Linux Audio

Check our new training course

Loading...
v6.8
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Memory Migration functionality - linux/mm/migrate.c
   4 *
   5 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
   6 *
   7 * Page migration was first developed in the context of the memory hotplug
   8 * project. The main authors of the migration code are:
   9 *
  10 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
  11 * Hirokazu Takahashi <taka@valinux.co.jp>
  12 * Dave Hansen <haveblue@us.ibm.com>
  13 * Christoph Lameter
  14 */
  15
  16#include <linux/migrate.h>
  17#include <linux/export.h>
  18#include <linux/swap.h>
  19#include <linux/swapops.h>
  20#include <linux/pagemap.h>
  21#include <linux/buffer_head.h>
  22#include <linux/mm_inline.h>
  23#include <linux/nsproxy.h>
  24#include <linux/ksm.h>
  25#include <linux/rmap.h>
  26#include <linux/topology.h>
  27#include <linux/cpu.h>
  28#include <linux/cpuset.h>
  29#include <linux/writeback.h>
  30#include <linux/mempolicy.h>
  31#include <linux/vmalloc.h>
  32#include <linux/security.h>
  33#include <linux/backing-dev.h>
  34#include <linux/compaction.h>
  35#include <linux/syscalls.h>
  36#include <linux/compat.h>
  37#include <linux/hugetlb.h>
  38#include <linux/hugetlb_cgroup.h>
  39#include <linux/gfp.h>
  40#include <linux/pfn_t.h>
  41#include <linux/memremap.h>
  42#include <linux/userfaultfd_k.h>
  43#include <linux/balloon_compaction.h>
  44#include <linux/page_idle.h>
  45#include <linux/page_owner.h>
  46#include <linux/sched/mm.h>
  47#include <linux/ptrace.h>
  48#include <linux/oom.h>
  49#include <linux/memory.h>
  50#include <linux/random.h>
  51#include <linux/sched/sysctl.h>
  52#include <linux/memory-tiers.h>
  53
  54#include <asm/tlbflush.h>
  55
  56#include <trace/events/migrate.h>
  57
  58#include "internal.h"
  59
  60bool isolate_movable_page(struct page *page, isolate_mode_t mode)
  61{
  62	struct folio *folio = folio_get_nontail_page(page);
  63	const struct movable_operations *mops;
  64
  65	/*
  66	 * Avoid burning cycles with pages that are yet under __free_pages(),
  67	 * or just got freed under us.
  68	 *
  69	 * In case we 'win' a race for a movable page being freed under us and
  70	 * raise its refcount preventing __free_pages() from doing its job
  71	 * the put_page() at the end of this block will take care of
  72	 * release this page, thus avoiding a nasty leakage.
  73	 */
  74	if (!folio)
  75		goto out;
  76
  77	if (unlikely(folio_test_slab(folio)))
  78		goto out_putfolio;
  79	/* Pairs with smp_wmb() in slab freeing, e.g. SLUB's __free_slab() */
  80	smp_rmb();
  81	/*
  82	 * Check movable flag before taking the page lock because
  83	 * we use non-atomic bitops on newly allocated page flags so
  84	 * unconditionally grabbing the lock ruins page's owner side.
  85	 */
  86	if (unlikely(!__folio_test_movable(folio)))
  87		goto out_putfolio;
  88	/* Pairs with smp_wmb() in slab allocation, e.g. SLUB's alloc_slab_page() */
  89	smp_rmb();
  90	if (unlikely(folio_test_slab(folio)))
  91		goto out_putfolio;
  92
  93	/*
  94	 * As movable pages are not isolated from LRU lists, concurrent
  95	 * compaction threads can race against page migration functions
  96	 * as well as race against the releasing a page.
  97	 *
  98	 * In order to avoid having an already isolated movable page
  99	 * being (wrongly) re-isolated while it is under migration,
 100	 * or to avoid attempting to isolate pages being released,
 101	 * lets be sure we have the page lock
 102	 * before proceeding with the movable page isolation steps.
 103	 */
 104	if (unlikely(!folio_trylock(folio)))
 105		goto out_putfolio;
 106
 107	if (!folio_test_movable(folio) || folio_test_isolated(folio))
 108		goto out_no_isolated;
 109
 110	mops = folio_movable_ops(folio);
 111	VM_BUG_ON_FOLIO(!mops, folio);
 112
 113	if (!mops->isolate_page(&folio->page, mode))
 114		goto out_no_isolated;
 115
 116	/* Driver shouldn't use PG_isolated bit of page->flags */
 117	WARN_ON_ONCE(folio_test_isolated(folio));
 118	folio_set_isolated(folio);
 119	folio_unlock(folio);
 120
 121	return true;
 122
 123out_no_isolated:
 124	folio_unlock(folio);
 125out_putfolio:
 126	folio_put(folio);
 127out:
 128	return false;
 129}
 130
 131static void putback_movable_folio(struct folio *folio)
 132{
 133	const struct movable_operations *mops = folio_movable_ops(folio);
 134
 135	mops->putback_page(&folio->page);
 136	folio_clear_isolated(folio);
 137}
 138
 139/*
 140 * Put previously isolated pages back onto the appropriate lists
 141 * from where they were once taken off for compaction/migration.
 142 *
 143 * This function shall be used whenever the isolated pageset has been
 144 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
 145 * and isolate_hugetlb().
 146 */
 147void putback_movable_pages(struct list_head *l)
 148{
 149	struct folio *folio;
 150	struct folio *folio2;
 151
 152	list_for_each_entry_safe(folio, folio2, l, lru) {
 153		if (unlikely(folio_test_hugetlb(folio))) {
 154			folio_putback_active_hugetlb(folio);
 155			continue;
 156		}
 157		list_del(&folio->lru);
 158		/*
 159		 * We isolated non-lru movable folio so here we can use
 160		 * __folio_test_movable because LRU folio's mapping cannot
 161		 * have PAGE_MAPPING_MOVABLE.
 162		 */
 163		if (unlikely(__folio_test_movable(folio))) {
 164			VM_BUG_ON_FOLIO(!folio_test_isolated(folio), folio);
 165			folio_lock(folio);
 166			if (folio_test_movable(folio))
 167				putback_movable_folio(folio);
 168			else
 169				folio_clear_isolated(folio);
 170			folio_unlock(folio);
 171			folio_put(folio);
 172		} else {
 173			node_stat_mod_folio(folio, NR_ISOLATED_ANON +
 174					folio_is_file_lru(folio), -folio_nr_pages(folio));
 175			folio_putback_lru(folio);
 176		}
 177	}
 178}
 179
 180/*
 181 * Restore a potential migration pte to a working pte entry
 182 */
 183static bool remove_migration_pte(struct folio *folio,
 184		struct vm_area_struct *vma, unsigned long addr, void *old)
 185{
 186	DEFINE_FOLIO_VMA_WALK(pvmw, old, vma, addr, PVMW_SYNC | PVMW_MIGRATION);
 187
 188	while (page_vma_mapped_walk(&pvmw)) {
 189		rmap_t rmap_flags = RMAP_NONE;
 190		pte_t old_pte;
 191		pte_t pte;
 192		swp_entry_t entry;
 193		struct page *new;
 194		unsigned long idx = 0;
 195
 196		/* pgoff is invalid for ksm pages, but they are never large */
 197		if (folio_test_large(folio) && !folio_test_hugetlb(folio))
 198			idx = linear_page_index(vma, pvmw.address) - pvmw.pgoff;
 199		new = folio_page(folio, idx);
 200
 201#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
 202		/* PMD-mapped THP migration entry */
 203		if (!pvmw.pte) {
 204			VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) ||
 205					!folio_test_pmd_mappable(folio), folio);
 206			remove_migration_pmd(&pvmw, new);
 207			continue;
 208		}
 209#endif
 210
 211		folio_get(folio);
 212		pte = mk_pte(new, READ_ONCE(vma->vm_page_prot));
 213		old_pte = ptep_get(pvmw.pte);
 214		if (pte_swp_soft_dirty(old_pte))
 215			pte = pte_mksoft_dirty(pte);
 216
 217		entry = pte_to_swp_entry(old_pte);
 218		if (!is_migration_entry_young(entry))
 219			pte = pte_mkold(pte);
 220		if (folio_test_dirty(folio) && is_migration_entry_dirty(entry))
 221			pte = pte_mkdirty(pte);
 
 
 
 
 
 222		if (is_writable_migration_entry(entry))
 223			pte = pte_mkwrite(pte, vma);
 224		else if (pte_swp_uffd_wp(old_pte))
 225			pte = pte_mkuffd_wp(pte);
 226
 227		if (folio_test_anon(folio) && !is_readable_migration_entry(entry))
 228			rmap_flags |= RMAP_EXCLUSIVE;
 229
 230		if (unlikely(is_device_private_page(new))) {
 231			if (pte_write(pte))
 232				entry = make_writable_device_private_entry(
 233							page_to_pfn(new));
 234			else
 235				entry = make_readable_device_private_entry(
 236							page_to_pfn(new));
 237			pte = swp_entry_to_pte(entry);
 238			if (pte_swp_soft_dirty(old_pte))
 239				pte = pte_swp_mksoft_dirty(pte);
 240			if (pte_swp_uffd_wp(old_pte))
 241				pte = pte_swp_mkuffd_wp(pte);
 242		}
 243
 244#ifdef CONFIG_HUGETLB_PAGE
 245		if (folio_test_hugetlb(folio)) {
 246			struct hstate *h = hstate_vma(vma);
 247			unsigned int shift = huge_page_shift(h);
 248			unsigned long psize = huge_page_size(h);
 249
 250			pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
 251			if (folio_test_anon(folio))
 252				hugetlb_add_anon_rmap(folio, vma, pvmw.address,
 253						      rmap_flags);
 254			else
 255				hugetlb_add_file_rmap(folio);
 256			set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte,
 257					psize);
 258		} else
 259#endif
 260		{
 261			if (folio_test_anon(folio))
 262				folio_add_anon_rmap_pte(folio, new, vma,
 263							pvmw.address, rmap_flags);
 264			else
 265				folio_add_file_rmap_pte(folio, new, vma);
 266			set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
 267		}
 268		if (vma->vm_flags & VM_LOCKED)
 269			mlock_drain_local();
 270
 271		trace_remove_migration_pte(pvmw.address, pte_val(pte),
 272					   compound_order(new));
 273
 274		/* No need to invalidate - it was non-present before */
 275		update_mmu_cache(vma, pvmw.address, pvmw.pte);
 276	}
 277
 278	return true;
 279}
 280
 281/*
 282 * Get rid of all migration entries and replace them by
 283 * references to the indicated page.
 284 */
 285void remove_migration_ptes(struct folio *src, struct folio *dst, bool locked)
 286{
 287	struct rmap_walk_control rwc = {
 288		.rmap_one = remove_migration_pte,
 289		.arg = src,
 290	};
 291
 292	if (locked)
 293		rmap_walk_locked(dst, &rwc);
 294	else
 295		rmap_walk(dst, &rwc);
 296}
 297
 298/*
 299 * Something used the pte of a page under migration. We need to
 300 * get to the page and wait until migration is finished.
 301 * When we return from this function the fault will be retried.
 302 */
 303void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
 304			  unsigned long address)
 305{
 306	spinlock_t *ptl;
 307	pte_t *ptep;
 308	pte_t pte;
 309	swp_entry_t entry;
 310
 311	ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
 312	if (!ptep)
 313		return;
 314
 315	pte = ptep_get(ptep);
 316	pte_unmap(ptep);
 317
 318	if (!is_swap_pte(pte))
 319		goto out;
 320
 321	entry = pte_to_swp_entry(pte);
 322	if (!is_migration_entry(entry))
 323		goto out;
 324
 325	migration_entry_wait_on_locked(entry, ptl);
 326	return;
 327out:
 328	spin_unlock(ptl);
 329}
 330
 331#ifdef CONFIG_HUGETLB_PAGE
 332/*
 333 * The vma read lock must be held upon entry. Holding that lock prevents either
 334 * the pte or the ptl from being freed.
 335 *
 336 * This function will release the vma lock before returning.
 337 */
 338void migration_entry_wait_huge(struct vm_area_struct *vma, pte_t *ptep)
 339{
 340	spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), vma->vm_mm, ptep);
 341	pte_t pte;
 342
 343	hugetlb_vma_assert_locked(vma);
 344	spin_lock(ptl);
 345	pte = huge_ptep_get(ptep);
 346
 347	if (unlikely(!is_hugetlb_entry_migration(pte))) {
 348		spin_unlock(ptl);
 349		hugetlb_vma_unlock_read(vma);
 350	} else {
 351		/*
 352		 * If migration entry existed, safe to release vma lock
 353		 * here because the pgtable page won't be freed without the
 354		 * pgtable lock released.  See comment right above pgtable
 355		 * lock release in migration_entry_wait_on_locked().
 356		 */
 357		hugetlb_vma_unlock_read(vma);
 358		migration_entry_wait_on_locked(pte_to_swp_entry(pte), ptl);
 359	}
 360}
 361#endif
 362
 363#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
 364void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
 365{
 366	spinlock_t *ptl;
 367
 368	ptl = pmd_lock(mm, pmd);
 369	if (!is_pmd_migration_entry(*pmd))
 370		goto unlock;
 371	migration_entry_wait_on_locked(pmd_to_swp_entry(*pmd), ptl);
 372	return;
 373unlock:
 374	spin_unlock(ptl);
 375}
 376#endif
 377
 378static int folio_expected_refs(struct address_space *mapping,
 379		struct folio *folio)
 380{
 381	int refs = 1;
 382	if (!mapping)
 383		return refs;
 384
 385	refs += folio_nr_pages(folio);
 386	if (folio_test_private(folio))
 387		refs++;
 388
 389	return refs;
 390}
 391
 392/*
 393 * Replace the page in the mapping.
 394 *
 395 * The number of remaining references must be:
 396 * 1 for anonymous pages without a mapping
 397 * 2 for pages with a mapping
 398 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
 399 */
 400int folio_migrate_mapping(struct address_space *mapping,
 401		struct folio *newfolio, struct folio *folio, int extra_count)
 402{
 403	XA_STATE(xas, &mapping->i_pages, folio_index(folio));
 404	struct zone *oldzone, *newzone;
 405	int dirty;
 406	int expected_count = folio_expected_refs(mapping, folio) + extra_count;
 407	long nr = folio_nr_pages(folio);
 408	long entries, i;
 409
 410	if (!mapping) {
 411		/* Anonymous page without mapping */
 412		if (folio_ref_count(folio) != expected_count)
 413			return -EAGAIN;
 414
 415		/* No turning back from here */
 416		newfolio->index = folio->index;
 417		newfolio->mapping = folio->mapping;
 418		if (folio_test_swapbacked(folio))
 419			__folio_set_swapbacked(newfolio);
 420
 421		return MIGRATEPAGE_SUCCESS;
 422	}
 423
 424	oldzone = folio_zone(folio);
 425	newzone = folio_zone(newfolio);
 426
 427	xas_lock_irq(&xas);
 428	if (!folio_ref_freeze(folio, expected_count)) {
 429		xas_unlock_irq(&xas);
 430		return -EAGAIN;
 431	}
 432
 433	/*
 434	 * Now we know that no one else is looking at the folio:
 435	 * no turning back from here.
 436	 */
 437	newfolio->index = folio->index;
 438	newfolio->mapping = folio->mapping;
 439	folio_ref_add(newfolio, nr); /* add cache reference */
 440	if (folio_test_swapbacked(folio)) {
 441		__folio_set_swapbacked(newfolio);
 442		if (folio_test_swapcache(folio)) {
 443			folio_set_swapcache(newfolio);
 444			newfolio->private = folio_get_private(folio);
 445		}
 446		entries = nr;
 447	} else {
 448		VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio);
 449		entries = 1;
 450	}
 451
 452	/* Move dirty while page refs frozen and newpage not yet exposed */
 453	dirty = folio_test_dirty(folio);
 454	if (dirty) {
 455		folio_clear_dirty(folio);
 456		folio_set_dirty(newfolio);
 457	}
 458
 459	/* Swap cache still stores N entries instead of a high-order entry */
 460	for (i = 0; i < entries; i++) {
 461		xas_store(&xas, newfolio);
 462		xas_next(&xas);
 463	}
 464
 465	/*
 466	 * Drop cache reference from old page by unfreezing
 467	 * to one less reference.
 468	 * We know this isn't the last reference.
 469	 */
 470	folio_ref_unfreeze(folio, expected_count - nr);
 471
 472	xas_unlock(&xas);
 473	/* Leave irq disabled to prevent preemption while updating stats */
 474
 475	/*
 476	 * If moved to a different zone then also account
 477	 * the page for that zone. Other VM counters will be
 478	 * taken care of when we establish references to the
 479	 * new page and drop references to the old page.
 480	 *
 481	 * Note that anonymous pages are accounted for
 482	 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
 483	 * are mapped to swap space.
 484	 */
 485	if (newzone != oldzone) {
 486		struct lruvec *old_lruvec, *new_lruvec;
 487		struct mem_cgroup *memcg;
 488
 489		memcg = folio_memcg(folio);
 490		old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat);
 491		new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat);
 492
 493		__mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr);
 494		__mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr);
 495		if (folio_test_swapbacked(folio) && !folio_test_swapcache(folio)) {
 496			__mod_lruvec_state(old_lruvec, NR_SHMEM, -nr);
 497			__mod_lruvec_state(new_lruvec, NR_SHMEM, nr);
 498
 499			if (folio_test_pmd_mappable(folio)) {
 500				__mod_lruvec_state(old_lruvec, NR_SHMEM_THPS, -nr);
 501				__mod_lruvec_state(new_lruvec, NR_SHMEM_THPS, nr);
 502			}
 503		}
 504#ifdef CONFIG_SWAP
 505		if (folio_test_swapcache(folio)) {
 506			__mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr);
 507			__mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr);
 508		}
 509#endif
 510		if (dirty && mapping_can_writeback(mapping)) {
 511			__mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr);
 512			__mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr);
 513			__mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr);
 514			__mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr);
 515		}
 516	}
 517	local_irq_enable();
 518
 519	return MIGRATEPAGE_SUCCESS;
 520}
 521EXPORT_SYMBOL(folio_migrate_mapping);
 522
 523/*
 524 * The expected number of remaining references is the same as that
 525 * of folio_migrate_mapping().
 526 */
 527int migrate_huge_page_move_mapping(struct address_space *mapping,
 528				   struct folio *dst, struct folio *src)
 529{
 530	XA_STATE(xas, &mapping->i_pages, folio_index(src));
 531	int expected_count;
 532
 533	xas_lock_irq(&xas);
 534	expected_count = folio_expected_refs(mapping, src);
 535	if (!folio_ref_freeze(src, expected_count)) {
 536		xas_unlock_irq(&xas);
 537		return -EAGAIN;
 538	}
 539
 540	dst->index = src->index;
 541	dst->mapping = src->mapping;
 542
 543	folio_ref_add(dst, folio_nr_pages(dst));
 544
 545	xas_store(&xas, dst);
 546
 547	folio_ref_unfreeze(src, expected_count - folio_nr_pages(src));
 548
 549	xas_unlock_irq(&xas);
 550
 551	return MIGRATEPAGE_SUCCESS;
 552}
 553
 554/*
 555 * Copy the flags and some other ancillary information
 556 */
 557void folio_migrate_flags(struct folio *newfolio, struct folio *folio)
 558{
 559	int cpupid;
 560
 561	if (folio_test_error(folio))
 562		folio_set_error(newfolio);
 563	if (folio_test_referenced(folio))
 564		folio_set_referenced(newfolio);
 565	if (folio_test_uptodate(folio))
 566		folio_mark_uptodate(newfolio);
 567	if (folio_test_clear_active(folio)) {
 568		VM_BUG_ON_FOLIO(folio_test_unevictable(folio), folio);
 569		folio_set_active(newfolio);
 570	} else if (folio_test_clear_unevictable(folio))
 571		folio_set_unevictable(newfolio);
 572	if (folio_test_workingset(folio))
 573		folio_set_workingset(newfolio);
 574	if (folio_test_checked(folio))
 575		folio_set_checked(newfolio);
 576	/*
 577	 * PG_anon_exclusive (-> PG_mappedtodisk) is always migrated via
 578	 * migration entries. We can still have PG_anon_exclusive set on an
 579	 * effectively unmapped and unreferenced first sub-pages of an
 580	 * anonymous THP: we can simply copy it here via PG_mappedtodisk.
 581	 */
 582	if (folio_test_mappedtodisk(folio))
 583		folio_set_mappedtodisk(newfolio);
 584
 585	/* Move dirty on pages not done by folio_migrate_mapping() */
 586	if (folio_test_dirty(folio))
 587		folio_set_dirty(newfolio);
 588
 589	if (folio_test_young(folio))
 590		folio_set_young(newfolio);
 591	if (folio_test_idle(folio))
 592		folio_set_idle(newfolio);
 593
 594	/*
 595	 * Copy NUMA information to the new page, to prevent over-eager
 596	 * future migrations of this same page.
 597	 */
 598	cpupid = folio_xchg_last_cpupid(folio, -1);
 599	/*
 600	 * For memory tiering mode, when migrate between slow and fast
 601	 * memory node, reset cpupid, because that is used to record
 602	 * page access time in slow memory node.
 603	 */
 604	if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) {
 605		bool f_toptier = node_is_toptier(folio_nid(folio));
 606		bool t_toptier = node_is_toptier(folio_nid(newfolio));
 607
 608		if (f_toptier != t_toptier)
 609			cpupid = -1;
 610	}
 611	folio_xchg_last_cpupid(newfolio, cpupid);
 612
 613	folio_migrate_ksm(newfolio, folio);
 614	/*
 615	 * Please do not reorder this without considering how mm/ksm.c's
 616	 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
 617	 */
 618	if (folio_test_swapcache(folio))
 619		folio_clear_swapcache(folio);
 620	folio_clear_private(folio);
 621
 622	/* page->private contains hugetlb specific flags */
 623	if (!folio_test_hugetlb(folio))
 624		folio->private = NULL;
 625
 626	/*
 627	 * If any waiters have accumulated on the new page then
 628	 * wake them up.
 629	 */
 630	if (folio_test_writeback(newfolio))
 631		folio_end_writeback(newfolio);
 632
 633	/*
 634	 * PG_readahead shares the same bit with PG_reclaim.  The above
 635	 * end_page_writeback() may clear PG_readahead mistakenly, so set the
 636	 * bit after that.
 637	 */
 638	if (folio_test_readahead(folio))
 639		folio_set_readahead(newfolio);
 640
 641	folio_copy_owner(newfolio, folio);
 642
 643	mem_cgroup_migrate(folio, newfolio);
 644}
 645EXPORT_SYMBOL(folio_migrate_flags);
 646
 647void folio_migrate_copy(struct folio *newfolio, struct folio *folio)
 648{
 649	folio_copy(newfolio, folio);
 650	folio_migrate_flags(newfolio, folio);
 651}
 652EXPORT_SYMBOL(folio_migrate_copy);
 653
 654/************************************************************
 655 *                    Migration functions
 656 ***********************************************************/
 657
 658int migrate_folio_extra(struct address_space *mapping, struct folio *dst,
 659		struct folio *src, enum migrate_mode mode, int extra_count)
 660{
 661	int rc;
 662
 663	BUG_ON(folio_test_writeback(src));	/* Writeback must be complete */
 664
 665	rc = folio_migrate_mapping(mapping, dst, src, extra_count);
 666
 667	if (rc != MIGRATEPAGE_SUCCESS)
 668		return rc;
 669
 670	if (mode != MIGRATE_SYNC_NO_COPY)
 671		folio_migrate_copy(dst, src);
 672	else
 673		folio_migrate_flags(dst, src);
 674	return MIGRATEPAGE_SUCCESS;
 675}
 676
 677/**
 678 * migrate_folio() - Simple folio migration.
