migrate_device.c

   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Device Memory Migration functionality.
   4 *
   5 * Originally written by Jérôme Glisse.
   6 */
   7#include <linux/export.h>
   8#include <linux/memremap.h>
   9#include <linux/migrate.h>
  10#include <linux/mm.h>
  11#include <linux/mm_inline.h>
  12#include <linux/mmu_notifier.h>
  13#include <linux/oom.h>
  14#include <linux/pagewalk.h>
  15#include <linux/rmap.h>
  16#include <linux/leafops.h>
  17#include <linux/pgalloc.h>
  18#include <asm/tlbflush.h>
  19#include "internal.h"
  20
  21static int migrate_vma_collect_skip(unsigned long start,
  22				    unsigned long end,
  23				    struct mm_walk *walk)
  24{
  25	struct migrate_vma *migrate = walk->private;
  26	unsigned long addr;
  27
  28	for (addr = start; addr < end; addr += PAGE_SIZE) {
  29		migrate->dst[migrate->npages] = 0;
  30		migrate->src[migrate->npages++] = 0;
  31	}
  32
  33	return 0;
  34}
  35
  36static int migrate_vma_collect_hole(unsigned long start,
  37				    unsigned long end,
  38				    __always_unused int depth,
  39				    struct mm_walk *walk)
  40{
  41	struct migrate_vma *migrate = walk->private;
  42	unsigned long addr;
  43
  44	/* Only allow populating anonymous memory. */
  45	if (!vma_is_anonymous(walk->vma))
  46		return migrate_vma_collect_skip(start, end, walk);
  47
  48	if (thp_migration_supported() &&
  49		(migrate->flags & MIGRATE_VMA_SELECT_COMPOUND) &&
  50		(IS_ALIGNED(start, HPAGE_PMD_SIZE) &&
  51		 IS_ALIGNED(end, HPAGE_PMD_SIZE))) {
  52		migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE |
  53						MIGRATE_PFN_COMPOUND;
  54		migrate->dst[migrate->npages] = 0;
  55		migrate->npages++;
  56		migrate->cpages++;
  57
  58		/*
  59		 * Collect the remaining entries as holes, in case we
  60		 * need to split later
  61		 */
  62		return migrate_vma_collect_skip(start + PAGE_SIZE, end, walk);
  63	}
  64
  65	for (addr = start; addr < end; addr += PAGE_SIZE) {
  66		migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
  67		migrate->dst[migrate->npages] = 0;
  68		migrate->npages++;
  69		migrate->cpages++;
  70	}
  71
  72	return 0;
  73}
  74
  75/**
  76 * migrate_vma_split_folio() - Helper function to split a THP folio
  77 * @folio: the folio to split
  78 * @fault_page: struct page associated with the fault if any
  79 *
  80 * Returns 0 on success
  81 */
  82static int migrate_vma_split_folio(struct folio *folio,
  83				   struct page *fault_page)
  84{
  85	int ret;
  86	struct folio *fault_folio = fault_page ? page_folio(fault_page) : NULL;
  87	struct folio *new_fault_folio = NULL;
  88
  89	if (folio != fault_folio) {
  90		folio_get(folio);
  91		folio_lock(folio);
  92	}
  93
  94	ret = split_folio(folio);
  95	if (ret) {
  96		if (folio != fault_folio) {
  97			folio_unlock(folio);
  98			folio_put(folio);
  99		}
 100		return ret;
 101	}
 102
 103	new_fault_folio = fault_page ? page_folio(fault_page) : NULL;
 104
 105	/*
 106	 * Ensure the lock is held on the correct
 107	 * folio after the split
 108	 */
 109	if (!new_fault_folio) {
 110		folio_unlock(folio);
 111		folio_put(folio);
 112	} else if (folio != new_fault_folio) {
 113		if (new_fault_folio != fault_folio) {
 114			folio_get(new_fault_folio);
 115			folio_lock(new_fault_folio);
 116		}
 117		folio_unlock(folio);
 118		folio_put(folio);
 119	}
 120
 121	return 0;
 122}
 123
 124/** migrate_vma_collect_huge_pmd - collect THP pages without splitting the
 125 * folio for device private pages.
 126 * @pmdp: pointer to pmd entry
 127 * @start: start address of the range for migration
 128 * @end: end address of the range for migration
 129 * @walk: mm_walk callback structure
 130 * @fault_folio: folio associated with the fault if any
 131 *
 132 * Collect the huge pmd entry at @pmdp for migration and set the
 133 * MIGRATE_PFN_COMPOUND flag in the migrate src entry to indicate that
 134 * migration will occur at HPAGE_PMD granularity
 135 */
 136static int migrate_vma_collect_huge_pmd(pmd_t *pmdp, unsigned long start,
 137					unsigned long end, struct mm_walk *walk,
 138					struct folio *fault_folio)
 139{
 140	struct mm_struct *mm = walk->mm;
 141	struct folio *folio;
 142	struct migrate_vma *migrate = walk->private;
 143	spinlock_t *ptl;
 144	int ret;
 145	unsigned long write = 0;
 146
 147	ptl = pmd_lock(mm, pmdp);
 148	if (pmd_none(*pmdp)) {
 149		spin_unlock(ptl);
 150		return migrate_vma_collect_hole(start, end, -1, walk);
 151	}
 152
 153	if (pmd_trans_huge(*pmdp)) {
 154		if (!(migrate->flags & MIGRATE_VMA_SELECT_SYSTEM)) {
 155			spin_unlock(ptl);
 156			return migrate_vma_collect_skip(start, end, walk);
 157		}
 158
 159		folio = pmd_folio(*pmdp);
 160		if (is_huge_zero_folio(folio)) {
 161			spin_unlock(ptl);
 162			return migrate_vma_collect_hole(start, end, -1, walk);
 163		}
 164		if (pmd_write(*pmdp))
 165			write = MIGRATE_PFN_WRITE;
 166	} else if (!pmd_present(*pmdp)) {
 167		const softleaf_t entry = softleaf_from_pmd(*pmdp);
 168
 169		folio = softleaf_to_folio(entry);
 170
 171		if (!softleaf_is_device_private(entry) ||
 172			!(migrate->flags & MIGRATE_VMA_SELECT_DEVICE_PRIVATE) ||
 173			(folio->pgmap->owner != migrate->pgmap_owner)) {
 174			spin_unlock(ptl);
 175			return migrate_vma_collect_skip(start, end, walk);
 176		}
 177
 178		if (softleaf_is_migration(entry)) {
 179			softleaf_entry_wait_on_locked(entry, ptl);
 180			spin_unlock(ptl);
 181			return -EAGAIN;
 182		}
 183
 184		if (softleaf_is_device_private_write(entry))
 185			write = MIGRATE_PFN_WRITE;
 186	} else {
 187		spin_unlock(ptl);
 188		return -EAGAIN;
 189	}
 190
 191	folio_get(folio);
 192	if (folio != fault_folio && unlikely(!folio_trylock(folio))) {
