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 <linux/cpuset.h>
27#include "internal.h"
28
29#ifdef CONFIG_COMPACTION
30/*
31 * Fragmentation score check interval for proactive compaction purposes.
32 */
33#define HPAGE_FRAG_CHECK_INTERVAL_MSEC (500)
34
35static inline void count_compact_event(enum vm_event_item item)
36{
37 count_vm_event(item);
38}
39
40static inline void count_compact_events(enum vm_event_item item, long delta)
41{
42 count_vm_events(item, delta);
43}
44
45/*
46 * order == -1 is expected when compacting proactively via
47 * 1. /proc/sys/vm/compact_memory
48 * 2. /sys/devices/system/node/nodex/compact
49 * 3. /proc/sys/vm/compaction_proactiveness
50 */
51static inline bool is_via_compact_memory(int order)
52{
53 return order == -1;
54}
55
56#else
57#define count_compact_event(item) do { } while (0)
58#define count_compact_events(item, delta) do { } while (0)
59static inline bool is_via_compact_memory(int order) { return false; }
60#endif
61
62#if defined CONFIG_COMPACTION || defined CONFIG_CMA
63
64#define CREATE_TRACE_POINTS
65#include <trace/events/compaction.h>
66
67#define block_start_pfn(pfn, order) round_down(pfn, 1UL << (order))
68#define block_end_pfn(pfn, order) ALIGN((pfn) + 1, 1UL << (order))
69
70/*
71 * Page order with-respect-to which proactive compaction
72 * calculates external fragmentation, which is used as
73 * the "fragmentation score" of a node/zone.
74 */
75#if defined CONFIG_TRANSPARENT_HUGEPAGE
76#define COMPACTION_HPAGE_ORDER HPAGE_PMD_ORDER
77#elif defined CONFIG_HUGETLBFS
78#define COMPACTION_HPAGE_ORDER HUGETLB_PAGE_ORDER
79#else
80#define COMPACTION_HPAGE_ORDER (PMD_SHIFT - PAGE_SHIFT)
81#endif
82
83static struct page *mark_allocated_noprof(struct page *page, unsigned int order, gfp_t gfp_flags)
84{
85 post_alloc_hook(page, order, __GFP_MOVABLE);
86 return page;
87}
88#define mark_allocated(...) alloc_hooks(mark_allocated_noprof(__VA_ARGS__))
89
90static unsigned long release_free_list(struct list_head *freepages)
91{
92 int order;
93 unsigned long high_pfn = 0;
94
95 for (order = 0; order < NR_PAGE_ORDERS; order++) {
96 struct page *page, *next;
97
98 list_for_each_entry_safe(page, next, &freepages[order], lru) {
99 unsigned long pfn = page_to_pfn(page);
100
101 list_del(&page->lru);
102 /*
103 * Convert free pages into post allocation pages, so
104 * that we can free them via __free_page.
105 */
106 mark_allocated(page, order, __GFP_MOVABLE);
107 __free_pages(page, order);
108 if (pfn > high_pfn)
109 high_pfn = pfn;
110 }
111 }
112 return high_pfn;
113}
114
115#ifdef CONFIG_COMPACTION
116bool PageMovable(struct page *page)
117{
118 const struct movable_operations *mops;
119
120 VM_BUG_ON_PAGE(!PageLocked(page), page);
121 if (!__PageMovable(page))
122 return false;
123
124 mops = page_movable_ops(page);
125 if (mops)
126 return true;
127
128 return false;
129}
130
131void __SetPageMovable(struct page *page, const struct movable_operations *mops)
132{
133 VM_BUG_ON_PAGE(!PageLocked(page), page);
134 VM_BUG_ON_PAGE((unsigned long)mops & PAGE_MAPPING_MOVABLE, page);
135 page->mapping = (void *)((unsigned long)mops | PAGE_MAPPING_MOVABLE);
136}
137EXPORT_SYMBOL(__SetPageMovable);
138
139void __ClearPageMovable(struct page *page)
140{
141 VM_BUG_ON_PAGE(!PageMovable(page), page);
142 /*
143 * This page still has the type of a movable page, but it's
144 * actually not movable any more.
145 */
146 page->mapping = (void *)PAGE_MAPPING_MOVABLE;
147}
148EXPORT_SYMBOL(__ClearPageMovable);
149
150/* Do not skip compaction more than 64 times */
151#define COMPACT_MAX_DEFER_SHIFT 6
152
153/*
154 * Compaction is deferred when compaction fails to result in a page
155 * allocation success. 1 << compact_defer_shift, compactions are skipped up
156 * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
157 */
158static void defer_compaction(struct zone *zone, int order)
159{
160 zone->compact_considered = 0;
161 zone->compact_defer_shift++;
162
163 if (order < zone->compact_order_failed)
164 zone->compact_order_failed = order;
165
166 if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
167 zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
168
169 trace_mm_compaction_defer_compaction(zone, order);
170}
171
172/* Returns true if compaction should be skipped this time */
173static bool compaction_deferred(struct zone *zone, int order)
174{
175 unsigned long defer_limit = 1UL << zone->compact_defer_shift;
176
177 if (order < zone->compact_order_failed)
178 return false;
179
180 /* Avoid possible overflow */
181 if (++zone->compact_considered >= defer_limit) {
182 zone->compact_considered = defer_limit;
183 return false;
184 }
185
186 trace_mm_compaction_deferred(zone, order);
187
188 return true;
189}
190
191/*
192 * Update defer tracking counters after successful compaction of given order,
193 * which means an allocation either succeeded (alloc_success == true) or is
194 * expected to succeed.
195 */
196void compaction_defer_reset(struct zone *zone, int order,
197 bool alloc_success)
198{
199 if (alloc_success) {
200 zone->compact_considered = 0;
201 zone->compact_defer_shift = 0;
202 }
203 if (order >= zone->compact_order_failed)
204 zone->compact_order_failed = order + 1;
205
206 trace_mm_compaction_defer_reset(zone, order);
207}
208
209/* Returns true if restarting compaction after many failures */
210static bool compaction_restarting(struct zone *zone, int order)
211{
212 if (order < zone->compact_order_failed)
213 return false;
214
215 return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
216 zone->compact_considered >= 1UL << zone->compact_defer_shift;
217}
218
219/* Returns true if the pageblock should be scanned for pages to isolate. */
220static inline bool isolation_suitable(struct compact_control *cc,
221 struct page *page)
222{
223 if (cc->ignore_skip_hint)
224 return true;
225
226 return !get_pageblock_skip(page);
227}
228
229static void reset_cached_positions(struct zone *zone)
230{
231 zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
232 zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
233 zone->compact_cached_free_pfn =
234 pageblock_start_pfn(zone_end_pfn(zone) - 1);
235}
236
237#ifdef CONFIG_SPARSEMEM
238/*
239 * If the PFN falls into an offline section, return the start PFN of the
240 * next online section. If the PFN falls into an online section or if
241 * there is no next online section, return 0.
242 */
243static unsigned long skip_offline_sections(unsigned long start_pfn)
244{
245 unsigned long start_nr = pfn_to_section_nr(start_pfn);
246
247 if (online_section_nr(start_nr))
248 return 0;
249
250 while (++start_nr <= __highest_present_section_nr) {
251 if (online_section_nr(start_nr))
252 return section_nr_to_pfn(start_nr);
253 }
254
255 return 0;
256}
257
258/*
259 * If the PFN falls into an offline section, return the end PFN of the
260 * next online section in reverse. If the PFN falls into an online section
261 * or if there is no next online section in reverse, return 0.
262 */
263static unsigned long skip_offline_sections_reverse(unsigned long start_pfn)
264{
265 unsigned long start_nr = pfn_to_section_nr(start_pfn);
266
267 if (!start_nr || online_section_nr(start_nr))
268 return 0;
269
270 while (start_nr-- > 0) {
271 if (online_section_nr(start_nr))
272 return section_nr_to_pfn(start_nr) + PAGES_PER_SECTION;
273 }
274
275 return 0;
276}
277#else
278static unsigned long skip_offline_sections(unsigned long start_pfn)
279{
280 return 0;
281}
282
283static unsigned long skip_offline_sections_reverse(unsigned long start_pfn)
284{
285 return 0;
286}
287#endif
288
289/*
290 * Compound pages of >= pageblock_order should consistently be skipped until
291 * released. It is always pointless to compact pages of such order (if they are
292 * migratable), and the pageblocks they occupy cannot contain any free pages.
293 */
294static bool pageblock_skip_persistent(struct page *page)
295{
296 if (!PageCompound(page))
297 return false;
298
299 page = compound_head(page);
300
301 if (compound_order(page) >= pageblock_order)
302 return true;
303
304 return false;
305}
306
307static bool
308__reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source,
309 bool check_target)
310{
311 struct page *page = pfn_to_online_page(pfn);
312 struct page *block_page;
313 struct page *end_page;
314 unsigned long block_pfn;
315
316 if (!page)
317 return false;
318 if (zone != page_zone(page))
319 return false;
320 if (pageblock_skip_persistent(page))
321 return false;
322
323 /*
324 * If skip is already cleared do no further checking once the
325 * restart points have been set.
326 */
327 if (check_source && check_target && !get_pageblock_skip(page))
328 return true;
329
330 /*
331 * If clearing skip for the target scanner, do not select a
332 * non-movable pageblock as the starting point.
333 */
334 if (!check_source && check_target &&
335 get_pageblock_migratetype(page) != MIGRATE_MOVABLE)
336 return false;
337
338 /* Ensure the start of the pageblock or zone is online and valid */
339 block_pfn = pageblock_start_pfn(pfn);
340 block_pfn = max(block_pfn, zone->zone_start_pfn);
341 block_page = pfn_to_online_page(block_pfn);
342 if (block_page) {
343 page = block_page;
344 pfn = block_pfn;
345 }
346
347 /* Ensure the end of the pageblock or zone is online and valid */
348 block_pfn = pageblock_end_pfn(pfn) - 1;
349 block_pfn = min(block_pfn, zone_end_pfn(zone) - 1);
350 end_page = pfn_to_online_page(block_pfn);
351 if (!end_page)
352 return false;
353
354 /*
355 * Only clear the hint if a sample indicates there is either a
356 * free page or an LRU page in the block. One or other condition
357 * is necessary for the block to be a migration source/target.
358 */
359 do {
360 if (check_source && PageLRU(page)) {
361 clear_pageblock_skip(page);
362 return true;
363 }
364
365 if (check_target && PageBuddy(page)) {
366 clear_pageblock_skip(page);
367 return true;
368 }
369
370 page += (1 << PAGE_ALLOC_COSTLY_ORDER);
371 } while (page <= end_page);
372
373 return false;
374}
375
376/*
377 * This function is called to clear all cached information on pageblocks that
378 * should be skipped for page isolation when the migrate and free page scanner
379 * meet.
380 */
381static void __reset_isolation_suitable(struct zone *zone)
382{
383 unsigned long migrate_pfn = zone->zone_start_pfn;
384 unsigned long free_pfn = zone_end_pfn(zone) - 1;
385 unsigned long reset_migrate = free_pfn;
386 unsigned long reset_free = migrate_pfn;
387 bool source_set = false;
388 bool free_set = false;
389
390 /* Only flush if a full compaction finished recently */
391 if (!zone->compact_blockskip_flush)
392 return;
393
394 zone->compact_blockskip_flush = false;
395
396 /*
397 * Walk the zone and update pageblock skip information. Source looks
398 * for PageLRU while target looks for PageBuddy. When the scanner
399 * is found, both PageBuddy and PageLRU are checked as the pageblock
400 * is suitable as both source and target.
401 */
402 for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages,
403 free_pfn -= pageblock_nr_pages) {
404 cond_resched();
405
406 /* Update the migrate PFN */
407 if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) &&
408 migrate_pfn < reset_migrate) {
409 source_set = true;
410 reset_migrate = migrate_pfn;
411 zone->compact_init_migrate_pfn = reset_migrate;
412 zone->compact_cached_migrate_pfn[0] = reset_migrate;
413 zone->compact_cached_migrate_pfn[1] = reset_migrate;
414 }
415
416 /* Update the free PFN */
417 if (__reset_isolation_pfn(zone, free_pfn, free_set, true) &&
418 free_pfn > reset_free) {
419 free_set = true;
420 reset_free = free_pfn;
421 zone->compact_init_free_pfn = reset_free;
422 zone->compact_cached_free_pfn = reset_free;
423 }
424 }
425
426 /* Leave no distance if no suitable block was reset */
427 if (reset_migrate >= reset_free) {
428 zone->compact_cached_migrate_pfn[0] = migrate_pfn;
429 zone->compact_cached_migrate_pfn[1] = migrate_pfn;
430 zone->compact_cached_free_pfn = free_pfn;
431 }
432}
433
434void reset_isolation_suitable(pg_data_t *pgdat)
435{
436 int zoneid;
437
438 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
439 struct zone *zone = &pgdat->node_zones[zoneid];
440 if (!populated_zone(zone))
441 continue;
442
443 __reset_isolation_suitable(zone);
444 }
445}
446
447/*
448 * Sets the pageblock skip bit if it was clear. Note that this is a hint as
449 * locks are not required for read/writers. Returns true if it was already set.
450 */
451static bool test_and_set_skip(struct compact_control *cc, struct page *page)
452{
453 bool skip;
454
455 /* Do not update if skip hint is being ignored */
456 if (cc->ignore_skip_hint)
457 return false;
458
459 skip = get_pageblock_skip(page);
460 if (!skip && !cc->no_set_skip_hint)
461 set_pageblock_skip(page);
462
463 return skip;
464}
465
466static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
467{
468 struct zone *zone = cc->zone;
469
470 /* Set for isolation rather than compaction */
471 if (cc->no_set_skip_hint)
472 return;
473
474 pfn = pageblock_end_pfn(pfn);
475
476 /* Update where async and sync compaction should restart */
477 if (pfn > zone->compact_cached_migrate_pfn[0])
478 zone->compact_cached_migrate_pfn[0] = pfn;
479 if (cc->mode != MIGRATE_ASYNC &&
480 pfn > zone->compact_cached_migrate_pfn[1])
481 zone->compact_cached_migrate_pfn[1] = pfn;
482}
483
484/*
485 * If no pages were isolated then mark this pageblock to be skipped in the
486 * future. The information is later cleared by __reset_isolation_suitable().
