound under lock */
964 if (!PageLRU(page))
965 goto isolate_fail;
966
967 /*
968 * Page become compound since the non-locked check,
969 * and it's on LRU. It can only be a THP so the order
970 * is safe to read and it's 0 for tail pages.
971 */
972 if (unlikely(PageCompound(page))) {
973 low_pfn += compound_nr(page) - 1;
974 goto isolate_fail;
975 }
976 }
977
978 lruvec = mem_cgroup_page_lruvec(page, pgdat);
979
980 /* Try isolate the page */
981 if (__isolate_lru_page(page, isolate_mode) != 0)
982 goto isolate_fail;
983
984 VM_BUG_ON_PAGE(PageCompound(page), page);
985
986 /* Successfully isolated */
987 del_page_from_lru_list(page, lruvec, page_lru(page));
988 inc_node_page_state(page,
989 NR_ISOLATED_ANON + page_is_file_cache(page));
990
991isolate_success:
992 list_add(&page->lru, &cc->migratepages);
993 cc->nr_migratepages++;
994 nr_isolated++;
995
996 /*
997 * Avoid isolating too much unless this block is being
998 * rescanned (e.g. dirty/writeback pages, parallel allocation)
999 * or a lock is contended. For contention, isolate quickly to
1000 * potentially remove one source of contention.
1001 */
1002 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX &&
1003 !cc->rescan && !cc->contended) {
1004 ++low_pfn;
1005 break;
1006 }
1007
1008 continue;
1009isolate_fail:
1010 if (!skip_on_failure)
1011 continue;
1012
1013 /*
1014 * We have isolated some pages, but then failed. Release them
1015 * instead of migrating, as we cannot form the cc->order buddy
1016 * page anyway.
1017 */
1018 if (nr_isolated) {
1019 if (locked) {
1020 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
1021 locked = false;
1022 }
1023 putback_movable_pages(&cc->migratepages);
1024 cc->nr_migratepages = 0;
1025 nr_isolated = 0;
1026 }
1027
1028 if (low_pfn < next_skip_pfn) {
1029 low_pfn = next_skip_pfn - 1;
1030 /*
1031 * The check near the loop beginning would have updated
1032 * next_skip_pfn too, but this is a bit simpler.
1033 */
1034 next_skip_pfn += 1UL << cc->order;
1035 }
1036 }
1037
1038 /*
1039 * The PageBuddy() check could have potentially brought us outside
1040 * the range to be scanned.
1041 */
1042 if (unlikely(low_pfn > end_pfn))
1043 low_pfn = end_pfn;
1044
1045isolate_abort:
1046 if (locked)
1047 spin_unlock_irqrestore(&pgdat->lru_lock, flags);
1048
1049 /*
1050 * Updated the cached scanner pfn once the pageblock has been scanned
1051 * Pages will either be migrated in which case there is no point
1052 * scanning in the near future or migration failed in which case the
1053 * failure reason may persist. The block is marked for skipping if
1054 * there were no pages isolated in the block or if the block is
1055 * rescanned twice in a row.
1056 */
1057 if (low_pfn == end_pfn && (!nr_isolated || cc->rescan)) {
1058 if (valid_page && !skip_updated)
1059 set_pageblock_skip(valid_page);
1060 update_cached_migrate(cc, low_pfn);
1061 }
1062
1063 trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
1064 nr_scanned, nr_isolated);
1065
1066fatal_pending:
1067 cc->total_migrate_scanned += nr_scanned;
1068 if (nr_isolated)
1069 count_compact_events(COMPACTISOLATED, nr_isolated);
1070
1071 return low_pfn;
1072}
1073
1074/**
1075 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
1076 * @cc: Compaction control structure.
1077 * @start_pfn: The first PFN to start isolating.
1078 * @end_pfn: The one-past-last PFN.
1079 *
1080 * Returns zero if isolation fails fatally due to e.g. pending signal.
1081 * Otherwise, function returns one-past-the-last PFN of isolated page
1082 * (which may be greater than end_pfn if end fell in a middle of a THP page).
1083 */
1084unsigned long
1085isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
1086 unsigned long end_pfn)
1087{
1088 unsigned long pfn, block_start_pfn, block_end_pfn;
1089
1090 /* Scan block by block. First and last block may be incomplete */
1091 pfn = start_pfn;
1092 block_start_pfn = pageblock_start_pfn(pfn);
1093 if (block_start_pfn < cc->zone->zone_start_pfn)
1094 block_start_pfn = cc->zone->zone_start_pfn;
1095 block_end_pfn = pageblock_end_pfn(pfn);
1096
1097 for (; pfn < end_pfn; pfn = block_end_pfn,
1098 block_start_pfn = block_end_pfn,
1099 block_end_pfn += pageblock_nr_pages) {
1100
1101 block_end_pfn = min(block_end_pfn, end_pfn);
1102
1103 if (!pageblock_pfn_to_page(block_start_pfn,
1104 block_end_pfn, cc->zone))
1105 continue;
1106
1107 pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
1108 ISOLATE_UNEVICTABLE);
1109
1110 if (!pfn)
1111 break;
1112
1113 if (cc->nr_migratepages == COMPACT_CLUSTER_MAX)
1114 break;
1115 }
1116
1117 return pfn;
1118}
1119
1120#endif /* CONFIG_COMPACTION || CONFIG_CMA */
1121#ifdef CONFIG_COMPACTION
1122
1123static bool suitable_migration_source(struct compact_control *cc,
1124 struct page *page)
1125{
1126 int block_mt;
1127
1128 if (pageblock_skip_persistent(page))
1129 return false;
1130
1131 if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction)
1132 return true;
1133
1134 block_mt = get_pageblock_migratetype(page);
1135
1136 if (cc->migratetype == MIGRATE_MOVABLE)
1137 return is_migrate_movable(block_mt);
1138 else
1139 return block_mt == cc->migratetype;
1140}
1141
1142/* Returns true if the page is within a block suitable for migration to */
1143static bool suitable_migration_target(struct compact_control *cc,
1144 struct page *page)
1145{
1146 /* If the page is a large free page, then disallow migration */
1147 if (PageBuddy(page)) {
1148 /*
1149 * We are checking page_order without zone->lock taken. But
1150 * the only small danger is that we skip a potentially suitable
1151 * pageblock, so it's not worth to check order for valid range.
1152 */
1153 if (page_order_unsafe(page) >= pageblock_order)
1154 return false;
1155 }
1156
1157 if (cc->ignore_block_suitable)
1158 return true;
1159
1160 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1161 if (is_migrate_movable(get_pageblock_migratetype(page)))
1162 return true;
1163
1164 /* Otherwise skip the block */
1165 return false;
1166}
1167
1168static inline unsigned int
1169freelist_scan_limit(struct compact_control *cc)
1170{
1171 unsigned short shift = BITS_PER_LONG - 1;
1172
1173 return (COMPACT_CLUSTER_MAX >> min(shift, cc->fast_search_fail)) + 1;
1174}
1175
1176/*
1177 * Test whether the free scanner has reached the same or lower pageblock than
1178 * the migration scanner, and compaction should thus terminate.