 193		spin_unlock(ptl);
 194		folio_put(folio);
 195		return migrate_vma_collect_skip(start, end, walk);
 196	}
 197
 198	if (thp_migration_supported() &&
 199		(migrate->flags & MIGRATE_VMA_SELECT_COMPOUND) &&
 200		(IS_ALIGNED(start, HPAGE_PMD_SIZE) &&
 201		 IS_ALIGNED(end, HPAGE_PMD_SIZE))) {
 202
 203		struct page_vma_mapped_walk pvmw = {
 204			.ptl = ptl,
 205			.address = start,
 206			.pmd = pmdp,
 207			.vma = walk->vma,
 208		};
 209
 210		unsigned long pfn = page_to_pfn(folio_page(folio, 0));
 211
 212		migrate->src[migrate->npages] = migrate_pfn(pfn) | write
 213						| MIGRATE_PFN_MIGRATE
 214						| MIGRATE_PFN_COMPOUND;
 215		migrate->dst[migrate->npages++] = 0;
 216		migrate->cpages++;
 217		ret = set_pmd_migration_entry(&pvmw, folio_page(folio, 0));
 218		if (ret) {
 219			migrate->npages--;
 220			migrate->cpages--;
 221			migrate->src[migrate->npages] = 0;
 222			migrate->dst[migrate->npages] = 0;
 223			goto fallback;
 224		}
 225		migrate_vma_collect_skip(start + PAGE_SIZE, end, walk);
 226		spin_unlock(ptl);
 227		return 0;
 228	}
 229
 230fallback:
 231	spin_unlock(ptl);
 232	if (!folio_test_large(folio))
 233		goto done;
 234	ret = split_folio(folio);
 235	if (fault_folio != folio)
 236		folio_unlock(folio);
 237	folio_put(folio);
 238	if (ret)
 239		return migrate_vma_collect_skip(start, end, walk);
 240	if (pmd_none(pmdp_get_lockless(pmdp)))
 241		return migrate_vma_collect_hole(start, end, -1, walk);
 242
 243done:
 244	return -ENOENT;
 245}
 246
 247static int migrate_vma_collect_pmd(pmd_t *pmdp,
 248				   unsigned long start,
 249				   unsigned long end,
 250				   struct mm_walk *walk)
 251{
 252	struct migrate_vma *migrate = walk->private;
 253	struct vm_area_struct *vma = walk->vma;
 254	struct mm_struct *mm = vma->vm_mm;
 255	unsigned long addr = start, unmapped = 0;
 256	spinlock_t *ptl;
 257	struct folio *fault_folio = migrate->fault_page ?
 258		page_folio(migrate->fault_page) : NULL;
 259	pte_t *ptep;
 260
 261again:
 262	if (pmd_trans_huge(*pmdp) || !pmd_present(*pmdp)) {
 263		int ret = migrate_vma_collect_huge_pmd(pmdp, start, end, walk, fault_folio);
 264
 265		if (ret == -EAGAIN)
 266			goto again;
 267		if (ret == 0)
 268			return 0;
 269	}
 270
 271	ptep = pte_offset_map_lock(mm, pmdp, start, &ptl);
 272	if (!ptep)
 273		goto again;
 274	arch_enter_lazy_mmu_mode();
 275	ptep += (addr - start) / PAGE_SIZE;
 276
 277	for (; addr < end; addr += PAGE_SIZE, ptep++) {
 278		struct dev_pagemap *pgmap;
 279		unsigned long mpfn = 0, pfn;
 280		struct folio *folio;
 281		struct page *page;
 282		softleaf_t entry;
 283		pte_t pte;
 284
 285		pte = ptep_get(ptep);
 286
 287		if (pte_none(pte)) {
 288			if (vma_is_anonymous(vma)) {
 289				mpfn = MIGRATE_PFN_MIGRATE;
 290				migrate->cpages++;
 291			}
 292			goto next;
 293		}
 294
 295		if (!pte_present(pte)) {
 296			/*
 297			 * Only care about unaddressable device page special
 298			 * page table entry. Other special swap entries are not
 299			 * migratable, and we ignore regular swapped page.
 300			 */
 301			entry = softleaf_from_pte(pte);
 302			if (!softleaf_is_device_private(entry))
 303				goto next;
 304
 305			page = softleaf_to_page(entry);
 306			pgmap = page_pgmap(page);
 307			if (!(migrate->flags &
 308				MIGRATE_VMA_SELECT_DEVICE_PRIVATE) ||
 309			    pgmap->owner != migrate->pgmap_owner)
 310				goto next;
 311
 312			folio = page_folio(page);
 313			if (folio_test_large(folio)) {
 314				int ret;
 315
 316				arch_leave_lazy_mmu_mode();
 317				pte_unmap_unlock(ptep, ptl);
 318				ret = migrate_vma_split_folio(folio,
 319							  migrate->fault_page);
 320
 321				if (ret) {
 322					if (unmapped)
 323						flush_tlb_range(walk->vma, start, end);
 324
 325					return migrate_vma_collect_skip(addr, end, walk);
 326				}
 327
 328				goto again;
 329			}
 330
 331			mpfn = migrate_pfn(page_to_pfn(page)) |
 332					MIGRATE_PFN_MIGRATE;
 333			if (softleaf_is_device_private_write(entry))
 334				mpfn |= MIGRATE_PFN_WRITE;
 335		} else {
 336			pfn = pte_pfn(pte);
 337			if (is_zero_pfn(pfn) &&
 338			    (migrate->flags & MIGRATE_VMA_SELECT_SYSTEM)) {
 339				mpfn = MIGRATE_PFN_MIGRATE;
 340				migrate->cpages++;
 341				goto next;
 342			}
 343			page = vm_normal_page(migrate->vma, addr, pte);
 344			if (page && !is_zone_device_page(page) &&
 345			    !(migrate->flags & MIGRATE_VMA_SELECT_SYSTEM)) {
 346				goto next;
 347			} else if (page && is_device_coherent_page(page)) {
 348				pgmap = page_pgmap(page);
 349
 350				if (!(migrate->flags &
 351					MIGRATE_VMA_SELECT_DEVICE_COHERENT) ||
 352					pgmap->owner != migrate->pgmap_owner)
 353					goto next;
 354			}
 355			folio = page ? page_folio(page) : NULL;
 356			if (folio && folio_test_large(folio)) {
 357				int ret;
 358
 359				arch_leave_lazy_mmu_mode();
 360				pte_unmap_unlock(ptep, ptl);
 361				ret = migrate_vma_split_folio(folio,
 362							  migrate->fault_page);
 363
 364				if (ret) {
 365					if (unmapped)
 366						flush_tlb_range(walk->vma, start, end);
 367
 368					return migrate_vma_collect_skip(addr, end, walk);
 369				}
 370
 371				goto again;
 372			}
 373			mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
 374			mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
 375		}
 376
 377		if (!page || !page->mapping) {
 378			mpfn = 0;
 379			goto next;
 380		}
 381
 382		/*
 383		 * By getting a reference on the folio we pin it and that blocks
 384		 * any kind of migration. Side effect is that it "freezes" the
 385		 * pte.
 386		 *
 387		 * We drop this reference after isolating the folio from the lru
 388		 * for non device folio (device folio are not on the lru and thus
 389		 * can't be dropped from it).
