Linux Audio

Check our new training course

Loading...
v6.8
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * linux/mm/compaction.c
   4 *
   5 * Memory compaction for the reduction of external fragmentation. Note that
   6 * this heavily depends upon page migration to do all the real heavy
   7 * lifting
   8 *
   9 * Copyright IBM Corp. 2007-2010 Mel Gorman <mel@csn.ul.ie>
  10 */
  11#include <linux/cpu.h>
  12#include <linux/swap.h>
  13#include <linux/migrate.h>
  14#include <linux/compaction.h>
  15#include <linux/mm_inline.h>
  16#include <linux/sched/signal.h>
  17#include <linux/backing-dev.h>
  18#include <linux/sysctl.h>
  19#include <linux/sysfs.h>
  20#include <linux/page-isolation.h>
  21#include <linux/kasan.h>
  22#include <linux/kthread.h>
  23#include <linux/freezer.h>
  24#include <linux/page_owner.h>
  25#include <linux/psi.h>
  26#include "internal.h"
  27
  28#ifdef CONFIG_COMPACTION
  29/*
  30 * Fragmentation score check interval for proactive compaction purposes.
  31 */
  32#define HPAGE_FRAG_CHECK_INTERVAL_MSEC	(500)
  33
  34static inline void count_compact_event(enum vm_event_item item)
  35{
  36	count_vm_event(item);
  37}
  38
  39static inline void count_compact_events(enum vm_event_item item, long delta)
  40{
  41	count_vm_events(item, delta);
  42}
 
 
 
 
 
 
 
 
 
 
 
 
  43#else
  44#define count_compact_event(item) do { } while (0)
  45#define count_compact_events(item, delta) do { } while (0)
 
  46#endif
  47
  48#if defined CONFIG_COMPACTION || defined CONFIG_CMA
  49
  50#define CREATE_TRACE_POINTS
  51#include <trace/events/compaction.h>
  52
  53#define block_start_pfn(pfn, order)	round_down(pfn, 1UL << (order))
  54#define block_end_pfn(pfn, order)	ALIGN((pfn) + 1, 1UL << (order))
  55
  56/*
  57 * Page order with-respect-to which proactive compaction
  58 * calculates external fragmentation, which is used as
  59 * the "fragmentation score" of a node/zone.
  60 */
  61#if defined CONFIG_TRANSPARENT_HUGEPAGE
  62#define COMPACTION_HPAGE_ORDER	HPAGE_PMD_ORDER
  63#elif defined CONFIG_HUGETLBFS
  64#define COMPACTION_HPAGE_ORDER	HUGETLB_PAGE_ORDER
  65#else
  66#define COMPACTION_HPAGE_ORDER	(PMD_SHIFT - PAGE_SHIFT)
  67#endif
  68
  69static unsigned long release_freepages(struct list_head *freelist)
  70{
 
  71	struct page *page, *next;
  72	unsigned long high_pfn = 0;
  73
  74	list_for_each_entry_safe(page, next, freelist, lru) {
  75		unsigned long pfn = page_to_pfn(page);
  76		list_del(&page->lru);
  77		__free_page(page);
  78		if (pfn > high_pfn)
  79			high_pfn = pfn;
  80	}
  81
  82	return high_pfn;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  83}
  84
  85static void split_map_pages(struct list_head *list)
  86{
  87	unsigned int i, order, nr_pages;
  88	struct page *page, *next;
  89	LIST_HEAD(tmp_list);
  90
  91	list_for_each_entry_safe(page, next, list, lru) {
  92		list_del(&page->lru);
  93
  94		order = page_private(page);
  95		nr_pages = 1 << order;
  96
  97		post_alloc_hook(page, order, __GFP_MOVABLE);
  98		if (order)
  99			split_page(page, order);
 100
 101		for (i = 0; i < nr_pages; i++) {
 102			list_add(&page->lru, &tmp_list);
 103			page++;
 
 
 
 
 
 
 104		}
 105	}
 106
 107	list_splice(&tmp_list, list);
 108}
 109
 110#ifdef CONFIG_COMPACTION
 111bool PageMovable(struct page *page)
 112{
 113	const struct movable_operations *mops;
 114
 115	VM_BUG_ON_PAGE(!PageLocked(page), page);
 116	if (!__PageMovable(page))
 117		return false;
 118
 119	mops = page_movable_ops(page);
 120	if (mops)
 121		return true;
 122
 123	return false;
 124}
 125
 126void __SetPageMovable(struct page *page, const struct movable_operations *mops)
 127{
 128	VM_BUG_ON_PAGE(!PageLocked(page), page);
 129	VM_BUG_ON_PAGE((unsigned long)mops & PAGE_MAPPING_MOVABLE, page);
 130	page->mapping = (void *)((unsigned long)mops | PAGE_MAPPING_MOVABLE);
 131}
 132EXPORT_SYMBOL(__SetPageMovable);
 133
 134void __ClearPageMovable(struct page *page)
 135{
 136	VM_BUG_ON_PAGE(!PageMovable(page), page);
 137	/*
 138	 * This page still has the type of a movable page, but it's
 139	 * actually not movable any more.
 140	 */
 141	page->mapping = (void *)PAGE_MAPPING_MOVABLE;
 142}
 143EXPORT_SYMBOL(__ClearPageMovable);
 144
 145/* Do not skip compaction more than 64 times */
 146#define COMPACT_MAX_DEFER_SHIFT 6
 147
 148/*
 149 * Compaction is deferred when compaction fails to result in a page
 150 * allocation success. 1 << compact_defer_shift, compactions are skipped up
 151 * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
 152 */
 153static void defer_compaction(struct zone *zone, int order)
 154{
 155	zone->compact_considered = 0;
 156	zone->compact_defer_shift++;
 157
 158	if (order < zone->compact_order_failed)
 159		zone->compact_order_failed = order;
 160
 161	if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
 162		zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
 163
 164	trace_mm_compaction_defer_compaction(zone, order);
 165}
 166
 167/* Returns true if compaction should be skipped this time */
 168static bool compaction_deferred(struct zone *zone, int order)
 169{
 170	unsigned long defer_limit = 1UL << zone->compact_defer_shift;
 171
 172	if (order < zone->compact_order_failed)
 173		return false;
 174
 175	/* Avoid possible overflow */
 176	if (++zone->compact_considered >= defer_limit) {
 177		zone->compact_considered = defer_limit;
 178		return false;
 179	}
 180
 181	trace_mm_compaction_deferred(zone, order);
 182
 183	return true;
 184}
 185
 186/*
 187 * Update defer tracking counters after successful compaction of given order,
 188 * which means an allocation either succeeded (alloc_success == true) or is
 189 * expected to succeed.
 190 */
 191void compaction_defer_reset(struct zone *zone, int order,
 192		bool alloc_success)
 193{
 194	if (alloc_success) {
 195		zone->compact_considered = 0;
 196		zone->compact_defer_shift = 0;
 197	}
 198	if (order >= zone->compact_order_failed)
 199		zone->compact_order_failed = order + 1;
 200
 201	trace_mm_compaction_defer_reset(zone, order);
 202}
 203
 204/* Returns true if restarting compaction after many failures */
 205static bool compaction_restarting(struct zone *zone, int order)
 206{
 207	if (order < zone->compact_order_failed)
 208		return false;
 209
 210	return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
 211		zone->compact_considered >= 1UL << zone->compact_defer_shift;
 212}
 213
 214/* Returns true if the pageblock should be scanned for pages to isolate. */
 215static inline bool isolation_suitable(struct compact_control *cc,
 216					struct page *page)
 217{
 218	if (cc->ignore_skip_hint)
 219		return true;
 220
 221	return !get_pageblock_skip(page);
 222}
 223
 224static void reset_cached_positions(struct zone *zone)
 225{
 226	zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
 227	zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
 228	zone->compact_cached_free_pfn =
 229				pageblock_start_pfn(zone_end_pfn(zone) - 1);
 230}
 231
 232#ifdef CONFIG_SPARSEMEM
 233/*
 234 * If the PFN falls into an offline section, return the start PFN of the
 235 * next online section. If the PFN falls into an online section or if
 236 * there is no next online section, return 0.
 237 */
 238static unsigned long skip_offline_sections(unsigned long start_pfn)
 239{
 240	unsigned long start_nr = pfn_to_section_nr(start_pfn);
 241
 242	if (online_section_nr(start_nr))
 243		return 0;
 244
 245	while (++start_nr <= __highest_present_section_nr) {
 246		if (online_section_nr(start_nr))
 247			return section_nr_to_pfn(start_nr);
 248	}
 249
 250	return 0;
 251}
 252
 253/*
 254 * If the PFN falls into an offline section, return the end PFN of the
 255 * next online section in reverse. If the PFN falls into an online section
 256 * or if there is no next online section in reverse, return 0.
 257 */
 258static unsigned long skip_offline_sections_reverse(unsigned long start_pfn)
 259{
 260	unsigned long start_nr = pfn_to_section_nr(start_pfn);
 261
 262	if (!start_nr || online_section_nr(start_nr))
 263		return 0;
 264
 265	while (start_nr-- > 0) {
 266		if (online_section_nr(start_nr))
 267			return section_nr_to_pfn(start_nr) + PAGES_PER_SECTION;
 268	}
 269
 270	return 0;
 271}
 272#else
 273static unsigned long skip_offline_sections(unsigned long start_pfn)
 274{
 275	return 0;
 276}
 277
 278static unsigned long skip_offline_sections_reverse(unsigned long start_pfn)
 279{
 280	return 0;
 281}
 282#endif
 283
 284/*
 285 * Compound pages of >= pageblock_order should consistently be skipped until
 286 * released. It is always pointless to compact pages of such order (if they are
 287 * migratable), and the pageblocks they occupy cannot contain any free pages.
 288 */
 289static bool pageblock_skip_persistent(struct page *page)
 290{
 291	if (!PageCompound(page))
 292		return false;
 293
 294	page = compound_head(page);
 295
 296	if (compound_order(page) >= pageblock_order)
 297		return true;
 298
 299	return false;
 300}
 301
 302static bool
 303__reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source,
 304							bool check_target)
 305{
 306	struct page *page = pfn_to_online_page(pfn);
 307	struct page *block_page;
 308	struct page *end_page;
 309	unsigned long block_pfn;
 310
 311	if (!page)
 312		return false;
 313	if (zone != page_zone(page))
 314		return false;
 315	if (pageblock_skip_persistent(page))
 316		return false;
 317
 318	/*
 319	 * If skip is already cleared do no further checking once the
 320	 * restart points have been set.
 321	 */
 322	if (check_source && check_target && !get_pageblock_skip(page))
 323		return true;
 324
 325	/*
 326	 * If clearing skip for the target scanner, do not select a
 327	 * non-movable pageblock as the starting point.
 328	 */
 329	if (!check_source && check_target &&
 330	    get_pageblock_migratetype(page) != MIGRATE_MOVABLE)
 331		return false;
 332
 333	/* Ensure the start of the pageblock or zone is online and valid */
 334	block_pfn = pageblock_start_pfn(pfn);
 335	block_pfn = max(block_pfn, zone->zone_start_pfn);
 336	block_page = pfn_to_online_page(block_pfn);
 337	if (block_page) {
 338		page = block_page;
 339		pfn = block_pfn;
 340	}
 341
 342	/* Ensure the end of the pageblock or zone is online and valid */
 343	block_pfn = pageblock_end_pfn(pfn) - 1;
 344	block_pfn = min(block_pfn, zone_end_pfn(zone) - 1);
 345	end_page = pfn_to_online_page(block_pfn);
 346	if (!end_page)
 347		return false;
 348
 349	/*
 350	 * Only clear the hint if a sample indicates there is either a
 351	 * free page or an LRU page in the block. One or other condition
 352	 * is necessary for the block to be a migration source/target.
 353	 */
 354	do {
 355		if (check_source && PageLRU(page)) {
 356			clear_pageblock_skip(page);
 357			return true;
 358		}
 359
 360		if (check_target && PageBuddy(page)) {
 361			clear_pageblock_skip(page);
 362			return true;
 363		}
 364
 365		page += (1 << PAGE_ALLOC_COSTLY_ORDER);
 366	} while (page <= end_page);
 367
 368	return false;
 369}
 370
 371/*
 372 * This function is called to clear all cached information on pageblocks that
 373 * should be skipped for page isolation when the migrate and free page scanner
 374 * meet.
 375 */
 376static void __reset_isolation_suitable(struct zone *zone)
 377{
 378	unsigned long migrate_pfn = zone->zone_start_pfn;
 379	unsigned long free_pfn = zone_end_pfn(zone) - 1;
 380	unsigned long reset_migrate = free_pfn;
 381	unsigned long reset_free = migrate_pfn;
 382	bool source_set = false;
 383	bool free_set = false;
 384
 385	/* Only flush if a full compaction finished recently */
 386	if (!zone->compact_blockskip_flush)
 387		return;
 388
 389	zone->compact_blockskip_flush = false;
 390
 391	/*
 392	 * Walk the zone and update pageblock skip information. Source looks
 393	 * for PageLRU while target looks for PageBuddy. When the scanner
 394	 * is found, both PageBuddy and PageLRU are checked as the pageblock
 395	 * is suitable as both source and target.
 396	 */
 397	for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages,
 398					free_pfn -= pageblock_nr_pages) {
 399		cond_resched();
 400
 401		/* Update the migrate PFN */
 402		if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) &&
 403		    migrate_pfn < reset_migrate) {
 404			source_set = true;
 405			reset_migrate = migrate_pfn;
 406			zone->compact_init_migrate_pfn = reset_migrate;
 407			zone->compact_cached_migrate_pfn[0] = reset_migrate;
 408			zone->compact_cached_migrate_pfn[1] = reset_migrate;
 409		}
 410
 411		/* Update the free PFN */
 412		if (__reset_isolation_pfn(zone, free_pfn, free_set, true) &&
 413		    free_pfn > reset_free) {
 414			free_set = true;
 415			reset_free = free_pfn;
 416			zone->compact_init_free_pfn = reset_free;
 417			zone->compact_cached_free_pfn = reset_free;
 418		}
 419	}
 420
 421	/* Leave no distance if no suitable block was reset */
 422	if (reset_migrate >= reset_free) {
 423		zone->compact_cached_migrate_pfn[0] = migrate_pfn;
 424		zone->compact_cached_migrate_pfn[1] = migrate_pfn;
 425		zone->compact_cached_free_pfn = free_pfn;
 426	}
 427}
 428
 429void reset_isolation_suitable(pg_data_t *pgdat)
 430{
 431	int zoneid;
 432
 433	for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
 434		struct zone *zone = &pgdat->node_zones[zoneid];
 435		if (!populated_zone(zone))
 436			continue;
 437
 438		__reset_isolation_suitable(zone);
 439	}
 440}
 441
 442/*
 443 * Sets the pageblock skip bit if it was clear. Note that this is a hint as
 444 * locks are not required for read/writers. Returns true if it was already set.
 445 */
 446static bool test_and_set_skip(struct compact_control *cc, struct page *page)
 447{
 448	bool skip;
 449
 450	/* Do not update if skip hint is being ignored */
 451	if (cc->ignore_skip_hint)
 452		return false;
 453
 454	skip = get_pageblock_skip(page);
 455	if (!skip && !cc->no_set_skip_hint)
 456		set_pageblock_skip(page);
 457
 458	return skip;
 459}
 460
 461static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
 462{
 463	struct zone *zone = cc->zone;
 464
 465	/* Set for isolation rather than compaction */
 466	if (cc->no_set_skip_hint)
 467		return;
 468
 469	pfn = pageblock_end_pfn(pfn);
 470
 471	/* Update where async and sync compaction should restart */
 472	if (pfn > zone->compact_cached_migrate_pfn[0])
 473		zone->compact_cached_migrate_pfn[0] = pfn;
 474	if (cc->mode != MIGRATE_ASYNC &&
 475	    pfn > zone->compact_cached_migrate_pfn[1])
 476		zone->compact_cached_migrate_pfn[1] = pfn;
 477}
 478
 479/*
 480 * If no pages were isolated then mark this pageblock to be skipped in the
 481 * future. The information is later cleared by __reset_isolation_suitable().
 482 */
 483static void update_pageblock_skip(struct compact_control *cc,
 484			struct page *page, unsigned long pfn)
 485{
 486	struct zone *zone = cc->zone;
 487
 488	if (cc->no_set_skip_hint)
 489		return;
 490
 491	set_pageblock_skip(page);
 492
 493	if (pfn < zone->compact_cached_free_pfn)
 494		zone->compact_cached_free_pfn = pfn;
 495}
 496#else
 497static inline bool isolation_suitable(struct compact_control *cc,
 498					struct page *page)
 499{
 500	return true;
 501}
 502
 503static inline bool pageblock_skip_persistent(struct page *page)
 504{
 505	return false;
 506}
 507
 508static inline void update_pageblock_skip(struct compact_control *cc,
 509			struct page *page, unsigned long pfn)
 510{
 511}
 512
 513static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
 514{
 515}
 516
 517static bool test_and_set_skip(struct compact_control *cc, struct page *page)
 518{
 519	return false;
 520}
 521#endif /* CONFIG_COMPACTION */
 522
 523/*
 524 * Compaction requires the taking of some coarse locks that are potentially
 525 * very heavily contended. For async compaction, trylock and record if the
 526 * lock is contended. The lock will still be acquired but compaction will
 527 * abort when the current block is finished regardless of success rate.
 528 * Sync compaction acquires the lock.
 529 *
 530 * Always returns true which makes it easier to track lock state in callers.
 531 */
 532static bool compact_lock_irqsave(spinlock_t *lock, unsigned long *flags,
 533						struct compact_control *cc)
 534	__acquires(lock)
 535{
 536	/* Track if the lock is contended in async mode */
 537	if (cc->mode == MIGRATE_ASYNC && !cc->contended) {
 538		if (spin_trylock_irqsave(lock, *flags))
 539			return true;
 540
 541		cc->contended = true;
 542	}
 543
 544	spin_lock_irqsave(lock, *flags);
 545	return true;
 546}
 547
 548/*
 549 * Compaction requires the taking of some coarse locks that are potentially
 550 * very heavily contended. The lock should be periodically unlocked to avoid
 551 * having disabled IRQs for a long time, even when there is nobody waiting on
 552 * the lock. It might also be that allowing the IRQs will result in
 553 * need_resched() becoming true. If scheduling is needed, compaction schedules.
 554 * Either compaction type will also abort if a fatal signal is pending.
 555 * In either case if the lock was locked, it is dropped and not regained.
 556 *
 557 * Returns true if compaction should abort due to fatal signal pending.
 558 * Returns false when compaction can continue.
 559 */
 560static bool compact_unlock_should_abort(spinlock_t *lock,
 561		unsigned long flags, bool *locked, struct compact_control *cc)
 562{
 563	if (*locked) {
 564		spin_unlock_irqrestore(lock, flags);
 565		*locked = false;
 566	}
 567
 568	if (fatal_signal_pending(current)) {
 569		cc->contended = true;
 570		return true;
 571	}
 572
 573	cond_resched();
 574
 575	return false;
 576}
 577
 578/*
 579 * Isolate free pages onto a private freelist. If @strict is true, will abort
 580 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
 581 * (even though it may still end up isolating some pages).
 582 */
 583static unsigned long isolate_freepages_block(struct compact_control *cc,
 584				unsigned long *start_pfn,
 585				unsigned long end_pfn,
 586				struct list_head *freelist,
 587				unsigned int stride,
 588				bool strict)
 589{
 590	int nr_scanned = 0, total_isolated = 0;
 591	struct page *page;
 592	unsigned long flags = 0;
 593	bool locked = false;
 594	unsigned long blockpfn = *start_pfn;
 595	unsigned int order;
 596
 597	/* Strict mode is for isolation, speed is secondary */
 598	if (strict)
 599		stride = 1;
 600
 601	page = pfn_to_page(blockpfn);
 602
 603	/* Isolate free pages. */
 604	for (; blockpfn < end_pfn; blockpfn += stride, page += stride) {
 605		int isolated;
 606
 607		/*
 608		 * Periodically drop the lock (if held) regardless of its
 609		 * contention, to give chance to IRQs. Abort if fatal signal
 610		 * pending.
 611		 */
 612		if (!(blockpfn % COMPACT_CLUSTER_MAX)
 613		    && compact_unlock_should_abort(&cc->zone->lock, flags,
 614								&locked, cc))
 615			break;
 616
 617		nr_scanned++;
 618
 619		/*
 620		 * For compound pages such as THP and hugetlbfs, we can save
 621		 * potentially a lot of iterations if we skip them at once.
 622		 * The check is racy, but we can consider only valid values
 623		 * and the only danger is skipping too much.
 624		 */
 625		if (PageCompound(page)) {
 626			const unsigned int order = compound_order(page);
 627
 628			if (blockpfn + (1UL << order) <= end_pfn) {
 629				blockpfn += (1UL << order) - 1;
 630				page += (1UL << order) - 1;
 631				nr_scanned += (1UL << order) - 1;
 632			}
 633
 634			goto isolate_fail;
 635		}
 636
 637		if (!PageBuddy(page))
 638			goto isolate_fail;
 639
 640		/* If we already hold the lock, we can skip some rechecking. */
 641		if (!locked) {
 642			locked = compact_lock_irqsave(&cc->zone->lock,
 643								&flags, cc);
 644
 645			/* Recheck this is a buddy page under lock */
 646			if (!PageBuddy(page))
 647				goto isolate_fail;
 648		}
 649
 650		/* Found a free page, will break it into order-0 pages */
 651		order = buddy_order(page);
 652		isolated = __isolate_free_page(page, order);
 653		if (!isolated)
 654			break;
 655		set_page_private(page, order);
 656
 657		nr_scanned += isolated - 1;
 658		total_isolated += isolated;
 659		cc->nr_freepages += isolated;
 660		list_add_tail(&page->lru, freelist);
 661
 662		if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
 663			blockpfn += isolated;
 664			break;
 665		}
 666		/* Advance to the end of split page */
 667		blockpfn += isolated - 1;
 668		page += isolated - 1;
 669		continue;
 670
 671isolate_fail:
 672		if (strict)
 673			break;
 674
 675	}
 676
 677	if (locked)
 678		spin_unlock_irqrestore(&cc->zone->lock, flags);
 679
 680	/*
 681	 * Be careful to not go outside of the pageblock.
 682	 */
 683	if (unlikely(blockpfn > end_pfn))
 684		blockpfn = end_pfn;
 685
 686	trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
 687					nr_scanned, total_isolated);
 688
 689	/* Record how far we have got within the block */
 690	*start_pfn = blockpfn;
 691
 692	/*
 693	 * If strict isolation is requested by CMA then check that all the
 694	 * pages requested were isolated. If there were any failures, 0 is
 695	 * returned and CMA will fail.
 696	 */
 697	if (strict && blockpfn < end_pfn)
 698		total_isolated = 0;
 699
 700	cc->total_free_scanned += nr_scanned;
 701	if (total_isolated)
 702		count_compact_events(COMPACTISOLATED, total_isolated);
 703	return total_isolated;
 704}
 705
 706/**
 707 * isolate_freepages_range() - isolate free pages.
 708 * @cc:        Compaction control structure.
 709 * @start_pfn: The first PFN to start isolating.
 710 * @end_pfn:   The one-past-last PFN.
 711 *
 712 * Non-free pages, invalid PFNs, or zone boundaries within the
 713 * [start_pfn, end_pfn) range are considered errors, cause function to
 714 * undo its actions and return zero.
 715 *
 716 * Otherwise, function returns one-past-the-last PFN of isolated page
 717 * (which may be greater then end_pfn if end fell in a middle of
 718 * a free page).
 719 */
 720unsigned long
 721isolate_freepages_range(struct compact_control *cc,
 722			unsigned long start_pfn, unsigned long end_pfn)
 723{
 724	unsigned long isolated, pfn, block_start_pfn, block_end_pfn;
 725	LIST_HEAD(freelist);
 
 
 
 
 726
 727	pfn = start_pfn;
 728	block_start_pfn = pageblock_start_pfn(pfn);
 729	if (block_start_pfn < cc->zone->zone_start_pfn)
 730		block_start_pfn = cc->zone->zone_start_pfn;
 731	block_end_pfn = pageblock_end_pfn(pfn);
 732
 733	for (; pfn < end_pfn; pfn += isolated,
 734				block_start_pfn = block_end_pfn,
 735				block_end_pfn += pageblock_nr_pages) {
 736		/* Protect pfn from changing by isolate_freepages_block */
 737		unsigned long isolate_start_pfn = pfn;
 738
 739		/*
 740		 * pfn could pass the block_end_pfn if isolated freepage
 741		 * is more than pageblock order. In this case, we adjust
 742		 * scanning range to right one.
 743		 */
 744		if (pfn >= block_end_pfn) {
 745			block_start_pfn = pageblock_start_pfn(pfn);
 746			block_end_pfn = pageblock_end_pfn(pfn);
 747		}
 748
 749		block_end_pfn = min(block_end_pfn, end_pfn);
 750
 751		if (!pageblock_pfn_to_page(block_start_pfn,
 752					block_end_pfn, cc->zone))
 753			break;
 754
 755		isolated = isolate_freepages_block(cc, &isolate_start_pfn,
 756					block_end_pfn, &freelist, 0, true);
 757
 758		/*
 759		 * In strict mode, isolate_freepages_block() returns 0 if
 760		 * there are any holes in the block (ie. invalid PFNs or
 761		 * non-free pages).
 762		 */
 763		if (!isolated)
 764			break;
 765
 766		/*
 767		 * If we managed to isolate pages, it is always (1 << n) *
 768		 * pageblock_nr_pages for some non-negative n.  (Max order
 769		 * page may span two pageblocks).
 770		 */
 771	}
 772
 773	/* __isolate_free_page() does not map the pages */
 774	split_map_pages(&freelist);
 775
 776	if (pfn < end_pfn) {
 777		/* Loop terminated early, cleanup. */
 778		release_freepages(&freelist);
 779		return 0;
 780	}
 781
 
 
 
