1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * linux/mm/compaction.c
4 *
5 * Memory compaction for the reduction of external fragmentation. Note that
6 * this heavily depends upon page migration to do all the real heavy
7 * lifting
8 *
9 * Copyright IBM Corp. 2007-2010 Mel Gorman <mel@csn.ul.ie>
10 */
11 #include <linux/cpu.h>
12 #include <linux/swap.h>
13 #include <linux/migrate.h>
14 #include <linux/compaction.h>
15 #include <linux/mm_inline.h>
16 #include <linux/sched/signal.h>
17 #include <linux/backing-dev.h>
18 #include <linux/sysctl.h>
19 #include <linux/sysfs.h>
20 #include <linux/page-isolation.h>
21 #include <linux/kasan.h>
22 #include <linux/kthread.h>
23 #include <linux/freezer.h>
24 #include <linux/page_owner.h>
25 #include <linux/psi.h>
26 #include "internal.h"
27
28 #ifdef CONFIG_COMPACTION
29 /*
30 * Fragmentation score check interval for proactive compaction purposes.
31 */
32 #define HPAGE_FRAG_CHECK_INTERVAL_MSEC (500)
33
34 static inline void count_compact_event(enum vm_event_item item)
35 {
36 count_vm_event(item);
37 }
38
39 static inline void count_compact_events(enum vm_event_item item, long delta)
40 {
41 count_vm_events(item, delta);
42 }
43
44 /*
45 * order == -1 is expected when compacting proactively via
46 * 1. /proc/sys/vm/compact_memory
47 * 2. /sys/devices/system/node/nodex/compact
48 * 3. /proc/sys/vm/compaction_proactiveness
49 */
50 static inline bool is_via_compact_memory(int order)
51 {
52 return order == -1;
53 }
54
55 #else
56 #define count_compact_event(item) do { } while (0)
57 #define count_compact_events(item, delta) do { } while (0)
58 static inline bool is_via_compact_memory(int order) { return false; }
59 #endif
60
61 #if defined CONFIG_COMPACTION || defined CONFIG_CMA
62
63 #define CREATE_TRACE_POINTS
64 #include <trace/events/compaction.h>
65
66 #define block_start_pfn(pfn, order) round_down(pfn, 1UL << (order))
67 #define block_end_pfn(pfn, order) ALIGN((pfn) + 1, 1UL << (order))
68
69 /*
70 * Page order with-respect-to which proactive compaction
71 * calculates external fragmentation, which is used as
72 * the "fragmentation score" of a node/zone.
73 */
74 #if defined CONFIG_TRANSPARENT_HUGEPAGE
75 #define COMPACTION_HPAGE_ORDER HPAGE_PMD_ORDER
76 #elif defined CONFIG_HUGETLBFS
77 #define COMPACTION_HPAGE_ORDER HUGETLB_PAGE_ORDER
78 #else
79 #define COMPACTION_HPAGE_ORDER (PMD_SHIFT - PAGE_SHIFT)
80 #endif
81
82 static struct page *mark_allocated_noprof(struct page *page, unsigned int order, gfp_t gfp_flags)
83 {
84 post_alloc_hook(page, order, __GFP_MOVABLE);
85 return page;
86 }
87 #define mark_allocated(...) alloc_hooks(mark_allocated_noprof(__VA_ARGS__))
88
89 static void split_map_pages(struct list_head *freepages)
90 {
91 unsigned int i, order;
92 struct page *page, *next;
93 LIST_HEAD(tmp_list);
94
95 for (order = 0; order < NR_PAGE_ORDERS; order++) {
96 list_for_each_entry_safe(page, next, &freepages[order], lru) {
97 unsigned int nr_pages;
98
99 list_del(&page->lru);
100
101 nr_pages = 1 << order;
102
103 mark_allocated(page, order, __GFP_MOVABLE);
104 if (order)
105 split_page(page, order);
106
107 for (i = 0; i < nr_pages; i++) {
108 list_add(&page->lru, &tmp_list);
109 page++;
110 }
111 }
112 list_splice_init(&tmp_list, &freepages[0]);
113 }
114 }
115
116 static unsigned long release_free_list(struct list_head *freepages)
117 {
118 int order;
119 unsigned long high_pfn = 0;
120
121 for (order = 0; order < NR_PAGE_ORDERS; order++) {
122 struct page *page, *next;
123
124 list_for_each_entry_safe(page, next, &freepages[order], lru) {
125 unsigned long pfn = page_to_pfn(page);
126
127 list_del(&page->lru);
128 /*
129 * Convert free pages into post allocation pages, so
130 * that we can free them via __free_page.
131 */
132 mark_allocated(page, order, __GFP_MOVABLE);
133 __free_pages(page, order);
134 if (pfn > high_pfn)
135 high_pfn = pfn;
136 }
137 }
138 return high_pfn;
139 }
140
141 #ifdef CONFIG_COMPACTION
142 bool PageMovable(struct page *page)
143 {
144 const struct movable_operations *mops;
145
146 VM_BUG_ON_PAGE(!PageLocked(page), page);
147 if (!__PageMovable(page))
148 return false;
149
150 mops = page_movable_ops(page);
151 if (mops)
152 return true;
153
154 return false;
155 }
156
157 void __SetPageMovable(struct page *page, const struct movable_operations *mops)
158 {
159 VM_BUG_ON_PAGE(!PageLocked(page), page);
160 VM_BUG_ON_PAGE((unsigned long)mops & PAGE_MAPPING_MOVABLE, page);
161 page->mapping = (void *)((unsigned long)mops | PAGE_MAPPING_MOVABLE);
162 }
163 EXPORT_SYMBOL(__SetPageMovable);
164
165 void __ClearPageMovable(struct page *page)
166 {
167 VM_BUG_ON_PAGE(!PageMovable(page), page);
168 /*
169 * This page still has the type of a movable page, but it's
170 * actually not movable any more.
171 */
172 page->mapping = (void *)PAGE_MAPPING_MOVABLE;
173 }
174 EXPORT_SYMBOL(__ClearPageMovable);
175
176 /* Do not skip compaction more than 64 times */
177 #define COMPACT_MAX_DEFER_SHIFT 6
178
179 /*
180 * Compaction is deferred when compaction fails to result in a page
181 * allocation success. 1 << compact_defer_shift, compactions are skipped up
182 * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
183 */
184 static void defer_compaction(struct zone *zone, int order)
185 {
186 zone->compact_considered = 0;
187 zone->compact_defer_shift++;
188
189 if (order < zone->compact_order_failed)
190 zone->compact_order_failed = order;
191
192 if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
193 zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
194
195 trace_mm_compaction_defer_compaction(zone, order);
196 }
197
198 /* Returns true if compaction should be skipped this time */
199 static bool compaction_deferred(struct zone *zone, int order)
200 {
201 unsigned long defer_limit = 1UL << zone->compact_defer_shift;
202
203 if (order < zone->compact_order_failed)
204 return false;
205
206 /* Avoid possible overflow */
207 if (++zone->compact_considered >= defer_limit) {
208 zone->compact_considered = defer_limit;
209 return false;
210 }
211
212 trace_mm_compaction_deferred(zone, order);
213
214 return true;
215 }
216
217 /*
218 * Update defer tracking counters after successful compaction of given order,
219 * which means an allocation either succeeded (alloc_success == true) or is
220 * expected to succeed.
221 */
222 void compaction_defer_reset(struct zone *zone, int order,
223 bool alloc_success)
224 {
225 if (alloc_success) {
226 zone->compact_considered = 0;
227 zone->compact_defer_shift = 0;
228 }
229 if (order >= zone->compact_order_failed)
230 zone->compact_order_failed = order + 1;
231
232 trace_mm_compaction_defer_reset(zone, order);
233 }
234
235 /* Returns true if restarting compaction after many failures */
236 static bool compaction_restarting(struct zone *zone, int order)
237 {
238 if (order < zone->compact_order_failed)
239 return false;
240
241 return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
242 zone->compact_considered >= 1UL << zone->compact_defer_shift;
243 }
244
245 /* Returns true if the pageblock should be scanned for pages to isolate. */
246 static inline bool isolation_suitable(struct compact_control *cc,
247 struct page *page)
248 {
249 if (cc->ignore_skip_hint)
250 return true;
251
252 return !get_pageblock_skip(page);
253 }
254
255 static void reset_cached_positions(struct zone *zone)
256 {
257 zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
258 zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
259 zone->compact_cached_free_pfn =
260 pageblock_start_pfn(zone_end_pfn(zone) - 1);
261 }
262
263 #ifdef CONFIG_SPARSEMEM
264 /*
265 * If the PFN falls into an offline section, return the start PFN of the
266 * next online section. If the PFN falls into an online section or if
267 * there is no next online section, return 0.
268 */
269 static unsigned long skip_offline_sections(unsigned long start_pfn)
270 {
271 unsigned long start_nr = pfn_to_section_nr(start_pfn);
272
273 if (online_section_nr(start_nr))
274 return 0;
275
276 while (++start_nr <= __highest_present_section_nr) {
277 if (online_section_nr(start_nr))
278 return section_nr_to_pfn(start_nr);
279 }
280
281 return 0;
282 }
283
284 /*
285 * If the PFN falls into an offline section, return the end PFN of the
286 * next online section in reverse. If the PFN falls into an online section
287 * or if there is no next online section in reverse, return 0.
288 */
289 static unsigned long skip_offline_sections_reverse(unsigned long start_pfn)
290 {
291 unsigned long start_nr = pfn_to_section_nr(start_pfn);
292
293 if (!start_nr || online_section_nr(start_nr))
294 return 0;
295
296 while (start_nr-- > 0) {
297 if (online_section_nr(start_nr))
298 return section_nr_to_pfn(start_nr) + PAGES_PER_SECTION;
299 }
300
301 return 0;
302 }
303 #else
304 static unsigned long skip_offline_sections(unsigned long start_pfn)
305 {
306 return 0;
307 }
308
309 static unsigned long skip_offline_sections_reverse(unsigned long start_pfn)
310 {
311 return 0;
312 }
313 #endif
314
315 /*
316 * Compound pages of >= pageblock_order should consistently be skipped until
317 * released. It is always pointless to compact pages of such order (if they are
318 * migratable), and the pageblocks they occupy cannot contain any free pages.
319 */
320 static bool pageblock_skip_persistent(struct page *page)
321 {
322 if (!PageCompound(page))
323 return false;
324
325 page = compound_head(page);
326
327 if (compound_order(page) >= pageblock_order)
328 return true;
329
330 return false;
331 }
332
333 static bool
334 __reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source,
335 bool check_target)
336 {
337 struct page *page = pfn_to_online_page(pfn);
338 struct page *block_page;
339 struct page *end_page;
340 unsigned long block_pfn;
341
342 if (!page)
343 return false;
344 if (zone != page_zone(page))
345 return false;
346 if (pageblock_skip_persistent(page))
347 return false;
348
349 /*
350 * If skip is already cleared do no further checking once the
351 * restart points have been set.
352 */
353 if (check_source && check_target && !get_pageblock_skip(page))
354 return true;
355
356 /*
357 * If clearing skip for the target scanner, do not select a
358 * non-movable pageblock as the starting point.
359 */
360 if (!check_source && check_target &&
361 get_pageblock_migratetype(page) != MIGRATE_MOVABLE)
362 return false;
363
364 /* Ensure the start of the pageblock or zone is online and valid */
365 block_pfn = pageblock_start_pfn(pfn);
366 block_pfn = max(block_pfn, zone->zone_start_pfn);
367 block_page = pfn_to_online_page(block_pfn);
368 if (block_page) {
369 page = block_page;
370 pfn = block_pfn;
371 }
372
373 /* Ensure the end of the pageblock or zone is online and valid */
374 block_pfn = pageblock_end_pfn(pfn) - 1;
375 block_pfn = min(block_pfn, zone_end_pfn(zone) - 1);
376 end_page = pfn_to_online_page(block_pfn);
377 if (!end_page)
378 return false;
379
380 /*
381 * Only clear the hint if a sample indicates there is either a
382 * free page or an LRU page in the block. One or other condition
383 * is necessary for the block to be a migration source/target.
384 */
385 do {
386 if (check_source && PageLRU(page)) {
387 clear_pageblock_skip(page);
388 return true;
389 }
390
391 if (check_target && PageBuddy(page)) {
392 clear_pageblock_skip(page);
393 return true;
394 }
395
396 page += (1 << PAGE_ALLOC_COSTLY_ORDER);
397 } while (page <= end_page);
398
399 return false;
400 }
401
402 /*
403 * This function is called to clear all cached information on pageblocks that
404 * should be skipped for page isolation when the migrate and free page scanner
405 * meet.
406 */
407 static void __reset_isolation_suitable(struct zone *zone)
408 {
409 unsigned long migrate_pfn = zone->zone_start_pfn;
410 unsigned long free_pfn = zone_end_pfn(zone) - 1;
411 unsigned long reset_migrate = free_pfn;
412 unsigned long reset_free = migrate_pfn;
413 bool source_set = false;
414 bool free_set = false;
415
416 /* Only flush if a full compaction finished recently */
417 if (!zone->compact_blockskip_flush)
418 return;
419
420 zone->compact_blockskip_flush = false;
421
422 /*
423 * Walk the zone and update pageblock skip information. Source looks
424 * for PageLRU while target looks for PageBuddy. When the scanner
425 * is found, both PageBuddy and PageLRU are checked as the pageblock
426 * is suitable as both source and target.
427 */
428 for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages,
429 free_pfn -= pageblock_nr_pages) {
430 cond_resched();
431
432 /* Update the migrate PFN */
433 if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) &&
434 migrate_pfn < reset_migrate) {
435 source_set = true;
436 reset_migrate = migrate_pfn;
437 zone->compact_init_migrate_pfn = reset_migrate;
438 zone->compact_cached_migrate_pfn[0] = reset_migrate;
439 zone->compact_cached_migrate_pfn[1] = reset_migrate;
440 }
441
442 /* Update the free PFN */
443 if (__reset_isolation_pfn(zone, free_pfn, free_set, true) &&
444 free_pfn > reset_free) {
445 free_set = true;
446 reset_free = free_pfn;
447 zone->compact_init_free_pfn = reset_free;
448 zone->compact_cached_free_pfn = reset_free;
449 }
450 }
451
452 /* Leave no distance if no suitable block was reset */
453 if (reset_migrate >= reset_free) {
454 zone->compact_cached_migrate_pfn[0] = migrate_pfn;
455 zone->compact_cached_migrate_pfn[1] = migrate_pfn;
456 zone->compact_cached_free_pfn = free_pfn;
457 }
458 }
459
460 void reset_isolation_suitable(pg_data_t *pgdat)
461 {
462 int zoneid;
463
464 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
465 struct zone *zone = &pgdat->node_zones[zoneid];
466 if (!populated_zone(zone))
467 continue;
468
469 __reset_isolation_suitable(zone);
470 }
471 }
472
473 /*
474 * Sets the pageblock skip bit if it was clear. Note that this is a hint as
475 * locks are not required for read/writers. Returns true if it was already set.