 390		 */
 391		folio = page_folio(page);
 392		folio_get(folio);
 393
 394		/*
 395		 * We rely on folio_trylock() to avoid deadlock between
 396		 * concurrent migrations where each is waiting on the others
 397		 * folio lock. If we can't immediately lock the folio we fail this
 398		 * migration as it is only best effort anyway.
 399		 *
 400		 * If we can lock the folio it's safe to set up a migration entry
 401		 * now. In the common case where the folio is mapped once in a
 402		 * single process setting up the migration entry now is an
 403		 * optimisation to avoid walking the rmap later with
 404		 * try_to_migrate().
 405		 */
 406		if (fault_folio == folio || folio_trylock(folio)) {
 407			bool anon_exclusive;
 408			pte_t swp_pte;
 409
 410			flush_cache_page(vma, addr, pte_pfn(pte));
 411			anon_exclusive = folio_test_anon(folio) &&
 412					  PageAnonExclusive(page);
 413			if (anon_exclusive) {
 414				pte = ptep_clear_flush(vma, addr, ptep);
 415
 416				if (folio_try_share_anon_rmap_pte(folio, page)) {
 417					set_pte_at(mm, addr, ptep, pte);
 418					if (fault_folio != folio)
 419						folio_unlock(folio);
 420					folio_put(folio);
 421					mpfn = 0;
 422					goto next;
 423				}
 424			} else {
 425				pte = ptep_get_and_clear(mm, addr, ptep);
 426			}
 427
 428			migrate->cpages++;
 429
 430			/* Set the dirty flag on the folio now the pte is gone. */
 431			if (pte_dirty(pte))
 432				folio_mark_dirty(folio);
 433
 434			/* Setup special migration page table entry */
 435			if (mpfn & MIGRATE_PFN_WRITE)
 436				entry = make_writable_migration_entry(
 437							page_to_pfn(page));
 438			else if (anon_exclusive)
 439				entry = make_readable_exclusive_migration_entry(
 440							page_to_pfn(page));
 441			else
 442				entry = make_readable_migration_entry(
 443							page_to_pfn(page));
 444			if (pte_present(pte)) {
 445				if (pte_young(pte))
 446					entry = make_migration_entry_young(entry);
 447				if (pte_dirty(pte))
 448					entry = make_migration_entry_dirty(entry);
 449			}
 450			swp_pte = swp_entry_to_pte(entry);
 451			if (pte_present(pte)) {
 452				if (pte_soft_dirty(pte))
 453					swp_pte = pte_swp_mksoft_dirty(swp_pte);
 454				if (pte_uffd_wp(pte))
 455					swp_pte = pte_swp_mkuffd_wp(swp_pte);
 456			} else {
 457				if (pte_swp_soft_dirty(pte))
 458					swp_pte = pte_swp_mksoft_dirty(swp_pte);
 459				if (pte_swp_uffd_wp(pte))
 460					swp_pte = pte_swp_mkuffd_wp(swp_pte);
 461			}
 462			set_pte_at(mm, addr, ptep, swp_pte);
 463
 464			/*
 465			 * This is like regular unmap: we remove the rmap and
 466			 * drop the folio refcount. The folio won't be freed, as
 467			 * we took a reference just above.
 468			 */
 469			folio_remove_rmap_pte(folio, page, vma);
 470			folio_put(folio);
 471
 472			if (pte_present(pte))
 473				unmapped++;
 474		} else {
 475			folio_put(folio);
 476			mpfn = 0;
 477		}
 478
 479next:
 480		migrate->dst[migrate->npages] = 0;
 481		migrate->src[migrate->npages++] = mpfn;
 482	}
 483
 484	/* Only flush the TLB if we actually modified any entries */
 485	if (unmapped)
 486		flush_tlb_range(walk->vma, start, end);
 487
 488	arch_leave_lazy_mmu_mode();
 489	pte_unmap_unlock(ptep - 1, ptl);
 490
 491	return 0;
 492}
 493
 494static const struct mm_walk_ops migrate_vma_walk_ops = {
 495	.pmd_entry		= migrate_vma_collect_pmd,
 496	.pte_hole		= migrate_vma_collect_hole,
 497	.walk_lock		= PGWALK_RDLOCK,
 498};
 499
 500/*
 501 * migrate_vma_collect() - collect pages over a range of virtual addresses
 502 * @migrate: migrate struct containing all migration information
 503 *
 504 * This will walk the CPU page table. For each virtual address backed by a
 505 * valid page, it updates the src array and takes a reference on the page, in
 506 * order to pin the page until we lock it and unmap it.
 507 */
 508static void migrate_vma_collect(struct migrate_vma *migrate)
 509{
 510	struct mmu_notifier_range range;
 511
 512	/*
 513	 * Note that the pgmap_owner is passed to the mmu notifier callback so
 514	 * that the registered device driver can skip invalidating device
 515	 * private page mappings that won't be migrated.
 516	 */
 517	mmu_notifier_range_init_owner(&range, MMU_NOTIFY_MIGRATE, 0,
 518		migrate->vma->vm_mm, migrate->start, migrate->end,
 519		migrate->pgmap_owner);
 520	mmu_notifier_invalidate_range_start(&range);
 521
 522	walk_page_range(migrate->vma->vm_mm, migrate->start, migrate->end,
 523			&migrate_vma_walk_ops, migrate);
 524
 525	mmu_notifier_invalidate_range_end(&range);
 526	migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
 527}
 528
 529/*
 530 * migrate_vma_check_page() - check if page is pinned or not
 531 * @page: struct page to check
 532 *
 533 * Pinned pages cannot be migrated. This is the same test as in
 534 * folio_migrate_mapping(), except that here we allow migration of a
 535 * ZONE_DEVICE page.
 536 */
 537static bool migrate_vma_check_page(struct page *page, struct page *fault_page)
 538{
 539	struct folio *folio = page_folio(page);
 540
 541	/*
 542	 * One extra ref because caller holds an extra reference, either from
 543	 * folio_isolate_lru() for a regular folio, or migrate_vma_collect() for
 544	 * a device folio.
 545	 */
 546	int extra = 1 + (page == fault_page);
 547
 548	/* Page from ZONE_DEVICE have one extra reference */
 549	if (folio_is_zone_device(folio))
 550		extra++;
 551
 552	/* For file back page */
 553	if (folio_mapping(folio))
 554		extra += 1 + folio_has_private(folio);
 555
 556	if ((folio_ref_count(folio) - extra) > folio_mapcount(folio))
 557		return false;
 558
 559	return true;
 560}
 561
 562/*
 563 * Unmaps pages for migration. Returns number of source pfns marked as
 564 * migrating.
 565 */
 566static unsigned long migrate_device_unmap(unsigned long *src_pfns,
 567					  unsigned long npages,
 568					  struct page *fault_page)
 569{
 570	struct folio *fault_folio = fault_page ?