 679 * @mapping: The address_space containing the folio.
 680 * @dst: The folio to migrate the data to.
 681 * @src: The folio containing the current data.
 682 * @mode: How to migrate the page.
 683 *
 684 * Common logic to directly migrate a single LRU folio suitable for
 685 * folios that do not use PagePrivate/PagePrivate2.
 686 *
 687 * Folios are locked upon entry and exit.
 688 */
 689int migrate_folio(struct address_space *mapping, struct folio *dst,
 690		struct folio *src, enum migrate_mode mode)
 691{
 692	return migrate_folio_extra(mapping, dst, src, mode, 0);
 693}
 694EXPORT_SYMBOL(migrate_folio);
 695
 696#ifdef CONFIG_BUFFER_HEAD
 697/* Returns true if all buffers are successfully locked */
 698static bool buffer_migrate_lock_buffers(struct buffer_head *head,
 699							enum migrate_mode mode)
 700{
 701	struct buffer_head *bh = head;
 702	struct buffer_head *failed_bh;
 703
 704	do {
 705		if (!trylock_buffer(bh)) {
 706			if (mode == MIGRATE_ASYNC)
 707				goto unlock;
 708			if (mode == MIGRATE_SYNC_LIGHT && !buffer_uptodate(bh))
 709				goto unlock;
 710			lock_buffer(bh);
 711		}
 712
 713		bh = bh->b_this_page;
 714	} while (bh != head);
 715
 716	return true;
 717
 718unlock:
 719	/* We failed to lock the buffer and cannot stall. */
 720	failed_bh = bh;
 721	bh = head;
 722	while (bh != failed_bh) {
 723		unlock_buffer(bh);
 724		bh = bh->b_this_page;
 725	}
 726
 727	return false;
 728}
 729
 730static int __buffer_migrate_folio(struct address_space *mapping,
 731		struct folio *dst, struct folio *src, enum migrate_mode mode,
 732		bool check_refs)
 733{
 734	struct buffer_head *bh, *head;
 735	int rc;
 736	int expected_count;
 737
 738	head = folio_buffers(src);
 739	if (!head)
 740		return migrate_folio(mapping, dst, src, mode);
 741
 742	/* Check whether page does not have extra refs before we do more work */
 743	expected_count = folio_expected_refs(mapping, src);
 744	if (folio_ref_count(src) != expected_count)
 745		return -EAGAIN;
 746
 747	if (!buffer_migrate_lock_buffers(head, mode))
 748		return -EAGAIN;
 749
 750	if (check_refs) {
 751		bool busy;
 752		bool invalidated = false;
 753
 754recheck_buffers:
 755		busy = false;
 756		spin_lock(&mapping->i_private_lock);
 757		bh = head;
 758		do {
 759			if (atomic_read(&bh->b_count)) {
 760				busy = true;
 761				break;
 762			}
 763			bh = bh->b_this_page;
 764		} while (bh != head);
 765		if (busy) {
 766			if (invalidated) {
 767				rc = -EAGAIN;
 768				goto unlock_buffers;
 769			}
 770			spin_unlock(&mapping->i_private_lock);
 771			invalidate_bh_lrus();
 772			invalidated = true;
 773			goto recheck_buffers;
 774		}
 775	}
 776
 777	rc = folio_migrate_mapping(mapping, dst, src, 0);
 778	if (rc != MIGRATEPAGE_SUCCESS)
 779		goto unlock_buffers;
 780
 781	folio_attach_private(dst, folio_detach_private(src));
 782
 783	bh = head;
 784	do {
 785		folio_set_bh(bh, dst, bh_offset(bh));
 786		bh = bh->b_this_page;
 787	} while (bh != head);
 788
 789	if (mode != MIGRATE_SYNC_NO_COPY)
 790		folio_migrate_copy(dst, src);
 791	else
 792		folio_migrate_flags(dst, src);
 793
 794	rc = MIGRATEPAGE_SUCCESS;
 795unlock_buffers:
 796	if (check_refs)
 797		spin_unlock(&mapping->i_private_lock);
 798	bh = head;
 799	do {
 800		unlock_buffer(bh);
 801		bh = bh->b_this_page;
 802	} while (bh != head);
 803
 804	return rc;
 805}
 806
 807/**
 808 * buffer_migrate_folio() - Migration function for folios with buffers.
 809 * @mapping: The address space containing @src.
 810 * @dst: The folio to migrate to.
 811 * @src: The folio to migrate from.
 812 * @mode: How to migrate the folio.
 813 *
 814 * This function can only be used if the underlying filesystem guarantees
 815 * that no other references to @src exist. For example attached buffer
 816 * heads are accessed only under the folio lock.  If your filesystem cannot
 817 * provide this guarantee, buffer_migrate_folio_norefs() may be more
 818 * appropriate.
 819 *
 820 * Return: 0 on success or a negative errno on failure.
 821 */
 822int buffer_migrate_folio(struct address_space *mapping,
 823		struct folio *dst, struct folio *src, enum migrate_mode mode)
 824{
 825	return __buffer_migrate_folio(mapping, dst, src, mode, false);
 826}
 827EXPORT_SYMBOL(buffer_migrate_folio);
 828
 829/**
 830 * buffer_migrate_folio_norefs() - Migration function for folios with buffers.
 831 * @mapping: The address space containing @src.
 832 * @dst: The folio to migrate to.
 833 * @src: The folio to migrate from.
 834 * @mode: How to migrate the folio.
 835 *
 836 * Like buffer_migrate_folio() except that this variant is more careful
 837 * and checks that there are also no buffer head references. This function
 838 * is the right one for mappings where buffer heads are directly looked
 839 * up and referenced (such as block device mappings).
 840 *
 841 * Return: 0 on success or a negative errno on failure.
 842 */
 843int buffer_migrate_folio_norefs(struct address_space *mapping,
 844		struct folio *dst, struct folio *src, enum migrate_mode mode)
 845{
 846	return __buffer_migrate_folio(mapping, dst, src, mode, true);
 847}
 848EXPORT_SYMBOL_GPL(buffer_migrate_folio_norefs);
 849#endif /* CONFIG_BUFFER_HEAD */
 850
 851int filemap_migrate_folio(struct address_space *mapping,
 852		struct folio *dst, struct folio *src, enum migrate_mode mode)
 853{
 854	int ret;
 855
 856	ret = folio_migrate_mapping(mapping, dst, src, 0);
 857	if (ret != MIGRATEPAGE_SUCCESS)
 858		return ret;
 859
 860	if (folio_get_private(src))
 861		folio_attach_private(dst, folio_detach_private(src));
 862
 863	if (mode != MIGRATE_SYNC_NO_COPY)
 864		folio_migrate_copy(dst, src);
 865	else
 866		folio_migrate_flags(dst, src);
 867	return MIGRATEPAGE_SUCCESS;
 868}
 869EXPORT_SYMBOL_GPL(filemap_migrate_folio);
 870
 871/*
 872 * Writeback a folio to clean the dirty state
 873 */
 874static int writeout(struct address_space *mapping, struct folio *folio)
 875{
 876	struct writeback_control wbc = {
 877		.sync_mode = WB_SYNC_NONE,
 878		.nr_to_write = 1,
 879		.range_start = 0,
 880		.range_end = LLONG_MAX,
 881		.for_reclaim = 1
 882	};
 883	int rc;
 884
 885	if (!mapping->a_ops->writepage)
 886		/* No write method for the address space */
 887		return -EINVAL;
 888
 889	if (!folio_clear_dirty_for_io(folio))
 890		/* Someone else already triggered a write */
 891		return -EAGAIN;
 892
 893	/*
 894	 * A dirty folio may imply that the underlying filesystem has
 895	 * the folio on some queue. So the folio must be clean for
 896	 * migration. Writeout may mean we lose the lock and the
 897	 * folio state is no longer what we checked for earlier.
 898	 * At this point we know that the migration attempt cannot
 899	 * be successful.
 900	 */
 901	remove_migration_ptes(folio, folio, false);
 902
 903	rc = mapping->a_ops->writepage(&folio->page, &wbc);
 904
 905	if (rc != AOP_WRITEPAGE_ACTIVATE)
 906		/* unlocked. Relock */
 907		folio_lock(folio);
 908
 909	return (rc < 0) ? -EIO : -EAGAIN;
 910}
 911
 912/*
 913 * Default handling if a filesystem does not provide a migration function.
 914 */
 915static int fallback_migrate_folio(struct address_space *mapping,
 916		struct folio *dst, struct folio *src, enum migrate_mode mode)
 917{
 918	if (folio_test_dirty(src)) {
 919		/* Only writeback folios in full synchronous migration */
 920		switch (mode) {
 921		case MIGRATE_SYNC:
 922		case MIGRATE_SYNC_NO_COPY:
 923			break;
 924		default:
 925			return -EBUSY;
 926		}
 927		return writeout(mapping, src);
 928	}
 929
 930	/*
 931	 * Buffers may be managed in a filesystem specific way.
 932	 * We must have no buffers or drop them.
 933	 */
 934	if (!filemap_release_folio(src, GFP_KERNEL))
 935		return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
 936
 937	return migrate_folio(mapping, dst, src, mode);
 938}
 939
 940/*
 941 * Move a page to a newly allocated page
 942 * The page is locked and all ptes have been successfully removed.
 943 *
 944 * The new page will have replaced the old page if this function
 945 * is successful.
 946 *
 947 * Return value:
 948 *   < 0 - error code
 949 *  MIGRATEPAGE_SUCCESS - success
 950 */
 951static int move_to_new_folio(struct folio *dst, struct folio *src,
 952				enum migrate_mode mode)
 953{
 954	int rc = -EAGAIN;
 955	bool is_lru = !__folio_test_movable(src);
 956
 957	VM_BUG_ON_FOLIO(!folio_test_locked(src), src);
 958	VM_BUG_ON_FOLIO(!folio_test_locked(dst), dst);
 959
 960	if (likely(is_lru)) {
 961		struct address_space *mapping = folio_mapping(src);
 962
 963		if (!mapping)
 964			rc = migrate_folio(mapping, dst, src, mode);
 965		else if (mapping_unmovable(mapping))
 966			rc = -EOPNOTSUPP;
 967		else if (mapping->a_ops->migrate_folio)
 968			/*
 969			 * Most folios have a mapping and most filesystems
 970			 * provide a migrate_folio callback. Anonymous folios
 971			 * are part of swap space which also has its own
 972			 * migrate_folio callback. This is the most common path
 973			 * for page migration.
 974			 */
 975			rc = mapping->a_ops->migrate_folio(mapping, dst, src,
 976								mode);
 977		else
 978			rc = fallback_migrate_folio(mapping, dst, src, mode);
 979	} else {
 980		const struct movable_operations *mops;
 981
 982		/*
 983		 * In case of non-lru page, it could be released after
 984		 * isolation step. In that case, we shouldn't try migration.
 985		 */
 986		VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
 987		if (!folio_test_movable(src)) {
 988			rc = MIGRATEPAGE_SUCCESS;
 989			folio_clear_isolated(src);
 990			goto out;
 991		}
 992
 993		mops = folio_movable_ops(src);
 994		rc = mops->migrate_page(&dst->page, &src->page, mode);
 995		WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
 996				!folio_test_isolated(src));
 997	}
 998
 999	/*
1000	 * When successful, old pagecache src->mapping must be cleared before
1001	 * src is freed; but stats require that PageAnon be left as PageAnon.
1002	 */
1003	if (rc == MIGRATEPAGE_SUCCESS) {
1004		if (__folio_test_movable(src)) {
1005			VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
1006
1007			/*
1008			 * We clear PG_movable under page_lock so any compactor
1009			 * cannot try to migrate this page.
1010			 */
1011			folio_clear_isolated(src);
1012		}
1013
1014		/*
1015		 * Anonymous and movable src->mapping will be cleared by
1016		 * free_pages_prepare so don't reset it here for keeping
1017		 * the type to work PageAnon, for example.
1018		 */
1019		if (!folio_mapping_flags(src))
1020			src->mapping = NULL;
1021
1022		if (likely(!folio_is_zone_device(dst)))
1023			flush_dcache_folio(dst);
1024	}
1025out:
1026	return rc;
1027}
1028
1029/*
1030 * To record some information during migration, we use unused private
1031 * field of struct folio of the newly allocated destination folio.
1032 * This is safe because nobody is using it except us.
1033 */
1034enum {
1035	PAGE_WAS_MAPPED = BIT(0),
1036	PAGE_WAS_MLOCKED = BIT(1),
1037	PAGE_OLD_STATES = PAGE_WAS_MAPPED | PAGE_WAS_MLOCKED,
1038};
1039
1040static void __migrate_folio_record(struct folio *dst,
1041				   int old_page_state,
1042				   struct anon_vma *anon_vma)
1043{
1044	dst->private = (void *)anon_vma + old_page_state;
1045}
1046
1047static void __migrate_folio_extract(struct folio *dst,
1048				   int *old_page_state,
1049				   struct anon_vma **anon_vmap)
1050{
1051	unsigned long private = (unsigned long)dst->private;
1052
1053	*anon_vmap = (struct anon_vma *)(private & ~PAGE_OLD_STATES);
1054	*old_page_state = private & PAGE_OLD_STATES;
1055	dst->private = NULL;
1056}
1057
1058/* Restore the source folio to the original state upon failure */
1059static void migrate_folio_undo_src(struct folio *src,
1060				   int page_was_mapped,
1061				   struct anon_vma *anon_vma,
1062				   bool locked,
1063				   struct list_head *ret)
1064{
1065	if (page_was_mapped)
1066		remove_migration_ptes(src, src, false);
1067	/* Drop an anon_vma reference if we took one */
1068	if (anon_vma)
1069		put_anon_vma(anon_vma);
1070	if (locked)
1071		folio_unlock(src);
1072	if (ret)
1073		list_move_tail(&src->lru, ret);
1074}
1075
1076/* Restore the destination folio to the original state upon failure */
1077static void migrate_folio_undo_dst(struct folio *dst, bool locked,
1078		free_folio_t put_new_folio, unsigned long private)
1079{
1080	if (locked)
1081		folio_unlock(dst);
1082	if (put_new_folio)
1083		put_new_folio(dst, private);
1084	else
1085		folio_put(dst);
1086}
1087
1088/* Cleanup src folio upon migration success */
1089static void migrate_folio_done(struct folio *src,
1090			       enum migrate_reason reason)
1091{
1092	/*
1093	 * Compaction can migrate also non-LRU pages which are
1094	 * not accounted to NR_ISOLATED_*. They can be recognized
1095	 * as __folio_test_movable
1096	 */
1097	if (likely(!__folio_test_movable(src)))
1098		mod_node_page_state(folio_pgdat(src), NR_ISOLATED_ANON +
1099				    folio_is_file_lru(src), -folio_nr_pages(src));
1100
1101	if (reason != MR_MEMORY_FAILURE)
1102		/* We release the page in page_handle_poison. */
1103		folio_put(src);
1104}
1105
1106/* Obtain the lock on page, remove all ptes. */
1107static int migrate_folio_unmap(new_folio_t get_new_folio,
1108		free_folio_t put_new_folio, unsigned long private,
1109		struct folio *src, struct folio **dstp, enum migrate_mode mode,
1110		enum migrate_reason reason, struct list_head *ret)
1111{
1112	struct folio *dst;
1113	int rc = -EAGAIN;
1114	int old_page_state = 0;
1115	struct anon_vma *anon_vma = NULL;
1116	bool is_lru = !__folio_test_movable(src);
1117	bool locked = false;
1118	bool dst_locked = false;
1119
1120	if (folio_ref_count(src) == 1) {
1121		/* Folio was freed from under us. So we are done. */
1122		folio_clear_active(src);
1123		folio_clear_unevictable(src);
1124		/* free_pages_prepare() will clear PG_isolated. */
1125		list_del(&src->lru);
1126		migrate_folio_done(src, reason);
1127		return MIGRATEPAGE_SUCCESS;
1128	}
1129
1130	dst = get_new_folio(src, private);
1131	if (!dst)
1132		return -ENOMEM;
1133	*dstp = dst;
1134
1135	dst->private = NULL;
1136
1137	if (!folio_trylock(src)) {
1138		if (mode == MIGRATE_ASYNC)
1139			goto out;
1140
1141		/*
1142		 * It's not safe for direct compaction to call lock_page.
1143		 * For example, during page readahead pages are added locked
1144		 * to the LRU. Later, when the IO completes the pages are
1145		 * marked uptodate and unlocked. However, the queueing
1146		 * could be merging multiple pages for one bio (e.g.
1147		 * mpage_readahead). If an allocation happens for the
1148		 * second or third page, the process can end up locking
1149		 * the same page twice and deadlocking. Rather than
1150		 * trying to be clever about what pages can be locked,
1151		 * avoid the use of lock_page for direct compaction
1152		 * altogether.
1153		 */
1154		if (current->flags & PF_MEMALLOC)
1155			goto out;
1156
1157		/*
1158		 * In "light" mode, we can wait for transient locks (eg
1159		 * inserting a page into the page table), but it's not
1160		 * worth waiting for I/O.
1161		 */
1162		if (mode == MIGRATE_SYNC_LIGHT && !folio_test_uptodate(src))
1163			goto out;
1164
1165		folio_lock(src);
1166	}
1167	locked = true;
1168	if (folio_test_mlocked(src))
1169		old_page_state |= PAGE_WAS_MLOCKED;
1170
1171	if (folio_test_writeback(src)) {
1172		/*
1173		 * Only in the case of a full synchronous migration is it
1174		 * necessary to wait for PageWriteback. In the async case,
1175		 * the retry loop is too short and in the sync-light case,
1176		 * the overhead of stalling is too much
1177		 */
1178		switch (mode) {
1179		case MIGRATE_SYNC:
1180		case MIGRATE_SYNC_NO_COPY:
1181			break;
1182		default:
1183			rc = -EBUSY;
1184			goto out;
1185		}
1186		folio_wait_writeback(src);
1187	}
1188
1189	/*
1190	 * By try_to_migrate(), src->mapcount goes down to 0 here. In this case,
1191	 * we cannot notice that anon_vma is freed while we migrate a page.
1192	 * This get_anon_vma() delays freeing anon_vma pointer until the end
1193	 * of migration. File cache pages are no problem because of page_lock()
1194	 * File Caches may use write_page() or lock_page() in migration, then,
1195	 * just care Anon page here.
1196	 *
1197	 * Only folio_get_anon_vma() understands the subtleties of
1198	 * getting a hold on an anon_vma from outside one of its mms.
1199	 * But if we cannot get anon_vma, then we won't need it anyway,
1200	 * because that implies that the anon page is no longer mapped
1201	 * (and cannot be remapped so long as we hold the page lock).
1202	 */
1203	if (folio_test_anon(src) && !folio_test_ksm(src))
1204		anon_vma = folio_get_anon_vma(src);
1205
1206	/*
1207	 * Block others from accessing the new page when we get around to
1208	 * establishing additional references. We are usually the only one
1209	 * holding a reference to dst at this point. We used to have a BUG
1210	 * here if folio_trylock(dst) fails, but would like to allow for
1211	 * cases where there might be a race with the previous use of dst.
1212	 * This is much like races on refcount of oldpage: just don't BUG().
1213	 */
1214	if (unlikely(!folio_trylock(dst)))
1215		goto out;
1216	dst_locked = true;
1217
1218	if (unlikely(!is_lru)) {
1219		__migrate_folio_record(dst, old_page_state, anon_vma);
1220		return MIGRATEPAGE_UNMAP;
1221	}
1222
1223	/*
1224	 * Corner case handling:
1225	 * 1. When a new swap-cache page is read into, it is added to the LRU
1226	 * and treated as swapcache but it has no rmap yet.
1227	 * Calling try_to_unmap() against a src->mapping==NULL page will
1228	 * trigger a BUG.  So handle it here.
1229	 * 2. An orphaned page (see truncate_cleanup_page) might have
1230	 * fs-private metadata. The page can be picked up due to memory
1231	 * offlining.  Everywhere else except page reclaim, the page is
1232	 * invisible to the vm, so the page can not be migrated.  So try to
1233	 * free the metadata, so the page can be freed.
1234	 */
1235	if (!src->mapping) {
1236		if (folio_test_private(src)) {
1237			try_to_free_buffers(src);
1238			goto out;
1239		}
1240	} else if (folio_mapped(src)) {
1241		/* Establish migration ptes */
1242		VM_BUG_ON_FOLIO(folio_test_anon(src) &&
1243			       !folio_test_ksm(src) && !anon_vma, src);
1244		try_to_migrate(src, mode == MIGRATE_ASYNC ? TTU_BATCH_FLUSH : 0);
1245		old_page_state |= PAGE_WAS_MAPPED;
1246	}
1247
1248	if (!folio_mapped(src)) {
1249		__migrate_folio_record(dst, old_page_state, anon_vma);
1250		return MIGRATEPAGE_UNMAP;
1251	}
1252
1253out:
1254	/*
1255	 * A folio that has not been unmapped will be restored to
1256	 * right list unless we want to retry.
1257	 */
1258	if (rc == -EAGAIN)
1259		ret = NULL;
1260
1261	migrate_folio_undo_src(src, old_page_state & PAGE_WAS_MAPPED,
1262			       anon_vma, locked, ret);
1263	migrate_folio_undo_dst(dst, dst_locked, put_new_folio, private);
1264
1265	return rc;
1266}
1267
1268/* Migrate the folio to the newly allocated folio in dst. */
1269static int migrate_folio_move(free_folio_t put_new_folio, unsigned long private,
1270			      struct folio *src, struct folio *dst,
1271			      enum migrate_mode mode, enum migrate_reason reason,
1272			      struct list_head *ret)
1273{
1274	int rc;
1275	int old_page_state = 0;
1276	struct anon_vma *anon_vma = NULL;
1277	bool is_lru = !__folio_test_movable(src);
1278	struct list_head *prev;
1279
1280	__migrate_folio_extract(dst, &old_page_state, &anon_vma);
1281	prev = dst->lru.prev;
1282	list_del(&dst->lru);
1283
1284	rc = move_to_new_folio(dst, src, mode);
1285	if (rc)
1286		goto out;
1287
1288	if (unlikely(!is_lru))
1289		goto out_unlock_both;
1290
1291	/*
1292	 * When successful, push dst to LRU immediately: so that if it
1293	 * turns out to be an mlocked page, remove_migration_ptes() will
1294	 * automatically build up the correct dst->mlock_count for it.
1295	 *
1296	 * We would like to do something similar for the old page, when
1297	 * unsuccessful, and other cases when a page has been temporarily
1298	 * isolated from the unevictable LRU: but this case is the easiest.
1299	 */
1300	folio_add_lru(dst);
1301	if (old_page_state & PAGE_WAS_MLOCKED)
1302		lru_add_drain();
1303
1304	if (old_page_state & PAGE_WAS_MAPPED)
1305		remove_migration_ptes(src, dst, false);
1306
1307out_unlock_both:
1308	folio_unlock(dst);
1309	set_page_owner_migrate_reason(&dst->page, reason);
1310	/*
1311	 * If migration is successful, decrease refcount of dst,
1312	 * which will not free the page because new page owner increased
1313	 * refcounter.
