Linux Audio

Check our new training course

Linux kernel drivers training

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