compaction.c

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