476 */
477 static bool test_and_set_skip(struct compact_control *cc, struct page *page)
478 {
479 bool skip;
480
481 /* Do not update if skip hint is being ignored */
482 if (cc->ignore_skip_hint)
483 return false;
484
485 skip = get_pageblock_skip(page);
486 if (!skip && !cc->no_set_skip_hint)
487 set_pageblock_skip(page);
488
489 return skip;
490 }
491
492 static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
493 {
494 struct zone *zone = cc->zone;
495
496 /* Set for isolation rather than compaction */
497 if (cc->no_set_skip_hint)
498 return;
499
500 pfn = pageblock_end_pfn(pfn);
501
502 /* Update where async and sync compaction should restart */
503 if (pfn > zone->compact_cached_migrate_pfn[0])
504 zone->compact_cached_migrate_pfn[0] = pfn;
505 if (cc->mode != MIGRATE_ASYNC &&
506 pfn > zone->compact_cached_migrate_pfn[1])
507 zone->compact_cached_migrate_pfn[1] = pfn;
508 }
509
510 /*
511 * If no pages were isolated then mark this pageblock to be skipped in the
512 * future. The information is later cleared by __reset_isolation_suitable().
513 */
514 static void update_pageblock_skip(struct compact_control *cc,
515 struct page *page, unsigned long pfn)
516 {
517 struct zone *zone = cc->zone;
518
519 if (cc->no_set_skip_hint)
520 return;
521
522 set_pageblock_skip(page);
523
524 if (pfn < zone->compact_cached_free_pfn)
525 zone->compact_cached_free_pfn = pfn;
526 }
527 #else
528 static inline bool isolation_suitable(struct compact_control *cc,
529 struct page *page)
530 {
531 return true;
532 }
533
534 static inline bool pageblock_skip_persistent(struct page *page)
535 {
536 return false;
537 }
538
539 static inline void update_pageblock_skip(struct compact_control *cc,
540 struct page *page, unsigned long pfn)
541 {
542 }
543
544 static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
545 {
546 }
547
548 static bool test_and_set_skip(struct compact_control *cc, struct page *page)
549 {
550 return false;
551 }
552 #endif /* CONFIG_COMPACTION */
553
554 /*
555 * Compaction requires the taking of some coarse locks that are potentially
556 * very heavily contended. For async compaction, trylock and record if the
557 * lock is contended. The lock will still be acquired but compaction will
558 * abort when the current block is finished regardless of success rate.
559 * Sync compaction acquires the lock.
560 *
561 * Always returns true which makes it easier to track lock state in callers.
562 */
563 static bool compact_lock_irqsave(spinlock_t *lock, unsigned long *flags,
564 struct compact_control *cc)
565 __acquires(lock)
566 {
567 /* Track if the lock is contended in async mode */
568 if (cc->mode == MIGRATE_ASYNC && !cc->contended) {
569 if (spin_trylock_irqsave(lock, *flags))
570 return true;
571
572 cc->contended = true;
573 }
574
575 spin_lock_irqsave(lock, *flags);
576 return true;
577 }
578
579 /*
580 * Compaction requires the taking of some coarse locks that are potentially
581 * very heavily contended. The lock should be periodically unlocked to avoid
582 * having disabled IRQs for a long time, even when there is nobody waiting on
583 * the lock. It might also be that allowing the IRQs will result in
584 * need_resched() becoming true. If scheduling is needed, compaction schedules.
585 * Either compaction type will also abort if a fatal signal is pending.
586 * In either case if the lock was locked, it is dropped and not regained.
587 *
588 * Returns true if compaction should abort due to fatal signal pending.
589 * Returns false when compaction can continue.
590 */
591 static bool compact_unlock_should_abort(spinlock_t *lock,
592 unsigned long flags, bool *locked, struct compact_control *cc)
593 {
594 if (*locked) {
595 spin_unlock_irqrestore(lock, flags);
596 *locked = false;
597 }
598
599 if (fatal_signal_pending(current)) {
600 cc->contended = true;
601 return true;
602 }
603
604 cond_resched();
605
606 return false;
607 }
608
609 /*
610 * Isolate free pages onto a private freelist. If @strict is true, will abort
611 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
612 * (even though it may still end up isolating some pages).
613 */
614 static unsigned long isolate_freepages_block(struct compact_control *cc,
615 unsigned long *start_pfn,
616 unsigned long end_pfn,
617 struct list_head *freelist,
618 unsigned int stride,
619 bool strict)
620 {
621 int nr_scanned = 0, total_isolated = 0;
622 struct page *page;
623 unsigned long flags = 0;
624 bool locked = false;
625 unsigned long blockpfn = *start_pfn;
626 unsigned int order;
627
628 /* Strict mode is for isolation, speed is secondary */
629 if (strict)
630 stride = 1;
631
632 page = pfn_to_page(blockpfn);
633
634 /* Isolate free pages. */
635 for (; blockpfn < end_pfn; blockpfn += stride, page += stride) {
636 int isolated;
637
638 /*
639 * Periodically drop the lock (if held) regardless of its
640 * contention, to give chance to IRQs. Abort if fatal signal
641 * pending.
642 */
643 if (!(blockpfn % COMPACT_CLUSTER_MAX)
644 && compact_unlock_should_abort(&cc->zone->lock, flags,
645 &locked, cc))
646 break;
647
648 nr_scanned++;
649
650 /*
651 * For compound pages such as THP and hugetlbfs, we can save
652 * potentially a lot of iterations if we skip them at once.
653 * The check is racy, but we can consider only valid values
654 * and the only danger is skipping too much.
655 */
656 if (PageCompound(page)) {
657 const unsigned int order = compound_order(page);
658
659 if (blockpfn + (1UL << order) <= end_pfn) {
660 blockpfn += (1UL << order) - 1;
661 page += (1UL << order) - 1;
662 nr_scanned += (1UL << order) - 1;
663 }
664
665 goto isolate_fail;
666 }
667
668 if (!PageBuddy(page))
669 goto isolate_fail;
670
671 /* If we already hold the lock, we can skip some rechecking. */
672 if (!locked) {
673 locked = compact_lock_irqsave(&cc->zone->lock,
674 &flags, cc);
675
676 /* Recheck this is a buddy page under lock */
677 if (!PageBuddy(page))
678 goto isolate_fail;
679 }
680
681 /* Found a free page, will break it into order-0 pages */
682 order = buddy_order(page);
683 isolated = __isolate_free_page(page, order);
684 if (!isolated)
685 break;
686 set_page_private(page, order);
687
688 nr_scanned += isolated - 1;
689 total_isolated += isolated;
690 cc->nr_freepages += isolated;
691 list_add_tail(&page->lru, &freelist[order]);
692
693 if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
694 blockpfn += isolated;
695 break;
696 }
697 /* Advance to the end of split page */
698 blockpfn += isolated - 1;
699 page += isolated - 1;
700 continue;
701
702 isolate_fail:
703 if (strict)
704 break;
705
706 }
707
708 if (locked)
709 spin_unlock_irqrestore(&cc->zone->lock, flags);
710
711 /*
712 * Be careful to not go outside of the pageblock.
713 */
714 if (unlikely(blockpfn > end_pfn))
715 blockpfn = end_pfn;
716
717 trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
718 nr_scanned, total_isolated);
719
720 /* Record how far we have got within the block */
721 *start_pfn = blockpfn;
722
723 /*
724 * If strict isolation is requested by CMA then check that all the
725 * pages requested were isolated. If there were any failures, 0 is
726 * returned and CMA will fail.
727 */
728 if (strict && blockpfn < end_pfn)
729 total_isolated = 0;
730
731 cc->total_free_scanned += nr_scanned;
732 if (total_isolated)
733 count_compact_events(COMPACTISOLATED, total_isolated);
734 return total_isolated;
735 }
736
737 /**
738 * isolate_freepages_range() - isolate free pages.
739 * @cc: Compaction control structure.
740 * @start_pfn: The first PFN to start isolating.
741 * @end_pfn: The one-past-last PFN.
742 *
743 * Non-free pages, invalid PFNs, or zone boundaries within the
744 * [start_pfn, end_pfn) range are considered errors, cause function to
745 * undo its actions and return zero.
746 *
747 * Otherwise, function returns one-past-the-last PFN of isolated page
748 * (which may be greater then end_pfn if end fell in a middle of
749 * a free page).
750 */
751 unsigned long
752 isolate_freepages_range(struct compact_control *cc,
753 unsigned long start_pfn, unsigned long end_pfn)
754 {
755 unsigned long isolated, pfn, block_start_pfn, block_end_pfn;
756 int order;
757 struct list_head tmp_freepages[NR_PAGE_ORDERS];
758
759 for (order = 0; order < NR_PAGE_ORDERS; order++)
760 INIT_LIST_HEAD(&tmp_freepages[order]);
761
762 pfn = start_pfn;
763 block_start_pfn = pageblock_start_pfn(pfn);
764 if (block_start_pfn < cc->zone->zone_start_pfn)
765 block_start_pfn = cc->zone->zone_start_pfn;
766 block_end_pfn = pageblock_end_pfn(pfn);
767
768 for (; pfn < end_pfn; pfn += isolated,
769 block_start_pfn = block_end_pfn,
770 block_end_pfn += pageblock_nr_pages) {
771 /* Protect pfn from changing by isolate_freepages_block */
772 unsigned long isolate_start_pfn = pfn;
773
774 /*
775 * pfn could pass the block_end_pfn if isolated freepage
776 * is more than pageblock order. In this case, we adjust
777 * scanning range to right one.
778 */
779 if (pfn >= block_end_pfn) {
780 block_start_pfn = pageblock_start_pfn(pfn);
781 block_end_pfn = pageblock_end_pfn(pfn);
782 }
783
784 block_end_pfn = min(block_end_pfn, end_pfn);
785
786 if (!pageblock_pfn_to_page(block_start_pfn,
787 block_end_pfn, cc->zone))
788 break;
789
790 isolated = isolate_freepages_block(cc, &isolate_start_pfn,
791 block_end_pfn, tmp_freepages, 0, true);
792
793 /*
794 * In strict mode, isolate_freepages_block() returns 0 if
795 * there are any holes in the block (ie. invalid PFNs or
796 * non-free pages).
797 */
798 if (!isolated)
799 break;
800
801 /*
802 * If we managed to isolate pages, it is always (1 << n) *
803 * pageblock_nr_pages for some non-negative n. (Max order
804 * page may span two pageblocks).
805 */
806 }
807
808 if (pfn < end_pfn) {
809 /* Loop terminated early, cleanup. */
810 release_free_list(tmp_freepages);
811 return 0;
812 }
813
814 /* __isolate_free_page() does not map the pages */
815 split_map_pages(tmp_freepages);
816
817 /* We don't use freelists for anything. */
818 return pfn;
819 }
820
821 /* Similar to reclaim, but different enough that they don't share logic */
822 static bool too_many_isolated(struct compact_control *cc)
823 {
824 pg_data_t *pgdat = cc->zone->zone_pgdat;
825 bool too_many;
826
827 unsigned long active, inactive, isolated;
828
829 inactive = node_page_state(pgdat, NR_INACTIVE_FILE) +
830 node_page_state(pgdat, NR_INACTIVE_ANON);
831 active = node_page_state(pgdat, NR_ACTIVE_FILE) +
832 node_page_state(pgdat, NR_ACTIVE_ANON);
833 isolated = node_page_state(pgdat, NR_ISOLATED_FILE) +
834 node_page_state(pgdat, NR_ISOLATED_ANON);
835
836 /*
837 * Allow GFP_NOFS to isolate past the limit set for regular
838 * compaction runs. This prevents an ABBA deadlock when other
839 * compactors have already isolated to the limit, but are
840 * blocked on filesystem locks held by the GFP_NOFS thread.
841 */
842 if (cc->gfp_mask & __GFP_FS) {
843 inactive >>= 3;
844 active >>= 3;
845 }
846
847 too_many = isolated > (inactive + active) / 2;
848 if (!too_many)
849 wake_throttle_isolated(pgdat);
850
851 return too_many;
852 }
853
854 /**
855 * skip_isolation_on_order() - determine when to skip folio isolation based on
856 * folio order and compaction target order
857 * @order: to-be-isolated folio order
858 * @target_order: compaction target order
859 *
860 * This avoids unnecessary folio isolations during compaction.
861 */
862 static bool skip_isolation_on_order(int order, int target_order)
863 {
864 /*
865 * Unless we are performing global compaction (i.e.,
866 * is_via_compact_memory), skip any folios that are larger than the
867 * target order: we wouldn't be here if we'd have a free folio with
868 * the desired target_order, so migrating this folio would likely fail
869 * later.
870 */
871 if (!is_via_compact_memory(target_order) && order >= target_order)
872 return true;
873 /*
874 * We limit memory compaction to pageblocks and won't try
875 * creating free blocks of memory that are larger than that.
876 */
877 return order >= pageblock_order;
878 }
879
880 /**
881 * isolate_migratepages_block() - isolate all migrate-able pages within
882 * a single pageblock
883 * @cc: Compaction control structure.
884 * @low_pfn: The first PFN to isolate
885 * @end_pfn: The one-past-the-last PFN to isolate, within same pageblock
886 * @mode: Isolation mode to be used.
887 *
888 * Isolate all pages that can be migrated from the range specified by
889 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
890 * Returns errno, like -EAGAIN or -EINTR in case e.g signal pending or congestion,
891 * -ENOMEM in case we could not allocate a page, or 0.
892 * cc->migrate_pfn will contain the next pfn to scan.
893 *
894 * The pages are isolated on cc->migratepages list (not required to be empty),
895 * and cc->nr_migratepages is updated accordingly.