 571		page_folio(fault_page) : NULL;
 572	unsigned long i, restore = 0;
 573	bool allow_drain = true;
 574	unsigned long unmapped = 0;
 575
 576	lru_add_drain();
 577
 578	for (i = 0; i < npages; ) {
 579		struct page *page = migrate_pfn_to_page(src_pfns[i]);
 580		struct folio *folio;
 581		unsigned int nr = 1;
 582
 583		if (!page) {
 584			if (src_pfns[i] & MIGRATE_PFN_MIGRATE)
 585				unmapped++;
 586			goto next;
 587		}
 588
 589		folio =	page_folio(page);
 590		nr = folio_nr_pages(folio);
 591
 592		if (nr > 1)
 593			src_pfns[i] |= MIGRATE_PFN_COMPOUND;
 594
 595
 596		/* ZONE_DEVICE folios are not on LRU */
 597		if (!folio_is_zone_device(folio)) {
 598			if (!folio_test_lru(folio) && allow_drain) {
 599				/* Drain CPU's lru cache */
 600				lru_add_drain_all();
 601				allow_drain = false;
 602			}
 603
 604			if (!folio_isolate_lru(folio)) {
 605				src_pfns[i] &= ~MIGRATE_PFN_MIGRATE;
 606				restore++;
 607				goto next;
 608			}
 609
 610			/* Drop the reference we took in collect */
 611			folio_put(folio);
 612		}
 613
 614		if (folio_mapped(folio))
 615			try_to_migrate(folio, 0);
 616
 617		if (folio_mapped(folio) ||
 618		    !migrate_vma_check_page(page, fault_page)) {
 619			if (!folio_is_zone_device(folio)) {
 620				folio_get(folio);
 621				folio_putback_lru(folio);
 622			}
 623
 624			src_pfns[i] &= ~MIGRATE_PFN_MIGRATE;
 625			restore++;
 626			goto next;
 627		}
 628
 629		unmapped++;
 630next:
 631		i += nr;
 632	}
 633
 634	for (i = 0; i < npages && restore; i++) {
 635		struct page *page = migrate_pfn_to_page(src_pfns[i]);
 636		struct folio *folio;
 637
 638		if (!page || (src_pfns[i] & MIGRATE_PFN_MIGRATE))
 639			continue;
 640
 641		folio = page_folio(page);
 642		remove_migration_ptes(folio, folio, 0);
 643
 644		src_pfns[i] = 0;
 645		if (fault_folio != folio)
 646			folio_unlock(folio);
 647		folio_put(folio);
 648		restore--;
 649	}
 650
 651	return unmapped;
 652}
 653
 654/*
 655 * migrate_vma_unmap() - replace page mapping with special migration pte entry
 656 * @migrate: migrate struct containing all migration information
 657 *
 658 * Isolate pages from the LRU and replace mappings (CPU page table pte) with a
 659 * special migration pte entry and check if it has been pinned. Pinned pages are
 660 * restored because we cannot migrate them.
 661 *
 662 * This is the last step before we call the device driver callback to allocate
 663 * destination memory and copy contents of original page over to new page.
 664 */
 665static void migrate_vma_unmap(struct migrate_vma *migrate)
 666{
 667	migrate->cpages = migrate_device_unmap(migrate->src, migrate->npages,
 668					migrate->fault_page);
 669}
 670
 671/**
 672 * migrate_vma_setup() - prepare to migrate a range of memory
 673 * @args: contains the vma, start, and pfns arrays for the migration
 674 *
 675 * Returns: negative errno on failures, 0 when 0 or more pages were migrated
 676 * without an error.
 677 *
 678 * Prepare to migrate a range of memory virtual address range by collecting all
 679 * the pages backing each virtual address in the range, saving them inside the
 680 * src array.  Then lock those pages and unmap them. Once the pages are locked
 681 * and unmapped, check whether each page is pinned or not.  Pages that aren't
 682 * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the
 683 * corresponding src array entry.  Then restores any pages that are pinned, by
 684 * remapping and unlocking those pages.
 685 *
 686 * The caller should then allocate destination memory and copy source memory to
 687 * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE
 688 * flag set).  Once these are allocated and copied, the caller must update each
 689 * corresponding entry in the dst array with the pfn value of the destination
 690 * page and with MIGRATE_PFN_VALID. Destination pages must be locked via
 691 * lock_page().
 692 *
 693 * Note that the caller does not have to migrate all the pages that are marked
 694 * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from
 695 * device memory to system memory.  If the caller cannot migrate a device page
 696 * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe
 697 * consequences for the userspace process, so it must be avoided if at all
 698 * possible.
 699 *
 700 * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we
 701 * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus
 702 * allowing the caller to allocate device memory for those unbacked virtual
 703 * addresses.  For this the caller simply has to allocate device memory and
 704 * properly set the destination entry like for regular migration.  Note that
 705 * this can still fail, and thus inside the device driver you must check if the
 706 * migration was successful for those entries after calling migrate_vma_pages(),
 707 * just like for regular migration.
 708 *
 709 * After that, the callers must call migrate_vma_pages() to go over each entry
 710 * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
 711 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
 712 * then migrate_vma_pages() to migrate struct page information from the source
 713 * struct page to the destination struct page.  If it fails to migrate the
 714 * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the
 715 * src array.
 716 *
 717 * At this point all successfully migrated pages have an entry in the src
 718 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
 719 * array entry with MIGRATE_PFN_VALID flag set.
 720 *
 721 * Once migrate_vma_pages() returns the caller may inspect which pages were
 722 * successfully migrated, and which were not.  Successfully migrated pages will
 723 * have the MIGRATE_PFN_MIGRATE flag set for their src array entry.
 724 *
 725 * It is safe to update device page table after migrate_vma_pages() because
 726 * both destination and source page are still locked, and the mmap_lock is held
 727 * in read mode (hence no one can unmap the range being migrated).
 728 *
 729 * Once the caller is done cleaning up things and updating its page table (if it
 730 * chose to do so, this is not an obligation) it finally calls
 731 * migrate_vma_finalize() to update the CPU page table to point to new pages
 732 * for successfully migrated pages or otherwise restore the CPU page table to
 733 * point to the original source pages.
 734 */
 735int migrate_vma_setup(struct migrate_vma *args)
 736{
 737	long nr_pages = (args->end - args->start) >> PAGE_SHIFT;
 738
 739	args->start &= PAGE_MASK;
 740	args->end &= PAGE_MASK;
 741	if (!args->vma || is_vm_hugetlb_page(args->vma) ||
 742	    (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma))
 743		return -EINVAL;
 744	if (nr_pages <= 0)
 745		return -EINVAL;
 746	if (args->start < args->vma->vm_start ||
 747	    args->start >= args->vma->vm_end)
 748		return -EINVAL;
 749	if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end)
 750		return -EINVAL;
 751	if (!args->src || !args->dst)
 752		return -EINVAL;
 753	if (args->fault_page && !is_device_private_page(args->fault_page))
 754		return -EINVAL;
 755	if (args->fault_page && !PageLocked(args->fault_page))
 756		return -EINVAL;
 757
 758	memset(args->src, 0, sizeof(*args->src) * nr_pages);
 759	args->cpages = 0;
 760	args->npages = 0;
 761
 762	migrate_vma_collect(args);
 763
 764	if (args->cpages)
 765		migrate_vma_unmap(args);
 766
 767	/*
 768	 * At this point pages are locked and unmapped, and thus they have
 769	 * stable content and can safely be copied to destination memory that
 770	 * is allocated by the drivers.