1314	 */
1315	folio_put(dst);
1316
1317	/*
1318	 * A folio that has been migrated has all references removed
1319	 * and will be freed.
1320	 */
1321	list_del(&src->lru);
1322	/* Drop an anon_vma reference if we took one */
1323	if (anon_vma)
1324		put_anon_vma(anon_vma);
1325	folio_unlock(src);
1326	migrate_folio_done(src, reason);
1327
1328	return rc;
1329out:
1330	/*
1331	 * A folio that has not been migrated will be restored to
1332	 * right list unless we want to retry.
1333	 */
1334	if (rc == -EAGAIN) {
1335		list_add(&dst->lru, prev);
1336		__migrate_folio_record(dst, old_page_state, anon_vma);
1337		return rc;
1338	}
1339
1340	migrate_folio_undo_src(src, old_page_state & PAGE_WAS_MAPPED,
1341			       anon_vma, true, ret);
1342	migrate_folio_undo_dst(dst, true, put_new_folio, private);
1343
1344	return rc;
1345}
1346
1347/*
1348 * Counterpart of unmap_and_move_page() for hugepage migration.
1349 *
1350 * This function doesn't wait the completion of hugepage I/O
1351 * because there is no race between I/O and migration for hugepage.
1352 * Note that currently hugepage I/O occurs only in direct I/O
1353 * where no lock is held and PG_writeback is irrelevant,
1354 * and writeback status of all subpages are counted in the reference
1355 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1356 * under direct I/O, the reference of the head page is 512 and a bit more.)
1357 * This means that when we try to migrate hugepage whose subpages are
1358 * doing direct I/O, some references remain after try_to_unmap() and
1359 * hugepage migration fails without data corruption.
1360 *
1361 * There is also no race when direct I/O is issued on the page under migration,
1362 * because then pte is replaced with migration swap entry and direct I/O code
1363 * will wait in the page fault for migration to complete.
1364 */
1365static int unmap_and_move_huge_page(new_folio_t get_new_folio,
1366		free_folio_t put_new_folio, unsigned long private,
1367		struct folio *src, int force, enum migrate_mode mode,
1368		int reason, struct list_head *ret)
1369{
1370	struct folio *dst;
1371	int rc = -EAGAIN;
1372	int page_was_mapped = 0;
1373	struct anon_vma *anon_vma = NULL;
1374	struct address_space *mapping = NULL;
1375
1376	if (folio_ref_count(src) == 1) {
1377		/* page was freed from under us. So we are done. */
1378		folio_putback_active_hugetlb(src);
1379		return MIGRATEPAGE_SUCCESS;
1380	}
1381
1382	dst = get_new_folio(src, private);
1383	if (!dst)
1384		return -ENOMEM;
1385
1386	if (!folio_trylock(src)) {
1387		if (!force)
1388			goto out;
1389		switch (mode) {
1390		case MIGRATE_SYNC:
1391		case MIGRATE_SYNC_NO_COPY:
1392			break;
1393		default:
1394			goto out;
1395		}
1396		folio_lock(src);
1397	}
1398
1399	/*
1400	 * Check for pages which are in the process of being freed.  Without
1401	 * folio_mapping() set, hugetlbfs specific move page routine will not
1402	 * be called and we could leak usage counts for subpools.
1403	 */
1404	if (hugetlb_folio_subpool(src) && !folio_mapping(src)) {
1405		rc = -EBUSY;
1406		goto out_unlock;
1407	}
1408
1409	if (folio_test_anon(src))
1410		anon_vma = folio_get_anon_vma(src);
1411
1412	if (unlikely(!folio_trylock(dst)))
1413		goto put_anon;
1414
1415	if (folio_mapped(src)) {
1416		enum ttu_flags ttu = 0;
1417
1418		if (!folio_test_anon(src)) {
1419			/*
1420			 * In shared mappings, try_to_unmap could potentially
1421			 * call huge_pmd_unshare.  Because of this, take
1422			 * semaphore in write mode here and set TTU_RMAP_LOCKED
1423			 * to let lower levels know we have taken the lock.
1424			 */
1425			mapping = hugetlb_page_mapping_lock_write(&src->page);
1426			if (unlikely(!mapping))
1427				goto unlock_put_anon;
1428
1429			ttu = TTU_RMAP_LOCKED;
1430		}
1431
1432		try_to_migrate(src, ttu);
1433		page_was_mapped = 1;
1434
1435		if (ttu & TTU_RMAP_LOCKED)
1436			i_mmap_unlock_write(mapping);
1437	}
1438
1439	if (!folio_mapped(src))
1440		rc = move_to_new_folio(dst, src, mode);
1441
1442	if (page_was_mapped)
1443		remove_migration_ptes(src,
1444			rc == MIGRATEPAGE_SUCCESS ? dst : src, false);
1445
1446unlock_put_anon:
1447	folio_unlock(dst);
1448
1449put_anon:
1450	if (anon_vma)
1451		put_anon_vma(anon_vma);
1452
1453	if (rc == MIGRATEPAGE_SUCCESS) {
1454		move_hugetlb_state(src, dst, reason);
1455		put_new_folio = NULL;
1456	}
1457
1458out_unlock:
1459	folio_unlock(src);
1460out:
1461	if (rc == MIGRATEPAGE_SUCCESS)
1462		folio_putback_active_hugetlb(src);
1463	else if (rc != -EAGAIN)
1464		list_move_tail(&src->lru, ret);
1465
1466	/*
1467	 * If migration was not successful and there's a freeing callback, use
1468	 * it.  Otherwise, put_page() will drop the reference grabbed during
1469	 * isolation.
1470	 */
1471	if (put_new_folio)
1472		put_new_folio(dst, private);
1473	else
1474		folio_putback_active_hugetlb(dst);
1475
1476	return rc;
1477}
1478
1479static inline int try_split_folio(struct folio *folio, struct list_head *split_folios)
1480{
1481	int rc;
1482
1483	folio_lock(folio);
1484	rc = split_folio_to_list(folio, split_folios);
1485	folio_unlock(folio);
1486	if (!rc)
1487		list_move_tail(&folio->lru, split_folios);
1488
1489	return rc;
1490}
1491
1492#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1493#define NR_MAX_BATCHED_MIGRATION	HPAGE_PMD_NR
1494#else
1495#define NR_MAX_BATCHED_MIGRATION	512
1496#endif
1497#define NR_MAX_MIGRATE_PAGES_RETRY	10
1498#define NR_MAX_MIGRATE_ASYNC_RETRY	3
1499#define NR_MAX_MIGRATE_SYNC_RETRY					\
1500	(NR_MAX_MIGRATE_PAGES_RETRY - NR_MAX_MIGRATE_ASYNC_RETRY)
1501
1502struct migrate_pages_stats {
1503	int nr_succeeded;	/* Normal and large folios migrated successfully, in
1504				   units of base pages */
1505	int nr_failed_pages;	/* Normal and large folios failed to be migrated, in
1506				   units of base pages.  Untried folios aren't counted */
1507	int nr_thp_succeeded;	/* THP migrated successfully */
1508	int nr_thp_failed;	/* THP failed to be migrated */
1509	int nr_thp_split;	/* THP split before migrating */
1510	int nr_split;	/* Large folio (include THP) split before migrating */
1511};
1512
1513/*
1514 * Returns the number of hugetlb folios that were not migrated, or an error code
1515 * after NR_MAX_MIGRATE_PAGES_RETRY attempts or if no hugetlb folios are movable
1516 * any more because the list has become empty or no retryable hugetlb folios
1517 * exist any more. It is caller's responsibility to call putback_movable_pages()
1518 * only if ret != 0.
1519 */
1520static int migrate_hugetlbs(struct list_head *from, new_folio_t get_new_folio,
1521			    free_folio_t put_new_folio, unsigned long private,
1522			    enum migrate_mode mode, int reason,
1523			    struct migrate_pages_stats *stats,
1524			    struct list_head *ret_folios)
1525{
1526	int retry = 1;
1527	int nr_failed = 0;
1528	int nr_retry_pages = 0;
1529	int pass = 0;
1530	struct folio *folio, *folio2;
1531	int rc, nr_pages;
1532
1533	for (pass = 0; pass < NR_MAX_MIGRATE_PAGES_RETRY && retry; pass++) {
1534		retry = 0;
1535		nr_retry_pages = 0;
1536
1537		list_for_each_entry_safe(folio, folio2, from, lru) {
1538			if (!folio_test_hugetlb(folio))
1539				continue;
1540
1541			nr_pages = folio_nr_pages(folio);
1542
1543			cond_resched();
1544
1545			/*
1546			 * Migratability of hugepages depends on architectures and
1547			 * their size.  This check is necessary because some callers
1548			 * of hugepage migration like soft offline and memory
1549			 * hotremove don't walk through page tables or check whether
1550			 * the hugepage is pmd-based or not before kicking migration.
1551			 */
1552			if (!hugepage_migration_supported(folio_hstate(folio))) {
1553				nr_failed++;
1554				stats->nr_failed_pages += nr_pages;
1555				list_move_tail(&folio->lru, ret_folios);
1556				continue;
1557			}
1558
1559			rc = unmap_and_move_huge_page(get_new_folio,
1560						      put_new_folio, private,
1561						      folio, pass > 2, mode,
1562						      reason, ret_folios);
1563			/*
1564			 * The rules are:
1565			 *	Success: hugetlb folio will be put back
1566			 *	-EAGAIN: stay on the from list
1567			 *	-ENOMEM: stay on the from list
1568			 *	Other errno: put on ret_folios list
1569			 */
1570			switch(rc) {
1571			case -ENOMEM:
1572				/*
1573				 * When memory is low, don't bother to try to migrate
1574				 * other folios, just exit.
1575				 */
1576				stats->nr_failed_pages += nr_pages + nr_retry_pages;
1577				return -ENOMEM;
1578			case -EAGAIN:
1579				retry++;
1580				nr_retry_pages += nr_pages;
1581				break;
1582			case MIGRATEPAGE_SUCCESS:
1583				stats->nr_succeeded += nr_pages;
1584				break;
1585			default:
1586				/*
1587				 * Permanent failure (-EBUSY, etc.):
1588				 * unlike -EAGAIN case, the failed folio is
1589				 * removed from migration folio list and not
1590				 * retried in the next outer loop.
1591				 */
1592				nr_failed++;
1593				stats->nr_failed_pages += nr_pages;
1594				break;
1595			}
1596		}
1597	}
1598	/*
1599	 * nr_failed is number of hugetlb folios failed to be migrated.  After
1600	 * NR_MAX_MIGRATE_PAGES_RETRY attempts, give up and count retried hugetlb
1601	 * folios as failed.
1602	 */
1603	nr_failed += retry;
1604	stats->nr_failed_pages += nr_retry_pages;
1605
1606	return nr_failed;
1607}
1608
1609/*
1610 * migrate_pages_batch() first unmaps folios in the from list as many as
1611 * possible, then move the unmapped folios.
1612 *
1613 * We only batch migration if mode == MIGRATE_ASYNC to avoid to wait a
1614 * lock or bit when we have locked more than one folio.  Which may cause
1615 * deadlock (e.g., for loop device).  So, if mode != MIGRATE_ASYNC, the
1616 * length of the from list must be <= 1.
1617 */
1618static int migrate_pages_batch(struct list_head *from,
1619		new_folio_t get_new_folio, free_folio_t put_new_folio,
1620		unsigned long private, enum migrate_mode mode, int reason,
1621		struct list_head *ret_folios, struct list_head *split_folios,
1622		struct migrate_pages_stats *stats, int nr_pass)
1623{
1624	int retry = 1;
1625	int thp_retry = 1;
1626	int nr_failed = 0;
1627	int nr_retry_pages = 0;
1628	int pass = 0;
1629	bool is_thp = false;
1630	bool is_large = false;
1631	struct folio *folio, *folio2, *dst = NULL, *dst2;
1632	int rc, rc_saved = 0, nr_pages;
1633	LIST_HEAD(unmap_folios);
1634	LIST_HEAD(dst_folios);
1635	bool nosplit = (reason == MR_NUMA_MISPLACED);
1636
1637	VM_WARN_ON_ONCE(mode != MIGRATE_ASYNC &&
1638			!list_empty(from) && !list_is_singular(from));
1639
1640	for (pass = 0; pass < nr_pass && retry; pass++) {
1641		retry = 0;
1642		thp_retry = 0;
1643		nr_retry_pages = 0;
1644
1645		list_for_each_entry_safe(folio, folio2, from, lru) {
1646			is_large = folio_test_large(folio);
1647			is_thp = is_large && folio_test_pmd_mappable(folio);
1648			nr_pages = folio_nr_pages(folio);
1649
1650			cond_resched();
1651
1652			/*
1653			 * Large folio migration might be unsupported or
1654			 * the allocation might be failed so we should retry
1655			 * on the same folio with the large folio split
1656			 * to normal folios.
1657			 *
1658			 * Split folios are put in split_folios, and
1659			 * we will migrate them after the rest of the
1660			 * list is processed.
1661			 */
1662			if (!thp_migration_supported() && is_thp) {
1663				nr_failed++;
1664				stats->nr_thp_failed++;
1665				if (!try_split_folio(folio, split_folios)) {
1666					stats->nr_thp_split++;
1667					stats->nr_split++;
1668					continue;
1669				}
1670				stats->nr_failed_pages += nr_pages;
1671				list_move_tail(&folio->lru, ret_folios);
1672				continue;
1673			}
1674
1675			rc = migrate_folio_unmap(get_new_folio, put_new_folio,
1676					private, folio, &dst, mode, reason,
1677					ret_folios);
1678			/*
1679			 * The rules are:
1680			 *	Success: folio will be freed
1681			 *	Unmap: folio will be put on unmap_folios list,
1682			 *	       dst folio put on dst_folios list
1683			 *	-EAGAIN: stay on the from list
1684			 *	-ENOMEM: stay on the from list
1685			 *	Other errno: put on ret_folios list
1686			 */
1687			switch(rc) {
1688			case -ENOMEM:
1689				/*
1690				 * When memory is low, don't bother to try to migrate
1691				 * other folios, move unmapped folios, then exit.
1692				 */
1693				nr_failed++;
1694				stats->nr_thp_failed += is_thp;
1695				/* Large folio NUMA faulting doesn't split to retry. */
1696				if (is_large && !nosplit) {
1697					int ret = try_split_folio(folio, split_folios);
1698
1699					if (!ret) {
1700						stats->nr_thp_split += is_thp;
1701						stats->nr_split++;
1702						break;
1703					} else if (reason == MR_LONGTERM_PIN &&
1704						   ret == -EAGAIN) {
1705						/*
1706						 * Try again to split large folio to
1707						 * mitigate the failure of longterm pinning.
1708						 */
1709						retry++;
1710						thp_retry += is_thp;
1711						nr_retry_pages += nr_pages;
1712						/* Undo duplicated failure counting. */
1713						nr_failed--;
1714						stats->nr_thp_failed -= is_thp;
1715						break;
1716					}
1717				}
1718
1719				stats->nr_failed_pages += nr_pages + nr_retry_pages;
1720				/* nr_failed isn't updated for not used */
1721				stats->nr_thp_failed += thp_retry;
1722				rc_saved = rc;
1723				if (list_empty(&unmap_folios))
1724					goto out;
1725				else
1726					goto move;
1727			case -EAGAIN:
1728				retry++;
1729				thp_retry += is_thp;
1730				nr_retry_pages += nr_pages;
1731				break;
1732			case MIGRATEPAGE_SUCCESS:
1733				stats->nr_succeeded += nr_pages;
1734				stats->nr_thp_succeeded += is_thp;
1735				break;
1736			case MIGRATEPAGE_UNMAP:
1737				list_move_tail(&folio->lru, &unmap_folios);
1738				list_add_tail(&dst->lru, &dst_folios);
1739				break;
1740			default:
1741				/*
1742				 * Permanent failure (-EBUSY, etc.):
1743				 * unlike -EAGAIN case, the failed folio is
1744				 * removed from migration folio list and not
1745				 * retried in the next outer loop.
1746				 */
1747				nr_failed++;
1748				stats->nr_thp_failed += is_thp;
1749				stats->nr_failed_pages += nr_pages;
1750				break;
1751			}
1752		}
1753	}
1754	nr_failed += retry;
1755	stats->nr_thp_failed += thp_retry;
1756	stats->nr_failed_pages += nr_retry_pages;
1757move:
1758	/* Flush TLBs for all unmapped folios */
1759	try_to_unmap_flush();
1760
1761	retry = 1;
1762	for (pass = 0; pass < nr_pass && retry; pass++) {
1763		retry = 0;
1764		thp_retry = 0;
1765		nr_retry_pages = 0;
1766
1767		dst = list_first_entry(&dst_folios, struct folio, lru);
1768		dst2 = list_next_entry(dst, lru);
1769		list_for_each_entry_safe(folio, folio2, &unmap_folios, lru) {
1770			is_thp = folio_test_large(folio) && folio_test_pmd_mappable(folio);
1771			nr_pages = folio_nr_pages(folio);
1772
1773			cond_resched();
1774
1775			rc = migrate_folio_move(put_new_folio, private,
1776						folio, dst, mode,
1777						reason, ret_folios);
1778			/*
1779			 * The rules are:
1780			 *	Success: folio will be freed
1781			 *	-EAGAIN: stay on the unmap_folios list
1782			 *	Other errno: put on ret_folios list
1783			 */
1784			switch(rc) {
1785			case -EAGAIN:
1786				retry++;
1787				thp_retry += is_thp;
1788				nr_retry_pages += nr_pages;
1789				break;
1790			case MIGRATEPAGE_SUCCESS:
1791				stats->nr_succeeded += nr_pages;
1792				stats->nr_thp_succeeded += is_thp;
1793				break;
1794			default:
1795				nr_failed++;
1796				stats->nr_thp_failed += is_thp;
1797				stats->nr_failed_pages += nr_pages;
1798				break;
1799			}
1800			dst = dst2;
1801			dst2 = list_next_entry(dst, lru);
1802		}
1803	}
1804	nr_failed += retry;
1805	stats->nr_thp_failed += thp_retry;
1806	stats->nr_failed_pages += nr_retry_pages;
1807
1808	rc = rc_saved ? : nr_failed;
1809out:
1810	/* Cleanup remaining folios */
1811	dst = list_first_entry(&dst_folios, struct folio, lru);
1812	dst2 = list_next_entry(dst, lru);
1813	list_for_each_entry_safe(folio, folio2, &unmap_folios, lru) {
1814		int old_page_state = 0;
1815		struct anon_vma *anon_vma = NULL;
1816
1817		__migrate_folio_extract(dst, &old_page_state, &anon_vma);
1818		migrate_folio_undo_src(folio, old_page_state & PAGE_WAS_MAPPED,
1819				       anon_vma, true, ret_folios);
1820		list_del(&dst->lru);
1821		migrate_folio_undo_dst(dst, true, put_new_folio, private);
1822		dst = dst2;
1823		dst2 = list_next_entry(dst, lru);
1824	}
1825
1826	return rc;
1827}
1828
1829static int migrate_pages_sync(struct list_head *from, new_folio_t get_new_folio,
1830		free_folio_t put_new_folio, unsigned long private,
1831		enum migrate_mode mode, int reason,
1832		struct list_head *ret_folios, struct list_head *split_folios,
1833		struct migrate_pages_stats *stats)
1834{
1835	int rc, nr_failed = 0;
1836	LIST_HEAD(folios);
1837	struct migrate_pages_stats astats;
1838
1839	memset(&astats, 0, sizeof(astats));
1840	/* Try to migrate in batch with MIGRATE_ASYNC mode firstly */
1841	rc = migrate_pages_batch(from, get_new_folio, put_new_folio, private, MIGRATE_ASYNC,
1842				 reason, &folios, split_folios, &astats,
1843				 NR_MAX_MIGRATE_ASYNC_RETRY);
1844	stats->nr_succeeded += astats.nr_succeeded;
1845	stats->nr_thp_succeeded += astats.nr_thp_succeeded;
1846	stats->nr_thp_split += astats.nr_thp_split;
1847	stats->nr_split += astats.nr_split;
1848	if (rc < 0) {
1849		stats->nr_failed_pages += astats.nr_failed_pages;
1850		stats->nr_thp_failed += astats.nr_thp_failed;
1851		list_splice_tail(&folios, ret_folios);
1852		return rc;
1853	}
1854	stats->nr_thp_failed += astats.nr_thp_split;
1855	/*
1856	 * Do not count rc, as pages will be retried below.
1857	 * Count nr_split only, since it includes nr_thp_split.
1858	 */
1859	nr_failed += astats.nr_split;
1860	/*
1861	 * Fall back to migrate all failed folios one by one synchronously. All
1862	 * failed folios except split THPs will be retried, so their failure
1863	 * isn't counted
1864	 */
1865	list_splice_tail_init(&folios, from);
1866	while (!list_empty(from)) {
1867		list_move(from->next, &folios);
1868		rc = migrate_pages_batch(&folios, get_new_folio, put_new_folio,
1869					 private, mode, reason, ret_folios,
1870					 split_folios, stats, NR_MAX_MIGRATE_SYNC_RETRY);
1871		list_splice_tail_init(&folios, ret_folios);
1872		if (rc < 0)
1873			return rc;
1874		nr_failed += rc;
1875	}
1876
1877	return nr_failed;
1878}
1879
1880/*
1881 * migrate_pages - migrate the folios specified in a list, to the free folios
1882 *		   supplied as the target for the page migration
1883 *
1884 * @from:		The list of folios to be migrated.
1885 * @get_new_folio:	The function used to allocate free folios to be used
1886 *			as the target of the folio migration.
1887 * @put_new_folio:	The function used to free target folios if migration
1888 *			fails, or NULL if no special handling is necessary.
1889 * @private:		Private data to be passed on to get_new_folio()
1890 * @mode:		The migration mode that specifies the constraints for
1891 *			folio migration, if any.
1892 * @reason:		The reason for folio migration.
1893 * @ret_succeeded:	Set to the number of folios migrated successfully if
1894 *			the caller passes a non-NULL pointer.
1895 *
1896 * The function returns after NR_MAX_MIGRATE_PAGES_RETRY attempts or if no folios
1897 * are movable any more because the list has become empty or no retryable folios
1898 * exist any more. It is caller's responsibility to call putback_movable_pages()
1899 * only if ret != 0.
1900 *
1901 * Returns the number of {normal folio, large folio, hugetlb} that were not
1902 * migrated, or an error code. The number of large folio splits will be
1903 * considered as the number of non-migrated large folio, no matter how many
1904 * split folios of the large folio are migrated successfully.