896 */
897 static int
898 isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
899 unsigned long end_pfn, isolate_mode_t mode)
900 {
901 pg_data_t *pgdat = cc->zone->zone_pgdat;
902 unsigned long nr_scanned = 0, nr_isolated = 0;
903 struct lruvec *lruvec;
904 unsigned long flags = 0;
905 struct lruvec *locked = NULL;
906 struct folio *folio = NULL;
907 struct page *page = NULL, *valid_page = NULL;
908 struct address_space *mapping;
909 unsigned long start_pfn = low_pfn;
910 bool skip_on_failure = false;
911 unsigned long next_skip_pfn = 0;
912 bool skip_updated = false;
913 int ret = 0;
914
915 cc->migrate_pfn = low_pfn;
916
917 /*
918 * Ensure that there are not too many pages isolated from the LRU
919 * list by either parallel reclaimers or compaction. If there are,
920 * delay for some time until fewer pages are isolated
921 */
922 while (unlikely(too_many_isolated(cc))) {
923 /* stop isolation if there are still pages not migrated */
924 if (cc->nr_migratepages)
925 return -EAGAIN;
926
927 /* async migration should just abort */
928 if (cc->mode == MIGRATE_ASYNC)
929 return -EAGAIN;
930
931 reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED);
932
933 if (fatal_signal_pending(current))
934 return -EINTR;
935 }
936
937 cond_resched();
938
939 if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
940 skip_on_failure = true;
941 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
942 }
943
944 /* Time to isolate some pages for migration */
945 for (; low_pfn < end_pfn; low_pfn++) {
946 bool is_dirty, is_unevictable;
947
948 if (skip_on_failure && low_pfn >= next_skip_pfn) {
949 /*
950 * We have isolated all migration candidates in the
951 * previous order-aligned block, and did not skip it due
952 * to failure. We should migrate the pages now and
953 * hopefully succeed compaction.
954 */
955 if (nr_isolated)
956 break;
957
958 /*
959 * We failed to isolate in the previous order-aligned
960 * block. Set the new boundary to the end of the
961 * current block. Note we can't simply increase
962 * next_skip_pfn by 1 << order, as low_pfn might have
963 * been incremented by a higher number due to skipping
964 * a compound or a high-order buddy page in the
965 * previous loop iteration.
966 */
967 next_skip_pfn = block_end_pfn(low_pfn, cc->order);
968 }
969
970 /*
971 * Periodically drop the lock (if held) regardless of its
972 * contention, to give chance to IRQs. Abort completely if
973 * a fatal signal is pending.
974 */
975 if (!(low_pfn % COMPACT_CLUSTER_MAX)) {
976 if (locked) {
977 unlock_page_lruvec_irqrestore(locked, flags);
978 locked = NULL;
979 }
980
981 if (fatal_signal_pending(current)) {
982 cc->contended = true;
983 ret = -EINTR;
984
985 goto fatal_pending;
986 }
987
988 cond_resched();
989 }
990
991 nr_scanned++;
992
993 page = pfn_to_page(low_pfn);
994
995 /*
996 * Check if the pageblock has already been marked skipped.
997 * Only the first PFN is checked as the caller isolates
998 * COMPACT_CLUSTER_MAX at a time so the second call must
999 * not falsely conclude that the block should be skipped.
1000 */
1001 if (!valid_page && (pageblock_aligned(low_pfn) ||
1002 low_pfn == cc->zone->zone_start_pfn)) {
1003 if (!isolation_suitable(cc, page)) {
1004 low_pfn = end_pfn;
1005 folio = NULL;
1006 goto isolate_abort;
1007 }
1008 valid_page = page;
1009 }
1010
1011 if (PageHuge(page)) {
1012 /*
1013 * skip hugetlbfs if we are not compacting for pages
1014 * bigger than its order. THPs and other compound pages
1015 * are handled below.
1016 */
1017 if (!cc->alloc_contig) {
1018 const unsigned int order = compound_order(page);
1019
1020 if (order <= MAX_PAGE_ORDER) {
1021 low_pfn += (1UL << order) - 1;
1022 nr_scanned += (1UL << order) - 1;
1023 }
1024 goto isolate_fail;
1025 }
1026 /* for alloc_contig case */
1027 if (locked) {
1028 unlock_page_lruvec_irqrestore(locked, flags);
1029 locked = NULL;
1030 }
1031
1032 ret = isolate_or_dissolve_huge_page(page, &cc->migratepages);
1033
1034 /*
1035 * Fail isolation in case isolate_or_dissolve_huge_page()
1036 * reports an error. In case of -ENOMEM, abort right away.
1037 */
1038 if (ret < 0) {
1039 /* Do not report -EBUSY down the chain */
1040 if (ret == -EBUSY)
1041 ret = 0;
1042 low_pfn += compound_nr(page) - 1;
1043 nr_scanned += compound_nr(page) - 1;
1044 goto isolate_fail;
1045 }
1046
1047 if (PageHuge(page)) {
1048 /*
1049 * Hugepage was successfully isolated and placed
1050 * on the cc->migratepages list.
1051 */
1052 folio = page_folio(page);
1053 low_pfn += folio_nr_pages(folio) - 1;
1054 goto isolate_success_no_list;
1055 }
1056
1057 /*
1058 * Ok, the hugepage was dissolved. Now these pages are
1059 * Buddy and cannot be re-allocated because they are
1060 * isolated. Fall-through as the check below handles
1061 * Buddy pages.
1062 */
1063 }
1064
1065 /*
1066 * Skip if free. We read page order here without zone lock
1067 * which is generally unsafe, but the race window is small and
1068 * the worst thing that can happen is that we skip some
1069 * potential isolation targets.
1070 */
1071 if (PageBuddy(page)) {
1072 unsigned long freepage_order = buddy_order_unsafe(page);
1073
1074 /*
1075 * Without lock, we cannot be sure that what we got is
1076 * a valid page order. Consider only values in the
1077 * valid order range to prevent low_pfn overflow.
1078 */
1079 if (freepage_order > 0 && freepage_order <= MAX_PAGE_ORDER) {
1080 low_pfn += (1UL << freepage_order) - 1;
1081 nr_scanned += (1UL << freepage_order) - 1;
1082 }
1083 continue;
1084 }
1085
1086 /*
1087 * Regardless of being on LRU, compound pages such as THP
1088 * (hugetlbfs is handled above) are not to be compacted unless
1089 * we are attempting an allocation larger than the compound
1090 * page size. We can potentially save a lot of iterations if we
1091 * skip them at once. The check is racy, but we can consider
1092 * only valid values and the only danger is skipping too much.
1093 */
1094 if (PageCompound(page) && !cc->alloc_contig) {
1095 const unsigned int order = compound_order(page);
1096
1097 /* Skip based on page order and compaction target order. */
1098 if (skip_isolation_on_order(order, cc->order)) {
1099 if (order <= MAX_PAGE_ORDER) {
1100 low_pfn += (1UL << order) - 1;
1101 nr_scanned += (1UL << order) - 1;
1102 }
1103 goto isolate_fail;
1104 }
1105 }
1106
1107 /*
1108 * Check may be lockless but that's ok as we recheck later.
1109 * It's possible to migrate LRU and non-lru movable pages.
1110 * Skip any other type of page
1111 */
1112 if (!PageLRU(page)) {
1113 /*
1114 * __PageMovable can return false positive so we need
1115 * to verify it under page_lock.
1116 */
1117 if (unlikely(__PageMovable(page)) &&
1118 !PageIsolated(page)) {
1119 if (locked) {
1120 unlock_page_lruvec_irqrestore(locked, flags);
1121 locked = NULL;
1122 }
1123
1124 if (isolate_movable_page(page, mode)) {
1125 folio = page_folio(page);
1126 goto isolate_success;
1127 }
1128 }
1129
1130 goto isolate_fail;
1131 }
1132
1133 /*
1134 * Be careful not to clear PageLRU until after we're
1135 * sure the page is not being freed elsewhere -- the
1136 * page release code relies on it.
1137 */
1138 folio = folio_get_nontail_page(page);
1139 if (unlikely(!folio))
1140 goto isolate_fail;
1141
1142 /*
1143 * Migration will fail if an anonymous page is pinned in memory,
1144 * so avoid taking lru_lock and isolating it unnecessarily in an
1145 * admittedly racy check.
1146 */
1147 mapping = folio_mapping(folio);
1148 if (!mapping && (folio_ref_count(folio) - 1) > folio_mapcount(folio))
1149 goto isolate_fail_put;
1150
1151 /*
1152 * Only allow to migrate anonymous pages in GFP_NOFS context
1153 * because those do not depend on fs locks.
1154 */
1155 if (!(cc->gfp_mask & __GFP_FS) && mapping)
1156 goto isolate_fail_put;
1157
1158 /* Only take pages on LRU: a check now makes later tests safe */
1159 if (!folio_test_lru(folio))
1160 goto isolate_fail_put;
1161
1162 is_unevictable = folio_test_unevictable(folio);
1163
1164 /* Compaction might skip unevictable pages but CMA takes them */
1165 if (!(mode & ISOLATE_UNEVICTABLE) && is_unevictable)
1166 goto isolate_fail_put;
1167
1168 /*
1169 * To minimise LRU disruption, the caller can indicate with
1170 * ISOLATE_ASYNC_MIGRATE that it only wants to isolate pages
1171 * it will be able to migrate without blocking - clean pages
1172 * for the most part. PageWriteback would require blocking.
1173 */
1174 if ((mode & ISOLATE_ASYNC_MIGRATE) && folio_test_writeback(folio))
1175 goto isolate_fail_put;
1176
1177 is_dirty = folio_test_dirty(folio);
1178
1179 if (((mode & ISOLATE_ASYNC_MIGRATE) && is_dirty) ||
1180 (mapping && is_unevictable)) {
1181 bool migrate_dirty = true;
1182 bool is_inaccessible;
1183
1184 /*
1185 * Only folios without mappings or that have
1186 * a ->migrate_folio callback are possible to migrate
1187 * without blocking.
1188 *
1189 * Folios from inaccessible mappings are not migratable.
1190 *
1191 * However, we can be racing with truncation, which can
1192 * free the mapping that we need to check. Truncation
1193 * holds the folio lock until after the folio is removed
1194 * from the page so holding it ourselves is sufficient.
1195 *
1196 * To avoid locking the folio just to check inaccessible,
1197 * assume every inaccessible folio is also unevictable,
1198 * which is a cheaper test. If our assumption goes
1199 * wrong, it's not a correctness bug, just potentially
1200 * wasted cycles.
1201 */
1202 if (!folio_trylock(folio))
1203 goto isolate_fail_put;
1204
1205 mapping = folio_mapping(folio);
1206 if ((mode & ISOLATE_ASYNC_MIGRATE) && is_dirty) {
1207 migrate_dirty = !mapping ||
1208 mapping->a_ops->migrate_folio;
1209 }
1210 is_inaccessible = mapping && mapping_inaccessible(mapping);
1211 folio_unlock(folio);
1212 if (!migrate_dirty || is_inaccessible)
1213 goto isolate_fail_put;
1214 }
1215
1216 /* Try isolate the folio */
1217 if (!folio_test_clear_lru(folio))
1218 goto isolate_fail_put;
1219
1220 lruvec = folio_lruvec(folio);
1221
1222 /* If we already hold the lock, we can skip some rechecking */
1223 if (lruvec != locked) {
1224 if (locked)
1225 unlock_page_lruvec_irqrestore(locked, flags);
1226
1227 compact_lock_irqsave(&lruvec->lru_lock, &flags, cc);
1228 locked = lruvec;
1229
1230 lruvec_memcg_debug(lruvec, folio);
1231
1232 /*
1233 * Try get exclusive access under lock. If marked for
1234 * skip, the scan is aborted unless the current context
1235 * is a rescan to reach the end of the pageblock.
1236 */
1237 if (!skip_updated && valid_page) {
1238 skip_updated = true;
1239 if (test_and_set_skip(cc, valid_page) &&
1240 !cc->finish_pageblock) {
1241 low_pfn = end_pfn;
1242 goto isolate_abort;
1243 }
1244 }
1245
1246 /*
1247 * Check LRU folio order under the lock
1248 */
1249 if (unlikely(skip_isolation_on_order(folio_order(folio),
1250 cc->order) &&
1251 !cc->alloc_contig)) {
1252 low_pfn += folio_nr_pages(folio) - 1;
1253 nr_scanned += folio_nr_pages(folio) - 1;
1254 folio_set_lru(folio);
1255 goto isolate_fail_put;
1256 }
1257 }
1258
1259 /* The folio is taken off the LRU */
1260 if (folio_test_large(folio))
1261 low_pfn += folio_nr_pages(folio) - 1;
1262
1263 /* Successfully isolated */
1264 lruvec_del_folio(lruvec, folio);
1265 node_stat_mod_folio(folio,
1266 NR_ISOLATED_ANON + folio_is_file_lru(folio),
1267 folio_nr_pages(folio));
1268
1269 isolate_success:
1270 list_add(&folio->lru, &cc->migratepages);
1271 isolate_success_no_list:
1272 cc->nr_migratepages += folio_nr_pages(folio);
1273 nr_isolated += folio_nr_pages(folio);
1274 nr_scanned += folio_nr_pages(folio) - 1;
1275
1276 /*
1277 * Avoid isolating too much unless this block is being
1278 * fully scanned (e.g. dirty/writeback pages, parallel allocation)
1279 * or a lock is contended. For contention, isolate quickly to
1280 * potentially remove one source of contention.
1281 */
1282 if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX &&
1283 !cc->finish_pageblock && !cc->contended) {
1284 ++low_pfn;
1285 break;
1286 }
1287
1288 continue;
1289
1290 isolate_fail_put:
1291 /* Avoid potential deadlock in freeing page under lru_lock */
1292 if (locked) {
1293 unlock_page_lruvec_irqrestore(locked, flags);
1294 locked = NULL;
1295 }
1296 folio_put(folio);
1297
1298 isolate_fail:
1299 if (!skip_on_failure && ret != -ENOMEM)
1300 continue;
1301
1302 /*
1303 * We have isolated some pages, but then failed. Release them
1304 * instead of migrating, as we cannot form the cc->order buddy
1305 * page anyway.