487 */
488static void update_pageblock_skip(struct compact_control *cc,
489 struct page *page, unsigned long pfn)
490{
491 struct zone *zone = cc->zone;
492
493 if (cc->no_set_skip_hint)
494 return;
495
496 set_pageblock_skip(page);
497
498 if (pfn < zone->compact_cached_free_pfn)
499 zone->compact_cached_free_pfn = pfn;
500}
501#else
502static inline bool isolation_suitable(struct compact_control *cc,
503 struct page *page)
504{
505 return true;
506}
507
508static inline bool pageblock_skip_persistent(struct page *page)
509{
510 return false;
511}
512
513static inline void update_pageblock_skip(struct compact_control *cc,
514 struct page *page, unsigned long pfn)
515{
516}
517
518static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
519{
520}
521
522static bool test_and_set_skip(struct compact_control *cc, struct page *page)
523{
524 return false;
525}
526#endif /* CONFIG_COMPACTION */
527
528/*
529 * Compaction requires the taking of some coarse locks that are potentially
530 * very heavily contended. For async compaction, trylock and record if the
531 * lock is contended. The lock will still be acquired but compaction will
532 * abort when the current block is finished regardless of success rate.
533 * Sync compaction acquires the lock.
534 *
535 * Always returns true which makes it easier to track lock state in callers.
536 */
537static bool compact_lock_irqsave(spinlock_t *lock, unsigned long *flags,
538 struct compact_control *cc)
539 __acquires(lock)
540{
541 /* Track if the lock is contended in async mode */
542 if (cc->mode == MIGRATE_ASYNC && !cc->contended) {
543 if (spin_trylock_irqsave(lock, *flags))
544 return true;
545
546 cc->contended = true;
547 }
548
549 spin_lock_irqsave(lock, *flags);
550 return true;
551}
552
553/*
554 * Compaction requires the taking of some coarse locks that are potentially
555 * very heavily contended. The lock should be periodically unlocked to avoid
556 * having disabled IRQs for a long time, even when there is nobody waiting on
557 * the lock. It might also be that allowing the IRQs will result in
558 * need_resched() becoming true. If scheduling is needed, compaction schedules.
559 * Either compaction type will also abort if a fatal signal is pending.
560 * In either case if the lock was locked, it is dropped and not regained.
561 *
562 * Returns true if compaction should abort due to fatal signal pending.
563 * Returns false when compaction can continue.
564 */
565static bool compact_unlock_should_abort(spinlock_t *lock,
566 unsigned long flags, bool *locked, struct compact_control *cc)
567{
568 if (*locked) {
569 spin_unlock_irqrestore(lock, flags);
570 *locked = false;
571 }
572
573 if (fatal_signal_pending(current)) {
574 cc->contended = true;
575 return true;
576 }
577
578 cond_resched();
579
580 return false;
581}
582
583/*
584 * Isolate free pages onto a private freelist. If @strict is true, will abort
585 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
586 * (even though it may still end up isolating some pages).
587 */
588static unsigned long isolate_freepages_block(struct compact_control *cc,
589 unsigned long *start_pfn,
590 unsigned long end_pfn,
591 struct list_head *freelist,
592 unsigned int stride,
593 bool strict)
594{
595 int nr_scanned = 0, total_isolated = 0;
596 struct page *page;
597 unsigned long flags = 0;
598 bool locked = false;
599 unsigned long blockpfn = *start_pfn;
600 unsigned int order;
601
602 /* Strict mode is for isolation, speed is secondary */
603 if (strict)
604 stride = 1;
605
606 page = pfn_to_page(blockpfn);
607
608 /* Isolate free pages. */
609 for (; blockpfn < end_pfn; blockpfn += stride, page += stride) {
610 int isolated;
611
612 /*
613 * Periodically drop the lock (if held) regardless of its
614 * contention, to give chance to IRQs. Abort if fatal signal
615 * pending.
616 */
617 if (!(blockpfn % COMPACT_CLUSTER_MAX)
618 && compact_unlock_should_abort(&cc->zone->lock, flags,
619 &locked, cc))
620 break;
621
622 nr_scanned++;
623
624 /*
625 * For compound pages such as THP and hugetlbfs, we can save
626 * potentially a lot of iterations if we skip them at once.
627 * The check is racy, but we can consider only valid values
628 * and the only danger is skipping too much.
629 */
630 if (PageCompound(page)) {
631 const unsigned int order = compound_order(page);
632
633 if ((order <= MAX_PAGE_ORDER) &&
634 (blockpfn + (1UL << order) <= end_pfn)) {
635 blockpfn += (1UL << order) - 1;
636 page += (1UL << order) - 1;
637 nr_scanned += (1UL << order) - 1;
638 }
639
640 goto isolate_fail;
641 }
642
643 if (!PageBuddy(page))
644 goto isolate_fail;
645
646 /* If we already hold the lock, we can skip some rechecking. */
647 if (!locked) {
648 locked = compact_lock_irqsave(&cc->zone->lock,
649 &flags, cc);
650
651 /* Recheck this is a buddy page under lock */
652 if (!PageBuddy(page))
653 goto isolate_fail;
654 }
655
656 /* Found a free page, will break it into order-0 pages */
657 order = buddy_order(page);
658 isolated = __isolate_free_page(page, order);
659 if (!isolated)
660 break;
661 set_page_private(page, order);
662
663 nr_scanned += isolated - 1;
664 total_isolated += isolated;
665 cc->nr_freepages += isolated;
666 list_add_tail(&page->lru, &freelist[order]);
667
668 if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
669 blockpfn += isolated;
670 break;
671 }
672 /* Advance to the end of split page */
673 blockpfn += isolated - 1;
674 page += isolated - 1;
675 continue;
676
677isolate_fail:
678 if (strict)
679 break;
680
681 }
682
683 if (locked)
684 spin_unlock_irqrestore(&cc->zone->lock, flags);
685
686 /*
687 * Be careful to not go outside of the pageblock.
688 */
689 if (unlikely(blockpfn > end_pfn))
690 blockpfn = end_pfn;
691
692 trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
693 nr_scanned, total_isolated);
694
695 /* Record how far we have got within the block */
696 *start_pfn = blockpfn;
697
698 /*
699 * If strict isolation is requested by CMA then check that all the
700 * pages requested were isolated. If there were any failures, 0 is
701 * returned and CMA will fail.
702 */
703 if (strict && blockpfn < end_pfn)
704 total_isolated = 0;
705
706 cc->total_free_scanned += nr_scanned;
707 if (total_isolated)
708 count_compact_events(COMPACTISOLATED, total_isolated);
709 return total_isolated;
710}
711
712/**
713 * isolate_freepages_range() - isolate free pages.
714 * @cc: Compaction control structure.
715 * @start_pfn: The first PFN to start isolating.
716 * @end_pfn: The one-past-last PFN.
717 *
718 * Non-free pages, invalid PFNs, or zone boundaries within the
719 * [start_pfn, end_pfn) range are considered errors, cause function to
720 * undo its actions and return zero. cc->freepages[] are empty.
721 *
722 * Otherwise, function returns one-past-the-last PFN of isolated page
723 * (which may be greater then end_pfn if end fell in a middle of
724 * a free page). cc->freepages[] contain free pages isolated.
725 */
726unsigned long
727isolate_freepages_range(struct compact_control *cc,
728 unsigned long start_pfn, unsigned long end_pfn)
729{
730 unsigned long isolated, pfn, block_start_pfn, block_end_pfn;
731 int order;
732
733 for (order = 0; order < NR_PAGE_ORDERS; order++)
734 INIT_LIST_HEAD(&cc->freepages[order]);
735
736 pfn = start_pfn;
737 block_start_pfn = pageblock_start_pfn(pfn);
738 if (block_start_pfn < cc->zone->zone_start_pfn)
739 block_start_pfn = cc->zone->zone_start_pfn;
740 block_end_pfn = pageblock_end_pfn(pfn);
741
742 for (; pfn < end_pfn; pfn += isolated,
743 block_start_pfn = block_end_pfn,
744 block_end_pfn += pageblock_nr_pages) {
745 /* Protect pfn from changing by isolate_freepages_block */
746 unsigned long isolate_start_pfn = pfn;
747
748 /*
749 * pfn could pass the block_end_pfn if isolated freepage
750 * is more than pageblock order. In this case, we adjust
751 * scanning range to right one.
752 */
753 if (pfn >= block_end_pfn) {
754 block_start_pfn = pageblock_start_pfn(pfn);
755 block_end_pfn = pageblock_end_pfn(pfn);
756 }
757
758 block_end_pfn = min(block_end_pfn, end_pfn);
759
760 if (!pageblock_pfn_to_page(block_start_pfn,
761 block_end_pfn, cc->zone))
762 break;
763
764 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
765 block_end_pfn, cc->freepages, 0, true);
766
767 /*
768 * In strict mode, isolate_freepages_block() returns 0 if
769 * there are any holes in the block (ie. invalid PFNs or
770 * non-free pages).
771 */
772 if (!isolated)
773 break;
774
775 /*
776 * If we managed to isolate pages, it is always (1 << n) *
777 * pageblock_nr_pages for some non-negative n. (Max order
778 * page may span two pageblocks).
779 */
780 }
781
782 if (pfn < end_pfn) {
783 /* Loop terminated early, cleanup. */
784 release_free_list(cc->freepages);
785 return 0;
786 }
787
788 /* We don't use freelists for anything. */
789 return pfn;
790}
791
792/* Similar to reclaim, but different enough that they don't share logic */
793static bool too_many_isolated(struct compact_control *cc)
794{
795 pg_data_t *pgdat = cc->zone->zone_pgdat;
796 bool too_many;
797
798 unsigned long active, inactive, isolated;
799
800 inactive = node_page_state(pgdat, NR_INACTIVE_FILE) +
801 node_page_state(pgdat, NR_INACTIVE_ANON);
802 active = node_page_state(pgdat, NR_ACTIVE_FILE) +
803 node_page_state(pgdat, NR_ACTIVE_ANON);
804 isolated = node_page_state(pgdat, NR_ISOLATED_FILE) +
805 node_page_state(pgdat, NR_ISOLATED_ANON);
806
807 /*
808 * Allow GFP_NOFS to isolate past the limit set for regular
809 * compaction runs. This prevents an ABBA deadlock when other
810 * compactors have already isolated to the limit, but are
811 * blocked on filesystem locks held by the GFP_NOFS thread.
812 */
813 if (cc->gfp_mask & __GFP_FS) {
814 inactive >>= 3;
815 active >>= 3;
816 }
817
818 too_many = isolated > (inactive + active) / 2;
819 if (!too_many)
820 wake_throttle_isolated(pgdat);
821
822 return too_many;
823}
824
825/**
826 * skip_isolation_on_order() - determine when to skip folio isolation based on
827 * folio order and compaction target order
828 * @order: to-be-isolated folio order
829 * @target_order: compaction target order
830 *
831 * This avoids unnecessary folio isolations during compaction.
832 */
833static bool skip_isolation_on_order(int order, int target_order)
834{
835 /*
836 * Unless we are performing global compaction (i.e.,
837 * is_via_compact_memory), skip any folios that are larger than the
838 * target order: we wouldn't be here if we'd have a free folio with
839 * the desired target_order, so migrating this folio would likely fail
840 * later.
841 */
842 if (!is_via_compact_memory(target_order) && order >= target_order)
843 return true;
844 /*
845 * We limit memory compaction to pageblocks and won't try
846 * creating free blocks of memory that are larger than that.
847 */
848 return order >= pageblock_order;
849}
850
851/**
852 * isolate_migratepages_block() - isolate all migrate-able pages within
853 * a single pageblock
854 * @cc: Compaction control structure.
855 * @low_pfn: The first PFN to isolate
856 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
857 * @mode: Isolation mode to be used.
858 *
859 * Isolate all pages that can be migrated from the range specified by
860 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
861 * Returns errno, like -EAGAIN or -EINTR in case e.g signal pending or congestion,
862 * -ENOMEM in case we could not allocate a page, or 0.
863 * cc->migrate_pfn will contain the next pfn to scan.
864 *
865 * The pages are isolated on cc->migratepages list (not required to be empty),
866 * and cc->nr_migratepages is updated accordingly.
867 */
868static int
869isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
870 unsigned long end_pfn, isolate_mode_t mode)
871{
872 pg_data_t *pgdat = cc->zone->zone_pgdat;
873 unsigned long nr_scanned = 0, nr_isolated = 0;
874 struct lruvec *lruvec;
875 unsigned long flags = 0;
876 struct lruvec *locked = NULL;
877 struct folio *folio = NULL;
878 struct page *page = NULL, *valid_page = NULL;
879 struct address_space *mapping;
880 unsigned long start_pfn = low_pfn;
881 bool skip_on_failure = false;
882 unsigned long next_skip_pfn = 0;
883 bool skip_updated = false;
884 int ret = 0;
885
886 cc->migrate_pfn = low_pfn;
887
888 /*
889 * Ensure that there are not too many pages isolated from the LRU
890 * list by either parallel reclaimers or compaction. If there are,
891 * delay for some time until fewer pages are isolated
892 */
893 while (unlikely(too_many_isolated(cc))) {
894 /* stop isolation if there are still pages not migrated */
895 if (cc->nr_migratepages)
896 return -EAGAIN;
897
898 /* async migration should just abort */
899 if (cc->mode == MIGRATE_ASYNC)
900 return -EAGAIN;
901
902 reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED);
903
904 if (fatal_signal_pending(current))
905 return -EINTR;
906 }
907
908 cond_resched();
909
910 if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
911 skip_on_failure = true;
912 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
913 }
914
915 /* Time to isolate some pages for migration */
916 for (; low_pfn < end_pfn; low_pfn++) {
917 bool is_dirty, is_unevictable;
918
919 if (skip_on_failure && low_pfn >= next_skip_pfn) {
920 /*
921 * We have isolated all migration candidates in the
922 * previous order-aligned block, and did not skip it due
923 * to failure. We should migrate the pages now and
924 * hopefully succeed compaction.