1905 */
1906int migrate_pages(struct list_head *from, new_folio_t get_new_folio,
1907		free_folio_t put_new_folio, unsigned long private,
1908		enum migrate_mode mode, int reason, unsigned int *ret_succeeded)
1909{
1910	int rc, rc_gather;
1911	int nr_pages;
1912	struct folio *folio, *folio2;
1913	LIST_HEAD(folios);
1914	LIST_HEAD(ret_folios);
1915	LIST_HEAD(split_folios);
1916	struct migrate_pages_stats stats;
1917
1918	trace_mm_migrate_pages_start(mode, reason);
1919
1920	memset(&stats, 0, sizeof(stats));
1921
1922	rc_gather = migrate_hugetlbs(from, get_new_folio, put_new_folio, private,
1923				     mode, reason, &stats, &ret_folios);
1924	if (rc_gather < 0)
1925		goto out;
1926
1927again:
1928	nr_pages = 0;
1929	list_for_each_entry_safe(folio, folio2, from, lru) {
1930		/* Retried hugetlb folios will be kept in list  */
1931		if (folio_test_hugetlb(folio)) {
1932			list_move_tail(&folio->lru, &ret_folios);
1933			continue;
1934		}
1935
1936		nr_pages += folio_nr_pages(folio);
1937		if (nr_pages >= NR_MAX_BATCHED_MIGRATION)
1938			break;
1939	}
1940	if (nr_pages >= NR_MAX_BATCHED_MIGRATION)
1941		list_cut_before(&folios, from, &folio2->lru);
1942	else
1943		list_splice_init(from, &folios);
1944	if (mode == MIGRATE_ASYNC)
1945		rc = migrate_pages_batch(&folios, get_new_folio, put_new_folio,
1946				private, mode, reason, &ret_folios,
1947				&split_folios, &stats,
1948				NR_MAX_MIGRATE_PAGES_RETRY);
1949	else
1950		rc = migrate_pages_sync(&folios, get_new_folio, put_new_folio,
1951				private, mode, reason, &ret_folios,
1952				&split_folios, &stats);
1953	list_splice_tail_init(&folios, &ret_folios);
1954	if (rc < 0) {
1955		rc_gather = rc;
1956		list_splice_tail(&split_folios, &ret_folios);
1957		goto out;
1958	}
1959	if (!list_empty(&split_folios)) {
1960		/*
1961		 * Failure isn't counted since all split folios of a large folio
1962		 * is counted as 1 failure already.  And, we only try to migrate
1963		 * with minimal effort, force MIGRATE_ASYNC mode and retry once.
1964		 */
1965		migrate_pages_batch(&split_folios, get_new_folio,
1966				put_new_folio, private, MIGRATE_ASYNC, reason,
1967				&ret_folios, NULL, &stats, 1);
1968		list_splice_tail_init(&split_folios, &ret_folios);
1969	}
1970	rc_gather += rc;
1971	if (!list_empty(from))
1972		goto again;
1973out:
1974	/*
1975	 * Put the permanent failure folio back to migration list, they
1976	 * will be put back to the right list by the caller.
1977	 */
1978	list_splice(&ret_folios, from);
1979
1980	/*
1981	 * Return 0 in case all split folios of fail-to-migrate large folios
1982	 * are migrated successfully.
1983	 */
1984	if (list_empty(from))
1985		rc_gather = 0;
1986
1987	count_vm_events(PGMIGRATE_SUCCESS, stats.nr_succeeded);
1988	count_vm_events(PGMIGRATE_FAIL, stats.nr_failed_pages);
1989	count_vm_events(THP_MIGRATION_SUCCESS, stats.nr_thp_succeeded);
1990	count_vm_events(THP_MIGRATION_FAIL, stats.nr_thp_failed);
1991	count_vm_events(THP_MIGRATION_SPLIT, stats.nr_thp_split);
1992	trace_mm_migrate_pages(stats.nr_succeeded, stats.nr_failed_pages,
1993			       stats.nr_thp_succeeded, stats.nr_thp_failed,
1994			       stats.nr_thp_split, stats.nr_split, mode,
1995			       reason);
1996
1997	if (ret_succeeded)
1998		*ret_succeeded = stats.nr_succeeded;
1999
2000	return rc_gather;
2001}
2002
2003struct folio *alloc_migration_target(struct folio *src, unsigned long private)
2004{
2005	struct migration_target_control *mtc;
2006	gfp_t gfp_mask;
2007	unsigned int order = 0;
2008	int nid;
2009	int zidx;
2010
2011	mtc = (struct migration_target_control *)private;
2012	gfp_mask = mtc->gfp_mask;
2013	nid = mtc->nid;
2014	if (nid == NUMA_NO_NODE)
2015		nid = folio_nid(src);
2016
2017	if (folio_test_hugetlb(src)) {
2018		struct hstate *h = folio_hstate(src);
2019
2020		gfp_mask = htlb_modify_alloc_mask(h, gfp_mask);
2021		return alloc_hugetlb_folio_nodemask(h, nid,
2022						mtc->nmask, gfp_mask);
2023	}
2024
2025	if (folio_test_large(src)) {
2026		/*
2027		 * clear __GFP_RECLAIM to make the migration callback
2028		 * consistent with regular THP allocations.
2029		 */
2030		gfp_mask &= ~__GFP_RECLAIM;
2031		gfp_mask |= GFP_TRANSHUGE;
2032		order = folio_order(src);
2033	}
2034	zidx = zone_idx(folio_zone(src));
2035	if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE)
2036		gfp_mask |= __GFP_HIGHMEM;
2037
2038	return __folio_alloc(gfp_mask, order, nid, mtc->nmask);
2039}
2040
2041#ifdef CONFIG_NUMA
2042
2043static int store_status(int __user *status, int start, int value, int nr)
2044{
2045	while (nr-- > 0) {
2046		if (put_user(value, status + start))
2047			return -EFAULT;
2048		start++;
2049	}
2050
2051	return 0;
2052}
2053
2054static int do_move_pages_to_node(struct list_head *pagelist, int node)
2055{
2056	int err;
2057	struct migration_target_control mtc = {
2058		.nid = node,
2059		.gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
2060	};
2061
2062	err = migrate_pages(pagelist, alloc_migration_target, NULL,
2063		(unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
2064	if (err)
2065		putback_movable_pages(pagelist);
2066	return err;
2067}
2068
2069/*
2070 * Resolves the given address to a struct page, isolates it from the LRU and
2071 * puts it to the given pagelist.
2072 * Returns:
2073 *     errno - if the page cannot be found/isolated
2074 *     0 - when it doesn't have to be migrated because it is already on the
2075 *         target node
2076 *     1 - when it has been queued
2077 */
2078static int add_page_for_migration(struct mm_struct *mm, const void __user *p,
2079		int node, struct list_head *pagelist, bool migrate_all)
2080{
2081	struct vm_area_struct *vma;
2082	unsigned long addr;
2083	struct page *page;
2084	struct folio *folio;
2085	int err;
2086
2087	mmap_read_lock(mm);
2088	addr = (unsigned long)untagged_addr_remote(mm, p);
2089
2090	err = -EFAULT;
2091	vma = vma_lookup(mm, addr);
2092	if (!vma || !vma_migratable(vma))
2093		goto out;
2094
2095	/* FOLL_DUMP to ignore special (like zero) pages */
2096	page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
2097
2098	err = PTR_ERR(page);
2099	if (IS_ERR(page))
2100		goto out;
2101
2102	err = -ENOENT;
2103	if (!page)
2104		goto out;
2105
2106	folio = page_folio(page);
2107	if (folio_is_zone_device(folio))
2108		goto out_putfolio;
2109
2110	err = 0;
2111	if (folio_nid(folio) == node)
2112		goto out_putfolio;
2113
2114	err = -EACCES;
2115	if (page_mapcount(page) > 1 && !migrate_all)
2116		goto out_putfolio;
2117
2118	err = -EBUSY;
2119	if (folio_test_hugetlb(folio)) {
2120		if (isolate_hugetlb(folio, pagelist))
2121			err = 1;
2122	} else {
2123		if (!folio_isolate_lru(folio))
2124			goto out_putfolio;
2125
2126		err = 1;
2127		list_add_tail(&folio->lru, pagelist);
2128		node_stat_mod_folio(folio,
2129			NR_ISOLATED_ANON + folio_is_file_lru(folio),
2130			folio_nr_pages(folio));
2131	}
2132out_putfolio:
2133	/*
2134	 * Either remove the duplicate refcount from folio_isolate_lru()
2135	 * or drop the folio ref if it was not isolated.
2136	 */
2137	folio_put(folio);
2138out:
2139	mmap_read_unlock(mm);
2140	return err;
2141}
2142
2143static int move_pages_and_store_status(int node,
2144		struct list_head *pagelist, int __user *status,
2145		int start, int i, unsigned long nr_pages)
2146{
2147	int err;
2148
2149	if (list_empty(pagelist))
2150		return 0;
2151
2152	err = do_move_pages_to_node(pagelist, node);
2153	if (err) {
2154		/*
2155		 * Positive err means the number of failed
2156		 * pages to migrate.  Since we are going to
2157		 * abort and return the number of non-migrated
2158		 * pages, so need to include the rest of the
2159		 * nr_pages that have not been attempted as
2160		 * well.
2161		 */
2162		if (err > 0)
2163			err += nr_pages - i;
2164		return err;
2165	}
2166	return store_status(status, start, node, i - start);
2167}
2168
2169/*
2170 * Migrate an array of page address onto an array of nodes and fill
2171 * the corresponding array of status.
2172 */
2173static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
2174			 unsigned long nr_pages,
2175			 const void __user * __user *pages,
2176			 const int __user *nodes,
2177			 int __user *status, int flags)
2178{
2179	compat_uptr_t __user *compat_pages = (void __user *)pages;
2180	int current_node = NUMA_NO_NODE;
2181	LIST_HEAD(pagelist);
2182	int start, i;
2183	int err = 0, err1;
2184
2185	lru_cache_disable();
2186
2187	for (i = start = 0; i < nr_pages; i++) {
2188		const void __user *p;
2189		int node;
2190
2191		err = -EFAULT;
2192		if (in_compat_syscall()) {
2193			compat_uptr_t cp;
2194
2195			if (get_user(cp, compat_pages + i))
2196				goto out_flush;
2197
2198			p = compat_ptr(cp);
2199		} else {
2200			if (get_user(p, pages + i))
2201				goto out_flush;
2202		}
2203		if (get_user(node, nodes + i))
2204			goto out_flush;
2205
2206		err = -ENODEV;
2207		if (node < 0 || node >= MAX_NUMNODES)
2208			goto out_flush;
2209		if (!node_state(node, N_MEMORY))
2210			goto out_flush;
2211
2212		err = -EACCES;
2213		if (!node_isset(node, task_nodes))
2214			goto out_flush;
2215
2216		if (current_node == NUMA_NO_NODE) {
2217			current_node = node;
2218			start = i;
2219		} else if (node != current_node) {
2220			err = move_pages_and_store_status(current_node,
2221					&pagelist, status, start, i, nr_pages);
2222			if (err)
2223				goto out;
2224			start = i;
2225			current_node = node;
2226		}
2227
2228		/*
2229		 * Errors in the page lookup or isolation are not fatal and we simply
2230		 * report them via status
2231		 */
2232		err = add_page_for_migration(mm, p, current_node, &pagelist,
2233					     flags & MPOL_MF_MOVE_ALL);
2234
2235		if (err > 0) {
2236			/* The page is successfully queued for migration */
2237			continue;
2238		}
2239
2240		/*
2241		 * The move_pages() man page does not have an -EEXIST choice, so
2242		 * use -EFAULT instead.
2243		 */
2244		if (err == -EEXIST)
2245			err = -EFAULT;
2246
2247		/*
2248		 * If the page is already on the target node (!err), store the
2249		 * node, otherwise, store the err.
2250		 */
2251		err = store_status(status, i, err ? : current_node, 1);
2252		if (err)
2253			goto out_flush;
2254
2255		err = move_pages_and_store_status(current_node, &pagelist,
2256				status, start, i, nr_pages);
2257		if (err) {
2258			/* We have accounted for page i */
2259			if (err > 0)
2260				err--;
2261			goto out;
2262		}
2263		current_node = NUMA_NO_NODE;
2264	}
2265out_flush:
2266	/* Make sure we do not overwrite the existing error */
2267	err1 = move_pages_and_store_status(current_node, &pagelist,
2268				status, start, i, nr_pages);
2269	if (err >= 0)
2270		err = err1;
2271out:
2272	lru_cache_enable();
2273	return err;
2274}
2275
2276/*
2277 * Determine the nodes of an array of pages and store it in an array of status.
2278 */
2279static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
2280				const void __user **pages, int *status)
2281{
2282	unsigned long i;
2283
2284	mmap_read_lock(mm);
2285
2286	for (i = 0; i < nr_pages; i++) {
2287		unsigned long addr = (unsigned long)(*pages);
2288		struct vm_area_struct *vma;
2289		struct page *page;
2290		int err = -EFAULT;
2291
2292		vma = vma_lookup(mm, addr);
2293		if (!vma)
2294			goto set_status;
2295
2296		/* FOLL_DUMP to ignore special (like zero) pages */
2297		page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
2298
2299		err = PTR_ERR(page);
2300		if (IS_ERR(page))
2301			goto set_status;
2302
2303		err = -ENOENT;
2304		if (!page)
2305			goto set_status;
2306
2307		if (!is_zone_device_page(page))
2308			err = page_to_nid(page);
2309
2310		put_page(page);
2311set_status:
2312		*status = err;
2313
2314		pages++;
2315		status++;
2316	}
2317
2318	mmap_read_unlock(mm);
2319}
2320
2321static int get_compat_pages_array(const void __user *chunk_pages[],
2322				  const void __user * __user *pages,
2323				  unsigned long chunk_nr)
2324{
2325	compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages;
2326	compat_uptr_t p;
2327	int i;
2328
2329	for (i = 0; i < chunk_nr; i++) {
2330		if (get_user(p, pages32 + i))
2331			return -EFAULT;
2332		chunk_pages[i] = compat_ptr(p);
2333	}
2334
2335	return 0;
2336}
2337
2338/*
2339 * Determine the nodes of a user array of pages and store it in
2340 * a user array of status.
2341 */
2342static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
2343			 const void __user * __user *pages,
2344			 int __user *status)
2345{
2346#define DO_PAGES_STAT_CHUNK_NR 16UL
2347	const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
2348	int chunk_status[DO_PAGES_STAT_CHUNK_NR];
2349
2350	while (nr_pages) {
2351		unsigned long chunk_nr = min(nr_pages, DO_PAGES_STAT_CHUNK_NR);
2352
2353		if (in_compat_syscall()) {
2354			if (get_compat_pages_array(chunk_pages, pages,
2355						   chunk_nr))
2356				break;
2357		} else {
2358			if (copy_from_user(chunk_pages, pages,
2359				      chunk_nr * sizeof(*chunk_pages)))
2360				break;
2361		}
2362
2363		do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
2364
2365		if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
2366			break;
2367
2368		pages += chunk_nr;
2369		status += chunk_nr;
2370		nr_pages -= chunk_nr;
2371	}
2372	return nr_pages ? -EFAULT : 0;
2373}
2374
2375static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes)
2376{
2377	struct task_struct *task;
2378	struct mm_struct *mm;
2379
2380	/*
2381	 * There is no need to check if current process has the right to modify
2382	 * the specified process when they are same.
2383	 */
2384	if (!pid) {
2385		mmget(current->mm);
2386		*mem_nodes = cpuset_mems_allowed(current);
2387		return current->mm;
2388	}
2389
2390	/* Find the mm_struct */
2391	rcu_read_lock();
2392	task = find_task_by_vpid(pid);
2393	if (!task) {
2394		rcu_read_unlock();
2395		return ERR_PTR(-ESRCH);
2396	}
2397	get_task_struct(task);
2398
2399	/*
2400	 * Check if this process has the right to modify the specified
2401	 * process. Use the regular "ptrace_may_access()" checks.
2402	 */
2403	if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
2404		rcu_read_unlock();
2405		mm = ERR_PTR(-EPERM);
2406		goto out;
2407	}
2408	rcu_read_unlock();
2409
2410	mm = ERR_PTR(security_task_movememory(task));
2411	if (IS_ERR(mm))
2412		goto out;
2413	*mem_nodes = cpuset_mems_allowed(task);
2414	mm = get_task_mm(task);
2415out:
2416	put_task_struct(task);
2417	if (!mm)
2418		mm = ERR_PTR(-EINVAL);
2419	return mm;
2420}
2421
2422/*
2423 * Move a list of pages in the address space of the currently executing
2424 * process.
2425 */
2426static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
2427			     const void __user * __user *pages,
2428			     const int __user *nodes,
2429			     int __user *status, int flags)
2430{
2431	struct mm_struct *mm;
2432	int err;
2433	nodemask_t task_nodes;
2434
2435	/* Check flags */
2436	if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
2437		return -EINVAL;
2438
2439	if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
2440		return -EPERM;
2441
2442	mm = find_mm_struct(pid, &task_nodes);
2443	if (IS_ERR(mm))
2444		return PTR_ERR(mm);
2445
2446	if (nodes)
2447		err = do_pages_move(mm, task_nodes, nr_pages, pages,
2448				    nodes, status, flags);
2449	else
2450		err = do_pages_stat(mm, nr_pages, pages, status);
2451
2452	mmput(mm);
2453	return err;
2454}
2455
2456SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
2457		const void __user * __user *, pages,
2458		const int __user *, nodes,
2459		int __user *, status, int, flags)
2460{
2461	return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
2462}
2463
2464#ifdef CONFIG_NUMA_BALANCING
2465/*
2466 * Returns true if this is a safe migration target node for misplaced NUMA
2467 * pages. Currently it only checks the watermarks which is crude.
2468 */
2469static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
2470				   unsigned long nr_migrate_pages)
2471{
2472	int z;
2473
2474	for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2475		struct zone *zone = pgdat->node_zones + z;
2476
2477		if (!managed_zone(zone))
2478			continue;
2479
2480		/* Avoid waking kswapd by allocating pages_to_migrate pages. */
2481		if (!zone_watermark_ok(zone, 0,
2482				       high_wmark_pages(zone) +
2483				       nr_migrate_pages,
2484				       ZONE_MOVABLE, 0))
2485			continue;
2486		return true;
2487	}
2488	return false;
2489}
2490
2491static struct folio *alloc_misplaced_dst_folio(struct folio *src,
2492					   unsigned long data)
2493{
2494	int nid = (int) data;
2495	int order = folio_order(src);
2496	gfp_t gfp = __GFP_THISNODE;
2497
2498	if (order > 0)
2499		gfp |= GFP_TRANSHUGE_LIGHT;
2500	else {
2501		gfp |= GFP_HIGHUSER_MOVABLE | __GFP_NOMEMALLOC | __GFP_NORETRY |
2502			__GFP_NOWARN;
2503		gfp &= ~__GFP_RECLAIM;
2504	}
2505	return __folio_alloc_node(gfp, order, nid);
2506}
2507
2508static int numamigrate_isolate_folio(pg_data_t *pgdat, struct folio *folio)
2509{
2510	int nr_pages = folio_nr_pages(folio);
2511
2512	/* Avoid migrating to a node that is nearly full */
2513	if (!migrate_balanced_pgdat(pgdat, nr_pages)) {
2514		int z;
2515
2516		if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING))
2517			return 0;
2518		for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2519			if (managed_zone(pgdat->node_zones + z))
2520				break;
2521		}
2522
2523		/*
2524		 * If there are no managed zones, it should not proceed
2525		 * further.
2526		 */
2527		if (z < 0)
2528			return 0;
2529
2530		wakeup_kswapd(pgdat->node_zones + z, 0,
2531			      folio_order(folio), ZONE_MOVABLE);
2532		return 0;
2533	}
2534
2535	if (!folio_isolate_lru(folio))
2536		return 0;
2537
2538	node_stat_mod_folio(folio, NR_ISOLATED_ANON + folio_is_file_lru(folio),
2539			    nr_pages);
2540
2541	/*
2542	 * Isolating the folio has taken another reference, so the
2543	 * caller's reference can be safely dropped without the folio
2544	 * disappearing underneath us during migration.
2545	 */
2546	folio_put(folio);
2547	return 1;
2548}
2549
2550/*
2551 * Attempt to migrate a misplaced folio to the specified destination
2552 * node. Caller is expected to have an elevated reference count on
2553 * the folio that will be dropped by this function before returning.
2554 */
2555int migrate_misplaced_folio(struct folio *folio, struct vm_area_struct *vma,
2556			    int node)
2557{
2558	pg_data_t *pgdat = NODE_DATA(node);
2559	int isolated;
2560	int nr_remaining;
2561	unsigned int nr_succeeded;
2562	LIST_HEAD(migratepages);
2563	int nr_pages = folio_nr_pages(folio);
2564
2565	/*
2566	 * Don't migrate file folios that are mapped in multiple processes
2567	 * with execute permissions as they are probably shared libraries.
2568	 * To check if the folio is shared, ideally we want to make sure
2569	 * every page is mapped to the same process. Doing that is very
2570	 * expensive, so check the estimated mapcount of the folio instead.
2571	 */
2572	if (folio_estimated_sharers(folio) != 1 && folio_is_file_lru(folio) &&
2573	    (vma->vm_flags & VM_EXEC))
2574		goto out;
2575
2576	/*
2577	 * Also do not migrate dirty folios as not all filesystems can move
2578	 * dirty folios in MIGRATE_ASYNC mode which is a waste of cycles.
2579	 */
2580	if (folio_is_file_lru(folio) && folio_test_dirty(folio))
2581		goto out;
2582
2583	isolated = numamigrate_isolate_folio(pgdat, folio);
2584	if (!isolated)
2585		goto out;
2586
2587	list_add(&folio->lru, &migratepages);
2588	nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_folio,
2589				     NULL, node, MIGRATE_ASYNC,
2590				     MR_NUMA_MISPLACED, &nr_succeeded);
2591	if (nr_remaining) {
2592		if (!list_empty(&migratepages)) {
2593			list_del(&folio->lru);
2594			node_stat_mod_folio(folio, NR_ISOLATED_ANON +
2595					folio_is_file_lru(folio), -nr_pages);
2596			folio_putback_lru(folio);
2597		}
2598		isolated = 0;
2599	}
2600	if (nr_succeeded) {
2601		count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_succeeded);
2602		if (!node_is_toptier(folio_nid(folio)) && node_is_toptier(node))
2603			mod_node_page_state(pgdat, PGPROMOTE_SUCCESS,
2604					    nr_succeeded);
2605	}
2606	BUG_ON(!list_empty(&migratepages));
2607	return isolated;
2608
2609out:
2610	folio_put(folio);
2611	return 0;
2612}
2613#endif /* CONFIG_NUMA_BALANCING */
2614#endif /* CONFIG_NUMA */
v6.9.4
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Memory Migration functionality - linux/mm/migrate.c
   4 *
   5 * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter
   6 *
   7 * Page migration was first developed in the context of the memory hotplug
   8 * project. The main authors of the migration code are:
   9 *
  10 * IWAMOTO Toshihiro <iwamoto@valinux.co.jp>
  11 * Hirokazu Takahashi <taka@valinux.co.jp>
  12 * Dave Hansen <haveblue@us.ibm.com>
  13 * Christoph Lameter
  14 */
  15
  16#include <linux/migrate.h>
  17#include <linux/export.h>
  18#include <linux/swap.h>
  19#include <linux/swapops.h>
  20#include <linux/pagemap.h>
  21#include <linux/buffer_head.h>
  22#include <linux/mm_inline.h>
  23#include <linux/nsproxy.h>
  24#include <linux/ksm.h>
  25#include <linux/rmap.h>
  26#include <linux/topology.h>
  27#include <linux/cpu.h>
  28#include <linux/cpuset.h>
  29#include <linux/writeback.h>
  30#include <linux/mempolicy.h>
  31#include <linux/vmalloc.h>
  32#include <linux/security.h>
  33#include <linux/backing-dev.h>
  34#include <linux/compaction.h>
  35#include <linux/syscalls.h>
  36#include <linux/compat.h>
  37#include <linux/hugetlb.h>
  38#include <linux/hugetlb_cgroup.h>
  39#include <linux/gfp.h>
  40#include <linux/pfn_t.h>
  41#include <linux/memremap.h>
  42#include <linux/userfaultfd_k.h>
  43#include <linux/balloon_compaction.h>
  44#include <linux/page_idle.h>
  45#include <linux/page_owner.h>
  46#include <linux/sched/mm.h>
  47#include <linux/ptrace.h>
  48#include <linux/oom.h>
  49#include <linux/memory.h>
  50#include <linux/random.h>
  51#include <linux/sched/sysctl.h>
  52#include <linux/memory-tiers.h>
  53
  54#include <asm/tlbflush.h>
  55
  56#include <trace/events/migrate.h>
  57
  58#include "internal.h"
  59
  60bool isolate_movable_page(struct page *page, isolate_mode_t mode)
  61{
  62	struct folio *folio = folio_get_nontail_page(page);
  63	const struct movable_operations *mops;
  64
  65	/*
  66	 * Avoid burning cycles with pages that are yet under __free_pages(),
  67	 * or just got freed under us.