 782	/* We don't use freelists for anything. */
 783	return pfn;
 784}
 785
 786/* Similar to reclaim, but different enough that they don't share logic */
 787static bool too_many_isolated(struct compact_control *cc)
 788{
 789	pg_data_t *pgdat = cc->zone->zone_pgdat;
 790	bool too_many;
 791
 792	unsigned long active, inactive, isolated;
 793
 794	inactive = node_page_state(pgdat, NR_INACTIVE_FILE) +
 795			node_page_state(pgdat, NR_INACTIVE_ANON);
 796	active = node_page_state(pgdat, NR_ACTIVE_FILE) +
 797			node_page_state(pgdat, NR_ACTIVE_ANON);
 798	isolated = node_page_state(pgdat, NR_ISOLATED_FILE) +
 799			node_page_state(pgdat, NR_ISOLATED_ANON);
 800
 801	/*
 802	 * Allow GFP_NOFS to isolate past the limit set for regular
 803	 * compaction runs. This prevents an ABBA deadlock when other
 804	 * compactors have already isolated to the limit, but are
 805	 * blocked on filesystem locks held by the GFP_NOFS thread.
 806	 */
 807	if (cc->gfp_mask & __GFP_FS) {
 808		inactive >>= 3;
 809		active >>= 3;
 810	}
 811
 812	too_many = isolated > (inactive + active) / 2;
 813	if (!too_many)
 814		wake_throttle_isolated(pgdat);
 815
 816	return too_many;
 817}
 818
 819/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 820 * isolate_migratepages_block() - isolate all migrate-able pages within
 821 *				  a single pageblock
 822 * @cc:		Compaction control structure.
 823 * @low_pfn:	The first PFN to isolate
 824 * @end_pfn:	The one-past-the-last PFN to isolate, within same pageblock
 825 * @mode:	Isolation mode to be used.
 826 *
 827 * Isolate all pages that can be migrated from the range specified by
 828 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
 829 * Returns errno, like -EAGAIN or -EINTR in case e.g signal pending or congestion,
 830 * -ENOMEM in case we could not allocate a page, or 0.
 831 * cc->migrate_pfn will contain the next pfn to scan.
 832 *
 833 * The pages are isolated on cc->migratepages list (not required to be empty),
 834 * and cc->nr_migratepages is updated accordingly.
 835 */
 836static int
 837isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
 838			unsigned long end_pfn, isolate_mode_t mode)
 839{
 840	pg_data_t *pgdat = cc->zone->zone_pgdat;
 841	unsigned long nr_scanned = 0, nr_isolated = 0;
 842	struct lruvec *lruvec;
 843	unsigned long flags = 0;
 844	struct lruvec *locked = NULL;
 845	struct folio *folio = NULL;
 846	struct page *page = NULL, *valid_page = NULL;
 847	struct address_space *mapping;
 848	unsigned long start_pfn = low_pfn;
 849	bool skip_on_failure = false;
 850	unsigned long next_skip_pfn = 0;
 851	bool skip_updated = false;
 852	int ret = 0;
 853
 854	cc->migrate_pfn = low_pfn;
 855
 856	/*
 857	 * Ensure that there are not too many pages isolated from the LRU
 858	 * list by either parallel reclaimers or compaction. If there are,
 859	 * delay for some time until fewer pages are isolated
 860	 */
 861	while (unlikely(too_many_isolated(cc))) {
 862		/* stop isolation if there are still pages not migrated */
 863		if (cc->nr_migratepages)
 864			return -EAGAIN;
 865
 866		/* async migration should just abort */
 867		if (cc->mode == MIGRATE_ASYNC)
 868			return -EAGAIN;
 869
 870		reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED);
 871
 872		if (fatal_signal_pending(current))
 873			return -EINTR;
 874	}
 875
 876	cond_resched();
 877
 878	if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
 879		skip_on_failure = true;
 880		next_skip_pfn = block_end_pfn(low_pfn, cc->order);
 881	}
 882
 883	/* Time to isolate some pages for migration */
 884	for (; low_pfn < end_pfn; low_pfn++) {
 885		bool is_dirty, is_unevictable;
 886
 887		if (skip_on_failure && low_pfn >= next_skip_pfn) {
 888			/*
 889			 * We have isolated all migration candidates in the
 890			 * previous order-aligned block, and did not skip it due
 891			 * to failure. We should migrate the pages now and
 892			 * hopefully succeed compaction.
 893			 */
 894			if (nr_isolated)
 895				break;
 896
 897			/*
 898			 * We failed to isolate in the previous order-aligned
 899			 * block. Set the new boundary to the end of the
 900			 * current block. Note we can't simply increase
 901			 * next_skip_pfn by 1 << order, as low_pfn might have
 902			 * been incremented by a higher number due to skipping
 903			 * a compound or a high-order buddy page in the
 904			 * previous loop iteration.
 905			 */
 906			next_skip_pfn = block_end_pfn(low_pfn, cc->order);
 907		}
 908
 909		/*
 910		 * Periodically drop the lock (if held) regardless of its
 911		 * contention, to give chance to IRQs. Abort completely if
 912		 * a fatal signal is pending.
 913		 */
 914		if (!(low_pfn % COMPACT_CLUSTER_MAX)) {
 915			if (locked) {
 916				unlock_page_lruvec_irqrestore(locked, flags);
 917				locked = NULL;
 918			}
 919
 920			if (fatal_signal_pending(current)) {
 921				cc->contended = true;
 922				ret = -EINTR;
 923
 924				goto fatal_pending;
 925			}
 926
 927			cond_resched();
 928		}
 929
 930		nr_scanned++;
 931
 932		page = pfn_to_page(low_pfn);
 933
 934		/*
 935		 * Check if the pageblock has already been marked skipped.
 936		 * Only the first PFN is checked as the caller isolates
 937		 * COMPACT_CLUSTER_MAX at a time so the second call must
 938		 * not falsely conclude that the block should be skipped.
 939		 */
 940		if (!valid_page && (pageblock_aligned(low_pfn) ||
 941				    low_pfn == cc->zone->zone_start_pfn)) {
 942			if (!isolation_suitable(cc, page)) {
 943				low_pfn = end_pfn;
 944				folio = NULL;
 945				goto isolate_abort;
 946			}
 947			valid_page = page;
 948		}
 949
 950		if (PageHuge(page) && cc->alloc_contig) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 951			if (locked) {
 952				unlock_page_lruvec_irqrestore(locked, flags);
 953				locked = NULL;
 954			}
 955
 956			ret = isolate_or_dissolve_huge_page(page, &cc->migratepages);
 957
 958			/*
 959			 * Fail isolation in case isolate_or_dissolve_huge_page()
 960			 * reports an error. In case of -ENOMEM, abort right away.
 961			 */
 962			if (ret < 0) {
 963				 /* Do not report -EBUSY down the chain */
 964				if (ret == -EBUSY)
 965					ret = 0;
 966				low_pfn += compound_nr(page) - 1;
 967				nr_scanned += compound_nr(page) - 1;
 968				goto isolate_fail;
 969			}
 970
 971			if (PageHuge(page)) {
 972				/*
 973				 * Hugepage was successfully isolated and placed
 974				 * on the cc->migratepages list.
 975				 */
 976				folio = page_folio(page);
 977				low_pfn += folio_nr_pages(folio) - 1;
 978				goto isolate_success_no_list;
 979			}
 980
 981			/*
 982			 * Ok, the hugepage was dissolved. Now these pages are
 983			 * Buddy and cannot be re-allocated because they are
 984			 * isolated. Fall-through as the check below handles
 985			 * Buddy pages.
 986			 */
 987		}
 988
 989		/*
 990		 * Skip if free. We read page order here without zone lock
 991		 * which is generally unsafe, but the race window is small and
 992		 * the worst thing that can happen is that we skip some
 993		 * potential isolation targets.
 994		 */
 995		if (PageBuddy(page)) {
 996			unsigned long freepage_order = buddy_order_unsafe(page);
 997
 998			/*
 999			 * Without lock, we cannot be sure that what we got is
1000			 * a valid page order. Consider only values in the
1001			 * valid order range to prevent low_pfn overflow.
1002			 */
1003			if (freepage_order > 0 && freepage_order <= MAX_PAGE_ORDER) {
1004				low_pfn += (1UL << freepage_order) - 1;
1005				nr_scanned += (1UL << freepage_order) - 1;
1006			}
1007			continue;
1008		}
1009
1010		/*
1011		 * Regardless of being on LRU, compound pages such as THP and
1012		 * hugetlbfs are not to be compacted unless we are attempting
1013		 * an allocation much larger than the huge page size (eg CMA).
1014		 * We can potentially save a lot of iterations if we skip them
1015		 * at once. The check is racy, but we can consider only valid
1016		 * values and the only danger is skipping too much.
1017		 */
1018		if (PageCompound(page) && !cc->alloc_contig) {
1019			const unsigned int order = compound_order(page);
1020
1021			if (likely(order <= MAX_PAGE_ORDER)) {
1022				low_pfn += (1UL << order) - 1;
1023				nr_scanned += (1UL << order) - 1;
 
 
 
 
1024			}
1025			goto isolate_fail;
1026		}
1027
1028		/*
1029		 * Check may be lockless but that's ok as we recheck later.
1030		 * It's possible to migrate LRU and non-lru movable pages.
1031		 * Skip any other type of page
1032		 */
1033		if (!PageLRU(page)) {
1034			/*
1035			 * __PageMovable can return false positive so we need
1036			 * to verify it under page_lock.
1037			 */
1038			if (unlikely(__PageMovable(page)) &&
1039					!PageIsolated(page)) {
1040				if (locked) {
1041					unlock_page_lruvec_irqrestore(locked, flags);
1042					locked = NULL;
1043				}
1044
1045				if (isolate_movable_page(page, mode)) {
1046					folio = page_folio(page);
1047					goto isolate_success;
1048				}
1049			}
1050
1051			goto isolate_fail;
1052		}
1053
1054		/*
1055		 * Be careful not to clear PageLRU until after we're
1056		 * sure the page is not being freed elsewhere -- the
1057		 * page release code relies on it.
1058		 */
1059		folio = folio_get_nontail_page(page);
1060		if (unlikely(!folio))
1061			goto isolate_fail;
1062
1063		/*
1064		 * Migration will fail if an anonymous page is pinned in memory,
1065		 * so avoid taking lru_lock and isolating it unnecessarily in an
1066		 * admittedly racy check.
1067		 */
1068		mapping = folio_mapping(folio);
1069		if (!mapping && (folio_ref_count(folio) - 1) > folio_mapcount(folio))
1070			goto isolate_fail_put;
1071
1072		/*
1073		 * Only allow to migrate anonymous pages in GFP_NOFS context
1074		 * because those do not depend on fs locks.
1075		 */
1076		if (!(cc->gfp_mask & __GFP_FS) && mapping)
1077			goto isolate_fail_put;
1078
1079		/* Only take pages on LRU: a check now makes later tests safe */
1080		if (!folio_test_lru(folio))
1081			goto isolate_fail_put;
1082
1083		is_unevictable = folio_test_unevictable(folio);
1084
1085		/* Compaction might skip unevictable pages but CMA takes them */
1086		if (!(mode & ISOLATE_UNEVICTABLE) && is_unevictable)
1087			goto isolate_fail_put;
1088
1089		/*
1090		 * To minimise LRU disruption, the caller can indicate with
1091		 * ISOLATE_ASYNC_MIGRATE that it only wants to isolate pages
1092		 * it will be able to migrate without blocking - clean pages
1093		 * for the most part.  PageWriteback would require blocking.
1094		 */
1095		if ((mode & ISOLATE_ASYNC_MIGRATE) && folio_test_writeback(folio))
1096			goto isolate_fail_put;
1097
1098		is_dirty = folio_test_dirty(folio);
1099
1100		if (((mode & ISOLATE_ASYNC_MIGRATE) && is_dirty) ||
1101		    (mapping && is_unevictable)) {
1102			bool migrate_dirty = true;
1103			bool is_unmovable;
1104
1105			/*
1106			 * Only folios without mappings or that have
1107			 * a ->migrate_folio callback are possible to migrate
1108			 * without blocking.
1109			 *
1110			 * Folios from unmovable mappings are not migratable.
1111			 *
1112			 * However, we can be racing with truncation, which can
1113			 * free the mapping that we need to check. Truncation
1114			 * holds the folio lock until after the folio is removed
1115			 * from the page so holding it ourselves is sufficient.
1116			 *
1117			 * To avoid locking the folio just to check unmovable,
1118			 * assume every unmovable folio is also unevictable,
1119			 * which is a cheaper test.  If our assumption goes
1120			 * wrong, it's not a correctness bug, just potentially
1121			 * wasted cycles.
1122			 */
1123			if (!folio_trylock(folio))
1124				goto isolate_fail_put;
1125
1126			mapping = folio_mapping(folio);
1127			if ((mode & ISOLATE_ASYNC_MIGRATE) && is_dirty) {
1128				migrate_dirty = !mapping ||
1129						mapping->a_ops->migrate_folio;
1130			}
1131			is_unmovable = mapping && mapping_unmovable(mapping);
1132			folio_unlock(folio);
1133			if (!migrate_dirty || is_unmovable)
1134				goto isolate_fail_put;
1135		}
1136
1137		/* Try isolate the folio */
1138		if (!folio_test_clear_lru(folio))
1139			goto isolate_fail_put;
1140
1141		lruvec = folio_lruvec(folio);
1142
1143		/* If we already hold the lock, we can skip some rechecking */
1144		if (lruvec != locked) {
1145			if (locked)
1146				unlock_page_lruvec_irqrestore(locked, flags);
1147
1148			compact_lock_irqsave(&lruvec->lru_lock, &flags, cc);
1149			locked = lruvec;
1150
1151			lruvec_memcg_debug(lruvec, folio);
1152
1153			/*
1154			 * Try get exclusive access under lock. If marked for
1155			 * skip, the scan is aborted unless the current context
1156			 * is a rescan to reach the end of the pageblock.
1157			 */
1158			if (!skip_updated && valid_page) {
1159				skip_updated = true;
1160				if (test_and_set_skip(cc, valid_page) &&
1161				    !cc->finish_pageblock) {
1162					low_pfn = end_pfn;
1163					goto isolate_abort;
1164				}
1165			}
1166
1167			/*
1168			 * folio become large since the non-locked check,
1169			 * and it's on LRU.
1170			 */
1171			if (unlikely(folio_test_large(folio) && !cc->alloc_contig)) {
 
 
1172				low_pfn += folio_nr_pages(folio) - 1;
1173				nr_scanned += folio_nr_pages(folio) - 1;
1174				folio_set_lru(folio);
1175				goto isolate_fail_put;
1176			}
1177		}
1178
1179		/* The folio is taken off the LRU */
1180		if (folio_test_large(folio))
1181			low_pfn += folio_nr_pages(folio) - 1;
1182
1183		/* Successfully isolated */
1184		lruvec_del_folio(lruvec, folio);
1185		node_stat_mod_folio(folio,
1186				NR_ISOLATED_ANON + folio_is_file_lru(folio),
1187				folio_nr_pages(folio));
1188
1189isolate_success:
1190		list_add(&folio->lru, &cc->migratepages);
1191isolate_success_no_list:
1192		cc->nr_migratepages += folio_nr_pages(folio);
1193		nr_isolated += folio_nr_pages(folio);
1194		nr_scanned += folio_nr_pages(folio) - 1;
1195
1196		/*
1197		 * Avoid isolating too much unless this block is being
1198		 * fully scanned (e.g. dirty/writeback pages, parallel allocation)
1199		 * or a lock is contended. For contention, isolate quickly to
1200		 * potentially remove one source of contention.
1201		 */
1202		if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX &&
1203		    !cc->finish_pageblock && !cc->contended) {
1204			++low_pfn;
1205			break;
1206		}
1207
1208		continue;
1209
1210isolate_fail_put:
1211		/* Avoid potential deadlock in freeing page under lru_lock */
1212		if (locked) {
1213			unlock_page_lruvec_irqrestore(locked, flags);
1214			locked = NULL;
1215		}
1216		folio_put(folio);
1217
1218isolate_fail:
1219		if (!skip_on_failure && ret != -ENOMEM)
1220			continue;
1221
1222		/*
1223		 * We have isolated some pages, but then failed. Release them
1224		 * instead of migrating, as we cannot form the cc->order buddy
1225		 * page anyway.
1226		 */
1227		if (nr_isolated) {
1228			if (locked) {
1229				unlock_page_lruvec_irqrestore(locked, flags);
1230				locked = NULL;
1231			}
1232			putback_movable_pages(&cc->migratepages);
1233			cc->nr_migratepages = 0;
1234			nr_isolated = 0;
1235		}
1236
1237		if (low_pfn < next_skip_pfn) {
1238			low_pfn = next_skip_pfn - 1;
1239			/*
1240			 * The check near the loop beginning would have updated
1241			 * next_skip_pfn too, but this is a bit simpler.
1242			 */
1243			next_skip_pfn += 1UL << cc->order;
1244		}
1245
1246		if (ret == -ENOMEM)
1247			break;
1248	}
1249
1250	/*
1251	 * The PageBuddy() check could have potentially brought us outside
1252	 * the range to be scanned.
1253	 */
1254	if (unlikely(low_pfn > end_pfn))
1255		low_pfn = end_pfn;
1256
1257	folio = NULL;
1258
1259isolate_abort:
1260	if (locked)
1261		unlock_page_lruvec_irqrestore(locked, flags);
1262	if (folio) {
1263		folio_set_lru(folio);
1264		folio_put(folio);
1265	}
1266
1267	/*
1268	 * Update the cached scanner pfn once the pageblock has been scanned.
1269	 * Pages will either be migrated in which case there is no point
1270	 * scanning in the near future or migration failed in which case the
1271	 * failure reason may persist. The block is marked for skipping if
1272	 * there were no pages isolated in the block or if the block is
1273	 * rescanned twice in a row.
1274	 */
1275	if (low_pfn == end_pfn && (!nr_isolated || cc->finish_pageblock)) {
1276		if (!cc->no_set_skip_hint && valid_page && !skip_updated)
1277			set_pageblock_skip(valid_page);
1278		update_cached_migrate(cc, low_pfn);
1279	}
1280
1281	trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
1282						nr_scanned, nr_isolated);
1283
1284fatal_pending:
1285	cc->total_migrate_scanned += nr_scanned;
1286	if (nr_isolated)
1287		count_compact_events(COMPACTISOLATED, nr_isolated);
1288
1289	cc->migrate_pfn = low_pfn;
1290
1291	return ret;
1292}
1293
1294/**
1295 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
1296 * @cc:        Compaction control structure.
1297 * @start_pfn: The first PFN to start isolating.
1298 * @end_pfn:   The one-past-last PFN.
1299 *
1300 * Returns -EAGAIN when contented, -EINTR in case of a signal pending, -ENOMEM
1301 * in case we could not allocate a page, or 0.
1302 */
1303int
1304isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
1305							unsigned long end_pfn)
1306{
1307	unsigned long pfn, block_start_pfn, block_end_pfn;
1308	int ret = 0;
1309
1310	/* Scan block by block. First and last block may be incomplete */
1311	pfn = start_pfn;
1312	block_start_pfn = pageblock_start_pfn(pfn);
1313	if (block_start_pfn < cc->zone->zone_start_pfn)
1314		block_start_pfn = cc->zone->zone_start_pfn;
1315	block_end_pfn = pageblock_end_pfn(pfn);
1316
1317	for (; pfn < end_pfn; pfn = block_end_pfn,
1318				block_start_pfn = block_end_pfn,
1319				block_end_pfn += pageblock_nr_pages) {
1320
1321		block_end_pfn = min(block_end_pfn, end_pfn);
1322
1323		if (!pageblock_pfn_to_page(block_start_pfn,
1324					block_end_pfn, cc->zone))
1325			continue;
1326
1327		ret = isolate_migratepages_block(cc, pfn, block_end_pfn,
1328						 ISOLATE_UNEVICTABLE);
1329
1330		if (ret)
1331			break;
1332
1333		if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX)
1334			break;
1335	}
1336
1337	return ret;
1338}
1339
1340#endif /* CONFIG_COMPACTION || CONFIG_CMA */
1341#ifdef CONFIG_COMPACTION
1342
1343static bool suitable_migration_source(struct compact_control *cc,
1344							struct page *page)
1345{
1346	int block_mt;
1347
1348	if (pageblock_skip_persistent(page))
1349		return false;
1350
1351	if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction)
1352		return true;
1353
1354	block_mt = get_pageblock_migratetype(page);
1355
1356	if (cc->migratetype == MIGRATE_MOVABLE)
1357		return is_migrate_movable(block_mt);
1358	else
1359		return block_mt == cc->migratetype;
1360}
1361
1362/* Returns true if the page is within a block suitable for migration to */
1363static bool suitable_migration_target(struct compact_control *cc,
1364							struct page *page)
1365{
1366	/* If the page is a large free page, then disallow migration */
1367	if (PageBuddy(page)) {
 
 
1368		/*
1369		 * We are checking page_order without zone->lock taken. But
1370		 * the only small danger is that we skip a potentially suitable
1371		 * pageblock, so it's not worth to check order for valid range.
1372		 */
1373		if (buddy_order_unsafe(page) >= pageblock_order)
1374			return false;
1375	}
1376
1377	if (cc->ignore_block_suitable)
1378		return true;
1379
1380	/* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1381	if (is_migrate_movable(get_pageblock_migratetype(page)))
1382		return true;
1383
1384	/* Otherwise skip the block */
1385	return false;
1386}
1387
1388static inline unsigned int
1389freelist_scan_limit(struct compact_control *cc)
1390{
1391	unsigned short shift = BITS_PER_LONG - 1;
1392
1393	return (COMPACT_CLUSTER_MAX >> min(shift, cc->fast_search_fail)) + 1;
1394}
1395
1396/*
1397 * Test whether the free scanner has reached the same or lower pageblock than
1398 * the migration scanner, and compaction should thus terminate.
1399 */
1400static inline bool compact_scanners_met(struct compact_control *cc)
1401{
1402	return (cc->free_pfn >> pageblock_order)
1403		<= (cc->migrate_pfn >> pageblock_order);
1404}
1405
1406/*
1407 * Used when scanning for a suitable migration target which scans freelists
1408 * in reverse. Reorders the list such as the unscanned pages are scanned
1409 * first on the next iteration of the free scanner
1410 */
1411static void
1412move_freelist_head(struct list_head *freelist, struct page *freepage)
1413{
1414	LIST_HEAD(sublist);
1415
1416	if (!list_is_first(&freepage->buddy_list, freelist)) {
1417		list_cut_before(&sublist, freelist, &freepage->buddy_list);
1418		list_splice_tail(&sublist, freelist);
1419	}
1420}
1421
1422/*
1423 * Similar to move_freelist_head except used by the migration scanner
1424 * when scanning forward. It's possible for these list operations to
1425 * move against each other if they search the free list exactly in
1426 * lockstep.
1427 */
1428static void
1429move_freelist_tail(struct list_head *freelist, struct page *freepage)
1430{
1431	LIST_HEAD(sublist);
1432
1433	if (!list_is_last(&freepage->buddy_list, freelist)) {
1434		list_cut_position(&sublist, freelist, &freepage->buddy_list);
1435		list_splice_tail(&sublist, freelist);
1436	}
1437}
1438
1439static void
1440fast_isolate_around(struct compact_control *cc, unsigned long pfn)
1441{
1442	unsigned long start_pfn, end_pfn;
1443	struct page *page;
1444
1445	/* Do not search around if there are enough pages already */
1446	if (cc->nr_freepages >= cc->nr_migratepages)
1447		return;
1448
1449	/* Minimise scanning during async compaction */
1450	if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC)
1451		return;
1452
1453	/* Pageblock boundaries */
1454	start_pfn = max(pageblock_start_pfn(pfn), cc->zone->zone_start_pfn);
1455	end_pfn = min(pageblock_end_pfn(pfn), zone_end_pfn(cc->zone));
1456
1457	page = pageblock_pfn_to_page(start_pfn, end_pfn, cc->zone);
1458	if (!page)
1459		return;
1460
1461	isolate_freepages_block(cc, &start_pfn, end_pfn, &cc->freepages, 1, false);
1462
1463	/* Skip this pageblock in the future as it's full or nearly full */
1464	if (start_pfn == end_pfn && !cc->no_set_skip_hint)
1465		set_pageblock_skip(page);
1466}
1467
1468/* Search orders in round-robin fashion */
1469static int next_search_order(struct compact_control *cc, int order)
1470{
1471	order--;
1472	if (order < 0)
1473		order = cc->order - 1;
1474
1475	/* Search wrapped around? */
1476	if (order == cc->search_order) {
1477		cc->search_order--;
1478		if (cc->search_order < 0)
1479			cc->search_order = cc->order - 1;
1480		return -1;
1481	}
1482
1483	return order;
1484}
1485
1486static void fast_isolate_freepages(struct compact_control *cc)
1487{
1488	unsigned int limit = max(1U, freelist_scan_limit(cc) >> 1);
1489	unsigned int nr_scanned = 0, total_isolated = 0;
1490	unsigned long low_pfn, min_pfn, highest = 0;
1491	unsigned long nr_isolated = 0;
1492	unsigned long distance;
1493	struct page *page = NULL;
1494	bool scan_start = false;
1495	int order;
1496
1497	/* Full compaction passes in a negative order */
1498	if (cc->order <= 0)
1499		return;
1500
1501	/*
1502	 * If starting the scan, use a deeper search and use the highest
1503	 * PFN found if a suitable one is not found.
1504	 */
1505	if (cc->free_pfn >= cc->zone->compact_init_free_pfn) {
1506		limit = pageblock_nr_pages >> 1;
1507		scan_start = true;
1508	}
1509
1510	/*
1511	 * Preferred point is in the top quarter of the scan space but take
1512	 * a pfn from the top half if the search is problematic.
1513	 */
1514	distance = (cc->free_pfn - cc->migrate_pfn);
1515	low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2));
1516	min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1));
1517
1518	if (WARN_ON_ONCE(min_pfn > low_pfn))
1519		low_pfn = min_pfn;
1520
1521	/*
1522	 * Search starts from the last successful isolation order or the next
1523	 * order to search after a previous failure
1524	 */
1525	cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order);
1526
1527	for (order = cc->search_order;
1528	     !page && order >= 0;
1529	     order = next_search_order(cc, order)) {
1530		struct free_area *area = &cc->zone->free_area[order];
1531		struct list_head *freelist;
1532		struct page *freepage;
1533		unsigned long flags;
1534		unsigned int order_scanned = 0;
1535		unsigned long high_pfn = 0;
1536
1537		if (!area->nr_free)
1538			continue;
1539
1540		spin_lock_irqsave(&cc->zone->lock, flags);
1541		freelist = &area->free_list[MIGRATE_MOVABLE];
1542		list_for_each_entry_reverse(freepage, freelist, buddy_list) {
1543			unsigned long pfn;
1544
1545			order_scanned++;
1546			nr_scanned++;
1547			pfn = page_to_pfn(freepage);
1548
1549			if (pfn >= highest)
1550				highest = max(pageblock_start_pfn(pfn),
1551					      cc->zone->zone_start_pfn);
1552
1553			if (pfn >= low_pfn) {
1554				cc->fast_search_fail = 0;
1555				cc->search_order = order;
1556				page = freepage;
1557				break;
1558			}
1559
1560			if (pfn >= min_pfn && pfn > high_pfn) {
1561				high_pfn = pfn;
1562
1563				/* Shorten the scan if a candidate is found */
1564				limit >>= 1;
1565			}
1566
1567			if (order_scanned >= limit)
1568				break;
1569		}
1570
1571		/* Use a maximum candidate pfn if a preferred one was not found */
1572		if (!page && high_pfn) {
1573			page = pfn_to_page(high_pfn);
1574
1575			/* Update freepage for the list reorder below */
1576			freepage = page;
1577		}
1578
1579		/* Reorder to so a future search skips recent pages */
1580		move_freelist_head(freelist, freepage);
1581
1582		/* Isolate the page if available */
1583		if (page) {
1584			if (__isolate_free_page(page, order)) {
1585				set_page_private(page, order);
1586				nr_isolated = 1 << order;
1587				nr_scanned += nr_isolated - 1;
1588				total_isolated += nr_isolated;
1589				cc->nr_freepages += nr_isolated;
1590				list_add_tail(&page->lru, &cc->freepages);
1591				count_compact_events(COMPACTISOLATED, nr_isolated);
1592			} else {
1593				/* If isolation fails, abort the search */
1594				order = cc->search_order + 1;
1595				page = NULL;
1596			}
1597		}
1598
1599		spin_unlock_irqrestore(&cc->zone->lock, flags);
1600
1601		/* Skip fast search if enough freepages isolated */
1602		if (cc->nr_freepages >= cc->nr_migratepages)
1603			break;
1604
1605		/*
1606		 * Smaller scan on next order so the total scan is related
1607		 * to freelist_scan_limit.
1608		 */
1609		if (order_scanned >= limit)
1610			limit = max(1U, limit >> 1);
1611	}
1612
1613	trace_mm_compaction_fast_isolate_freepages(min_pfn, cc->free_pfn,
1614						   nr_scanned, total_isolated);
1615
1616	if (!page) {
1617		cc->fast_search_fail++;
1618		if (scan_start) {
1619			/*
1620			 * Use the highest PFN found above min. If one was
1621			 * not found, be pessimistic for direct compaction
1622			 * and use the min mark.
1623			 */
1624			if (highest >= min_pfn) {
1625				page = pfn_to_page(highest);
1626				cc->free_pfn = highest;
1627			} else {
1628				if (cc->direct_compaction && pfn_valid(min_pfn)) {
1629					page = pageblock_pfn_to_page(min_pfn,
1630						min(pageblock_end_pfn(min_pfn),
1631						    zone_end_pfn(cc->zone)),
1632						cc->zone);
1633					if (page && !suitable_migration_target(cc, page))
1634						page = NULL;
1635
1636					cc->free_pfn = min_pfn;
1637				}
1638			}
1639		}
1640	}
1641
1642	if (highest && highest >= cc->zone->compact_cached_free_pfn) {
1643		highest -= pageblock_nr_pages;
1644		cc->zone->compact_cached_free_pfn = highest;
1645	}
1646
1647	cc->total_free_scanned += nr_scanned;
1648	if (!page)
1649		return;
1650
1651	low_pfn = page_to_pfn(page);
1652	fast_isolate_around(cc, low_pfn);
1653}
1654
1655/*
1656 * Based on information in the current compact_control, find blocks
1657 * suitable for isolating free pages from and then isolate them.
1658 */
1659static void isolate_freepages(struct compact_control *cc)
1660{
1661	struct zone *zone = cc->zone;
1662	struct page *page;
1663	unsigned long block_start_pfn;	/* start of current pageblock */
1664	unsigned long isolate_start_pfn; /* exact pfn we start at */
1665	unsigned long block_end_pfn;	/* end of current pageblock */
1666	unsigned long low_pfn;	     /* lowest pfn scanner is able to scan */
1667	struct list_head *freelist = &cc->freepages;
1668	unsigned int stride;
1669
1670	/* Try a small search of the free lists for a candidate */
1671	fast_isolate_freepages(cc);
1672	if (cc->nr_freepages)
1673		goto splitmap;
1674
1675	/*
1676	 * Initialise the free scanner. The starting point is where we last
1677	 * successfully isolated from, zone-cached value, or the end of the
1678	 * zone when isolating for the first time. For looping we also need
1679	 * this pfn aligned down to the pageblock boundary, because we do
1680	 * block_start_pfn -= pageblock_nr_pages in the for loop.
1681	 * For ending point, take care when isolating in last pageblock of a
1682	 * zone which ends in the middle of a pageblock.
1683	 * The low boundary is the end of the pageblock the migration scanner
1684	 * is using.
1685	 */
1686	isolate_start_pfn = cc->free_pfn;
1687	block_start_pfn = pageblock_start_pfn(isolate_start_pfn);
1688	block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1689						zone_end_pfn(zone));
1690	low_pfn = pageblock_end_pfn(cc->migrate_pfn);
1691	stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1;
1692
1693	/*
1694	 * Isolate free pages until enough are available to migrate the
1695	 * pages on cc->migratepages. We stop searching if the migrate
1696	 * and free page scanners meet or enough free pages are isolated.
1697	 */
1698	for (; block_start_pfn >= low_pfn;
1699				block_end_pfn = block_start_pfn,
1700				block_start_pfn -= pageblock_nr_pages,
1701				isolate_start_pfn = block_start_pfn) {
1702		unsigned long nr_isolated;
1703
1704		/*
1705		 * This can iterate a massively long zone without finding any
1706		 * suitable migration targets, so periodically check resched.
1707		 */
1708		if (!(block_start_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages)))
1709			cond_resched();
1710
1711		page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1712									zone);
1713		if (!page) {
1714			unsigned long next_pfn;
1715
1716			next_pfn = skip_offline_sections_reverse(block_start_pfn);
1717			if (next_pfn)
1718				block_start_pfn = max(next_pfn, low_pfn);
1719
1720			continue;
1721		}
1722
1723		/* Check the block is suitable for migration */
1724		if (!suitable_migration_target(cc, page))
1725			continue;
1726
1727		/* If isolation recently failed, do not retry */
1728		if (!isolation_suitable(cc, page))
1729			continue;
1730
1731		/* Found a block suitable for isolating free pages from. */
1732		nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn,
1733					block_end_pfn, freelist, stride, false);
1734
1735		/* Update the skip hint if the full pageblock was scanned */
1736		if (isolate_start_pfn == block_end_pfn)
1737			update_pageblock_skip(cc, page, block_start_pfn -
1738					      pageblock_nr_pages);
1739
1740		/* Are enough freepages isolated? */
1741		if (cc->nr_freepages >= cc->nr_migratepages) {
1742			if (isolate_start_pfn >= block_end_pfn) {
1743				/*
1744				 * Restart at previous pageblock if more
1745				 * freepages can be isolated next time.
1746				 */
1747				isolate_start_pfn =
1748					block_start_pfn - pageblock_nr_pages;
1749			}
1750			break;
1751		} else if (isolate_start_pfn < block_end_pfn) {
1752			/*
1753			 * If isolation failed early, do not continue
1754			 * needlessly.
1755			 */
1756			break;
1757		}
1758
1759		/* Adjust stride depending on isolation */
1760		if (nr_isolated) {
1761			stride = 1;
1762			continue;
1763		}
1764		stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1);
1765	}
1766
1767	/*
1768	 * Record where the free scanner will restart next time. Either we
1769	 * broke from the loop and set isolate_start_pfn based on the last
1770	 * call to isolate_freepages_block(), or we met the migration scanner
1771	 * and the loop terminated due to isolate_start_pfn < low_pfn
1772	 */
1773	cc->free_pfn = isolate_start_pfn;
1774
1775splitmap:
1776	/* __isolate_free_page() does not map the pages */
1777	split_map_pages(freelist);
1778}
1779
1780/*
1781 * This is a migrate-callback that "allocates" freepages by taking pages
1782 * from the isolated freelists in the block we are migrating to.
1783 */
1784static struct folio *compaction_alloc(struct folio *src, unsigned long data)
1785{
1786	struct compact_control *cc = (struct compact_control *)data;
1787	struct folio *dst;
 
