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