1179 */
1180static inline bool compact_scanners_met(struct compact_control *cc)
1181{
1182 return (cc->free_pfn >> pageblock_order)
1183 <= (cc->migrate_pfn >> pageblock_order);
1184}
1185
1186/*
1187 * Used when scanning for a suitable migration target which scans freelists
1188 * in reverse. Reorders the list such as the unscanned pages are scanned
1189 * first on the next iteration of the free scanner
1190 */
1191static void
1192move_freelist_head(struct list_head *freelist, struct page *freepage)
1193{
1194 LIST_HEAD(sublist);
1195
1196 if (!list_is_last(freelist, &freepage->lru)) {
1197 list_cut_before(&sublist, freelist, &freepage->lru);
1198 if (!list_empty(&sublist))
1199 list_splice_tail(&sublist, freelist);
1200 }
1201}
1202
1203/*
1204 * Similar to move_freelist_head except used by the migration scanner
1205 * when scanning forward. It's possible for these list operations to
1206 * move against each other if they search the free list exactly in
1207 * lockstep.
1208 */
1209static void
1210move_freelist_tail(struct list_head *freelist, struct page *freepage)
1211{
1212 LIST_HEAD(sublist);
1213
1214 if (!list_is_first(freelist, &freepage->lru)) {
1215 list_cut_position(&sublist, freelist, &freepage->lru);
1216 if (!list_empty(&sublist))
1217 list_splice_tail(&sublist, freelist);
1218 }
1219}
1220
1221static void
1222fast_isolate_around(struct compact_control *cc, unsigned long pfn, unsigned long nr_isolated)
1223{
1224 unsigned long start_pfn, end_pfn;
1225 struct page *page = pfn_to_page(pfn);
1226
1227 /* Do not search around if there are enough pages already */
1228 if (cc->nr_freepages >= cc->nr_migratepages)
1229 return;
1230
1231 /* Minimise scanning during async compaction */
1232 if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC)
1233 return;
1234
1235 /* Pageblock boundaries */
1236 start_pfn = pageblock_start_pfn(pfn);
1237 end_pfn = min(pageblock_end_pfn(pfn), zone_end_pfn(cc->zone)) - 1;
1238
1239 /* Scan before */
1240 if (start_pfn != pfn) {
1241 isolate_freepages_block(cc, &start_pfn, pfn, &cc->freepages, 1, false);
1242 if (cc->nr_freepages >= cc->nr_migratepages)
1243 return;
1244 }
1245
1246 /* Scan after */
1247 start_pfn = pfn + nr_isolated;
1248 if (start_pfn < end_pfn)
1249 isolate_freepages_block(cc, &start_pfn, end_pfn, &cc->freepages, 1, false);
1250
1251 /* Skip this pageblock in the future as it's full or nearly full */
1252 if (cc->nr_freepages < cc->nr_migratepages)
1253 set_pageblock_skip(page);
1254}
1255
1256/* Search orders in round-robin fashion */
1257static int next_search_order(struct compact_control *cc, int order)
1258{
1259 order--;
1260 if (order < 0)
1261 order = cc->order - 1;
1262
1263 /* Search wrapped around? */
1264 if (order == cc->search_order) {
1265 cc->search_order--;
1266 if (cc->search_order < 0)
1267 cc->search_order = cc->order - 1;
1268 return -1;
1269 }
1270
1271 return order;
1272}
1273
1274static unsigned long
1275fast_isolate_freepages(struct compact_control *cc)
1276{
1277 unsigned int limit = min(1U, freelist_scan_limit(cc) >> 1);
1278 unsigned int nr_scanned = 0;
1279 unsigned long low_pfn, min_pfn, high_pfn = 0, highest = 0;
1280 unsigned long nr_isolated = 0;
1281 unsigned long distance;
1282 struct page *page = NULL;
1283 bool scan_start = false;
1284 int order;
1285
1286 /* Full compaction passes in a negative order */
1287 if (cc->order <= 0)
1288 return cc->free_pfn;
1289
1290 /*
1291 * If starting the scan, use a deeper search and use the highest
1292 * PFN found if a suitable one is not found.
1293 */
1294 if (cc->free_pfn >= cc->zone->compact_init_free_pfn) {
1295 limit = pageblock_nr_pages >> 1;
1296 scan_start = true;
1297 }
1298
1299 /*
1300 * Preferred point is in the top quarter of the scan space but take
1301 * a pfn from the top half if the search is problematic.
1302 */
1303 distance = (cc->free_pfn - cc->migrate_pfn);
1304 low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2));
1305 min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1));
1306
1307 if (WARN_ON_ONCE(min_pfn > low_pfn))
1308 low_pfn = min_pfn;
1309
1310 /*
1311 * Search starts from the last successful isolation order or the next
1312 * order to search after a previous failure
1313 */
1314 cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order);
1315
1316 for (order = cc->search_order;
1317 !page && order >= 0;
1318 order = next_search_order(cc, order)) {
1319 struct free_area *area = &cc->zone->free_area[order];
1320 struct list_head *freelist;
1321 struct page *freepage;
1322 unsigned long flags;
1323 unsigned int order_scanned = 0;
1324
1325 if (!area->nr_free)
1326 continue;
1327
1328 spin_lock_irqsave(&cc->zone->lock, flags);
1329 freelist = &area->free_list[MIGRATE_MOVABLE];
1330 list_for_each_entry_reverse(freepage, freelist, lru) {
1331 unsigned long pfn;
1332
1333 order_scanned++;
1334 nr_scanned++;
1335 pfn = page_to_pfn(freepage);
1336
1337 if (pfn >= highest)
1338 highest = pageblock_start_pfn(pfn);
1339
1340 if (pfn >= low_pfn) {
1341 cc->fast_search_fail = 0;
1342 cc->search_order = order;
1343 page = freepage;
1344 break;
1345 }
1346
1347 if (pfn >= min_pfn && pfn > high_pfn) {
1348 high_pfn = pfn;
1349
1350 /* Shorten the scan if a candidate is found */
1351 limit >>= 1;
1352 }
1353
1354 if (order_scanned >= limit)
1355 break;
1356 }
1357
1358 /* Use a minimum pfn if a preferred one was not found */
1359 if (!page && high_pfn) {
1360 page = pfn_to_page(high_pfn);
1361
1362 /* Update freepage for the list reorder below */
1363 freepage = page;
1364 }
1365
1366 /* Reorder to so a future search skips recent pages */
1367 move_freelist_head(freelist, freepage);
1368
1369 /* Isolate the page if available */
1370 if (page) {
1371 if (__isolate_free_page(page, order)) {
1372 set_page_private(page, order);
1373 nr_isolated = 1 << order;
1374 cc->nr_freepages += nr_isolated;
1375 list_add_tail(&page->lru, &cc->freepages);
1376 count_compact_events(COMPACTISOLATED, nr_isolated);
1377 } else {
1378 /* If isolation fails, abort the search */
1379 order = cc->search_order + 1;
1380 page = NULL;
1381 }
1382 }
1383
1384 spin_unlock_irqrestore(&cc->zone->lock, flags);
1385
1386 /*
1387 * Smaller scan on next order so the total scan ig related
1388 * to freelist_scan_limit.
1389 */
1390 if (order_scanned >= limit)
1391 limit = min(1U, limit >> 1);
1392 }
1393
1394 if (!page) {
1395 cc->fast_search_fail++;
1396 if (scan_start) {
1397 /*
1398 * Use the highest PFN found above min. If one was
1399 * not found, be pessemistic for direct compaction
1400 * and use the min mark.
1401 */
1402 if (highest) {
1403 page = pfn_to_page(highest);
1404 cc->free_pfn = highest;
1405 } else {
1406 if (cc->direct_compaction && pfn_valid(min_pfn)) {
1407 page = pfn_to_page(min_pfn);
1408 cc->free_pfn = min_pfn;
1409 }
1410 }
1411 }
1412 }
1413
1414 if (highest && highest >= cc->zone->compact_cached_free_pfn) {
1415 highest -= pageblock_nr_pages;
1416 cc->zone->compact_cached_free_pfn = highest;
1417 }
1418
1419 cc->total_free_scanned += nr_scanned;
1420 if (!page)
1421 return cc->free_pfn;
1422
1423 low_pfn = page_to_pfn(page);
1424 fast_isolate_around(cc, low_pfn, nr_isolated);
1425 return low_pfn;
1426}
1427
1428/*
1429 * Based on information in the current compact_control, find blocks
1430 * suitable for isolating free pages from and then isolate them.