925 */
926 if (nr_isolated)
927 break;
928
929 /*
930 * We failed to isolate in the previous order-aligned
931 * block. Set the new boundary to the end of the
932 * current block. Note we can't simply increase
933 * next_skip_pfn by 1 << order, as low_pfn might have
934 * been incremented by a higher number due to skipping
935 * a compound or a high-order buddy page in the
936 * previous loop iteration.
937 */
938 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
939 }
940
941 /*
942 * Periodically drop the lock (if held) regardless of its
943 * contention, to give chance to IRQs. Abort completely if
944 * a fatal signal is pending.
945 */
946 if (!(low_pfn % COMPACT_CLUSTER_MAX)) {
947 if (locked) {
948 unlock_page_lruvec_irqrestore(locked, flags);
949 locked = NULL;
950 }
951
952 if (fatal_signal_pending(current)) {
953 cc->contended = true;
954 ret = -EINTR;
955
956 goto fatal_pending;
957 }
958
959 cond_resched();
960 }
961
962 nr_scanned++;
963
964 page = pfn_to_page(low_pfn);
965
966 /*
967 * Check if the pageblock has already been marked skipped.
968 * Only the first PFN is checked as the caller isolates
969 * COMPACT_CLUSTER_MAX at a time so the second call must
970 * not falsely conclude that the block should be skipped.
971 */
972 if (!valid_page && (pageblock_aligned(low_pfn) ||
973 low_pfn == cc->zone->zone_start_pfn)) {
974 if (!isolation_suitable(cc, page)) {
975 low_pfn = end_pfn;
976 folio = NULL;
977 goto isolate_abort;
978 }
979 valid_page = page;
980 }
981
982 if (PageHuge(page)) {
983 /*
984 * skip hugetlbfs if we are not compacting for pages
985 * bigger than its order. THPs and other compound pages
986 * are handled below.
987 */
988 if (!cc->alloc_contig) {
989 const unsigned int order = compound_order(page);
990
991 if (order <= MAX_PAGE_ORDER) {
992 low_pfn += (1UL << order) - 1;
993 nr_scanned += (1UL << order) - 1;
994 }
995 goto isolate_fail;
996 }
997 /* for alloc_contig case */
998 if (locked) {
999 unlock_page_lruvec_irqrestore(locked, flags);
1000 locked = NULL;
1001 }
1002
1003 ret = isolate_or_dissolve_huge_page(page, &cc->migratepages);
1004
1005 /*
1006 * Fail isolation in case isolate_or_dissolve_huge_page()
1007 * reports an error. In case of -ENOMEM, abort right away.
1008 */
1009 if (ret < 0) {
1010 /* Do not report -EBUSY down the chain */
1011 if (ret == -EBUSY)
1012 ret = 0;
1013 low_pfn += compound_nr(page) - 1;
1014 nr_scanned += compound_nr(page) - 1;
1015 goto isolate_fail;
1016 }
1017
1018 if (PageHuge(page)) {
1019 /*
1020 * Hugepage was successfully isolated and placed
1021 * on the cc->migratepages list.
1022 */
1023 folio = page_folio(page);
1024 low_pfn += folio_nr_pages(folio) - 1;
1025 goto isolate_success_no_list;
1026 }
1027
1028 /*
1029 * Ok, the hugepage was dissolved. Now these pages are
1030 * Buddy and cannot be re-allocated because they are
1031 * isolated. Fall-through as the check below handles
1032 * Buddy pages.
1033 */
1034 }
1035
1036 /*
1037 * Skip if free. We read page order here without zone lock
1038 * which is generally unsafe, but the race window is small and
1039 * the worst thing that can happen is that we skip some
1040 * potential isolation targets.
1041 */
1042 if (PageBuddy(page)) {
1043 unsigned long freepage_order = buddy_order_unsafe(page);
1044
1045 /*
1046 * Without lock, we cannot be sure that what we got is
1047 * a valid page order. Consider only values in the
1048 * valid order range to prevent low_pfn overflow.
1049 */
1050 if (freepage_order > 0 && freepage_order <= MAX_PAGE_ORDER) {
1051 low_pfn += (1UL << freepage_order) - 1;
1052 nr_scanned += (1UL << freepage_order) - 1;
1053 }
1054 continue;
1055 }
1056
1057 /*
1058 * Regardless of being on LRU, compound pages such as THP
1059 * (hugetlbfs is handled above) are not to be compacted unless
1060 * we are attempting an allocation larger than the compound
1061 * page size. We can potentially save a lot of iterations if we
1062 * skip them at once. The check is racy, but we can consider
1063 * only valid values and the only danger is skipping too much.
1064 */
1065 if (PageCompound(page) && !cc->alloc_contig) {
1066 const unsigned int order = compound_order(page);
1067
1068 /* Skip based on page order and compaction target order. */
1069 if (skip_isolation_on_order(order, cc->order)) {
1070 if (order <= MAX_PAGE_ORDER) {
1071 low_pfn += (1UL << order) - 1;
1072 nr_scanned += (1UL << order) - 1;
1073 }
1074 goto isolate_fail;
1075 }
1076 }
1077
1078 /*
1079 * Check may be lockless but that's ok as we recheck later.
1080 * It's possible to migrate LRU and non-lru movable pages.
1081 * Skip any other type of page
1082 */
1083 if (!PageLRU(page)) {
1084 /*
1085 * __PageMovable can return false positive so we need
1086 * to verify it under page_lock.
1087 */
1088 if (unlikely(__PageMovable(page)) &&
1089 !PageIsolated(page)) {
1090 if (locked) {
1091 unlock_page_lruvec_irqrestore(locked, flags);
1092 locked = NULL;
1093 }
1094
1095 if (isolate_movable_page(page, mode)) {
1096 folio = page_folio(page);
1097 goto isolate_success;
1098 }
1099 }
1100
1101 goto isolate_fail;
1102 }
1103
1104 /*
1105 * Be careful not to clear PageLRU until after we're
1106 * sure the page is not being freed elsewhere -- the
1107 * page release code relies on it.
1108 */
1109 folio = folio_get_nontail_page(page);
1110 if (unlikely(!folio))
1111 goto isolate_fail;
1112
1113 /*
1114 * Migration will fail if an anonymous page is pinned in memory,
1115 * so avoid taking lru_lock and isolating it unnecessarily in an
1116 * admittedly racy check.
1117 */
1118 mapping = folio_mapping(folio);
1119 if (!mapping && (folio_ref_count(folio) - 1) > folio_mapcount(folio))
1120 goto isolate_fail_put;
1121
1122 /*
1123 * Only allow to migrate anonymous pages in GFP_NOFS context
1124 * because those do not depend on fs locks.
1125 */
1126 if (!(cc->gfp_mask & __GFP_FS) && mapping)
1127 goto isolate_fail_put;
1128
1129 /* Only take pages on LRU: a check now makes later tests safe */
1130 if (!folio_test_lru(folio))
1131 goto isolate_fail_put;
1132
1133 is_unevictable = folio_test_unevictable(folio);
1134
1135 /* Compaction might skip unevictable pages but CMA takes them */
1136 if (!(mode & ISOLATE_UNEVICTABLE) && is_unevictable)
1137 goto isolate_fail_put;
1138
1139 /*
1140 * To minimise LRU disruption, the caller can indicate with
1141 * ISOLATE_ASYNC_MIGRATE that it only wants to isolate pages
1142 * it will be able to migrate without blocking - clean pages
1143 * for the most part. PageWriteback would require blocking.
1144 */
1145 if ((mode & ISOLATE_ASYNC_MIGRATE) && folio_test_writeback(folio))
1146 goto isolate_fail_put;
1147
1148 is_dirty = folio_test_dirty(folio);
1149
1150 if (((mode & ISOLATE_ASYNC_MIGRATE) && is_dirty) ||
1151 (mapping && is_unevictable)) {
1152 bool migrate_dirty = true;
1153 bool is_inaccessible;
1154
1155 /*
1156 * Only folios without mappings or that have
1157 * a ->migrate_folio callback are possible to migrate
1158 * without blocking.
1159 *
1160 * Folios from inaccessible mappings are not migratable.
1161 *
1162 * However, we can be racing with truncation, which can
1163 * free the mapping that we need to check. Truncation
1164 * holds the folio lock until after the folio is removed
1165 * from the page so holding it ourselves is sufficient.
1166 *
1167 * To avoid locking the folio just to check inaccessible,
1168 * assume every inaccessible folio is also unevictable,
1169 * which is a cheaper test. If our assumption goes
1170 * wrong, it's not a correctness bug, just potentially
1171 * wasted cycles.
1172 */
1173 if (!folio_trylock(folio))
1174 goto isolate_fail_put;
1175
1176 mapping = folio_mapping(folio);
1177 if ((mode & ISOLATE_ASYNC_MIGRATE) && is_dirty) {
1178 migrate_dirty = !mapping ||
1179 mapping->a_ops->migrate_folio;
1180 }
1181 is_inaccessible = mapping && mapping_inaccessible(mapping);
1182 folio_unlock(folio);
1183 if (!migrate_dirty || is_inaccessible)
1184 goto isolate_fail_put;
1185 }
1186
1187 /* Try isolate the folio */
1188 if (!folio_test_clear_lru(folio))
1189 goto isolate_fail_put;
1190
1191 lruvec = folio_lruvec(folio);
1192
1193 /* If we already hold the lock, we can skip some rechecking */
1194 if (lruvec != locked) {
1195 if (locked)
1196 unlock_page_lruvec_irqrestore(locked, flags);
1197
1198 compact_lock_irqsave(&lruvec->lru_lock, &flags, cc);
1199 locked = lruvec;
1200
1201 lruvec_memcg_debug(lruvec, folio);
1202
1203 /*
1204 * Try get exclusive access under lock. If marked for
1205 * skip, the scan is aborted unless the current context
1206 * is a rescan to reach the end of the pageblock.
1207 */
1208 if (!skip_updated && valid_page) {
1209 skip_updated = true;
1210 if (test_and_set_skip(cc, valid_page) &&
1211 !cc->finish_pageblock) {
1212 low_pfn = end_pfn;
1213 goto isolate_abort;
1214 }
1215 }
1216
1217 /*
1218 * Check LRU folio order under the lock
1219 */
1220 if (unlikely(skip_isolation_on_order(folio_order(folio),
1221 cc->order) &&
1222 !cc->alloc_contig)) {
1223 low_pfn += folio_nr_pages(folio) - 1;
1224 nr_scanned += folio_nr_pages(folio) - 1;
1225 folio_set_lru(folio);
1226 goto isolate_fail_put;
1227 }
1228 }
1229
1230 /* The folio is taken off the LRU */
1231 if (folio_test_large(folio))
1232 low_pfn += folio_nr_pages(folio) - 1;
1233
1234 /* Successfully isolated */
1235 lruvec_del_folio(lruvec, folio);
1236 node_stat_mod_folio(folio,
1237 NR_ISOLATED_ANON + folio_is_file_lru(folio),
1238 folio_nr_pages(folio));
1239
1240isolate_success:
1241 list_add(&folio->lru, &cc->migratepages);
1242isolate_success_no_list:
1243 cc->nr_migratepages += folio_nr_pages(folio);
1244 nr_isolated += folio_nr_pages(folio);
1245 nr_scanned += folio_nr_pages(folio) - 1;
1246
1247 /*
1248 * Avoid isolating too much unless this block is being
1249 * fully scanned (e.g. dirty/writeback pages, parallel allocation)
1250 * or a lock is contended. For contention, isolate quickly to
1251 * potentially remove one source of contention.
1252 */
1253 if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX &&
1254 !cc->finish_pageblock && !cc->contended) {
1255 ++low_pfn;
1256 break;
1257 }
1258
1259 continue;
1260
1261isolate_fail_put:
1262 /* Avoid potential deadlock in freeing page under lru_lock */
1263 if (locked) {
1264 unlock_page_lruvec_irqrestore(locked, flags);
1265 locked = NULL;
1266 }
1267 folio_put(folio);
1268
1269isolate_fail:
1270 if (!skip_on_failure && ret != -ENOMEM)
1271 continue;
1272
1273 /*
1274 * We have isolated some pages, but then failed. Release them
1275 * instead of migrating, as we cannot form the cc->order buddy
1276 * page anyway.
1277 */
1278 if (nr_isolated) {
1279 if (locked) {
1280 unlock_page_lruvec_irqrestore(locked, flags);
1281 locked = NULL;
1282 }
1283 putback_movable_pages(&cc->migratepages);
1284 cc->nr_migratepages = 0;
1285 nr_isolated = 0;
1286 }
1287
1288 if (low_pfn < next_skip_pfn) {
1289 low_pfn = next_skip_pfn - 1;
1290 /*
1291 * The check near the loop beginning would have updated
1292 * next_skip_pfn too, but this is a bit simpler.
1293 */
1294 next_skip_pfn += 1UL << cc->order;
1295 }
1296
1297 if (ret == -ENOMEM)
1298 break;
1299 }
1300
1301 /*
1302 * The PageBuddy() check could have potentially brought us outside
1303 * the range to be scanned.
1304 */
1305 if (unlikely(low_pfn > end_pfn))
1306 low_pfn = end_pfn;
1307
1308 folio = NULL;
1309
1310isolate_abort:
1311 if (locked)
1312 unlock_page_lruvec_irqrestore(locked, flags);
1313 if (folio) {
1314 folio_set_lru(folio);
1315 folio_put(folio);
1316 }
1317
1318 /*
1319 * Update the cached scanner pfn once the pageblock has been scanned.
1320 * Pages will either be migrated in which case there is no point
1321 * scanning in the near future or migration failed in which case the
1322 * failure reason may persist. The block is marked for skipping if
1323 * there were no pages isolated in the block or if the block is
1324 * rescanned twice in a row.
