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