  68	 *
  69	 * In case we 'win' a race for a movable page being freed under us and
  70	 * raise its refcount preventing __free_pages() from doing its job
  71	 * the put_page() at the end of this block will take care of
  72	 * release this page, thus avoiding a nasty leakage.
  73	 */
  74	if (!folio)
  75		goto out;
  76
  77	if (unlikely(folio_test_slab(folio)))
  78		goto out_putfolio;
  79	/* Pairs with smp_wmb() in slab freeing, e.g. SLUB's __free_slab() */
  80	smp_rmb();
  81	/*
  82	 * Check movable flag before taking the page lock because
  83	 * we use non-atomic bitops on newly allocated page flags so
  84	 * unconditionally grabbing the lock ruins page's owner side.
  85	 */
  86	if (unlikely(!__folio_test_movable(folio)))
  87		goto out_putfolio;
  88	/* Pairs with smp_wmb() in slab allocation, e.g. SLUB's alloc_slab_page() */
  89	smp_rmb();
  90	if (unlikely(folio_test_slab(folio)))
  91		goto out_putfolio;
  92
  93	/*
  94	 * As movable pages are not isolated from LRU lists, concurrent
  95	 * compaction threads can race against page migration functions
  96	 * as well as race against the releasing a page.
  97	 *
  98	 * In order to avoid having an already isolated movable page
  99	 * being (wrongly) re-isolated while it is under migration,
 100	 * or to avoid attempting to isolate pages being released,
 101	 * lets be sure we have the page lock
 102	 * before proceeding with the movable page isolation steps.
 103	 */
 104	if (unlikely(!folio_trylock(folio)))
 105		goto out_putfolio;
 106
 107	if (!folio_test_movable(folio) || folio_test_isolated(folio))
 108		goto out_no_isolated;
 109
 110	mops = folio_movable_ops(folio);
 111	VM_BUG_ON_FOLIO(!mops, folio);
 112
 113	if (!mops->isolate_page(&folio->page, mode))
 114		goto out_no_isolated;
 115
 116	/* Driver shouldn't use PG_isolated bit of page->flags */
 117	WARN_ON_ONCE(folio_test_isolated(folio));
 118	folio_set_isolated(folio);
 119	folio_unlock(folio);
 120
 121	return true;
 122
 123out_no_isolated:
 124	folio_unlock(folio);
 125out_putfolio:
 126	folio_put(folio);
 127out:
 128	return false;
 129}
 130
 131static void putback_movable_folio(struct folio *folio)
 132{
 133	const struct movable_operations *mops = folio_movable_ops(folio);
 134
 135	mops->putback_page(&folio->page);
 136	folio_clear_isolated(folio);
 137}
 138
 139/*
 140 * Put previously isolated pages back onto the appropriate lists
 141 * from where they were once taken off for compaction/migration.
 142 *
 143 * This function shall be used whenever the isolated pageset has been
 144 * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
 145 * and isolate_hugetlb().
 146 */
 147void putback_movable_pages(struct list_head *l)
 148{
 149	struct folio *folio;
 150	struct folio *folio2;
 151
 152	list_for_each_entry_safe(folio, folio2, l, lru) {
 153		if (unlikely(folio_test_hugetlb(folio))) {
 154			folio_putback_active_hugetlb(folio);
 155			continue;
 156		}
 157		list_del(&folio->lru);
 158		/*
 159		 * We isolated non-lru movable folio so here we can use
 160		 * __folio_test_movable because LRU folio's mapping cannot
 161		 * have PAGE_MAPPING_MOVABLE.
 162		 */
 163		if (unlikely(__folio_test_movable(folio))) {
 164			VM_BUG_ON_FOLIO(!folio_test_isolated(folio), folio);
 165			folio_lock(folio);
 166			if (folio_test_movable(folio))
 167				putback_movable_folio(folio);
 168			else
 169				folio_clear_isolated(folio);
 170			folio_unlock(folio);
 171			folio_put(folio);
 172		} else {
 173			node_stat_mod_folio(folio, NR_ISOLATED_ANON +
 174					folio_is_file_lru(folio), -folio_nr_pages(folio));
 175			folio_putback_lru(folio);
 176		}
 177	}
 178}
 179
 180/*
 181 * Restore a potential migration pte to a working pte entry
 182 */
 183static bool remove_migration_pte(struct folio *folio,
 184		struct vm_area_struct *vma, unsigned long addr, void *old)
 185{
 186	DEFINE_FOLIO_VMA_WALK(pvmw, old, vma, addr, PVMW_SYNC | PVMW_MIGRATION);
 187
 188	while (page_vma_mapped_walk(&pvmw)) {
 189		rmap_t rmap_flags = RMAP_NONE;
 190		pte_t old_pte;
 191		pte_t pte;
 192		swp_entry_t entry;
 193		struct page *new;
 194		unsigned long idx = 0;
 195
 196		/* pgoff is invalid for ksm pages, but they are never large */
 197		if (folio_test_large(folio) && !folio_test_hugetlb(folio))
 198			idx = linear_page_index(vma, pvmw.address) - pvmw.pgoff;
 199		new = folio_page(folio, idx);
 200
 201#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
 202		/* PMD-mapped THP migration entry */
 203		if (!pvmw.pte) {
 204			VM_BUG_ON_FOLIO(folio_test_hugetlb(folio) ||
 205					!folio_test_pmd_mappable(folio), folio);
 206			remove_migration_pmd(&pvmw, new);
 207			continue;
 208		}
 209#endif
 210
 211		folio_get(folio);
 212		pte = mk_pte(new, READ_ONCE(vma->vm_page_prot));
 213		old_pte = ptep_get(pvmw.pte);
 
 
 214
 215		entry = pte_to_swp_entry(old_pte);
 216		if (!is_migration_entry_young(entry))
 217			pte = pte_mkold(pte);
 218		if (folio_test_dirty(folio) && is_migration_entry_dirty(entry))
 219			pte = pte_mkdirty(pte);
 220		if (pte_swp_soft_dirty(old_pte))
 221			pte = pte_mksoft_dirty(pte);
 222		else
 223			pte = pte_clear_soft_dirty(pte);
 224
 225		if (is_writable_migration_entry(entry))
 226			pte = pte_mkwrite(pte, vma);
 227		else if (pte_swp_uffd_wp(old_pte))
 228			pte = pte_mkuffd_wp(pte);
 229
 230		if (folio_test_anon(folio) && !is_readable_migration_entry(entry))
 231			rmap_flags |= RMAP_EXCLUSIVE;
 232
 233		if (unlikely(is_device_private_page(new))) {
 234			if (pte_write(pte))
 235				entry = make_writable_device_private_entry(
 236							page_to_pfn(new));
 237			else
 238				entry = make_readable_device_private_entry(
 239							page_to_pfn(new));
 240			pte = swp_entry_to_pte(entry);
 241			if (pte_swp_soft_dirty(old_pte))
 242				pte = pte_swp_mksoft_dirty(pte);
 243			if (pte_swp_uffd_wp(old_pte))
 244				pte = pte_swp_mkuffd_wp(pte);
 245		}
 246
 247#ifdef CONFIG_HUGETLB_PAGE
 248		if (folio_test_hugetlb(folio)) {
 249			struct hstate *h = hstate_vma(vma);
 250			unsigned int shift = huge_page_shift(h);
 251			unsigned long psize = huge_page_size(h);
 252
 253			pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
 254			if (folio_test_anon(folio))
 255				hugetlb_add_anon_rmap(folio, vma, pvmw.address,
 256						      rmap_flags);
 257			else
 258				hugetlb_add_file_rmap(folio);
 259			set_huge_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte,
 260					psize);
 261		} else
 262#endif
 263		{
 264			if (folio_test_anon(folio))
 265				folio_add_anon_rmap_pte(folio, new, vma,
 266							pvmw.address, rmap_flags);
 267			else
 268				folio_add_file_rmap_pte(folio, new, vma);
 269			set_pte_at(vma->vm_mm, pvmw.address, pvmw.pte, pte);
 270		}
 271		if (vma->vm_flags & VM_LOCKED)
 272			mlock_drain_local();
 273
 274		trace_remove_migration_pte(pvmw.address, pte_val(pte),
 275					   compound_order(new));
 276
 277		/* No need to invalidate - it was non-present before */
 278		update_mmu_cache(vma, pvmw.address, pvmw.pte);
 279	}
 280
 281	return true;
 282}
 283
 284/*
 285 * Get rid of all migration entries and replace them by
 286 * references to the indicated page.
 287 */
 288void remove_migration_ptes(struct folio *src, struct folio *dst, bool locked)
 289{
 290	struct rmap_walk_control rwc = {
 291		.rmap_one = remove_migration_pte,
 292		.arg = src,
 293	};
 294
 295	if (locked)
 296		rmap_walk_locked(dst, &rwc);
 297	else
 298		rmap_walk(dst, &rwc);
 299}
 300
 301/*
 302 * Something used the pte of a page under migration. We need to
 303 * get to the page and wait until migration is finished.
 304 * When we return from this function the fault will be retried.
 305 */
 306void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
 307			  unsigned long address)
 308{
 309	spinlock_t *ptl;
 310	pte_t *ptep;
 311	pte_t pte;
 312	swp_entry_t entry;
 313
 314	ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
 315	if (!ptep)
 316		return;
 317
 318	pte = ptep_get(ptep);
 319	pte_unmap(ptep);
 320
 321	if (!is_swap_pte(pte))
 322		goto out;
 323
 324	entry = pte_to_swp_entry(pte);
 325	if (!is_migration_entry(entry))
 326		goto out;
 327
 328	migration_entry_wait_on_locked(entry, ptl);
 329	return;
 330out:
 331	spin_unlock(ptl);
 332}
 333
 334#ifdef CONFIG_HUGETLB_PAGE
 335/*
 336 * The vma read lock must be held upon entry. Holding that lock prevents either
 337 * the pte or the ptl from being freed.
 338 *
 339 * This function will release the vma lock before returning.
 340 */
 341void migration_entry_wait_huge(struct vm_area_struct *vma, pte_t *ptep)
 342{
 343	spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), vma->vm_mm, ptep);
 344	pte_t pte;
 345
 346	hugetlb_vma_assert_locked(vma);
 347	spin_lock(ptl);
 348	pte = huge_ptep_get(ptep);
 349
 350	if (unlikely(!is_hugetlb_entry_migration(pte))) {
 351		spin_unlock(ptl);
 352		hugetlb_vma_unlock_read(vma);
 353	} else {
 354		/*
 355		 * If migration entry existed, safe to release vma lock
 356		 * here because the pgtable page won't be freed without the
 357		 * pgtable lock released.  See comment right above pgtable
 358		 * lock release in migration_entry_wait_on_locked().
 359		 */
 360		hugetlb_vma_unlock_read(vma);
 361		migration_entry_wait_on_locked(pte_to_swp_entry(pte), ptl);
 362	}
 363}
 364#endif
 365
 366#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
 367void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
 368{
 369	spinlock_t *ptl;
 370
 371	ptl = pmd_lock(mm, pmd);
 372	if (!is_pmd_migration_entry(*pmd))
 373		goto unlock;
 374	migration_entry_wait_on_locked(pmd_to_swp_entry(*pmd), ptl);
 375	return;
 376unlock:
 377	spin_unlock(ptl);
 378}
 379#endif
 380
 381static int folio_expected_refs(struct address_space *mapping,
 382		struct folio *folio)
 383{
 384	int refs = 1;
 385	if (!mapping)
 386		return refs;
 387
 388	refs += folio_nr_pages(folio);
 389	if (folio_test_private(folio))
 390		refs++;
 391
 392	return refs;
 393}
 394
 395/*
 396 * Replace the page in the mapping.
 397 *
 398 * The number of remaining references must be:
 399 * 1 for anonymous pages without a mapping
 400 * 2 for pages with a mapping
 401 * 3 for pages with a mapping and PagePrivate/PagePrivate2 set.
 402 */
 403int folio_migrate_mapping(struct address_space *mapping,
 404		struct folio *newfolio, struct folio *folio, int extra_count)
 405{
 406	XA_STATE(xas, &mapping->i_pages, folio_index(folio));
 407	struct zone *oldzone, *newzone;
 408	int dirty;
 409	int expected_count = folio_expected_refs(mapping, folio) + extra_count;
 410	long nr = folio_nr_pages(folio);
 411	long entries, i;
 412
 413	if (!mapping) {
 414		/* Anonymous page without mapping */
 415		if (folio_ref_count(folio) != expected_count)
 416			return -EAGAIN;
 417
 418		/* No turning back from here */
 419		newfolio->index = folio->index;
 420		newfolio->mapping = folio->mapping;
 421		if (folio_test_swapbacked(folio))
 422			__folio_set_swapbacked(newfolio);
 423
 424		return MIGRATEPAGE_SUCCESS;
 425	}
 426
 427	oldzone = folio_zone(folio);
 428	newzone = folio_zone(newfolio);
 429
 430	xas_lock_irq(&xas);
 431	if (!folio_ref_freeze(folio, expected_count)) {
 432		xas_unlock_irq(&xas);
 433		return -EAGAIN;
 434	}
 435
 436	/*
 437	 * Now we know that no one else is looking at the folio:
 438	 * no turning back from here.
 439	 */
 440	newfolio->index = folio->index;
 441	newfolio->mapping = folio->mapping;
 442	folio_ref_add(newfolio, nr); /* add cache reference */
 443	if (folio_test_swapbacked(folio)) {
 444		__folio_set_swapbacked(newfolio);
 445		if (folio_test_swapcache(folio)) {
 446			folio_set_swapcache(newfolio);
 447			newfolio->private = folio_get_private(folio);
 448		}
 449		entries = nr;
 450	} else {
 451		VM_BUG_ON_FOLIO(folio_test_swapcache(folio), folio);
 452		entries = 1;
 453	}
 454
 455	/* Move dirty while page refs frozen and newpage not yet exposed */
 456	dirty = folio_test_dirty(folio);
 457	if (dirty) {
 458		folio_clear_dirty(folio);
 459		folio_set_dirty(newfolio);
 460	}
 461
 462	/* Swap cache still stores N entries instead of a high-order entry */
 463	for (i = 0; i < entries; i++) {
 464		xas_store(&xas, newfolio);
 465		xas_next(&xas);
 466	}
 467
 468	/*
 469	 * Drop cache reference from old page by unfreezing
 470	 * to one less reference.
 471	 * We know this isn't the last reference.
 472	 */
 473	folio_ref_unfreeze(folio, expected_count - nr);
 474
 475	xas_unlock(&xas);
 476	/* Leave irq disabled to prevent preemption while updating stats */
 477
 478	/*
 479	 * If moved to a different zone then also account
 480	 * the page for that zone. Other VM counters will be
 481	 * taken care of when we establish references to the
 482	 * new page and drop references to the old page.
 483	 *
 484	 * Note that anonymous pages are accounted for
 485	 * via NR_FILE_PAGES and NR_ANON_MAPPED if they
 486	 * are mapped to swap space.
 487	 */
 488	if (newzone != oldzone) {
 489		struct lruvec *old_lruvec, *new_lruvec;
 490		struct mem_cgroup *memcg;
 491
 492		memcg = folio_memcg(folio);
 493		old_lruvec = mem_cgroup_lruvec(memcg, oldzone->zone_pgdat);
 494		new_lruvec = mem_cgroup_lruvec(memcg, newzone->zone_pgdat);
 495
 496		__mod_lruvec_state(old_lruvec, NR_FILE_PAGES, -nr);
 497		__mod_lruvec_state(new_lruvec, NR_FILE_PAGES, nr);
 498		if (folio_test_swapbacked(folio) && !folio_test_swapcache(folio)) {
 499			__mod_lruvec_state(old_lruvec, NR_SHMEM, -nr);
 500			__mod_lruvec_state(new_lruvec, NR_SHMEM, nr);
 501
 502			if (folio_test_pmd_mappable(folio)) {
 503				__mod_lruvec_state(old_lruvec, NR_SHMEM_THPS, -nr);
 504				__mod_lruvec_state(new_lruvec, NR_SHMEM_THPS, nr);
 505			}
 506		}
 507#ifdef CONFIG_SWAP
 508		if (folio_test_swapcache(folio)) {
 509			__mod_lruvec_state(old_lruvec, NR_SWAPCACHE, -nr);
 510			__mod_lruvec_state(new_lruvec, NR_SWAPCACHE, nr);
 511		}
 512#endif
 513		if (dirty && mapping_can_writeback(mapping)) {
 514			__mod_lruvec_state(old_lruvec, NR_FILE_DIRTY, -nr);
 515			__mod_zone_page_state(oldzone, NR_ZONE_WRITE_PENDING, -nr);
 516			__mod_lruvec_state(new_lruvec, NR_FILE_DIRTY, nr);
 517			__mod_zone_page_state(newzone, NR_ZONE_WRITE_PENDING, nr);
 518		}
 519	}
 520	local_irq_enable();
 521
 522	return MIGRATEPAGE_SUCCESS;
 523}
 524EXPORT_SYMBOL(folio_migrate_mapping);
 525
 526/*
 527 * The expected number of remaining references is the same as that
 528 * of folio_migrate_mapping().
 529 */
 530int migrate_huge_page_move_mapping(struct address_space *mapping,
 531				   struct folio *dst, struct folio *src)
 532{
 533	XA_STATE(xas, &mapping->i_pages, folio_index(src));
 534	int expected_count;
 535
 536	xas_lock_irq(&xas);
 537	expected_count = folio_expected_refs(mapping, src);
 538	if (!folio_ref_freeze(src, expected_count)) {
 539		xas_unlock_irq(&xas);
 540		return -EAGAIN;
 541	}
 542
 543	dst->index = src->index;
 544	dst->mapping = src->mapping;
 545
 546	folio_ref_add(dst, folio_nr_pages(dst));
 547
 548	xas_store(&xas, dst);
 549
 550	folio_ref_unfreeze(src, expected_count - folio_nr_pages(src));
 551
 552	xas_unlock_irq(&xas);
 553
 554	return MIGRATEPAGE_SUCCESS;
 555}
 556
 557/*
 558 * Copy the flags and some other ancillary information
 559 */
 560void folio_migrate_flags(struct folio *newfolio, struct folio *folio)
 561{
 562	int cpupid;
 563
 564	if (folio_test_error(folio))
 565		folio_set_error(newfolio);
 566	if (folio_test_referenced(folio))
 567		folio_set_referenced(newfolio);
 568	if (folio_test_uptodate(folio))
 569		folio_mark_uptodate(newfolio);
 570	if (folio_test_clear_active(folio)) {
 571		VM_BUG_ON_FOLIO(folio_test_unevictable(folio), folio);
 572		folio_set_active(newfolio);
 573	} else if (folio_test_clear_unevictable(folio))
 574		folio_set_unevictable(newfolio);
 575	if (folio_test_workingset(folio))
 576		folio_set_workingset(newfolio);
 577	if (folio_test_checked(folio))
 578		folio_set_checked(newfolio);
 579	/*
 580	 * PG_anon_exclusive (-> PG_mappedtodisk) is always migrated via
 581	 * migration entries. We can still have PG_anon_exclusive set on an
 582	 * effectively unmapped and unreferenced first sub-pages of an
 583	 * anonymous THP: we can simply copy it here via PG_mappedtodisk.
 584	 */
 585	if (folio_test_mappedtodisk(folio))
 586		folio_set_mappedtodisk(newfolio);
 587
 588	/* Move dirty on pages not done by folio_migrate_mapping() */
 589	if (folio_test_dirty(folio))
 590		folio_set_dirty(newfolio);
 591
 592	if (folio_test_young(folio))
 593		folio_set_young(newfolio);
 594	if (folio_test_idle(folio))
 595		folio_set_idle(newfolio);
 596
 597	/*
 598	 * Copy NUMA information to the new page, to prevent over-eager
 599	 * future migrations of this same page.
 600	 */
 601	cpupid = folio_xchg_last_cpupid(folio, -1);
 602	/*
 603	 * For memory tiering mode, when migrate between slow and fast
 604	 * memory node, reset cpupid, because that is used to record
 605	 * page access time in slow memory node.
 606	 */
 607	if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) {
 608		bool f_toptier = node_is_toptier(folio_nid(folio));
 609		bool t_toptier = node_is_toptier(folio_nid(newfolio));
 610
 611		if (f_toptier != t_toptier)
 612			cpupid = -1;
 613	}
 614	folio_xchg_last_cpupid(newfolio, cpupid);
 615
 616	folio_migrate_ksm(newfolio, folio);
 617	/*
 618	 * Please do not reorder this without considering how mm/ksm.c's
 619	 * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache().
 620	 */
 621	if (folio_test_swapcache(folio))
 622		folio_clear_swapcache(folio);
 623	folio_clear_private(folio);
 624
 625	/* page->private contains hugetlb specific flags */
 626	if (!folio_test_hugetlb(folio))
 627		folio->private = NULL;
 628
 629	/*
 630	 * If any waiters have accumulated on the new page then
 631	 * wake them up.
 632	 */
 633	if (folio_test_writeback(newfolio))
 634		folio_end_writeback(newfolio);
 635
 636	/*
 637	 * PG_readahead shares the same bit with PG_reclaim.  The above
 638	 * end_page_writeback() may clear PG_readahead mistakenly, so set the
 639	 * bit after that.
 640	 */
 641	if (folio_test_readahead(folio))
 642		folio_set_readahead(newfolio);
 643
 644	folio_copy_owner(newfolio, folio);
 645
 646	mem_cgroup_migrate(folio, newfolio);
 647}
 648EXPORT_SYMBOL(folio_migrate_flags);
 649
 650void folio_migrate_copy(struct folio *newfolio, struct folio *folio)
 651{
 652	folio_copy(newfolio, folio);
 653	folio_migrate_flags(newfolio, folio);
 654}
 655EXPORT_SYMBOL(folio_migrate_copy);
 656
 657/************************************************************
 658 *                    Migration functions
 659 ***********************************************************/
 660
 661int migrate_folio_extra(struct address_space *mapping, struct folio *dst,
 662		struct folio *src, enum migrate_mode mode, int extra_count)
 663{
 664	int rc;
 665
 666	BUG_ON(folio_test_writeback(src));	/* Writeback must be complete */
 667
 668	rc = folio_migrate_mapping(mapping, dst, src, extra_count);
 669
 670	if (rc != MIGRATEPAGE_SUCCESS)
 671		return rc;
 672
 673	if (mode != MIGRATE_SYNC_NO_COPY)
 674		folio_migrate_copy(dst, src);
 675	else
 676		folio_migrate_flags(dst, src);
 677	return MIGRATEPAGE_SUCCESS;
 678}
 679
 680/**
 681 * migrate_folio() - Simple folio migration.