 
 
 
 
 
 
 
 
 
1788
1789	if (list_empty(&cc->freepages)) {
1790		isolate_freepages(cc);
1791
1792		if (list_empty(&cc->freepages))
1793			return NULL;
 
 
 
1794	}
1795
1796	dst = list_entry(cc->freepages.next, struct folio, lru);
1797	list_del(&dst->lru);
1798	cc->nr_freepages--;
 
 
 
 
 
 
 
 
 
 
 
1799
1800	return dst;
 
 
 
 
 
1801}
1802
1803/*
1804 * This is a migrate-callback that "frees" freepages back to the isolated
1805 * freelist.  All pages on the freelist are from the same zone, so there is no
1806 * special handling needed for NUMA.
1807 */
1808static void compaction_free(struct folio *dst, unsigned long data)
1809{
1810	struct compact_control *cc = (struct compact_control *)data;
 
 
1811
1812	list_add(&dst->lru, &cc->freepages);
1813	cc->nr_freepages++;
 
 
 
 
 
 
 
 
1814}
1815
1816/* possible outcome of isolate_migratepages */
1817typedef enum {
1818	ISOLATE_ABORT,		/* Abort compaction now */
1819	ISOLATE_NONE,		/* No pages isolated, continue scanning */
1820	ISOLATE_SUCCESS,	/* Pages isolated, migrate */
1821} isolate_migrate_t;
1822
1823/*
1824 * Allow userspace to control policy on scanning the unevictable LRU for
1825 * compactable pages.
1826 */
1827static int sysctl_compact_unevictable_allowed __read_mostly = CONFIG_COMPACT_UNEVICTABLE_DEFAULT;
1828/*
1829 * Tunable for proactive compaction. It determines how
1830 * aggressively the kernel should compact memory in the
1831 * background. It takes values in the range [0, 100].
1832 */
1833static unsigned int __read_mostly sysctl_compaction_proactiveness = 20;
1834static int sysctl_extfrag_threshold = 500;
1835static int __read_mostly sysctl_compact_memory;
1836
1837static inline void
1838update_fast_start_pfn(struct compact_control *cc, unsigned long pfn)
1839{
1840	if (cc->fast_start_pfn == ULONG_MAX)
1841		return;
1842
1843	if (!cc->fast_start_pfn)
1844		cc->fast_start_pfn = pfn;
1845
1846	cc->fast_start_pfn = min(cc->fast_start_pfn, pfn);
1847}
1848
1849static inline unsigned long
1850reinit_migrate_pfn(struct compact_control *cc)
1851{
1852	if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX)
1853		return cc->migrate_pfn;
1854
1855	cc->migrate_pfn = cc->fast_start_pfn;
1856	cc->fast_start_pfn = ULONG_MAX;
1857
1858	return cc->migrate_pfn;
1859}
1860
1861/*
1862 * Briefly search the free lists for a migration source that already has
1863 * some free pages to reduce the number of pages that need migration
1864 * before a pageblock is free.
1865 */
1866static unsigned long fast_find_migrateblock(struct compact_control *cc)
1867{
1868	unsigned int limit = freelist_scan_limit(cc);
1869	unsigned int nr_scanned = 0;
1870	unsigned long distance;
1871	unsigned long pfn = cc->migrate_pfn;
1872	unsigned long high_pfn;
1873	int order;
1874	bool found_block = false;
1875
1876	/* Skip hints are relied on to avoid repeats on the fast search */
1877	if (cc->ignore_skip_hint)
1878		return pfn;
1879
1880	/*
1881	 * If the pageblock should be finished then do not select a different
1882	 * pageblock.
1883	 */
1884	if (cc->finish_pageblock)
1885		return pfn;
1886
1887	/*
1888	 * If the migrate_pfn is not at the start of a zone or the start
1889	 * of a pageblock then assume this is a continuation of a previous
1890	 * scan restarted due to COMPACT_CLUSTER_MAX.
1891	 */
1892	if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn))
1893		return pfn;
1894
1895	/*
1896	 * For smaller orders, just linearly scan as the number of pages
1897	 * to migrate should be relatively small and does not necessarily
1898	 * justify freeing up a large block for a small allocation.
1899	 */
1900	if (cc->order <= PAGE_ALLOC_COSTLY_ORDER)
1901		return pfn;
1902
1903	/*
1904	 * Only allow kcompactd and direct requests for movable pages to
1905	 * quickly clear out a MOVABLE pageblock for allocation. This
1906	 * reduces the risk that a large movable pageblock is freed for
1907	 * an unmovable/reclaimable small allocation.
1908	 */
1909	if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE)
1910		return pfn;
1911
1912	/*
1913	 * When starting the migration scanner, pick any pageblock within the
1914	 * first half of the search space. Otherwise try and pick a pageblock
1915	 * within the first eighth to reduce the chances that a migration
1916	 * target later becomes a source.
1917	 */
1918	distance = (cc->free_pfn - cc->migrate_pfn) >> 1;
1919	if (cc->migrate_pfn != cc->zone->zone_start_pfn)
1920		distance >>= 2;
1921	high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance);
1922
1923	for (order = cc->order - 1;
1924	     order >= PAGE_ALLOC_COSTLY_ORDER && !found_block && nr_scanned < limit;
1925	     order--) {
1926		struct free_area *area = &cc->zone->free_area[order];
1927		struct list_head *freelist;
1928		unsigned long flags;
1929		struct page *freepage;
1930
1931		if (!area->nr_free)
1932			continue;
1933
1934		spin_lock_irqsave(&cc->zone->lock, flags);
1935		freelist = &area->free_list[MIGRATE_MOVABLE];
1936		list_for_each_entry(freepage, freelist, buddy_list) {
1937			unsigned long free_pfn;
1938
1939			if (nr_scanned++ >= limit) {
1940				move_freelist_tail(freelist, freepage);
1941				break;
1942			}
1943
1944			free_pfn = page_to_pfn(freepage);
1945			if (free_pfn < high_pfn) {
1946				/*
1947				 * Avoid if skipped recently. Ideally it would
1948				 * move to the tail but even safe iteration of
1949				 * the list assumes an entry is deleted, not
1950				 * reordered.
1951				 */
1952				if (get_pageblock_skip(freepage))
1953					continue;
1954
1955				/* Reorder to so a future search skips recent pages */
1956				move_freelist_tail(freelist, freepage);
1957
1958				update_fast_start_pfn(cc, free_pfn);
1959				pfn = pageblock_start_pfn(free_pfn);
1960				if (pfn < cc->zone->zone_start_pfn)
1961					pfn = cc->zone->zone_start_pfn;
1962				cc->fast_search_fail = 0;
1963				found_block = true;
1964				break;
1965			}
1966		}
1967		spin_unlock_irqrestore(&cc->zone->lock, flags);
1968	}
1969
1970	cc->total_migrate_scanned += nr_scanned;
1971
1972	/*
1973	 * If fast scanning failed then use a cached entry for a page block
1974	 * that had free pages as the basis for starting a linear scan.
1975	 */
1976	if (!found_block) {
1977		cc->fast_search_fail++;
1978		pfn = reinit_migrate_pfn(cc);
1979	}
1980	return pfn;
1981}
1982
1983/*
1984 * Isolate all pages that can be migrated from the first suitable block,
1985 * starting at the block pointed to by the migrate scanner pfn within
1986 * compact_control.
1987 */
1988static isolate_migrate_t isolate_migratepages(struct compact_control *cc)
1989{
1990	unsigned long block_start_pfn;
1991	unsigned long block_end_pfn;
1992	unsigned long low_pfn;
1993	struct page *page;
1994	const isolate_mode_t isolate_mode =
1995		(sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
1996		(cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
1997	bool fast_find_block;
1998
1999	/*
2000	 * Start at where we last stopped, or beginning of the zone as
2001	 * initialized by compact_zone(). The first failure will use
2002	 * the lowest PFN as the starting point for linear scanning.
2003	 */
2004	low_pfn = fast_find_migrateblock(cc);
2005	block_start_pfn = pageblock_start_pfn(low_pfn);
2006	if (block_start_pfn < cc->zone->zone_start_pfn)
2007		block_start_pfn = cc->zone->zone_start_pfn;
2008
2009	/*
2010	 * fast_find_migrateblock() has already ensured the pageblock is not
2011	 * set with a skipped flag, so to avoid the isolation_suitable check
2012	 * below again, check whether the fast search was successful.
2013	 */
2014	fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail;
2015
2016	/* Only scan within a pageblock boundary */
2017	block_end_pfn = pageblock_end_pfn(low_pfn);
2018
2019	/*
2020	 * Iterate over whole pageblocks until we find the first suitable.
2021	 * Do not cross the free scanner.
2022	 */
2023	for (; block_end_pfn <= cc->free_pfn;
2024			fast_find_block = false,
2025			cc->migrate_pfn = low_pfn = block_end_pfn,
2026			block_start_pfn = block_end_pfn,
2027			block_end_pfn += pageblock_nr_pages) {
2028
2029		/*
2030		 * This can potentially iterate a massively long zone with
2031		 * many pageblocks unsuitable, so periodically check if we
2032		 * need to schedule.
2033		 */
2034		if (!(low_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages)))
2035			cond_resched();
2036
2037		page = pageblock_pfn_to_page(block_start_pfn,
2038						block_end_pfn, cc->zone);
2039		if (!page) {
2040			unsigned long next_pfn;
2041
2042			next_pfn = skip_offline_sections(block_start_pfn);
2043			if (next_pfn)
2044				block_end_pfn = min(next_pfn, cc->free_pfn);
2045			continue;
2046		}
2047
2048		/*
2049		 * If isolation recently failed, do not retry. Only check the
2050		 * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock
2051		 * to be visited multiple times. Assume skip was checked
2052		 * before making it "skip" so other compaction instances do
2053		 * not scan the same block.
2054		 */
2055		if ((pageblock_aligned(low_pfn) ||
2056		     low_pfn == cc->zone->zone_start_pfn) &&
2057		    !fast_find_block && !isolation_suitable(cc, page))
2058			continue;
2059
2060		/*
2061		 * For async direct compaction, only scan the pageblocks of the
2062		 * same migratetype without huge pages. Async direct compaction
2063		 * is optimistic to see if the minimum amount of work satisfies
2064		 * the allocation. The cached PFN is updated as it's possible
2065		 * that all remaining blocks between source and target are
2066		 * unsuitable and the compaction scanners fail to meet.
2067		 */
2068		if (!suitable_migration_source(cc, page)) {
2069			update_cached_migrate(cc, block_end_pfn);
2070			continue;
2071		}
2072
2073		/* Perform the isolation */
2074		if (isolate_migratepages_block(cc, low_pfn, block_end_pfn,
2075						isolate_mode))
2076			return ISOLATE_ABORT;
2077
2078		/*
2079		 * Either we isolated something and proceed with migration. Or
2080		 * we failed and compact_zone should decide if we should
2081		 * continue or not.
2082		 */
2083		break;
2084	}
2085
2086	return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
2087}
2088
2089/*
2090 * order == -1 is expected when compacting proactively via
2091 * 1. /proc/sys/vm/compact_memory
2092 * 2. /sys/devices/system/node/nodex/compact
2093 * 3. /proc/sys/vm/compaction_proactiveness
2094 */
2095static inline bool is_via_compact_memory(int order)
2096{
2097	return order == -1;
2098}
2099
2100/*
2101 * Determine whether kswapd is (or recently was!) running on this node.
2102 *
2103 * pgdat_kswapd_lock() pins pgdat->kswapd, so a concurrent kswapd_stop() can't
2104 * zero it.
2105 */
2106static bool kswapd_is_running(pg_data_t *pgdat)
2107{
2108	bool running;
2109
2110	pgdat_kswapd_lock(pgdat);
2111	running = pgdat->kswapd && task_is_running(pgdat->kswapd);
2112	pgdat_kswapd_unlock(pgdat);
2113
2114	return running;
2115}
2116
2117/*
2118 * A zone's fragmentation score is the external fragmentation wrt to the
2119 * COMPACTION_HPAGE_ORDER. It returns a value in the range [0, 100].
2120 */
2121static unsigned int fragmentation_score_zone(struct zone *zone)
2122{
2123	return extfrag_for_order(zone, COMPACTION_HPAGE_ORDER);
2124}
2125
2126/*
2127 * A weighted zone's fragmentation score is the external fragmentation
2128 * wrt to the COMPACTION_HPAGE_ORDER scaled by the zone's size. It
2129 * returns a value in the range [0, 100].
2130 *
2131 * The scaling factor ensures that proactive compaction focuses on larger
2132 * zones like ZONE_NORMAL, rather than smaller, specialized zones like
2133 * ZONE_DMA32. For smaller zones, the score value remains close to zero,
2134 * and thus never exceeds the high threshold for proactive compaction.
2135 */
2136static unsigned int fragmentation_score_zone_weighted(struct zone *zone)
2137{
2138	unsigned long score;
2139
2140	score = zone->present_pages * fragmentation_score_zone(zone);
2141	return div64_ul(score, zone->zone_pgdat->node_present_pages + 1);
2142}
2143
2144/*
2145 * The per-node proactive (background) compaction process is started by its
2146 * corresponding kcompactd thread when the node's fragmentation score
2147 * exceeds the high threshold. The compaction process remains active till
2148 * the node's score falls below the low threshold, or one of the back-off
2149 * conditions is met.
2150 */
2151static unsigned int fragmentation_score_node(pg_data_t *pgdat)
2152{
2153	unsigned int score = 0;
2154	int zoneid;
2155
2156	for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2157		struct zone *zone;
2158
2159		zone = &pgdat->node_zones[zoneid];
2160		if (!populated_zone(zone))
2161			continue;
2162		score += fragmentation_score_zone_weighted(zone);
2163	}
2164
2165	return score;
2166}
2167
2168static unsigned int fragmentation_score_wmark(bool low)
2169{
2170	unsigned int wmark_low;
2171
2172	/*
2173	 * Cap the low watermark to avoid excessive compaction
2174	 * activity in case a user sets the proactiveness tunable
2175	 * close to 100 (maximum).
2176	 */
2177	wmark_low = max(100U - sysctl_compaction_proactiveness, 5U);
2178	return low ? wmark_low : min(wmark_low + 10, 100U);
2179}
2180
2181static bool should_proactive_compact_node(pg_data_t *pgdat)
2182{
2183	int wmark_high;
2184
2185	if (!sysctl_compaction_proactiveness || kswapd_is_running(pgdat))
2186		return false;
2187
2188	wmark_high = fragmentation_score_wmark(false);
2189	return fragmentation_score_node(pgdat) > wmark_high;
2190}
2191
2192static enum compact_result __compact_finished(struct compact_control *cc)
2193{
2194	unsigned int order;
2195	const int migratetype = cc->migratetype;
2196	int ret;
2197
2198	/* Compaction run completes if the migrate and free scanner meet */
2199	if (compact_scanners_met(cc)) {
2200		/* Let the next compaction start anew. */
2201		reset_cached_positions(cc->zone);
2202
2203		/*
2204		 * Mark that the PG_migrate_skip information should be cleared
2205		 * by kswapd when it goes to sleep. kcompactd does not set the
2206		 * flag itself as the decision to be clear should be directly
2207		 * based on an allocation request.
2208		 */
2209		if (cc->direct_compaction)
2210			cc->zone->compact_blockskip_flush = true;
2211
2212		if (cc->whole_zone)
2213			return COMPACT_COMPLETE;
2214		else
2215			return COMPACT_PARTIAL_SKIPPED;
2216	}
2217
2218	if (cc->proactive_compaction) {
2219		int score, wmark_low;
2220		pg_data_t *pgdat;
2221
2222		pgdat = cc->zone->zone_pgdat;
2223		if (kswapd_is_running(pgdat))
2224			return COMPACT_PARTIAL_SKIPPED;
2225
2226		score = fragmentation_score_zone(cc->zone);
2227		wmark_low = fragmentation_score_wmark(true);
2228
2229		if (score > wmark_low)
2230			ret = COMPACT_CONTINUE;
2231		else
2232			ret = COMPACT_SUCCESS;
2233
2234		goto out;
2235	}
2236
2237	if (is_via_compact_memory(cc->order))
2238		return COMPACT_CONTINUE;
2239
2240	/*
2241	 * Always finish scanning a pageblock to reduce the possibility of
2242	 * fallbacks in the future. This is particularly important when
2243	 * migration source is unmovable/reclaimable but it's not worth
2244	 * special casing.
2245	 */
2246	if (!pageblock_aligned(cc->migrate_pfn))
2247		return COMPACT_CONTINUE;
2248
2249	/* Direct compactor: Is a suitable page free? */
2250	ret = COMPACT_NO_SUITABLE_PAGE;
2251	for (order = cc->order; order < NR_PAGE_ORDERS; order++) {
2252		struct free_area *area = &cc->zone->free_area[order];
2253		bool can_steal;
2254
2255		/* Job done if page is free of the right migratetype */
2256		if (!free_area_empty(area, migratetype))
2257			return COMPACT_SUCCESS;
2258
2259#ifdef CONFIG_CMA
2260		/* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
2261		if (migratetype == MIGRATE_MOVABLE &&
2262			!free_area_empty(area, MIGRATE_CMA))
2263			return COMPACT_SUCCESS;
2264#endif
2265		/*
2266		 * Job done if allocation would steal freepages from
2267		 * other migratetype buddy lists.
2268		 */
2269		if (find_suitable_fallback(area, order, migratetype,
2270						true, &can_steal) != -1)
2271			/*
2272			 * Movable pages are OK in any pageblock. If we are
2273			 * stealing for a non-movable allocation, make sure
2274			 * we finish compacting the current pageblock first
2275			 * (which is assured by the above migrate_pfn align
2276			 * check) so it is as free as possible and we won't
2277			 * have to steal another one soon.
2278			 */
2279			return COMPACT_SUCCESS;
2280	}
2281
2282out:
2283	if (cc->contended || fatal_signal_pending(current))
2284		ret = COMPACT_CONTENDED;
2285
2286	return ret;
2287}
2288
2289static enum compact_result compact_finished(struct compact_control *cc)
2290{
2291	int ret;
2292
2293	ret = __compact_finished(cc);
2294	trace_mm_compaction_finished(cc->zone, cc->order, ret);
2295	if (ret == COMPACT_NO_SUITABLE_PAGE)
2296		ret = COMPACT_CONTINUE;
2297
2298	return ret;
2299}
2300
2301static bool __compaction_suitable(struct zone *zone, int order,
2302				  int highest_zoneidx,
2303				  unsigned long wmark_target)
2304{
2305	unsigned long watermark;
2306	/*
2307	 * Watermarks for order-0 must be met for compaction to be able to
2308	 * isolate free pages for migration targets. This means that the
2309	 * watermark and alloc_flags have to match, or be more pessimistic than
2310	 * the check in __isolate_free_page(). We don't use the direct
2311	 * compactor's alloc_flags, as they are not relevant for freepage
2312	 * isolation. We however do use the direct compactor's highest_zoneidx
2313	 * to skip over zones where lowmem reserves would prevent allocation
2314	 * even if compaction succeeds.
2315	 * For costly orders, we require low watermark instead of min for
2316	 * compaction to proceed to increase its chances.
2317	 * ALLOC_CMA is used, as pages in CMA pageblocks are considered
2318	 * suitable migration targets
2319	 */
2320	watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
2321				low_wmark_pages(zone) : min_wmark_pages(zone);
2322	watermark += compact_gap(order);
2323	return __zone_watermark_ok(zone, 0, watermark, highest_zoneidx,
2324				   ALLOC_CMA, wmark_target);
2325}
2326
2327/*
2328 * compaction_suitable: Is this suitable to run compaction on this zone now?
2329 */
2330bool compaction_suitable(struct zone *zone, int order, int highest_zoneidx)
2331{
2332	enum compact_result compact_result;
2333	bool suitable;
2334
2335	suitable = __compaction_suitable(zone, order, highest_zoneidx,
2336					 zone_page_state(zone, NR_FREE_PAGES));
2337	/*
2338	 * fragmentation index determines if allocation failures are due to
2339	 * low memory or external fragmentation
2340	 *
2341	 * index of -1000 would imply allocations might succeed depending on
2342	 * watermarks, but we already failed the high-order watermark check
2343	 * index towards 0 implies failure is due to lack of memory
2344	 * index towards 1000 implies failure is due to fragmentation
2345	 *
2346	 * Only compact if a failure would be due to fragmentation. Also
2347	 * ignore fragindex for non-costly orders where the alternative to
2348	 * a successful reclaim/compaction is OOM. Fragindex and the
2349	 * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
2350	 * excessive compaction for costly orders, but it should not be at the
2351	 * expense of system stability.
2352	 */
2353	if (suitable) {
2354		compact_result = COMPACT_CONTINUE;
2355		if (order > PAGE_ALLOC_COSTLY_ORDER) {
2356			int fragindex = fragmentation_index(zone, order);
2357
2358			if (fragindex >= 0 &&
2359			    fragindex <= sysctl_extfrag_threshold) {
2360				suitable = false;
2361				compact_result = COMPACT_NOT_SUITABLE_ZONE;
2362			}
2363		}
2364	} else {
2365		compact_result = COMPACT_SKIPPED;
2366	}
2367
2368	trace_mm_compaction_suitable(zone, order, compact_result);
2369
2370	return suitable;
2371}
2372
2373bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
2374		int alloc_flags)
2375{
2376	struct zone *zone;
2377	struct zoneref *z;
2378
2379	/*
2380	 * Make sure at least one zone would pass __compaction_suitable if we continue
2381	 * retrying the reclaim.
2382	 */
2383	for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2384				ac->highest_zoneidx, ac->nodemask) {
2385		unsigned long available;
2386
2387		/*
2388		 * Do not consider all the reclaimable memory because we do not
2389		 * want to trash just for a single high order allocation which
2390		 * is even not guaranteed to appear even if __compaction_suitable
2391		 * is happy about the watermark check.
2392		 */
2393		available = zone_reclaimable_pages(zone) / order;
2394		available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
2395		if (__compaction_suitable(zone, order, ac->highest_zoneidx,
2396					  available))
2397			return true;
2398	}
2399
2400	return false;
2401}
2402
2403/*
2404 * Should we do compaction for target allocation order.
2405 * Return COMPACT_SUCCESS if allocation for target order can be already
2406 * satisfied
2407 * Return COMPACT_SKIPPED if compaction for target order is likely to fail
2408 * Return COMPACT_CONTINUE if compaction for target order should be ran
2409 */
2410static enum compact_result
2411compaction_suit_allocation_order(struct zone *zone, unsigned int order,
2412				 int highest_zoneidx, unsigned int alloc_flags)
2413{
2414	unsigned long watermark;
2415
2416	watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
2417	if (zone_watermark_ok(zone, order, watermark, highest_zoneidx,
2418			      alloc_flags))
2419		return COMPACT_SUCCESS;
2420
2421	if (!compaction_suitable(zone, order, highest_zoneidx))
2422		return COMPACT_SKIPPED;
2423
2424	return COMPACT_CONTINUE;
2425}
2426
2427static enum compact_result
2428compact_zone(struct compact_control *cc, struct capture_control *capc)
2429{
2430	enum compact_result ret;
2431	unsigned long start_pfn = cc->zone->zone_start_pfn;
2432	unsigned long end_pfn = zone_end_pfn(cc->zone);
2433	unsigned long last_migrated_pfn;
2434	const bool sync = cc->mode != MIGRATE_ASYNC;
2435	bool update_cached;
2436	unsigned int nr_succeeded = 0;
 
2437
2438	/*
2439	 * These counters track activities during zone compaction.  Initialize
2440	 * them before compacting a new zone.
2441	 */
2442	cc->total_migrate_scanned = 0;
2443	cc->total_free_scanned = 0;
2444	cc->nr_migratepages = 0;
2445	cc->nr_freepages = 0;
2446	INIT_LIST_HEAD(&cc->freepages);
 
2447	INIT_LIST_HEAD(&cc->migratepages);
2448
2449	cc->migratetype = gfp_migratetype(cc->gfp_mask);
2450
2451	if (!is_via_compact_memory(cc->order)) {
2452		ret = compaction_suit_allocation_order(cc->zone, cc->order,
2453						       cc->highest_zoneidx,
2454						       cc->alloc_flags);
2455		if (ret != COMPACT_CONTINUE)
2456			return ret;
2457	}
2458
2459	/*
2460	 * Clear pageblock skip if there were failures recently and compaction
2461	 * is about to be retried after being deferred.
2462	 */
2463	if (compaction_restarting(cc->zone, cc->order))
2464		__reset_isolation_suitable(cc->zone);
2465
2466	/*
2467	 * Setup to move all movable pages to the end of the zone. Used cached
2468	 * information on where the scanners should start (unless we explicitly
2469	 * want to compact the whole zone), but check that it is initialised
2470	 * by ensuring the values are within zone boundaries.
2471	 */
2472	cc->fast_start_pfn = 0;
2473	if (cc->whole_zone) {
2474		cc->migrate_pfn = start_pfn;
2475		cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2476	} else {
2477		cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync];
2478		cc->free_pfn = cc->zone->compact_cached_free_pfn;
2479		if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
2480			cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2481			cc->zone->compact_cached_free_pfn = cc->free_pfn;
2482		}
2483		if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
2484			cc->migrate_pfn = start_pfn;
2485			cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
2486			cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
2487		}
2488
2489		if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn)
2490			cc->whole_zone = true;
2491	}
2492
2493	last_migrated_pfn = 0;
2494
2495	/*
2496	 * Migrate has separate cached PFNs for ASYNC and SYNC* migration on
2497	 * the basis that some migrations will fail in ASYNC mode. However,
2498	 * if the cached PFNs match and pageblocks are skipped due to having
2499	 * no isolation candidates, then the sync state does not matter.
2500	 * Until a pageblock with isolation candidates is found, keep the
2501	 * cached PFNs in sync to avoid revisiting the same blocks.
2502	 */
2503	update_cached = !sync &&
2504		cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1];
2505
2506	trace_mm_compaction_begin(cc, start_pfn, end_pfn, sync);
2507
2508	/* lru_add_drain_all could be expensive with involving other CPUs */
2509	lru_add_drain();
2510
2511	while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) {
2512		int err;
2513		unsigned long iteration_start_pfn = cc->migrate_pfn;
2514
2515		/*
2516		 * Avoid multiple rescans of the same pageblock which can
2517		 * happen if a page cannot be isolated (dirty/writeback in
2518		 * async mode) or if the migrated pages are being allocated
2519		 * before the pageblock is cleared.  The first rescan will
2520		 * capture the entire pageblock for migration. If it fails,
2521		 * it'll be marked skip and scanning will proceed as normal.
2522		 */
2523		cc->finish_pageblock = false;
2524		if (pageblock_start_pfn(last_migrated_pfn) ==
2525		    pageblock_start_pfn(iteration_start_pfn)) {
2526			cc->finish_pageblock = true;
2527		}
2528
2529rescan:
2530		switch (isolate_migratepages(cc)) {
2531		case ISOLATE_ABORT:
2532			ret = COMPACT_CONTENDED;
2533			putback_movable_pages(&cc->migratepages);
2534			cc->nr_migratepages = 0;
2535			goto out;
2536		case ISOLATE_NONE:
2537			if (update_cached) {
2538				cc->zone->compact_cached_migrate_pfn[1] =
2539					cc->zone->compact_cached_migrate_pfn[0];
2540			}
2541
2542			/*
2543			 * We haven't isolated and migrated anything, but
2544			 * there might still be unflushed migrations from
2545			 * previous cc->order aligned block.
2546			 */
2547			goto check_drain;
2548		case ISOLATE_SUCCESS:
2549			update_cached = false;
2550			last_migrated_pfn = max(cc->zone->zone_start_pfn,
2551				pageblock_start_pfn(cc->migrate_pfn - 1));
2552		}
2553
 