1325 */
1326 if (low_pfn == end_pfn && (!nr_isolated || cc->finish_pageblock)) {
1327 if (!cc->no_set_skip_hint && valid_page && !skip_updated)
1328 set_pageblock_skip(valid_page);
1329 update_cached_migrate(cc, low_pfn);
1330 }
1331
1332 trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
1333 nr_scanned, nr_isolated);
1334
1335fatal_pending:
1336 cc->total_migrate_scanned += nr_scanned;
1337 if (nr_isolated)
1338 count_compact_events(COMPACTISOLATED, nr_isolated);
1339
1340 cc->migrate_pfn = low_pfn;
1341
1342 return ret;
1343}
1344
1345/**
1346 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
1347 * @cc: Compaction control structure.
1348 * @start_pfn: The first PFN to start isolating.
1349 * @end_pfn: The one-past-last PFN.
1350 *
1351 * Returns -EAGAIN when contented, -EINTR in case of a signal pending, -ENOMEM
1352 * in case we could not allocate a page, or 0.
1353 */
1354int
1355isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
1356 unsigned long end_pfn)
1357{
1358 unsigned long pfn, block_start_pfn, block_end_pfn;
1359 int ret = 0;
1360
1361 /* Scan block by block. First and last block may be incomplete */
1362 pfn = start_pfn;
1363 block_start_pfn = pageblock_start_pfn(pfn);
1364 if (block_start_pfn < cc->zone->zone_start_pfn)
1365 block_start_pfn = cc->zone->zone_start_pfn;
1366 block_end_pfn = pageblock_end_pfn(pfn);
1367
1368 for (; pfn < end_pfn; pfn = block_end_pfn,
1369 block_start_pfn = block_end_pfn,
1370 block_end_pfn += pageblock_nr_pages) {
1371
1372 block_end_pfn = min(block_end_pfn, end_pfn);
1373
1374 if (!pageblock_pfn_to_page(block_start_pfn,
1375 block_end_pfn, cc->zone))
1376 continue;
1377
1378 ret = isolate_migratepages_block(cc, pfn, block_end_pfn,
1379 ISOLATE_UNEVICTABLE);
1380
1381 if (ret)
1382 break;
1383
1384 if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX)
1385 break;
1386 }
1387
1388 return ret;
1389}
1390
1391#endif /* CONFIG_COMPACTION || CONFIG_CMA */
1392#ifdef CONFIG_COMPACTION
1393
1394static bool suitable_migration_source(struct compact_control *cc,
1395 struct page *page)
1396{
1397 int block_mt;
1398
1399 if (pageblock_skip_persistent(page))
1400 return false;
1401
1402 if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction)
1403 return true;
1404
1405 block_mt = get_pageblock_migratetype(page);
1406
1407 if (cc->migratetype == MIGRATE_MOVABLE)
1408 return is_migrate_movable(block_mt);
1409 else
1410 return block_mt == cc->migratetype;
1411}
1412
1413/* Returns true if the page is within a block suitable for migration to */
1414static bool suitable_migration_target(struct compact_control *cc,
1415 struct page *page)
1416{
1417 /* If the page is a large free page, then disallow migration */
1418 if (PageBuddy(page)) {
1419 int order = cc->order > 0 ? cc->order : pageblock_order;
1420
1421 /*
1422 * We are checking page_order without zone->lock taken. But
1423 * the only small danger is that we skip a potentially suitable
1424 * pageblock, so it's not worth to check order for valid range.
1425 */
1426 if (buddy_order_unsafe(page) >= order)
1427 return false;
1428 }
1429
1430 if (cc->ignore_block_suitable)
1431 return true;
1432
1433 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1434 if (is_migrate_movable(get_pageblock_migratetype(page)))
1435 return true;
1436
1437 /* Otherwise skip the block */
1438 return false;
1439}
1440
1441static inline unsigned int
1442freelist_scan_limit(struct compact_control *cc)
1443{
1444 unsigned short shift = BITS_PER_LONG - 1;
1445
1446 return (COMPACT_CLUSTER_MAX >> min(shift, cc->fast_search_fail)) + 1;
1447}
1448
1449/*
1450 * Test whether the free scanner has reached the same or lower pageblock than
1451 * the migration scanner, and compaction should thus terminate.
1452 */
1453static inline bool compact_scanners_met(struct compact_control *cc)
1454{
1455 return (cc->free_pfn >> pageblock_order)
1456 <= (cc->migrate_pfn >> pageblock_order);
1457}
1458
1459/*
1460 * Used when scanning for a suitable migration target which scans freelists
1461 * in reverse. Reorders the list such as the unscanned pages are scanned
1462 * first on the next iteration of the free scanner
1463 */
1464static void
1465move_freelist_head(struct list_head *freelist, struct page *freepage)
1466{
1467 LIST_HEAD(sublist);
1468
1469 if (!list_is_first(&freepage->buddy_list, freelist)) {
1470 list_cut_before(&sublist, freelist, &freepage->buddy_list);
1471 list_splice_tail(&sublist, freelist);
1472 }
1473}
1474
1475/*
1476 * Similar to move_freelist_head except used by the migration scanner
1477 * when scanning forward. It's possible for these list operations to
1478 * move against each other if they search the free list exactly in
1479 * lockstep.
1480 */
1481static void
1482move_freelist_tail(struct list_head *freelist, struct page *freepage)
1483{
1484 LIST_HEAD(sublist);
1485
1486 if (!list_is_last(&freepage->buddy_list, freelist)) {
1487 list_cut_position(&sublist, freelist, &freepage->buddy_list);
1488 list_splice_tail(&sublist, freelist);
1489 }
1490}
1491
1492static void
1493fast_isolate_around(struct compact_control *cc, unsigned long pfn)
1494{
1495 unsigned long start_pfn, end_pfn;
1496 struct page *page;
1497
1498 /* Do not search around if there are enough pages already */
1499 if (cc->nr_freepages >= cc->nr_migratepages)
1500 return;
1501
1502 /* Minimise scanning during async compaction */
1503 if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC)
1504 return;
1505
1506 /* Pageblock boundaries */
1507 start_pfn = max(pageblock_start_pfn(pfn), cc->zone->zone_start_pfn);
1508 end_pfn = min(pageblock_end_pfn(pfn), zone_end_pfn(cc->zone));
1509
1510 page = pageblock_pfn_to_page(start_pfn, end_pfn, cc->zone);
1511 if (!page)
1512 return;
1513
1514 isolate_freepages_block(cc, &start_pfn, end_pfn, cc->freepages, 1, false);
1515
1516 /* Skip this pageblock in the future as it's full or nearly full */
1517 if (start_pfn == end_pfn && !cc->no_set_skip_hint)
1518 set_pageblock_skip(page);
1519}
1520
1521/* Search orders in round-robin fashion */
1522static int next_search_order(struct compact_control *cc, int order)
1523{
1524 order--;
1525 if (order < 0)
1526 order = cc->order - 1;
1527
1528 /* Search wrapped around? */
1529 if (order == cc->search_order) {
1530 cc->search_order--;
1531 if (cc->search_order < 0)
1532 cc->search_order = cc->order - 1;
1533 return -1;
1534 }
1535
1536 return order;
1537}
1538
1539static void fast_isolate_freepages(struct compact_control *cc)
1540{
1541 unsigned int limit = max(1U, freelist_scan_limit(cc) >> 1);
1542 unsigned int nr_scanned = 0, total_isolated = 0;
1543 unsigned long low_pfn, min_pfn, highest = 0;
1544 unsigned long nr_isolated = 0;
1545 unsigned long distance;
1546 struct page *page = NULL;
1547 bool scan_start = false;
1548 int order;
1549
1550 /* Full compaction passes in a negative order */
1551 if (cc->order <= 0)
1552 return;
1553
1554 /*
1555 * If starting the scan, use a deeper search and use the highest
1556 * PFN found if a suitable one is not found.
1557 */
1558 if (cc->free_pfn >= cc->zone->compact_init_free_pfn) {
1559 limit = pageblock_nr_pages >> 1;
1560 scan_start = true;
1561 }
1562
1563 /*
1564 * Preferred point is in the top quarter of the scan space but take
1565 * a pfn from the top half if the search is problematic.
1566 */
1567 distance = (cc->free_pfn - cc->migrate_pfn);
1568 low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2));
1569 min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1));
1570
1571 if (WARN_ON_ONCE(min_pfn > low_pfn))
1572 low_pfn = min_pfn;
1573
1574 /*
1575 * Search starts from the last successful isolation order or the next
1576 * order to search after a previous failure
1577 */
1578 cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order);
1579
1580 for (order = cc->search_order;
1581 !page && order >= 0;
1582 order = next_search_order(cc, order)) {
1583 struct free_area *area = &cc->zone->free_area[order];
1584 struct list_head *freelist;
1585 struct page *freepage;
1586 unsigned long flags;
1587 unsigned int order_scanned = 0;
1588 unsigned long high_pfn = 0;
1589
1590 if (!area->nr_free)
1591 continue;
1592
1593 spin_lock_irqsave(&cc->zone->lock, flags);
1594 freelist = &area->free_list[MIGRATE_MOVABLE];
1595 list_for_each_entry_reverse(freepage, freelist, buddy_list) {
1596 unsigned long pfn;
1597
1598 order_scanned++;
1599 nr_scanned++;
1600 pfn = page_to_pfn(freepage);
1601
1602 if (pfn >= highest)
1603 highest = max(pageblock_start_pfn(pfn),
1604 cc->zone->zone_start_pfn);
1605
1606 if (pfn >= low_pfn) {
1607 cc->fast_search_fail = 0;
1608 cc->search_order = order;
1609 page = freepage;
1610 break;
1611 }
1612
1613 if (pfn >= min_pfn && pfn > high_pfn) {
1614 high_pfn = pfn;
1615
1616 /* Shorten the scan if a candidate is found */
1617 limit >>= 1;
1618 }
1619
1620 if (order_scanned >= limit)
1621 break;
1622 }
1623
1624 /* Use a maximum candidate pfn if a preferred one was not found */
1625 if (!page && high_pfn) {
1626 page = pfn_to_page(high_pfn);
1627
1628 /* Update freepage for the list reorder below */
1629 freepage = page;
1630 }
1631
1632 /* Reorder to so a future search skips recent pages */
1633 move_freelist_head(freelist, freepage);
1634
1635 /* Isolate the page if available */
1636 if (page) {
1637 if (__isolate_free_page(page, order)) {
1638 set_page_private(page, order);
1639 nr_isolated = 1 << order;
1640 nr_scanned += nr_isolated - 1;
1641 total_isolated += nr_isolated;
1642 cc->nr_freepages += nr_isolated;
1643 list_add_tail(&page->lru, &cc->freepages[order]);
1644 count_compact_events(COMPACTISOLATED, nr_isolated);
1645 } else {
1646 /* If isolation fails, abort the search */
1647 order = cc->search_order + 1;
1648 page = NULL;
1649 }
1650 }
1651
1652 spin_unlock_irqrestore(&cc->zone->lock, flags);
1653
1654 /* Skip fast search if enough freepages isolated */
1655 if (cc->nr_freepages >= cc->nr_migratepages)
1656 break;
1657
1658 /*
1659 * Smaller scan on next order so the total scan is related
1660 * to freelist_scan_limit.
1661 */
1662 if (order_scanned >= limit)
1663 limit = max(1U, limit >> 1);
1664 }
1665
1666 trace_mm_compaction_fast_isolate_freepages(min_pfn, cc->free_pfn,
1667 nr_scanned, total_isolated);
1668
1669 if (!page) {
1670 cc->fast_search_fail++;
1671 if (scan_start) {
1672 /*
1673 * Use the highest PFN found above min. If one was
1674 * not found, be pessimistic for direct compaction
1675 * and use the min mark.
1676 */
1677 if (highest >= min_pfn) {
1678 page = pfn_to_page(highest);
1679 cc->free_pfn = highest;
1680 } else {
1681 if (cc->direct_compaction && pfn_valid(min_pfn)) {
1682 page = pageblock_pfn_to_page(min_pfn,
1683 min(pageblock_end_pfn(min_pfn),
1684 zone_end_pfn(cc->zone)),
1685 cc->zone);
1686 if (page && !suitable_migration_target(cc, page))
1687 page = NULL;
1688
1689 cc->free_pfn = min_pfn;
1690 }
1691 }
1692 }
1693 }
1694
1695 if (highest && highest >= cc->zone->compact_cached_free_pfn) {
1696 highest -= pageblock_nr_pages;
1697 cc->zone->compact_cached_free_pfn = highest;
1698 }
1699
1700 cc->total_free_scanned += nr_scanned;
1701 if (!page)
1702 return;
1703
1704 low_pfn = page_to_pfn(page);
1705 fast_isolate_around(cc, low_pfn);
1706}
1707
1708/*
1709 * Based on information in the current compact_control, find blocks
1710 * suitable for isolating free pages from and then isolate them.
1711 */
1712static void isolate_freepages(struct compact_control *cc)
1713{
1714 struct zone *zone = cc->zone;
1715 struct page *page;
1716 unsigned long block_start_pfn; /* start of current pageblock */
1717 unsigned long isolate_start_pfn; /* exact pfn we start at */
1718 unsigned long block_end_pfn; /* end of current pageblock */
1719 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
1720 unsigned int stride;
1721
1722 /* Try a small search of the free lists for a candidate */
1723 fast_isolate_freepages(cc);
1724 if (cc->nr_freepages)
1725 return;
1726
1727 /*
1728 * Initialise the free scanner. The starting point is where we last
1729 * successfully isolated from, zone-cached value, or the end of the
1730 * zone when isolating for the first time. For looping we also need
1731 * this pfn aligned down to the pageblock boundary, because we do
1732 * block_start_pfn -= pageblock_nr_pages in the for loop.
1733 * For ending point, take care when isolating in last pageblock of a
1734 * zone which ends in the middle of a pageblock.
1735 * The low boundary is the end of the pageblock the migration scanner
1736 * is using.