 771	 */
 772	return 0;
 773
 774}
 775EXPORT_SYMBOL(migrate_vma_setup);
 776
 777#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
 778/**
 779 * migrate_vma_insert_huge_pmd_page: Insert a huge folio into @migrate->vma->vm_mm
 780 * at @addr. folio is already allocated as a part of the migration process with
 781 * large page.
 782 *
 783 * @page needs to be initialized and setup after it's allocated. The code bits
 784 * here follow closely the code in __do_huge_pmd_anonymous_page(). This API does
 785 * not support THP zero pages.
 786 *
 787 * @migrate: migrate_vma arguments
 788 * @addr: address where the folio will be inserted
 789 * @page: page to be inserted at @addr
 790 * @src: src pfn which is being migrated
 791 * @pmdp: pointer to the pmd
 792 */
 793static int migrate_vma_insert_huge_pmd_page(struct migrate_vma *migrate,
 794					 unsigned long addr,
 795					 struct page *page,
 796					 unsigned long *src,
 797					 pmd_t *pmdp)
 798{
 799	struct vm_area_struct *vma = migrate->vma;
 800	gfp_t gfp = vma_thp_gfp_mask(vma);
 801	struct folio *folio = page_folio(page);
 802	int ret;
 803	vm_fault_t csa_ret;
 804	spinlock_t *ptl;
 805	pgtable_t pgtable;
 806	pmd_t entry;
 807	bool flush = false;
 808	unsigned long i;
 809
 810	VM_WARN_ON_FOLIO(!folio, folio);
 811	VM_WARN_ON_ONCE(!pmd_none(*pmdp) && !is_huge_zero_pmd(*pmdp));
 812
 813	if (!thp_vma_suitable_order(vma, addr, HPAGE_PMD_ORDER))
 814		return -EINVAL;
 815
 816	ret = anon_vma_prepare(vma);
 817	if (ret)
 818		return ret;
 819
 820	folio_set_order(folio, HPAGE_PMD_ORDER);
 821	folio_set_large_rmappable(folio);
 822
 823	if (mem_cgroup_charge(folio, migrate->vma->vm_mm, gfp)) {
 824		count_vm_event(THP_FAULT_FALLBACK);
 825		count_mthp_stat(HPAGE_PMD_ORDER, MTHP_STAT_ANON_FAULT_FALLBACK_CHARGE);
 826		ret = -ENOMEM;
 827		goto abort;
 828	}
 829
 830	__folio_mark_uptodate(folio);
 831
 832	pgtable = pte_alloc_one(vma->vm_mm);
 833	if (unlikely(!pgtable))
 834		goto abort;
 835
 836	if (folio_is_device_private(folio)) {
 837		swp_entry_t swp_entry;
 838
 839		if (vma->vm_flags & VM_WRITE)
 840			swp_entry = make_writable_device_private_entry(
 841						page_to_pfn(page));
 842		else
 843			swp_entry = make_readable_device_private_entry(
 844						page_to_pfn(page));
 845		entry = swp_entry_to_pmd(swp_entry);
 846	} else {
 847		if (folio_is_zone_device(folio) &&
 848		    !folio_is_device_coherent(folio)) {
 849			goto abort;
 850		}
 851		entry = folio_mk_pmd(folio, vma->vm_page_prot);
 852		if (vma->vm_flags & VM_WRITE)
 853			entry = pmd_mkwrite(pmd_mkdirty(entry), vma);
 854	}
 855
 856	ptl = pmd_lock(vma->vm_mm, pmdp);
 857	csa_ret = check_stable_address_space(vma->vm_mm);
 858	if (csa_ret)
 859		goto abort;
 860
 861	/*
 862	 * Check for userfaultfd but do not deliver the fault. Instead,
 863	 * just back off.
 864	 */
 865	if (userfaultfd_missing(vma))
 866		goto unlock_abort;
 867
 868	if (!pmd_none(*pmdp)) {
 869		if (!is_huge_zero_pmd(*pmdp))
 870			goto unlock_abort;
 871		flush = true;
 872	} else if (!pmd_none(*pmdp))
 873		goto unlock_abort;
 874
 875	add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 876	folio_add_new_anon_rmap(folio, vma, addr, RMAP_EXCLUSIVE);
 877	if (!folio_is_zone_device(folio))
 878		folio_add_lru_vma(folio, vma);
 879	folio_get(folio);
 880
 881	if (flush) {
 882		pte_free(vma->vm_mm, pgtable);
 883		flush_cache_page(vma, addr, addr + HPAGE_PMD_SIZE);
 884		pmdp_invalidate(vma, addr, pmdp);
 885	} else {
 886		pgtable_trans_huge_deposit(vma->vm_mm, pmdp, pgtable);
 887		mm_inc_nr_ptes(vma->vm_mm);
 888	}
 889	set_pmd_at(vma->vm_mm, addr, pmdp, entry);
 890	update_mmu_cache_pmd(vma, addr, pmdp);
 891
 892	spin_unlock(ptl);
 893
 894	count_vm_event(THP_FAULT_ALLOC);
 895	count_mthp_stat(HPAGE_PMD_ORDER, MTHP_STAT_ANON_FAULT_ALLOC);
 896	count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
 897
 898	return 0;
 899
 900unlock_abort:
 901	spin_unlock(ptl);
 902abort:
 903	for (i = 0; i < HPAGE_PMD_NR; i++)
 904		src[i] &= ~MIGRATE_PFN_MIGRATE;
 905	return 0;
 906}
 907
 908static int migrate_vma_split_unmapped_folio(struct migrate_vma *migrate,
 909					    unsigned long idx, unsigned long addr,
 910					    struct folio *folio)
 911{
 912	unsigned long i;
 913	unsigned long pfn;
 914	unsigned long flags;
 915	int ret = 0;
 916
 917	folio_get(folio);
 918	split_huge_pmd_address(migrate->vma, addr, true);
 919	ret = folio_split_unmapped(folio, 0);
 920	if (ret)
 921		return ret;
 922	migrate->src[idx] &= ~MIGRATE_PFN_COMPOUND;
 923	flags = migrate->src[idx] & ((1UL << MIGRATE_PFN_SHIFT) - 1);
 924	pfn = migrate->src[idx] >> MIGRATE_PFN_SHIFT;
 925	for (i = 1; i < HPAGE_PMD_NR; i++)
 926		migrate->src[i+idx] = migrate_pfn(pfn + i) | flags;
 927	return ret;
 928}
 929#else /* !CONFIG_ARCH_ENABLE_THP_MIGRATION */
 930static int migrate_vma_insert_huge_pmd_page(struct migrate_vma *migrate,
 931					 unsigned long addr,
 932					 struct page *page,
 933					 unsigned long *src,
 934					 pmd_t *pmdp)
 935{
 936	return 0;
 937}
 938
 939static int migrate_vma_split_unmapped_folio(struct migrate_vma *migrate,
 940					    unsigned long idx, unsigned long addr,
 941					    struct folio *folio)
 942{
 943	return 0;
 944}
 945#endif
 946
 947static unsigned long migrate_vma_nr_pages(unsigned long *src)
 948{
 949	unsigned long nr = 1;
 950#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
 951	if (*src & MIGRATE_PFN_COMPOUND)
 952		nr = HPAGE_PMD_NR;
 953#else
 954	if (*src & MIGRATE_PFN_COMPOUND)
 955		VM_WARN_ON_ONCE(true);
 956#endif
 957	return nr;
 958}
 959
 960/*
 961 * This code closely matches the code in:
 962 *   __handle_mm_fault()
 963 *     handle_pte_fault()
 964 *       do_anonymous_page()
 965 * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE
 966 * private or coherent page.