 
 
 
 
 
2554		err = migrate_pages(&cc->migratepages, compaction_alloc,
2555				compaction_free, (unsigned long)cc, cc->mode,
2556				MR_COMPACTION, &nr_succeeded);
2557
2558		trace_mm_compaction_migratepages(cc, nr_succeeded);
2559
2560		/* All pages were either migrated or will be released */
2561		cc->nr_migratepages = 0;
2562		if (err) {
2563			putback_movable_pages(&cc->migratepages);
2564			/*
2565			 * migrate_pages() may return -ENOMEM when scanners meet
2566			 * and we want compact_finished() to detect it
2567			 */
2568			if (err == -ENOMEM && !compact_scanners_met(cc)) {
2569				ret = COMPACT_CONTENDED;
2570				goto out;
2571			}
2572			/*
2573			 * If an ASYNC or SYNC_LIGHT fails to migrate a page
2574			 * within the pageblock_order-aligned block and
2575			 * fast_find_migrateblock may be used then scan the
2576			 * remainder of the pageblock. This will mark the
2577			 * pageblock "skip" to avoid rescanning in the near
2578			 * future. This will isolate more pages than necessary
2579			 * for the request but avoid loops due to
2580			 * fast_find_migrateblock revisiting blocks that were
2581			 * recently partially scanned.
2582			 */
2583			if (!pageblock_aligned(cc->migrate_pfn) &&
2584			    !cc->ignore_skip_hint && !cc->finish_pageblock &&
2585			    (cc->mode < MIGRATE_SYNC)) {
2586				cc->finish_pageblock = true;
2587
2588				/*
2589				 * Draining pcplists does not help THP if
2590				 * any page failed to migrate. Even after
2591				 * drain, the pageblock will not be free.
2592				 */
2593				if (cc->order == COMPACTION_HPAGE_ORDER)
2594					last_migrated_pfn = 0;
2595
2596				goto rescan;
2597			}
2598		}
2599
2600		/* Stop if a page has been captured */
2601		if (capc && capc->page) {
2602			ret = COMPACT_SUCCESS;
2603			break;
2604		}
2605
2606check_drain:
2607		/*
2608		 * Has the migration scanner moved away from the previous
2609		 * cc->order aligned block where we migrated from? If yes,
2610		 * flush the pages that were freed, so that they can merge and
2611		 * compact_finished() can detect immediately if allocation
2612		 * would succeed.
2613		 */
2614		if (cc->order > 0 && last_migrated_pfn) {
2615			unsigned long current_block_start =
2616				block_start_pfn(cc->migrate_pfn, cc->order);
2617
2618			if (last_migrated_pfn < current_block_start) {
2619				lru_add_drain_cpu_zone(cc->zone);
2620				/* No more flushing until we migrate again */
2621				last_migrated_pfn = 0;
2622			}
2623		}
2624	}
2625
2626out:
2627	/*
2628	 * Release free pages and update where the free scanner should restart,
2629	 * so we don't leave any returned pages behind in the next attempt.
2630	 */
2631	if (cc->nr_freepages > 0) {
2632		unsigned long free_pfn = release_freepages(&cc->freepages);
2633
2634		cc->nr_freepages = 0;
2635		VM_BUG_ON(free_pfn == 0);
2636		/* The cached pfn is always the first in a pageblock */
2637		free_pfn = pageblock_start_pfn(free_pfn);
2638		/*
2639		 * Only go back, not forward. The cached pfn might have been
2640		 * already reset to zone end in compact_finished()
2641		 */
2642		if (free_pfn > cc->zone->compact_cached_free_pfn)
2643			cc->zone->compact_cached_free_pfn = free_pfn;
2644	}
2645
2646	count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned);
2647	count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned);
2648
2649	trace_mm_compaction_end(cc, start_pfn, end_pfn, sync, ret);
2650
2651	VM_BUG_ON(!list_empty(&cc->freepages));
2652	VM_BUG_ON(!list_empty(&cc->migratepages));
2653
2654	return ret;
2655}
2656
2657static enum compact_result compact_zone_order(struct zone *zone, int order,
2658		gfp_t gfp_mask, enum compact_priority prio,
2659		unsigned int alloc_flags, int highest_zoneidx,
2660		struct page **capture)
2661{
2662	enum compact_result ret;
2663	struct compact_control cc = {
2664		.order = order,
2665		.search_order = order,
2666		.gfp_mask = gfp_mask,
2667		.zone = zone,
2668		.mode = (prio == COMPACT_PRIO_ASYNC) ?
2669					MIGRATE_ASYNC :	MIGRATE_SYNC_LIGHT,
2670		.alloc_flags = alloc_flags,
2671		.highest_zoneidx = highest_zoneidx,
2672		.direct_compaction = true,
2673		.whole_zone = (prio == MIN_COMPACT_PRIORITY),
2674		.ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
2675		.ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
2676	};
2677	struct capture_control capc = {
2678		.cc = &cc,
2679		.page = NULL,
2680	};
2681
2682	/*
2683	 * Make sure the structs are really initialized before we expose the
2684	 * capture control, in case we are interrupted and the interrupt handler
2685	 * frees a page.
2686	 */
2687	barrier();
2688	WRITE_ONCE(current->capture_control, &capc);
2689
2690	ret = compact_zone(&cc, &capc);
2691
2692	/*
2693	 * Make sure we hide capture control first before we read the captured
2694	 * page pointer, otherwise an interrupt could free and capture a page
2695	 * and we would leak it.
2696	 */
2697	WRITE_ONCE(current->capture_control, NULL);
2698	*capture = READ_ONCE(capc.page);
2699	/*
2700	 * Technically, it is also possible that compaction is skipped but
2701	 * the page is still captured out of luck(IRQ came and freed the page).
2702	 * Returning COMPACT_SUCCESS in such cases helps in properly accounting
2703	 * the COMPACT[STALL|FAIL] when compaction is skipped.
2704	 */
2705	if (*capture)
2706		ret = COMPACT_SUCCESS;
2707
2708	return ret;
2709}
2710
2711/**
2712 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
2713 * @gfp_mask: The GFP mask of the current allocation
2714 * @order: The order of the current allocation
2715 * @alloc_flags: The allocation flags of the current allocation
2716 * @ac: The context of current allocation
2717 * @prio: Determines how hard direct compaction should try to succeed
2718 * @capture: Pointer to free page created by compaction will be stored here
2719 *
2720 * This is the main entry point for direct page compaction.
2721 */
2722enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
2723		unsigned int alloc_flags, const struct alloc_context *ac,
2724		enum compact_priority prio, struct page **capture)
2725{
2726	struct zoneref *z;
2727	struct zone *zone;
2728	enum compact_result rc = COMPACT_SKIPPED;
2729
2730	if (!gfp_compaction_allowed(gfp_mask))
2731		return COMPACT_SKIPPED;
2732
2733	trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
2734
2735	/* Compact each zone in the list */
2736	for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2737					ac->highest_zoneidx, ac->nodemask) {
2738		enum compact_result status;
2739
2740		if (prio > MIN_COMPACT_PRIORITY
2741					&& compaction_deferred(zone, order)) {
2742			rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
2743			continue;
2744		}
2745
2746		status = compact_zone_order(zone, order, gfp_mask, prio,
2747				alloc_flags, ac->highest_zoneidx, capture);
2748		rc = max(status, rc);
2749
2750		/* The allocation should succeed, stop compacting */
2751		if (status == COMPACT_SUCCESS) {
2752			/*
2753			 * We think the allocation will succeed in this zone,
2754			 * but it is not certain, hence the false. The caller
2755			 * will repeat this with true if allocation indeed
2756			 * succeeds in this zone.
2757			 */
2758			compaction_defer_reset(zone, order, false);
2759
2760			break;
2761		}
2762
2763		if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
2764					status == COMPACT_PARTIAL_SKIPPED))
2765			/*
2766			 * We think that allocation won't succeed in this zone
2767			 * so we defer compaction there. If it ends up
2768			 * succeeding after all, it will be reset.
2769			 */
2770			defer_compaction(zone, order);
2771
2772		/*
2773		 * We might have stopped compacting due to need_resched() in
2774		 * async compaction, or due to a fatal signal detected. In that
2775		 * case do not try further zones
2776		 */
2777		if ((prio == COMPACT_PRIO_ASYNC && need_resched())
2778					|| fatal_signal_pending(current))
2779			break;
2780	}
2781
2782	return rc;
2783}
2784
2785/*
2786 * Compact all zones within a node till each zone's fragmentation score
2787 * reaches within proactive compaction thresholds (as determined by the
2788 * proactiveness tunable).
2789 *
2790 * It is possible that the function returns before reaching score targets
2791 * due to various back-off conditions, such as, contention on per-node or
2792 * per-zone locks.
 
 
2793 */
2794static void proactive_compact_node(pg_data_t *pgdat)
2795{
2796	int zoneid;
2797	struct zone *zone;
2798	struct compact_control cc = {
2799		.order = -1,
2800		.mode = MIGRATE_SYNC_LIGHT,
2801		.ignore_skip_hint = true,
2802		.whole_zone = true,
2803		.gfp_mask = GFP_KERNEL,
2804		.proactive_compaction = true,
2805	};
2806
2807	for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2808		zone = &pgdat->node_zones[zoneid];
2809		if (!populated_zone(zone))
2810			continue;
2811
 
 
 
2812		cc.zone = zone;
2813
2814		compact_zone(&cc, NULL);
2815
2816		count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
2817				     cc.total_migrate_scanned);
2818		count_compact_events(KCOMPACTD_FREE_SCANNED,
2819				     cc.total_free_scanned);
 
 
2820	}
2821}
2822
2823/* Compact all zones within a node */
2824static void compact_node(int nid)
2825{
2826	pg_data_t *pgdat = NODE_DATA(nid);
2827	int zoneid;
2828	struct zone *zone;
2829	struct compact_control cc = {
2830		.order = -1,
2831		.mode = MIGRATE_SYNC,
2832		.ignore_skip_hint = true,
2833		.whole_zone = true,
2834		.gfp_mask = GFP_KERNEL,
2835	};
2836
2837
2838	for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2839
2840		zone = &pgdat->node_zones[zoneid];
2841		if (!populated_zone(zone))
2842			continue;
2843
2844		cc.zone = zone;
2845
2846		compact_zone(&cc, NULL);
2847	}
2848}
2849
2850/* Compact all nodes in the system */
2851static void compact_nodes(void)
2852{
2853	int nid;
2854
2855	/* Flush pending updates to the LRU lists */
2856	lru_add_drain_all();
2857
2858	for_each_online_node(nid)
2859		compact_node(nid);
 
 
 
 
 
2860}
2861
2862static int compaction_proactiveness_sysctl_handler(struct ctl_table *table, int write,
2863		void *buffer, size_t *length, loff_t *ppos)
2864{
2865	int rc, nid;
2866
2867	rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
2868	if (rc)
2869		return rc;
2870
2871	if (write && sysctl_compaction_proactiveness) {
2872		for_each_online_node(nid) {
2873			pg_data_t *pgdat = NODE_DATA(nid);
2874
2875			if (pgdat->proactive_compact_trigger)
2876				continue;
2877
2878			pgdat->proactive_compact_trigger = true;
2879			trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, -1,
2880							     pgdat->nr_zones - 1);
2881			wake_up_interruptible(&pgdat->kcompactd_wait);
2882		}
2883	}
2884
2885	return 0;
2886}
2887
2888/*
2889 * This is the entry point for compacting all nodes via
2890 * /proc/sys/vm/compact_memory
2891 */
2892static int sysctl_compaction_handler(struct ctl_table *table, int write,
2893			void *buffer, size_t *length, loff_t *ppos)
2894{
2895	int ret;
2896
2897	ret = proc_dointvec(table, write, buffer, length, ppos);
2898	if (ret)
2899		return ret;
2900
2901	if (sysctl_compact_memory != 1)
2902		return -EINVAL;
2903
2904	if (write)
2905		compact_nodes();
2906
2907	return 0;
2908}
2909
2910#if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
2911static ssize_t compact_store(struct device *dev,
2912			     struct device_attribute *attr,
2913			     const char *buf, size_t count)
2914{
2915	int nid = dev->id;
2916
2917	if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
2918		/* Flush pending updates to the LRU lists */
2919		lru_add_drain_all();
2920
2921		compact_node(nid);
2922	}
2923
2924	return count;
2925}
2926static DEVICE_ATTR_WO(compact);
2927
2928int compaction_register_node(struct node *node)
2929{
2930	return device_create_file(&node->dev, &dev_attr_compact);
2931}
2932
2933void compaction_unregister_node(struct node *node)
2934{
2935	device_remove_file(&node->dev, &dev_attr_compact);
2936}
2937#endif /* CONFIG_SYSFS && CONFIG_NUMA */
2938
2939static inline bool kcompactd_work_requested(pg_data_t *pgdat)
2940{
2941	return pgdat->kcompactd_max_order > 0 || kthread_should_stop() ||
2942		pgdat->proactive_compact_trigger;
2943}
2944
2945static bool kcompactd_node_suitable(pg_data_t *pgdat)
2946{
2947	int zoneid;
2948	struct zone *zone;
2949	enum zone_type highest_zoneidx = pgdat->kcompactd_highest_zoneidx;
2950	enum compact_result ret;
2951
2952	for (zoneid = 0; zoneid <= highest_zoneidx; zoneid++) {
2953		zone = &pgdat->node_zones[zoneid];
2954
2955		if (!populated_zone(zone))
2956			continue;
2957
2958		ret = compaction_suit_allocation_order(zone,
2959				pgdat->kcompactd_max_order,
2960				highest_zoneidx, ALLOC_WMARK_MIN);
2961		if (ret == COMPACT_CONTINUE)
2962			return true;
2963	}
2964
2965	return false;
2966}
2967
2968static void kcompactd_do_work(pg_data_t *pgdat)
2969{
2970	/*
2971	 * With no special task, compact all zones so that a page of requested
2972	 * order is allocatable.
2973	 */
2974	int zoneid;
2975	struct zone *zone;
2976	struct compact_control cc = {
2977		.order = pgdat->kcompactd_max_order,
2978		.search_order = pgdat->kcompactd_max_order,
2979		.highest_zoneidx = pgdat->kcompactd_highest_zoneidx,
2980		.mode = MIGRATE_SYNC_LIGHT,
2981		.ignore_skip_hint = false,
2982		.gfp_mask = GFP_KERNEL,
2983	};
2984	enum compact_result ret;
2985
2986	trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
2987							cc.highest_zoneidx);
2988	count_compact_event(KCOMPACTD_WAKE);
2989
2990	for (zoneid = 0; zoneid <= cc.highest_zoneidx; zoneid++) {
2991		int status;
2992
2993		zone = &pgdat->node_zones[zoneid];
2994		if (!populated_zone(zone))
2995			continue;
2996
2997		if (compaction_deferred(zone, cc.order))
2998			continue;
2999
3000		ret = compaction_suit_allocation_order(zone,
3001				cc.order, zoneid, ALLOC_WMARK_MIN);
3002		if (ret != COMPACT_CONTINUE)
3003			continue;
3004
3005		if (kthread_should_stop())
3006			return;
3007
3008		cc.zone = zone;
3009		status = compact_zone(&cc, NULL);
3010
3011		if (status == COMPACT_SUCCESS) {
3012			compaction_defer_reset(zone, cc.order, false);
3013		} else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
3014			/*
3015			 * Buddy pages may become stranded on pcps that could
3016			 * otherwise coalesce on the zone's free area for
3017			 * order >= cc.order.  This is ratelimited by the
3018			 * upcoming deferral.
3019			 */
3020			drain_all_pages(zone);
3021
3022			/*
3023			 * We use sync migration mode here, so we defer like
3024			 * sync direct compaction does.
3025			 */
3026			defer_compaction(zone, cc.order);
3027		}
3028
3029		count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
3030				     cc.total_migrate_scanned);
3031		count_compact_events(KCOMPACTD_FREE_SCANNED,
3032				     cc.total_free_scanned);
3033	}
3034
3035	/*
3036	 * Regardless of success, we are done until woken up next. But remember
3037	 * the requested order/highest_zoneidx in case it was higher/tighter
3038	 * than our current ones
3039	 */
3040	if (pgdat->kcompactd_max_order <= cc.order)
3041		pgdat->kcompactd_max_order = 0;
3042	if (pgdat->kcompactd_highest_zoneidx >= cc.highest_zoneidx)
3043		pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
3044}
3045
3046void wakeup_kcompactd(pg_data_t *pgdat, int order, int highest_zoneidx)
3047{
3048	if (!order)
3049		return;
3050
3051	if (pgdat->kcompactd_max_order < order)
3052		pgdat->kcompactd_max_order = order;
3053
3054	if (pgdat->kcompactd_highest_zoneidx > highest_zoneidx)
3055		pgdat->kcompactd_highest_zoneidx = highest_zoneidx;
3056
3057	/*
3058	 * Pairs with implicit barrier in wait_event_freezable()
3059	 * such that wakeups are not missed.
3060	 */
3061	if (!wq_has_sleeper(&pgdat->kcompactd_wait))
3062		return;
3063
3064	if (!kcompactd_node_suitable(pgdat))
3065		return;
3066
3067	trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
3068							highest_zoneidx);
3069	wake_up_interruptible(&pgdat->kcompactd_wait);
3070}
3071
3072/*
3073 * The background compaction daemon, started as a kernel thread
3074 * from the init process.
3075 */
3076static int kcompactd(void *p)
3077{
3078	pg_data_t *pgdat = (pg_data_t *)p;
3079	struct task_struct *tsk = current;
3080	long default_timeout = msecs_to_jiffies(HPAGE_FRAG_CHECK_INTERVAL_MSEC);
3081	long timeout = default_timeout;
3082
3083	const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3084
3085	if (!cpumask_empty(cpumask))
3086		set_cpus_allowed_ptr(tsk, cpumask);
3087
3088	set_freezable();
3089
3090	pgdat->kcompactd_max_order = 0;
3091	pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
3092
3093	while (!kthread_should_stop()) {
3094		unsigned long pflags;
3095
3096		/*
3097		 * Avoid the unnecessary wakeup for proactive compaction
3098		 * when it is disabled.
3099		 */
3100		if (!sysctl_compaction_proactiveness)
3101			timeout = MAX_SCHEDULE_TIMEOUT;
3102		trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
3103		if (wait_event_freezable_timeout(pgdat->kcompactd_wait,
3104			kcompactd_work_requested(pgdat), timeout) &&
3105			!pgdat->proactive_compact_trigger) {
3106
3107			psi_memstall_enter(&pflags);
3108			kcompactd_do_work(pgdat);
3109			psi_memstall_leave(&pflags);
3110			/*
3111			 * Reset the timeout value. The defer timeout from
3112			 * proactive compaction is lost here but that is fine
3113			 * as the condition of the zone changing substantionally
3114			 * then carrying on with the previous defer interval is
3115			 * not useful.
3116			 */
3117			timeout = default_timeout;
3118			continue;
3119		}
3120
3121		/*
3122		 * Start the proactive work with default timeout. Based
3123		 * on the fragmentation score, this timeout is updated.
3124		 */
3125		timeout = default_timeout;
3126		if (should_proactive_compact_node(pgdat)) {
3127			unsigned int prev_score, score;
3128
3129			prev_score = fragmentation_score_node(pgdat);
3130			proactive_compact_node(pgdat);
3131			score = fragmentation_score_node(pgdat);
3132			/*
3133			 * Defer proactive compaction if the fragmentation
3134			 * score did not go down i.e. no progress made.
3135			 */
3136			if (unlikely(score >= prev_score))
3137				timeout =
3138				   default_timeout << COMPACT_MAX_DEFER_SHIFT;
3139		}
3140		if (unlikely(pgdat->proactive_compact_trigger))
3141			pgdat->proactive_compact_trigger = false;
3142	}
3143
3144	return 0;
3145}
3146
3147/*
3148 * This kcompactd start function will be called by init and node-hot-add.
3149 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
3150 */
3151void __meminit kcompactd_run(int nid)
3152{
3153	pg_data_t *pgdat = NODE_DATA(nid);
3154
3155	if (pgdat->kcompactd)
3156		return;
3157
3158	pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
3159	if (IS_ERR(pgdat->kcompactd)) {
3160		pr_err("Failed to start kcompactd on node %d\n", nid);
3161		pgdat->kcompactd = NULL;
3162	}
3163}
3164
3165/*
3166 * Called by memory hotplug when all memory in a node is offlined. Caller must
3167 * be holding mem_hotplug_begin/done().
3168 */
3169void __meminit kcompactd_stop(int nid)
3170{
3171	struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
3172
3173	if (kcompactd) {
3174		kthread_stop(kcompactd);
3175		NODE_DATA(nid)->kcompactd = NULL;
3176	}
3177}
3178
3179/*
3180 * It's optimal to keep kcompactd on the same CPUs as their memory, but
3181 * not required for correctness. So if the last cpu in a node goes
3182 * away, we get changed to run anywhere: as the first one comes back,
3183 * restore their cpu bindings.
3184 */
3185static int kcompactd_cpu_online(unsigned int cpu)
3186{
3187	int nid;
3188
3189	for_each_node_state(nid, N_MEMORY) {
3190		pg_data_t *pgdat = NODE_DATA(nid);
3191		const struct cpumask *mask;
3192
3193		mask = cpumask_of_node(pgdat->node_id);
3194
3195		if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3196			/* One of our CPUs online: restore mask */
3197			if (pgdat->kcompactd)
3198				set_cpus_allowed_ptr(pgdat->kcompactd, mask);
3199	}
3200	return 0;
3201}
3202
3203static int proc_dointvec_minmax_warn_RT_change(struct ctl_table *table,
3204		int write, void *buffer, size_t *lenp, loff_t *ppos)
3205{
3206	int ret, old;
3207
3208	if (!IS_ENABLED(CONFIG_PREEMPT_RT) || !write)
3209		return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
3210
3211	old = *(int *)table->data;
3212	ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
3213	if (ret)
3214		return ret;
3215	if (old != *(int *)table->data)
3216		pr_warn_once("sysctl attribute %s changed by %s[%d]\n",
3217			     table->procname, current->comm,
3218			     task_pid_nr(current));
3219	return ret;
3220}
3221
3222static struct ctl_table vm_compaction[] = {
3223	{
3224		.procname	= "compact_memory",
3225		.data		= &sysctl_compact_memory,
3226		.maxlen		= sizeof(int),
3227		.mode		= 0200,
3228		.proc_handler	= sysctl_compaction_handler,
3229	},
3230	{
3231		.procname	= "compaction_proactiveness",
3232		.data		= &sysctl_compaction_proactiveness,
3233		.maxlen		= sizeof(sysctl_compaction_proactiveness),
3234		.mode		= 0644,
3235		.proc_handler	= compaction_proactiveness_sysctl_handler,
3236		.extra1		= SYSCTL_ZERO,
3237		.extra2		= SYSCTL_ONE_HUNDRED,
3238	},
3239	{
3240		.procname	= "extfrag_threshold",
3241		.data		= &sysctl_extfrag_threshold,
3242		.maxlen		= sizeof(int),
3243		.mode		= 0644,
3244		.proc_handler	= proc_dointvec_minmax,
3245		.extra1		= SYSCTL_ZERO,
3246		.extra2		= SYSCTL_ONE_THOUSAND,
3247	},
3248	{
3249		.procname	= "compact_unevictable_allowed",
3250		.data		= &sysctl_compact_unevictable_allowed,
3251		.maxlen		= sizeof(int),
3252		.mode		= 0644,
3253		.proc_handler	= proc_dointvec_minmax_warn_RT_change,
3254		.extra1		= SYSCTL_ZERO,
3255		.extra2		= SYSCTL_ONE,
3256	},
3257	{ }
3258};
3259
3260static int __init kcompactd_init(void)
3261{
3262	int nid;
3263	int ret;
3264
3265	ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3266					"mm/compaction:online",
3267					kcompactd_cpu_online, NULL);
3268	if (ret < 0) {
3269		pr_err("kcompactd: failed to register hotplug callbacks.\n");
3270		return ret;
3271	}
3272
3273	for_each_node_state(nid, N_MEMORY)
3274		kcompactd_run(nid);
3275	register_sysctl_init("vm", vm_compaction);
3276	return 0;
3277}
3278subsys_initcall(kcompactd_init)
3279
3280#endif /* CONFIG_COMPACTION */
v6.9.4
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * linux/mm/compaction.c
   4 *
   5 * Memory compaction for the reduction of external fragmentation. Note that
   6 * this heavily depends upon page migration to do all the real heavy
   7 * lifting
   8 *
   9 * Copyright IBM Corp. 2007-2010 Mel Gorman <mel@csn.ul.ie>
  10 */
  11#include <linux/cpu.h>
  12#include <linux/swap.h>
  13#include <linux/migrate.h>
  14#include <linux/compaction.h>
  15#include <linux/mm_inline.h>
  16#include <linux/sched/signal.h>
  17#include <linux/backing-dev.h>
  18#include <linux/sysctl.h>
  19#include <linux/sysfs.h>
  20#include <linux/page-isolation.h>
  21#include <linux/kasan.h>
  22#include <linux/kthread.h>
  23#include <linux/freezer.h>
  24#include <linux/page_owner.h>
  25#include <linux/psi.h>
  26#include "internal.h"
  27
  28#ifdef CONFIG_COMPACTION
  29/*
  30 * Fragmentation score check interval for proactive compaction purposes.
  31 */
  32#define HPAGE_FRAG_CHECK_INTERVAL_MSEC	(500)
  33
  34static inline void count_compact_event(enum vm_event_item item)
  35{
  36	count_vm_event(item);
  37}
  38
  39static inline void count_compact_events(enum vm_event_item item, long delta)
  40{
  41	count_vm_events(item, delta);
  42}
  43
  44/*
  45 * order == -1 is expected when compacting proactively via
  46 * 1. /proc/sys/vm/compact_memory
  47 * 2. /sys/devices/system/node/nodex/compact
  48 * 3. /proc/sys/vm/compaction_proactiveness
  49 */
  50static inline bool is_via_compact_memory(int order)
  51{
  52	return order == -1;
  53}
  54
  55#else
  56#define count_compact_event(item) do { } while (0)
  57#define count_compact_events(item, delta) do { } while (0)
  58static inline bool is_via_compact_memory(int order) { return false; }
  59#endif
  60
  61#if defined CONFIG_COMPACTION || defined CONFIG_CMA
  62
  63#define CREATE_TRACE_POINTS
  64#include <trace/events/compaction.h>
  65
  66#define block_start_pfn(pfn, order)	round_down(pfn, 1UL << (order))
  67#define block_end_pfn(pfn, order)	ALIGN((pfn) + 1, 1UL << (order))
  68
  69/*
  70 * Page order with-respect-to which proactive compaction
  71 * calculates external fragmentation, which is used as
  72 * the "fragmentation score" of a node/zone.
  73 */
  74#if defined CONFIG_TRANSPARENT_HUGEPAGE
  75#define COMPACTION_HPAGE_ORDER	HPAGE_PMD_ORDER
  76#elif defined CONFIG_HUGETLBFS
  77#define COMPACTION_HPAGE_ORDER	HUGETLB_PAGE_ORDER
  78#else
  79#define COMPACTION_HPAGE_ORDER	(PMD_SHIFT - PAGE_SHIFT)
  80#endif
  81
  82static void split_map_pages(struct list_head *freepages)
  83{
  84	unsigned int i, order;
  85	struct page *page, *next;
  86	LIST_HEAD(tmp_list);
  87
  88	for (order = 0; order < NR_PAGE_ORDERS; order++) {
  89		list_for_each_entry_safe(page, next, &freepages[order], lru) {
  90			unsigned int nr_pages;
 
 
 
 
  91
  92			list_del(&page->lru);
  93
  94			nr_pages = 1 << order;
  95
  96			post_alloc_hook(page, order, __GFP_MOVABLE);
  97			if (order)
  98				split_page(page, order);
  99
 100			for (i = 0; i < nr_pages; i++) {
 101				list_add(&page->lru, &tmp_list);
 102				page++;
 103			}
 104		}
 105		list_splice_init(&tmp_list, &freepages[0]);
 106	}
 107}
 108
 109static unsigned long release_free_list(struct list_head *freepages)
 110{
 111	int order;
 112	unsigned long high_pfn = 0;
 
 
 