1306 */
1307 if (nr_isolated) {
1308 if (locked) {
1309 unlock_page_lruvec_irqrestore(locked, flags);
1310 locked = NULL;
1311 }
1312 putback_movable_pages(&cc->migratepages);
1313 cc->nr_migratepages = 0;
1314 nr_isolated = 0;
1315 }
1316
1317 if (low_pfn < next_skip_pfn) {
1318 low_pfn = next_skip_pfn - 1;
1319 /*
1320 * The check near the loop beginning would have updated
1321 * next_skip_pfn too, but this is a bit simpler.
1322 */
1323 next_skip_pfn += 1UL << cc->order;
1324 }
1325
1326 if (ret == -ENOMEM)
1327 break;
1328 }
1329
1330 /*
1331 * The PageBuddy() check could have potentially brought us outside
1332 * the range to be scanned.
1333 */
1334 if (unlikely(low_pfn > end_pfn))
1335 low_pfn = end_pfn;
1336
1337 folio = NULL;
1338
1339 isolate_abort:
1340 if (locked)
1341 unlock_page_lruvec_irqrestore(locked, flags);
1342 if (folio) {
1343 folio_set_lru(folio);
1344 folio_put(folio);
1345 }
1346
1347 /*
1348 * Update the cached scanner pfn once the pageblock has been scanned.
1349 * Pages will either be migrated in which case there is no point
1350 * scanning in the near future or migration failed in which case the
1351 * failure reason may persist. The block is marked for skipping if
1352 * there were no pages isolated in the block or if the block is
1353 * rescanned twice in a row.
1354 */
1355 if (low_pfn == end_pfn && (!nr_isolated || cc->finish_pageblock)) {
1356 if (!cc->no_set_skip_hint && valid_page && !skip_updated)
1357 set_pageblock_skip(valid_page);
1358 update_cached_migrate(cc, low_pfn);
1359 }
1360
1361 trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
1362 nr_scanned, nr_isolated);
1363
1364 fatal_pending:
1365 cc->total_migrate_scanned += nr_scanned;
1366 if (nr_isolated)
1367 count_compact_events(COMPACTISOLATED, nr_isolated);
1368
1369 cc->migrate_pfn = low_pfn;
1370
1371 return ret;
1372 }
1373
1374 /**
1375 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
1376 * @cc: Compaction control structure.
1377 * @start_pfn: The first PFN to start isolating.
1378 * @end_pfn: The one-past-last PFN.
1379 *
1380 * Returns -EAGAIN when contented, -EINTR in case of a signal pending, -ENOMEM
1381 * in case we could not allocate a page, or 0.
1382 */
1383 int
1384 isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
1385 unsigned long end_pfn)
1386 {
1387 unsigned long pfn, block_start_pfn, block_end_pfn;
1388 int ret = 0;
1389
1390 /* Scan block by block. First and last block may be incomplete */
1391 pfn = start_pfn;
1392 block_start_pfn = pageblock_start_pfn(pfn);
1393 if (block_start_pfn < cc->zone->zone_start_pfn)
1394 block_start_pfn = cc->zone->zone_start_pfn;
1395 block_end_pfn = pageblock_end_pfn(pfn);
1396
1397 for (; pfn < end_pfn; pfn = block_end_pfn,
1398 block_start_pfn = block_end_pfn,
1399 block_end_pfn += pageblock_nr_pages) {
1400
1401 block_end_pfn = min(block_end_pfn, end_pfn);
1402
1403 if (!pageblock_pfn_to_page(block_start_pfn,
1404 block_end_pfn, cc->zone))
1405 continue;
1406
1407 ret = isolate_migratepages_block(cc, pfn, block_end_pfn,
1408 ISOLATE_UNEVICTABLE);
1409
1410 if (ret)
1411 break;
1412
1413 if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX)
1414 break;
1415 }
1416
1417 return ret;
1418 }
1419
1420 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
1421 #ifdef CONFIG_COMPACTION
1422
1423 static bool suitable_migration_source(struct compact_control *cc,
1424 struct page *page)
1425 {
1426 int block_mt;
1427
1428 if (pageblock_skip_persistent(page))
1429 return false;
1430
1431 if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction)
1432 return true;
1433
1434 block_mt = get_pageblock_migratetype(page);
1435
1436 if (cc->migratetype == MIGRATE_MOVABLE)
1437 return is_migrate_movable(block_mt);
1438 else
1439 return block_mt == cc->migratetype;
1440 }
1441
1442 /* Returns true if the page is within a block suitable for migration to */
1443 static bool suitable_migration_target(struct compact_control *cc,
1444 struct page *page)
1445 {
1446 /* If the page is a large free page, then disallow migration */
1447 if (PageBuddy(page)) {
1448 int order = cc->order > 0 ? cc->order : pageblock_order;
1449
1450 /*
1451 * We are checking page_order without zone->lock taken. But
1452 * the only small danger is that we skip a potentially suitable
1453 * pageblock, so it's not worth to check order for valid range.
1454 */
1455 if (buddy_order_unsafe(page) >= order)
1456 return false;
1457 }
1458
1459 if (cc->ignore_block_suitable)
1460 return true;
1461
1462 /* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1463 if (is_migrate_movable(get_pageblock_migratetype(page)))
1464 return true;
1465
1466 /* Otherwise skip the block */
1467 return false;
1468 }
1469
1470 static inline unsigned int
1471 freelist_scan_limit(struct compact_control *cc)
1472 {
1473 unsigned short shift = BITS_PER_LONG - 1;
1474
1475 return (COMPACT_CLUSTER_MAX >> min(shift, cc->fast_search_fail)) + 1;
1476 }
1477
1478 /*
1479 * Test whether the free scanner has reached the same or lower pageblock than
1480 * the migration scanner, and compaction should thus terminate.
1481 */
1482 static inline bool compact_scanners_met(struct compact_control *cc)
1483 {
1484 return (cc->free_pfn >> pageblock_order)
1485 <= (cc->migrate_pfn >> pageblock_order);
1486 }
1487
1488 /*
1489 * Used when scanning for a suitable migration target which scans freelists
1490 * in reverse. Reorders the list such as the unscanned pages are scanned
1491 * first on the next iteration of the free scanner
1492 */
1493 static void
1494 move_freelist_head(struct list_head *freelist, struct page *freepage)
1495 {
1496 LIST_HEAD(sublist);
1497
1498 if (!list_is_first(&freepage->buddy_list, freelist)) {
1499 list_cut_before(&sublist, freelist, &freepage->buddy_list);
1500 list_splice_tail(&sublist, freelist);
1501 }
1502 }
1503
1504 /*
1505 * Similar to move_freelist_head except used by the migration scanner
1506 * when scanning forward. It's possible for these list operations to
1507 * move against each other if they search the free list exactly in
1508 * lockstep.
1509 */
1510 static void
1511 move_freelist_tail(struct list_head *freelist, struct page *freepage)
1512 {
1513 LIST_HEAD(sublist);
1514
1515 if (!list_is_last(&freepage->buddy_list, freelist)) {
1516 list_cut_position(&sublist, freelist, &freepage->buddy_list);
1517 list_splice_tail(&sublist, freelist);
1518 }
1519 }
1520
1521 static void
1522 fast_isolate_around(struct compact_control *cc, unsigned long pfn)
1523 {
1524 unsigned long start_pfn, end_pfn;
1525 struct page *page;
1526
1527 /* Do not search around if there are enough pages already */
1528 if (cc->nr_freepages >= cc->nr_migratepages)
1529 return;
1530
1531 /* Minimise scanning during async compaction */
1532 if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC)
1533 return;
1534
1535 /* Pageblock boundaries */
1536 start_pfn = max(pageblock_start_pfn(pfn), cc->zone->zone_start_pfn);
1537 end_pfn = min(pageblock_end_pfn(pfn), zone_end_pfn(cc->zone));
1538
1539 page = pageblock_pfn_to_page(start_pfn, end_pfn, cc->zone);
1540 if (!page)
1541 return;
1542
1543 isolate_freepages_block(cc, &start_pfn, end_pfn, cc->freepages, 1, false);
1544
1545 /* Skip this pageblock in the future as it's full or nearly full */
1546 if (start_pfn == end_pfn && !cc->no_set_skip_hint)
1547 set_pageblock_skip(page);
1548 }
1549
1550 /* Search orders in round-robin fashion */
1551 static int next_search_order(struct compact_control *cc, int order)
1552 {
1553 order--;
1554 if (order < 0)
1555 order = cc->order - 1;
1556
1557 /* Search wrapped around? */
1558 if (order == cc->search_order) {
1559 cc->search_order--;
1560 if (cc->search_order < 0)
1561 cc->search_order = cc->order - 1;
1562 return -1;
1563 }
1564
1565 return order;
1566 }
1567
1568 static void fast_isolate_freepages(struct compact_control *cc)
1569 {
1570 unsigned int limit = max(1U, freelist_scan_limit(cc) >> 1);
1571 unsigned int nr_scanned = 0, total_isolated = 0;
1572 unsigned long low_pfn, min_pfn, highest = 0;
1573 unsigned long nr_isolated = 0;
1574 unsigned long distance;
1575 struct page *page = NULL;
1576 bool scan_start = false;
1577 int order;
1578
1579 /* Full compaction passes in a negative order */
1580 if (cc->order <= 0)
1581 return;
1582
1583 /*
1584 * If starting the scan, use a deeper search and use the highest
1585 * PFN found if a suitable one is not found.
1586 */
1587 if (cc->free_pfn >= cc->zone->compact_init_free_pfn) {
1588 limit = pageblock_nr_pages >> 1;
1589 scan_start = true;
1590 }
1591
1592 /*
1593 * Preferred point is in the top quarter of the scan space but take
1594 * a pfn from the top half if the search is problematic.
1595 */
1596 distance = (cc->free_pfn - cc->migrate_pfn);
1597 low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2));
1598 min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1));
1599
1600 if (WARN_ON_ONCE(min_pfn > low_pfn))
1601 low_pfn = min_pfn;
1602
1603 /*
1604 * Search starts from the last successful isolation order or the next
1605 * order to search after a previous failure
1606 */
1607 cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order);
1608
1609 for (order = cc->search_order;
1610 !page && order >= 0;
1611 order = next_search_order(cc, order)) {
1612 struct free_area *area = &cc->zone->free_area[order];
1613 struct list_head *freelist;
1614 struct page *freepage;
1615 unsigned long flags;
1616 unsigned int order_scanned = 0;
1617 unsigned long high_pfn = 0;
1618
1619 if (!area->nr_free)
1620 continue;
1621
1622 spin_lock_irqsave(&cc->zone->lock, flags);
1623 freelist = &area->free_list[MIGRATE_MOVABLE];
1624 list_for_each_entry_reverse(freepage, freelist, buddy_list) {
1625 unsigned long pfn;
1626
1627 order_scanned++;
1628 nr_scanned++;
1629 pfn = page_to_pfn(freepage);
1630
1631 if (pfn >= highest)
1632 highest = max(pageblock_start_pfn(pfn),
1633 cc->zone->zone_start_pfn);
1634
1635 if (pfn >= low_pfn) {
1636 cc->fast_search_fail = 0;
1637 cc->search_order = order;
1638 page = freepage;
1639 break;
1640 }
1641
1642 if (pfn >= min_pfn && pfn > high_pfn) {
1643 high_pfn = pfn;
1644
1645 /* Shorten the scan if a candidate is found */
1646 limit >>= 1;
1647 }
1648
1649 if (order_scanned >= limit)
1650 break;
1651 }
1652
1653 /* Use a maximum candidate pfn if a preferred one was not found */
1654 if (!page && high_pfn) {
1655 page = pfn_to_page(high_pfn);
1656
1657 /* Update freepage for the list reorder below */
1658 freepage = page;
1659 }
1660
1661 /* Reorder to so a future search skips recent pages */
1662 move_freelist_head(freelist, freepage);
1663
1664 /* Isolate the page if available */
1665 if (page) {
1666 if (__isolate_free_page(page, order)) {
1667 set_page_private(page, order);
1668 nr_isolated = 1 << order;
1669 nr_scanned += nr_isolated - 1;
1670 total_isolated += nr_isolated;
1671 cc->nr_freepages += nr_isolated;
1672 list_add_tail(&page->lru, &cc->freepages[order]);
1673 count_compact_events(COMPACTISOLATED, nr_isolated);
1674 } else {
1675 /* If isolation fails, abort the search */
1676 order = cc->search_order + 1;
1677 page = NULL;
1678 }
1679 }
1680
1681 spin_unlock_irqrestore(&cc->zone->lock, flags);
1682
1683 /* Skip fast search if enough freepages isolated */
1684 if (cc->nr_freepages >= cc->nr_migratepages)
1685 break;
1686
1687 /*
1688 * Smaller scan on next order so the total scan is related
1689 * to freelist_scan_limit.
1690 */
1691 if (order_scanned >= limit)
1692 limit = max(1U, limit >> 1);
1693 }
1694
1695 trace_mm_compaction_fast_isolate_freepages(min_pfn, cc->free_pfn,
1696 nr_scanned, total_isolated);
1697
1698 if (!page) {
1699 cc->fast_search_fail++;
1700 if (scan_start) {
1701 /*
1702 * Use the highest PFN found above min. If one was
1703 * not found, be pessimistic for direct compaction
1704 * and use the min mark.
1705 */
1706 if (highest >= min_pfn) {
1707 page = pfn_to_page(highest);
1708 cc->free_pfn = highest;
1709 } else {
1710 if (cc->direct_compaction && pfn_valid(min_pfn)) {
1711 page = pageblock_pfn_to_page(min_pfn,
1712 min(pageblock_end_pfn(min_pfn),
1713 zone_end_pfn(cc->zone)),
1714 cc->zone);
1715 if (page && !suitable_migration_target(cc, page))
1716 page = NULL;
1717
1718 cc->free_pfn = min_pfn;
1719 }
1720 }
1721 }
1722 }
1723
1724 if (highest && highest >= cc->zone->compact_cached_free_pfn) {
1725 highest -= pageblock_nr_pages;
1726 cc->zone->compact_cached_free_pfn = highest;
1727 }
1728
1729 cc->total_free_scanned += nr_scanned;
1730 if (!page)
1731 return;
1732
1733 low_pfn = page_to_pfn(page);
1734 fast_isolate_around(cc, low_pfn);
1735 }
1736
1737 /*
1738 * Based on information in the current compact_control, find blocks
1739 * suitable for isolating free pages from and then isolate them.