 967 */
 968static void migrate_vma_insert_page(struct migrate_vma *migrate,
 969				    unsigned long addr,
 970				    unsigned long *dst,
 971				    unsigned long *src)
 972{
 973	struct page *page = migrate_pfn_to_page(*dst);
 974	struct folio *folio = page_folio(page);
 975	struct vm_area_struct *vma = migrate->vma;
 976	struct mm_struct *mm = vma->vm_mm;
 977	bool flush = false;
 978	spinlock_t *ptl;
 979	pte_t entry;
 980	pgd_t *pgdp;
 981	p4d_t *p4dp;
 982	pud_t *pudp;
 983	pmd_t *pmdp;
 984	pte_t *ptep;
 985	pte_t orig_pte;
 986
 987	/* Only allow populating anonymous memory */
 988	if (!vma_is_anonymous(vma))
 989		goto abort;
 990
 991	pgdp = pgd_offset(mm, addr);
 992	p4dp = p4d_alloc(mm, pgdp, addr);
 993	if (!p4dp)
 994		goto abort;
 995	pudp = pud_alloc(mm, p4dp, addr);
 996	if (!pudp)
 997		goto abort;
 998	pmdp = pmd_alloc(mm, pudp, addr);
 999	if (!pmdp)
1000		goto abort;
1001
1002	if (thp_migration_supported() && (*dst & MIGRATE_PFN_COMPOUND)) {
1003		int ret = migrate_vma_insert_huge_pmd_page(migrate, addr, page,
1004								src, pmdp);
1005		if (ret)
1006			goto abort;
1007		return;
1008	}
1009
1010	if (!pmd_none(*pmdp)) {
1011		if (pmd_trans_huge(*pmdp)) {
1012			if (!is_huge_zero_pmd(*pmdp))
1013				goto abort;
1014			split_huge_pmd(vma, pmdp, addr);
1015		} else if (pmd_leaf(*pmdp))
1016			goto abort;
1017	}
1018
1019	if (pte_alloc(mm, pmdp))
1020		goto abort;
1021	if (unlikely(anon_vma_prepare(vma)))
1022		goto abort;
1023	if (mem_cgroup_charge(folio, vma->vm_mm, GFP_KERNEL))
1024		goto abort;
1025
1026	/*
1027	 * The memory barrier inside __folio_mark_uptodate makes sure that
1028	 * preceding stores to the folio contents become visible before
1029	 * the set_pte_at() write.
1030	 */
1031	__folio_mark_uptodate(folio);
1032
1033	if (folio_is_device_private(folio)) {
1034		swp_entry_t swp_entry;
1035
1036		if (vma->vm_flags & VM_WRITE)
1037			swp_entry = make_writable_device_private_entry(
1038						page_to_pfn(page));
1039		else
1040			swp_entry = make_readable_device_private_entry(
1041						page_to_pfn(page));
1042		entry = swp_entry_to_pte(swp_entry);
1043	} else {
1044		if (folio_is_zone_device(folio) &&
1045		    !folio_is_device_coherent(folio)) {
1046			pr_warn_once("Unsupported ZONE_DEVICE page type.\n");
1047			goto abort;
1048		}
1049		entry = mk_pte(page, vma->vm_page_prot);
1050		if (vma->vm_flags & VM_WRITE)
1051			entry = pte_mkwrite(pte_mkdirty(entry), vma);
1052	}
1053
1054	ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
1055	if (!ptep)
1056		goto abort;
1057	orig_pte = ptep_get(ptep);
1058
1059	if (check_stable_address_space(mm))
1060		goto unlock_abort;
1061
1062	if (pte_present(orig_pte)) {
1063		unsigned long pfn = pte_pfn(orig_pte);
1064
1065		if (!is_zero_pfn(pfn))
1066			goto unlock_abort;
1067		flush = true;
1068	} else if (!pte_none(orig_pte))
1069		goto unlock_abort;
1070
1071	/*
1072	 * Check for userfaultfd but do not deliver the fault. Instead,
1073	 * just back off.
1074	 */
1075	if (userfaultfd_missing(vma))
1076		goto unlock_abort;
1077
1078	inc_mm_counter(mm, MM_ANONPAGES);
1079	folio_add_new_anon_rmap(folio, vma, addr, RMAP_EXCLUSIVE);
1080	if (!folio_is_zone_device(folio))
1081		folio_add_lru_vma(folio, vma);
1082	folio_get(folio);
1083
1084	if (flush) {
1085		flush_cache_page(vma, addr, pte_pfn(orig_pte));
1086		ptep_clear_flush(vma, addr, ptep);
1087	}
1088	set_pte_at(mm, addr, ptep, entry);
1089	update_mmu_cache(vma, addr, ptep);
1090
1091	pte_unmap_unlock(ptep, ptl);
1092	*src = MIGRATE_PFN_MIGRATE;
1093	return;
1094
1095unlock_abort:
1096	pte_unmap_unlock(ptep, ptl);
1097abort:
1098	*src &= ~MIGRATE_PFN_MIGRATE;
1099}
1100
1101static void __migrate_device_pages(unsigned long *src_pfns,
1102				unsigned long *dst_pfns, unsigned long npages,
1103				struct migrate_vma *migrate)
1104{
1105	struct mmu_notifier_range range;
1106	unsigned long i, j;
1107	bool notified = false;
1108	unsigned long addr;
1109
1110	for (i = 0; i < npages; ) {
1111		struct page *newpage = migrate_pfn_to_page(dst_pfns[i]);
1112		struct page *page = migrate_pfn_to_page(src_pfns[i]);
1113		struct address_space *mapping;
1114		struct folio *newfolio, *folio;
1115		int r, extra_cnt = 0;
1116		unsigned long nr = 1;
1117
1118		if (!newpage) {
1119			src_pfns[i] &= ~MIGRATE_PFN_MIGRATE;
1120			goto next;
1121		}
1122
1123		if (!page) {
1124			unsigned long addr;
1125
1126			if (!(src_pfns[i] & MIGRATE_PFN_MIGRATE))
1127				goto next;
1128
1129			/*
1130			 * The only time there is no vma is when called from
1131			 * migrate_device_coherent_folio(). However this isn't
1132			 * called if the page could not be unmapped.