 682 * @mapping: The address_space containing the folio.
 683 * @dst: The folio to migrate the data to.
 684 * @src: The folio containing the current data.
 685 * @mode: How to migrate the page.
 686 *
 687 * Common logic to directly migrate a single LRU folio suitable for
 688 * folios that do not use PagePrivate/PagePrivate2.
 689 *
 690 * Folios are locked upon entry and exit.
 691 */
 692int migrate_folio(struct address_space *mapping, struct folio *dst,
 693		struct folio *src, enum migrate_mode mode)
 694{
 695	return migrate_folio_extra(mapping, dst, src, mode, 0);
 696}
 697EXPORT_SYMBOL(migrate_folio);
 698
 699#ifdef CONFIG_BUFFER_HEAD
 700/* Returns true if all buffers are successfully locked */
 701static bool buffer_migrate_lock_buffers(struct buffer_head *head,
 702							enum migrate_mode mode)
 703{
 704	struct buffer_head *bh = head;
 705	struct buffer_head *failed_bh;
 706
 707	do {
 708		if (!trylock_buffer(bh)) {
 709			if (mode == MIGRATE_ASYNC)
 710				goto unlock;
 711			if (mode == MIGRATE_SYNC_LIGHT && !buffer_uptodate(bh))
 712				goto unlock;
 713			lock_buffer(bh);
 714		}
 715
 716		bh = bh->b_this_page;
 717	} while (bh != head);
 718
 719	return true;
 720
 721unlock:
 722	/* We failed to lock the buffer and cannot stall. */
 723	failed_bh = bh;
 724	bh = head;
 725	while (bh != failed_bh) {
 726		unlock_buffer(bh);
 727		bh = bh->b_this_page;
 728	}
 729
 730	return false;
 731}
 732
 733static int __buffer_migrate_folio(struct address_space *mapping,
 734		struct folio *dst, struct folio *src, enum migrate_mode mode,
 735		bool check_refs)
 736{
 737	struct buffer_head *bh, *head;
 738	int rc;
 739	int expected_count;
 740
 741	head = folio_buffers(src);
 742	if (!head)
 743		return migrate_folio(mapping, dst, src, mode);
 744
 745	/* Check whether page does not have extra refs before we do more work */
 746	expected_count = folio_expected_refs(mapping, src);
 747	if (folio_ref_count(src) != expected_count)
 748		return -EAGAIN;
 749
 750	if (!buffer_migrate_lock_buffers(head, mode))
 751		return -EAGAIN;
 752
 753	if (check_refs) {
 754		bool busy;
 755		bool invalidated = false;
 756
 757recheck_buffers:
 758		busy = false;
 759		spin_lock(&mapping->i_private_lock);
 760		bh = head;
 761		do {
 762			if (atomic_read(&bh->b_count)) {
 763				busy = true;
 764				break;
 765			}
 766			bh = bh->b_this_page;
 767		} while (bh != head);
 768		if (busy) {
 769			if (invalidated) {
 770				rc = -EAGAIN;
 771				goto unlock_buffers;
 772			}
 773			spin_unlock(&mapping->i_private_lock);
 774			invalidate_bh_lrus();
 775			invalidated = true;
 776			goto recheck_buffers;
 777		}
 778	}
 779
 780	rc = folio_migrate_mapping(mapping, dst, src, 0);
 781	if (rc != MIGRATEPAGE_SUCCESS)
 782		goto unlock_buffers;
 783
 784	folio_attach_private(dst, folio_detach_private(src));
 785
 786	bh = head;
 787	do {
 788		folio_set_bh(bh, dst, bh_offset(bh));
 789		bh = bh->b_this_page;
 790	} while (bh != head);
 791
 792	if (mode != MIGRATE_SYNC_NO_COPY)
 793		folio_migrate_copy(dst, src);
 794	else
 795		folio_migrate_flags(dst, src);
 796
 797	rc = MIGRATEPAGE_SUCCESS;
 798unlock_buffers:
 799	if (check_refs)
 800		spin_unlock(&mapping->i_private_lock);
 801	bh = head;
 802	do {
 803		unlock_buffer(bh);
 804		bh = bh->b_this_page;
 805	} while (bh != head);
 806
 807	return rc;
 808}
 809
 810/**
 811 * buffer_migrate_folio() - Migration function for folios with buffers.
 812 * @mapping: The address space containing @src.
 813 * @dst: The folio to migrate to.
 814 * @src: The folio to migrate from.
 815 * @mode: How to migrate the folio.
 816 *
 817 * This function can only be used if the underlying filesystem guarantees
 818 * that no other references to @src exist. For example attached buffer
 819 * heads are accessed only under the folio lock.  If your filesystem cannot
 820 * provide this guarantee, buffer_migrate_folio_norefs() may be more
 821 * appropriate.
 822 *
 823 * Return: 0 on success or a negative errno on failure.
 824 */
 825int buffer_migrate_folio(struct address_space *mapping,
 826		struct folio *dst, struct folio *src, enum migrate_mode mode)
 827{
 828	return __buffer_migrate_folio(mapping, dst, src, mode, false);
 829}
 830EXPORT_SYMBOL(buffer_migrate_folio);
 831
 832/**
 833 * buffer_migrate_folio_norefs() - Migration function for folios with buffers.
 834 * @mapping: The address space containing @src.
 835 * @dst: The folio to migrate to.
 836 * @src: The folio to migrate from.
 837 * @mode: How to migrate the folio.
 838 *
 839 * Like buffer_migrate_folio() except that this variant is more careful
 840 * and checks that there are also no buffer head references. This function
 841 * is the right one for mappings where buffer heads are directly looked
 842 * up and referenced (such as block device mappings).
 843 *
 844 * Return: 0 on success or a negative errno on failure.
 845 */
 846int buffer_migrate_folio_norefs(struct address_space *mapping,
 847		struct folio *dst, struct folio *src, enum migrate_mode mode)
 848{
 849	return __buffer_migrate_folio(mapping, dst, src, mode, true);
 850}
 851EXPORT_SYMBOL_GPL(buffer_migrate_folio_norefs);
 852#endif /* CONFIG_BUFFER_HEAD */
 853
 854int filemap_migrate_folio(struct address_space *mapping,
 855		struct folio *dst, struct folio *src, enum migrate_mode mode)
 856{
 857	int ret;
 858
 859	ret = folio_migrate_mapping(mapping, dst, src, 0);
 860	if (ret != MIGRATEPAGE_SUCCESS)
 861		return ret;
 862
 863	if (folio_get_private(src))
 864		folio_attach_private(dst, folio_detach_private(src));
 865
 866	if (mode != MIGRATE_SYNC_NO_COPY)
 867		folio_migrate_copy(dst, src);
 868	else
 869		folio_migrate_flags(dst, src);
 870	return MIGRATEPAGE_SUCCESS;
 871}
 872EXPORT_SYMBOL_GPL(filemap_migrate_folio);
 873
 874/*
 875 * Writeback a folio to clean the dirty state
 876 */
 877static int writeout(struct address_space *mapping, struct folio *folio)
 878{
 879	struct writeback_control wbc = {
 880		.sync_mode = WB_SYNC_NONE,
 881		.nr_to_write = 1,
 882		.range_start = 0,
 883		.range_end = LLONG_MAX,
 884		.for_reclaim = 1
 885	};
 886	int rc;
 887
 888	if (!mapping->a_ops->writepage)
 889		/* No write method for the address space */
 890		return -EINVAL;
 891
 892	if (!folio_clear_dirty_for_io(folio))
 893		/* Someone else already triggered a write */
 894		return -EAGAIN;
 895
 896	/*
 897	 * A dirty folio may imply that the underlying filesystem has
 898	 * the folio on some queue. So the folio must be clean for
 899	 * migration. Writeout may mean we lose the lock and the
 900	 * folio state is no longer what we checked for earlier.
 901	 * At this point we know that the migration attempt cannot
 902	 * be successful.
 903	 */
 904	remove_migration_ptes(folio, folio, false);
 905
 906	rc = mapping->a_ops->writepage(&folio->page, &wbc);
 907
 908	if (rc != AOP_WRITEPAGE_ACTIVATE)
 909		/* unlocked. Relock */
 910		folio_lock(folio);
 911
 912	return (rc < 0) ? -EIO : -EAGAIN;
 913}
 914
 915/*
 916 * Default handling if a filesystem does not provide a migration function.
 917 */
 918static int fallback_migrate_folio(struct address_space *mapping,
 919		struct folio *dst, struct folio *src, enum migrate_mode mode)
 920{
 921	if (folio_test_dirty(src)) {
 922		/* Only writeback folios in full synchronous migration */
 923		switch (mode) {
 924		case MIGRATE_SYNC:
 925		case MIGRATE_SYNC_NO_COPY:
 926			break;
 927		default:
 928			return -EBUSY;
 929		}
 930		return writeout(mapping, src);
 931	}
 932
 933	/*
 934	 * Buffers may be managed in a filesystem specific way.
 935	 * We must have no buffers or drop them.
 936	 */
 937	if (!filemap_release_folio(src, GFP_KERNEL))
 938		return mode == MIGRATE_SYNC ? -EAGAIN : -EBUSY;
 939
 940	return migrate_folio(mapping, dst, src, mode);
 941}
 942
 943/*
 944 * Move a page to a newly allocated page
 945 * The page is locked and all ptes have been successfully removed.
 946 *
 947 * The new page will have replaced the old page if this function
 948 * is successful.
 949 *
 950 * Return value:
 951 *   < 0 - error code
 952 *  MIGRATEPAGE_SUCCESS - success
 953 */
 954static int move_to_new_folio(struct folio *dst, struct folio *src,
 955				enum migrate_mode mode)
 956{
 957	int rc = -EAGAIN;
 958	bool is_lru = !__folio_test_movable(src);
 959
 960	VM_BUG_ON_FOLIO(!folio_test_locked(src), src);
 961	VM_BUG_ON_FOLIO(!folio_test_locked(dst), dst);
 962
 963	if (likely(is_lru)) {
 964		struct address_space *mapping = folio_mapping(src);
 965
 966		if (!mapping)
 967			rc = migrate_folio(mapping, dst, src, mode);
 968		else if (mapping_unmovable(mapping))
 969			rc = -EOPNOTSUPP;
 970		else if (mapping->a_ops->migrate_folio)
 971			/*
 972			 * Most folios have a mapping and most filesystems
 973			 * provide a migrate_folio callback. Anonymous folios
 974			 * are part of swap space which also has its own
 975			 * migrate_folio callback. This is the most common path
 976			 * for page migration.
 977			 */
 978			rc = mapping->a_ops->migrate_folio(mapping, dst, src,
 979								mode);
 980		else
 981			rc = fallback_migrate_folio(mapping, dst, src, mode);
 982	} else {
 983		const struct movable_operations *mops;
 984
 985		/*
 986		 * In case of non-lru page, it could be released after
 987		 * isolation step. In that case, we shouldn't try migration.
 988		 */
 989		VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
 990		if (!folio_test_movable(src)) {
 991			rc = MIGRATEPAGE_SUCCESS;
 992			folio_clear_isolated(src);
 993			goto out;
 994		}
 995
 996		mops = folio_movable_ops(src);
 997		rc = mops->migrate_page(&dst->page, &src->page, mode);
 998		WARN_ON_ONCE(rc == MIGRATEPAGE_SUCCESS &&
 999				!folio_test_isolated(src));
1000	}
1001
1002	/*
1003	 * When successful, old pagecache src->mapping must be cleared before
1004	 * src is freed; but stats require that PageAnon be left as PageAnon.
1005	 */
1006	if (rc == MIGRATEPAGE_SUCCESS) {
1007		if (__folio_test_movable(src)) {
1008			VM_BUG_ON_FOLIO(!folio_test_isolated(src), src);
1009
1010			/*
1011			 * We clear PG_movable under page_lock so any compactor
1012			 * cannot try to migrate this page.
1013			 */
1014			folio_clear_isolated(src);
1015		}
1016
1017		/*
1018		 * Anonymous and movable src->mapping will be cleared by
1019		 * free_pages_prepare so don't reset it here for keeping
1020		 * the type to work PageAnon, for example.
1021		 */
1022		if (!folio_mapping_flags(src))
1023			src->mapping = NULL;
1024
1025		if (likely(!folio_is_zone_device(dst)))
1026			flush_dcache_folio(dst);
1027	}
1028out:
1029	return rc;
1030}
1031
1032/*
1033 * To record some information during migration, we use unused private
1034 * field of struct folio of the newly allocated destination folio.
1035 * This is safe because nobody is using it except us.
1036 */
1037enum {
1038	PAGE_WAS_MAPPED = BIT(0),
1039	PAGE_WAS_MLOCKED = BIT(1),
1040	PAGE_OLD_STATES = PAGE_WAS_MAPPED | PAGE_WAS_MLOCKED,
1041};
1042
1043static void __migrate_folio_record(struct folio *dst,
1044				   int old_page_state,
1045				   struct anon_vma *anon_vma)
1046{
1047	dst->private = (void *)anon_vma + old_page_state;
1048}
1049
1050static void __migrate_folio_extract(struct folio *dst,
1051				   int *old_page_state,
1052				   struct anon_vma **anon_vmap)
1053{
1054	unsigned long private = (unsigned long)dst->private;
1055
1056	*anon_vmap = (struct anon_vma *)(private & ~PAGE_OLD_STATES);
1057	*old_page_state = private & PAGE_OLD_STATES;
1058	dst->private = NULL;
1059}
1060
1061/* Restore the source folio to the original state upon failure */
1062static void migrate_folio_undo_src(struct folio *src,
1063				   int page_was_mapped,
1064				   struct anon_vma *anon_vma,
1065				   bool locked,
1066				   struct list_head *ret)
1067{
1068	if (page_was_mapped)
1069		remove_migration_ptes(src, src, false);
1070	/* Drop an anon_vma reference if we took one */
1071	if (anon_vma)
1072		put_anon_vma(anon_vma);
1073	if (locked)
1074		folio_unlock(src);
1075	if (ret)
1076		list_move_tail(&src->lru, ret);
1077}
1078
1079/* Restore the destination folio to the original state upon failure */
1080static void migrate_folio_undo_dst(struct folio *dst, bool locked,
1081		free_folio_t put_new_folio, unsigned long private)
1082{
1083	if (locked)
1084		folio_unlock(dst);
1085	if (put_new_folio)
1086		put_new_folio(dst, private);
1087	else
1088		folio_put(dst);
1089}
1090
1091/* Cleanup src folio upon migration success */
1092static void migrate_folio_done(struct folio *src,
1093			       enum migrate_reason reason)
1094{
1095	/*
1096	 * Compaction can migrate also non-LRU pages which are
1097	 * not accounted to NR_ISOLATED_*. They can be recognized
1098	 * as __folio_test_movable
1099	 */
1100	if (likely(!__folio_test_movable(src)))
1101		mod_node_page_state(folio_pgdat(src), NR_ISOLATED_ANON +
1102				    folio_is_file_lru(src), -folio_nr_pages(src));
1103
1104	if (reason != MR_MEMORY_FAILURE)
1105		/* We release the page in page_handle_poison. */
1106		folio_put(src);
1107}
1108
1109/* Obtain the lock on page, remove all ptes. */
1110static int migrate_folio_unmap(new_folio_t get_new_folio,
1111		free_folio_t put_new_folio, unsigned long private,
1112		struct folio *src, struct folio **dstp, enum migrate_mode mode,
1113		enum migrate_reason reason, struct list_head *ret)
1114{
1115	struct folio *dst;
1116	int rc = -EAGAIN;
1117	int old_page_state = 0;
1118	struct anon_vma *anon_vma = NULL;
1119	bool is_lru = !__folio_test_movable(src);
1120	bool locked = false;
1121	bool dst_locked = false;
1122
1123	if (folio_ref_count(src) == 1) {
1124		/* Folio was freed from under us. So we are done. */
1125		folio_clear_active(src);
1126		folio_clear_unevictable(src);
1127		/* free_pages_prepare() will clear PG_isolated. */
1128		list_del(&src->lru);
1129		migrate_folio_done(src, reason);
1130		return MIGRATEPAGE_SUCCESS;
1131	}
1132
1133	dst = get_new_folio(src, private);
1134	if (!dst)
1135		return -ENOMEM;
1136	*dstp = dst;
1137
1138	dst->private = NULL;
1139
1140	if (!folio_trylock(src)) {
1141		if (mode == MIGRATE_ASYNC)
1142			goto out;
1143
1144		/*
1145		 * It's not safe for direct compaction to call lock_page.
1146		 * For example, during page readahead pages are added locked
1147		 * to the LRU. Later, when the IO completes the pages are
1148		 * marked uptodate and unlocked. However, the queueing
1149		 * could be merging multiple pages for one bio (e.g.
1150		 * mpage_readahead). If an allocation happens for the
1151		 * second or third page, the process can end up locking
1152		 * the same page twice and deadlocking. Rather than
1153		 * trying to be clever about what pages can be locked,
1154		 * avoid the use of lock_page for direct compaction
1155		 * altogether.
1156		 */
1157		if (current->flags & PF_MEMALLOC)
1158			goto out;
1159
1160		/*
1161		 * In "light" mode, we can wait for transient locks (eg
1162		 * inserting a page into the page table), but it's not
1163		 * worth waiting for I/O.
1164		 */
1165		if (mode == MIGRATE_SYNC_LIGHT && !folio_test_uptodate(src))
1166			goto out;
1167
1168		folio_lock(src);
1169	}
1170	locked = true;
1171	if (folio_test_mlocked(src))
1172		old_page_state |= PAGE_WAS_MLOCKED;
1173
1174	if (folio_test_writeback(src)) {
1175		/*
1176		 * Only in the case of a full synchronous migration is it
1177		 * necessary to wait for PageWriteback. In the async case,
1178		 * the retry loop is too short and in the sync-light case,
1179		 * the overhead of stalling is too much
1180		 */
1181		switch (mode) {
1182		case MIGRATE_SYNC:
1183		case MIGRATE_SYNC_NO_COPY:
1184			break;
1185		default:
1186			rc = -EBUSY;
1187			goto out;
1188		}
1189		folio_wait_writeback(src);
1190	}
1191
1192	/*
1193	 * By try_to_migrate(), src->mapcount goes down to 0 here. In this case,
1194	 * we cannot notice that anon_vma is freed while we migrate a page.
1195	 * This get_anon_vma() delays freeing anon_vma pointer until the end
1196	 * of migration. File cache pages are no problem because of page_lock()
1197	 * File Caches may use write_page() or lock_page() in migration, then,
1198	 * just care Anon page here.
1199	 *
1200	 * Only folio_get_anon_vma() understands the subtleties of
1201	 * getting a hold on an anon_vma from outside one of its mms.
1202	 * But if we cannot get anon_vma, then we won't need it anyway,
1203	 * because that implies that the anon page is no longer mapped
1204	 * (and cannot be remapped so long as we hold the page lock).
1205	 */
1206	if (folio_test_anon(src) && !folio_test_ksm(src))
1207		anon_vma = folio_get_anon_vma(src);
1208
1209	/*
1210	 * Block others from accessing the new page when we get around to
1211	 * establishing additional references. We are usually the only one
1212	 * holding a reference to dst at this point. We used to have a BUG
1213	 * here if folio_trylock(dst) fails, but would like to allow for
1214	 * cases where there might be a race with the previous use of dst.
1215	 * This is much like races on refcount of oldpage: just don't BUG().
1216	 */
1217	if (unlikely(!folio_trylock(dst)))
1218		goto out;
1219	dst_locked = true;
1220
1221	if (unlikely(!is_lru)) {
1222		__migrate_folio_record(dst, old_page_state, anon_vma);
1223		return MIGRATEPAGE_UNMAP;
1224	}
1225
1226	/*
1227	 * Corner case handling:
1228	 * 1. When a new swap-cache page is read into, it is added to the LRU
1229	 * and treated as swapcache but it has no rmap yet.
1230	 * Calling try_to_unmap() against a src->mapping==NULL page will
1231	 * trigger a BUG.  So handle it here.
1232	 * 2. An orphaned page (see truncate_cleanup_page) might have
1233	 * fs-private metadata. The page can be picked up due to memory
1234	 * offlining.  Everywhere else except page reclaim, the page is
1235	 * invisible to the vm, so the page can not be migrated.  So try to
1236	 * free the metadata, so the page can be freed.
1237	 */
1238	if (!src->mapping) {
1239		if (folio_test_private(src)) {
1240			try_to_free_buffers(src);
1241			goto out;
1242		}
1243	} else if (folio_mapped(src)) {
1244		/* Establish migration ptes */
1245		VM_BUG_ON_FOLIO(folio_test_anon(src) &&
1246			       !folio_test_ksm(src) && !anon_vma, src);
1247		try_to_migrate(src, mode == MIGRATE_ASYNC ? TTU_BATCH_FLUSH : 0);
1248		old_page_state |= PAGE_WAS_MAPPED;
1249	}
1250
1251	if (!folio_mapped(src)) {
1252		__migrate_folio_record(dst, old_page_state, anon_vma);
1253		return MIGRATEPAGE_UNMAP;
1254	}
1255
1256out:
1257	/*
1258	 * A folio that has not been unmapped will be restored to
1259	 * right list unless we want to retry.
1260	 */
1261	if (rc == -EAGAIN)
1262		ret = NULL;
1263
1264	migrate_folio_undo_src(src, old_page_state & PAGE_WAS_MAPPED,
1265			       anon_vma, locked, ret);
1266	migrate_folio_undo_dst(dst, dst_locked, put_new_folio, private);
1267
1268	return rc;
1269}
1270
1271/* Migrate the folio to the newly allocated folio in dst. */
1272static int migrate_folio_move(free_folio_t put_new_folio, unsigned long private,
1273			      struct folio *src, struct folio *dst,
1274			      enum migrate_mode mode, enum migrate_reason reason,
1275			      struct list_head *ret)
1276{
1277	int rc;
1278	int old_page_state = 0;
1279	struct anon_vma *anon_vma = NULL;
1280	bool is_lru = !__folio_test_movable(src);
1281	struct list_head *prev;
1282
1283	__migrate_folio_extract(dst, &old_page_state, &anon_vma);
1284	prev = dst->lru.prev;
1285	list_del(&dst->lru);
1286
1287	rc = move_to_new_folio(dst, src, mode);
1288	if (rc)
1289		goto out;
1290
1291	if (unlikely(!is_lru))
1292		goto out_unlock_both;
1293
1294	/*
1295	 * When successful, push dst to LRU immediately: so that if it
1296	 * turns out to be an mlocked page, remove_migration_ptes() will
1297	 * automatically build up the correct dst->mlock_count for it.
1298	 *
1299	 * We would like to do something similar for the old page, when
1300	 * unsuccessful, and other cases when a page has been temporarily
1301	 * isolated from the unevictable LRU: but this case is the easiest.