1740 */
1741 static void isolate_freepages(struct compact_control *cc)
1742 {
1743 struct zone *zone = cc->zone;
1744 struct page *page;
1745 unsigned long block_start_pfn; /* start of current pageblock */
1746 unsigned long isolate_start_pfn; /* exact pfn we start at */
1747 unsigned long block_end_pfn; /* end of current pageblock */
1748 unsigned long low_pfn; /* lowest pfn scanner is able to scan */
1749 unsigned int stride;
1750
1751 /* Try a small search of the free lists for a candidate */
1752 fast_isolate_freepages(cc);
1753 if (cc->nr_freepages)
1754 return;
1755
1756 /*
1757 * Initialise the free scanner. The starting point is where we last
1758 * successfully isolated from, zone-cached value, or the end of the
1759 * zone when isolating for the first time. For looping we also need
1760 * this pfn aligned down to the pageblock boundary, because we do
1761 * block_start_pfn -= pageblock_nr_pages in the for loop.
1762 * For ending point, take care when isolating in last pageblock of a
1763 * zone which ends in the middle of a pageblock.
1764 * The low boundary is the end of the pageblock the migration scanner
1765 * is using.
1766 */
1767 isolate_start_pfn = cc->free_pfn;
1768 block_start_pfn = pageblock_start_pfn(isolate_start_pfn);
1769 block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1770 zone_end_pfn(zone));
1771 low_pfn = pageblock_end_pfn(cc->migrate_pfn);
1772 stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1;
1773
1774 /*
1775 * Isolate free pages until enough are available to migrate the
1776 * pages on cc->migratepages. We stop searching if the migrate
1777 * and free page scanners meet or enough free pages are isolated.
1778 */
1779 for (; block_start_pfn >= low_pfn;
1780 block_end_pfn = block_start_pfn,
1781 block_start_pfn -= pageblock_nr_pages,
1782 isolate_start_pfn = block_start_pfn) {
1783 unsigned long nr_isolated;
1784
1785 /*
1786 * This can iterate a massively long zone without finding any
1787 * suitable migration targets, so periodically check resched.
1788 */
1789 if (!(block_start_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages)))
1790 cond_resched();
1791
1792 page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1793 zone);
1794 if (!page) {
1795 unsigned long next_pfn;
1796
1797 next_pfn = skip_offline_sections_reverse(block_start_pfn);
1798 if (next_pfn)
1799 block_start_pfn = max(next_pfn, low_pfn);
1800
1801 continue;
1802 }
1803
1804 /* Check the block is suitable for migration */
1805 if (!suitable_migration_target(cc, page))
1806 continue;
1807
1808 /* If isolation recently failed, do not retry */
1809 if (!isolation_suitable(cc, page))
1810 continue;
1811
1812 /* Found a block suitable for isolating free pages from. */
1813 nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn,
1814 block_end_pfn, cc->freepages, stride, false);
1815
1816 /* Update the skip hint if the full pageblock was scanned */
1817 if (isolate_start_pfn == block_end_pfn)
1818 update_pageblock_skip(cc, page, block_start_pfn -
1819 pageblock_nr_pages);
1820
1821 /* Are enough freepages isolated? */
1822 if (cc->nr_freepages >= cc->nr_migratepages) {
1823 if (isolate_start_pfn >= block_end_pfn) {
1824 /*
1825 * Restart at previous pageblock if more
1826 * freepages can be isolated next time.
1827 */
1828 isolate_start_pfn =
1829 block_start_pfn - pageblock_nr_pages;
1830 }
1831 break;
1832 } else if (isolate_start_pfn < block_end_pfn) {
1833 /*
1834 * If isolation failed early, do not continue
1835 * needlessly.
1836 */
1837 break;
1838 }
1839
1840 /* Adjust stride depending on isolation */
1841 if (nr_isolated) {
1842 stride = 1;
1843 continue;
1844 }
1845 stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1);
1846 }
1847
1848 /*
1849 * Record where the free scanner will restart next time. Either we
1850 * broke from the loop and set isolate_start_pfn based on the last
1851 * call to isolate_freepages_block(), or we met the migration scanner
1852 * and the loop terminated due to isolate_start_pfn < low_pfn
1853 */
1854 cc->free_pfn = isolate_start_pfn;
1855 }
1856
1857 /*
1858 * This is a migrate-callback that "allocates" freepages by taking pages
1859 * from the isolated freelists in the block we are migrating to.
1860 */
1861 static struct folio *compaction_alloc_noprof(struct folio *src, unsigned long data)
1862 {
1863 struct compact_control *cc = (struct compact_control *)data;
1864 struct folio *dst;
1865 int order = folio_order(src);
1866 bool has_isolated_pages = false;
1867 int start_order;
1868 struct page *freepage;
1869 unsigned long size;
1870
1871 again:
1872 for (start_order = order; start_order < NR_PAGE_ORDERS; start_order++)
1873 if (!list_empty(&cc->freepages[start_order]))
1874 break;
1875
1876 /* no free pages in the list */
1877 if (start_order == NR_PAGE_ORDERS) {
1878 if (has_isolated_pages)
1879 return NULL;
1880 isolate_freepages(cc);
1881 has_isolated_pages = true;
1882 goto again;
1883 }
1884
1885 freepage = list_first_entry(&cc->freepages[start_order], struct page,
1886 lru);
1887 size = 1 << start_order;
1888
1889 list_del(&freepage->lru);
1890
1891 while (start_order > order) {
1892 start_order--;
1893 size >>= 1;
1894
1895 list_add(&freepage[size].lru, &cc->freepages[start_order]);
1896 set_page_private(&freepage[size], start_order);
1897 }
1898 dst = (struct folio *)freepage;
1899
1900 post_alloc_hook(&dst->page, order, __GFP_MOVABLE);
1901 if (order)
1902 prep_compound_page(&dst->page, order);
1903 cc->nr_freepages -= 1 << order;
1904 cc->nr_migratepages -= 1 << order;
1905 return page_rmappable_folio(&dst->page);
1906 }
1907
1908 static struct folio *compaction_alloc(struct folio *src, unsigned long data)
1909 {
1910 return alloc_hooks(compaction_alloc_noprof(src, data));
1911 }
1912
1913 /*
1914 * This is a migrate-callback that "frees" freepages back to the isolated
1915 * freelist. All pages on the freelist are from the same zone, so there is no
1916 * special handling needed for NUMA.
1917 */
1918 static void compaction_free(struct folio *dst, unsigned long data)
1919 {
1920 struct compact_control *cc = (struct compact_control *)data;
1921 int order = folio_order(dst);
1922 struct page *page = &dst->page;
1923
1924 if (folio_put_testzero(dst)) {
1925 free_pages_prepare(page, order);
1926 list_add(&dst->lru, &cc->freepages[order]);
1927 cc->nr_freepages += 1 << order;
1928 }
1929 cc->nr_migratepages += 1 << order;
1930 /*
1931 * someone else has referenced the page, we cannot take it back to our
1932 * free list.
1933 */
1934 }
1935
1936 /* possible outcome of isolate_migratepages */
1937 typedef enum {
1938 ISOLATE_ABORT, /* Abort compaction now */
1939 ISOLATE_NONE, /* No pages isolated, continue scanning */
1940 ISOLATE_SUCCESS, /* Pages isolated, migrate */
1941 } isolate_migrate_t;
1942
1943 /*
1944 * Allow userspace to control policy on scanning the unevictable LRU for
1945 * compactable pages.
1946 */
1947 static int sysctl_compact_unevictable_allowed __read_mostly = CONFIG_COMPACT_UNEVICTABLE_DEFAULT;
1948 /*
1949 * Tunable for proactive compaction. It determines how
1950 * aggressively the kernel should compact memory in the
1951 * background. It takes values in the range [0, 100].
1952 */
1953 static unsigned int __read_mostly sysctl_compaction_proactiveness = 20;
1954 static int sysctl_extfrag_threshold = 500;
1955 static int __read_mostly sysctl_compact_memory;
1956
1957 static inline void
1958 update_fast_start_pfn(struct compact_control *cc, unsigned long pfn)
1959 {
1960 if (cc->fast_start_pfn == ULONG_MAX)
1961 return;
1962
1963 if (!cc->fast_start_pfn)
1964 cc->fast_start_pfn = pfn;
1965
1966 cc->fast_start_pfn = min(cc->fast_start_pfn, pfn);
1967 }
1968
1969 static inline unsigned long
1970 reinit_migrate_pfn(struct compact_control *cc)
1971 {
1972 if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX)
1973 return cc->migrate_pfn;
1974
1975 cc->migrate_pfn = cc->fast_start_pfn;
1976 cc->fast_start_pfn = ULONG_MAX;
1977
1978 return cc->migrate_pfn;
1979 }
1980
1981 /*
1982 * Briefly search the free lists for a migration source that already has
1983 * some free pages to reduce the number of pages that need migration
1984 * before a pageblock is free.
1985 */
1986 static unsigned long fast_find_migrateblock(struct compact_control *cc)
1987 {
1988 unsigned int limit = freelist_scan_limit(cc);
1989 unsigned int nr_scanned = 0;
1990 unsigned long distance;
1991 unsigned long pfn = cc->migrate_pfn;
1992 unsigned long high_pfn;
1993 int order;
1994 bool found_block = false;
1995
1996 /* Skip hints are relied on to avoid repeats on the fast search */
1997 if (cc->ignore_skip_hint)
1998 return pfn;
1999
2000 /*
2001 * If the pageblock should be finished then do not select a different
2002 * pageblock.
2003 */
2004 if (cc->finish_pageblock)
2005 return pfn;
2006
2007 /*
2008 * If the migrate_pfn is not at the start of a zone or the start
2009 * of a pageblock then assume this is a continuation of a previous
2010 * scan restarted due to COMPACT_CLUSTER_MAX.
2011 */
2012 if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn))
2013 return pfn;
2014
2015 /*
2016 * For smaller orders, just linearly scan as the number of pages
2017 * to migrate should be relatively small and does not necessarily
2018 * justify freeing up a large block for a small allocation.
2019 */
2020 if (cc->order <= PAGE_ALLOC_COSTLY_ORDER)
2021 return pfn;
2022
2023 /*
2024 * Only allow kcompactd and direct requests for movable pages to
2025 * quickly clear out a MOVABLE pageblock for allocation. This
2026 * reduces the risk that a large movable pageblock is freed for
2027 * an unmovable/reclaimable small allocation.
2028 */
2029 if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE)
2030 return pfn;
2031
2032 /*
2033 * When starting the migration scanner, pick any pageblock within the
2034 * first half of the search space. Otherwise try and pick a pageblock
2035 * within the first eighth to reduce the chances that a migration
2036 * target later becomes a source.
2037 */
2038 distance = (cc->free_pfn - cc->migrate_pfn) >> 1;
2039 if (cc->migrate_pfn != cc->zone->zone_start_pfn)
2040 distance >>= 2;
2041 high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance);
2042
2043 for (order = cc->order - 1;
2044 order >= PAGE_ALLOC_COSTLY_ORDER && !found_block && nr_scanned < limit;
2045 order--) {
2046 struct free_area *area = &cc->zone->free_area[order];
2047 struct list_head *freelist;
2048 unsigned long flags;
2049 struct page *freepage;
2050
2051 if (!area->nr_free)
2052 continue;
2053
2054 spin_lock_irqsave(&cc->zone->lock, flags);
2055 freelist = &area->free_list[MIGRATE_MOVABLE];
2056 list_for_each_entry(freepage, freelist, buddy_list) {
2057 unsigned long free_pfn;
2058
2059 if (nr_scanned++ >= limit) {
2060 move_freelist_tail(freelist, freepage);
2061 break;
2062 }
2063
2064 free_pfn = page_to_pfn(freepage);
2065 if (free_pfn < high_pfn) {
2066 /*
2067 * Avoid if skipped recently. Ideally it would
2068 * move to the tail but even safe iteration of
2069 * the list assumes an entry is deleted, not
2070 * reordered.
2071 */
2072 if (get_pageblock_skip(freepage))
2073 continue;
2074
2075 /* Reorder to so a future search skips recent pages */
2076 move_freelist_tail(freelist, freepage);
2077
2078 update_fast_start_pfn(cc, free_pfn);
2079 pfn = pageblock_start_pfn(free_pfn);
2080 if (pfn < cc->zone->zone_start_pfn)
2081 pfn = cc->zone->zone_start_pfn;
2082 cc->fast_search_fail = 0;
2083 found_block = true;
2084 break;
2085 }
2086 }
2087 spin_unlock_irqrestore(&cc->zone->lock, flags);
2088 }
2089
2090 cc->total_migrate_scanned += nr_scanned;
2091
2092 /*
2093 * If fast scanning failed then use a cached entry for a page block
2094 * that had free pages as the basis for starting a linear scan.
2095 */
2096 if (!found_block) {
2097 cc->fast_search_fail++;
2098 pfn = reinit_migrate_pfn(cc);
2099 }
2100 return pfn;
2101 }
2102
2103 /*
2104 * Isolate all pages that can be migrated from the first suitable block,
2105 * starting at the block pointed to by the migrate scanner pfn within
2106 * compact_control.
2107 */
2108 static isolate_migrate_t isolate_migratepages(struct compact_control *cc)
2109 {
2110 unsigned long block_start_pfn;
2111 unsigned long block_end_pfn;
2112 unsigned long low_pfn;
2113 struct page *page;
2114 const isolate_mode_t isolate_mode =
2115 (sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
2116 (cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
2117 bool fast_find_block;
2118
2119 /*
2120 * Start at where we last stopped, or beginning of the zone as
2121 * initialized by compact_zone(). The first failure will use
2122 * the lowest PFN as the starting point for linear scanning.
2123 */
2124 low_pfn = fast_find_migrateblock(cc);
2125 block_start_pfn = pageblock_start_pfn(low_pfn);
2126 if (block_start_pfn < cc->zone->zone_start_pfn)
2127 block_start_pfn = cc->zone->zone_start_pfn;
2128
2129 /*
2130 * fast_find_migrateblock() has already ensured the pageblock is not
2131 * set with a skipped flag, so to avoid the isolation_suitable check
2132 * below again, check whether the fast search was successful.
2133 */
2134 fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail;
2135
2136 /* Only scan within a pageblock boundary */
2137 block_end_pfn = pageblock_end_pfn(low_pfn);
2138
2139 /*
2140 * Iterate over whole pageblocks until we find the first suitable.
2141 * Do not cross the free scanner.