 
 113
 114	for (order = 0; order < NR_PAGE_ORDERS; order++) {
 115		struct page *page, *next;
 116
 117		list_for_each_entry_safe(page, next, &freepages[order], lru) {
 118			unsigned long pfn = page_to_pfn(page);
 
 119
 120			list_del(&page->lru);
 121			/*
 122			 * Convert free pages into post allocation pages, so
 123			 * that we can free them via __free_page.
 124			 */
 125			post_alloc_hook(page, order, __GFP_MOVABLE);
 126			__free_pages(page, order);
 127			if (pfn > high_pfn)
 128				high_pfn = pfn;
 129		}
 130	}
 131	return high_pfn;
 
 132}
 133
 134#ifdef CONFIG_COMPACTION
 135bool PageMovable(struct page *page)
 136{
 137	const struct movable_operations *mops;
 138
 139	VM_BUG_ON_PAGE(!PageLocked(page), page);
 140	if (!__PageMovable(page))
 141		return false;
 142
 143	mops = page_movable_ops(page);
 144	if (mops)
 145		return true;
 146
 147	return false;
 148}
 149
 150void __SetPageMovable(struct page *page, const struct movable_operations *mops)
 151{
 152	VM_BUG_ON_PAGE(!PageLocked(page), page);
 153	VM_BUG_ON_PAGE((unsigned long)mops & PAGE_MAPPING_MOVABLE, page);
 154	page->mapping = (void *)((unsigned long)mops | PAGE_MAPPING_MOVABLE);
 155}
 156EXPORT_SYMBOL(__SetPageMovable);
 157
 158void __ClearPageMovable(struct page *page)
 159{
 160	VM_BUG_ON_PAGE(!PageMovable(page), page);
 161	/*
 162	 * This page still has the type of a movable page, but it's
 163	 * actually not movable any more.
 164	 */
 165	page->mapping = (void *)PAGE_MAPPING_MOVABLE;
 166}
 167EXPORT_SYMBOL(__ClearPageMovable);
 168
 169/* Do not skip compaction more than 64 times */
 170#define COMPACT_MAX_DEFER_SHIFT 6
 171
 172/*
 173 * Compaction is deferred when compaction fails to result in a page
 174 * allocation success. 1 << compact_defer_shift, compactions are skipped up
 175 * to a limit of 1 << COMPACT_MAX_DEFER_SHIFT
 176 */
 177static void defer_compaction(struct zone *zone, int order)
 178{
 179	zone->compact_considered = 0;
 180	zone->compact_defer_shift++;
 181
 182	if (order < zone->compact_order_failed)
 183		zone->compact_order_failed = order;
 184
 185	if (zone->compact_defer_shift > COMPACT_MAX_DEFER_SHIFT)
 186		zone->compact_defer_shift = COMPACT_MAX_DEFER_SHIFT;
 187
 188	trace_mm_compaction_defer_compaction(zone, order);
 189}
 190
 191/* Returns true if compaction should be skipped this time */
 192static bool compaction_deferred(struct zone *zone, int order)
 193{
 194	unsigned long defer_limit = 1UL << zone->compact_defer_shift;
 195
 196	if (order < zone->compact_order_failed)
 197		return false;
 198
 199	/* Avoid possible overflow */
 200	if (++zone->compact_considered >= defer_limit) {
 201		zone->compact_considered = defer_limit;
 202		return false;
 203	}
 204
 205	trace_mm_compaction_deferred(zone, order);
 206
 207	return true;
 208}
 209
 210/*
 211 * Update defer tracking counters after successful compaction of given order,
 212 * which means an allocation either succeeded (alloc_success == true) or is
 213 * expected to succeed.
 214 */
 215void compaction_defer_reset(struct zone *zone, int order,
 216		bool alloc_success)
 217{
 218	if (alloc_success) {
 219		zone->compact_considered = 0;
 220		zone->compact_defer_shift = 0;
 221	}
 222	if (order >= zone->compact_order_failed)
 223		zone->compact_order_failed = order + 1;
 224
 225	trace_mm_compaction_defer_reset(zone, order);
 226}
 227
 228/* Returns true if restarting compaction after many failures */
 229static bool compaction_restarting(struct zone *zone, int order)
 230{
 231	if (order < zone->compact_order_failed)
 232		return false;
 233
 234	return zone->compact_defer_shift == COMPACT_MAX_DEFER_SHIFT &&
 235		zone->compact_considered >= 1UL << zone->compact_defer_shift;
 236}
 237
 238/* Returns true if the pageblock should be scanned for pages to isolate. */
 239static inline bool isolation_suitable(struct compact_control *cc,
 240					struct page *page)
 241{
 242	if (cc->ignore_skip_hint)
 243		return true;
 244
 245	return !get_pageblock_skip(page);
 246}
 247
 248static void reset_cached_positions(struct zone *zone)
 249{
 250	zone->compact_cached_migrate_pfn[0] = zone->zone_start_pfn;
 251	zone->compact_cached_migrate_pfn[1] = zone->zone_start_pfn;
 252	zone->compact_cached_free_pfn =
 253				pageblock_start_pfn(zone_end_pfn(zone) - 1);
 254}
 255
 256#ifdef CONFIG_SPARSEMEM
 257/*
 258 * If the PFN falls into an offline section, return the start PFN of the
 259 * next online section. If the PFN falls into an online section or if
 260 * there is no next online section, return 0.
 261 */
 262static unsigned long skip_offline_sections(unsigned long start_pfn)
 263{
 264	unsigned long start_nr = pfn_to_section_nr(start_pfn);
 265
 266	if (online_section_nr(start_nr))
 267		return 0;
 268
 269	while (++start_nr <= __highest_present_section_nr) {
 270		if (online_section_nr(start_nr))
 271			return section_nr_to_pfn(start_nr);
 272	}
 273
 274	return 0;
 275}
 276
 277/*
 278 * If the PFN falls into an offline section, return the end PFN of the
 279 * next online section in reverse. If the PFN falls into an online section
 280 * or if there is no next online section in reverse, return 0.
 281 */
 282static unsigned long skip_offline_sections_reverse(unsigned long start_pfn)
 283{
 284	unsigned long start_nr = pfn_to_section_nr(start_pfn);
 285
 286	if (!start_nr || online_section_nr(start_nr))
 287		return 0;
 288
 289	while (start_nr-- > 0) {
 290		if (online_section_nr(start_nr))
 291			return section_nr_to_pfn(start_nr) + PAGES_PER_SECTION;
 292	}
 293
 294	return 0;
 295}
 296#else
 297static unsigned long skip_offline_sections(unsigned long start_pfn)
 298{
 299	return 0;
 300}
 301
 302static unsigned long skip_offline_sections_reverse(unsigned long start_pfn)
 303{
 304	return 0;
 305}
 306#endif
 307
 308/*
 309 * Compound pages of >= pageblock_order should consistently be skipped until
 310 * released. It is always pointless to compact pages of such order (if they are
 311 * migratable), and the pageblocks they occupy cannot contain any free pages.
 312 */
 313static bool pageblock_skip_persistent(struct page *page)
 314{
 315	if (!PageCompound(page))
 316		return false;
 317
 318	page = compound_head(page);
 319
 320	if (compound_order(page) >= pageblock_order)
 321		return true;
 322
 323	return false;
 324}
 325
 326static bool
 327__reset_isolation_pfn(struct zone *zone, unsigned long pfn, bool check_source,
 328							bool check_target)
 329{
 330	struct page *page = pfn_to_online_page(pfn);
 331	struct page *block_page;
 332	struct page *end_page;
 333	unsigned long block_pfn;
 334
 335	if (!page)
 336		return false;
 337	if (zone != page_zone(page))
 338		return false;
 339	if (pageblock_skip_persistent(page))
 340		return false;
 341
 342	/*
 343	 * If skip is already cleared do no further checking once the
 344	 * restart points have been set.
 345	 */
 346	if (check_source && check_target && !get_pageblock_skip(page))
 347		return true;
 348
 349	/*
 350	 * If clearing skip for the target scanner, do not select a
 351	 * non-movable pageblock as the starting point.
 352	 */
 353	if (!check_source && check_target &&
 354	    get_pageblock_migratetype(page) != MIGRATE_MOVABLE)
 355		return false;
 356
 357	/* Ensure the start of the pageblock or zone is online and valid */
 358	block_pfn = pageblock_start_pfn(pfn);
 359	block_pfn = max(block_pfn, zone->zone_start_pfn);
 360	block_page = pfn_to_online_page(block_pfn);
 361	if (block_page) {
 362		page = block_page;
 363		pfn = block_pfn;
 364	}
 365
 366	/* Ensure the end of the pageblock or zone is online and valid */
 367	block_pfn = pageblock_end_pfn(pfn) - 1;
 368	block_pfn = min(block_pfn, zone_end_pfn(zone) - 1);
 369	end_page = pfn_to_online_page(block_pfn);
 370	if (!end_page)
 371		return false;
 372
 373	/*
 374	 * Only clear the hint if a sample indicates there is either a
 375	 * free page or an LRU page in the block. One or other condition
 376	 * is necessary for the block to be a migration source/target.
 377	 */
 378	do {
 379		if (check_source && PageLRU(page)) {
 380			clear_pageblock_skip(page);
 381			return true;
 382		}
 383
 384		if (check_target && PageBuddy(page)) {
 385			clear_pageblock_skip(page);
 386			return true;
 387		}
 388
 389		page += (1 << PAGE_ALLOC_COSTLY_ORDER);
 390	} while (page <= end_page);
 391
 392	return false;
 393}
 394
 395/*
 396 * This function is called to clear all cached information on pageblocks that
 397 * should be skipped for page isolation when the migrate and free page scanner
 398 * meet.
 399 */
 400static void __reset_isolation_suitable(struct zone *zone)
 401{
 402	unsigned long migrate_pfn = zone->zone_start_pfn;
 403	unsigned long free_pfn = zone_end_pfn(zone) - 1;
 404	unsigned long reset_migrate = free_pfn;
 405	unsigned long reset_free = migrate_pfn;
 406	bool source_set = false;
 407	bool free_set = false;
 408
 409	/* Only flush if a full compaction finished recently */
 410	if (!zone->compact_blockskip_flush)
 411		return;
 412
 413	zone->compact_blockskip_flush = false;
 414
 415	/*
 416	 * Walk the zone and update pageblock skip information. Source looks
 417	 * for PageLRU while target looks for PageBuddy. When the scanner
 418	 * is found, both PageBuddy and PageLRU are checked as the pageblock
 419	 * is suitable as both source and target.
 420	 */
 421	for (; migrate_pfn < free_pfn; migrate_pfn += pageblock_nr_pages,
 422					free_pfn -= pageblock_nr_pages) {
 423		cond_resched();
 424
 425		/* Update the migrate PFN */
 426		if (__reset_isolation_pfn(zone, migrate_pfn, true, source_set) &&
 427		    migrate_pfn < reset_migrate) {
 428			source_set = true;
 429			reset_migrate = migrate_pfn;
 430			zone->compact_init_migrate_pfn = reset_migrate;
 431			zone->compact_cached_migrate_pfn[0] = reset_migrate;
 432			zone->compact_cached_migrate_pfn[1] = reset_migrate;
 433		}
 434
 435		/* Update the free PFN */
 436		if (__reset_isolation_pfn(zone, free_pfn, free_set, true) &&
 437		    free_pfn > reset_free) {
 438			free_set = true;
 439			reset_free = free_pfn;
 440			zone->compact_init_free_pfn = reset_free;
 441			zone->compact_cached_free_pfn = reset_free;
 442		}
 443	}
 444
 445	/* Leave no distance if no suitable block was reset */
 446	if (reset_migrate >= reset_free) {
 447		zone->compact_cached_migrate_pfn[0] = migrate_pfn;
 448		zone->compact_cached_migrate_pfn[1] = migrate_pfn;
 449		zone->compact_cached_free_pfn = free_pfn;
 450	}
 451}
 452
 453void reset_isolation_suitable(pg_data_t *pgdat)
 454{
 455	int zoneid;
 456
 457	for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
 458		struct zone *zone = &pgdat->node_zones[zoneid];
 459		if (!populated_zone(zone))
 460			continue;
 461
 462		__reset_isolation_suitable(zone);
 463	}
 464}
 465
 466/*
 467 * Sets the pageblock skip bit if it was clear. Note that this is a hint as
 468 * locks are not required for read/writers. Returns true if it was already set.
 469 */
 470static bool test_and_set_skip(struct compact_control *cc, struct page *page)
 471{
 472	bool skip;
 473
 474	/* Do not update if skip hint is being ignored */
 475	if (cc->ignore_skip_hint)
 476		return false;
 477
 478	skip = get_pageblock_skip(page);
 479	if (!skip && !cc->no_set_skip_hint)
 480		set_pageblock_skip(page);
 481
 482	return skip;
 483}
 484
 485static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
 486{
 487	struct zone *zone = cc->zone;
 488
 489	/* Set for isolation rather than compaction */
 490	if (cc->no_set_skip_hint)
 491		return;
 492
 493	pfn = pageblock_end_pfn(pfn);
 494
 495	/* Update where async and sync compaction should restart */
 496	if (pfn > zone->compact_cached_migrate_pfn[0])
 497		zone->compact_cached_migrate_pfn[0] = pfn;
 498	if (cc->mode != MIGRATE_ASYNC &&
 499	    pfn > zone->compact_cached_migrate_pfn[1])
 500		zone->compact_cached_migrate_pfn[1] = pfn;
 501}
 502
 503/*
 504 * If no pages were isolated then mark this pageblock to be skipped in the
 505 * future. The information is later cleared by __reset_isolation_suitable().
 506 */
 507static void update_pageblock_skip(struct compact_control *cc,
 508			struct page *page, unsigned long pfn)
 509{
 510	struct zone *zone = cc->zone;
 511
 512	if (cc->no_set_skip_hint)
 513		return;
 514
 515	set_pageblock_skip(page);
 516
 517	if (pfn < zone->compact_cached_free_pfn)
 518		zone->compact_cached_free_pfn = pfn;
 519}
 520#else
 521static inline bool isolation_suitable(struct compact_control *cc,
 522					struct page *page)
 523{
 524	return true;
 525}
 526
 527static inline bool pageblock_skip_persistent(struct page *page)
 528{
 529	return false;
 530}
 531
 532static inline void update_pageblock_skip(struct compact_control *cc,
 533			struct page *page, unsigned long pfn)
 534{
 535}
 536
 537static void update_cached_migrate(struct compact_control *cc, unsigned long pfn)
 538{
 539}
 540
 541static bool test_and_set_skip(struct compact_control *cc, struct page *page)
 542{
 543	return false;
 544}
 545#endif /* CONFIG_COMPACTION */
 546
 547/*
 548 * Compaction requires the taking of some coarse locks that are potentially
 549 * very heavily contended. For async compaction, trylock and record if the
 550 * lock is contended. The lock will still be acquired but compaction will
 551 * abort when the current block is finished regardless of success rate.
 552 * Sync compaction acquires the lock.
 553 *
 554 * Always returns true which makes it easier to track lock state in callers.
 555 */
 556static bool compact_lock_irqsave(spinlock_t *lock, unsigned long *flags,
 557						struct compact_control *cc)
 558	__acquires(lock)
 559{
 560	/* Track if the lock is contended in async mode */
 561	if (cc->mode == MIGRATE_ASYNC && !cc->contended) {
 562		if (spin_trylock_irqsave(lock, *flags))
 563			return true;
 564
 565		cc->contended = true;
 566	}
 567
 568	spin_lock_irqsave(lock, *flags);
 569	return true;
 570}
 571
 572/*
 573 * Compaction requires the taking of some coarse locks that are potentially
 574 * very heavily contended. The lock should be periodically unlocked to avoid
 575 * having disabled IRQs for a long time, even when there is nobody waiting on
 576 * the lock. It might also be that allowing the IRQs will result in
 577 * need_resched() becoming true. If scheduling is needed, compaction schedules.
 578 * Either compaction type will also abort if a fatal signal is pending.
 579 * In either case if the lock was locked, it is dropped and not regained.
 580 *
 581 * Returns true if compaction should abort due to fatal signal pending.
 582 * Returns false when compaction can continue.
 583 */
 584static bool compact_unlock_should_abort(spinlock_t *lock,
 585		unsigned long flags, bool *locked, struct compact_control *cc)
 586{
 587	if (*locked) {
 588		spin_unlock_irqrestore(lock, flags);
 589		*locked = false;
 590	}
 591
 592	if (fatal_signal_pending(current)) {
 593		cc->contended = true;
 594		return true;
 595	}
 596
 597	cond_resched();
 598
 599	return false;
 600}
 601
 602/*
 603 * Isolate free pages onto a private freelist. If @strict is true, will abort
 604 * returning 0 on any invalid PFNs or non-free pages inside of the pageblock
 605 * (even though it may still end up isolating some pages).
 606 */
 607static unsigned long isolate_freepages_block(struct compact_control *cc,
 608				unsigned long *start_pfn,
 609				unsigned long end_pfn,
 610				struct list_head *freelist,
 611				unsigned int stride,
 612				bool strict)
 613{
 614	int nr_scanned = 0, total_isolated = 0;
 615	struct page *page;
 616	unsigned long flags = 0;
 617	bool locked = false;
 618	unsigned long blockpfn = *start_pfn;
 619	unsigned int order;
 620
 621	/* Strict mode is for isolation, speed is secondary */
 622	if (strict)
 623		stride = 1;
 624
 625	page = pfn_to_page(blockpfn);
 626
 627	/* Isolate free pages. */
 628	for (; blockpfn < end_pfn; blockpfn += stride, page += stride) {
 629		int isolated;
 630
 631		/*
 632		 * Periodically drop the lock (if held) regardless of its
 633		 * contention, to give chance to IRQs. Abort if fatal signal
 634		 * pending.
 635		 */
 636		if (!(blockpfn % COMPACT_CLUSTER_MAX)
 637		    && compact_unlock_should_abort(&cc->zone->lock, flags,
 638								&locked, cc))
 639			break;
 640
 641		nr_scanned++;
 642
 643		/*
 644		 * For compound pages such as THP and hugetlbfs, we can save
 645		 * potentially a lot of iterations if we skip them at once.
 646		 * The check is racy, but we can consider only valid values
 647		 * and the only danger is skipping too much.
 648		 */
 649		if (PageCompound(page)) {
 650			const unsigned int order = compound_order(page);
 651
 652			if (blockpfn + (1UL << order) <= end_pfn) {
 653				blockpfn += (1UL << order) - 1;
 654				page += (1UL << order) - 1;
 655				nr_scanned += (1UL << order) - 1;
 656			}
 657
 658			goto isolate_fail;
 659		}
 660
 661		if (!PageBuddy(page))
 662			goto isolate_fail;
 663
 664		/* If we already hold the lock, we can skip some rechecking. */
 665		if (!locked) {
 666			locked = compact_lock_irqsave(&cc->zone->lock,
 667								&flags, cc);
 668
 669			/* Recheck this is a buddy page under lock */
 670			if (!PageBuddy(page))
 671				goto isolate_fail;
 672		}
 673
 674		/* Found a free page, will break it into order-0 pages */
 675		order = buddy_order(page);
 676		isolated = __isolate_free_page(page, order);
 677		if (!isolated)
 678			break;
 679		set_page_private(page, order);
 680
 681		nr_scanned += isolated - 1;
 682		total_isolated += isolated;
 683		cc->nr_freepages += isolated;
 684		list_add_tail(&page->lru, &freelist[order]);
 685
 686		if (!strict && cc->nr_migratepages <= cc->nr_freepages) {
 687			blockpfn += isolated;
 688			break;
 689		}
 690		/* Advance to the end of split page */
 691		blockpfn += isolated - 1;
 692		page += isolated - 1;
 693		continue;
 694
 695isolate_fail:
 696		if (strict)
 697			break;
 698
 699	}
 700
 701	if (locked)
 702		spin_unlock_irqrestore(&cc->zone->lock, flags);
 703
 704	/*
 705	 * Be careful to not go outside of the pageblock.
 706	 */
 707	if (unlikely(blockpfn > end_pfn))
 708		blockpfn = end_pfn;
 709
 710	trace_mm_compaction_isolate_freepages(*start_pfn, blockpfn,
 711					nr_scanned, total_isolated);
 712
 713	/* Record how far we have got within the block */
 714	*start_pfn = blockpfn;
 715
 716	/*
 717	 * If strict isolation is requested by CMA then check that all the
 718	 * pages requested were isolated. If there were any failures, 0 is
 719	 * returned and CMA will fail.
 720	 */
 721	if (strict && blockpfn < end_pfn)
 722		total_isolated = 0;
 723
 724	cc->total_free_scanned += nr_scanned;
 725	if (total_isolated)
 726		count_compact_events(COMPACTISOLATED, total_isolated);
 727	return total_isolated;
 728}
 729
 730/**
 731 * isolate_freepages_range() - isolate free pages.
 732 * @cc:        Compaction control structure.
 733 * @start_pfn: The first PFN to start isolating.
 734 * @end_pfn:   The one-past-last PFN.
 735 *
 736 * Non-free pages, invalid PFNs, or zone boundaries within the
 737 * [start_pfn, end_pfn) range are considered errors, cause function to
 738 * undo its actions and return zero.
 739 *
 740 * Otherwise, function returns one-past-the-last PFN of isolated page
 741 * (which may be greater then end_pfn if end fell in a middle of
 742 * a free page).
 743 */
 744unsigned long
 745isolate_freepages_range(struct compact_control *cc,
 746			unsigned long start_pfn, unsigned long end_pfn)
 747{
 748	unsigned long isolated, pfn, block_start_pfn, block_end_pfn;
 749	int order;
 750	struct list_head tmp_freepages[NR_PAGE_ORDERS];
 751
 752	for (order = 0; order < NR_PAGE_ORDERS; order++)
 753		INIT_LIST_HEAD(&tmp_freepages[order]);
 754
 755	pfn = start_pfn;
 756	block_start_pfn = pageblock_start_pfn(pfn);
 757	if (block_start_pfn < cc->zone->zone_start_pfn)
 758		block_start_pfn = cc->zone->zone_start_pfn;
 759	block_end_pfn = pageblock_end_pfn(pfn);
 760
 761	for (; pfn < end_pfn; pfn += isolated,
 762				block_start_pfn = block_end_pfn,
 763				block_end_pfn += pageblock_nr_pages) {
 764		/* Protect pfn from changing by isolate_freepages_block */
 765		unsigned long isolate_start_pfn = pfn;
 766
 767		/*
 768		 * pfn could pass the block_end_pfn if isolated freepage
 769		 * is more than pageblock order. In this case, we adjust
 770		 * scanning range to right one.
 771		 */
 772		if (pfn >= block_end_pfn) {
 773			block_start_pfn = pageblock_start_pfn(pfn);
 774			block_end_pfn = pageblock_end_pfn(pfn);
 775		}
 776
 777		block_end_pfn = min(block_end_pfn, end_pfn);
 778
 779		if (!pageblock_pfn_to_page(block_start_pfn,
 780					block_end_pfn, cc->zone))
 781			break;
 782
 783		isolated = isolate_freepages_block(cc, &isolate_start_pfn,
 784					block_end_pfn, tmp_freepages, 0, true);
 785
 786		/*
 787		 * In strict mode, isolate_freepages_block() returns 0 if
 788		 * there are any holes in the block (ie. invalid PFNs or
 789		 * non-free pages).
 790		 */
 791		if (!isolated)
 792			break;
 793
 794		/*
 795		 * If we managed to isolate pages, it is always (1 << n) *
 796		 * pageblock_nr_pages for some non-negative n.  (Max order
 797		 * page may span two pageblocks).
 798		 */
 799	}
 800
 
 
 
 801	if (pfn < end_pfn) {
 802		/* Loop terminated early, cleanup. */
 803		release_free_list(tmp_freepages);
 804		return 0;
 805	}
 806
 807	/* __isolate_free_page() does not map the pages */
 808	split_map_pages(tmp_freepages);
 809
 810	/* We don't use freelists for anything. */
 811	return pfn;
 812}
 813
 814/* Similar to reclaim, but different enough that they don't share logic */
 815static bool too_many_isolated(struct compact_control *cc)
 816{
 817	pg_data_t *pgdat = cc->zone->zone_pgdat;
 818	bool too_many;
 819
 820	unsigned long active, inactive, isolated;
 821
 822	inactive = node_page_state(pgdat, NR_INACTIVE_FILE) +
 823			node_page_state(pgdat, NR_INACTIVE_ANON);
 824	active = node_page_state(pgdat, NR_ACTIVE_FILE) +
 825			node_page_state(pgdat, NR_ACTIVE_ANON);
 826	isolated = node_page_state(pgdat, NR_ISOLATED_FILE) +
 827			node_page_state(pgdat, NR_ISOLATED_ANON);
 828
 829	/*
 830	 * Allow GFP_NOFS to isolate past the limit set for regular
 831	 * compaction runs. This prevents an ABBA deadlock when other
 832	 * compactors have already isolated to the limit, but are
 833	 * blocked on filesystem locks held by the GFP_NOFS thread.
 834	 */
 835	if (cc->gfp_mask & __GFP_FS) {
 836		inactive >>= 3;
 837		active >>= 3;
 838	}
 839
 840	too_many = isolated > (inactive + active) / 2;
 841	if (!too_many)
 842		wake_throttle_isolated(pgdat);
 843
 844	return too_many;
 845}
 846
 847/**
 848 * skip_isolation_on_order() - determine when to skip folio isolation based on
 849 *			       folio order and compaction target order
 850 * @order:		to-be-isolated folio order
 851 * @target_order:	compaction target order
 852 *
 853 * This avoids unnecessary folio isolations during compaction.
 854 */
 855static bool skip_isolation_on_order(int order, int target_order)
 856{
 857	/*
 858	 * Unless we are performing global compaction (i.e.,
 859	 * is_via_compact_memory), skip any folios that are larger than the
 860	 * target order: we wouldn't be here if we'd have a free folio with
 861	 * the desired target_order, so migrating this folio would likely fail
 862	 * later.
 863	 */
 864	if (!is_via_compact_memory(target_order) && order >= target_order)
 865		return true;
 866	/*
 867	 * We limit memory compaction to pageblocks and won't try
 868	 * creating free blocks of memory that are larger than that.
 869	 */
 870	return order >= pageblock_order;
 871}
 872
 873/**
 874 * isolate_migratepages_block() - isolate all migrate-able pages within
 875 *				  a single pageblock
 876 * @cc:		Compaction control structure.
 877 * @low_pfn:	The first PFN to isolate
 878 * @end_pfn:	The one-past-the-last PFN to isolate, within same pageblock
 879 * @mode:	Isolation mode to be used.
 880 *
 881 * Isolate all pages that can be migrated from the range specified by
 882 * [low_pfn, end_pfn). The range is expected to be within same pageblock.
 883 * Returns errno, like -EAGAIN or -EINTR in case e.g signal pending or congestion,
 884 * -ENOMEM in case we could not allocate a page, or 0.
 885 * cc->migrate_pfn will contain the next pfn to scan.
 886 *
 887 * The pages are isolated on cc->migratepages list (not required to be empty),
 888 * and cc->nr_migratepages is updated accordingly.
 889 */
 890static int
 891isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
 892			unsigned long end_pfn, isolate_mode_t mode)
 893{
 894	pg_data_t *pgdat = cc->zone->zone_pgdat;
 895	unsigned long nr_scanned = 0, nr_isolated = 0;
 896	struct lruvec *lruvec;
 897	unsigned long flags = 0;
 898	struct lruvec *locked = NULL;
 899	struct folio *folio = NULL;
 900	struct page *page = NULL, *valid_page = NULL;
 901	struct address_space *mapping;
 902	unsigned long start_pfn = low_pfn;
 903	bool skip_on_failure = false;
 904	unsigned long next_skip_pfn = 0;
 905	bool skip_updated = false;
 906	int ret = 0;
 907
 908	cc->migrate_pfn = low_pfn;
 909
 910	/*
 911	 * Ensure that there are not too many pages isolated from the LRU
 912	 * list by either parallel reclaimers or compaction. If there are,
 913	 * delay for some time until fewer pages are isolated
 914	 */
 915	while (unlikely(too_many_isolated(cc))) {
 916		/* stop isolation if there are still pages not migrated */
 917		if (cc->nr_migratepages)
 918			return -EAGAIN;
 919
 920		/* async migration should just abort */
 921		if (cc->mode == MIGRATE_ASYNC)
 922			return -EAGAIN;
 923
 924		reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED);
 925
 926		if (fatal_signal_pending(current))
 927			return -EINTR;
 928	}
 929
 930	cond_resched();
 931
 932	if (cc->direct_compaction && (cc->mode == MIGRATE_ASYNC)) {
 933		skip_on_failure = true;
 934		next_skip_pfn = block_end_pfn(low_pfn, cc->order);
 935	}
 936
 937	/* Time to isolate some pages for migration */
 938	for (; low_pfn < end_pfn; low_pfn++) {
 939		bool is_dirty, is_unevictable;
 940
 941		if (skip_on_failure && low_pfn >= next_skip_pfn) {
 942			/*
 943			 * We have isolated all migration candidates in the
 944			 * previous order-aligned block, and did not skip it due
 945			 * to failure. We should migrate the pages now and
 946			 * hopefully succeed compaction.
 947			 */
 948			if (nr_isolated)
 949				break;
 950
 951			/*
 952			 * We failed to isolate in the previous order-aligned
 953			 * block. Set the new boundary to the end of the
 954			 * current block. Note we can't simply increase
 955			 * next_skip_pfn by 1 << order, as low_pfn might have
 956			 * been incremented by a higher number due to skipping
 957			 * a compound or a high-order buddy page in the
 958			 * previous loop iteration.
 959			 */
 960			next_skip_pfn = block_end_pfn(low_pfn, cc->order);
 961		}
 962
 963		/*
 964		 * Periodically drop the lock (if held) regardless of its
 965		 * contention, to give chance to IRQs. Abort completely if
 966		 * a fatal signal is pending.
 967		 */
 968		if (!(low_pfn % COMPACT_CLUSTER_MAX)) {
 969			if (locked) {
 970				unlock_page_lruvec_irqrestore(locked, flags);
 971				locked = NULL;
 972			}
 973
 974			if (fatal_signal_pending(current)) {
 975				cc->contended = true;
 976				ret = -EINTR;
 977
 978				goto fatal_pending;
 979			}
 980
 981			cond_resched();
 982		}
 983
 984		nr_scanned++;
 985
 986		page = pfn_to_page(low_pfn);
 987
 988		/*
 989		 * Check if the pageblock has already been marked skipped.
 990		 * Only the first PFN is checked as the caller isolates
 991		 * COMPACT_CLUSTER_MAX at a time so the second call must
 992		 * not falsely conclude that the block should be skipped.
 993		 */
 994		if (!valid_page && (pageblock_aligned(low_pfn) ||
 995				    low_pfn == cc->zone->zone_start_pfn)) {
 996			if (!isolation_suitable(cc, page)) {
 997				low_pfn = end_pfn;
 998				folio = NULL;
 999				goto isolate_abort;
1000			}
1001			valid_page = page;
1002		}
1003
1004		if (PageHuge(page)) {
1005			/*
1006			 * skip hugetlbfs if we are not compacting for pages
1007			 * bigger than its order. THPs and other compound pages
1008			 * are handled below.
1009			 */
1010			if (!cc->alloc_contig) {
1011				const unsigned int order = compound_order(page);
1012
1013				if (order <= MAX_PAGE_ORDER) {
1014					low_pfn += (1UL << order) - 1;
1015					nr_scanned += (1UL << order) - 1;
1016				}
1017				goto isolate_fail;
1018			}
1019			/* for alloc_contig case */
1020			if (locked) {
1021				unlock_page_lruvec_irqrestore(locked, flags);
1022				locked = NULL;
1023			}
1024
1025			ret = isolate_or_dissolve_huge_page(page, &cc->migratepages);
1026
1027			/*
1028			 * Fail isolation in case isolate_or_dissolve_huge_page()
1029			 * reports an error. In case of -ENOMEM, abort right away.
1030			 */
1031			if (ret < 0) {
1032				 /* Do not report -EBUSY down the chain */
1033				if (ret == -EBUSY)
1034					ret = 0;
1035				low_pfn += compound_nr(page) - 1;
1036				nr_scanned += compound_nr(page) - 1;
1037				goto isolate_fail;
1038			}
1039
1040			if (PageHuge(page)) {
1041				/*
1042				 * Hugepage was successfully isolated and placed
1043				 * on the cc->migratepages list.
1044				 */
1045				folio = page_folio(page);
1046				low_pfn += folio_nr_pages(folio) - 1;
1047				goto isolate_success_no_list;
1048			}
1049
1050			/*
1051			 * Ok, the hugepage was dissolved. Now these pages are
1052			 * Buddy and cannot be re-allocated because they are
1053			 * isolated. Fall-through as the check below handles
1054			 * Buddy pages.
1055			 */
1056		}
1057
1058		/*
1059		 * Skip if free. We read page order here without zone lock
1060		 * which is generally unsafe, but the race window is small and
1061		 * the worst thing that can happen is that we skip some
1062		 * potential isolation targets.
1063		 */
1064		if (PageBuddy(page)) {
1065			unsigned long freepage_order = buddy_order_unsafe(page);
1066
1067			/*
1068			 * Without lock, we cannot be sure that what we got is
1069			 * a valid page order. Consider only values in the
1070			 * valid order range to prevent low_pfn overflow.
1071			 */
1072			if (freepage_order > 0 && freepage_order <= MAX_PAGE_ORDER) {
1073				low_pfn += (1UL << freepage_order) - 1;
1074				nr_scanned += (1UL << freepage_order) - 1;
1075			}
1076			continue;
1077		}
1078
1079		/*
1080		 * Regardless of being on LRU, compound pages such as THP
1081		 * (hugetlbfs is handled above) are not to be compacted unless
1082		 * we are attempting an allocation larger than the compound
1083		 * page size. We can potentially save a lot of iterations if we
1084		 * skip them at once. The check is racy, but we can consider
1085		 * only valid values and the only danger is skipping too much.
1086		 */
1087		if (PageCompound(page) && !cc->alloc_contig) {
1088			const unsigned int order = compound_order(page);
1089
1090			/* Skip based on page order and compaction target order. */
1091			if (skip_isolation_on_order(order, cc->order)) {
1092				if (order <= MAX_PAGE_ORDER) {
1093					low_pfn += (1UL << order) - 1;
1094					nr_scanned += (1UL << order) - 1;
1095				}
1096				goto isolate_fail;
1097			}
 
