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