2142 */
2143 for (; block_end_pfn <= cc->free_pfn;
2144 fast_find_block = false,
2145 cc->migrate_pfn = low_pfn = block_end_pfn,
2146 block_start_pfn = block_end_pfn,
2147 block_end_pfn += pageblock_nr_pages) {
2148
2149 /*
2150 * This can potentially iterate a massively long zone with
2151 * many pageblocks unsuitable, so periodically check if we
2152 * need to schedule.
2153 */
2154 if (!(low_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages)))
2155 cond_resched();
2156
2157 page = pageblock_pfn_to_page(block_start_pfn,
2158 block_end_pfn, cc->zone);
2159 if (!page) {
2160 unsigned long next_pfn;
2161
2162 next_pfn = skip_offline_sections(block_start_pfn);
2163 if (next_pfn)
2164 block_end_pfn = min(next_pfn, cc->free_pfn);
2165 continue;
2166 }
2167
2168 /*
2169 * If isolation recently failed, do not retry. Only check the
2170 * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock
2171 * to be visited multiple times. Assume skip was checked
2172 * before making it "skip" so other compaction instances do
2173 * not scan the same block.
2174 */
2175 if ((pageblock_aligned(low_pfn) ||
2176 low_pfn == cc->zone->zone_start_pfn) &&
2177 !fast_find_block && !isolation_suitable(cc, page))
2178 continue;
2179
2180 /*
2181 * For async direct compaction, only scan the pageblocks of the
2182 * same migratetype without huge pages. Async direct compaction
2183 * is optimistic to see if the minimum amount of work satisfies
2184 * the allocation. The cached PFN is updated as it's possible
2185 * that all remaining blocks between source and target are
2186 * unsuitable and the compaction scanners fail to meet.
2187 */
2188 if (!suitable_migration_source(cc, page)) {
2189 update_cached_migrate(cc, block_end_pfn);
2190 continue;
2191 }
2192
2193 /* Perform the isolation */
2194 if (isolate_migratepages_block(cc, low_pfn, block_end_pfn,
2195 isolate_mode))
2196 return ISOLATE_ABORT;
2197
2198 /*
2199 * Either we isolated something and proceed with migration. Or
2200 * we failed and compact_zone should decide if we should
2201 * continue or not.
2202 */
2203 break;
2204 }
2205
2206 return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
2207 }
2208
2209 /*
2210 * Determine whether kswapd is (or recently was!) running on this node.
2211 *
2212 * pgdat_kswapd_lock() pins pgdat->kswapd, so a concurrent kswapd_stop() can't
2213 * zero it.
2214 */
2215 static bool kswapd_is_running(pg_data_t *pgdat)
2216 {
2217 bool running;
2218
2219 pgdat_kswapd_lock(pgdat);
2220 running = pgdat->kswapd && task_is_running(pgdat->kswapd);
2221 pgdat_kswapd_unlock(pgdat);
2222
2223 return running;
2224 }
2225
2226 /*
2227 * A zone's fragmentation score is the external fragmentation wrt to the
2228 * COMPACTION_HPAGE_ORDER. It returns a value in the range [0, 100].
2229 */
2230 static unsigned int fragmentation_score_zone(struct zone *zone)
2231 {
2232 return extfrag_for_order(zone, COMPACTION_HPAGE_ORDER);
2233 }
2234
2235 /*
2236 * A weighted zone's fragmentation score is the external fragmentation
2237 * wrt to the COMPACTION_HPAGE_ORDER scaled by the zone's size. It
2238 * returns a value in the range [0, 100].
2239 *
2240 * The scaling factor ensures that proactive compaction focuses on larger
2241 * zones like ZONE_NORMAL, rather than smaller, specialized zones like
2242 * ZONE_DMA32. For smaller zones, the score value remains close to zero,
2243 * and thus never exceeds the high threshold for proactive compaction.
2244 */
2245 static unsigned int fragmentation_score_zone_weighted(struct zone *zone)
2246 {
2247 unsigned long score;
2248
2249 score = zone->present_pages * fragmentation_score_zone(zone);
2250 return div64_ul(score, zone->zone_pgdat->node_present_pages + 1);
2251 }
2252
2253 /*
2254 * The per-node proactive (background) compaction process is started by its
2255 * corresponding kcompactd thread when the node's fragmentation score
2256 * exceeds the high threshold. The compaction process remains active till
2257 * the node's score falls below the low threshold, or one of the back-off
2258 * conditions is met.
2259 */
2260 static unsigned int fragmentation_score_node(pg_data_t *pgdat)
2261 {
2262 unsigned int score = 0;
2263 int zoneid;
2264
2265 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2266 struct zone *zone;
2267
2268 zone = &pgdat->node_zones[zoneid];
2269 if (!populated_zone(zone))
2270 continue;
2271 score += fragmentation_score_zone_weighted(zone);
2272 }
2273
2274 return score;
2275 }
2276
2277 static unsigned int fragmentation_score_wmark(bool low)
2278 {
2279 unsigned int wmark_low;
2280
2281 /*
2282 * Cap the low watermark to avoid excessive compaction
2283 * activity in case a user sets the proactiveness tunable
2284 * close to 100 (maximum).
2285 */
2286 wmark_low = max(100U - sysctl_compaction_proactiveness, 5U);
2287 return low ? wmark_low : min(wmark_low + 10, 100U);
2288 }
2289
2290 static bool should_proactive_compact_node(pg_data_t *pgdat)
2291 {
2292 int wmark_high;
2293
2294 if (!sysctl_compaction_proactiveness || kswapd_is_running(pgdat))
2295 return false;
2296
2297 wmark_high = fragmentation_score_wmark(false);
2298 return fragmentation_score_node(pgdat) > wmark_high;
2299 }
2300
2301 static enum compact_result __compact_finished(struct compact_control *cc)
2302 {
2303 unsigned int order;
2304 const int migratetype = cc->migratetype;
2305 int ret;
2306
2307 /* Compaction run completes if the migrate and free scanner meet */
2308 if (compact_scanners_met(cc)) {
2309 /* Let the next compaction start anew. */
2310 reset_cached_positions(cc->zone);
2311
2312 /*
2313 * Mark that the PG_migrate_skip information should be cleared
2314 * by kswapd when it goes to sleep. kcompactd does not set the
2315 * flag itself as the decision to be clear should be directly
2316 * based on an allocation request.
2317 */
2318 if (cc->direct_compaction)
2319 cc->zone->compact_blockskip_flush = true;
2320
2321 if (cc->whole_zone)
2322 return COMPACT_COMPLETE;
2323 else
2324 return COMPACT_PARTIAL_SKIPPED;
2325 }
2326
2327 if (cc->proactive_compaction) {
2328 int score, wmark_low;
2329 pg_data_t *pgdat;
2330
2331 pgdat = cc->zone->zone_pgdat;
2332 if (kswapd_is_running(pgdat))
2333 return COMPACT_PARTIAL_SKIPPED;
2334
2335 score = fragmentation_score_zone(cc->zone);
2336 wmark_low = fragmentation_score_wmark(true);
2337
2338 if (score > wmark_low)
2339 ret = COMPACT_CONTINUE;
2340 else
2341 ret = COMPACT_SUCCESS;
2342
2343 goto out;
2344 }
2345
2346 if (is_via_compact_memory(cc->order))
2347 return COMPACT_CONTINUE;
2348
2349 /*
2350 * Always finish scanning a pageblock to reduce the possibility of
2351 * fallbacks in the future. This is particularly important when
2352 * migration source is unmovable/reclaimable but it's not worth
2353 * special casing.
2354 */
2355 if (!pageblock_aligned(cc->migrate_pfn))
2356 return COMPACT_CONTINUE;
2357
2358 /* Direct compactor: Is a suitable page free? */
2359 ret = COMPACT_NO_SUITABLE_PAGE;
2360 for (order = cc->order; order < NR_PAGE_ORDERS; order++) {
2361 struct free_area *area = &cc->zone->free_area[order];
2362 bool can_steal;
2363
2364 /* Job done if page is free of the right migratetype */
2365 if (!free_area_empty(area, migratetype))
2366 return COMPACT_SUCCESS;
2367
2368 #ifdef CONFIG_CMA
2369 /* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
2370 if (migratetype == MIGRATE_MOVABLE &&
2371 !free_area_empty(area, MIGRATE_CMA))
2372 return COMPACT_SUCCESS;
2373 #endif
2374 /*
2375 * Job done if allocation would steal freepages from
2376 * other migratetype buddy lists.
2377 */
2378 if (find_suitable_fallback(area, order, migratetype,
2379 true, &can_steal) != -1)
2380 /*
2381 * Movable pages are OK in any pageblock. If we are
2382 * stealing for a non-movable allocation, make sure
2383 * we finish compacting the current pageblock first
2384 * (which is assured by the above migrate_pfn align
2385 * check) so it is as free as possible and we won't
2386 * have to steal another one soon.
2387 */
2388 return COMPACT_SUCCESS;
2389 }
2390
2391 out:
2392 if (cc->contended || fatal_signal_pending(current))
2393 ret = COMPACT_CONTENDED;
2394
2395 return ret;
2396 }
2397
2398 static enum compact_result compact_finished(struct compact_control *cc)
2399 {
2400 int ret;
2401
2402 ret = __compact_finished(cc);
2403 trace_mm_compaction_finished(cc->zone, cc->order, ret);
2404 if (ret == COMPACT_NO_SUITABLE_PAGE)
2405 ret = COMPACT_CONTINUE;
2406
2407 return ret;
2408 }
2409
2410 static bool __compaction_suitable(struct zone *zone, int order,
2411 int highest_zoneidx,
2412 unsigned long wmark_target)
2413 {
2414 unsigned long watermark;
2415 /*
2416 * Watermarks for order-0 must be met for compaction to be able to
2417 * isolate free pages for migration targets. This means that the
2418 * watermark and alloc_flags have to match, or be more pessimistic than
2419 * the check in __isolate_free_page(). We don't use the direct
2420 * compactor's alloc_flags, as they are not relevant for freepage
2421 * isolation. We however do use the direct compactor's highest_zoneidx
2422 * to skip over zones where lowmem reserves would prevent allocation
2423 * even if compaction succeeds.
2424 * For costly orders, we require low watermark instead of min for
2425 * compaction to proceed to increase its chances.
2426 * ALLOC_CMA is used, as pages in CMA pageblocks are considered
2427 * suitable migration targets
2428 */
2429 watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
2430 low_wmark_pages(zone) : min_wmark_pages(zone);
2431 watermark += compact_gap(order);
2432 return __zone_watermark_ok(zone, 0, watermark, highest_zoneidx,
2433 ALLOC_CMA, wmark_target);
2434 }
2435
2436 /*
2437 * compaction_suitable: Is this suitable to run compaction on this zone now?
2438 */
2439 bool compaction_suitable(struct zone *zone, int order, int highest_zoneidx)
2440 {
2441 enum compact_result compact_result;
2442 bool suitable;
2443
2444 suitable = __compaction_suitable(zone, order, highest_zoneidx,
2445 zone_page_state(zone, NR_FREE_PAGES));
2446 /*
2447 * fragmentation index determines if allocation failures are due to
2448 * low memory or external fragmentation
2449 *
2450 * index of -1000 would imply allocations might succeed depending on
2451 * watermarks, but we already failed the high-order watermark check
2452 * index towards 0 implies failure is due to lack of memory
2453 * index towards 1000 implies failure is due to fragmentation
2454 *
2455 * Only compact if a failure would be due to fragmentation. Also
2456 * ignore fragindex for non-costly orders where the alternative to
2457 * a successful reclaim/compaction is OOM. Fragindex and the
2458 * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
2459 * excessive compaction for costly orders, but it should not be at the
2460 * expense of system stability.
2461 */
2462 if (suitable) {
2463 compact_result = COMPACT_CONTINUE;
2464 if (order > PAGE_ALLOC_COSTLY_ORDER) {
2465 int fragindex = fragmentation_index(zone, order);
2466
2467 if (fragindex >= 0 &&
2468 fragindex <= sysctl_extfrag_threshold) {
2469 suitable = false;
2470 compact_result = COMPACT_NOT_SUITABLE_ZONE;
2471 }
2472 }
2473 } else {
2474 compact_result = COMPACT_SKIPPED;
2475 }
2476
2477 trace_mm_compaction_suitable(zone, order, compact_result);
2478
2479 return suitable;
2480 }
2481
2482 bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
2483 int alloc_flags)
2484 {
2485 struct zone *zone;
2486 struct zoneref *z;
2487
2488 /*
2489 * Make sure at least one zone would pass __compaction_suitable if we continue
2490 * retrying the reclaim.
2491 */
2492 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2493 ac->highest_zoneidx, ac->nodemask) {
2494 unsigned long available;
2495
2496 /*
2497 * Do not consider all the reclaimable memory because we do not
2498 * want to trash just for a single high order allocation which
2499 * is even not guaranteed to appear even if __compaction_suitable
2500 * is happy about the watermark check.
2501 */
2502 available = zone_reclaimable_pages(zone) / order;
2503 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
2504 if (__compaction_suitable(zone, order, ac->highest_zoneidx,
2505 available))
2506 return true;
2507 }
2508
2509 return false;
2510 }
2511
2512 /*
2513 * Should we do compaction for target allocation order.
2514 * Return COMPACT_SUCCESS if allocation for target order can be already
2515 * satisfied
2516 * Return COMPACT_SKIPPED if compaction for target order is likely to fail
2517 * Return COMPACT_CONTINUE if compaction for target order should be ran
2518 */
2519 static enum compact_result
2520 compaction_suit_allocation_order(struct zone *zone, unsigned int order,
2521 int highest_zoneidx, unsigned int alloc_flags)
2522 {
2523 unsigned long watermark;
2524
2525 watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
2526 if (zone_watermark_ok(zone, order, watermark, highest_zoneidx,
2527 alloc_flags))
2528 return COMPACT_SUCCESS;
2529
2530 if (!compaction_suitable(zone, order, highest_zoneidx))
2531 return COMPACT_SKIPPED;
2532
2533 return COMPACT_CONTINUE;
2534 }
2535
2536 static enum compact_result
2537 compact_zone(struct compact_control *cc, struct capture_control *capc)
2538 {
2539 enum compact_result ret;
2540 unsigned long start_pfn = cc->zone->zone_start_pfn;
2541 unsigned long end_pfn = zone_end_pfn(cc->zone);
2542 unsigned long last_migrated_pfn;
2543 const bool sync = cc->mode != MIGRATE_ASYNC;
2544 bool update_cached;
2545 unsigned int nr_succeeded = 0, nr_migratepages;
2546 int order;
2547
2548 /*
2549 * These counters track activities during zone compaction. Initialize
2550 * them before compacting a new zone.
