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