1133			 */
1134			VM_BUG_ON(!migrate);
1135			addr = migrate->start + i*PAGE_SIZE;
1136			if (!notified) {
1137				notified = true;
1138
1139				mmu_notifier_range_init_owner(&range,
1140					MMU_NOTIFY_MIGRATE, 0,
1141					migrate->vma->vm_mm, addr, migrate->end,
1142					migrate->pgmap_owner);
1143				mmu_notifier_invalidate_range_start(&range);
1144			}
1145
1146			if ((src_pfns[i] & MIGRATE_PFN_COMPOUND) &&
1147				(!(dst_pfns[i] & MIGRATE_PFN_COMPOUND))) {
1148				nr = migrate_vma_nr_pages(&src_pfns[i]);
1149				src_pfns[i] &= ~MIGRATE_PFN_COMPOUND;
1150			} else {
1151				nr = 1;
1152			}
1153
1154			for (j = 0; j < nr && i + j < npages; j++) {
1155				src_pfns[i+j] |= MIGRATE_PFN_MIGRATE;
1156				migrate_vma_insert_page(migrate,
1157					addr + j * PAGE_SIZE,
1158					&dst_pfns[i+j], &src_pfns[i+j]);
1159			}
1160			goto next;
1161		}
1162
1163		newfolio = page_folio(newpage);
1164		folio = page_folio(page);
1165		mapping = folio_mapping(folio);
1166
1167		/*
1168		 * If THP migration is enabled, check if both src and dst
1169		 * can migrate large pages
1170		 */
1171		if (thp_migration_supported()) {
1172			if ((src_pfns[i] & MIGRATE_PFN_MIGRATE) &&
1173				(src_pfns[i] & MIGRATE_PFN_COMPOUND) &&
1174				!(dst_pfns[i] & MIGRATE_PFN_COMPOUND)) {
1175
1176				if (!migrate) {
1177					src_pfns[i] &= ~(MIGRATE_PFN_MIGRATE |
1178							 MIGRATE_PFN_COMPOUND);
1179					goto next;
1180				}
1181				nr = 1 << folio_order(folio);
1182				addr = migrate->start + i * PAGE_SIZE;
1183				if (migrate_vma_split_unmapped_folio(migrate, i, addr, folio)) {
1184					src_pfns[i] &= ~(MIGRATE_PFN_MIGRATE |
1185							 MIGRATE_PFN_COMPOUND);
1186					goto next;
1187				}
1188			} else if ((src_pfns[i] & MIGRATE_PFN_MIGRATE) &&
1189				(dst_pfns[i] & MIGRATE_PFN_COMPOUND) &&
1190				!(src_pfns[i] & MIGRATE_PFN_COMPOUND)) {
1191				src_pfns[i] &= ~MIGRATE_PFN_MIGRATE;
1192			}
1193		}
1194
1195
1196		if (folio_is_device_private(newfolio) ||
1197		    folio_is_device_coherent(newfolio)) {
1198			if (mapping) {
1199				/*
1200				 * For now only support anonymous memory migrating to
1201				 * device private or coherent memory.
1202				 *
1203				 * Try to get rid of swap cache if possible.
1204				 */
1205				if (!folio_test_anon(folio) ||
1206				    !folio_free_swap(folio)) {
1207					src_pfns[i] &= ~MIGRATE_PFN_MIGRATE;
1208					goto next;
1209				}
1210			}
1211		} else if (folio_is_zone_device(newfolio)) {
1212			/*
1213			 * Other types of ZONE_DEVICE page are not supported.
1214			 */
1215			src_pfns[i] &= ~MIGRATE_PFN_MIGRATE;
1216			goto next;
1217		}
1218
1219		BUG_ON(folio_test_writeback(folio));
1220
1221		if (migrate && migrate->fault_page == page)
1222			extra_cnt = 1;
1223		for (j = 0; j < nr && i + j < npages; j++) {
1224			folio = page_folio(migrate_pfn_to_page(src_pfns[i+j]));
1225			newfolio = page_folio(migrate_pfn_to_page(dst_pfns[i+j]));
1226
1227			r = folio_migrate_mapping(mapping, newfolio, folio, extra_cnt);
1228			if (r)
1229				src_pfns[i+j] &= ~MIGRATE_PFN_MIGRATE;
1230			else
1231				folio_migrate_flags(newfolio, folio);
1232		}
1233next:
1234		i += nr;
1235	}
1236
1237	if (notified)
1238		mmu_notifier_invalidate_range_end(&range);
1239}
1240
1241/**
1242 * migrate_device_pages() - migrate meta-data from src page to dst page
1243 * @src_pfns: src_pfns returned from migrate_device_range()
1244 * @dst_pfns: array of pfns allocated by the driver to migrate memory to
1245 * @npages: number of pages in the range
1246 *
1247 * Equivalent to migrate_vma_pages(). This is called to migrate struct page
1248 * meta-data from source struct page to destination.
1249 */
1250void migrate_device_pages(unsigned long *src_pfns, unsigned long *dst_pfns,
1251			unsigned long npages)
1252{
1253	__migrate_device_pages(src_pfns, dst_pfns, npages, NULL);
1254}
1255EXPORT_SYMBOL(migrate_device_pages);
1256
1257/**
1258 * migrate_vma_pages() - migrate meta-data from src page to dst page
1259 * @migrate: migrate struct containing all migration information
1260 *
1261 * This migrates struct page meta-data from source struct page to destination
1262 * struct page. This effectively finishes the migration from source page to the
1263 * destination page.
1264 */
1265void migrate_vma_pages(struct migrate_vma *migrate)
1266{
1267	__migrate_device_pages(migrate->src, migrate->dst, migrate->npages, migrate);
1268}
1269EXPORT_SYMBOL(migrate_vma_pages);
1270
1271static void __migrate_device_finalize(unsigned long *src_pfns,
1272				      unsigned long *dst_pfns,
1273				      unsigned long npages,
1274				      struct page *fault_page)
1275{
1276	struct folio *fault_folio = fault_page ?
1277		page_folio(fault_page) : NULL;
1278	unsigned long i;
1279
1280	for (i = 0; i < npages; i++) {
1281		struct folio *dst = NULL, *src = NULL;
1282		struct page *newpage = migrate_pfn_to_page(dst_pfns[i]);
1283		struct page *page = migrate_pfn_to_page(src_pfns[i]);
1284
1285		if (newpage)
1286			dst = page_folio(newpage);
1287
1288		if (!page) {
1289			if (dst) {
1290				WARN_ON_ONCE(fault_folio == dst);
1291				folio_unlock(dst);
1292				folio_put(dst);
1293			}
1294			continue;
1295		}
1296
1297		src = page_folio(page);
1298
1299		if (!(src_pfns[i] & MIGRATE_PFN_MIGRATE) || !dst) {
1300			if (dst) {
1301				WARN_ON_ONCE(fault_folio == dst);
1302				folio_unlock(dst);
1303				folio_put(dst);
1304			}
1305			dst = src;
1306		}
1307
1308		if (!folio_is_zone_device(dst))
1309			folio_add_lru(dst);
1310		remove_migration_ptes(src, dst, 0);
1311		if (fault_folio != src)
1312			folio_unlock(src);
1313		folio_put(src);
1314
1315		if (dst != src) {
1316			WARN_ON_ONCE(fault_folio == dst);
1317			folio_unlock(dst);
1318			folio_put(dst);
1319		}
1320	}
1321}
1322
1323/*
1324 * migrate_device_finalize() - complete page migration
1325 * @src_pfns: src_pfns returned from migrate_device_range()
1326 * @dst_pfns: array of pfns allocated by the driver to migrate memory to
1327 * @npages: number of pages in the range
1328 *
1329 * Completes migration of the page by removing special migration entries.