1737 */
1738 isolate_start_pfn = cc->free_pfn;
1739 block_start_pfn = pageblock_start_pfn(isolate_start_pfn);
1740 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1741 zone_end_pfn(zone));
1742 low_pfn = pageblock_end_pfn(cc->migrate_pfn);
1743 stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1;
1744
1745 /*
1746 * Isolate free pages until enough are available to migrate the
1747 * pages on cc->migratepages. We stop searching if the migrate
1748 * and free page scanners meet or enough free pages are isolated.
1749 */
1750 for (; block_start_pfn >= low_pfn;
1751 block_end_pfn = block_start_pfn,
1752 block_start_pfn -= pageblock_nr_pages,
1753 isolate_start_pfn = block_start_pfn) {
1754 unsigned long nr_isolated;
1755
1756 /*
1757 * This can iterate a massively long zone without finding any
1758 * suitable migration targets, so periodically check resched.
1759 */
1760 if (!(block_start_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages)))
1761 cond_resched();
1762
1763 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1764 zone);
1765 if (!page) {
1766 unsigned long next_pfn;
1767
1768 next_pfn = skip_offline_sections_reverse(block_start_pfn);
1769 if (next_pfn)
1770 block_start_pfn = max(next_pfn, low_pfn);
1771
1772 continue;
1773 }
1774
1775 /* Check the block is suitable for migration */
1776 if (!suitable_migration_target(cc, page))
1777 continue;
1778
1779 /* If isolation recently failed, do not retry */
1780 if (!isolation_suitable(cc, page))
1781 continue;
1782
1783 /* Found a block suitable for isolating free pages from. */
1784 nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn,
1785 block_end_pfn, cc->freepages, stride, false);
1786
1787 /* Update the skip hint if the full pageblock was scanned */
1788 if (isolate_start_pfn == block_end_pfn)
1789 update_pageblock_skip(cc, page, block_start_pfn -
1790 pageblock_nr_pages);
1791
1792 /* Are enough freepages isolated? */
1793 if (cc->nr_freepages >= cc->nr_migratepages) {
1794 if (isolate_start_pfn >= block_end_pfn) {
1795 /*
1796 * Restart at previous pageblock if more
1797 * freepages can be isolated next time.
1798 */
1799 isolate_start_pfn =
1800 block_start_pfn - pageblock_nr_pages;
1801 }
1802 break;
1803 } else if (isolate_start_pfn < block_end_pfn) {
1804 /*
1805 * If isolation failed early, do not continue
1806 * needlessly.
1807 */
1808 break;
1809 }
1810
1811 /* Adjust stride depending on isolation */
1812 if (nr_isolated) {
1813 stride = 1;
1814 continue;
1815 }
1816 stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1);
1817 }
1818
1819 /*
1820 * Record where the free scanner will restart next time. Either we
1821 * broke from the loop and set isolate_start_pfn based on the last
1822 * call to isolate_freepages_block(), or we met the migration scanner
1823 * and the loop terminated due to isolate_start_pfn < low_pfn
1824 */
1825 cc->free_pfn = isolate_start_pfn;
1826}
1827
1828/*
1829 * This is a migrate-callback that "allocates" freepages by taking pages
1830 * from the isolated freelists in the block we are migrating to.
1831 */
1832static struct folio *compaction_alloc_noprof(struct folio *src, unsigned long data)
1833{
1834 struct compact_control *cc = (struct compact_control *)data;
1835 struct folio *dst;
1836 int order = folio_order(src);
1837 bool has_isolated_pages = false;
1838 int start_order;
1839 struct page *freepage;
1840 unsigned long size;
1841
1842again:
1843 for (start_order = order; start_order < NR_PAGE_ORDERS; start_order++)
1844 if (!list_empty(&cc->freepages[start_order]))
1845 break;
1846
1847 /* no free pages in the list */
1848 if (start_order == NR_PAGE_ORDERS) {
1849 if (has_isolated_pages)
1850 return NULL;
1851 isolate_freepages(cc);
1852 has_isolated_pages = true;
1853 goto again;
1854 }
1855
1856 freepage = list_first_entry(&cc->freepages[start_order], struct page,
1857 lru);
1858 size = 1 << start_order;
1859
1860 list_del(&freepage->lru);
1861
1862 while (start_order > order) {
1863 start_order--;
1864 size >>= 1;
1865
1866 list_add(&freepage[size].lru, &cc->freepages[start_order]);
1867 set_page_private(&freepage[size], start_order);
1868 }
1869 dst = (struct folio *)freepage;
1870
1871 post_alloc_hook(&dst->page, order, __GFP_MOVABLE);
1872 if (order)
1873 prep_compound_page(&dst->page, order);
1874 cc->nr_freepages -= 1 << order;
1875 cc->nr_migratepages -= 1 << order;
1876 return page_rmappable_folio(&dst->page);
1877}
1878
1879static struct folio *compaction_alloc(struct folio *src, unsigned long data)
1880{
1881 return alloc_hooks(compaction_alloc_noprof(src, data));
1882}
1883
1884/*
1885 * This is a migrate-callback that "frees" freepages back to the isolated
1886 * freelist. All pages on the freelist are from the same zone, so there is no
1887 * special handling needed for NUMA.
1888 */
1889static void compaction_free(struct folio *dst, unsigned long data)
1890{
1891 struct compact_control *cc = (struct compact_control *)data;
1892 int order = folio_order(dst);
1893 struct page *page = &dst->page;
1894
1895 if (folio_put_testzero(dst)) {
1896 free_pages_prepare(page, order);
1897 list_add(&dst->lru, &cc->freepages[order]);
1898 cc->nr_freepages += 1 << order;
1899 }
1900 cc->nr_migratepages += 1 << order;
1901 /*
1902 * someone else has referenced the page, we cannot take it back to our
1903 * free list.
1904 */
1905}
1906
1907/* possible outcome of isolate_migratepages */
1908typedef enum {
1909 ISOLATE_ABORT, /* Abort compaction now */
1910 ISOLATE_NONE, /* No pages isolated, continue scanning */
1911 ISOLATE_SUCCESS, /* Pages isolated, migrate */
1912} isolate_migrate_t;
1913
1914/*
1915 * Allow userspace to control policy on scanning the unevictable LRU for
1916 * compactable pages.
1917 */
1918static int sysctl_compact_unevictable_allowed __read_mostly = CONFIG_COMPACT_UNEVICTABLE_DEFAULT;
1919/*
1920 * Tunable for proactive compaction. It determines how
1921 * aggressively the kernel should compact memory in the
1922 * background. It takes values in the range [0, 100].
1923 */
1924static unsigned int __read_mostly sysctl_compaction_proactiveness = 20;
1925static int sysctl_extfrag_threshold = 500;
1926static int __read_mostly sysctl_compact_memory;
1927
1928static inline void
1929update_fast_start_pfn(struct compact_control *cc, unsigned long pfn)
1930{
1931 if (cc->fast_start_pfn == ULONG_MAX)
1932 return;
1933
1934 if (!cc->fast_start_pfn)
1935 cc->fast_start_pfn = pfn;
1936
1937 cc->fast_start_pfn = min(cc->fast_start_pfn, pfn);
1938}
1939
1940static inline unsigned long
1941reinit_migrate_pfn(struct compact_control *cc)
1942{
1943 if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX)
1944 return cc->migrate_pfn;
1945
1946 cc->migrate_pfn = cc->fast_start_pfn;
1947 cc->fast_start_pfn = ULONG_MAX;
1948
1949 return cc->migrate_pfn;
1950}
1951
1952/*
1953 * Briefly search the free lists for a migration source that already has
1954 * some free pages to reduce the number of pages that need migration
1955 * before a pageblock is free.
1956 */
1957static unsigned long fast_find_migrateblock(struct compact_control *cc)
1958{
1959 unsigned int limit = freelist_scan_limit(cc);
1960 unsigned int nr_scanned = 0;
1961 unsigned long distance;
1962 unsigned long pfn = cc->migrate_pfn;
1963 unsigned long high_pfn;
1964 int order;
1965 bool found_block = false;
1966
1967 /* Skip hints are relied on to avoid repeats on the fast search */
1968 if (cc->ignore_skip_hint)
1969 return pfn;
1970
1971 /*
1972 * If the pageblock should be finished then do not select a different
1973 * pageblock.
1974 */
1975 if (cc->finish_pageblock)
1976 return pfn;
1977
1978 /*
1979 * If the migrate_pfn is not at the start of a zone or the start
1980 * of a pageblock then assume this is a continuation of a previous
1981 * scan restarted due to COMPACT_CLUSTER_MAX.
1982 */
1983 if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn))
1984 return pfn;
1985
1986 /*
1987 * For smaller orders, just linearly scan as the number of pages
1988 * to migrate should be relatively small and does not necessarily
1989 * justify freeing up a large block for a small allocation.
1990 */
1991 if (cc->order <= PAGE_ALLOC_COSTLY_ORDER)
1992 return pfn;
1993
1994 /*
1995 * Only allow kcompactd and direct requests for movable pages to
1996 * quickly clear out a MOVABLE pageblock for allocation. This
1997 * reduces the risk that a large movable pageblock is freed for
1998 * an unmovable/reclaimable small allocation.
1999 */
2000 if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE)
2001 return pfn;
2002
2003 /*
2004 * When starting the migration scanner, pick any pageblock within the
2005 * first half of the search space. Otherwise try and pick a pageblock
2006 * within the first eighth to reduce the chances that a migration
2007 * target later becomes a source.
2008 */
2009 distance = (cc->free_pfn - cc->migrate_pfn) >> 1;
2010 if (cc->migrate_pfn != cc->zone->zone_start_pfn)
2011 distance >>= 2;
2012 high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance);
2013
2014 for (order = cc->order - 1;
2015 order >= PAGE_ALLOC_COSTLY_ORDER && !found_block && nr_scanned < limit;
2016 order--) {
2017 struct free_area *area = &cc->zone->free_area[order];
2018 struct list_head *freelist;
2019 unsigned long flags;
2020 struct page *freepage;
2021
2022 if (!area->nr_free)
2023 continue;
2024
2025 spin_lock_irqsave(&cc->zone->lock, flags);
2026 freelist = &area->free_list[MIGRATE_MOVABLE];
2027 list_for_each_entry(freepage, freelist, buddy_list) {
2028 unsigned long free_pfn;
2029
2030 if (nr_scanned++ >= limit) {
2031 move_freelist_tail(freelist, freepage);
2032 break;
2033 }
2034
2035 free_pfn = page_to_pfn(freepage);
2036 if (free_pfn < high_pfn) {
2037 /*
2038 * Avoid if skipped recently. Ideally it would
2039 * move to the tail but even safe iteration of
2040 * the list assumes an entry is deleted, not
2041 * reordered.
2042 */
2043 if (get_pageblock_skip(freepage))
2044 continue;
2045
2046 /* Reorder to so a future search skips recent pages */
2047 move_freelist_tail(freelist, freepage);
2048
2049 update_fast_start_pfn(cc, free_pfn);
2050 pfn = pageblock_start_pfn(free_pfn);
2051 if (pfn < cc->zone->zone_start_pfn)
2052 pfn = cc->zone->zone_start_pfn;
2053 cc->fast_search_fail = 0;
2054 found_block = true;
2055 break;
2056 }
2057 }
2058 spin_unlock_irqrestore(&cc->zone->lock, flags);
2059 }
2060
2061 cc->total_migrate_scanned += nr_scanned;
2062
2063 /*
2064 * If fast scanning failed then use a cached entry for a page block
2065 * that had free pages as the basis for starting a linear scan.
2066 */
2067 if (!found_block) {
2068 cc->fast_search_fail++;
2069 pfn = reinit_migrate_pfn(cc);
2070 }
2071 return pfn;
2072}
2073
2074/*
2075 * Isolate all pages that can be migrated from the first suitable block,
2076 * starting at the block pointed to by the migrate scanner pfn within
2077 * compact_control.
2078 */
2079static isolate_migrate_t isolate_migratepages(struct compact_control *cc)
2080{
2081 unsigned long block_start_pfn;
2082 unsigned long block_end_pfn;
2083 unsigned long low_pfn;
2084 struct page *page;
2085 const isolate_mode_t isolate_mode =
2086 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
2087 (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
2088 bool fast_find_block;
2089
2090 /*
2091 * Start at where we last stopped, or beginning of the zone as
2092 * initialized by compact_zone(). The first failure will use
2093 * the lowest PFN as the starting point for linear scanning.
2094 */
2095 low_pfn = fast_find_migrateblock(cc);
2096 block_start_pfn = pageblock_start_pfn(low_pfn);
2097 if (block_start_pfn < cc->zone->zone_start_pfn)
2098 block_start_pfn = cc->zone->zone_start_pfn;
2099
2100 /*
2101 * fast_find_migrateblock() has already ensured the pageblock is not
2102 * set with a skipped flag, so to avoid the isolation_suitable check
2103 * below again, check whether the fast search was successful.
2104 */
2105 fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail;
2106
2107 /* Only scan within a pageblock boundary */
2108 block_end_pfn = pageblock_end_pfn(low_pfn);
2109
2110 /*
2111 * Iterate over whole pageblocks until we find the first suitable.
2112 * Do not cross the free scanner.
2113 */
2114 for (; block_end_pfn <= cc->free_pfn;
2115 fast_find_block = false,
2116 cc->migrate_pfn = low_pfn = block_end_pfn,
2117 block_start_pfn = block_end_pfn,
2118 block_end_pfn += pageblock_nr_pages) {
2119
2120 /*
2121 * This can potentially iterate a massively long zone with
2122 * many pageblocks unsuitable, so periodically check if we
2123 * need to schedule.
2124 */
2125 if (!(low_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages)))
2126 cond_resched();
2127
2128 page = pageblock_pfn_to_page(block_start_pfn,
2129 block_end_pfn, cc->zone);
2130 if (!page) {
2131 unsigned long next_pfn;
2132
2133 next_pfn = skip_offline_sections(block_start_pfn);
2134 if (next_pfn)
2135 block_end_pfn = min(next_pfn, cc->free_pfn);
2136 continue;
2137 }
2138
2139 /*
2140 * If isolation recently failed, do not retry. Only check the
2141 * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock
2142 * to be visited multiple times. Assume skip was checked
2143 * before making it "skip" so other compaction instances do
2144 * not scan the same block.