1098		}
1099
1100		/*
1101		 * Check may be lockless but that's ok as we recheck later.
1102		 * It's possible to migrate LRU and non-lru movable pages.
1103		 * Skip any other type of page
1104		 */
1105		if (!PageLRU(page)) {
1106			/*
1107			 * __PageMovable can return false positive so we need
1108			 * to verify it under page_lock.
1109			 */
1110			if (unlikely(__PageMovable(page)) &&
1111					!PageIsolated(page)) {
1112				if (locked) {
1113					unlock_page_lruvec_irqrestore(locked, flags);
1114					locked = NULL;
1115				}
1116
1117				if (isolate_movable_page(page, mode)) {
1118					folio = page_folio(page);
1119					goto isolate_success;
1120				}
1121			}
1122
1123			goto isolate_fail;
1124		}
1125
1126		/*
1127		 * Be careful not to clear PageLRU until after we're
1128		 * sure the page is not being freed elsewhere -- the
1129		 * page release code relies on it.
1130		 */
1131		folio = folio_get_nontail_page(page);
1132		if (unlikely(!folio))
1133			goto isolate_fail;
1134
1135		/*
1136		 * Migration will fail if an anonymous page is pinned in memory,
1137		 * so avoid taking lru_lock and isolating it unnecessarily in an
1138		 * admittedly racy check.
1139		 */
1140		mapping = folio_mapping(folio);
1141		if (!mapping && (folio_ref_count(folio) - 1) > folio_mapcount(folio))
1142			goto isolate_fail_put;
1143
1144		/*
1145		 * Only allow to migrate anonymous pages in GFP_NOFS context
1146		 * because those do not depend on fs locks.
1147		 */
1148		if (!(cc->gfp_mask & __GFP_FS) && mapping)
1149			goto isolate_fail_put;
1150
1151		/* Only take pages on LRU: a check now makes later tests safe */
1152		if (!folio_test_lru(folio))
1153			goto isolate_fail_put;
1154
1155		is_unevictable = folio_test_unevictable(folio);
1156
1157		/* Compaction might skip unevictable pages but CMA takes them */
1158		if (!(mode & ISOLATE_UNEVICTABLE) && is_unevictable)
1159			goto isolate_fail_put;
1160
1161		/*
1162		 * To minimise LRU disruption, the caller can indicate with
1163		 * ISOLATE_ASYNC_MIGRATE that it only wants to isolate pages
1164		 * it will be able to migrate without blocking - clean pages
1165		 * for the most part.  PageWriteback would require blocking.
1166		 */
1167		if ((mode & ISOLATE_ASYNC_MIGRATE) && folio_test_writeback(folio))
1168			goto isolate_fail_put;
1169
1170		is_dirty = folio_test_dirty(folio);
1171
1172		if (((mode & ISOLATE_ASYNC_MIGRATE) && is_dirty) ||
1173		    (mapping && is_unevictable)) {
1174			bool migrate_dirty = true;
1175			bool is_unmovable;
1176
1177			/*
1178			 * Only folios without mappings or that have
1179			 * a ->migrate_folio callback are possible to migrate
1180			 * without blocking.
1181			 *
1182			 * Folios from unmovable mappings are not migratable.
1183			 *
1184			 * However, we can be racing with truncation, which can
1185			 * free the mapping that we need to check. Truncation
1186			 * holds the folio lock until after the folio is removed
1187			 * from the page so holding it ourselves is sufficient.
1188			 *
1189			 * To avoid locking the folio just to check unmovable,
1190			 * assume every unmovable folio is also unevictable,
1191			 * which is a cheaper test.  If our assumption goes
1192			 * wrong, it's not a correctness bug, just potentially
1193			 * wasted cycles.
1194			 */
1195			if (!folio_trylock(folio))
1196				goto isolate_fail_put;
1197
1198			mapping = folio_mapping(folio);
1199			if ((mode & ISOLATE_ASYNC_MIGRATE) && is_dirty) {
1200				migrate_dirty = !mapping ||
1201						mapping->a_ops->migrate_folio;
1202			}
1203			is_unmovable = mapping && mapping_unmovable(mapping);
1204			folio_unlock(folio);
1205			if (!migrate_dirty || is_unmovable)
1206				goto isolate_fail_put;
1207		}
1208
1209		/* Try isolate the folio */
1210		if (!folio_test_clear_lru(folio))
1211			goto isolate_fail_put;
1212
1213		lruvec = folio_lruvec(folio);
1214
1215		/* If we already hold the lock, we can skip some rechecking */
1216		if (lruvec != locked) {
1217			if (locked)
1218				unlock_page_lruvec_irqrestore(locked, flags);
1219
1220			compact_lock_irqsave(&lruvec->lru_lock, &flags, cc);
1221			locked = lruvec;
1222
1223			lruvec_memcg_debug(lruvec, folio);
1224
1225			/*
1226			 * Try get exclusive access under lock. If marked for
1227			 * skip, the scan is aborted unless the current context
1228			 * is a rescan to reach the end of the pageblock.
1229			 */
1230			if (!skip_updated && valid_page) {
1231				skip_updated = true;
1232				if (test_and_set_skip(cc, valid_page) &&
1233				    !cc->finish_pageblock) {
1234					low_pfn = end_pfn;
1235					goto isolate_abort;
1236				}
1237			}
1238
1239			/*
1240			 * Check LRU folio order under the lock
 
1241			 */
1242			if (unlikely(skip_isolation_on_order(folio_order(folio),
1243							     cc->order) &&
1244				     !cc->alloc_contig)) {
1245				low_pfn += folio_nr_pages(folio) - 1;
1246				nr_scanned += folio_nr_pages(folio) - 1;
1247				folio_set_lru(folio);
1248				goto isolate_fail_put;
1249			}
1250		}
1251
1252		/* The folio is taken off the LRU */
1253		if (folio_test_large(folio))
1254			low_pfn += folio_nr_pages(folio) - 1;
1255
1256		/* Successfully isolated */
1257		lruvec_del_folio(lruvec, folio);
1258		node_stat_mod_folio(folio,
1259				NR_ISOLATED_ANON + folio_is_file_lru(folio),
1260				folio_nr_pages(folio));
1261
1262isolate_success:
1263		list_add(&folio->lru, &cc->migratepages);
1264isolate_success_no_list:
1265		cc->nr_migratepages += folio_nr_pages(folio);
1266		nr_isolated += folio_nr_pages(folio);
1267		nr_scanned += folio_nr_pages(folio) - 1;
1268
1269		/*
1270		 * Avoid isolating too much unless this block is being
1271		 * fully scanned (e.g. dirty/writeback pages, parallel allocation)
1272		 * or a lock is contended. For contention, isolate quickly to
1273		 * potentially remove one source of contention.
1274		 */
1275		if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX &&
1276		    !cc->finish_pageblock && !cc->contended) {
1277			++low_pfn;
1278			break;
1279		}
1280
1281		continue;
1282
1283isolate_fail_put:
1284		/* Avoid potential deadlock in freeing page under lru_lock */
1285		if (locked) {
1286			unlock_page_lruvec_irqrestore(locked, flags);
1287			locked = NULL;
1288		}
1289		folio_put(folio);
1290
1291isolate_fail:
1292		if (!skip_on_failure && ret != -ENOMEM)
1293			continue;
1294
1295		/*
1296		 * We have isolated some pages, but then failed. Release them
1297		 * instead of migrating, as we cannot form the cc->order buddy
1298		 * page anyway.
1299		 */
1300		if (nr_isolated) {
1301			if (locked) {
1302				unlock_page_lruvec_irqrestore(locked, flags);
1303				locked = NULL;
1304			}
1305			putback_movable_pages(&cc->migratepages);
1306			cc->nr_migratepages = 0;
1307			nr_isolated = 0;
1308		}
1309
1310		if (low_pfn < next_skip_pfn) {
1311			low_pfn = next_skip_pfn - 1;
1312			/*
1313			 * The check near the loop beginning would have updated
1314			 * next_skip_pfn too, but this is a bit simpler.
1315			 */
1316			next_skip_pfn += 1UL << cc->order;
1317		}
1318
1319		if (ret == -ENOMEM)
1320			break;
1321	}
1322
1323	/*
1324	 * The PageBuddy() check could have potentially brought us outside
1325	 * the range to be scanned.
1326	 */
1327	if (unlikely(low_pfn > end_pfn))
1328		low_pfn = end_pfn;
1329
1330	folio = NULL;
1331
1332isolate_abort:
1333	if (locked)
1334		unlock_page_lruvec_irqrestore(locked, flags);
1335	if (folio) {
1336		folio_set_lru(folio);
1337		folio_put(folio);
1338	}
1339
1340	/*
1341	 * Update the cached scanner pfn once the pageblock has been scanned.
1342	 * Pages will either be migrated in which case there is no point
1343	 * scanning in the near future or migration failed in which case the
1344	 * failure reason may persist. The block is marked for skipping if
1345	 * there were no pages isolated in the block or if the block is
1346	 * rescanned twice in a row.
1347	 */
1348	if (low_pfn == end_pfn && (!nr_isolated || cc->finish_pageblock)) {
1349		if (!cc->no_set_skip_hint && valid_page && !skip_updated)
1350			set_pageblock_skip(valid_page);
1351		update_cached_migrate(cc, low_pfn);
1352	}
1353
1354	trace_mm_compaction_isolate_migratepages(start_pfn, low_pfn,
1355						nr_scanned, nr_isolated);
1356
1357fatal_pending:
1358	cc->total_migrate_scanned += nr_scanned;
1359	if (nr_isolated)
1360		count_compact_events(COMPACTISOLATED, nr_isolated);
1361
1362	cc->migrate_pfn = low_pfn;
1363
1364	return ret;
1365}
1366
1367/**
1368 * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
1369 * @cc:        Compaction control structure.
1370 * @start_pfn: The first PFN to start isolating.
1371 * @end_pfn:   The one-past-last PFN.
1372 *
1373 * Returns -EAGAIN when contented, -EINTR in case of a signal pending, -ENOMEM
1374 * in case we could not allocate a page, or 0.
1375 */
1376int
1377isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
1378							unsigned long end_pfn)
1379{
1380	unsigned long pfn, block_start_pfn, block_end_pfn;
1381	int ret = 0;
1382
1383	/* Scan block by block. First and last block may be incomplete */
1384	pfn = start_pfn;
1385	block_start_pfn = pageblock_start_pfn(pfn);
1386	if (block_start_pfn < cc->zone->zone_start_pfn)
1387		block_start_pfn = cc->zone->zone_start_pfn;
1388	block_end_pfn = pageblock_end_pfn(pfn);
1389
1390	for (; pfn < end_pfn; pfn = block_end_pfn,
1391				block_start_pfn = block_end_pfn,
1392				block_end_pfn += pageblock_nr_pages) {
1393
1394		block_end_pfn = min(block_end_pfn, end_pfn);
1395
1396		if (!pageblock_pfn_to_page(block_start_pfn,
1397					block_end_pfn, cc->zone))
1398			continue;
1399
1400		ret = isolate_migratepages_block(cc, pfn, block_end_pfn,
1401						 ISOLATE_UNEVICTABLE);
1402
1403		if (ret)
1404			break;
1405
1406		if (cc->nr_migratepages >= COMPACT_CLUSTER_MAX)
1407			break;
1408	}
1409
1410	return ret;
1411}
1412
1413#endif /* CONFIG_COMPACTION || CONFIG_CMA */
1414#ifdef CONFIG_COMPACTION
1415
1416static bool suitable_migration_source(struct compact_control *cc,
1417							struct page *page)
1418{
1419	int block_mt;
1420
1421	if (pageblock_skip_persistent(page))
1422		return false;
1423
1424	if ((cc->mode != MIGRATE_ASYNC) || !cc->direct_compaction)
1425		return true;
1426
1427	block_mt = get_pageblock_migratetype(page);
1428
1429	if (cc->migratetype == MIGRATE_MOVABLE)
1430		return is_migrate_movable(block_mt);
1431	else
1432		return block_mt == cc->migratetype;
1433}
1434
1435/* Returns true if the page is within a block suitable for migration to */
1436static bool suitable_migration_target(struct compact_control *cc,
1437							struct page *page)
1438{
1439	/* If the page is a large free page, then disallow migration */
1440	if (PageBuddy(page)) {
1441		int order = cc->order > 0 ? cc->order : pageblock_order;
1442
1443		/*
1444		 * We are checking page_order without zone->lock taken. But
1445		 * the only small danger is that we skip a potentially suitable
1446		 * pageblock, so it's not worth to check order for valid range.
1447		 */
1448		if (buddy_order_unsafe(page) >= order)
1449			return false;
1450	}
1451
1452	if (cc->ignore_block_suitable)
1453		return true;
1454
1455	/* If the block is MIGRATE_MOVABLE or MIGRATE_CMA, allow migration */
1456	if (is_migrate_movable(get_pageblock_migratetype(page)))
1457		return true;
1458
1459	/* Otherwise skip the block */
1460	return false;
1461}
1462
1463static inline unsigned int
1464freelist_scan_limit(struct compact_control *cc)
1465{
1466	unsigned short shift = BITS_PER_LONG - 1;
1467
1468	return (COMPACT_CLUSTER_MAX >> min(shift, cc->fast_search_fail)) + 1;
1469}
1470
1471/*
1472 * Test whether the free scanner has reached the same or lower pageblock than
1473 * the migration scanner, and compaction should thus terminate.
1474 */
1475static inline bool compact_scanners_met(struct compact_control *cc)
1476{
1477	return (cc->free_pfn >> pageblock_order)
1478		<= (cc->migrate_pfn >> pageblock_order);
1479}
1480
1481/*
1482 * Used when scanning for a suitable migration target which scans freelists
1483 * in reverse. Reorders the list such as the unscanned pages are scanned
1484 * first on the next iteration of the free scanner
1485 */
1486static void
1487move_freelist_head(struct list_head *freelist, struct page *freepage)
1488{
1489	LIST_HEAD(sublist);
1490
1491	if (!list_is_first(&freepage->buddy_list, freelist)) {
1492		list_cut_before(&sublist, freelist, &freepage->buddy_list);
1493		list_splice_tail(&sublist, freelist);
1494	}
1495}
1496
1497/*
1498 * Similar to move_freelist_head except used by the migration scanner
1499 * when scanning forward. It's possible for these list operations to
1500 * move against each other if they search the free list exactly in
1501 * lockstep.
1502 */
1503static void
1504move_freelist_tail(struct list_head *freelist, struct page *freepage)
1505{
1506	LIST_HEAD(sublist);
1507
1508	if (!list_is_last(&freepage->buddy_list, freelist)) {
1509		list_cut_position(&sublist, freelist, &freepage->buddy_list);
1510		list_splice_tail(&sublist, freelist);
1511	}
1512}
1513
1514static void
1515fast_isolate_around(struct compact_control *cc, unsigned long pfn)
1516{
1517	unsigned long start_pfn, end_pfn;
1518	struct page *page;
1519
1520	/* Do not search around if there are enough pages already */
1521	if (cc->nr_freepages >= cc->nr_migratepages)
1522		return;
1523
1524	/* Minimise scanning during async compaction */
1525	if (cc->direct_compaction && cc->mode == MIGRATE_ASYNC)
1526		return;
1527
1528	/* Pageblock boundaries */
1529	start_pfn = max(pageblock_start_pfn(pfn), cc->zone->zone_start_pfn);
1530	end_pfn = min(pageblock_end_pfn(pfn), zone_end_pfn(cc->zone));
1531
1532	page = pageblock_pfn_to_page(start_pfn, end_pfn, cc->zone);
1533	if (!page)
1534		return;
1535
1536	isolate_freepages_block(cc, &start_pfn, end_pfn, cc->freepages, 1, false);
1537
1538	/* Skip this pageblock in the future as it's full or nearly full */
1539	if (start_pfn == end_pfn && !cc->no_set_skip_hint)
1540		set_pageblock_skip(page);
1541}
1542
1543/* Search orders in round-robin fashion */
1544static int next_search_order(struct compact_control *cc, int order)
1545{
1546	order--;
1547	if (order < 0)
1548		order = cc->order - 1;
1549
1550	/* Search wrapped around? */
1551	if (order == cc->search_order) {
1552		cc->search_order--;
1553		if (cc->search_order < 0)
1554			cc->search_order = cc->order - 1;
1555		return -1;
1556	}
1557
1558	return order;
1559}
1560
1561static void fast_isolate_freepages(struct compact_control *cc)
1562{
1563	unsigned int limit = max(1U, freelist_scan_limit(cc) >> 1);
1564	unsigned int nr_scanned = 0, total_isolated = 0;
1565	unsigned long low_pfn, min_pfn, highest = 0;
1566	unsigned long nr_isolated = 0;
1567	unsigned long distance;
1568	struct page *page = NULL;
1569	bool scan_start = false;
1570	int order;
1571
1572	/* Full compaction passes in a negative order */
1573	if (cc->order <= 0)
1574		return;
1575
1576	/*
1577	 * If starting the scan, use a deeper search and use the highest
1578	 * PFN found if a suitable one is not found.
1579	 */
1580	if (cc->free_pfn >= cc->zone->compact_init_free_pfn) {
1581		limit = pageblock_nr_pages >> 1;
1582		scan_start = true;
1583	}
1584
1585	/*
1586	 * Preferred point is in the top quarter of the scan space but take
1587	 * a pfn from the top half if the search is problematic.
1588	 */
1589	distance = (cc->free_pfn - cc->migrate_pfn);
1590	low_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 2));
1591	min_pfn = pageblock_start_pfn(cc->free_pfn - (distance >> 1));
1592
1593	if (WARN_ON_ONCE(min_pfn > low_pfn))
1594		low_pfn = min_pfn;
1595
1596	/*
1597	 * Search starts from the last successful isolation order or the next
1598	 * order to search after a previous failure
1599	 */
1600	cc->search_order = min_t(unsigned int, cc->order - 1, cc->search_order);
1601
1602	for (order = cc->search_order;
1603	     !page && order >= 0;
1604	     order = next_search_order(cc, order)) {
1605		struct free_area *area = &cc->zone->free_area[order];
1606		struct list_head *freelist;
1607		struct page *freepage;
1608		unsigned long flags;
1609		unsigned int order_scanned = 0;
1610		unsigned long high_pfn = 0;
1611
1612		if (!area->nr_free)
1613			continue;
1614
1615		spin_lock_irqsave(&cc->zone->lock, flags);
1616		freelist = &area->free_list[MIGRATE_MOVABLE];
1617		list_for_each_entry_reverse(freepage, freelist, buddy_list) {
1618			unsigned long pfn;
1619
1620			order_scanned++;
1621			nr_scanned++;
1622			pfn = page_to_pfn(freepage);
1623
1624			if (pfn >= highest)
1625				highest = max(pageblock_start_pfn(pfn),
1626					      cc->zone->zone_start_pfn);
1627
1628			if (pfn >= low_pfn) {
1629				cc->fast_search_fail = 0;
1630				cc->search_order = order;
1631				page = freepage;
1632				break;
1633			}
1634
1635			if (pfn >= min_pfn && pfn > high_pfn) {
1636				high_pfn = pfn;
1637
1638				/* Shorten the scan if a candidate is found */
1639				limit >>= 1;
1640			}
1641
1642			if (order_scanned >= limit)
1643				break;
1644		}
1645
1646		/* Use a maximum candidate pfn if a preferred one was not found */
1647		if (!page && high_pfn) {
1648			page = pfn_to_page(high_pfn);
1649
1650			/* Update freepage for the list reorder below */
1651			freepage = page;
1652		}
1653
1654		/* Reorder to so a future search skips recent pages */
1655		move_freelist_head(freelist, freepage);
1656
1657		/* Isolate the page if available */
1658		if (page) {
1659			if (__isolate_free_page(page, order)) {
1660				set_page_private(page, order);
1661				nr_isolated = 1 << order;
1662				nr_scanned += nr_isolated - 1;
1663				total_isolated += nr_isolated;
1664				cc->nr_freepages += nr_isolated;
1665				list_add_tail(&page->lru, &cc->freepages[order]);
1666				count_compact_events(COMPACTISOLATED, nr_isolated);
1667			} else {
1668				/* If isolation fails, abort the search */
1669				order = cc->search_order + 1;
1670				page = NULL;
1671			}
1672		}
1673
1674		spin_unlock_irqrestore(&cc->zone->lock, flags);
1675
1676		/* Skip fast search if enough freepages isolated */
1677		if (cc->nr_freepages >= cc->nr_migratepages)
1678			break;
1679
1680		/*
1681		 * Smaller scan on next order so the total scan is related
1682		 * to freelist_scan_limit.
1683		 */
1684		if (order_scanned >= limit)
1685			limit = max(1U, limit >> 1);
1686	}
1687
1688	trace_mm_compaction_fast_isolate_freepages(min_pfn, cc->free_pfn,
1689						   nr_scanned, total_isolated);
1690
1691	if (!page) {
1692		cc->fast_search_fail++;
1693		if (scan_start) {
1694			/*
1695			 * Use the highest PFN found above min. If one was
1696			 * not found, be pessimistic for direct compaction
1697			 * and use the min mark.
1698			 */
1699			if (highest >= min_pfn) {
1700				page = pfn_to_page(highest);
1701				cc->free_pfn = highest;
1702			} else {
1703				if (cc->direct_compaction && pfn_valid(min_pfn)) {
1704					page = pageblock_pfn_to_page(min_pfn,
1705						min(pageblock_end_pfn(min_pfn),
1706						    zone_end_pfn(cc->zone)),
1707						cc->zone);
1708					if (page && !suitable_migration_target(cc, page))
1709						page = NULL;
1710
1711					cc->free_pfn = min_pfn;
1712				}
1713			}
1714		}
1715	}
1716
1717	if (highest && highest >= cc->zone->compact_cached_free_pfn) {
1718		highest -= pageblock_nr_pages;
1719		cc->zone->compact_cached_free_pfn = highest;
1720	}
1721
1722	cc->total_free_scanned += nr_scanned;
1723	if (!page)
1724		return;
1725
1726	low_pfn = page_to_pfn(page);
1727	fast_isolate_around(cc, low_pfn);
1728}
1729
1730/*
1731 * Based on information in the current compact_control, find blocks
1732 * suitable for isolating free pages from and then isolate them.
1733 */
1734static void isolate_freepages(struct compact_control *cc)
1735{
1736	struct zone *zone = cc->zone;
1737	struct page *page;
1738	unsigned long block_start_pfn;	/* start of current pageblock */
1739	unsigned long isolate_start_pfn; /* exact pfn we start at */
1740	unsigned long block_end_pfn;	/* end of current pageblock */
1741	unsigned long low_pfn;	     /* lowest pfn scanner is able to scan */
 