1431 */
1432static void isolate_freepages(struct compact_control *cc)
1433{
1434 struct zone *zone = cc->zone;
1435 struct page *page;
1436 unsigned long block_start_pfn; /* start of current pageblock */
1437 unsigned long isolate_start_pfn; /* exact pfn we start at */
1438 unsigned long block_end_pfn; /* end of current pageblock */
1439 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
1440 struct list_head *freelist = &cc->freepages;
1441 unsigned int stride;
1442
1443 /* Try a small search of the free lists for a candidate */
1444 isolate_start_pfn = fast_isolate_freepages(cc);
1445 if (cc->nr_freepages)
1446 goto splitmap;
1447
1448 /*
1449 * Initialise the free scanner. The starting point is where we last
1450 * successfully isolated from, zone-cached value, or the end of the
1451 * zone when isolating for the first time. For looping we also need
1452 * this pfn aligned down to the pageblock boundary, because we do
1453 * block_start_pfn -= pageblock_nr_pages in the for loop.
1454 * For ending point, take care when isolating in last pageblock of a
1455 * a zone which ends in the middle of a pageblock.
1456 * The low boundary is the end of the pageblock the migration scanner
1457 * is using.
1458 */
1459 isolate_start_pfn = cc->free_pfn;
1460 block_start_pfn = pageblock_start_pfn(isolate_start_pfn);
1461 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1462 zone_end_pfn(zone));
1463 low_pfn = pageblock_end_pfn(cc->migrate_pfn);
1464 stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1;
1465
1466 /*
1467 * Isolate free pages until enough are available to migrate the
1468 * pages on cc->migratepages. We stop searching if the migrate
1469 * and free page scanners meet or enough free pages are isolated.
1470 */
1471 for (; block_start_pfn >= low_pfn;
1472 block_end_pfn = block_start_pfn,
1473 block_start_pfn -= pageblock_nr_pages,
1474 isolate_start_pfn = block_start_pfn) {
1475 unsigned long nr_isolated;
1476
1477 /*
1478 * This can iterate a massively long zone without finding any
1479 * suitable migration targets, so periodically check resched.
1480 */
1481 if (!(block_start_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)))
1482 cond_resched();
1483
1484 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1485 zone);
1486 if (!page)
1487 continue;
1488
1489 /* Check the block is suitable for migration */
1490 if (!suitable_migration_target(cc, page))
1491 continue;
1492
1493 /* If isolation recently failed, do not retry */
1494 if (!isolation_suitable(cc, page))
1495 continue;
1496
1497 /* Found a block suitable for isolating free pages from. */
1498 nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn,
1499 block_end_pfn, freelist, stride, false);
1500
1501 /* Update the skip hint if the full pageblock was scanned */
1502 if (isolate_start_pfn == block_end_pfn)
1503 update_pageblock_skip(cc, page, block_start_pfn);
1504
1505 /* Are enough freepages isolated? */
1506 if (cc->nr_freepages >= cc->nr_migratepages) {
1507 if (isolate_start_pfn >= block_end_pfn) {
1508 /*
1509 * Restart at previous pageblock if more
1510 * freepages can be isolated next time.
1511 */
1512 isolate_start_pfn =
1513 block_start_pfn - pageblock_nr_pages;
1514 }
1515 break;
1516 } else if (isolate_start_pfn < block_end_pfn) {
1517 /*
1518 * If isolation failed early, do not continue
1519 * needlessly.
1520 */
1521 break;
1522 }
1523
1524 /* Adjust stride depending on isolation */
1525 if (nr_isolated) {
1526 stride = 1;
1527 continue;
1528 }
1529 stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1);
1530 }
1531
1532 /*
1533 * Record where the free scanner will restart next time. Either we
1534 * broke from the loop and set isolate_start_pfn based on the last
1535 * call to isolate_freepages_block(), or we met the migration scanner
1536 * and the loop terminated due to isolate_start_pfn < low_pfn
1537 */
1538 cc->free_pfn = isolate_start_pfn;
1539
1540splitmap:
1541 /* __isolate_free_page() does not map the pages */
1542 split_map_pages(freelist);
1543}
1544
1545/*
1546 * This is a migrate-callback that "allocates" freepages by taking pages
1547 * from the isolated freelists in the block we are migrating to.
1548 */
1549static struct page *compaction_alloc(struct page *migratepage,
1550 unsigned long data)
1551{
1552 struct compact_control *cc = (struct compact_control *)data;
1553 struct page *freepage;
1554
1555 if (list_empty(&cc->freepages)) {
1556 isolate_freepages(cc);
1557
1558 if (list_empty(&cc->freepages))
1559 return NULL;
1560 }
1561
1562 freepage = list_entry(cc->freepages.next, struct page, lru);
1563 list_del(&freepage->lru);
1564 cc->nr_freepages--;
1565
1566 return freepage;
1567}
1568
1569/*
1570 * This is a migrate-callback that "frees" freepages back to the isolated
1571 * freelist. All pages on the freelist are from the same zone, so there is no
1572 * special handling needed for NUMA.
1573 */
1574static void compaction_free(struct page *page, unsigned long data)
1575{
1576 struct compact_control *cc = (struct compact_control *)data;
1577
1578 list_add(&page->lru, &cc->freepages);
1579 cc->nr_freepages++;
1580}
1581
1582/* possible outcome of isolate_migratepages */
1583typedef enum {
1584 ISOLATE_ABORT, /* Abort compaction now */
1585 ISOLATE_NONE, /* No pages isolated, continue scanning */
1586 ISOLATE_SUCCESS, /* Pages isolated, migrate */
1587} isolate_migrate_t;
1588
1589/*
1590 * Allow userspace to control policy on scanning the unevictable LRU for
1591 * compactable pages.
1592 */
1593int sysctl_compact_unevictable_allowed __read_mostly = 1;
1594
1595static inline void
1596update_fast_start_pfn(struct compact_control *cc, unsigned long pfn)
1597{
1598 if (cc->fast_start_pfn == ULONG_MAX)
1599 return;
1600
1601 if (!cc->fast_start_pfn)
1602 cc->fast_start_pfn = pfn;
1603
1604 cc->fast_start_pfn = min(cc->fast_start_pfn, pfn);
1605}
1606
1607static inline unsigned long
1608reinit_migrate_pfn(struct compact_control *cc)
1609{
1610 if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX)
1611 return cc->migrate_pfn;
1612
1613 cc->migrate_pfn = cc->fast_start_pfn;
1614 cc->fast_start_pfn = ULONG_MAX;
1615
1616 return cc->migrate_pfn;
1617}
1618
1619/*
1620 * Briefly search the free lists for a migration source that already has
1621 * some free pages to reduce the number of pages that need migration
1622 * before a pageblock is free.
1623 */
1624static unsigned long fast_find_migrateblock(struct compact_control *cc)
1625{
1626 unsigned int limit = freelist_scan_limit(cc);
1627 unsigned int nr_scanned = 0;
1628 unsigned long distance;
1629 unsigned long pfn = cc->migrate_pfn;
1630 unsigned long high_pfn;
1631 int order;
1632
1633 /* Skip hints are relied on to avoid repeats on the fast search */
1634 if (cc->ignore_skip_hint)
1635 return pfn;
1636
1637 /*
1638 * If the migrate_pfn is not at the start of a zone or the start
1639 * of a pageblock then assume this is a continuation of a previous
1640 * scan restarted due to COMPACT_CLUSTER_MAX.