2551 */
2552 cc->total_migrate_scanned = 0;
2553 cc->total_free_scanned = 0;
2554 cc->nr_migratepages = 0;
2555 cc->nr_freepages = 0;
2556 for (order = 0; order < NR_PAGE_ORDERS; order++)
2557 INIT_LIST_HEAD(&cc->freepages[order]);
2558 INIT_LIST_HEAD(&cc->migratepages);
2559
2560 cc->migratetype = gfp_migratetype(cc->gfp_mask);
2561
2562 if (!is_via_compact_memory(cc->order)) {
2563 ret = compaction_suit_allocation_order(cc->zone, cc->order,
2564 cc->highest_zoneidx,
2565 cc->alloc_flags);
2566 if (ret != COMPACT_CONTINUE)
2567 return ret;
2568 }
2569
2570 /*
2571 * Clear pageblock skip if there were failures recently and compaction
2572 * is about to be retried after being deferred.
2573 */
2574 if (compaction_restarting(cc->zone, cc->order))
2575 __reset_isolation_suitable(cc->zone);
2576
2577 /*
2578 * Setup to move all movable pages to the end of the zone. Used cached
2579 * information on where the scanners should start (unless we explicitly
2580 * want to compact the whole zone), but check that it is initialised
2581 * by ensuring the values are within zone boundaries.
2582 */
2583 cc->fast_start_pfn = 0;
2584 if (cc->whole_zone) {
2585 cc->migrate_pfn = start_pfn;
2586 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2587 } else {
2588 cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync];
2589 cc->free_pfn = cc->zone->compact_cached_free_pfn;
2590 if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
2591 cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2592 cc->zone->compact_cached_free_pfn = cc->free_pfn;
2593 }
2594 if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
2595 cc->migrate_pfn = start_pfn;
2596 cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
2597 cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
2598 }
2599
2600 if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn)
2601 cc->whole_zone = true;
2602 }
2603
2604 last_migrated_pfn = 0;
2605
2606 /*
2607 * Migrate has separate cached PFNs for ASYNC and SYNC* migration on
2608 * the basis that some migrations will fail in ASYNC mode. However,
2609 * if the cached PFNs match and pageblocks are skipped due to having
2610 * no isolation candidates, then the sync state does not matter.
2611 * Until a pageblock with isolation candidates is found, keep the
2612 * cached PFNs in sync to avoid revisiting the same blocks.
2613 */
2614 update_cached = !sync &&
2615 cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1];
2616
2617 trace_mm_compaction_begin(cc, start_pfn, end_pfn, sync);
2618
2619 /* lru_add_drain_all could be expensive with involving other CPUs */
2620 lru_add_drain();
2621
2622 while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) {
2623 int err;
2624 unsigned long iteration_start_pfn = cc->migrate_pfn;
2625
2626 /*
2627 * Avoid multiple rescans of the same pageblock which can
2628 * happen if a page cannot be isolated (dirty/writeback in
2629 * async mode) or if the migrated pages are being allocated
2630 * before the pageblock is cleared. The first rescan will
2631 * capture the entire pageblock for migration. If it fails,
2632 * it'll be marked skip and scanning will proceed as normal.
2633 */
2634 cc->finish_pageblock = false;
2635 if (pageblock_start_pfn(last_migrated_pfn) ==
2636 pageblock_start_pfn(iteration_start_pfn)) {
2637 cc->finish_pageblock = true;
2638 }
2639
2640 rescan:
2641 switch (isolate_migratepages(cc)) {
2642 case ISOLATE_ABORT:
2643 ret = COMPACT_CONTENDED;
2644 putback_movable_pages(&cc->migratepages);
2645 cc->nr_migratepages = 0;
2646 goto out;
2647 case ISOLATE_NONE:
2648 if (update_cached) {
2649 cc->zone->compact_cached_migrate_pfn[1] =
2650 cc->zone->compact_cached_migrate_pfn[0];
2651 }
2652
2653 /*
2654 * We haven't isolated and migrated anything, but
2655 * there might still be unflushed migrations from
2656 * previous cc->order aligned block.
2657 */
2658 goto check_drain;
2659 case ISOLATE_SUCCESS:
2660 update_cached = false;
2661 last_migrated_pfn = max(cc->zone->zone_start_pfn,
2662 pageblock_start_pfn(cc->migrate_pfn - 1));
2663 }
2664
2665 /*
2666 * Record the number of pages to migrate since the
2667 * compaction_alloc/free() will update cc->nr_migratepages
2668 * properly.
2669 */
2670 nr_migratepages = cc->nr_migratepages;
2671 err = migrate_pages(&cc->migratepages, compaction_alloc,
2672 compaction_free, (unsigned long)cc, cc->mode,
2673 MR_COMPACTION, &nr_succeeded);
2674
2675 trace_mm_compaction_migratepages(nr_migratepages, nr_succeeded);
2676
2677 /* All pages were either migrated or will be released */
2678 cc->nr_migratepages = 0;
2679 if (err) {
2680 putback_movable_pages(&cc->migratepages);
2681 /*
2682 * migrate_pages() may return -ENOMEM when scanners meet
2683 * and we want compact_finished() to detect it
2684 */
2685 if (err == -ENOMEM && !compact_scanners_met(cc)) {
2686 ret = COMPACT_CONTENDED;
2687 goto out;
2688 }
2689 /*
2690 * If an ASYNC or SYNC_LIGHT fails to migrate a page
2691 * within the pageblock_order-aligned block and
2692 * fast_find_migrateblock may be used then scan the
2693 * remainder of the pageblock. This will mark the
2694 * pageblock "skip" to avoid rescanning in the near
2695 * future. This will isolate more pages than necessary
2696 * for the request but avoid loops due to
2697 * fast_find_migrateblock revisiting blocks that were
2698 * recently partially scanned.
2699 */
2700 if (!pageblock_aligned(cc->migrate_pfn) &&
2701 !cc->ignore_skip_hint && !cc->finish_pageblock &&
2702 (cc->mode < MIGRATE_SYNC)) {
2703 cc->finish_pageblock = true;
2704
2705 /*
2706 * Draining pcplists does not help THP if
2707 * any page failed to migrate. Even after
2708 * drain, the pageblock will not be free.
2709 */
2710 if (cc->order == COMPACTION_HPAGE_ORDER)
2711 last_migrated_pfn = 0;
2712
2713 goto rescan;
2714 }
2715 }
2716
2717 /* Stop if a page has been captured */
2718 if (capc && capc->page) {
2719 ret = COMPACT_SUCCESS;
2720 break;
2721 }
2722
2723 check_drain:
2724 /*
2725 * Has the migration scanner moved away from the previous
2726 * cc->order aligned block where we migrated from? If yes,
2727 * flush the pages that were freed, so that they can merge and
2728 * compact_finished() can detect immediately if allocation
2729 * would succeed.
2730 */
2731 if (cc->order > 0 && last_migrated_pfn) {
2732 unsigned long current_block_start =
2733 block_start_pfn(cc->migrate_pfn, cc->order);
2734
2735 if (last_migrated_pfn < current_block_start) {
2736 lru_add_drain_cpu_zone(cc->zone);
2737 /* No more flushing until we migrate again */
2738 last_migrated_pfn = 0;
2739 }
2740 }
2741 }
2742
2743 out:
2744 /*
2745 * Release free pages and update where the free scanner should restart,
2746 * so we don't leave any returned pages behind in the next attempt.
2747 */
2748 if (cc->nr_freepages > 0) {
2749 unsigned long free_pfn = release_free_list(cc->freepages);
2750
2751 cc->nr_freepages = 0;
2752 VM_BUG_ON(free_pfn == 0);
2753 /* The cached pfn is always the first in a pageblock */
2754 free_pfn = pageblock_start_pfn(free_pfn);
2755 /*
2756 * Only go back, not forward. The cached pfn might have been
2757 * already reset to zone end in compact_finished()
2758 */
2759 if (free_pfn > cc->zone->compact_cached_free_pfn)
2760 cc->zone->compact_cached_free_pfn = free_pfn;
2761 }
2762
2763 count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned);
2764 count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned);
2765
2766 trace_mm_compaction_end(cc, start_pfn, end_pfn, sync, ret);
2767
2768 VM_BUG_ON(!list_empty(&cc->migratepages));
2769
2770 return ret;
2771 }
2772
2773 static enum compact_result compact_zone_order(struct zone *zone, int order,
2774 gfp_t gfp_mask, enum compact_priority prio,
2775 unsigned int alloc_flags, int highest_zoneidx,
2776 struct page **capture)
2777 {
2778 enum compact_result ret;
2779 struct compact_control cc = {
2780 .order = order,
2781 .search_order = order,
2782 .gfp_mask = gfp_mask,
2783 .zone = zone,
2784 .mode = (prio == COMPACT_PRIO_ASYNC) ?
2785 MIGRATE_ASYNC : MIGRATE_SYNC_LIGHT,
2786 .alloc_flags = alloc_flags,
2787 .highest_zoneidx = highest_zoneidx,
2788 .direct_compaction = true,
2789 .whole_zone = (prio == MIN_COMPACT_PRIORITY),
2790 .ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
2791 .ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
2792 };
2793 struct capture_control capc = {
2794 .cc = &cc,
2795 .page = NULL,
2796 };
2797
2798 /*
2799 * Make sure the structs are really initialized before we expose the
2800 * capture control, in case we are interrupted and the interrupt handler
2801 * frees a page.
2802 */
2803 barrier();
2804 WRITE_ONCE(current->capture_control, &capc);
2805
2806 ret = compact_zone(&cc, &capc);
2807
2808 /*
2809 * Make sure we hide capture control first before we read the captured
2810 * page pointer, otherwise an interrupt could free and capture a page
2811 * and we would leak it.
2812 */
2813 WRITE_ONCE(current->capture_control, NULL);
2814 *capture = READ_ONCE(capc.page);
2815 /*
2816 * Technically, it is also possible that compaction is skipped but
2817 * the page is still captured out of luck(IRQ came and freed the page).
2818 * Returning COMPACT_SUCCESS in such cases helps in properly accounting
2819 * the COMPACT[STALL|FAIL] when compaction is skipped.
2820 */
2821 if (*capture)
2822 ret = COMPACT_SUCCESS;
2823
2824 return ret;
2825 }
2826
2827 /**
2828 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
2829 * @gfp_mask: The GFP mask of the current allocation
2830 * @order: The order of the current allocation
2831 * @alloc_flags: The allocation flags of the current allocation
2832 * @ac: The context of current allocation
2833 * @prio: Determines how hard direct compaction should try to succeed
2834 * @capture: Pointer to free page created by compaction will be stored here
2835 *
2836 * This is the main entry point for direct page compaction.
2837 */
2838 enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
2839 unsigned int alloc_flags, const struct alloc_context *ac,
2840 enum compact_priority prio, struct page **capture)
2841 {
2842 struct zoneref *z;
2843 struct zone *zone;
2844 enum compact_result rc = COMPACT_SKIPPED;
2845
2846 if (!gfp_compaction_allowed(gfp_mask))
2847 return COMPACT_SKIPPED;
2848
2849 trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
2850
2851 /* Compact each zone in the list */
2852 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2853 ac->highest_zoneidx, ac->nodemask) {
2854 enum compact_result status;
2855
2856 if (prio > MIN_COMPACT_PRIORITY
2857 && compaction_deferred(zone, order)) {
2858 rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
2859 continue;
2860 }
2861
2862 status = compact_zone_order(zone, order, gfp_mask, prio,
2863 alloc_flags, ac->highest_zoneidx, capture);
2864 rc = max(status, rc);
2865
2866 /* The allocation should succeed, stop compacting */
2867 if (status == COMPACT_SUCCESS) {
2868 /*
2869 * We think the allocation will succeed in this zone,
2870 * but it is not certain, hence the false. The caller
2871 * will repeat this with true if allocation indeed
2872 * succeeds in this zone.
2873 */
2874 compaction_defer_reset(zone, order, false);
2875
2876 break;
2877 }
2878
2879 if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
2880 status == COMPACT_PARTIAL_SKIPPED))
2881 /*
2882 * We think that allocation won't succeed in this zone
2883 * so we defer compaction there. If it ends up
2884 * succeeding after all, it will be reset.
2885 */
2886 defer_compaction(zone, order);
2887
2888 /*
2889 * We might have stopped compacting due to need_resched() in
2890 * async compaction, or due to a fatal signal detected. In that
2891 * case do not try further zones
2892 */
2893 if ((prio == COMPACT_PRIO_ASYNC && need_resched())
2894 || fatal_signal_pending(current))
2895 break;
2896 }
2897
2898 return rc;
2899 }
2900
2901 /*
2902 * compact_node() - compact all zones within a node
2903 * @pgdat: The node page data
2904 * @proactive: Whether the compaction is proactive
2905 *
2906 * For proactive compaction, compact till each zone's fragmentation score
2907 * reaches within proactive compaction thresholds (as determined by the
2908 * proactiveness tunable), it is possible that the function returns before
2909 * reaching score targets due to various back-off conditions, such as,
2910 * contention on per-node or per-zone locks.