2145 */
2146 if ((pageblock_aligned(low_pfn) ||
2147 low_pfn == cc->zone->zone_start_pfn) &&
2148 !fast_find_block && !isolation_suitable(cc, page))
2149 continue;
2150
2151 /*
2152 * For async direct compaction, only scan the pageblocks of the
2153 * same migratetype without huge pages. Async direct compaction
2154 * is optimistic to see if the minimum amount of work satisfies
2155 * the allocation. The cached PFN is updated as it's possible
2156 * that all remaining blocks between source and target are
2157 * unsuitable and the compaction scanners fail to meet.
2158 */
2159 if (!suitable_migration_source(cc, page)) {
2160 update_cached_migrate(cc, block_end_pfn);
2161 continue;
2162 }
2163
2164 /* Perform the isolation */
2165 if (isolate_migratepages_block(cc, low_pfn, block_end_pfn,
2166 isolate_mode))
2167 return ISOLATE_ABORT;
2168
2169 /*
2170 * Either we isolated something and proceed with migration. Or
2171 * we failed and compact_zone should decide if we should
2172 * continue or not.
2173 */
2174 break;
2175 }
2176
2177 return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
2178}
2179
2180/*
2181 * Determine whether kswapd is (or recently was!) running on this node.
2182 *
2183 * pgdat_kswapd_lock() pins pgdat->kswapd, so a concurrent kswapd_stop() can't
2184 * zero it.
2185 */
2186static bool kswapd_is_running(pg_data_t *pgdat)
2187{
2188 bool running;
2189
2190 pgdat_kswapd_lock(pgdat);
2191 running = pgdat->kswapd && task_is_running(pgdat->kswapd);
2192 pgdat_kswapd_unlock(pgdat);
2193
2194 return running;
2195}
2196
2197/*
2198 * A zone's fragmentation score is the external fragmentation wrt to the
2199 * COMPACTION_HPAGE_ORDER. It returns a value in the range [0, 100].
2200 */
2201static unsigned int fragmentation_score_zone(struct zone *zone)
2202{
2203 return extfrag_for_order(zone, COMPACTION_HPAGE_ORDER);
2204}
2205
2206/*
2207 * A weighted zone's fragmentation score is the external fragmentation
2208 * wrt to the COMPACTION_HPAGE_ORDER scaled by the zone's size. It
2209 * returns a value in the range [0, 100].
2210 *
2211 * The scaling factor ensures that proactive compaction focuses on larger
2212 * zones like ZONE_NORMAL, rather than smaller, specialized zones like
2213 * ZONE_DMA32. For smaller zones, the score value remains close to zero,
2214 * and thus never exceeds the high threshold for proactive compaction.
2215 */
2216static unsigned int fragmentation_score_zone_weighted(struct zone *zone)
2217{
2218 unsigned long score;
2219
2220 score = zone->present_pages * fragmentation_score_zone(zone);
2221 return div64_ul(score, zone->zone_pgdat->node_present_pages + 1);
2222}
2223
2224/*
2225 * The per-node proactive (background) compaction process is started by its
2226 * corresponding kcompactd thread when the node's fragmentation score
2227 * exceeds the high threshold. The compaction process remains active till
2228 * the node's score falls below the low threshold, or one of the back-off
2229 * conditions is met.
2230 */
2231static unsigned int fragmentation_score_node(pg_data_t *pgdat)
2232{
2233 unsigned int score = 0;
2234 int zoneid;
2235
2236 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2237 struct zone *zone;
2238
2239 zone = &pgdat->node_zones[zoneid];
2240 if (!populated_zone(zone))
2241 continue;
2242 score += fragmentation_score_zone_weighted(zone);
2243 }
2244
2245 return score;
2246}
2247
2248static unsigned int fragmentation_score_wmark(bool low)
2249{
2250 unsigned int wmark_low;
2251
2252 /*
2253 * Cap the low watermark to avoid excessive compaction
2254 * activity in case a user sets the proactiveness tunable
2255 * close to 100 (maximum).
2256 */
2257 wmark_low = max(100U - sysctl_compaction_proactiveness, 5U);
2258 return low ? wmark_low : min(wmark_low + 10, 100U);
2259}
2260
2261static bool should_proactive_compact_node(pg_data_t *pgdat)
2262{
2263 int wmark_high;
2264
2265 if (!sysctl_compaction_proactiveness || kswapd_is_running(pgdat))
2266 return false;
2267
2268 wmark_high = fragmentation_score_wmark(false);
2269 return fragmentation_score_node(pgdat) > wmark_high;
2270}
2271
2272static enum compact_result __compact_finished(struct compact_control *cc)
2273{
2274 unsigned int order;
2275 const int migratetype = cc->migratetype;
2276 int ret;
2277
2278 /* Compaction run completes if the migrate and free scanner meet */
2279 if (compact_scanners_met(cc)) {
2280 /* Let the next compaction start anew. */
2281 reset_cached_positions(cc->zone);
2282
2283 /*
2284 * Mark that the PG_migrate_skip information should be cleared
2285 * by kswapd when it goes to sleep. kcompactd does not set the
2286 * flag itself as the decision to be clear should be directly
2287 * based on an allocation request.
2288 */
2289 if (cc->direct_compaction)
2290 cc->zone->compact_blockskip_flush = true;
2291
2292 if (cc->whole_zone)
2293 return COMPACT_COMPLETE;
2294 else
2295 return COMPACT_PARTIAL_SKIPPED;
2296 }
2297
2298 if (cc->proactive_compaction) {
2299 int score, wmark_low;
2300 pg_data_t *pgdat;
2301
2302 pgdat = cc->zone->zone_pgdat;
2303 if (kswapd_is_running(pgdat))
2304 return COMPACT_PARTIAL_SKIPPED;
2305
2306 score = fragmentation_score_zone(cc->zone);
2307 wmark_low = fragmentation_score_wmark(true);
2308
2309 if (score > wmark_low)
2310 ret = COMPACT_CONTINUE;
2311 else
2312 ret = COMPACT_SUCCESS;
2313
2314 goto out;
2315 }
2316
2317 if (is_via_compact_memory(cc->order))
2318 return COMPACT_CONTINUE;
2319
2320 /*
2321 * Always finish scanning a pageblock to reduce the possibility of
2322 * fallbacks in the future. This is particularly important when
2323 * migration source is unmovable/reclaimable but it's not worth
2324 * special casing.
2325 */
2326 if (!pageblock_aligned(cc->migrate_pfn))
2327 return COMPACT_CONTINUE;
2328
2329 /* Direct compactor: Is a suitable page free? */
2330 ret = COMPACT_NO_SUITABLE_PAGE;
2331 for (order = cc->order; order < NR_PAGE_ORDERS; order++) {
2332 struct free_area *area = &cc->zone->free_area[order];
2333 bool can_steal;
2334
2335 /* Job done if page is free of the right migratetype */
2336 if (!free_area_empty(area, migratetype))
2337 return COMPACT_SUCCESS;
2338
2339#ifdef CONFIG_CMA
2340 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
2341 if (migratetype == MIGRATE_MOVABLE &&
2342 !free_area_empty(area, MIGRATE_CMA))
2343 return COMPACT_SUCCESS;
2344#endif
2345 /*
2346 * Job done if allocation would steal freepages from
2347 * other migratetype buddy lists.
2348 */
2349 if (find_suitable_fallback(area, order, migratetype,
2350 true, &can_steal) != -1)
2351 /*
2352 * Movable pages are OK in any pageblock. If we are
2353 * stealing for a non-movable allocation, make sure
2354 * we finish compacting the current pageblock first
2355 * (which is assured by the above migrate_pfn align
2356 * check) so it is as free as possible and we won't
2357 * have to steal another one soon.
2358 */
2359 return COMPACT_SUCCESS;
2360 }
2361
2362out:
2363 if (cc->contended || fatal_signal_pending(current))
2364 ret = COMPACT_CONTENDED;
2365
2366 return ret;
2367}
2368
2369static enum compact_result compact_finished(struct compact_control *cc)
2370{
2371 int ret;
2372
2373 ret = __compact_finished(cc);
2374 trace_mm_compaction_finished(cc->zone, cc->order, ret);
2375 if (ret == COMPACT_NO_SUITABLE_PAGE)
2376 ret = COMPACT_CONTINUE;
2377
2378 return ret;
2379}
2380
2381static bool __compaction_suitable(struct zone *zone, int order,
2382 int highest_zoneidx,
2383 unsigned long wmark_target)
2384{
2385 unsigned long watermark;
2386 /*
2387 * Watermarks for order-0 must be met for compaction to be able to
2388 * isolate free pages for migration targets. This means that the
2389 * watermark and alloc_flags have to match, or be more pessimistic than
2390 * the check in __isolate_free_page(). We don't use the direct
2391 * compactor's alloc_flags, as they are not relevant for freepage
2392 * isolation. We however do use the direct compactor's highest_zoneidx
2393 * to skip over zones where lowmem reserves would prevent allocation
2394 * even if compaction succeeds.
2395 * For costly orders, we require low watermark instead of min for
2396 * compaction to proceed to increase its chances.
2397 * ALLOC_CMA is used, as pages in CMA pageblocks are considered
2398 * suitable migration targets
2399 */
2400 watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
2401 low_wmark_pages(zone) : min_wmark_pages(zone);
2402 watermark += compact_gap(order);
2403 return __zone_watermark_ok(zone, 0, watermark, highest_zoneidx,
2404 ALLOC_CMA, wmark_target);
2405}
2406
2407/*
2408 * compaction_suitable: Is this suitable to run compaction on this zone now?
2409 */
2410bool compaction_suitable(struct zone *zone, int order, int highest_zoneidx)
2411{
2412 enum compact_result compact_result;
2413 bool suitable;
2414
2415 suitable = __compaction_suitable(zone, order, highest_zoneidx,
2416 zone_page_state(zone, NR_FREE_PAGES));
2417 /*
2418 * fragmentation index determines if allocation failures are due to
2419 * low memory or external fragmentation
2420 *
2421 * index of -1000 would imply allocations might succeed depending on
2422 * watermarks, but we already failed the high-order watermark check
2423 * index towards 0 implies failure is due to lack of memory
2424 * index towards 1000 implies failure is due to fragmentation
2425 *
2426 * Only compact if a failure would be due to fragmentation. Also
2427 * ignore fragindex for non-costly orders where the alternative to
2428 * a successful reclaim/compaction is OOM. Fragindex and the
2429 * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
2430 * excessive compaction for costly orders, but it should not be at the
2431 * expense of system stability.
2432 */
2433 if (suitable) {
2434 compact_result = COMPACT_CONTINUE;
2435 if (order > PAGE_ALLOC_COSTLY_ORDER) {
2436 int fragindex = fragmentation_index(zone, order);
2437
2438 if (fragindex >= 0 &&
2439 fragindex <= sysctl_extfrag_threshold) {
2440 suitable = false;
2441 compact_result = COMPACT_NOT_SUITABLE_ZONE;
2442 }
2443 }
2444 } else {
2445 compact_result = COMPACT_SKIPPED;
2446 }
2447
2448 trace_mm_compaction_suitable(zone, order, compact_result);
2449
2450 return suitable;
2451}
2452
2453bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
2454 int alloc_flags)
2455{
2456 struct zone *zone;
2457 struct zoneref *z;
2458
2459 /*
2460 * Make sure at least one zone would pass __compaction_suitable if we continue
2461 * retrying the reclaim.
2462 */
2463 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2464 ac->highest_zoneidx, ac->nodemask) {
2465 unsigned long available;
2466
2467 /*
2468 * Do not consider all the reclaimable memory because we do not
2469 * want to trash just for a single high order allocation which
2470 * is even not guaranteed to appear even if __compaction_suitable
2471 * is happy about the watermark check.
2472 */
2473 available = zone_reclaimable_pages(zone) / order;
2474 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
2475 if (__compaction_suitable(zone, order, ac->highest_zoneidx,
2476 available))
2477 return true;
2478 }
2479
2480 return false;
2481}
2482
2483/*
2484 * Should we do compaction for target allocation order.
2485 * Return COMPACT_SUCCESS if allocation for target order can be already
2486 * satisfied
2487 * Return COMPACT_SKIPPED if compaction for target order is likely to fail
2488 * Return COMPACT_CONTINUE if compaction for target order should be ran
2489 */
2490static enum compact_result
2491compaction_suit_allocation_order(struct zone *zone, unsigned int order,
2492 int highest_zoneidx, unsigned int alloc_flags)
2493{
2494 unsigned long watermark;
2495
2496 watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
2497 if (zone_watermark_ok(zone, order, watermark, highest_zoneidx,
2498 alloc_flags))
2499 return COMPACT_SUCCESS;
2500
2501 if (!compaction_suitable(zone, order, highest_zoneidx))
2502 return COMPACT_SKIPPED;
2503
2504 return COMPACT_CONTINUE;
2505}
2506
2507static enum compact_result
2508compact_zone(struct compact_control *cc, struct capture_control *capc)
2509{
2510 enum compact_result ret;
2511 unsigned long start_pfn = cc->zone->zone_start_pfn;
2512 unsigned long end_pfn = zone_end_pfn(cc->zone);
2513 unsigned long last_migrated_pfn;
2514 const bool sync = cc->mode != MIGRATE_ASYNC;
2515 bool update_cached;
2516 unsigned int nr_succeeded = 0, nr_migratepages;
2517 int order;
2518
2519 /*
2520 * These counters track activities during zone compaction. Initialize
2521 * them before compacting a new zone.