1641 */
1642 if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn))
1643 return pfn;
1644
1645 /*
1646 * For smaller orders, just linearly scan as the number of pages
1647 * to migrate should be relatively small and does not necessarily
1648 * justify freeing up a large block for a small allocation.
1649 */
1650 if (cc->order <= PAGE_ALLOC_COSTLY_ORDER)
1651 return pfn;
1652
1653 /*
1654 * Only allow kcompactd and direct requests for movable pages to
1655 * quickly clear out a MOVABLE pageblock for allocation. This
1656 * reduces the risk that a large movable pageblock is freed for
1657 * an unmovable/reclaimable small allocation.
1658 */
1659 if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE)
1660 return pfn;
1661
1662 /*
1663 * When starting the migration scanner, pick any pageblock within the
1664 * first half of the search space. Otherwise try and pick a pageblock
1665 * within the first eighth to reduce the chances that a migration
1666 * target later becomes a source.
1667 */
1668 distance = (cc->free_pfn - cc->migrate_pfn) >> 1;
1669 if (cc->migrate_pfn != cc->zone->zone_start_pfn)
1670 distance >>= 2;
1671 high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance);
1672
1673 for (order = cc->order - 1;
1674 order >= PAGE_ALLOC_COSTLY_ORDER && pfn == cc->migrate_pfn && nr_scanned < limit;
1675 order--) {
1676 struct free_area *area = &cc->zone->free_area[order];
1677 struct list_head *freelist;
1678 unsigned long flags;
1679 struct page *freepage;
1680
1681 if (!area->nr_free)
1682 continue;
1683
1684 spin_lock_irqsave(&cc->zone->lock, flags);
1685 freelist = &area->free_list[MIGRATE_MOVABLE];
1686 list_for_each_entry(freepage, freelist, lru) {
1687 unsigned long free_pfn;
1688
1689 nr_scanned++;
1690 free_pfn = page_to_pfn(freepage);
1691 if (free_pfn < high_pfn) {
1692 /*
1693 * Avoid if skipped recently. Ideally it would
1694 * move to the tail but even safe iteration of
1695 * the list assumes an entry is deleted, not
1696 * reordered.
1697 */
1698 if (get_pageblock_skip(freepage)) {
1699 if (list_is_last(freelist, &freepage->lru))
1700 break;
1701
1702 continue;
1703 }
1704
1705 /* Reorder to so a future search skips recent pages */
1706 move_freelist_tail(freelist, freepage);
1707
1708 update_fast_start_pfn(cc, free_pfn);
1709 pfn = pageblock_start_pfn(free_pfn);
1710 cc->fast_search_fail = 0;
1711 set_pageblock_skip(freepage);
1712 break;
1713 }
1714
1715 if (nr_scanned >= limit) {
1716 cc->fast_search_fail++;
1717 move_freelist_tail(freelist, freepage);
1718 break;
1719 }
1720 }
1721 spin_unlock_irqrestore(&cc->zone->lock, flags);
1722 }
1723
1724 cc->total_migrate_scanned += nr_scanned;
1725
1726 /*
1727 * If fast scanning failed then use a cached entry for a page block
1728 * that had free pages as the basis for starting a linear scan.
1729 */
1730 if (pfn == cc->migrate_pfn)
1731 pfn = reinit_migrate_pfn(cc);
1732
1733 return pfn;
1734}
1735
1736/*
1737 * Isolate all pages that can be migrated from the first suitable block,
1738 * starting at the block pointed to by the migrate scanner pfn within
1739 * compact_control.
1740 */
1741static isolate_migrate_t isolate_migratepages(struct compact_control *cc)
1742{
1743 unsigned long block_start_pfn;
1744 unsigned long block_end_pfn;
1745 unsigned long low_pfn;
1746 struct page *page;
1747 const isolate_mode_t isolate_mode =
1748 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1749 (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1750 bool fast_find_block;
1751
1752 /*
1753 * Start at where we last stopped, or beginning of the zone as
1754 * initialized by compact_zone(). The first failure will use
1755 * the lowest PFN as the starting point for linear scanning.
1756 */
1757 low_pfn = fast_find_migrateblock(cc);
1758 block_start_pfn = pageblock_start_pfn(low_pfn);
1759 if (block_start_pfn < cc->zone->zone_start_pfn)
1760 block_start_pfn = cc->zone->zone_start_pfn;
1761
1762 /*
1763 * fast_find_migrateblock marks a pageblock skipped so to avoid
1764 * the isolation_suitable check below, check whether the fast
1765 * search was successful.
1766 */
1767 fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail;
1768
1769 /* Only scan within a pageblock boundary */
1770 block_end_pfn = pageblock_end_pfn(low_pfn);
1771
1772 /*
1773 * Iterate over whole pageblocks until we find the first suitable.
1774 * Do not cross the free scanner.
1775 */
1776 for (; block_end_pfn <= cc->free_pfn;
1777 fast_find_block = false,
1778 low_pfn = block_end_pfn,
1779 block_start_pfn = block_end_pfn,
1780 block_end_pfn += pageblock_nr_pages) {
1781
1782 /*
1783 * This can potentially iterate a massively long zone with
1784 * many pageblocks unsuitable, so periodically check if we
1785 * need to schedule.
1786 */
1787 if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages)))
1788 cond_resched();
1789
1790 page = pageblock_pfn_to_page(block_start_pfn,
1791 block_end_pfn, cc->zone);
1792 if (!page)
1793 continue;
1794
1795 /*
1796 * If isolation recently failed, do not retry. Only check the
1797 * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock
1798 * to be visited multiple times. Assume skip was checked
1799 * before making it "skip" so other compaction instances do
1800 * not scan the same block.
1801 */
1802 if (IS_ALIGNED(low_pfn, pageblock_nr_pages) &&
1803 !fast_find_block && !isolation_suitable(cc, page))
1804 continue;
1805
1806 /*
1807 * For async compaction, also only scan in MOVABLE blocks
1808 * without huge pages. Async compaction is optimistic to see
1809 * if the minimum amount of work satisfies the allocation.
1810 * The cached PFN is updated as it's possible that all
1811 * remaining blocks between source and target are unsuitable
1812 * and the compaction scanners fail to meet.
1813 */
1814 if (!suitable_migration_source(cc, page)) {
1815 update_cached_migrate(cc, block_end_pfn);
1816 continue;
1817 }
1818
1819 /* Perform the isolation */
1820 low_pfn = isolate_migratepages_block(cc, low_pfn,
1821 block_end_pfn, isolate_mode);
1822
1823 if (!low_pfn)
1824 return ISOLATE_ABORT;
1825
1826 /*
1827 * Either we isolated something and proceed with migration. Or
1828 * we failed and compact_zone should decide if we should
1829 * continue or not.