2911 */
2912 static int compact_node(pg_data_t *pgdat, bool proactive)
2913 {
2914 int zoneid;
2915 struct zone *zone;
2916 struct compact_control cc = {
2917 .order = -1,
2918 .mode = proactive ? MIGRATE_SYNC_LIGHT : MIGRATE_SYNC,
2919 .ignore_skip_hint = true,
2920 .whole_zone = true,
2921 .gfp_mask = GFP_KERNEL,
2922 .proactive_compaction = proactive,
2923 };
2924
2925 for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2926 zone = &pgdat->node_zones[zoneid];
2927 if (!populated_zone(zone))
2928 continue;
2929
2930 if (fatal_signal_pending(current))
2931 return -EINTR;
2932
2933 cc.zone = zone;
2934
2935 compact_zone(&cc, NULL);
2936
2937 if (proactive) {
2938 count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
2939 cc.total_migrate_scanned);
2940 count_compact_events(KCOMPACTD_FREE_SCANNED,
2941 cc.total_free_scanned);
2942 }
2943 }
2944
2945 return 0;
2946 }
2947
2948 /* Compact all zones of all nodes in the system */
2949 static int compact_nodes(void)
2950 {
2951 int ret, nid;
2952
2953 /* Flush pending updates to the LRU lists */
2954 lru_add_drain_all();
2955
2956 for_each_online_node(nid) {
2957 ret = compact_node(NODE_DATA(nid), false);
2958 if (ret)
2959 return ret;
2960 }
2961
2962 return 0;
2963 }
2964
2965 static int compaction_proactiveness_sysctl_handler(const struct ctl_table *table, int write,
2966 void *buffer, size_t *length, loff_t *ppos)
2967 {
2968 int rc, nid;
2969
2970 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
2971 if (rc)
2972 return rc;
2973
2974 if (write && sysctl_compaction_proactiveness) {
2975 for_each_online_node(nid) {
2976 pg_data_t *pgdat = NODE_DATA(nid);
2977
2978 if (pgdat->proactive_compact_trigger)
2979 continue;
2980
2981 pgdat->proactive_compact_trigger = true;
2982 trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, -1,
2983 pgdat->nr_zones - 1);
2984 wake_up_interruptible(&pgdat->kcompactd_wait);
2985 }
2986 }
2987
2988 return 0;
2989 }
2990
2991 /*
2992 * This is the entry point for compacting all nodes via
2993 * /proc/sys/vm/compact_memory
2994 */
2995 static int sysctl_compaction_handler(const struct ctl_table *table, int write,
2996 void *buffer, size_t *length, loff_t *ppos)
2997 {
2998 int ret;
2999
3000 ret = proc_dointvec(table, write, buffer, length, ppos);
3001 if (ret)
3002 return ret;
3003
3004 if (sysctl_compact_memory != 1)
3005 return -EINVAL;
3006
3007 if (write)
3008 ret = compact_nodes();
3009
3010 return ret;
3011 }
3012
3013 #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
3014 static ssize_t compact_store(struct device *dev,
3015 struct device_attribute *attr,
3016 const char *buf, size_t count)
3017 {
3018 int nid = dev->id;
3019
3020 if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
3021 /* Flush pending updates to the LRU lists */
3022 lru_add_drain_all();
3023
3024 compact_node(NODE_DATA(nid), false);
3025 }
3026
3027 return count;
3028 }
3029 static DEVICE_ATTR_WO(compact);
3030
3031 int compaction_register_node(struct node *node)
3032 {
3033 return device_create_file(&node->dev, &dev_attr_compact);
3034 }
3035
3036 void compaction_unregister_node(struct node *node)
3037 {
3038 device_remove_file(&node->dev, &dev_attr_compact);
3039 }
3040 #endif /* CONFIG_SYSFS && CONFIG_NUMA */
3041
3042 static inline bool kcompactd_work_requested(pg_data_t *pgdat)
3043 {
3044 return pgdat->kcompactd_max_order > 0 || kthread_should_stop() ||
3045 pgdat->proactive_compact_trigger;
3046 }
3047
3048 static bool kcompactd_node_suitable(pg_data_t *pgdat)
3049 {
3050 int zoneid;
3051 struct zone *zone;
3052 enum zone_type highest_zoneidx = pgdat->kcompactd_highest_zoneidx;
3053 enum compact_result ret;
3054
3055 for (zoneid = 0; zoneid <= highest_zoneidx; zoneid++) {
3056 zone = &pgdat->node_zones[zoneid];
3057
3058 if (!populated_zone(zone))
3059 continue;
3060
3061 ret = compaction_suit_allocation_order(zone,
3062 pgdat->kcompactd_max_order,
3063 highest_zoneidx, ALLOC_WMARK_MIN);
3064 if (ret == COMPACT_CONTINUE)
3065 return true;
3066 }
3067
3068 return false;
3069 }
3070
3071 static void kcompactd_do_work(pg_data_t *pgdat)
3072 {
3073 /*
3074 * With no special task, compact all zones so that a page of requested
3075 * order is allocatable.
3076 */
3077 int zoneid;
3078 struct zone *zone;
3079 struct compact_control cc = {
3080 .order = pgdat->kcompactd_max_order,
3081 .search_order = pgdat->kcompactd_max_order,
3082 .highest_zoneidx = pgdat->kcompactd_highest_zoneidx,
3083 .mode = MIGRATE_SYNC_LIGHT,
3084 .ignore_skip_hint = false,
3085 .gfp_mask = GFP_KERNEL,
3086 };
3087 enum compact_result ret;
3088
3089 trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
3090 cc.highest_zoneidx);
3091 count_compact_event(KCOMPACTD_WAKE);
3092
3093 for (zoneid = 0; zoneid <= cc.highest_zoneidx; zoneid++) {
3094 int status;
3095
3096 zone = &pgdat->node_zones[zoneid];
3097 if (!populated_zone(zone))
3098 continue;
3099
3100 if (compaction_deferred(zone, cc.order))
3101 continue;
3102
3103 ret = compaction_suit_allocation_order(zone,
3104 cc.order, zoneid, ALLOC_WMARK_MIN);
3105 if (ret != COMPACT_CONTINUE)
3106 continue;
3107
3108 if (kthread_should_stop())
3109 return;
3110
3111 cc.zone = zone;
3112 status = compact_zone(&cc, NULL);
3113
3114 if (status == COMPACT_SUCCESS) {
3115 compaction_defer_reset(zone, cc.order, false);
3116 } else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
3117 /*
3118 * Buddy pages may become stranded on pcps that could
3119 * otherwise coalesce on the zone's free area for
3120 * order >= cc.order. This is ratelimited by the
3121 * upcoming deferral.
3122 */
3123 drain_all_pages(zone);
3124
3125 /*
3126 * We use sync migration mode here, so we defer like
3127 * sync direct compaction does.
3128 */
3129 defer_compaction(zone, cc.order);
3130 }
3131
3132 count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
3133 cc.total_migrate_scanned);
3134 count_compact_events(KCOMPACTD_FREE_SCANNED,
3135 cc.total_free_scanned);
3136 }
3137
3138 /*
3139 * Regardless of success, we are done until woken up next. But remember
3140 * the requested order/highest_zoneidx in case it was higher/tighter
3141 * than our current ones
3142 */
3143 if (pgdat->kcompactd_max_order <= cc.order)
3144 pgdat->kcompactd_max_order = 0;
3145 if (pgdat->kcompactd_highest_zoneidx >= cc.highest_zoneidx)
3146 pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
3147 }
3148
3149 void wakeup_kcompactd(pg_data_t *pgdat, int order, int highest_zoneidx)
3150 {
3151 if (!order)
3152 return;
3153
3154 if (pgdat->kcompactd_max_order < order)
3155 pgdat->kcompactd_max_order = order;
3156
3157 if (pgdat->kcompactd_highest_zoneidx > highest_zoneidx)
3158 pgdat->kcompactd_highest_zoneidx = highest_zoneidx;
3159
3160 /*
3161 * Pairs with implicit barrier in wait_event_freezable()
3162 * such that wakeups are not missed.
3163 */
3164 if (!wq_has_sleeper(&pgdat->kcompactd_wait))
3165 return;
3166
3167 if (!kcompactd_node_suitable(pgdat))
3168 return;
3169
3170 trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
3171 highest_zoneidx);
3172 wake_up_interruptible(&pgdat->kcompactd_wait);
3173 }
3174
3175 /*
3176 * The background compaction daemon, started as a kernel thread
3177 * from the init process.
3178 */
3179 static int kcompactd(void *p)
3180 {
3181 pg_data_t *pgdat = (pg_data_t *)p;
3182 struct task_struct *tsk = current;
3183 long default_timeout = msecs_to_jiffies(HPAGE_FRAG_CHECK_INTERVAL_MSEC);
3184 long timeout = default_timeout;
3185
3186 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3187
3188 if (!cpumask_empty(cpumask))
3189 set_cpus_allowed_ptr(tsk, cpumask);
3190
3191 set_freezable();
3192
3193 pgdat->kcompactd_max_order = 0;
3194 pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
3195
3196 while (!kthread_should_stop()) {
3197 unsigned long pflags;
3198
3199 /*
3200 * Avoid the unnecessary wakeup for proactive compaction
3201 * when it is disabled.
3202 */
3203 if (!sysctl_compaction_proactiveness)
3204 timeout = MAX_SCHEDULE_TIMEOUT;
3205 trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
3206 if (wait_event_freezable_timeout(pgdat->kcompactd_wait,
3207 kcompactd_work_requested(pgdat), timeout) &&
3208 !pgdat->proactive_compact_trigger) {
3209
3210 psi_memstall_enter(&pflags);
3211 kcompactd_do_work(pgdat);
3212 psi_memstall_leave(&pflags);
3213 /*
3214 * Reset the timeout value. The defer timeout from
3215 * proactive compaction is lost here but that is fine
3216 * as the condition of the zone changing substantionally
3217 * then carrying on with the previous defer interval is
3218 * not useful.
3219 */
3220 timeout = default_timeout;
3221 continue;
3222 }
3223
3224 /*
3225 * Start the proactive work with default timeout. Based
3226 * on the fragmentation score, this timeout is updated.
3227 */
3228 timeout = default_timeout;
3229 if (should_proactive_compact_node(pgdat)) {
3230 unsigned int prev_score, score;
3231
3232 prev_score = fragmentation_score_node(pgdat);
3233 compact_node(pgdat, true);
3234 score = fragmentation_score_node(pgdat);
3235 /*
3236 * Defer proactive compaction if the fragmentation
3237 * score did not go down i.e. no progress made.
3238 */
3239 if (unlikely(score >= prev_score))
3240 timeout =
3241 default_timeout << COMPACT_MAX_DEFER_SHIFT;
3242 }
3243 if (unlikely(pgdat->proactive_compact_trigger))
3244 pgdat->proactive_compact_trigger = false;
3245 }
3246
3247 return 0;
3248 }
3249
3250 /*
3251 * This kcompactd start function will be called by init and node-hot-add.
3252 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
3253 */
3254 void __meminit kcompactd_run(int nid)
3255 {
3256 pg_data_t *pgdat = NODE_DATA(nid);
3257
3258 if (pgdat->kcompactd)
3259 return;
3260
3261 pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
3262 if (IS_ERR(pgdat->kcompactd)) {
3263 pr_err("Failed to start kcompactd on node %d\n", nid);
3264 pgdat->kcompactd = NULL;
3265 }
3266 }
3267
3268 /*
3269 * Called by memory hotplug when all memory in a node is offlined. Caller must
3270 * be holding mem_hotplug_begin/done().
3271 */
3272 void __meminit kcompactd_stop(int nid)
3273 {
3274 struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
3275
3276 if (kcompactd) {
3277 kthread_stop(kcompactd);
3278 NODE_DATA(nid)->kcompactd = NULL;
3279 }
3280 }
3281
3282 /*
3283 * It's optimal to keep kcompactd on the same CPUs as their memory, but
3284 * not required for correctness. So if the last cpu in a node goes
3285 * away, we get changed to run anywhere: as the first one comes back,
3286 * restore their cpu bindings.
3287 */
3288 static int kcompactd_cpu_online(unsigned int cpu)
3289 {
3290 int nid;
3291
3292 for_each_node_state(nid, N_MEMORY) {
3293 pg_data_t *pgdat = NODE_DATA(nid);
3294 const struct cpumask *mask;
3295
3296 mask = cpumask_of_node(pgdat->node_id);
3297
3298 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3299 /* One of our CPUs online: restore mask */
3300 if (pgdat->kcompactd)
3301 set_cpus_allowed_ptr(pgdat->kcompactd, mask);
3302 }
3303 return 0;
3304 }
3305
3306 static int proc_dointvec_minmax_warn_RT_change(const struct ctl_table *table,
3307 int write, void *buffer, size_t *lenp, loff_t *ppos)
3308 {
3309 int ret, old;
3310
3311 if (!IS_ENABLED(CONFIG_PREEMPT_RT) || !write)
3312 return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
3313
3314 old = *(int *)table->data;
3315 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
3316 if (ret)
3317 return ret;
3318 if (old != *(int *)table->data)
3319 pr_warn_once("sysctl attribute %s changed by %s[%d]\n",
3320 table->procname, current->comm,
3321 task_pid_nr(current));
3322 return ret;
3323 }
3324
3325 static struct ctl_table vm_compaction[] = {
3326 {
3327 .procname = "compact_memory",
3328 .data = &sysctl_compact_memory,
3329 .maxlen = sizeof(int),
3330 .mode = 0200,
3331 .proc_handler = sysctl_compaction_handler,
3332 },
3333 {
3334 .procname = "compaction_proactiveness",
3335 .data = &sysctl_compaction_proactiveness,
3336 .maxlen = sizeof(sysctl_compaction_proactiveness),
3337 .mode = 0644,
3338 .proc_handler = compaction_proactiveness_sysctl_handler,
3339 .extra1 = SYSCTL_ZERO,
3340 .extra2 = SYSCTL_ONE_HUNDRED,
3341 },
3342 {
3343 .procname = "extfrag_threshold",
3344 .data = &sysctl_extfrag_threshold,
3345 .maxlen = sizeof(int),
3346 .mode = 0644,
3347 .proc_handler = proc_dointvec_minmax,
3348 .extra1 = SYSCTL_ZERO,
3349 .extra2 = SYSCTL_ONE_THOUSAND,
3350 },
3351 {
3352 .procname = "compact_unevictable_allowed",
3353 .data = &sysctl_compact_unevictable_allowed,
3354 .maxlen = sizeof(int),
3355 .mode = 0644,
3356 .proc_handler = proc_dointvec_minmax_warn_RT_change,
3357 .extra1 = SYSCTL_ZERO,
3358 .extra2 = SYSCTL_ONE,
3359 },
3360 };
3361
3362 static int __init kcompactd_init(void)
3363 {
3364 int nid;
3365 int ret;
3366
3367 ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3368 "mm/compaction:online",
3369 kcompactd_cpu_online, NULL);
3370 if (ret < 0) {
3371 pr_err("kcompactd: failed to register hotplug callbacks.\n");
3372 return ret;
3373 }
3374
3375 for_each_node_state(nid, N_MEMORY)
3376 kcompactd_run(nid);
3377 register_sysctl_init("vm", vm_compaction);
3378 return 0;
3379 }
3380 subsys_initcall(kcompactd_init)
3381
3382 #endif /* CONFIG_COMPACTION */
3383
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.