1302	 */
1303	folio_add_lru(dst);
1304	if (old_page_state & PAGE_WAS_MLOCKED)
1305		lru_add_drain();
1306
1307	if (old_page_state & PAGE_WAS_MAPPED)
1308		remove_migration_ptes(src, dst, false);
1309
1310out_unlock_both:
1311	folio_unlock(dst);
1312	set_page_owner_migrate_reason(&dst->page, reason);
1313	/*
1314	 * If migration is successful, decrease refcount of dst,
1315	 * which will not free the page because new page owner increased
1316	 * refcounter.
1317	 */
1318	folio_put(dst);
1319
1320	/*
1321	 * A folio that has been migrated has all references removed
1322	 * and will be freed.
1323	 */
1324	list_del(&src->lru);
1325	/* Drop an anon_vma reference if we took one */
1326	if (anon_vma)
1327		put_anon_vma(anon_vma);
1328	folio_unlock(src);
1329	migrate_folio_done(src, reason);
1330
1331	return rc;
1332out:
1333	/*
1334	 * A folio that has not been migrated will be restored to
1335	 * right list unless we want to retry.
1336	 */
1337	if (rc == -EAGAIN) {
1338		list_add(&dst->lru, prev);
1339		__migrate_folio_record(dst, old_page_state, anon_vma);
1340		return rc;
1341	}
1342
1343	migrate_folio_undo_src(src, old_page_state & PAGE_WAS_MAPPED,
1344			       anon_vma, true, ret);
1345	migrate_folio_undo_dst(dst, true, put_new_folio, private);
1346
1347	return rc;
1348}
1349
1350/*
1351 * Counterpart of unmap_and_move_page() for hugepage migration.
1352 *
1353 * This function doesn't wait the completion of hugepage I/O
1354 * because there is no race between I/O and migration for hugepage.
1355 * Note that currently hugepage I/O occurs only in direct I/O
1356 * where no lock is held and PG_writeback is irrelevant,
1357 * and writeback status of all subpages are counted in the reference
1358 * count of the head page (i.e. if all subpages of a 2MB hugepage are
1359 * under direct I/O, the reference of the head page is 512 and a bit more.)
1360 * This means that when we try to migrate hugepage whose subpages are
1361 * doing direct I/O, some references remain after try_to_unmap() and
1362 * hugepage migration fails without data corruption.
1363 *
1364 * There is also no race when direct I/O is issued on the page under migration,
1365 * because then pte is replaced with migration swap entry and direct I/O code
1366 * will wait in the page fault for migration to complete.
1367 */
1368static int unmap_and_move_huge_page(new_folio_t get_new_folio,
1369		free_folio_t put_new_folio, unsigned long private,
1370		struct folio *src, int force, enum migrate_mode mode,
1371		int reason, struct list_head *ret)
1372{
1373	struct folio *dst;
1374	int rc = -EAGAIN;
1375	int page_was_mapped = 0;
1376	struct anon_vma *anon_vma = NULL;
1377	struct address_space *mapping = NULL;
1378
1379	if (folio_ref_count(src) == 1) {
1380		/* page was freed from under us. So we are done. */
1381		folio_putback_active_hugetlb(src);
1382		return MIGRATEPAGE_SUCCESS;
1383	}
1384
1385	dst = get_new_folio(src, private);
1386	if (!dst)
1387		return -ENOMEM;
1388
1389	if (!folio_trylock(src)) {
1390		if (!force)
1391			goto out;
1392		switch (mode) {
1393		case MIGRATE_SYNC:
1394		case MIGRATE_SYNC_NO_COPY:
1395			break;
1396		default:
1397			goto out;
1398		}
1399		folio_lock(src);
1400	}
1401
1402	/*
1403	 * Check for pages which are in the process of being freed.  Without
1404	 * folio_mapping() set, hugetlbfs specific move page routine will not
1405	 * be called and we could leak usage counts for subpools.
1406	 */
1407	if (hugetlb_folio_subpool(src) && !folio_mapping(src)) {
1408		rc = -EBUSY;
1409		goto out_unlock;
1410	}
1411
1412	if (folio_test_anon(src))
1413		anon_vma = folio_get_anon_vma(src);
1414
1415	if (unlikely(!folio_trylock(dst)))
1416		goto put_anon;
1417
1418	if (folio_mapped(src)) {
1419		enum ttu_flags ttu = 0;
1420
1421		if (!folio_test_anon(src)) {
1422			/*
1423			 * In shared mappings, try_to_unmap could potentially
1424			 * call huge_pmd_unshare.  Because of this, take
1425			 * semaphore in write mode here and set TTU_RMAP_LOCKED
1426			 * to let lower levels know we have taken the lock.
1427			 */
1428			mapping = hugetlb_page_mapping_lock_write(&src->page);
1429			if (unlikely(!mapping))
1430				goto unlock_put_anon;
1431
1432			ttu = TTU_RMAP_LOCKED;
1433		}
1434
1435		try_to_migrate(src, ttu);
1436		page_was_mapped = 1;
1437
1438		if (ttu & TTU_RMAP_LOCKED)
1439			i_mmap_unlock_write(mapping);
1440	}
1441
1442	if (!folio_mapped(src))
1443		rc = move_to_new_folio(dst, src, mode);
1444
1445	if (page_was_mapped)
1446		remove_migration_ptes(src,
1447			rc == MIGRATEPAGE_SUCCESS ? dst : src, false);
1448
1449unlock_put_anon:
1450	folio_unlock(dst);
1451
1452put_anon:
1453	if (anon_vma)
1454		put_anon_vma(anon_vma);
1455
1456	if (rc == MIGRATEPAGE_SUCCESS) {
1457		move_hugetlb_state(src, dst, reason);
1458		put_new_folio = NULL;
1459	}
1460
1461out_unlock:
1462	folio_unlock(src);
1463out:
1464	if (rc == MIGRATEPAGE_SUCCESS)
1465		folio_putback_active_hugetlb(src);
1466	else if (rc != -EAGAIN)
1467		list_move_tail(&src->lru, ret);
1468
1469	/*
1470	 * If migration was not successful and there's a freeing callback, use
1471	 * it.  Otherwise, put_page() will drop the reference grabbed during
1472	 * isolation.
1473	 */
1474	if (put_new_folio)
1475		put_new_folio(dst, private);
1476	else
1477		folio_putback_active_hugetlb(dst);
1478
1479	return rc;
1480}
1481
1482static inline int try_split_folio(struct folio *folio, struct list_head *split_folios)
1483{
1484	int rc;
1485
1486	folio_lock(folio);
1487	rc = split_folio_to_list(folio, split_folios);
1488	folio_unlock(folio);
1489	if (!rc)
1490		list_move_tail(&folio->lru, split_folios);
1491
1492	return rc;
1493}
1494
1495#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1496#define NR_MAX_BATCHED_MIGRATION	HPAGE_PMD_NR
1497#else
1498#define NR_MAX_BATCHED_MIGRATION	512
1499#endif
1500#define NR_MAX_MIGRATE_PAGES_RETRY	10
1501#define NR_MAX_MIGRATE_ASYNC_RETRY	3
1502#define NR_MAX_MIGRATE_SYNC_RETRY					\
1503	(NR_MAX_MIGRATE_PAGES_RETRY - NR_MAX_MIGRATE_ASYNC_RETRY)
1504
1505struct migrate_pages_stats {
1506	int nr_succeeded;	/* Normal and large folios migrated successfully, in
1507				   units of base pages */
1508	int nr_failed_pages;	/* Normal and large folios failed to be migrated, in
1509				   units of base pages.  Untried folios aren't counted */
1510	int nr_thp_succeeded;	/* THP migrated successfully */
1511	int nr_thp_failed;	/* THP failed to be migrated */
1512	int nr_thp_split;	/* THP split before migrating */
1513	int nr_split;	/* Large folio (include THP) split before migrating */
1514};
1515
1516/*
1517 * Returns the number of hugetlb folios that were not migrated, or an error code
1518 * after NR_MAX_MIGRATE_PAGES_RETRY attempts or if no hugetlb folios are movable
1519 * any more because the list has become empty or no retryable hugetlb folios
1520 * exist any more. It is caller's responsibility to call putback_movable_pages()
1521 * only if ret != 0.
1522 */
1523static int migrate_hugetlbs(struct list_head *from, new_folio_t get_new_folio,
1524			    free_folio_t put_new_folio, unsigned long private,
1525			    enum migrate_mode mode, int reason,
1526			    struct migrate_pages_stats *stats,
1527			    struct list_head *ret_folios)
1528{
1529	int retry = 1;
1530	int nr_failed = 0;
1531	int nr_retry_pages = 0;
1532	int pass = 0;
1533	struct folio *folio, *folio2;
1534	int rc, nr_pages;
1535
1536	for (pass = 0; pass < NR_MAX_MIGRATE_PAGES_RETRY && retry; pass++) {
1537		retry = 0;
1538		nr_retry_pages = 0;
1539
1540		list_for_each_entry_safe(folio, folio2, from, lru) {
1541			if (!folio_test_hugetlb(folio))
1542				continue;
1543
1544			nr_pages = folio_nr_pages(folio);
1545
1546			cond_resched();
1547
1548			/*
1549			 * Migratability of hugepages depends on architectures and
1550			 * their size.  This check is necessary because some callers
1551			 * of hugepage migration like soft offline and memory
1552			 * hotremove don't walk through page tables or check whether
1553			 * the hugepage is pmd-based or not before kicking migration.
1554			 */
1555			if (!hugepage_migration_supported(folio_hstate(folio))) {
1556				nr_failed++;
1557				stats->nr_failed_pages += nr_pages;
1558				list_move_tail(&folio->lru, ret_folios);
1559				continue;
1560			}
1561
1562			rc = unmap_and_move_huge_page(get_new_folio,
1563						      put_new_folio, private,
1564						      folio, pass > 2, mode,
1565						      reason, ret_folios);
1566			/*
1567			 * The rules are:
1568			 *	Success: hugetlb folio will be put back
1569			 *	-EAGAIN: stay on the from list
1570			 *	-ENOMEM: stay on the from list
1571			 *	Other errno: put on ret_folios list
1572			 */
1573			switch(rc) {
1574			case -ENOMEM:
1575				/*
1576				 * When memory is low, don't bother to try to migrate
1577				 * other folios, just exit.
1578				 */
1579				stats->nr_failed_pages += nr_pages + nr_retry_pages;
1580				return -ENOMEM;
1581			case -EAGAIN:
1582				retry++;
1583				nr_retry_pages += nr_pages;
1584				break;
1585			case MIGRATEPAGE_SUCCESS:
1586				stats->nr_succeeded += nr_pages;
1587				break;
1588			default:
1589				/*
1590				 * Permanent failure (-EBUSY, etc.):
1591				 * unlike -EAGAIN case, the failed folio is
1592				 * removed from migration folio list and not
1593				 * retried in the next outer loop.
1594				 */
1595				nr_failed++;
1596				stats->nr_failed_pages += nr_pages;
1597				break;
1598			}
1599		}
1600	}
1601	/*
1602	 * nr_failed is number of hugetlb folios failed to be migrated.  After
1603	 * NR_MAX_MIGRATE_PAGES_RETRY attempts, give up and count retried hugetlb
1604	 * folios as failed.
1605	 */
1606	nr_failed += retry;
1607	stats->nr_failed_pages += nr_retry_pages;
1608
1609	return nr_failed;
1610}
1611
1612/*
1613 * migrate_pages_batch() first unmaps folios in the from list as many as
1614 * possible, then move the unmapped folios.
1615 *
1616 * We only batch migration if mode == MIGRATE_ASYNC to avoid to wait a
1617 * lock or bit when we have locked more than one folio.  Which may cause
1618 * deadlock (e.g., for loop device).  So, if mode != MIGRATE_ASYNC, the
1619 * length of the from list must be <= 1.
1620 */
1621static int migrate_pages_batch(struct list_head *from,
1622		new_folio_t get_new_folio, free_folio_t put_new_folio,
1623		unsigned long private, enum migrate_mode mode, int reason,
1624		struct list_head *ret_folios, struct list_head *split_folios,
1625		struct migrate_pages_stats *stats, int nr_pass)
1626{
1627	int retry = 1;
1628	int thp_retry = 1;
1629	int nr_failed = 0;
1630	int nr_retry_pages = 0;
1631	int pass = 0;
1632	bool is_thp = false;
1633	bool is_large = false;
1634	struct folio *folio, *folio2, *dst = NULL, *dst2;
1635	int rc, rc_saved = 0, nr_pages;
1636	LIST_HEAD(unmap_folios);
1637	LIST_HEAD(dst_folios);
1638	bool nosplit = (reason == MR_NUMA_MISPLACED);
1639
1640	VM_WARN_ON_ONCE(mode != MIGRATE_ASYNC &&
1641			!list_empty(from) && !list_is_singular(from));
1642
1643	for (pass = 0; pass < nr_pass && retry; pass++) {
1644		retry = 0;
1645		thp_retry = 0;
1646		nr_retry_pages = 0;
1647
1648		list_for_each_entry_safe(folio, folio2, from, lru) {
1649			is_large = folio_test_large(folio);
1650			is_thp = is_large && folio_test_pmd_mappable(folio);
1651			nr_pages = folio_nr_pages(folio);
1652
1653			cond_resched();
1654
1655			/*
1656			 * Large folio migration might be unsupported or
1657			 * the allocation might be failed so we should retry
1658			 * on the same folio with the large folio split
1659			 * to normal folios.
1660			 *
1661			 * Split folios are put in split_folios, and
1662			 * we will migrate them after the rest of the
1663			 * list is processed.
1664			 */
1665			if (!thp_migration_supported() && is_thp) {
1666				nr_failed++;
1667				stats->nr_thp_failed++;
1668				if (!try_split_folio(folio, split_folios)) {
1669					stats->nr_thp_split++;
1670					stats->nr_split++;
1671					continue;
1672				}
1673				stats->nr_failed_pages += nr_pages;
1674				list_move_tail(&folio->lru, ret_folios);
1675				continue;
1676			}
1677
1678			rc = migrate_folio_unmap(get_new_folio, put_new_folio,
1679					private, folio, &dst, mode, reason,
1680					ret_folios);
1681			/*
1682			 * The rules are:
1683			 *	Success: folio will be freed
1684			 *	Unmap: folio will be put on unmap_folios list,
1685			 *	       dst folio put on dst_folios list
1686			 *	-EAGAIN: stay on the from list
1687			 *	-ENOMEM: stay on the from list
1688			 *	Other errno: put on ret_folios list
1689			 */
1690			switch(rc) {
1691			case -ENOMEM:
1692				/*
1693				 * When memory is low, don't bother to try to migrate
1694				 * other folios, move unmapped folios, then exit.
1695				 */
1696				nr_failed++;
1697				stats->nr_thp_failed += is_thp;
1698				/* Large folio NUMA faulting doesn't split to retry. */
1699				if (is_large && !nosplit) {
1700					int ret = try_split_folio(folio, split_folios);
1701
1702					if (!ret) {
1703						stats->nr_thp_split += is_thp;
1704						stats->nr_split++;
1705						break;
1706					} else if (reason == MR_LONGTERM_PIN &&
1707						   ret == -EAGAIN) {
1708						/*
1709						 * Try again to split large folio to
1710						 * mitigate the failure of longterm pinning.
1711						 */
1712						retry++;
1713						thp_retry += is_thp;
1714						nr_retry_pages += nr_pages;
1715						/* Undo duplicated failure counting. */
1716						nr_failed--;
1717						stats->nr_thp_failed -= is_thp;
1718						break;
1719					}
1720				}
1721
1722				stats->nr_failed_pages += nr_pages + nr_retry_pages;
1723				/* nr_failed isn't updated for not used */
1724				stats->nr_thp_failed += thp_retry;
1725				rc_saved = rc;
1726				if (list_empty(&unmap_folios))
1727					goto out;
1728				else
1729					goto move;
1730			case -EAGAIN:
1731				retry++;
1732				thp_retry += is_thp;
1733				nr_retry_pages += nr_pages;
1734				break;
1735			case MIGRATEPAGE_SUCCESS:
1736				stats->nr_succeeded += nr_pages;
1737				stats->nr_thp_succeeded += is_thp;
1738				break;
1739			case MIGRATEPAGE_UNMAP:
1740				list_move_tail(&folio->lru, &unmap_folios);
1741				list_add_tail(&dst->lru, &dst_folios);
1742				break;
1743			default:
1744				/*
1745				 * Permanent failure (-EBUSY, etc.):
1746				 * unlike -EAGAIN case, the failed folio is
1747				 * removed from migration folio list and not
1748				 * retried in the next outer loop.
1749				 */
1750				nr_failed++;
1751				stats->nr_thp_failed += is_thp;
1752				stats->nr_failed_pages += nr_pages;
1753				break;
1754			}
1755		}
1756	}
1757	nr_failed += retry;
1758	stats->nr_thp_failed += thp_retry;
1759	stats->nr_failed_pages += nr_retry_pages;
1760move:
1761	/* Flush TLBs for all unmapped folios */
1762	try_to_unmap_flush();
1763
1764	retry = 1;
1765	for (pass = 0; pass < nr_pass && retry; pass++) {
1766		retry = 0;
1767		thp_retry = 0;
1768		nr_retry_pages = 0;
1769
1770		dst = list_first_entry(&dst_folios, struct folio, lru);
1771		dst2 = list_next_entry(dst, lru);
1772		list_for_each_entry_safe(folio, folio2, &unmap_folios, lru) {
1773			is_thp = folio_test_large(folio) && folio_test_pmd_mappable(folio);
1774			nr_pages = folio_nr_pages(folio);
1775
1776			cond_resched();
1777
1778			rc = migrate_folio_move(put_new_folio, private,
1779						folio, dst, mode,
1780						reason, ret_folios);
1781			/*
1782			 * The rules are:
1783			 *	Success: folio will be freed
1784			 *	-EAGAIN: stay on the unmap_folios list
1785			 *	Other errno: put on ret_folios list
1786			 */
1787			switch(rc) {
1788			case -EAGAIN:
1789				retry++;
1790				thp_retry += is_thp;
1791				nr_retry_pages += nr_pages;
1792				break;
1793			case MIGRATEPAGE_SUCCESS:
1794				stats->nr_succeeded += nr_pages;
1795				stats->nr_thp_succeeded += is_thp;
1796				break;
1797			default:
1798				nr_failed++;
1799				stats->nr_thp_failed += is_thp;
1800				stats->nr_failed_pages += nr_pages;
1801				break;
1802			}
1803			dst = dst2;
1804			dst2 = list_next_entry(dst, lru);
1805		}
1806	}
1807	nr_failed += retry;
1808	stats->nr_thp_failed += thp_retry;
1809	stats->nr_failed_pages += nr_retry_pages;
1810
1811	rc = rc_saved ? : nr_failed;
1812out:
1813	/* Cleanup remaining folios */
1814	dst = list_first_entry(&dst_folios, struct folio, lru);
1815	dst2 = list_next_entry(dst, lru);
1816	list_for_each_entry_safe(folio, folio2, &unmap_folios, lru) {
1817		int old_page_state = 0;
1818		struct anon_vma *anon_vma = NULL;
1819
1820		__migrate_folio_extract(dst, &old_page_state, &anon_vma);
1821		migrate_folio_undo_src(folio, old_page_state & PAGE_WAS_MAPPED,
1822				       anon_vma, true, ret_folios);
1823		list_del(&dst->lru);
1824		migrate_folio_undo_dst(dst, true, put_new_folio, private);
1825		dst = dst2;
1826		dst2 = list_next_entry(dst, lru);
1827	}
1828
1829	return rc;
1830}
1831
1832static int migrate_pages_sync(struct list_head *from, new_folio_t get_new_folio,
1833		free_folio_t put_new_folio, unsigned long private,
1834		enum migrate_mode mode, int reason,
1835		struct list_head *ret_folios, struct list_head *split_folios,
1836		struct migrate_pages_stats *stats)
1837{
1838	int rc, nr_failed = 0;
1839	LIST_HEAD(folios);
1840	struct migrate_pages_stats astats;
1841
1842	memset(&astats, 0, sizeof(astats));
1843	/* Try to migrate in batch with MIGRATE_ASYNC mode firstly */
1844	rc = migrate_pages_batch(from, get_new_folio, put_new_folio, private, MIGRATE_ASYNC,
1845				 reason, &folios, split_folios, &astats,
1846				 NR_MAX_MIGRATE_ASYNC_RETRY);
1847	stats->nr_succeeded += astats.nr_succeeded;
1848	stats->nr_thp_succeeded += astats.nr_thp_succeeded;
1849	stats->nr_thp_split += astats.nr_thp_split;
1850	stats->nr_split += astats.nr_split;
1851	if (rc < 0) {
1852		stats->nr_failed_pages += astats.nr_failed_pages;
1853		stats->nr_thp_failed += astats.nr_thp_failed;
1854		list_splice_tail(&folios, ret_folios);
1855		return rc;
1856	}
1857	stats->nr_thp_failed += astats.nr_thp_split;
1858	/*
1859	 * Do not count rc, as pages will be retried below.
1860	 * Count nr_split only, since it includes nr_thp_split.
1861	 */
1862	nr_failed += astats.nr_split;
1863	/*
1864	 * Fall back to migrate all failed folios one by one synchronously. All
1865	 * failed folios except split THPs will be retried, so their failure
1866	 * isn't counted
1867	 */
1868	list_splice_tail_init(&folios, from);
1869	while (!list_empty(from)) {
1870		list_move(from->next, &folios);
1871		rc = migrate_pages_batch(&folios, get_new_folio, put_new_folio,
1872					 private, mode, reason, ret_folios,
1873					 split_folios, stats, NR_MAX_MIGRATE_SYNC_RETRY);
1874		list_splice_tail_init(&folios, ret_folios);
1875		if (rc < 0)
1876			return rc;
1877		nr_failed += rc;
1878	}
1879
1880	return nr_failed;
1881}
1882
1883/*
1884 * migrate_pages - migrate the folios specified in a list, to the free folios
1885 *		   supplied as the target for the page migration
1886 *
1887 * @from:		The list of folios to be migrated.
1888 * @get_new_folio:	The function used to allocate free folios to be used
1889 *			as the target of the folio migration.
1890 * @put_new_folio:	The function used to free target folios if migration
1891 *			fails, or NULL if no special handling is necessary.
1892 * @private:		Private data to be passed on to get_new_folio()
1893 * @mode:		The migration mode that specifies the constraints for
1894 *			folio migration, if any.
1895 * @reason:		The reason for folio migration.
1896 * @ret_succeeded:	Set to the number of folios migrated successfully if
1897 *			the caller passes a non-NULL pointer.
1898 *
1899 * The function returns after NR_MAX_MIGRATE_PAGES_RETRY attempts or if no folios
1900 * are movable any more because the list has become empty or no retryable folios
1901 * exist any more. It is caller's responsibility to call putback_movable_pages()
1902 * only if ret != 0.
1903 *
1904 * Returns the number of {normal folio, large folio, hugetlb} that were not
1905 * migrated, or an error code. The number of large folio splits will be
1906 * considered as the number of non-migrated large folio, no matter how many
1907 * split folios of the large folio are migrated successfully.
