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