1830 */
1831 break;
1832 }
1833
1834 /* Record where migration scanner will be restarted. */
1835 cc->migrate_pfn = low_pfn;
1836
1837 return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
1838}
1839
1840/*
1841 * order == -1 is expected when compacting via
1842 * /proc/sys/vm/compact_memory
1843 */
1844static inline bool is_via_compact_memory(int order)
1845{
1846 return order == -1;
1847}
1848
1849static enum compact_result __compact_finished(struct compact_control *cc)
1850{
1851 unsigned int order;
1852 const int migratetype = cc->migratetype;
1853 int ret;
1854
1855 /* Compaction run completes if the migrate and free scanner meet */
1856 if (compact_scanners_met(cc)) {
1857 /* Let the next compaction start anew. */
1858 reset_cached_positions(cc->zone);
1859
1860 /*
1861 * Mark that the PG_migrate_skip information should be cleared
1862 * by kswapd when it goes to sleep. kcompactd does not set the
1863 * flag itself as the decision to be clear should be directly
1864 * based on an allocation request.
1865 */
1866 if (cc->direct_compaction)
1867 cc->zone->compact_blockskip_flush = true;
1868
1869 if (cc->whole_zone)
1870 return COMPACT_COMPLETE;
1871 else
1872 return COMPACT_PARTIAL_SKIPPED;
1873 }
1874
1875 if (is_via_compact_memory(cc->order))
1876 return COMPACT_CONTINUE;
1877
1878 /*
1879 * Always finish scanning a pageblock to reduce the possibility of
1880 * fallbacks in the future. This is particularly important when
1881 * migration source is unmovable/reclaimable but it's not worth
1882 * special casing.
1883 */
1884 if (!IS_ALIGNED(cc->migrate_pfn, pageblock_nr_pages))
1885 return COMPACT_CONTINUE;
1886
1887 /* Direct compactor: Is a suitable page free? */
1888 ret = COMPACT_NO_SUITABLE_PAGE;
1889 for (order = cc->order; order < MAX_ORDER; order++) {
1890 struct free_area *area = &cc->zone->free_area[order];
1891 bool can_steal;
1892
1893 /* Job done if page is free of the right migratetype */
1894 if (!free_area_empty(area, migratetype))
1895 return COMPACT_SUCCESS;
1896
1897#ifdef CONFIG_CMA
1898 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
1899 if (migratetype == MIGRATE_MOVABLE &&
1900 !free_area_empty(area, MIGRATE_CMA))
1901 return COMPACT_SUCCESS;
1902#endif
1903 /*
1904 * Job done if allocation would steal freepages from
1905 * other migratetype buddy lists.
1906 */
1907 if (find_suitable_fallback(area, order, migratetype,
1908 true, &can_steal) != -1) {
1909
1910 /* movable pages are OK in any pageblock */
1911 if (migratetype == MIGRATE_MOVABLE)
1912 return COMPACT_SUCCESS;
1913
1914 /*
1915 * We are stealing for a non-movable allocation. Make
1916 * sure we finish compacting the current pageblock
1917 * first so it is as free as possible and we won't
1918 * have to steal another one soon. This only applies
1919 * to sync compaction, as async compaction operates
1920 * on pageblocks of the same migratetype.
1921 */
1922 if (cc->mode == MIGRATE_ASYNC ||
1923 IS_ALIGNED(cc->migrate_pfn,
1924 pageblock_nr_pages)) {
1925 return COMPACT_SUCCESS;
1926 }
1927
1928 ret = COMPACT_CONTINUE;
1929 break;
1930 }
1931 }
1932
1933 if (cc->contended || fatal_signal_pending(current))
1934 ret = COMPACT_CONTENDED;
1935
1936 return ret;
1937}
1938
1939static enum compact_result compact_finished(struct compact_control *cc)
1940{
1941 int ret;
1942
1943 ret = __compact_finished(cc);
1944 trace_mm_compaction_finished(cc->zone, cc->order, ret);
1945 if (ret == COMPACT_NO_SUITABLE_PAGE)
1946 ret = COMPACT_CONTINUE;
1947
1948 return ret;
1949}
1950
1951/*
1952 * compaction_suitable: Is this suitable to run compaction on this zone now?
1953 * Returns
1954 * COMPACT_SKIPPED - If there are too few free pages for compaction
1955 * COMPACT_SUCCESS - If the allocation would succeed without compaction
1956 * COMPACT_CONTINUE - If compaction should run now
1957 */
1958static enum compact_result __compaction_suitable(struct zone *zone, int order,
1959 unsigned int alloc_flags,
1960 int classzone_idx,
1961 unsigned long wmark_target)
1962{
1963 unsigned long watermark;
1964
1965 if (is_via_compact_memory(order))
1966 return COMPACT_CONTINUE;
1967
1968 watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
1969 /*
1970 * If watermarks for high-order allocation are already met, there
1971 * should be no need for compaction at all.
1972 */
1973 if (zone_watermark_ok(zone, order, watermark, classzone_idx,
1974 alloc_flags))
1975 return COMPACT_SUCCESS;
1976
1977 /*
1978 * Watermarks for order-0 must be met for compaction to be able to
1979 * isolate free pages for migration targets. This means that the
1980 * watermark and alloc_flags have to match, or be more pessimistic than
1981 * the check in __isolate_free_page(). We don't use the direct
1982 * compactor's alloc_flags, as they are not relevant for freepage
1983 * isolation. We however do use the direct compactor's classzone_idx to
1984 * skip over zones where lowmem reserves would prevent allocation even
1985 * if compaction succeeds.
1986 * For costly orders, we require low watermark instead of min for
1987 * compaction to proceed to increase its chances.
1988 * ALLOC_CMA is used, as pages in CMA pageblocks are considered
1989 * suitable migration targets
1990 */
1991 watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
1992 low_wmark_pages(zone) : min_wmark_pages(zone);
1993 watermark += compact_gap(order);
1994 if (!__zone_watermark_ok(zone, 0, watermark, classzone_idx,
1995 ALLOC_CMA, wmark_target))
1996 return COMPACT_SKIPPED;
1997
1998 return COMPACT_CONTINUE;
1999}
2000
2001enum compact_result compaction_suitable(struct zone *zone, int order,
2002 unsigned int alloc_flags,
2003 int classzone_idx)
2004{
2005 enum compact_result ret;
2006 int fragindex;
2007
2008 ret = __compaction_suitable(zone, order, alloc_flags, classzone_idx,
2009 zone_page_state(zone, NR_FREE_PAGES));
2010 /*
2011 * fragmentation index determines if allocation failures are due to
2012 * low memory or external fragmentation
2013 *
2014 * index of -1000 would imply allocations might succeed depending on
2015 * watermarks, but we already failed the high-order watermark check
2016 * index towards 0 implies failure is due to lack of memory
2017 * index towards 1000 implies failure is due to fragmentation
2018 *
2019 * Only compact if a failure would be due to fragmentation. Also
2020 * ignore fragindex for non-costly orders where the alternative to
2021 * a successful reclaim/compaction is OOM. Fragindex and the
2022 * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
2023 * excessive compaction for costly orders, but it should not be at the
2024 * expense of system stability.
2025 */
2026 if (ret == COMPACT_CONTINUE && (order > PAGE_ALLOC_COSTLY_ORDER)) {
2027 fragindex = fragmentation_index(zone, order);
2028 if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
2029 ret = COMPACT_NOT_SUITABLE_ZONE;
2030 }
2031
2032 trace_mm_compaction_suitable(zone, order, ret);
2033 if (ret == COMPACT_NOT_SUITABLE_ZONE)
2034 ret = COMPACT_SKIPPED;
2035
2036 return ret;
2037}
2038
2039bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
2040 int alloc_flags)
2041{
2042 struct zone *zone;
2043 struct zoneref *z;
2044
2045 /*
2046 * Make sure at least one zone would pass __compaction_suitable if we continue
2047 * retrying the reclaim.