1908 */
1909int migrate_pages(struct list_head *from, new_folio_t get_new_folio,
1910		free_folio_t put_new_folio, unsigned long private,
1911		enum migrate_mode mode, int reason, unsigned int *ret_succeeded)
1912{
1913	int rc, rc_gather;
1914	int nr_pages;
1915	struct folio *folio, *folio2;
1916	LIST_HEAD(folios);
1917	LIST_HEAD(ret_folios);
1918	LIST_HEAD(split_folios);
1919	struct migrate_pages_stats stats;
1920
1921	trace_mm_migrate_pages_start(mode, reason);
1922
1923	memset(&stats, 0, sizeof(stats));
1924
1925	rc_gather = migrate_hugetlbs(from, get_new_folio, put_new_folio, private,
1926				     mode, reason, &stats, &ret_folios);
1927	if (rc_gather < 0)
1928		goto out;
1929
1930again:
1931	nr_pages = 0;
1932	list_for_each_entry_safe(folio, folio2, from, lru) {
1933		/* Retried hugetlb folios will be kept in list  */
1934		if (folio_test_hugetlb(folio)) {
1935			list_move_tail(&folio->lru, &ret_folios);
1936			continue;
1937		}
1938
1939		nr_pages += folio_nr_pages(folio);
1940		if (nr_pages >= NR_MAX_BATCHED_MIGRATION)
1941			break;
1942	}
1943	if (nr_pages >= NR_MAX_BATCHED_MIGRATION)
1944		list_cut_before(&folios, from, &folio2->lru);
1945	else
1946		list_splice_init(from, &folios);
1947	if (mode == MIGRATE_ASYNC)
1948		rc = migrate_pages_batch(&folios, get_new_folio, put_new_folio,
1949				private, mode, reason, &ret_folios,
1950				&split_folios, &stats,
1951				NR_MAX_MIGRATE_PAGES_RETRY);
1952	else
1953		rc = migrate_pages_sync(&folios, get_new_folio, put_new_folio,
1954				private, mode, reason, &ret_folios,
1955				&split_folios, &stats);
1956	list_splice_tail_init(&folios, &ret_folios);
1957	if (rc < 0) {
1958		rc_gather = rc;
1959		list_splice_tail(&split_folios, &ret_folios);
1960		goto out;
1961	}
1962	if (!list_empty(&split_folios)) {
1963		/*
1964		 * Failure isn't counted since all split folios of a large folio
1965		 * is counted as 1 failure already.  And, we only try to migrate
1966		 * with minimal effort, force MIGRATE_ASYNC mode and retry once.
1967		 */
1968		migrate_pages_batch(&split_folios, get_new_folio,
1969				put_new_folio, private, MIGRATE_ASYNC, reason,
1970				&ret_folios, NULL, &stats, 1);
1971		list_splice_tail_init(&split_folios, &ret_folios);
1972	}
1973	rc_gather += rc;
1974	if (!list_empty(from))
1975		goto again;
1976out:
1977	/*
1978	 * Put the permanent failure folio back to migration list, they
1979	 * will be put back to the right list by the caller.
1980	 */
1981	list_splice(&ret_folios, from);
1982
1983	/*
1984	 * Return 0 in case all split folios of fail-to-migrate large folios
1985	 * are migrated successfully.
1986	 */
1987	if (list_empty(from))
1988		rc_gather = 0;
1989
1990	count_vm_events(PGMIGRATE_SUCCESS, stats.nr_succeeded);
1991	count_vm_events(PGMIGRATE_FAIL, stats.nr_failed_pages);
1992	count_vm_events(THP_MIGRATION_SUCCESS, stats.nr_thp_succeeded);
1993	count_vm_events(THP_MIGRATION_FAIL, stats.nr_thp_failed);
1994	count_vm_events(THP_MIGRATION_SPLIT, stats.nr_thp_split);
1995	trace_mm_migrate_pages(stats.nr_succeeded, stats.nr_failed_pages,
1996			       stats.nr_thp_succeeded, stats.nr_thp_failed,
1997			       stats.nr_thp_split, stats.nr_split, mode,
1998			       reason);
1999
2000	if (ret_succeeded)
2001		*ret_succeeded = stats.nr_succeeded;
2002
2003	return rc_gather;
2004}
2005
2006struct folio *alloc_migration_target(struct folio *src, unsigned long private)
2007{
2008	struct migration_target_control *mtc;
2009	gfp_t gfp_mask;
2010	unsigned int order = 0;
2011	int nid;
2012	int zidx;
2013
2014	mtc = (struct migration_target_control *)private;
2015	gfp_mask = mtc->gfp_mask;
2016	nid = mtc->nid;
2017	if (nid == NUMA_NO_NODE)
2018		nid = folio_nid(src);
2019
2020	if (folio_test_hugetlb(src)) {
2021		struct hstate *h = folio_hstate(src);
2022
2023		gfp_mask = htlb_modify_alloc_mask(h, gfp_mask);
2024		return alloc_hugetlb_folio_nodemask(h, nid,
2025						mtc->nmask, gfp_mask);
2026	}
2027
2028	if (folio_test_large(src)) {
2029		/*
2030		 * clear __GFP_RECLAIM to make the migration callback
2031		 * consistent with regular THP allocations.
2032		 */
2033		gfp_mask &= ~__GFP_RECLAIM;
2034		gfp_mask |= GFP_TRANSHUGE;
2035		order = folio_order(src);
2036	}
2037	zidx = zone_idx(folio_zone(src));
2038	if (is_highmem_idx(zidx) || zidx == ZONE_MOVABLE)
2039		gfp_mask |= __GFP_HIGHMEM;
2040
2041	return __folio_alloc(gfp_mask, order, nid, mtc->nmask);
2042}
2043
2044#ifdef CONFIG_NUMA
2045
2046static int store_status(int __user *status, int start, int value, int nr)
2047{
2048	while (nr-- > 0) {
2049		if (put_user(value, status + start))
2050			return -EFAULT;
2051		start++;
2052	}
2053
2054	return 0;
2055}
2056
2057static int do_move_pages_to_node(struct list_head *pagelist, int node)
2058{
2059	int err;
2060	struct migration_target_control mtc = {
2061		.nid = node,
2062		.gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
2063	};
2064
2065	err = migrate_pages(pagelist, alloc_migration_target, NULL,
2066		(unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
2067	if (err)
2068		putback_movable_pages(pagelist);
2069	return err;
2070}
2071
2072/*
2073 * Resolves the given address to a struct page, isolates it from the LRU and
2074 * puts it to the given pagelist.
2075 * Returns:
2076 *     errno - if the page cannot be found/isolated
2077 *     0 - when it doesn't have to be migrated because it is already on the
2078 *         target node
2079 *     1 - when it has been queued
2080 */
2081static int add_page_for_migration(struct mm_struct *mm, const void __user *p,
2082		int node, struct list_head *pagelist, bool migrate_all)
2083{
2084	struct vm_area_struct *vma;
2085	unsigned long addr;
2086	struct page *page;
2087	struct folio *folio;
2088	int err;
2089
2090	mmap_read_lock(mm);
2091	addr = (unsigned long)untagged_addr_remote(mm, p);
2092
2093	err = -EFAULT;
2094	vma = vma_lookup(mm, addr);
2095	if (!vma || !vma_migratable(vma))
2096		goto out;
2097
2098	/* FOLL_DUMP to ignore special (like zero) pages */
2099	page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
2100
2101	err = PTR_ERR(page);
2102	if (IS_ERR(page))
2103		goto out;
2104
2105	err = -ENOENT;
2106	if (!page)
2107		goto out;
2108
2109	folio = page_folio(page);
2110	if (folio_is_zone_device(folio))
2111		goto out_putfolio;
2112
2113	err = 0;
2114	if (folio_nid(folio) == node)
2115		goto out_putfolio;
2116
2117	err = -EACCES;
2118	if (page_mapcount(page) > 1 && !migrate_all)
2119		goto out_putfolio;
2120
2121	err = -EBUSY;
2122	if (folio_test_hugetlb(folio)) {
2123		if (isolate_hugetlb(folio, pagelist))
2124			err = 1;
2125	} else {
2126		if (!folio_isolate_lru(folio))
2127			goto out_putfolio;
2128
2129		err = 1;
2130		list_add_tail(&folio->lru, pagelist);
2131		node_stat_mod_folio(folio,
2132			NR_ISOLATED_ANON + folio_is_file_lru(folio),
2133			folio_nr_pages(folio));
2134	}
2135out_putfolio:
2136	/*
2137	 * Either remove the duplicate refcount from folio_isolate_lru()
2138	 * or drop the folio ref if it was not isolated.
2139	 */
2140	folio_put(folio);
2141out:
2142	mmap_read_unlock(mm);
2143	return err;
2144}
2145
2146static int move_pages_and_store_status(int node,
2147		struct list_head *pagelist, int __user *status,
2148		int start, int i, unsigned long nr_pages)
2149{
2150	int err;
2151
2152	if (list_empty(pagelist))
2153		return 0;
2154
2155	err = do_move_pages_to_node(pagelist, node);
2156	if (err) {
2157		/*
2158		 * Positive err means the number of failed
2159		 * pages to migrate.  Since we are going to
2160		 * abort and return the number of non-migrated
2161		 * pages, so need to include the rest of the
2162		 * nr_pages that have not been attempted as
2163		 * well.
2164		 */
2165		if (err > 0)
2166			err += nr_pages - i;
2167		return err;
2168	}
2169	return store_status(status, start, node, i - start);
2170}
2171
2172/*
2173 * Migrate an array of page address onto an array of nodes and fill
2174 * the corresponding array of status.
2175 */
2176static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
2177			 unsigned long nr_pages,
2178			 const void __user * __user *pages,
2179			 const int __user *nodes,
2180			 int __user *status, int flags)
2181{
2182	compat_uptr_t __user *compat_pages = (void __user *)pages;
2183	int current_node = NUMA_NO_NODE;
2184	LIST_HEAD(pagelist);
2185	int start, i;
2186	int err = 0, err1;
2187
2188	lru_cache_disable();
2189
2190	for (i = start = 0; i < nr_pages; i++) {
2191		const void __user *p;
2192		int node;
2193
2194		err = -EFAULT;
2195		if (in_compat_syscall()) {
2196			compat_uptr_t cp;
2197
2198			if (get_user(cp, compat_pages + i))
2199				goto out_flush;
2200
2201			p = compat_ptr(cp);
2202		} else {
2203			if (get_user(p, pages + i))
2204				goto out_flush;
2205		}
2206		if (get_user(node, nodes + i))
2207			goto out_flush;
2208
2209		err = -ENODEV;
2210		if (node < 0 || node >= MAX_NUMNODES)
2211			goto out_flush;
2212		if (!node_state(node, N_MEMORY))
2213			goto out_flush;
2214
2215		err = -EACCES;
2216		if (!node_isset(node, task_nodes))
2217			goto out_flush;
2218
2219		if (current_node == NUMA_NO_NODE) {
2220			current_node = node;
2221			start = i;
2222		} else if (node != current_node) {
2223			err = move_pages_and_store_status(current_node,
2224					&pagelist, status, start, i, nr_pages);
2225			if (err)
2226				goto out;
2227			start = i;
2228			current_node = node;
2229		}
2230
2231		/*
2232		 * Errors in the page lookup or isolation are not fatal and we simply
2233		 * report them via status
2234		 */
2235		err = add_page_for_migration(mm, p, current_node, &pagelist,
2236					     flags & MPOL_MF_MOVE_ALL);
2237
2238		if (err > 0) {
2239			/* The page is successfully queued for migration */
2240			continue;
2241		}
2242
2243		/*
2244		 * The move_pages() man page does not have an -EEXIST choice, so
2245		 * use -EFAULT instead.
2246		 */
2247		if (err == -EEXIST)
2248			err = -EFAULT;
2249
2250		/*
2251		 * If the page is already on the target node (!err), store the
2252		 * node, otherwise, store the err.
2253		 */
2254		err = store_status(status, i, err ? : current_node, 1);
2255		if (err)
2256			goto out_flush;
2257
2258		err = move_pages_and_store_status(current_node, &pagelist,
2259				status, start, i, nr_pages);
2260		if (err) {
2261			/* We have accounted for page i */
2262			if (err > 0)
2263				err--;
2264			goto out;
2265		}
2266		current_node = NUMA_NO_NODE;
2267	}
2268out_flush:
2269	/* Make sure we do not overwrite the existing error */
2270	err1 = move_pages_and_store_status(current_node, &pagelist,
2271				status, start, i, nr_pages);
2272	if (err >= 0)
2273		err = err1;
2274out:
2275	lru_cache_enable();
2276	return err;
2277}
2278
2279/*
2280 * Determine the nodes of an array of pages and store it in an array of status.
2281 */
2282static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages,
2283				const void __user **pages, int *status)
2284{
2285	unsigned long i;
2286
2287	mmap_read_lock(mm);
2288
2289	for (i = 0; i < nr_pages; i++) {
2290		unsigned long addr = (unsigned long)(*pages);
2291		struct vm_area_struct *vma;
2292		struct page *page;
2293		int err = -EFAULT;
2294
2295		vma = vma_lookup(mm, addr);
2296		if (!vma)
2297			goto set_status;
2298
2299		/* FOLL_DUMP to ignore special (like zero) pages */
2300		page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
2301
2302		err = PTR_ERR(page);
2303		if (IS_ERR(page))
2304			goto set_status;
2305
2306		err = -ENOENT;
2307		if (!page)
2308			goto set_status;
2309
2310		if (!is_zone_device_page(page))
2311			err = page_to_nid(page);
2312
2313		put_page(page);
2314set_status:
2315		*status = err;
2316
2317		pages++;
2318		status++;
2319	}
2320
2321	mmap_read_unlock(mm);
2322}
2323
2324static int get_compat_pages_array(const void __user *chunk_pages[],
2325				  const void __user * __user *pages,
2326				  unsigned long chunk_nr)
2327{
2328	compat_uptr_t __user *pages32 = (compat_uptr_t __user *)pages;
2329	compat_uptr_t p;
2330	int i;
2331
2332	for (i = 0; i < chunk_nr; i++) {
2333		if (get_user(p, pages32 + i))
2334			return -EFAULT;
2335		chunk_pages[i] = compat_ptr(p);
2336	}
2337
2338	return 0;
2339}
2340
2341/*
2342 * Determine the nodes of a user array of pages and store it in
2343 * a user array of status.
2344 */
2345static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages,
2346			 const void __user * __user *pages,
2347			 int __user *status)
2348{
2349#define DO_PAGES_STAT_CHUNK_NR 16UL
2350	const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR];
2351	int chunk_status[DO_PAGES_STAT_CHUNK_NR];
2352
2353	while (nr_pages) {
2354		unsigned long chunk_nr = min(nr_pages, DO_PAGES_STAT_CHUNK_NR);
2355
2356		if (in_compat_syscall()) {
2357			if (get_compat_pages_array(chunk_pages, pages,
2358						   chunk_nr))
2359				break;
2360		} else {
2361			if (copy_from_user(chunk_pages, pages,
2362				      chunk_nr * sizeof(*chunk_pages)))
2363				break;
2364		}
2365
2366		do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status);
2367
2368		if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status)))
2369			break;
2370
2371		pages += chunk_nr;
2372		status += chunk_nr;
2373		nr_pages -= chunk_nr;
2374	}
2375	return nr_pages ? -EFAULT : 0;
2376}
2377
2378static struct mm_struct *find_mm_struct(pid_t pid, nodemask_t *mem_nodes)
2379{
2380	struct task_struct *task;
2381	struct mm_struct *mm;
2382
2383	/*
2384	 * There is no need to check if current process has the right to modify
2385	 * the specified process when they are same.
2386	 */
2387	if (!pid) {
2388		mmget(current->mm);
2389		*mem_nodes = cpuset_mems_allowed(current);
2390		return current->mm;
2391	}
2392
2393	/* Find the mm_struct */
2394	rcu_read_lock();
2395	task = find_task_by_vpid(pid);
2396	if (!task) {
2397		rcu_read_unlock();
2398		return ERR_PTR(-ESRCH);
2399	}
2400	get_task_struct(task);
2401
2402	/*
2403	 * Check if this process has the right to modify the specified
2404	 * process. Use the regular "ptrace_may_access()" checks.
2405	 */
2406	if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
2407		rcu_read_unlock();
2408		mm = ERR_PTR(-EPERM);
2409		goto out;
2410	}
2411	rcu_read_unlock();
2412
2413	mm = ERR_PTR(security_task_movememory(task));
2414	if (IS_ERR(mm))
2415		goto out;
2416	*mem_nodes = cpuset_mems_allowed(task);
2417	mm = get_task_mm(task);
2418out:
2419	put_task_struct(task);
2420	if (!mm)
2421		mm = ERR_PTR(-EINVAL);
2422	return mm;
2423}
2424
2425/*
2426 * Move a list of pages in the address space of the currently executing
2427 * process.
2428 */
2429static int kernel_move_pages(pid_t pid, unsigned long nr_pages,
2430			     const void __user * __user *pages,
2431			     const int __user *nodes,
2432			     int __user *status, int flags)
2433{
2434	struct mm_struct *mm;
2435	int err;
2436	nodemask_t task_nodes;
2437
2438	/* Check flags */
2439	if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
2440		return -EINVAL;
2441
2442	if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
2443		return -EPERM;
2444
2445	mm = find_mm_struct(pid, &task_nodes);
2446	if (IS_ERR(mm))
2447		return PTR_ERR(mm);
2448
2449	if (nodes)
2450		err = do_pages_move(mm, task_nodes, nr_pages, pages,
2451				    nodes, status, flags);
2452	else
2453		err = do_pages_stat(mm, nr_pages, pages, status);
2454
2455	mmput(mm);
2456	return err;
2457}
2458
2459SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages,
2460		const void __user * __user *, pages,
2461		const int __user *, nodes,
2462		int __user *, status, int, flags)
2463{
2464	return kernel_move_pages(pid, nr_pages, pages, nodes, status, flags);
2465}
2466
2467#ifdef CONFIG_NUMA_BALANCING
2468/*
2469 * Returns true if this is a safe migration target node for misplaced NUMA
2470 * pages. Currently it only checks the watermarks which is crude.
2471 */
2472static bool migrate_balanced_pgdat(struct pglist_data *pgdat,
2473				   unsigned long nr_migrate_pages)
2474{
2475	int z;
2476
2477	for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2478		struct zone *zone = pgdat->node_zones + z;
2479
2480		if (!managed_zone(zone))
2481			continue;
2482
2483		/* Avoid waking kswapd by allocating pages_to_migrate pages. */
2484		if (!zone_watermark_ok(zone, 0,
2485				       high_wmark_pages(zone) +
2486				       nr_migrate_pages,
2487				       ZONE_MOVABLE, 0))
2488			continue;
2489		return true;
2490	}
2491	return false;
2492}
2493
2494static struct folio *alloc_misplaced_dst_folio(struct folio *src,
2495					   unsigned long data)
2496{
2497	int nid = (int) data;
2498	int order = folio_order(src);
2499	gfp_t gfp = __GFP_THISNODE;
2500
2501	if (order > 0)
2502		gfp |= GFP_TRANSHUGE_LIGHT;
2503	else {
2504		gfp |= GFP_HIGHUSER_MOVABLE | __GFP_NOMEMALLOC | __GFP_NORETRY |
2505			__GFP_NOWARN;
2506		gfp &= ~__GFP_RECLAIM;
2507	}
2508	return __folio_alloc_node(gfp, order, nid);
2509}
2510
2511static int numamigrate_isolate_folio(pg_data_t *pgdat, struct folio *folio)
2512{
2513	int nr_pages = folio_nr_pages(folio);
2514
2515	/* Avoid migrating to a node that is nearly full */
2516	if (!migrate_balanced_pgdat(pgdat, nr_pages)) {
2517		int z;
2518
2519		if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING))
2520			return 0;
2521		for (z = pgdat->nr_zones - 1; z >= 0; z--) {
2522			if (managed_zone(pgdat->node_zones + z))
2523				break;
2524		}
2525
2526		/*
2527		 * If there are no managed zones, it should not proceed
2528		 * further.
2529		 */
2530		if (z < 0)
2531			return 0;
2532
2533		wakeup_kswapd(pgdat->node_zones + z, 0,
2534			      folio_order(folio), ZONE_MOVABLE);
2535		return 0;
2536	}
2537
2538	if (!folio_isolate_lru(folio))
2539		return 0;
2540
2541	node_stat_mod_folio(folio, NR_ISOLATED_ANON + folio_is_file_lru(folio),
2542			    nr_pages);
2543
2544	/*
2545	 * Isolating the folio has taken another reference, so the
2546	 * caller's reference can be safely dropped without the folio
2547	 * disappearing underneath us during migration.
2548	 */
2549	folio_put(folio);
2550	return 1;
2551}
2552
2553/*
2554 * Attempt to migrate a misplaced folio to the specified destination
2555 * node. Caller is expected to have an elevated reference count on
2556 * the folio that will be dropped by this function before returning.
2557 */
2558int migrate_misplaced_folio(struct folio *folio, struct vm_area_struct *vma,
2559			    int node)
2560{
2561	pg_data_t *pgdat = NODE_DATA(node);
2562	int isolated;
2563	int nr_remaining;
2564	unsigned int nr_succeeded;
2565	LIST_HEAD(migratepages);
2566	int nr_pages = folio_nr_pages(folio);
2567
2568	/*
2569	 * Don't migrate file folios that are mapped in multiple processes
2570	 * with execute permissions as they are probably shared libraries.
2571	 * To check if the folio is shared, ideally we want to make sure
2572	 * every page is mapped to the same process. Doing that is very
2573	 * expensive, so check the estimated mapcount of the folio instead.
2574	 */
2575	if (folio_estimated_sharers(folio) != 1 && folio_is_file_lru(folio) &&
2576	    (vma->vm_flags & VM_EXEC))
2577		goto out;
2578
2579	/*
2580	 * Also do not migrate dirty folios as not all filesystems can move
2581	 * dirty folios in MIGRATE_ASYNC mode which is a waste of cycles.
2582	 */
2583	if (folio_is_file_lru(folio) && folio_test_dirty(folio))
2584		goto out;
2585
2586	isolated = numamigrate_isolate_folio(pgdat, folio);
2587	if (!isolated)
2588		goto out;
2589
2590	list_add(&folio->lru, &migratepages);
2591	nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_folio,
2592				     NULL, node, MIGRATE_ASYNC,
2593				     MR_NUMA_MISPLACED, &nr_succeeded);
2594	if (nr_remaining) {
2595		if (!list_empty(&migratepages)) {
2596			list_del(&folio->lru);
2597			node_stat_mod_folio(folio, NR_ISOLATED_ANON +
2598					folio_is_file_lru(folio), -nr_pages);
2599			folio_putback_lru(folio);
2600		}
2601		isolated = 0;
2602	}
2603	if (nr_succeeded) {
2604		count_vm_numa_events(NUMA_PAGE_MIGRATE, nr_succeeded);
2605		if (!node_is_toptier(folio_nid(folio)) && node_is_toptier(node))
2606			mod_node_page_state(pgdat, PGPROMOTE_SUCCESS,
2607					    nr_succeeded);
2608	}
2609	BUG_ON(!list_empty(&migratepages));
2610	return isolated;
2611
2612out:
2613	folio_put(folio);
2614	return 0;
2615}
2616#endif /* CONFIG_NUMA_BALANCING */
2617#endif /* CONFIG_NUMA */