2522 */
2523 cc->total_migrate_scanned = 0;
2524 cc->total_free_scanned = 0;
2525 cc->nr_migratepages = 0;
2526 cc->nr_freepages = 0;
2527 for (order = 0; order < NR_PAGE_ORDERS; order++)
2528 INIT_LIST_HEAD(&cc->freepages[order]);
2529 INIT_LIST_HEAD(&cc->migratepages);
2530
2531 cc->migratetype = gfp_migratetype(cc->gfp_mask);
2532
2533 if (!is_via_compact_memory(cc->order)) {
2534 ret = compaction_suit_allocation_order(cc->zone, cc->order,
2535 cc->highest_zoneidx,
2536 cc->alloc_flags);
2537 if (ret != COMPACT_CONTINUE)
2538 return ret;
2539 }
2540
2541 /*
2542 * Clear pageblock skip if there were failures recently and compaction
2543 * is about to be retried after being deferred.
2544 */
2545 if (compaction_restarting(cc->zone, cc->order))
2546 __reset_isolation_suitable(cc->zone);
2547
2548 /*
2549 * Setup to move all movable pages to the end of the zone. Used cached
2550 * information on where the scanners should start (unless we explicitly
2551 * want to compact the whole zone), but check that it is initialised
2552 * by ensuring the values are within zone boundaries.
2553 */
2554 cc->fast_start_pfn = 0;
2555 if (cc->whole_zone) {
2556 cc->migrate_pfn = start_pfn;
2557 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2558 } else {
2559 cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync];
2560 cc->free_pfn = cc->zone->compact_cached_free_pfn;
2561 if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
2562 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2563 cc->zone->compact_cached_free_pfn = cc->free_pfn;
2564 }
2565 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
2566 cc->migrate_pfn = start_pfn;
2567 cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
2568 cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
2569 }
2570
2571 if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn)
2572 cc->whole_zone = true;
2573 }
2574
2575 last_migrated_pfn = 0;
2576
2577 /*
2578 * Migrate has separate cached PFNs for ASYNC and SYNC* migration on
2579 * the basis that some migrations will fail in ASYNC mode. However,
2580 * if the cached PFNs match and pageblocks are skipped due to having
2581 * no isolation candidates, then the sync state does not matter.
2582 * Until a pageblock with isolation candidates is found, keep the
2583 * cached PFNs in sync to avoid revisiting the same blocks.
2584 */
2585 update_cached = !sync &&
2586 cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1];
2587
2588 trace_mm_compaction_begin(cc, start_pfn, end_pfn, sync);
2589
2590 /* lru_add_drain_all could be expensive with involving other CPUs */
2591 lru_add_drain();
2592
2593 while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) {
2594 int err;
2595 unsigned long iteration_start_pfn = cc->migrate_pfn;
2596
2597 /*
2598 * Avoid multiple rescans of the same pageblock which can
2599 * happen if a page cannot be isolated (dirty/writeback in
2600 * async mode) or if the migrated pages are being allocated
2601 * before the pageblock is cleared. The first rescan will
2602 * capture the entire pageblock for migration. If it fails,
2603 * it'll be marked skip and scanning will proceed as normal.
2604 */
2605 cc->finish_pageblock = false;
2606 if (pageblock_start_pfn(last_migrated_pfn) ==
2607 pageblock_start_pfn(iteration_start_pfn)) {
2608 cc->finish_pageblock = true;
2609 }
2610
2611rescan:
2612 switch (isolate_migratepages(cc)) {
2613 case ISOLATE_ABORT:
2614 ret = COMPACT_CONTENDED;
2615 putback_movable_pages(&cc->migratepages);
2616 cc->nr_migratepages = 0;
2617 goto out;
2618 case ISOLATE_NONE:
2619 if (update_cached) {
2620 cc->zone->compact_cached_migrate_pfn[1] =
2621 cc->zone->compact_cached_migrate_pfn[0];
2622 }
2623
2624 /*
2625 * We haven't isolated and migrated anything, but
2626 * there might still be unflushed migrations from
2627 * previous cc->order aligned block.
2628 */
2629 goto check_drain;
2630 case ISOLATE_SUCCESS:
2631 update_cached = false;
2632 last_migrated_pfn = max(cc->zone->zone_start_pfn,
2633 pageblock_start_pfn(cc->migrate_pfn - 1));
2634 }
2635
2636 /*
2637 * Record the number of pages to migrate since the
2638 * compaction_alloc/free() will update cc->nr_migratepages
2639 * properly.
2640 */
2641 nr_migratepages = cc->nr_migratepages;
2642 err = migrate_pages(&cc->migratepages, compaction_alloc,
2643 compaction_free, (unsigned long)cc, cc->mode,
2644 MR_COMPACTION, &nr_succeeded);
2645
2646 trace_mm_compaction_migratepages(nr_migratepages, nr_succeeded);
2647
2648 /* All pages were either migrated or will be released */
2649 cc->nr_migratepages = 0;
2650 if (err) {
2651 putback_movable_pages(&cc->migratepages);
2652 /*
2653 * migrate_pages() may return -ENOMEM when scanners meet
2654 * and we want compact_finished() to detect it
2655 */
2656 if (err == -ENOMEM && !compact_scanners_met(cc)) {
2657 ret = COMPACT_CONTENDED;
2658 goto out;
2659 }
2660 /*
2661 * If an ASYNC or SYNC_LIGHT fails to migrate a page
2662 * within the pageblock_order-aligned block and
2663 * fast_find_migrateblock may be used then scan the
2664 * remainder of the pageblock. This will mark the
2665 * pageblock "skip" to avoid rescanning in the near
2666 * future. This will isolate more pages than necessary
2667 * for the request but avoid loops due to
2668 * fast_find_migrateblock revisiting blocks that were
2669 * recently partially scanned.
2670 */
2671 if (!pageblock_aligned(cc->migrate_pfn) &&
2672 !cc->ignore_skip_hint && !cc->finish_pageblock &&
2673 (cc->mode < MIGRATE_SYNC)) {
2674 cc->finish_pageblock = true;
2675
2676 /*
2677 * Draining pcplists does not help THP if
2678 * any page failed to migrate. Even after
2679 * drain, the pageblock will not be free.
2680 */
2681 if (cc->order == COMPACTION_HPAGE_ORDER)
2682 last_migrated_pfn = 0;
2683
2684 goto rescan;
2685 }
2686 }
2687
2688 /* Stop if a page has been captured */
2689 if (capc && capc->page) {
2690 ret = COMPACT_SUCCESS;
2691 break;
2692 }
2693
2694check_drain:
2695 /*
2696 * Has the migration scanner moved away from the previous
2697 * cc->order aligned block where we migrated from? If yes,
2698 * flush the pages that were freed, so that they can merge and
2699 * compact_finished() can detect immediately if allocation
2700 * would succeed.
2701 */
2702 if (cc->order > 0 && last_migrated_pfn) {
2703 unsigned long current_block_start =
2704 block_start_pfn(cc->migrate_pfn, cc->order);
2705
2706 if (last_migrated_pfn < current_block_start) {
2707 lru_add_drain_cpu_zone(cc->zone);
2708 /* No more flushing until we migrate again */
2709 last_migrated_pfn = 0;
2710 }
2711 }
2712 }
2713
2714out:
2715 /*
2716 * Release free pages and update where the free scanner should restart,
2717 * so we don't leave any returned pages behind in the next attempt.
2718 */
2719 if (cc->nr_freepages > 0) {
2720 unsigned long free_pfn = release_free_list(cc->freepages);
2721
2722 cc->nr_freepages = 0;
2723 VM_BUG_ON(free_pfn == 0);
2724 /* The cached pfn is always the first in a pageblock */
2725 free_pfn = pageblock_start_pfn(free_pfn);
2726 /*
2727 * Only go back, not forward. The cached pfn might have been
2728 * already reset to zone end in compact_finished()
2729 */
2730 if (free_pfn > cc->zone->compact_cached_free_pfn)
2731 cc->zone->compact_cached_free_pfn = free_pfn;
2732 }
2733
2734 count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned);
2735 count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned);
2736
2737 trace_mm_compaction_end(cc, start_pfn, end_pfn, sync, ret);
2738
2739 VM_BUG_ON(!list_empty(&cc->migratepages));
2740
2741 return ret;
2742}
2743
2744static enum compact_result compact_zone_order(struct zone *zone, int order,
2745 gfp_t gfp_mask, enum compact_priority prio,
2746 unsigned int alloc_flags, int highest_zoneidx,
2747 struct page **capture)
2748{
2749 enum compact_result ret;
2750 struct compact_control cc = {
2751 .order = order,
2752 .search_order = order,
2753 .gfp_mask = gfp_mask,
2754 .zone = zone,
2755 .mode = (prio == COMPACT_PRIO_ASYNC) ?
2756 MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT,
2757 .alloc_flags = alloc_flags,
2758 .highest_zoneidx = highest_zoneidx,
2759 .direct_compaction = true,
2760 .whole_zone = (prio == MIN_COMPACT_PRIORITY),
2761 .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
2762 .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
2763 };
2764 struct capture_control capc = {
2765 .cc = &cc,
2766 .page = NULL,
2767 };
2768
2769 /*
2770 * Make sure the structs are really initialized before we expose the
2771 * capture control, in case we are interrupted and the interrupt handler
2772 * frees a page.
2773 */
2774 barrier();
2775 WRITE_ONCE(current->capture_control, &capc);
2776
2777 ret = compact_zone(&cc, &capc);
2778
2779 /*
2780 * Make sure we hide capture control first before we read the captured
2781 * page pointer, otherwise an interrupt could free and capture a page
2782 * and we would leak it.
2783 */
2784 WRITE_ONCE(current->capture_control, NULL);
2785 *capture = READ_ONCE(capc.page);
2786 /*
2787 * Technically, it is also possible that compaction is skipped but
2788 * the page is still captured out of luck(IRQ came and freed the page).
2789 * Returning COMPACT_SUCCESS in such cases helps in properly accounting
2790 * the COMPACT[STALL|FAIL] when compaction is skipped.
2791 */
2792 if (*capture)
2793 ret = COMPACT_SUCCESS;
2794
2795 return ret;
2796}
2797
2798/**
2799 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
2800 * @gfp_mask: The GFP mask of the current allocation
2801 * @order: The order of the current allocation
2802 * @alloc_flags: The allocation flags of the current allocation
2803 * @ac: The context of current allocation
2804 * @prio: Determines how hard direct compaction should try to succeed
2805 * @capture: Pointer to free page created by compaction will be stored here
2806 *
2807 * This is the main entry point for direct page compaction.
2808 */
2809enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
2810 unsigned int alloc_flags, const struct alloc_context *ac,
2811 enum compact_priority prio, struct page **capture)
2812{
2813 struct zoneref *z;
2814 struct zone *zone;
2815 enum compact_result rc = COMPACT_SKIPPED;
2816
2817 if (!gfp_compaction_allowed(gfp_mask))
2818 return COMPACT_SKIPPED;
2819
2820 trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
2821
2822 /* Compact each zone in the list */
2823 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2824 ac->highest_zoneidx, ac->nodemask) {
2825 enum compact_result status;
2826
2827 if (cpusets_enabled() &&
2828 (alloc_flags & ALLOC_CPUSET) &&
2829 !__cpuset_zone_allowed(zone, gfp_mask))
2830 continue;
2831
2832 if (prio > MIN_COMPACT_PRIORITY
2833 && compaction_deferred(zone, order)) {
2834 rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
2835 continue;
2836 }
2837
2838 status = compact_zone_order(zone, order, gfp_mask, prio,
2839 alloc_flags, ac->highest_zoneidx, capture);
2840 rc = max(status, rc);
2841
2842 /* The allocation should succeed, stop compacting */
2843 if (status == COMPACT_SUCCESS) {
2844 /*
2845 * We think the allocation will succeed in this zone,
2846 * but it is not certain, hence the false. The caller
2847 * will repeat this with true if allocation indeed
2848 * succeeds in this zone.
2849 */
2850 compaction_defer_reset(zone, order, false);
2851
2852 break;
2853 }
2854
2855 if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
2856 status == COMPACT_PARTIAL_SKIPPED))
2857 /*
2858 * We think that allocation won't succeed in this zone
2859 * so we defer compaction there. If it ends up
2860 * succeeding after all, it will be reset.
2861 */
2862 defer_compaction(zone, order);
2863
2864 /*
2865 * We might have stopped compacting due to need_resched() in
2866 * async compaction, or due to a fatal signal detected. In that
2867 * case do not try further zones
2868 */
2869 if ((prio == COMPACT_PRIO_ASYNC && need_resched())
2870 || fatal_signal_pending(current))
2871 break;
2872 }
2873
2874 return rc;
2875}
2876
2877/*
2878 * compact_node() - compact all zones within a node
2879 * @pgdat: The node page data
2880 * @proactive: Whether the compaction is proactive
2881 *
2882 * For proactive compaction, compact till each zone's fragmentation score
2883 * reaches within proactive compaction thresholds (as determined by the
2884 * proactiveness tunable), it is possible that the function returns before
2885 * reaching score targets due to various back-off conditions, such as,
2886 * contention on per-node or per-zone locks.