2048 */
2049 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
2050 ac->nodemask) {
2051 unsigned long available;
2052 enum compact_result compact_result;
2053
2054 /*
2055 * Do not consider all the reclaimable memory because we do not
2056 * want to trash just for a single high order allocation which
2057 * is even not guaranteed to appear even if __compaction_suitable
2058 * is happy about the watermark check.
2059 */
2060 available = zone_reclaimable_pages(zone) / order;
2061 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
2062 compact_result = __compaction_suitable(zone, order, alloc_flags,
2063 ac_classzone_idx(ac), available);
2064 if (compact_result != COMPACT_SKIPPED)
2065 return true;
2066 }
2067
2068 return false;
2069}
2070
2071static enum compact_result
2072compact_zone(struct compact_control *cc, struct capture_control *capc)
2073{
2074 enum compact_result ret;
2075 unsigned long start_pfn = cc->zone->zone_start_pfn;
2076 unsigned long end_pfn = zone_end_pfn(cc->zone);
2077 unsigned long last_migrated_pfn;
2078 const bool sync = cc->mode != MIGRATE_ASYNC;
2079 bool update_cached;
2080
2081 /*
2082 * These counters track activities during zone compaction. Initialize
2083 * them before compacting a new zone.
2084 */
2085 cc->total_migrate_scanned = 0;
2086 cc->total_free_scanned = 0;
2087 cc->nr_migratepages = 0;
2088 cc->nr_freepages = 0;
2089 INIT_LIST_HEAD(&cc->freepages);
2090 INIT_LIST_HEAD(&cc->migratepages);
2091
2092 cc->migratetype = gfpflags_to_migratetype(cc->gfp_mask);
2093 ret = compaction_suitable(cc->zone, cc->order, cc->alloc_flags,
2094 cc->classzone_idx);
2095 /* Compaction is likely to fail */
2096 if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED)
2097 return ret;
2098
2099 /* huh, compaction_suitable is returning something unexpected */
2100 VM_BUG_ON(ret != COMPACT_CONTINUE);
2101
2102 /*
2103 * Clear pageblock skip if there were failures recently and compaction
2104 * is about to be retried after being deferred.
2105 */
2106 if (compaction_restarting(cc->zone, cc->order))
2107 __reset_isolation_suitable(cc->zone);
2108
2109 /*
2110 * Setup to move all movable pages to the end of the zone. Used cached
2111 * information on where the scanners should start (unless we explicitly
2112 * want to compact the whole zone), but check that it is initialised
2113 * by ensuring the values are within zone boundaries.
2114 */
2115 cc->fast_start_pfn = 0;
2116 if (cc->whole_zone) {
2117 cc->migrate_pfn = start_pfn;
2118 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2119 } else {
2120 cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync];
2121 cc->free_pfn = cc->zone->compact_cached_free_pfn;
2122 if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
2123 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2124 cc->zone->compact_cached_free_pfn = cc->free_pfn;
2125 }
2126 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
2127 cc->migrate_pfn = start_pfn;
2128 cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
2129 cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
2130 }
2131
2132 if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn)
2133 cc->whole_zone = true;
2134 }
2135
2136 last_migrated_pfn = 0;
2137
2138 /*
2139 * Migrate has separate cached PFNs for ASYNC and SYNC* migration on
2140 * the basis that some migrations will fail in ASYNC mode. However,
2141 * if the cached PFNs match and pageblocks are skipped due to having
2142 * no isolation candidates, then the sync state does not matter.
2143 * Until a pageblock with isolation candidates is found, keep the
2144 * cached PFNs in sync to avoid revisiting the same blocks.
2145 */
2146 update_cached = !sync &&
2147 cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1];
2148
2149 trace_mm_compaction_begin(start_pfn, cc->migrate_pfn,
2150 cc->free_pfn, end_pfn, sync);
2151
2152 migrate_prep_local();
2153
2154 while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) {
2155 int err;
2156 unsigned long start_pfn = cc->migrate_pfn;
2157
2158 /*
2159 * Avoid multiple rescans which can happen if a page cannot be
2160 * isolated (dirty/writeback in async mode) or if the migrated
2161 * pages are being allocated before the pageblock is cleared.
2162 * The first rescan will capture the entire pageblock for
2163 * migration. If it fails, it'll be marked skip and scanning
2164 * will proceed as normal.
2165 */
2166 cc->rescan = false;
2167 if (pageblock_start_pfn(last_migrated_pfn) ==
2168 pageblock_start_pfn(start_pfn)) {
2169 cc->rescan = true;
2170 }
2171
2172 switch (isolate_migratepages(cc)) {
2173 case ISOLATE_ABORT:
2174 ret = COMPACT_CONTENDED;
2175 putback_movable_pages(&cc->migratepages);
2176 cc->nr_migratepages = 0;
2177 last_migrated_pfn = 0;
2178 goto out;
2179 case ISOLATE_NONE:
2180 if (update_cached) {
2181 cc->zone->compact_cached_migrate_pfn[1] =
2182 cc->zone->compact_cached_migrate_pfn[0];
2183 }
2184
2185 /*
2186 * We haven't isolated and migrated anything, but
2187 * there might still be unflushed migrations from
2188 * previous cc->order aligned block.
2189 */
2190 goto check_drain;
2191 case ISOLATE_SUCCESS:
2192 update_cached = false;
2193 last_migrated_pfn = start_pfn;
2194 ;
2195 }
2196
2197 err = migrate_pages(&cc->migratepages, compaction_alloc,
2198 compaction_free, (unsigned long)cc, cc->mode,
2199 MR_COMPACTION);
2200
2201 trace_mm_compaction_migratepages(cc->nr_migratepages, err,
2202 &cc->migratepages);
2203
2204 /* All pages were either migrated or will be released */
2205 cc->nr_migratepages = 0;
2206 if (err) {
2207 putback_movable_pages(&cc->migratepages);
2208 /*
2209 * migrate_pages() may return -ENOMEM when scanners meet
2210 * and we want compact_finished() to detect it
2211 */
2212 if (err == -ENOMEM && !compact_scanners_met(cc)) {
2213 ret = COMPACT_CONTENDED;
2214 goto out;
2215 }
2216 /*
2217 * We failed to migrate at least one page in the current
2218 * order-aligned block, so skip the rest of it.
2219 */
2220 if (cc->direct_compaction &&
2221 (cc->mode == MIGRATE_ASYNC)) {
2222 cc->migrate_pfn = block_end_pfn(
2223 cc->migrate_pfn - 1, cc->order);
2224 /* Draining pcplists is useless in this case */
2225 last_migrated_pfn = 0;
2226 }
2227 }
2228
2229check_drain:
2230 /*
2231 * Has the migration scanner moved away from the previous
2232 * cc->order aligned block where we migrated from? If yes,
2233 * flush the pages that were freed, so that they can merge and
2234 * compact_finished() can detect immediately if allocation
2235 * would succeed.
2236 */
2237 if (cc->order > 0 && last_migrated_pfn) {
2238 int cpu;
2239 unsigned long current_block_start =
2240 block_start_pfn(cc->migrate_pfn, cc->order);
2241
2242 if (last_migrated_pfn < current_block_start) {
2243 cpu = get_cpu();
2244 lru_add_drain_cpu(cpu);
2245 drain_local_pages(cc->zone);
2246 put_cpu();
2247 /* No more flushing until we migrate again */
2248 last_migrated_pfn = 0;
2249 }
2250 }
2251
2252 /* Stop if a page has been captured */
2253 if (capc && capc->page) {
2254 ret = COMPACT_SUCCESS;
2255 break;
2256 }
2257 }
2258
2259out:
2260 /*
2261 * Release free pages and update where the free scanner should restart,
2262 * so we don't leave any returned pages behind in the next attempt.