1742	unsigned int stride;
1743
1744	/* Try a small search of the free lists for a candidate */
1745	fast_isolate_freepages(cc);
1746	if (cc->nr_freepages)
1747		return;
1748
1749	/*
1750	 * Initialise the free scanner. The starting point is where we last
1751	 * successfully isolated from, zone-cached value, or the end of the
1752	 * zone when isolating for the first time. For looping we also need
1753	 * this pfn aligned down to the pageblock boundary, because we do
1754	 * block_start_pfn -= pageblock_nr_pages in the for loop.
1755	 * For ending point, take care when isolating in last pageblock of a
1756	 * zone which ends in the middle of a pageblock.
1757	 * The low boundary is the end of the pageblock the migration scanner
1758	 * is using.
1759	 */
1760	isolate_start_pfn = cc->free_pfn;
1761	block_start_pfn = pageblock_start_pfn(isolate_start_pfn);
1762	block_end_pfn = min(block_start_pfn + pageblock_nr_pages,
1763						zone_end_pfn(zone));
1764	low_pfn = pageblock_end_pfn(cc->migrate_pfn);
1765	stride = cc->mode == MIGRATE_ASYNC ? COMPACT_CLUSTER_MAX : 1;
1766
1767	/*
1768	 * Isolate free pages until enough are available to migrate the
1769	 * pages on cc->migratepages. We stop searching if the migrate
1770	 * and free page scanners meet or enough free pages are isolated.
1771	 */
1772	for (; block_start_pfn >= low_pfn;
1773				block_end_pfn = block_start_pfn,
1774				block_start_pfn -= pageblock_nr_pages,
1775				isolate_start_pfn = block_start_pfn) {
1776		unsigned long nr_isolated;
1777
1778		/*
1779		 * This can iterate a massively long zone without finding any
1780		 * suitable migration targets, so periodically check resched.
1781		 */
1782		if (!(block_start_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages)))
1783			cond_resched();
1784
1785		page = pageblock_pfn_to_page(block_start_pfn, block_end_pfn,
1786									zone);
1787		if (!page) {
1788			unsigned long next_pfn;
1789
1790			next_pfn = skip_offline_sections_reverse(block_start_pfn);
1791			if (next_pfn)
1792				block_start_pfn = max(next_pfn, low_pfn);
1793
1794			continue;
1795		}
1796
1797		/* Check the block is suitable for migration */
1798		if (!suitable_migration_target(cc, page))
1799			continue;
1800
1801		/* If isolation recently failed, do not retry */
1802		if (!isolation_suitable(cc, page))
1803			continue;
1804
1805		/* Found a block suitable for isolating free pages from. */
1806		nr_isolated = isolate_freepages_block(cc, &isolate_start_pfn,
1807					block_end_pfn, cc->freepages, stride, false);
1808
1809		/* Update the skip hint if the full pageblock was scanned */
1810		if (isolate_start_pfn == block_end_pfn)
1811			update_pageblock_skip(cc, page, block_start_pfn -
1812					      pageblock_nr_pages);
1813
1814		/* Are enough freepages isolated? */
1815		if (cc->nr_freepages >= cc->nr_migratepages) {
1816			if (isolate_start_pfn >= block_end_pfn) {
1817				/*
1818				 * Restart at previous pageblock if more
1819				 * freepages can be isolated next time.
1820				 */
1821				isolate_start_pfn =
1822					block_start_pfn - pageblock_nr_pages;
1823			}
1824			break;
1825		} else if (isolate_start_pfn < block_end_pfn) {
1826			/*
1827			 * If isolation failed early, do not continue
1828			 * needlessly.
1829			 */
1830			break;
1831		}
1832
1833		/* Adjust stride depending on isolation */
1834		if (nr_isolated) {
1835			stride = 1;
1836			continue;
1837		}
1838		stride = min_t(unsigned int, COMPACT_CLUSTER_MAX, stride << 1);
1839	}
1840
1841	/*
1842	 * Record where the free scanner will restart next time. Either we
1843	 * broke from the loop and set isolate_start_pfn based on the last
1844	 * call to isolate_freepages_block(), or we met the migration scanner
1845	 * and the loop terminated due to isolate_start_pfn < low_pfn
1846	 */
1847	cc->free_pfn = isolate_start_pfn;
 
 
 
 
1848}
1849
1850/*
1851 * This is a migrate-callback that "allocates" freepages by taking pages
1852 * from the isolated freelists in the block we are migrating to.
1853 */
1854static struct folio *compaction_alloc(struct folio *src, unsigned long data)
1855{
1856	struct compact_control *cc = (struct compact_control *)data;
1857	struct folio *dst;
1858	int order = folio_order(src);
1859	bool has_isolated_pages = false;
1860	int start_order;
1861	struct page *freepage;
1862	unsigned long size;
1863
1864again:
1865	for (start_order = order; start_order < NR_PAGE_ORDERS; start_order++)
1866		if (!list_empty(&cc->freepages[start_order]))
1867			break;
1868
1869	/* no free pages in the list */
1870	if (start_order == NR_PAGE_ORDERS) {
1871		if (has_isolated_pages)
 
1872			return NULL;
1873		isolate_freepages(cc);
1874		has_isolated_pages = true;
1875		goto again;
1876	}
1877
1878	freepage = list_first_entry(&cc->freepages[start_order], struct page,
1879				lru);
1880	size = 1 << start_order;
1881
1882	list_del(&freepage->lru);
1883
1884	while (start_order > order) {
1885		start_order--;
1886		size >>= 1;
1887
1888		list_add(&freepage[size].lru, &cc->freepages[start_order]);
1889		set_page_private(&freepage[size], start_order);
1890	}
1891	dst = (struct folio *)freepage;
1892
1893	post_alloc_hook(&dst->page, order, __GFP_MOVABLE);
1894	if (order)
1895		prep_compound_page(&dst->page, order);
1896	cc->nr_freepages -= 1 << order;
1897	cc->nr_migratepages -= 1 << order;
1898	return page_rmappable_folio(&dst->page);
1899}
1900
1901/*
1902 * This is a migrate-callback that "frees" freepages back to the isolated
1903 * freelist.  All pages on the freelist are from the same zone, so there is no
1904 * special handling needed for NUMA.
1905 */
1906static void compaction_free(struct folio *dst, unsigned long data)
1907{
1908	struct compact_control *cc = (struct compact_control *)data;
1909	int order = folio_order(dst);
1910	struct page *page = &dst->page;
1911
1912	if (folio_put_testzero(dst)) {
1913		free_pages_prepare(page, order);
1914		list_add(&dst->lru, &cc->freepages[order]);
1915		cc->nr_freepages += 1 << order;
1916	}
1917	cc->nr_migratepages += 1 << order;
1918	/*
1919	 * someone else has referenced the page, we cannot take it back to our
1920	 * free list.
1921	 */
1922}
1923
1924/* possible outcome of isolate_migratepages */
1925typedef enum {
1926	ISOLATE_ABORT,		/* Abort compaction now */
1927	ISOLATE_NONE,		/* No pages isolated, continue scanning */
1928	ISOLATE_SUCCESS,	/* Pages isolated, migrate */
1929} isolate_migrate_t;
1930
1931/*
1932 * Allow userspace to control policy on scanning the unevictable LRU for
1933 * compactable pages.
1934 */
1935static int sysctl_compact_unevictable_allowed __read_mostly = CONFIG_COMPACT_UNEVICTABLE_DEFAULT;
1936/*
1937 * Tunable for proactive compaction. It determines how
1938 * aggressively the kernel should compact memory in the
1939 * background. It takes values in the range [0, 100].
1940 */
1941static unsigned int __read_mostly sysctl_compaction_proactiveness = 20;
1942static int sysctl_extfrag_threshold = 500;
1943static int __read_mostly sysctl_compact_memory;
1944
1945static inline void
1946update_fast_start_pfn(struct compact_control *cc, unsigned long pfn)
1947{
1948	if (cc->fast_start_pfn == ULONG_MAX)
1949		return;
1950
1951	if (!cc->fast_start_pfn)
1952		cc->fast_start_pfn = pfn;
1953
1954	cc->fast_start_pfn = min(cc->fast_start_pfn, pfn);
1955}
1956
1957static inline unsigned long
1958reinit_migrate_pfn(struct compact_control *cc)
1959{
1960	if (!cc->fast_start_pfn || cc->fast_start_pfn == ULONG_MAX)
1961		return cc->migrate_pfn;
1962
1963	cc->migrate_pfn = cc->fast_start_pfn;
1964	cc->fast_start_pfn = ULONG_MAX;
1965
1966	return cc->migrate_pfn;
1967}
1968
1969/*
1970 * Briefly search the free lists for a migration source that already has
1971 * some free pages to reduce the number of pages that need migration
1972 * before a pageblock is free.
1973 */
1974static unsigned long fast_find_migrateblock(struct compact_control *cc)
1975{
1976	unsigned int limit = freelist_scan_limit(cc);
1977	unsigned int nr_scanned = 0;
1978	unsigned long distance;
1979	unsigned long pfn = cc->migrate_pfn;
1980	unsigned long high_pfn;
1981	int order;
1982	bool found_block = false;
1983
1984	/* Skip hints are relied on to avoid repeats on the fast search */
1985	if (cc->ignore_skip_hint)
1986		return pfn;
1987
1988	/*
1989	 * If the pageblock should be finished then do not select a different
1990	 * pageblock.
1991	 */
1992	if (cc->finish_pageblock)
1993		return pfn;
1994
1995	/*
1996	 * If the migrate_pfn is not at the start of a zone or the start
1997	 * of a pageblock then assume this is a continuation of a previous
1998	 * scan restarted due to COMPACT_CLUSTER_MAX.
1999	 */
2000	if (pfn != cc->zone->zone_start_pfn && pfn != pageblock_start_pfn(pfn))
2001		return pfn;
2002
2003	/*
2004	 * For smaller orders, just linearly scan as the number of pages
2005	 * to migrate should be relatively small and does not necessarily
2006	 * justify freeing up a large block for a small allocation.
2007	 */
2008	if (cc->order <= PAGE_ALLOC_COSTLY_ORDER)
2009		return pfn;
2010
2011	/*
2012	 * Only allow kcompactd and direct requests for movable pages to
2013	 * quickly clear out a MOVABLE pageblock for allocation. This
2014	 * reduces the risk that a large movable pageblock is freed for
2015	 * an unmovable/reclaimable small allocation.
2016	 */
2017	if (cc->direct_compaction && cc->migratetype != MIGRATE_MOVABLE)
2018		return pfn;
2019
2020	/*
2021	 * When starting the migration scanner, pick any pageblock within the
2022	 * first half of the search space. Otherwise try and pick a pageblock
2023	 * within the first eighth to reduce the chances that a migration
2024	 * target later becomes a source.
2025	 */
2026	distance = (cc->free_pfn - cc->migrate_pfn) >> 1;
2027	if (cc->migrate_pfn != cc->zone->zone_start_pfn)
2028		distance >>= 2;
2029	high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance);
2030
2031	for (order = cc->order - 1;
2032	     order >= PAGE_ALLOC_COSTLY_ORDER && !found_block && nr_scanned < limit;
2033	     order--) {
2034		struct free_area *area = &cc->zone->free_area[order];
2035		struct list_head *freelist;
2036		unsigned long flags;
2037		struct page *freepage;
2038
2039		if (!area->nr_free)
2040			continue;
2041
2042		spin_lock_irqsave(&cc->zone->lock, flags);
2043		freelist = &area->free_list[MIGRATE_MOVABLE];
2044		list_for_each_entry(freepage, freelist, buddy_list) {
2045			unsigned long free_pfn;
2046
2047			if (nr_scanned++ >= limit) {
2048				move_freelist_tail(freelist, freepage);
2049				break;
2050			}
2051
2052			free_pfn = page_to_pfn(freepage);
2053			if (free_pfn < high_pfn) {
2054				/*
2055				 * Avoid if skipped recently. Ideally it would
2056				 * move to the tail but even safe iteration of
2057				 * the list assumes an entry is deleted, not
2058				 * reordered.
2059				 */
2060				if (get_pageblock_skip(freepage))
2061					continue;
2062
2063				/* Reorder to so a future search skips recent pages */
2064				move_freelist_tail(freelist, freepage);
2065
2066				update_fast_start_pfn(cc, free_pfn);
2067				pfn = pageblock_start_pfn(free_pfn);
2068				if (pfn < cc->zone->zone_start_pfn)
2069					pfn = cc->zone->zone_start_pfn;
2070				cc->fast_search_fail = 0;
2071				found_block = true;
2072				break;
2073			}
2074		}
2075		spin_unlock_irqrestore(&cc->zone->lock, flags);
2076	}
2077
2078	cc->total_migrate_scanned += nr_scanned;
2079
2080	/*
2081	 * If fast scanning failed then use a cached entry for a page block
2082	 * that had free pages as the basis for starting a linear scan.
2083	 */
2084	if (!found_block) {
2085		cc->fast_search_fail++;
2086		pfn = reinit_migrate_pfn(cc);
2087	}
2088	return pfn;
2089}
2090
2091/*
2092 * Isolate all pages that can be migrated from the first suitable block,
2093 * starting at the block pointed to by the migrate scanner pfn within
2094 * compact_control.
2095 */
2096static isolate_migrate_t isolate_migratepages(struct compact_control *cc)
2097{
2098	unsigned long block_start_pfn;
2099	unsigned long block_end_pfn;
2100	unsigned long low_pfn;
2101	struct page *page;
2102	const isolate_mode_t isolate_mode =
2103		(sysctl_compact_unevictable_allowed ? ISOLATE_UNEVICTABLE : 0) |
2104		(cc->mode != MIGRATE_SYNC ? ISOLATE_ASYNC_MIGRATE : 0);
2105	bool fast_find_block;
2106
2107	/*
2108	 * Start at where we last stopped, or beginning of the zone as
2109	 * initialized by compact_zone(). The first failure will use
2110	 * the lowest PFN as the starting point for linear scanning.
2111	 */
2112	low_pfn = fast_find_migrateblock(cc);
2113	block_start_pfn = pageblock_start_pfn(low_pfn);
2114	if (block_start_pfn < cc->zone->zone_start_pfn)
2115		block_start_pfn = cc->zone->zone_start_pfn;
2116
2117	/*
2118	 * fast_find_migrateblock() has already ensured the pageblock is not
2119	 * set with a skipped flag, so to avoid the isolation_suitable check
2120	 * below again, check whether the fast search was successful.
2121	 */
2122	fast_find_block = low_pfn != cc->migrate_pfn && !cc->fast_search_fail;
2123
2124	/* Only scan within a pageblock boundary */
2125	block_end_pfn = pageblock_end_pfn(low_pfn);
2126
2127	/*
2128	 * Iterate over whole pageblocks until we find the first suitable.
2129	 * Do not cross the free scanner.
2130	 */
2131	for (; block_end_pfn <= cc->free_pfn;
2132			fast_find_block = false,
2133			cc->migrate_pfn = low_pfn = block_end_pfn,
2134			block_start_pfn = block_end_pfn,
2135			block_end_pfn += pageblock_nr_pages) {
2136
2137		/*
2138		 * This can potentially iterate a massively long zone with
2139		 * many pageblocks unsuitable, so periodically check if we
2140		 * need to schedule.
2141		 */
2142		if (!(low_pfn % (COMPACT_CLUSTER_MAX * pageblock_nr_pages)))
2143			cond_resched();
2144
2145		page = pageblock_pfn_to_page(block_start_pfn,
2146						block_end_pfn, cc->zone);
2147		if (!page) {
2148			unsigned long next_pfn;
2149
2150			next_pfn = skip_offline_sections(block_start_pfn);
2151			if (next_pfn)
2152				block_end_pfn = min(next_pfn, cc->free_pfn);
2153			continue;
2154		}
2155
2156		/*
2157		 * If isolation recently failed, do not retry. Only check the
2158		 * pageblock once. COMPACT_CLUSTER_MAX causes a pageblock
2159		 * to be visited multiple times. Assume skip was checked
2160		 * before making it "skip" so other compaction instances do
2161		 * not scan the same block.
2162		 */
2163		if ((pageblock_aligned(low_pfn) ||
2164		     low_pfn == cc->zone->zone_start_pfn) &&
2165		    !fast_find_block && !isolation_suitable(cc, page))
2166			continue;
2167
2168		/*
2169		 * For async direct compaction, only scan the pageblocks of the
2170		 * same migratetype without huge pages. Async direct compaction
2171		 * is optimistic to see if the minimum amount of work satisfies
2172		 * the allocation. The cached PFN is updated as it's possible
2173		 * that all remaining blocks between source and target are
2174		 * unsuitable and the compaction scanners fail to meet.
2175		 */
2176		if (!suitable_migration_source(cc, page)) {
2177			update_cached_migrate(cc, block_end_pfn);
2178			continue;
2179		}
2180
2181		/* Perform the isolation */
2182		if (isolate_migratepages_block(cc, low_pfn, block_end_pfn,
2183						isolate_mode))
2184			return ISOLATE_ABORT;
2185
2186		/*
2187		 * Either we isolated something and proceed with migration. Or
2188		 * we failed and compact_zone should decide if we should
2189		 * continue or not.
2190		 */
2191		break;
2192	}
2193
2194	return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
2195}
2196
2197/*
 
 
 
 
 
 
 
 
 
 
 
2198 * Determine whether kswapd is (or recently was!) running on this node.
2199 *
2200 * pgdat_kswapd_lock() pins pgdat->kswapd, so a concurrent kswapd_stop() can't
2201 * zero it.
2202 */
2203static bool kswapd_is_running(pg_data_t *pgdat)
2204{
2205	bool running;
2206
2207	pgdat_kswapd_lock(pgdat);
2208	running = pgdat->kswapd && task_is_running(pgdat->kswapd);
2209	pgdat_kswapd_unlock(pgdat);
2210
2211	return running;
2212}
2213
2214/*
2215 * A zone's fragmentation score is the external fragmentation wrt to the
2216 * COMPACTION_HPAGE_ORDER. It returns a value in the range [0, 100].
2217 */
2218static unsigned int fragmentation_score_zone(struct zone *zone)
2219{
2220	return extfrag_for_order(zone, COMPACTION_HPAGE_ORDER);
2221}
2222
2223/*
2224 * A weighted zone's fragmentation score is the external fragmentation
2225 * wrt to the COMPACTION_HPAGE_ORDER scaled by the zone's size. It
2226 * returns a value in the range [0, 100].
2227 *
2228 * The scaling factor ensures that proactive compaction focuses on larger
2229 * zones like ZONE_NORMAL, rather than smaller, specialized zones like
2230 * ZONE_DMA32. For smaller zones, the score value remains close to zero,
2231 * and thus never exceeds the high threshold for proactive compaction.
2232 */
2233static unsigned int fragmentation_score_zone_weighted(struct zone *zone)
2234{
2235	unsigned long score;
2236
2237	score = zone->present_pages * fragmentation_score_zone(zone);
2238	return div64_ul(score, zone->zone_pgdat->node_present_pages + 1);
2239}
2240
2241/*
2242 * The per-node proactive (background) compaction process is started by its
2243 * corresponding kcompactd thread when the node's fragmentation score
2244 * exceeds the high threshold. The compaction process remains active till
2245 * the node's score falls below the low threshold, or one of the back-off
2246 * conditions is met.
2247 */
2248static unsigned int fragmentation_score_node(pg_data_t *pgdat)
2249{
2250	unsigned int score = 0;
2251	int zoneid;
2252
2253	for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2254		struct zone *zone;
2255
2256		zone = &pgdat->node_zones[zoneid];
2257		if (!populated_zone(zone))
2258			continue;
2259		score += fragmentation_score_zone_weighted(zone);
2260	}
2261
2262	return score;
2263}
2264
2265static unsigned int fragmentation_score_wmark(bool low)
2266{
2267	unsigned int wmark_low;
2268
2269	/*
2270	 * Cap the low watermark to avoid excessive compaction
2271	 * activity in case a user sets the proactiveness tunable
2272	 * close to 100 (maximum).
2273	 */
2274	wmark_low = max(100U - sysctl_compaction_proactiveness, 5U);
2275	return low ? wmark_low : min(wmark_low + 10, 100U);
2276}
2277
2278static bool should_proactive_compact_node(pg_data_t *pgdat)
2279{
2280	int wmark_high;
2281
2282	if (!sysctl_compaction_proactiveness || kswapd_is_running(pgdat))
2283		return false;
2284
2285	wmark_high = fragmentation_score_wmark(false);
2286	return fragmentation_score_node(pgdat) > wmark_high;
2287}
2288
2289static enum compact_result __compact_finished(struct compact_control *cc)
2290{
2291	unsigned int order;
2292	const int migratetype = cc->migratetype;
2293	int ret;
2294
2295	/* Compaction run completes if the migrate and free scanner meet */
2296	if (compact_scanners_met(cc)) {
2297		/* Let the next compaction start anew. */
2298		reset_cached_positions(cc->zone);
2299
2300		/*
2301		 * Mark that the PG_migrate_skip information should be cleared
2302		 * by kswapd when it goes to sleep. kcompactd does not set the
2303		 * flag itself as the decision to be clear should be directly
2304		 * based on an allocation request.
2305		 */
2306		if (cc->direct_compaction)
2307			cc->zone->compact_blockskip_flush = true;
2308
2309		if (cc->whole_zone)
2310			return COMPACT_COMPLETE;
2311		else
2312			return COMPACT_PARTIAL_SKIPPED;
2313	}
2314
2315	if (cc->proactive_compaction) {
2316		int score, wmark_low;
2317		pg_data_t *pgdat;
2318
2319		pgdat = cc->zone->zone_pgdat;
2320		if (kswapd_is_running(pgdat))
2321			return COMPACT_PARTIAL_SKIPPED;
2322
2323		score = fragmentation_score_zone(cc->zone);
2324		wmark_low = fragmentation_score_wmark(true);
2325
2326		if (score > wmark_low)
2327			ret = COMPACT_CONTINUE;
2328		else
2329			ret = COMPACT_SUCCESS;
2330
2331		goto out;
2332	}
2333
2334	if (is_via_compact_memory(cc->order))
2335		return COMPACT_CONTINUE;
2336
2337	/*
2338	 * Always finish scanning a pageblock to reduce the possibility of
2339	 * fallbacks in the future. This is particularly important when
2340	 * migration source is unmovable/reclaimable but it's not worth
2341	 * special casing.
2342	 */
2343	if (!pageblock_aligned(cc->migrate_pfn))
2344		return COMPACT_CONTINUE;
2345
2346	/* Direct compactor: Is a suitable page free? */
2347	ret = COMPACT_NO_SUITABLE_PAGE;
2348	for (order = cc->order; order < NR_PAGE_ORDERS; order++) {
2349		struct free_area *area = &cc->zone->free_area[order];
2350		bool can_steal;
2351
2352		/* Job done if page is free of the right migratetype */
2353		if (!free_area_empty(area, migratetype))
2354			return COMPACT_SUCCESS;
2355
2356#ifdef CONFIG_CMA
2357		/* MIGRATE_MOVABLE can fallback on MIGRATE_CMA */
2358		if (migratetype == MIGRATE_MOVABLE &&
2359			!free_area_empty(area, MIGRATE_CMA))
2360			return COMPACT_SUCCESS;
2361#endif
2362		/*
2363		 * Job done if allocation would steal freepages from
2364		 * other migratetype buddy lists.
2365		 */
2366		if (find_suitable_fallback(area, order, migratetype,
2367						true, &can_steal) != -1)
2368			/*
2369			 * Movable pages are OK in any pageblock. If we are
2370			 * stealing for a non-movable allocation, make sure
2371			 * we finish compacting the current pageblock first
2372			 * (which is assured by the above migrate_pfn align
2373			 * check) so it is as free as possible and we won't
2374			 * have to steal another one soon.
2375			 */
2376			return COMPACT_SUCCESS;
2377	}
2378
2379out:
2380	if (cc->contended || fatal_signal_pending(current))
2381		ret = COMPACT_CONTENDED;
2382
2383	return ret;
2384}
2385
2386static enum compact_result compact_finished(struct compact_control *cc)
2387{
2388	int ret;
2389
2390	ret = __compact_finished(cc);
2391	trace_mm_compaction_finished(cc->zone, cc->order, ret);
2392	if (ret == COMPACT_NO_SUITABLE_PAGE)
2393		ret = COMPACT_CONTINUE;
2394
2395	return ret;
2396}
2397
2398static bool __compaction_suitable(struct zone *zone, int order,
2399				  int highest_zoneidx,
2400				  unsigned long wmark_target)
2401{
2402	unsigned long watermark;
2403	/*
2404	 * Watermarks for order-0 must be met for compaction to be able to
2405	 * isolate free pages for migration targets. This means that the
2406	 * watermark and alloc_flags have to match, or be more pessimistic than
2407	 * the check in __isolate_free_page(). We don't use the direct
2408	 * compactor's alloc_flags, as they are not relevant for freepage
2409	 * isolation. We however do use the direct compactor's highest_zoneidx
2410	 * to skip over zones where lowmem reserves would prevent allocation
2411	 * even if compaction succeeds.
2412	 * For costly orders, we require low watermark instead of min for
2413	 * compaction to proceed to increase its chances.
2414	 * ALLOC_CMA is used, as pages in CMA pageblocks are considered
2415	 * suitable migration targets
2416	 */
2417	watermark = (order > PAGE_ALLOC_COSTLY_ORDER) ?
2418				low_wmark_pages(zone) : min_wmark_pages(zone);
2419	watermark += compact_gap(order);
2420	return __zone_watermark_ok(zone, 0, watermark, highest_zoneidx,
2421				   ALLOC_CMA, wmark_target);
2422}
2423
2424/*
2425 * compaction_suitable: Is this suitable to run compaction on this zone now?
2426 */
2427bool compaction_suitable(struct zone *zone, int order, int highest_zoneidx)
2428{
2429	enum compact_result compact_result;
2430	bool suitable;
2431
2432	suitable = __compaction_suitable(zone, order, highest_zoneidx,
2433					 zone_page_state(zone, NR_FREE_PAGES));
2434	/*
2435	 * fragmentation index determines if allocation failures are due to
2436	 * low memory or external fragmentation
2437	 *
2438	 * index of -1000 would imply allocations might succeed depending on
2439	 * watermarks, but we already failed the high-order watermark check
2440	 * index towards 0 implies failure is due to lack of memory
2441	 * index towards 1000 implies failure is due to fragmentation
2442	 *
2443	 * Only compact if a failure would be due to fragmentation. Also
2444	 * ignore fragindex for non-costly orders where the alternative to
2445	 * a successful reclaim/compaction is OOM. Fragindex and the
2446	 * vm.extfrag_threshold sysctl is meant as a heuristic to prevent
2447	 * excessive compaction for costly orders, but it should not be at the
2448	 * expense of system stability.
2449	 */
2450	if (suitable) {
2451		compact_result = COMPACT_CONTINUE;
2452		if (order > PAGE_ALLOC_COSTLY_ORDER) {
2453			int fragindex = fragmentation_index(zone, order);
2454
2455			if (fragindex >= 0 &&
2456			    fragindex <= sysctl_extfrag_threshold) {
2457				suitable = false;
2458				compact_result = COMPACT_NOT_SUITABLE_ZONE;
2459			}
2460		}
2461	} else {
2462		compact_result = COMPACT_SKIPPED;
2463	}
2464
2465	trace_mm_compaction_suitable(zone, order, compact_result);
2466
2467	return suitable;
2468}
2469
2470bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
2471		int alloc_flags)
2472{
2473	struct zone *zone;
2474	struct zoneref *z;
2475
2476	/*
2477	 * Make sure at least one zone would pass __compaction_suitable if we continue
2478	 * retrying the reclaim.
2479	 */
2480	for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2481				ac->highest_zoneidx, ac->nodemask) {
2482		unsigned long available;
2483
2484		/*
2485		 * Do not consider all the reclaimable memory because we do not
2486		 * want to trash just for a single high order allocation which
2487		 * is even not guaranteed to appear even if __compaction_suitable
2488		 * is happy about the watermark check.
2489		 */
2490		available = zone_reclaimable_pages(zone) / order;
2491		available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
2492		if (__compaction_suitable(zone, order, ac->highest_zoneidx,
2493					  available))
2494			return true;
2495	}
2496
2497	return false;
2498}
2499
2500/*
2501 * Should we do compaction for target allocation order.
2502 * Return COMPACT_SUCCESS if allocation for target order can be already
2503 * satisfied
2504 * Return COMPACT_SKIPPED if compaction for target order is likely to fail
2505 * Return COMPACT_CONTINUE if compaction for target order should be ran
2506 */
2507static enum compact_result
2508compaction_suit_allocation_order(struct zone *zone, unsigned int order,
2509				 int highest_zoneidx, unsigned int alloc_flags)
2510{
2511	unsigned long watermark;
2512
2513	watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
2514	if (zone_watermark_ok(zone, order, watermark, highest_zoneidx,
2515			      alloc_flags))
2516		return COMPACT_SUCCESS;
2517
2518	if (!compaction_suitable(zone, order, highest_zoneidx))
2519		return COMPACT_SKIPPED;
2520
2521	return COMPACT_CONTINUE;
2522}
2523
2524static enum compact_result
2525compact_zone(struct compact_control *cc, struct capture_control *capc)
2526{
2527	enum compact_result ret;
2528	unsigned long start_pfn = cc->zone->zone_start_pfn;
2529	unsigned long end_pfn = zone_end_pfn(cc->zone);
2530	unsigned long last_migrated_pfn;
2531	const bool sync = cc->mode != MIGRATE_ASYNC;
2532	bool update_cached;
2533	unsigned int nr_succeeded = 0, nr_migratepages;
2534	int order;
2535
2536	/*
2537	 * These counters track activities during zone compaction.  Initialize
2538	 * them before compacting a new zone.
2539	 */
2540	cc->total_migrate_scanned = 0;
2541	cc->total_free_scanned = 0;
2542	cc->nr_migratepages = 0;
2543	cc->nr_freepages = 0;
2544	for (order = 0; order < NR_PAGE_ORDERS; order++)
2545		INIT_LIST_HEAD(&cc->freepages[order]);
2546	INIT_LIST_HEAD(&cc->migratepages);
2547
2548	cc->migratetype = gfp_migratetype(cc->gfp_mask);
2549
2550	if (!is_via_compact_memory(cc->order)) {
2551		ret = compaction_suit_allocation_order(cc->zone, cc->order,
2552						       cc->highest_zoneidx,
2553						       cc->alloc_flags);
2554		if (ret != COMPACT_CONTINUE)
2555			return ret;
2556	}
2557
2558	/*
2559	 * Clear pageblock skip if there were failures recently and compaction
2560	 * is about to be retried after being deferred.
2561	 */
2562	if (compaction_restarting(cc->zone, cc->order))
2563		__reset_isolation_suitable(cc->zone);
2564
2565	/*
2566	 * Setup to move all movable pages to the end of the zone. Used cached
2567	 * information on where the scanners should start (unless we explicitly
2568	 * want to compact the whole zone), but check that it is initialised
2569	 * by ensuring the values are within zone boundaries.
2570	 */
2571	cc->fast_start_pfn = 0;
2572	if (cc->whole_zone) {
2573		cc->migrate_pfn = start_pfn;
2574		cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2575	} else {
2576		cc->migrate_pfn = cc->zone->compact_cached_migrate_pfn[sync];
2577		cc->free_pfn = cc->zone->compact_cached_free_pfn;
2578		if (cc->free_pfn < start_pfn || cc->free_pfn >= end_pfn) {
2579			cc->free_pfn = pageblock_start_pfn(end_pfn - 1);
2580			cc->zone->compact_cached_free_pfn = cc->free_pfn;
2581		}
2582		if (cc->migrate_pfn < start_pfn || cc->migrate_pfn >= end_pfn) {
2583			cc->migrate_pfn = start_pfn;
2584			cc->zone->compact_cached_migrate_pfn[0] = cc->migrate_pfn;
2585			cc->zone->compact_cached_migrate_pfn[1] = cc->migrate_pfn;
2586		}
2587
2588		if (cc->migrate_pfn <= cc->zone->compact_init_migrate_pfn)
2589			cc->whole_zone = true;
2590	}
2591
2592	last_migrated_pfn = 0;
2593
2594	/*
2595	 * Migrate has separate cached PFNs for ASYNC and SYNC* migration on
2596	 * the basis that some migrations will fail in ASYNC mode. However,
2597	 * if the cached PFNs match and pageblocks are skipped due to having
2598	 * no isolation candidates, then the sync state does not matter.
2599	 * Until a pageblock with isolation candidates is found, keep the
2600	 * cached PFNs in sync to avoid revisiting the same blocks.
2601	 */
2602	update_cached = !sync &&
2603		cc->zone->compact_cached_migrate_pfn[0] == cc->zone->compact_cached_migrate_pfn[1];
2604
2605	trace_mm_compaction_begin(cc, start_pfn, end_pfn, sync);
2606
2607	/* lru_add_drain_all could be expensive with involving other CPUs */
2608	lru_add_drain();
2609
2610	while ((ret = compact_finished(cc)) == COMPACT_CONTINUE) {
2611		int err;
2612		unsigned long iteration_start_pfn = cc->migrate_pfn;
2613
2614		/*
2615		 * Avoid multiple rescans of the same pageblock which can
2616		 * happen if a page cannot be isolated (dirty/writeback in
2617		 * async mode) or if the migrated pages are being allocated
2618		 * before the pageblock is cleared.  The first rescan will
2619		 * capture the entire pageblock for migration. If it fails,
2620		 * it'll be marked skip and scanning will proceed as normal.
2621		 */
2622		cc->finish_pageblock = false;
2623		if (pageblock_start_pfn(last_migrated_pfn) ==
2624		    pageblock_start_pfn(iteration_start_pfn)) {
2625			cc->finish_pageblock = true;
2626		}
2627
2628rescan:
2629		switch (isolate_migratepages(cc)) {
2630		case ISOLATE_ABORT:
2631			ret = COMPACT_CONTENDED;
2632			putback_movable_pages(&cc->migratepages);
2633			cc->nr_migratepages = 0;
2634			goto out;
2635		case ISOLATE_NONE:
2636			if (update_cached) {
2637				cc->zone->compact_cached_migrate_pfn[1] =
2638					cc->zone->compact_cached_migrate_pfn[0];
2639			}
2640
2641			/*
2642			 * We haven't isolated and migrated anything, but
2643			 * there might still be unflushed migrations from
2644			 * previous cc->order aligned block.
2645			 */
2646			goto check_drain;
2647		case ISOLATE_SUCCESS:
2648			update_cached = false;
2649			last_migrated_pfn = max(cc->zone->zone_start_pfn,
2650				pageblock_start_pfn(cc->migrate_pfn - 1));
2651		}
2652
2653		/*
2654		 * Record the number of pages to migrate since the
2655		 * compaction_alloc/free() will update cc->nr_migratepages
2656		 * properly.
2657		 */
2658		nr_migratepages = cc->nr_migratepages;
2659		err = migrate_pages(&cc->migratepages, compaction_alloc,
2660				compaction_free, (unsigned long)cc, cc->mode,
2661				MR_COMPACTION, &nr_succeeded);
2662
2663		trace_mm_compaction_migratepages(nr_migratepages, nr_succeeded);
2664
2665		/* All pages were either migrated or will be released */
2666		cc->nr_migratepages = 0;
2667		if (err) {
2668			putback_movable_pages(&cc->migratepages);
2669			/*
2670			 * migrate_pages() may return -ENOMEM when scanners meet
2671			 * and we want compact_finished() to detect it
2672			 */
2673			if (err == -ENOMEM && !compact_scanners_met(cc)) {
2674				ret = COMPACT_CONTENDED;
2675				goto out;
2676			}
2677			/*
2678			 * If an ASYNC or SYNC_LIGHT fails to migrate a page
2679			 * within the pageblock_order-aligned block and
2680			 * fast_find_migrateblock may be used then scan the
2681			 * remainder of the pageblock. This will mark the
2682			 * pageblock "skip" to avoid rescanning in the near
2683			 * future. This will isolate more pages than necessary
2684			 * for the request but avoid loops due to
2685			 * fast_find_migrateblock revisiting blocks that were
2686			 * recently partially scanned.
2687			 */
2688			if (!pageblock_aligned(cc->migrate_pfn) &&
2689			    !cc->ignore_skip_hint && !cc->finish_pageblock &&
2690			    (cc->mode < MIGRATE_SYNC)) {
2691				cc->finish_pageblock = true;
2692
2693				/*
2694				 * Draining pcplists does not help THP if
2695				 * any page failed to migrate. Even after
2696				 * drain, the pageblock will not be free.
2697				 */
2698				if (cc->order == COMPACTION_HPAGE_ORDER)
2699					last_migrated_pfn = 0;
2700
2701				goto rescan;
2702			}
2703		}
2704
2705		/* Stop if a page has been captured */
2706		if (capc && capc->page) {
2707			ret = COMPACT_SUCCESS;
2708			break;
2709		}
2710
2711check_drain:
2712		/*
2713		 * Has the migration scanner moved away from the previous
2714		 * cc->order aligned block where we migrated from? If yes,
2715		 * flush the pages that were freed, so that they can merge and
2716		 * compact_finished() can detect immediately if allocation
2717		 * would succeed.
2718		 */
2719		if (cc->order > 0 && last_migrated_pfn) {
2720			unsigned long current_block_start =
2721				block_start_pfn(cc->migrate_pfn, cc->order);
2722
2723			if (last_migrated_pfn < current_block_start) {
2724				lru_add_drain_cpu_zone(cc->zone);
2725				/* No more flushing until we migrate again */
2726				last_migrated_pfn = 0;
2727			}
2728		}
2729	}
2730
2731out:
2732	/*
2733	 * Release free pages and update where the free scanner should restart,
2734	 * so we don't leave any returned pages behind in the next attempt.
2735	 */
2736	if (cc->nr_freepages > 0) {
2737		unsigned long free_pfn = release_free_list(cc->freepages);
2738
2739		cc->nr_freepages = 0;
2740		VM_BUG_ON(free_pfn == 0);
2741		/* The cached pfn is always the first in a pageblock */
2742		free_pfn = pageblock_start_pfn(free_pfn);
2743		/*
2744		 * Only go back, not forward. The cached pfn might have been
2745		 * already reset to zone end in compact_finished()
2746		 */
2747		if (free_pfn > cc->zone->compact_cached_free_pfn)
2748			cc->zone->compact_cached_free_pfn = free_pfn;
2749	}
2750
2751	count_compact_events(COMPACTMIGRATE_SCANNED, cc->total_migrate_scanned);
2752	count_compact_events(COMPACTFREE_SCANNED, cc->total_free_scanned);
2753
2754	trace_mm_compaction_end(cc, start_pfn, end_pfn, sync, ret);
2755
 