1330 * Drivers must ensure copying of page data is complete and visible to the CPU
1331 * before calling this.
1332 */
1333void migrate_device_finalize(unsigned long *src_pfns,
1334			     unsigned long *dst_pfns, unsigned long npages)
1335{
1336	return __migrate_device_finalize(src_pfns, dst_pfns, npages, NULL);
1337}
1338EXPORT_SYMBOL(migrate_device_finalize);
1339
1340/**
1341 * migrate_vma_finalize() - restore CPU page table entry
1342 * @migrate: migrate struct containing all migration information
1343 *
1344 * This replaces the special migration pte entry with either a mapping to the
1345 * new page if migration was successful for that page, or to the original page
1346 * otherwise.
1347 *
1348 * This also unlocks the pages and puts them back on the lru, or drops the extra
1349 * refcount, for device pages.
1350 */
1351void migrate_vma_finalize(struct migrate_vma *migrate)
1352{
1353	__migrate_device_finalize(migrate->src, migrate->dst, migrate->npages,
1354				  migrate->fault_page);
1355}
1356EXPORT_SYMBOL(migrate_vma_finalize);
1357
1358static unsigned long migrate_device_pfn_lock(unsigned long pfn)
1359{
1360	struct folio *folio;
1361
1362	folio = folio_get_nontail_page(pfn_to_page(pfn));
1363	if (!folio)
1364		return 0;
1365
1366	if (!folio_trylock(folio)) {
1367		folio_put(folio);
1368		return 0;
1369	}
1370
1371	return migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
1372}
1373
1374/**
1375 * migrate_device_range() - migrate device private pfns to normal memory.
1376 * @src_pfns: array large enough to hold migrating source device private pfns.
1377 * @start: starting pfn in the range to migrate.
1378 * @npages: number of pages to migrate.
1379 *
1380 * migrate_vma_setup() is similar in concept to migrate_vma_setup() except that
1381 * instead of looking up pages based on virtual address mappings a range of
1382 * device pfns that should be migrated to system memory is used instead.
1383 *
1384 * This is useful when a driver needs to free device memory but doesn't know the
1385 * virtual mappings of every page that may be in device memory. For example this
1386 * is often the case when a driver is being unloaded or unbound from a device.
1387 *
1388 * Like migrate_vma_setup() this function will take a reference and lock any
1389 * migrating pages that aren't free before unmapping them. Drivers may then
1390 * allocate destination pages and start copying data from the device to CPU
1391 * memory before calling migrate_device_pages().
1392 */
1393int migrate_device_range(unsigned long *src_pfns, unsigned long start,
1394			unsigned long npages)
1395{
1396	unsigned long i, j, pfn;
1397
1398	for (pfn = start, i = 0; i < npages; pfn++, i++) {
1399		struct page *page = pfn_to_page(pfn);
1400		struct folio *folio = page_folio(page);
1401		unsigned int nr = 1;
1402
1403		src_pfns[i] = migrate_device_pfn_lock(pfn);
1404		nr = folio_nr_pages(folio);
1405		if (nr > 1) {
1406			src_pfns[i] |= MIGRATE_PFN_COMPOUND;
1407			for (j = 1; j < nr; j++)
1408				src_pfns[i+j] = 0;
1409			i += j - 1;
1410			pfn += j - 1;
1411		}
1412	}
1413
1414	migrate_device_unmap(src_pfns, npages, NULL);
1415
1416	return 0;
1417}
1418EXPORT_SYMBOL(migrate_device_range);
1419
1420/**
1421 * migrate_device_pfns() - migrate device private pfns to normal memory.
1422 * @src_pfns: pre-popluated array of source device private pfns to migrate.
1423 * @npages: number of pages to migrate.
1424 *
1425 * Similar to migrate_device_range() but supports non-contiguous pre-popluated
1426 * array of device pages to migrate.
1427 */
1428int migrate_device_pfns(unsigned long *src_pfns, unsigned long npages)
1429{
1430	unsigned long i, j;
1431
1432	for (i = 0; i < npages; i++) {
1433		struct page *page = pfn_to_page(src_pfns[i]);
1434		struct folio *folio = page_folio(page);
1435		unsigned int nr = 1;
1436
1437		src_pfns[i] = migrate_device_pfn_lock(src_pfns[i]);
1438		nr = folio_nr_pages(folio);
1439		if (nr > 1) {
1440			src_pfns[i] |= MIGRATE_PFN_COMPOUND;
1441			for (j = 1; j < nr; j++)
1442				src_pfns[i+j] = 0;
1443			i += j - 1;
1444		}
1445	}
1446
1447	migrate_device_unmap(src_pfns, npages, NULL);
1448
1449	return 0;
1450}
1451EXPORT_SYMBOL(migrate_device_pfns);
1452
1453/*
1454 * Migrate a device coherent folio back to normal memory. The caller should have
1455 * a reference on folio which will be copied to the new folio if migration is
1456 * successful or dropped on failure.
1457 */
1458int migrate_device_coherent_folio(struct folio *folio)
1459{
1460	unsigned long src_pfn, dst_pfn = 0;
1461	struct folio *dfolio;
1462
1463	WARN_ON_ONCE(folio_test_large(folio));
1464
1465	folio_lock(folio);
1466	src_pfn = migrate_pfn(folio_pfn(folio)) | MIGRATE_PFN_MIGRATE;
1467
1468	/*
1469	 * We don't have a VMA and don't need to walk the page tables to find
1470	 * the source folio. So call migrate_vma_unmap() directly to unmap the
1471	 * folio as migrate_vma_setup() will fail if args.vma == NULL.
1472	 */
1473	migrate_device_unmap(&src_pfn, 1, NULL);
1474	if (!(src_pfn & MIGRATE_PFN_MIGRATE))
1475		return -EBUSY;
1476
1477	dfolio = folio_alloc(GFP_USER | __GFP_NOWARN, 0);
1478	if (dfolio) {
1479		folio_lock(dfolio);
1480		dst_pfn = migrate_pfn(folio_pfn(dfolio));
1481	}
1482
1483	migrate_device_pages(&src_pfn, &dst_pfn, 1);
1484	if (src_pfn & MIGRATE_PFN_MIGRATE)
1485		folio_copy(dfolio, folio);
1486	migrate_device_finalize(&src_pfn, &dst_pfn, 1);
1487
1488	if (src_pfn & MIGRATE_PFN_MIGRATE)
1489		return 0;
1490	return -EBUSY;
1491}