2263 */
2264 if (cc->nr_freepages > 0) {
2265 unsigned long free_pfn = release_freepages(&cc->freepages);
2266
2267 cc->nr_freepages = 0;
2268 VM_BUG_ON(free_pfn == 0);
2269 /* The cached pfn is always the first in a pageblock */
2270 free_pfn = pageblock_start_pfn(free_pfn);
2271 /*
2272 * Only go back, not forward. The cached pfn might have been
2273 * already reset to zone end in compact_finished()
2274 */
2275 if (free_pfn > cc->zone->compact_cached_free_pfn)
2276 cc->zone->compact_cached_free_pfn = free_pfn;
2277 }
2278
2279 count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned);
2280 count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned);
2281
2282 trace_mm_compaction_end(start_pfn, cc->migrate_pfn,
2283 cc->free_pfn, end_pfn, sync, ret);
2284
2285 return ret;
2286}
2287
2288static enum compact_result compact_zone_order(struct zone *zone, int order,
2289 gfp_t gfp_mask, enum compact_priority prio,
2290 unsigned int alloc_flags, int classzone_idx,
2291 struct page **capture)
2292{
2293 enum compact_result ret;
2294 struct compact_control cc = {
2295 .order = order,
2296 .search_order = order,
2297 .gfp_mask = gfp_mask,
2298 .zone = zone,
2299 .mode = (prio == COMPACT_PRIO_ASYNC) ?
2300 MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT,
2301 .alloc_flags = alloc_flags,
2302 .classzone_idx = classzone_idx,
2303 .direct_compaction = true,
2304 .whole_zone = (prio == MIN_COMPACT_PRIORITY),
2305 .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
2306 .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
2307 };
2308 struct capture_control capc = {
2309 .cc = &cc,
2310 .page = NULL,
2311 };
2312
2313 if (capture)
2314 current->capture_control = &capc;
2315
2316 ret = compact_zone(&cc, &capc);
2317
2318 VM_BUG_ON(!list_empty(&cc.freepages));
2319 VM_BUG_ON(!list_empty(&cc.migratepages));
2320
2321 *capture = capc.page;
2322 current->capture_control = NULL;
2323
2324 return ret;
2325}
2326
2327int sysctl_extfrag_threshold = 500;
2328
2329/**
2330 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
2331 * @gfp_mask: The GFP mask of the current allocation
2332 * @order: The order of the current allocation
2333 * @alloc_flags: The allocation flags of the current allocation
2334 * @ac: The context of current allocation
2335 * @prio: Determines how hard direct compaction should try to succeed
2336 *
2337 * This is the main entry point for direct page compaction.
2338 */
2339enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
2340 unsigned int alloc_flags, const struct alloc_context *ac,
2341 enum compact_priority prio, struct page **capture)
2342{
2343 int may_perform_io = gfp_mask & __GFP_IO;
2344 struct zoneref *z;
2345 struct zone *zone;
2346 enum compact_result rc = COMPACT_SKIPPED;
2347
2348 /*
2349 * Check if the GFP flags allow compaction - GFP_NOIO is really
2350 * tricky context because the migration might require IO
2351 */
2352 if (!may_perform_io)
2353 return COMPACT_SKIPPED;
2354
2355 trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
2356
2357 /* Compact each zone in the list */
2358 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
2359 ac->nodemask) {
2360 enum compact_result status;
2361
2362 if (prio > MIN_COMPACT_PRIORITY
2363 && compaction_deferred(zone, order)) {
2364 rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
2365 continue;
2366 }
2367
2368 status = compact_zone_order(zone, order, gfp_mask, prio,
2369 alloc_flags, ac_classzone_idx(ac), capture);
2370 rc = max(status, rc);
2371
2372 /* The allocation should succeed, stop compacting */
2373 if (status == COMPACT_SUCCESS) {
2374 /*
2375 * We think the allocation will succeed in this zone,
2376 * but it is not certain, hence the false. The caller
2377 * will repeat this with true if allocation indeed
2378 * succeeds in this zone.
2379 */
2380 compaction_defer_reset(zone, order, false);
2381
2382 break;
2383 }
2384
2385 if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
2386 status == COMPACT_PARTIAL_SKIPPED))
2387 /*
2388 * We think that allocation won't succeed in this zone
2389 * so we defer compaction there. If it ends up
2390 * succeeding after all, it will be reset.
2391 */
2392 defer_compaction(zone, order);
2393
2394 /*
2395 * We might have stopped compacting due to need_resched() in
2396 * async compaction, or due to a fatal signal detected. In that
2397 * case do not try further zones
2398 */
2399 if ((prio == COMPACT_PRIO_ASYNC && need_resched())
2400 || fatal_signal_pending(current))
2401 break;
2402 }
2403
2404 return rc;
2405}
2406
2407
2408/* Compact all zones within a node */
2409static void compact_node(int nid)
2410{
2411 pg_data_t *pgdat = NODE_DATA(nid);
2412 int zoneid;
2413 struct zone *zone;
2414 struct compact_control cc = {
2415 .order = -1,
2416 .mode = MIGRATE_SYNC,
2417 .ignore_skip_hint = true,
2418 .whole_zone = true,
2419 .gfp_mask = GFP_KERNEL,
2420 };
2421
2422
2423 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2424
2425 zone = &pgdat->node_zones[zoneid];
2426 if (!populated_zone(zone))
2427 continue;
2428
2429 cc.zone = zone;
2430
2431 compact_zone(&cc, NULL);
2432
2433 VM_BUG_ON(!list_empty(&cc.freepages));
2434 VM_BUG_ON(!list_empty(&cc.migratepages));
2435 }
2436}
2437
2438/* Compact all nodes in the system */
2439static void compact_nodes(void)
2440{
2441 int nid;
2442
2443 /* Flush pending updates to the LRU lists */
2444 lru_add_drain_all();
2445
2446 for_each_online_node(nid)
2447 compact_node(nid);
2448}
2449
2450/* The written value is actually unused, all memory is compacted */
2451int sysctl_compact_memory;
2452
2453/*
2454 * This is the entry point for compacting all nodes via
2455 * /proc/sys/vm/compact_memory
2456 */
2457int sysctl_compaction_handler(struct ctl_table *table, int write,
2458 void __user *buffer, size_t *length, loff_t *ppos)
2459{
2460 if (write)
2461 compact_nodes();
2462
2463 return 0;
2464}
2465
2466#if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
2467static ssize_t sysfs_compact_node(struct device *dev,
2468 struct device_attribute *attr,
2469 const char *buf, size_t count)
2470{
2471 int nid = dev->id;
2472
2473 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
2474 /* Flush pending updates to the LRU lists */
2475 lru_add_drain_all();
2476
2477 compact_node(nid);
2478 }
2479
2480 return count;
2481}
2482static DEVICE_ATTR(compact, 0200, NULL, sysfs_compact_node);
2483
2484int compaction_register_node(struct node *node)
2485{
2486 return device_create_file(&node->dev, &dev_attr_compact);
2487}
2488
2489void compaction_unregister_node(struct node *node)
2490{
2491 return device_remove_file(&node->dev, &dev_attr_compact);
2492}
2493#endif /* CONFIG_SYSFS && CONFIG_NUMA */
2494
2495static inline bool kcompactd_work_requested(pg_data_t *pgdat)
2496{
2497 return pgdat->kcompactd_max_order > 0 || kthread_should_stop();
2498}
2499
2500static bool kcompactd_node_suitable(pg_data_t *pgdat)
2501{
2502 int zoneid;
2503 struct zone *zone;
2504 enum zone_type classzone_idx = pgdat->kcompactd_classzone_idx;
2505
2506 for (zoneid = 0; zoneid <= classzone_idx; zoneid++) {
2507 zone = &pgdat->node_zones[zoneid];
2508
2509 if (!populated_zone(zone))
2510 continue;
2511
2512 if (compaction_suitable(zone, pgdat->kcompactd_max_order, 0,
2513 classzone_idx) == COMPACT_CONTINUE)
2514 return true;
2515 }
2516
2517 return false;
2518}
2519
2520static void kcompactd_do_work(pg_data_t *pgdat)
2521{
2522 /*
2523 * With no special task, compact all zones so that a page of requested
2524 * order is allocatable.