2887 */
2888static int compact_node(pg_data_t *pgdat, bool proactive)
2889{
2890 int zoneid;
2891 struct zone *zone;
2892 struct compact_control cc = {
2893 .order = -1,
2894 .mode = proactive ? MIGRATE_SYNC_LIGHT : MIGRATE_SYNC,
2895 .ignore_skip_hint = true,
2896 .whole_zone = true,
2897 .gfp_mask = GFP_KERNEL,
2898 .proactive_compaction = proactive,
2899 };
2900
2901 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2902 zone = &pgdat->node_zones[zoneid];
2903 if (!populated_zone(zone))
2904 continue;
2905
2906 if (fatal_signal_pending(current))
2907 return -EINTR;
2908
2909 cc.zone = zone;
2910
2911 compact_zone(&cc, NULL);
2912
2913 if (proactive) {
2914 count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
2915 cc.total_migrate_scanned);
2916 count_compact_events(KCOMPACTD_FREE_SCANNED,
2917 cc.total_free_scanned);
2918 }
2919 }
2920
2921 return 0;
2922}
2923
2924/* Compact all zones of all nodes in the system */
2925static int compact_nodes(void)
2926{
2927 int ret, nid;
2928
2929 /* Flush pending updates to the LRU lists */
2930 lru_add_drain_all();
2931
2932 for_each_online_node(nid) {
2933 ret = compact_node(NODE_DATA(nid), false);
2934 if (ret)
2935 return ret;
2936 }
2937
2938 return 0;
2939}
2940
2941static int compaction_proactiveness_sysctl_handler(const struct ctl_table *table, int write,
2942 void *buffer, size_t *length, loff_t *ppos)
2943{
2944 int rc, nid;
2945
2946 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
2947 if (rc)
2948 return rc;
2949
2950 if (write && sysctl_compaction_proactiveness) {
2951 for_each_online_node(nid) {
2952 pg_data_t *pgdat = NODE_DATA(nid);
2953
2954 if (pgdat->proactive_compact_trigger)
2955 continue;
2956
2957 pgdat->proactive_compact_trigger = true;
2958 trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, -1,
2959 pgdat->nr_zones - 1);
2960 wake_up_interruptible(&pgdat->kcompactd_wait);
2961 }
2962 }
2963
2964 return 0;
2965}
2966
2967/*
2968 * This is the entry point for compacting all nodes via
2969 * /proc/sys/vm/compact_memory
2970 */
2971static int sysctl_compaction_handler(const struct ctl_table *table, int write,
2972 void *buffer, size_t *length, loff_t *ppos)
2973{
2974 int ret;
2975
2976 ret = proc_dointvec(table, write, buffer, length, ppos);
2977 if (ret)
2978 return ret;
2979
2980 if (sysctl_compact_memory != 1)
2981 return -EINVAL;
2982
2983 if (write)
2984 ret = compact_nodes();
2985
2986 return ret;
2987}
2988
2989#if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
2990static ssize_t compact_store(struct device *dev,
2991 struct device_attribute *attr,
2992 const char *buf, size_t count)
2993{
2994 int nid = dev->id;
2995
2996 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
2997 /* Flush pending updates to the LRU lists */
2998 lru_add_drain_all();
2999
3000 compact_node(NODE_DATA(nid), false);
3001 }
3002
3003 return count;
3004}
3005static DEVICE_ATTR_WO(compact);
3006
3007int compaction_register_node(struct node *node)
3008{
3009 return device_create_file(&node->dev, &dev_attr_compact);
3010}
3011
3012void compaction_unregister_node(struct node *node)
3013{
3014 device_remove_file(&node->dev, &dev_attr_compact);
3015}
3016#endif /* CONFIG_SYSFS && CONFIG_NUMA */
3017
3018static inline bool kcompactd_work_requested(pg_data_t *pgdat)
3019{
3020 return pgdat->kcompactd_max_order > 0 || kthread_should_stop() ||
3021 pgdat->proactive_compact_trigger;
3022}
3023
3024static bool kcompactd_node_suitable(pg_data_t *pgdat)
3025{
3026 int zoneid;
3027 struct zone *zone;
3028 enum zone_type highest_zoneidx = pgdat->kcompactd_highest_zoneidx;
3029 enum compact_result ret;
3030
3031 for (zoneid = 0; zoneid <= highest_zoneidx; zoneid++) {
3032 zone = &pgdat->node_zones[zoneid];
3033
3034 if (!populated_zone(zone))
3035 continue;
3036
3037 ret = compaction_suit_allocation_order(zone,
3038 pgdat->kcompactd_max_order,
3039 highest_zoneidx, ALLOC_WMARK_MIN);
3040 if (ret == COMPACT_CONTINUE)
3041 return true;
3042 }
3043
3044 return false;
3045}
3046
3047static void kcompactd_do_work(pg_data_t *pgdat)
3048{
3049 /*
3050 * With no special task, compact all zones so that a page of requested
3051 * order is allocatable.
3052 */
3053 int zoneid;
3054 struct zone *zone;
3055 struct compact_control cc = {
3056 .order = pgdat->kcompactd_max_order,
3057 .search_order = pgdat->kcompactd_max_order,
3058 .highest_zoneidx = pgdat->kcompactd_highest_zoneidx,
3059 .mode = MIGRATE_SYNC_LIGHT,
3060 .ignore_skip_hint = false,
3061 .gfp_mask = GFP_KERNEL,
3062 };
3063 enum compact_result ret;
3064
3065 trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
3066 cc.highest_zoneidx);
3067 count_compact_event(KCOMPACTD_WAKE);
3068
3069 for (zoneid = 0; zoneid <= cc.highest_zoneidx; zoneid++) {
3070 int status;
3071
3072 zone = &pgdat->node_zones[zoneid];
3073 if (!populated_zone(zone))
3074 continue;
3075
3076 if (compaction_deferred(zone, cc.order))
3077 continue;
3078
3079 ret = compaction_suit_allocation_order(zone,
3080 cc.order, zoneid, ALLOC_WMARK_MIN);
3081 if (ret != COMPACT_CONTINUE)
3082 continue;
3083
3084 if (kthread_should_stop())
3085 return;
3086
3087 cc.zone = zone;
3088 status = compact_zone(&cc, NULL);
3089
3090 if (status == COMPACT_SUCCESS) {
3091 compaction_defer_reset(zone, cc.order, false);
3092 } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
3093 /*
3094 * Buddy pages may become stranded on pcps that could
3095 * otherwise coalesce on the zone's free area for
3096 * order >= cc.order. This is ratelimited by the
3097 * upcoming deferral.
3098 */
3099 drain_all_pages(zone);
3100
3101 /*
3102 * We use sync migration mode here, so we defer like
3103 * sync direct compaction does.
3104 */
3105 defer_compaction(zone, cc.order);
3106 }
3107
3108 count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
3109 cc.total_migrate_scanned);
3110 count_compact_events(KCOMPACTD_FREE_SCANNED,
3111 cc.total_free_scanned);
3112 }
3113
3114 /*
3115 * Regardless of success, we are done until woken up next. But remember
3116 * the requested order/highest_zoneidx in case it was higher/tighter
3117 * than our current ones
3118 */
3119 if (pgdat->kcompactd_max_order <= cc.order)
3120 pgdat->kcompactd_max_order = 0;
3121 if (pgdat->kcompactd_highest_zoneidx >= cc.highest_zoneidx)
3122 pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
3123}
3124
3125void wakeup_kcompactd(pg_data_t *pgdat, int order, int highest_zoneidx)
3126{
3127 if (!order)
3128 return;
3129
3130 if (pgdat->kcompactd_max_order < order)
3131 pgdat->kcompactd_max_order = order;
3132
3133 if (pgdat->kcompactd_highest_zoneidx > highest_zoneidx)
3134 pgdat->kcompactd_highest_zoneidx = highest_zoneidx;
3135
3136 /*
3137 * Pairs with implicit barrier in wait_event_freezable()
3138 * such that wakeups are not missed.
3139 */
3140 if (!wq_has_sleeper(&pgdat->kcompactd_wait))
3141 return;
3142
3143 if (!kcompactd_node_suitable(pgdat))
3144 return;
3145
3146 trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
3147 highest_zoneidx);
3148 wake_up_interruptible(&pgdat->kcompactd_wait);
3149}
3150
3151/*
3152 * The background compaction daemon, started as a kernel thread
3153 * from the init process.
3154 */
3155static int kcompactd(void *p)
3156{
3157 pg_data_t *pgdat = (pg_data_t *)p;
3158 struct task_struct *tsk = current;
3159 long default_timeout = msecs_to_jiffies(HPAGE_FRAG_CHECK_INTERVAL_MSEC);
3160 long timeout = default_timeout;
3161
3162 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3163
3164 if (!cpumask_empty(cpumask))
3165 set_cpus_allowed_ptr(tsk, cpumask);
3166
3167 current->flags |= PF_KCOMPACTD;
3168 set_freezable();
3169
3170 pgdat->kcompactd_max_order = 0;
3171 pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
3172
3173 while (!kthread_should_stop()) {
3174 unsigned long pflags;
3175
3176 /*
3177 * Avoid the unnecessary wakeup for proactive compaction
3178 * when it is disabled.
3179 */
3180 if (!sysctl_compaction_proactiveness)
3181 timeout = MAX_SCHEDULE_TIMEOUT;
3182 trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
3183 if (wait_event_freezable_timeout(pgdat->kcompactd_wait,
3184 kcompactd_work_requested(pgdat), timeout) &&
3185 !pgdat->proactive_compact_trigger) {
3186
3187 psi_memstall_enter(&pflags);
3188 kcompactd_do_work(pgdat);
3189 psi_memstall_leave(&pflags);
3190 /*
3191 * Reset the timeout value. The defer timeout from
3192 * proactive compaction is lost here but that is fine
3193 * as the condition of the zone changing substantionally
3194 * then carrying on with the previous defer interval is
3195 * not useful.
3196 */
3197 timeout = default_timeout;
3198 continue;
3199 }
3200
3201 /*
3202 * Start the proactive work with default timeout. Based
3203 * on the fragmentation score, this timeout is updated.
3204 */
3205 timeout = default_timeout;
3206 if (should_proactive_compact_node(pgdat)) {
3207 unsigned int prev_score, score;
3208
3209 prev_score = fragmentation_score_node(pgdat);
3210 compact_node(pgdat, true);
3211 score = fragmentation_score_node(pgdat);
3212 /*
3213 * Defer proactive compaction if the fragmentation
3214 * score did not go down i.e. no progress made.
3215 */
3216 if (unlikely(score >= prev_score))
3217 timeout =
3218 default_timeout << COMPACT_MAX_DEFER_SHIFT;
3219 }
3220 if (unlikely(pgdat->proactive_compact_trigger))
3221 pgdat->proactive_compact_trigger = false;
3222 }
3223
3224 current->flags &= ~PF_KCOMPACTD;
3225
3226 return 0;
3227}
3228
3229/*
3230 * This kcompactd start function will be called by init and node-hot-add.
3231 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
3232 */
3233void __meminit kcompactd_run(int nid)
3234{
3235 pg_data_t *pgdat = NODE_DATA(nid);
3236
3237 if (pgdat->kcompactd)
3238 return;
3239
3240 pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
3241 if (IS_ERR(pgdat->kcompactd)) {
3242 pr_err("Failed to start kcompactd on node %d\n", nid);
3243 pgdat->kcompactd = NULL;
3244 }
3245}
3246
3247/*
3248 * Called by memory hotplug when all memory in a node is offlined. Caller must
3249 * be holding mem_hotplug_begin/done().
3250 */
3251void __meminit kcompactd_stop(int nid)
3252{
3253 struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
3254
3255 if (kcompactd) {
3256 kthread_stop(kcompactd);
3257 NODE_DATA(nid)->kcompactd = NULL;
3258 }
3259}
3260
3261/*
3262 * It's optimal to keep kcompactd on the same CPUs as their memory, but
3263 * not required for correctness. So if the last cpu in a node goes
3264 * away, we get changed to run anywhere: as the first one comes back,
3265 * restore their cpu bindings.
3266 */
3267static int kcompactd_cpu_online(unsigned int cpu)
3268{
3269 int nid;
3270
3271 for_each_node_state(nid, N_MEMORY) {
3272 pg_data_t *pgdat = NODE_DATA(nid);
3273 const struct cpumask *mask;
3274
3275 mask = cpumask_of_node(pgdat->node_id);
3276
3277 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3278 /* One of our CPUs online: restore mask */
3279 if (pgdat->kcompactd)
3280 set_cpus_allowed_ptr(pgdat->kcompactd, mask);
3281 }
3282 return 0;
3283}
3284
3285static int proc_dointvec_minmax_warn_RT_change(const struct ctl_table *table,
3286 int write, void *buffer, size_t *lenp, loff_t *ppos)
3287{
3288 int ret, old;
3289
3290 if (!IS_ENABLED(CONFIG_PREEMPT_RT) || !write)
3291 return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
3292
3293 old = *(int *)table->data;
3294 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
3295 if (ret)
3296 return ret;
3297 if (old != *(int *)table->data)
3298 pr_warn_once("sysctl attribute %s changed by %s[%d]\n",
3299 table->procname, current->comm,
3300 task_pid_nr(current));
3301 return ret;
3302}
3303
3304static struct ctl_table vm_compaction[] = {
3305 {
3306 .procname = "compact_memory",
3307 .data = &sysctl_compact_memory,
3308 .maxlen = sizeof(int),
3309 .mode = 0200,
3310 .proc_handler = sysctl_compaction_handler,
3311 },
3312 {
3313 .procname = "compaction_proactiveness",
3314 .data = &sysctl_compaction_proactiveness,
3315 .maxlen = sizeof(sysctl_compaction_proactiveness),
3316 .mode = 0644,
3317 .proc_handler = compaction_proactiveness_sysctl_handler,
3318 .extra1 = SYSCTL_ZERO,
3319 .extra2 = SYSCTL_ONE_HUNDRED,
3320 },
3321 {
3322 .procname = "extfrag_threshold",
3323 .data = &sysctl_extfrag_threshold,
3324 .maxlen = sizeof(int),
3325 .mode = 0644,
3326 .proc_handler = proc_dointvec_minmax,
3327 .extra1 = SYSCTL_ZERO,
3328 .extra2 = SYSCTL_ONE_THOUSAND,
3329 },
3330 {
3331 .procname = "compact_unevictable_allowed",
3332 .data = &sysctl_compact_unevictable_allowed,
3333 .maxlen = sizeof(int),
3334 .mode = 0644,
3335 .proc_handler = proc_dointvec_minmax_warn_RT_change,
3336 .extra1 = SYSCTL_ZERO,
3337 .extra2 = SYSCTL_ONE,
3338 },
3339};
3340
3341static int __init kcompactd_init(void)
3342{
3343 int nid;
3344 int ret;
3345
3346 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3347 "mm/compaction:online",
3348 kcompactd_cpu_online, NULL);
3349 if (ret < 0) {
3350 pr_err("kcompactd: failed to register hotplug callbacks.\n");
3351 return ret;
3352 }
3353
3354 for_each_node_state(nid, N_MEMORY)
3355 kcompactd_run(nid);
3356 register_sysctl_init("vm", vm_compaction);
3357 return 0;
3358}
3359subsys_initcall(kcompactd_init)
3360
3361#endif /* CONFIG_COMPACTION */