2756	VM_BUG_ON(!list_empty(&cc->migratepages));
2757
2758	return ret;
2759}
2760
2761static enum compact_result compact_zone_order(struct zone *zone, int order,
2762		gfp_t gfp_mask, enum compact_priority prio,
2763		unsigned int alloc_flags, int highest_zoneidx,
2764		struct page **capture)
2765{
2766	enum compact_result ret;
2767	struct compact_control cc = {
2768		.order = order,
2769		.search_order = order,
2770		.gfp_mask = gfp_mask,
2771		.zone = zone,
2772		.mode = (prio == COMPACT_PRIO_ASYNC) ?
2773					MIGRATE_ASYNC :	MIGRATE_SYNC_LIGHT,
2774		.alloc_flags = alloc_flags,
2775		.highest_zoneidx = highest_zoneidx,
2776		.direct_compaction = true,
2777		.whole_zone = (prio == MIN_COMPACT_PRIORITY),
2778		.ignore_skip_hint = (prio == MIN_COMPACT_PRIORITY),
2779		.ignore_block_suitable = (prio == MIN_COMPACT_PRIORITY)
2780	};
2781	struct capture_control capc = {
2782		.cc = &cc,
2783		.page = NULL,
2784	};
2785
2786	/*
2787	 * Make sure the structs are really initialized before we expose the
2788	 * capture control, in case we are interrupted and the interrupt handler
2789	 * frees a page.
2790	 */
2791	barrier();
2792	WRITE_ONCE(current->capture_control, &capc);
2793
2794	ret = compact_zone(&cc, &capc);
2795
2796	/*
2797	 * Make sure we hide capture control first before we read the captured
2798	 * page pointer, otherwise an interrupt could free and capture a page
2799	 * and we would leak it.
2800	 */
2801	WRITE_ONCE(current->capture_control, NULL);
2802	*capture = READ_ONCE(capc.page);
2803	/*
2804	 * Technically, it is also possible that compaction is skipped but
2805	 * the page is still captured out of luck(IRQ came and freed the page).
2806	 * Returning COMPACT_SUCCESS in such cases helps in properly accounting
2807	 * the COMPACT[STALL|FAIL] when compaction is skipped.
2808	 */
2809	if (*capture)
2810		ret = COMPACT_SUCCESS;
2811
2812	return ret;
2813}
2814
2815/**
2816 * try_to_compact_pages - Direct compact to satisfy a high-order allocation
2817 * @gfp_mask: The GFP mask of the current allocation
2818 * @order: The order of the current allocation
2819 * @alloc_flags: The allocation flags of the current allocation
2820 * @ac: The context of current allocation
2821 * @prio: Determines how hard direct compaction should try to succeed
2822 * @capture: Pointer to free page created by compaction will be stored here
2823 *
2824 * This is the main entry point for direct page compaction.
2825 */
2826enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
2827		unsigned int alloc_flags, const struct alloc_context *ac,
2828		enum compact_priority prio, struct page **capture)
2829{
2830	struct zoneref *z;
2831	struct zone *zone;
2832	enum compact_result rc = COMPACT_SKIPPED;
2833
2834	if (!gfp_compaction_allowed(gfp_mask))
2835		return COMPACT_SKIPPED;
2836
2837	trace_mm_compaction_try_to_compact_pages(order, gfp_mask, prio);
2838
2839	/* Compact each zone in the list */
2840	for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2841					ac->highest_zoneidx, ac->nodemask) {
2842		enum compact_result status;
2843
2844		if (prio > MIN_COMPACT_PRIORITY
2845					&& compaction_deferred(zone, order)) {
2846			rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
2847			continue;
2848		}
2849
2850		status = compact_zone_order(zone, order, gfp_mask, prio,
2851				alloc_flags, ac->highest_zoneidx, capture);
2852		rc = max(status, rc);
2853
2854		/* The allocation should succeed, stop compacting */
2855		if (status == COMPACT_SUCCESS) {
2856			/*
2857			 * We think the allocation will succeed in this zone,
2858			 * but it is not certain, hence the false. The caller
2859			 * will repeat this with true if allocation indeed
2860			 * succeeds in this zone.
2861			 */
2862			compaction_defer_reset(zone, order, false);
2863
2864			break;
2865		}
2866
2867		if (prio != COMPACT_PRIO_ASYNC && (status == COMPACT_COMPLETE ||
2868					status == COMPACT_PARTIAL_SKIPPED))
2869			/*
2870			 * We think that allocation won't succeed in this zone
2871			 * so we defer compaction there. If it ends up
2872			 * succeeding after all, it will be reset.
2873			 */
2874			defer_compaction(zone, order);
2875
2876		/*
2877		 * We might have stopped compacting due to need_resched() in
2878		 * async compaction, or due to a fatal signal detected. In that
2879		 * case do not try further zones
2880		 */
2881		if ((prio == COMPACT_PRIO_ASYNC && need_resched())
2882					|| fatal_signal_pending(current))
2883			break;
2884	}
2885
2886	return rc;
2887}
2888
2889/*
2890 * compact_node() - compact all zones within a node
2891 * @pgdat: The node page data
2892 * @proactive: Whether the compaction is proactive
2893 *
2894 * For proactive compaction, compact till each zone's fragmentation score
2895 * reaches within proactive compaction thresholds (as determined by the
2896 * proactiveness tunable), it is possible that the function returns before
2897 * reaching score targets due to various back-off conditions, such as,
2898 * contention on per-node or per-zone locks.
2899 */
2900static int compact_node(pg_data_t *pgdat, bool proactive)
2901{
2902	int zoneid;
2903	struct zone *zone;
2904	struct compact_control cc = {
2905		.order = -1,
2906		.mode = proactive ? MIGRATE_SYNC_LIGHT : MIGRATE_SYNC,
2907		.ignore_skip_hint = true,
2908		.whole_zone = true,
2909		.gfp_mask = GFP_KERNEL,
2910		.proactive_compaction = proactive,
2911	};
2912
2913	for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
2914		zone = &pgdat->node_zones[zoneid];
2915		if (!populated_zone(zone))
2916			continue;
2917
2918		if (fatal_signal_pending(current))
2919			return -EINTR;
2920
2921		cc.zone = zone;
2922
2923		compact_zone(&cc, NULL);
2924
2925		if (proactive) {
2926			count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
2927					     cc.total_migrate_scanned);
2928			count_compact_events(KCOMPACTD_FREE_SCANNED,
2929					     cc.total_free_scanned);
2930		}
2931	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2932
2933	return 0;
 
 
 
 
 
 
 
 
 
 
2934}
2935
2936/* Compact all zones of all nodes in the system */
2937static int compact_nodes(void)
2938{
2939	int ret, nid;
2940
2941	/* Flush pending updates to the LRU lists */
2942	lru_add_drain_all();
2943
2944	for_each_online_node(nid) {
2945		ret = compact_node(NODE_DATA(nid), false);
2946		if (ret)
2947			return ret;
2948	}
2949
2950	return 0;
2951}
2952
2953static int compaction_proactiveness_sysctl_handler(struct ctl_table *table, int write,
2954		void *buffer, size_t *length, loff_t *ppos)
2955{
2956	int rc, nid;
2957
2958	rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
2959	if (rc)
2960		return rc;
2961
2962	if (write && sysctl_compaction_proactiveness) {
2963		for_each_online_node(nid) {
2964			pg_data_t *pgdat = NODE_DATA(nid);
2965
2966			if (pgdat->proactive_compact_trigger)
2967				continue;
2968
2969			pgdat->proactive_compact_trigger = true;
2970			trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, -1,
2971							     pgdat->nr_zones - 1);
2972			wake_up_interruptible(&pgdat->kcompactd_wait);
2973		}
2974	}
2975
2976	return 0;
2977}
2978
2979/*
2980 * This is the entry point for compacting all nodes via
2981 * /proc/sys/vm/compact_memory
2982 */
2983static int sysctl_compaction_handler(struct ctl_table *table, int write,
2984			void *buffer, size_t *length, loff_t *ppos)
2985{
2986	int ret;
2987
2988	ret = proc_dointvec(table, write, buffer, length, ppos);
2989	if (ret)
2990		return ret;
2991
2992	if (sysctl_compact_memory != 1)
2993		return -EINVAL;
2994
2995	if (write)
2996		ret = compact_nodes();
2997
2998	return ret;
2999}
3000
3001#if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
3002static ssize_t compact_store(struct device *dev,
3003			     struct device_attribute *attr,
3004			     const char *buf, size_t count)
3005{
3006	int nid = dev->id;
3007
3008	if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
3009		/* Flush pending updates to the LRU lists */
3010		lru_add_drain_all();
3011
3012		compact_node(NODE_DATA(nid), false);
3013	}
3014
3015	return count;
3016}
3017static DEVICE_ATTR_WO(compact);
3018
3019int compaction_register_node(struct node *node)
3020{
3021	return device_create_file(&node->dev, &dev_attr_compact);
3022}
3023
3024void compaction_unregister_node(struct node *node)
3025{
3026	device_remove_file(&node->dev, &dev_attr_compact);
3027}
3028#endif /* CONFIG_SYSFS && CONFIG_NUMA */
3029
3030static inline bool kcompactd_work_requested(pg_data_t *pgdat)
3031{
3032	return pgdat->kcompactd_max_order > 0 || kthread_should_stop() ||
3033		pgdat->proactive_compact_trigger;
3034}
3035
3036static bool kcompactd_node_suitable(pg_data_t *pgdat)
3037{
3038	int zoneid;
3039	struct zone *zone;
3040	enum zone_type highest_zoneidx = pgdat->kcompactd_highest_zoneidx;
3041	enum compact_result ret;
3042
3043	for (zoneid = 0; zoneid <= highest_zoneidx; zoneid++) {
3044		zone = &pgdat->node_zones[zoneid];
3045
3046		if (!populated_zone(zone))
3047			continue;
3048
3049		ret = compaction_suit_allocation_order(zone,
3050				pgdat->kcompactd_max_order,
3051				highest_zoneidx, ALLOC_WMARK_MIN);
3052		if (ret == COMPACT_CONTINUE)
3053			return true;
3054	}
3055
3056	return false;
3057}
3058
3059static void kcompactd_do_work(pg_data_t *pgdat)
3060{
3061	/*
3062	 * With no special task, compact all zones so that a page of requested
3063	 * order is allocatable.
3064	 */
3065	int zoneid;
3066	struct zone *zone;
3067	struct compact_control cc = {
3068		.order = pgdat->kcompactd_max_order,
3069		.search_order = pgdat->kcompactd_max_order,
3070		.highest_zoneidx = pgdat->kcompactd_highest_zoneidx,
3071		.mode = MIGRATE_SYNC_LIGHT,
3072		.ignore_skip_hint = false,
3073		.gfp_mask = GFP_KERNEL,
3074	};
3075	enum compact_result ret;
3076
3077	trace_mm_compaction_kcompactd_wake(pgdat->node_id, cc.order,
3078							cc.highest_zoneidx);
3079	count_compact_event(KCOMPACTD_WAKE);
3080
3081	for (zoneid = 0; zoneid <= cc.highest_zoneidx; zoneid++) {
3082		int status;
3083
3084		zone = &pgdat->node_zones[zoneid];
3085		if (!populated_zone(zone))
3086			continue;
3087
3088		if (compaction_deferred(zone, cc.order))
3089			continue;
3090
3091		ret = compaction_suit_allocation_order(zone,
3092				cc.order, zoneid, ALLOC_WMARK_MIN);
3093		if (ret != COMPACT_CONTINUE)
3094			continue;
3095
3096		if (kthread_should_stop())
3097			return;
3098
3099		cc.zone = zone;
3100		status = compact_zone(&cc, NULL);
3101
3102		if (status == COMPACT_SUCCESS) {
3103			compaction_defer_reset(zone, cc.order, false);
3104		} else if (status == COMPACT_PARTIAL_SKIPPED || status == COMPACT_COMPLETE) {
3105			/*
3106			 * Buddy pages may become stranded on pcps that could
3107			 * otherwise coalesce on the zone's free area for
3108			 * order >= cc.order.  This is ratelimited by the
3109			 * upcoming deferral.
3110			 */
3111			drain_all_pages(zone);
3112
3113			/*
3114			 * We use sync migration mode here, so we defer like
3115			 * sync direct compaction does.
3116			 */
3117			defer_compaction(zone, cc.order);
3118		}
3119
3120		count_compact_events(KCOMPACTD_MIGRATE_SCANNED,
3121				     cc.total_migrate_scanned);
3122		count_compact_events(KCOMPACTD_FREE_SCANNED,
3123				     cc.total_free_scanned);
3124	}
3125
3126	/*
3127	 * Regardless of success, we are done until woken up next. But remember
3128	 * the requested order/highest_zoneidx in case it was higher/tighter
3129	 * than our current ones
3130	 */
3131	if (pgdat->kcompactd_max_order <= cc.order)
3132		pgdat->kcompactd_max_order = 0;
3133	if (pgdat->kcompactd_highest_zoneidx >= cc.highest_zoneidx)
3134		pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
3135}
3136
3137void wakeup_kcompactd(pg_data_t *pgdat, int order, int highest_zoneidx)
3138{
3139	if (!order)
3140		return;
3141
3142	if (pgdat->kcompactd_max_order < order)
3143		pgdat->kcompactd_max_order = order;
3144
3145	if (pgdat->kcompactd_highest_zoneidx > highest_zoneidx)
3146		pgdat->kcompactd_highest_zoneidx = highest_zoneidx;
3147
3148	/*
3149	 * Pairs with implicit barrier in wait_event_freezable()
3150	 * such that wakeups are not missed.
3151	 */
3152	if (!wq_has_sleeper(&pgdat->kcompactd_wait))
3153		return;
3154
3155	if (!kcompactd_node_suitable(pgdat))
3156		return;
3157
3158	trace_mm_compaction_wakeup_kcompactd(pgdat->node_id, order,
3159							highest_zoneidx);
3160	wake_up_interruptible(&pgdat->kcompactd_wait);
3161}
3162
3163/*
3164 * The background compaction daemon, started as a kernel thread
3165 * from the init process.
3166 */
3167static int kcompactd(void *p)
3168{
3169	pg_data_t *pgdat = (pg_data_t *)p;
3170	struct task_struct *tsk = current;
3171	long default_timeout = msecs_to_jiffies(HPAGE_FRAG_CHECK_INTERVAL_MSEC);
3172	long timeout = default_timeout;
3173
3174	const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
3175
3176	if (!cpumask_empty(cpumask))
3177		set_cpus_allowed_ptr(tsk, cpumask);
3178
3179	set_freezable();
3180
3181	pgdat->kcompactd_max_order = 0;
3182	pgdat->kcompactd_highest_zoneidx = pgdat->nr_zones - 1;
3183
3184	while (!kthread_should_stop()) {
3185		unsigned long pflags;
3186
3187		/*
3188		 * Avoid the unnecessary wakeup for proactive compaction
3189		 * when it is disabled.
3190		 */
3191		if (!sysctl_compaction_proactiveness)
3192			timeout = MAX_SCHEDULE_TIMEOUT;
3193		trace_mm_compaction_kcompactd_sleep(pgdat->node_id);
3194		if (wait_event_freezable_timeout(pgdat->kcompactd_wait,
3195			kcompactd_work_requested(pgdat), timeout) &&
3196			!pgdat->proactive_compact_trigger) {
3197
3198			psi_memstall_enter(&pflags);
3199			kcompactd_do_work(pgdat);
3200			psi_memstall_leave(&pflags);
3201			/*
3202			 * Reset the timeout value. The defer timeout from
3203			 * proactive compaction is lost here but that is fine
3204			 * as the condition of the zone changing substantionally
3205			 * then carrying on with the previous defer interval is
3206			 * not useful.
3207			 */
3208			timeout = default_timeout;
3209			continue;
3210		}
3211
3212		/*
3213		 * Start the proactive work with default timeout. Based
3214		 * on the fragmentation score, this timeout is updated.
3215		 */
3216		timeout = default_timeout;
3217		if (should_proactive_compact_node(pgdat)) {
3218			unsigned int prev_score, score;
3219
3220			prev_score = fragmentation_score_node(pgdat);
3221			compact_node(pgdat, true);
3222			score = fragmentation_score_node(pgdat);
3223			/*
3224			 * Defer proactive compaction if the fragmentation
3225			 * score did not go down i.e. no progress made.
3226			 */
3227			if (unlikely(score >= prev_score))
3228				timeout =
3229				   default_timeout << COMPACT_MAX_DEFER_SHIFT;
3230		}
3231		if (unlikely(pgdat->proactive_compact_trigger))
3232			pgdat->proactive_compact_trigger = false;
3233	}
3234
3235	return 0;
3236}
3237
3238/*
3239 * This kcompactd start function will be called by init and node-hot-add.
3240 * On node-hot-add, kcompactd will moved to proper cpus if cpus are hot-added.
3241 */
3242void __meminit kcompactd_run(int nid)
3243{
3244	pg_data_t *pgdat = NODE_DATA(nid);
3245
3246	if (pgdat->kcompactd)
3247		return;
3248
3249	pgdat->kcompactd = kthread_run(kcompactd, pgdat, "kcompactd%d", nid);
3250	if (IS_ERR(pgdat->kcompactd)) {
3251		pr_err("Failed to start kcompactd on node %d\n", nid);
3252		pgdat->kcompactd = NULL;
3253	}
3254}
3255
3256/*
3257 * Called by memory hotplug when all memory in a node is offlined. Caller must
3258 * be holding mem_hotplug_begin/done().
3259 */
3260void __meminit kcompactd_stop(int nid)
3261{
3262	struct task_struct *kcompactd = NODE_DATA(nid)->kcompactd;
3263
3264	if (kcompactd) {
3265		kthread_stop(kcompactd);
3266		NODE_DATA(nid)->kcompactd = NULL;
3267	}
3268}
3269
3270/*
3271 * It's optimal to keep kcompactd on the same CPUs as their memory, but
3272 * not required for correctness. So if the last cpu in a node goes
3273 * away, we get changed to run anywhere: as the first one comes back,
3274 * restore their cpu bindings.
3275 */
3276static int kcompactd_cpu_online(unsigned int cpu)
3277{
3278	int nid;
3279
3280	for_each_node_state(nid, N_MEMORY) {
3281		pg_data_t *pgdat = NODE_DATA(nid);
3282		const struct cpumask *mask;
3283
3284		mask = cpumask_of_node(pgdat->node_id);
3285
3286		if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
3287			/* One of our CPUs online: restore mask */
3288			if (pgdat->kcompactd)
3289				set_cpus_allowed_ptr(pgdat->kcompactd, mask);
3290	}
3291	return 0;
3292}
3293
3294static int proc_dointvec_minmax_warn_RT_change(struct ctl_table *table,
3295		int write, void *buffer, size_t *lenp, loff_t *ppos)
3296{
3297	int ret, old;
3298
3299	if (!IS_ENABLED(CONFIG_PREEMPT_RT) || !write)
3300		return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
3301
3302	old = *(int *)table->data;
3303	ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
3304	if (ret)
3305		return ret;
3306	if (old != *(int *)table->data)
3307		pr_warn_once("sysctl attribute %s changed by %s[%d]\n",
3308			     table->procname, current->comm,
3309			     task_pid_nr(current));
3310	return ret;
3311}
3312
3313static struct ctl_table vm_compaction[] = {
3314	{
3315		.procname	= "compact_memory",
3316		.data		= &sysctl_compact_memory,
3317		.maxlen		= sizeof(int),
3318		.mode		= 0200,
3319		.proc_handler	= sysctl_compaction_handler,
3320	},
3321	{
3322		.procname	= "compaction_proactiveness",
3323		.data		= &sysctl_compaction_proactiveness,
3324		.maxlen		= sizeof(sysctl_compaction_proactiveness),
3325		.mode		= 0644,
3326		.proc_handler	= compaction_proactiveness_sysctl_handler,
3327		.extra1		= SYSCTL_ZERO,
3328		.extra2		= SYSCTL_ONE_HUNDRED,
3329	},
3330	{
3331		.procname	= "extfrag_threshold",
3332		.data		= &sysctl_extfrag_threshold,
3333		.maxlen		= sizeof(int),
3334		.mode		= 0644,
3335		.proc_handler	= proc_dointvec_minmax,
3336		.extra1		= SYSCTL_ZERO,
3337		.extra2		= SYSCTL_ONE_THOUSAND,
3338	},
3339	{
3340		.procname	= "compact_unevictable_allowed",
3341		.data		= &sysctl_compact_unevictable_allowed,
3342		.maxlen		= sizeof(int),
3343		.mode		= 0644,
3344		.proc_handler	= proc_dointvec_minmax_warn_RT_change,
3345		.extra1		= SYSCTL_ZERO,
3346		.extra2		= SYSCTL_ONE,
3347	},
3348	{ }
3349};
3350
3351static int __init kcompactd_init(void)
3352{
3353	int nid;
3354	int ret;
3355
3356	ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
3357					"mm/compaction:online",
3358					kcompactd_cpu_online, NULL);
3359	if (ret < 0) {
3360		pr_err("kcompactd: failed to register hotplug callbacks.\n");
3361		return ret;
3362	}
3363
3364	for_each_node_state(nid, N_MEMORY)
3365		kcompactd_run(nid);
3366	register_sysctl_init("vm", vm_compaction);
3367	return 0;
3368}
3369subsys_initcall(kcompactd_init)
3370
3371#endif /* CONFIG_COMPACTION */