2525 */
2526 int zoneid;
2527 struct zone *zone;
2528 struct compact_control cc = {
2529 .order = pgdat->kcompactd_max_order,
2530 .search_order = pgdat->kcompactd_max_order,
2531 .classzone_idx = pgdat->kcompactd_classzone_idx,
2532 .mode = MIGRATE_SYNC_LIGHT,
2533 .ignore_skip_hint = false,
2534 .gfp_mask = GFP_KERNEL,
2535 };
2536 trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
2537 cc.classzone_idx);
2538 count_compact_event(KCOMPACTD_WAKE);
2539
2540 for (zoneid = 0; zoneid <= cc.classzone_idx; zoneid++) {
2541 int status;
2542
2543 zone = &pgdat->node_zones[zoneid];
2544 if (!populated_zone(zone))
2545 continue;
2546
2547 if (compaction_deferred(zone, cc.order))
2548 continue;
2549
2550 if (compaction_suitable(zone, cc.order, 0, zoneid) !=
2551 COMPACT_CONTINUE)
2552 continue;
2553
2554 if (kthread_should_stop())
2555 return;
2556
2557 cc.zone = zone;
2558 status = compact_zone(&cc, NULL);
2559
2560 if (status == COMPACT_SUCCESS) {
2561 compaction_defer_reset(zone, cc.order, false);
2562 } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
2563 /*
2564 * Buddy pages may become stranded on pcps that could
2565 * otherwise coalesce on the zone's free area for
2566 * order >= cc.order. This is ratelimited by the
2567 * upcoming deferral.
2568 */
2569 drain_all_pages(zone);
2570
2571 /*
2572 * We use sync migration mode here, so we defer like
2573 * sync direct compaction does.
2574 */
2575 defer_compaction(zone, cc.order);
2576 }
2577
2578 count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
2579 cc.total_migrate_scanned);
2580 count_compact_events(KCOMPACTD_FREE_SCANNED,
2581 cc.total_free_scanned);
2582
2583 VM_BUG_ON(!list_empty(&cc.freepages));
2584 VM_BUG_ON(!list_empty(&cc.migratepages));
2585 }
2586
2587 /*
2588 * Regardless of success, we are done until woken up next. But remember
2589 * the requested order/classzone_idx in case it was higher/tighter than
2590 * our current ones
2591 */
2592 if (pgdat->kcompactd_max_order <= cc.order)
2593 pgdat->kcompactd_max_order = 0;
2594 if (pgdat->kcompactd_classzone_idx >= cc.classzone_idx)
2595 pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
2596}
2597
2598void wakeup_kcompactd(pg_data_t *pgdat, int order, int classzone_idx)
2599{
2600 if (!order)
2601 return;
2602
2603 if (pgdat->kcompactd_max_order < order)
2604 pgdat->kcompactd_max_order = order;
2605
2606 if (pgdat->kcompactd_classzone_idx > classzone_idx)
2607 pgdat->kcompactd_classzone_idx = classzone_idx;
2608
2609 /*
2610 * Pairs with implicit barrier in wait_event_freezable()
2611 * such that wakeups are not missed.
2612 */
2613 if (!wq_has_sleeper(&pgdat->kcompactd_wait))
2614 return;
2615
2616 if (!kcompactd_node_suitable(pgdat))
2617 return;
2618
2619 trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
2620 classzone_idx);
2621 wake_up_interruptible(&pgdat->kcompactd_wait);
2622}
2623
2624/*
2625 * The background compaction daemon, started as a kernel thread
2626 * from the init process.
2627 */
2628static int kcompactd(void *p)
2629{
2630 pg_data_t *pgdat = (pg_data_t*)p;
2631 struct task_struct *tsk = current;
2632
2633 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2634
2635 if (!cpumask_empty(cpumask))
2636 set_cpus_allowed_ptr(tsk, cpumask);
2637
2638 set_freezable();
2639
2640 pgdat->kcompactd_max_order = 0;
2641 pgdat->kcompactd_classzone_idx = pgdat->nr_zones - 1;
2642
2643 while (!kthread_should_stop()) {
2644 unsigned long pflags;
2645
2646 trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
2647 wait_event_freezable(pgdat->kcompactd_wait,
2648 kcompactd_work_requested(pgdat));
2649
2650 psi_memstall_enter(&pflags);
2651 kcompactd_do_work(pgdat);
2652 psi_memstall_leave(&pflags);
2653 }
2654
2655 return 0;
2656}
2657
2658/*
2659 * This kcompactd start function will be called by init and node-hot-add.
2660 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
2661 */
2662int kcompactd_run(int nid)
2663{
2664 pg_data_t *pgdat = NODE_DATA(nid);
2665 int ret = 0;
2666
2667 if (pgdat->kcompactd)
2668 return 0;
2669
2670 pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
2671 if (IS_ERR(pgdat->kcompactd)) {
2672 pr_err("Failed to start kcompactd on node %d\n", nid);
2673 ret = PTR_ERR(pgdat->kcompactd);
2674 pgdat->kcompactd = NULL;
2675 }
2676 return ret;
2677}
2678
2679/*
2680 * Called by memory hotplug when all memory in a node is offlined. Caller must
2681 * hold mem_hotplug_begin/end().
2682 */
2683void kcompactd_stop(int nid)
2684{
2685 struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
2686
2687 if (kcompactd) {
2688 kthread_stop(kcompactd);
2689 NODE_DATA(nid)->kcompactd = NULL;
2690 }
2691}
2692
2693/*
2694 * It's optimal to keep kcompactd on the same CPUs as their memory, but
2695 * not required for correctness. So if the last cpu in a node goes
2696 * away, we get changed to run anywhere: as the first one comes back,
2697 * restore their cpu bindings.
2698 */
2699static int kcompactd_cpu_online(unsigned int cpu)
2700{
2701 int nid;
2702
2703 for_each_node_state(nid, N_MEMORY) {
2704 pg_data_t *pgdat = NODE_DATA(nid);
2705 const struct cpumask *mask;
2706
2707 mask = cpumask_of_node(pgdat->node_id);
2708
2709 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
2710 /* One of our CPUs online: restore mask */
2711 set_cpus_allowed_ptr(pgdat->kcompactd, mask);
2712 }
2713 return 0;
2714}
2715
2716static int __init kcompactd_init(void)
2717{
2718 int nid;
2719 int ret;
2720
2721 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
2722 "mm/compaction:online",
2723 kcompactd_cpu_online, NULL);
2724 if (ret < 0) {
2725 pr_err("kcompactd: failed to register hotplug callbacks.\n");
2726 return ret;
2727 }
2728
2729 for_each_node_state(nid, N_MEMORY)
2730 kcompactd_run(nid);
2731 return 0;
2732}
2733subsys_initcall(kcompactd_init)
2734
2735#endif /* CONFIG_COMPACTION */