1 // SPDX-License-Identifier: GPL-2.0
3 * Device Memory Migration functionality.
5 * Originally written by Jérôme Glisse.
7 #include <linux/export.h>
8 #include <linux/memremap.h>
9 #include <linux/migrate.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/swapops.h>
17 #include <asm/tlbflush.h>
20 static int migrate_vma_collect_skip(unsigned long start,
24 struct migrate_vma *migrate = walk->private;
27 for (addr = start; addr < end; addr += PAGE_SIZE) {
28 migrate->dst[migrate->npages] = 0;
29 migrate->src[migrate->npages++] = 0;
35 static int migrate_vma_collect_hole(unsigned long start,
37 __always_unused int depth,
40 struct migrate_vma *migrate = walk->private;
43 /* Only allow populating anonymous memory. */
44 if (!vma_is_anonymous(walk->vma))
45 return migrate_vma_collect_skip(start, end, walk);
47 for (addr = start; addr < end; addr += PAGE_SIZE) {
48 migrate->src[migrate->npages] = MIGRATE_PFN_MIGRATE;
49 migrate->dst[migrate->npages] = 0;
57 static int migrate_vma_collect_pmd(pmd_t *pmdp,
62 struct migrate_vma *migrate = walk->private;
63 struct vm_area_struct *vma = walk->vma;
64 struct mm_struct *mm = vma->vm_mm;
65 unsigned long addr = start, unmapped = 0;
71 return migrate_vma_collect_hole(start, end, -1, walk);
73 if (pmd_trans_huge(*pmdp)) {
76 ptl = pmd_lock(mm, pmdp);
77 if (unlikely(!pmd_trans_huge(*pmdp))) {
82 page = pmd_page(*pmdp);
83 if (is_huge_zero_page(page)) {
85 split_huge_pmd(vma, pmdp, addr);
86 if (pmd_trans_unstable(pmdp))
87 return migrate_vma_collect_skip(start, end,
94 if (unlikely(!trylock_page(page)))
95 return migrate_vma_collect_skip(start, end,
97 ret = split_huge_page(page);
101 return migrate_vma_collect_skip(start, end,
104 return migrate_vma_collect_hole(start, end, -1,
109 if (unlikely(pmd_bad(*pmdp)))
110 return migrate_vma_collect_skip(start, end, walk);
112 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
113 arch_enter_lazy_mmu_mode();
115 for (; addr < end; addr += PAGE_SIZE, ptep++) {
116 unsigned long mpfn = 0, pfn;
124 if (vma_is_anonymous(vma)) {
125 mpfn = MIGRATE_PFN_MIGRATE;
131 if (!pte_present(pte)) {
133 * Only care about unaddressable device page special
134 * page table entry. Other special swap entries are not
135 * migratable, and we ignore regular swapped page.
137 entry = pte_to_swp_entry(pte);
138 if (!is_device_private_entry(entry))
141 page = pfn_swap_entry_to_page(entry);
142 if (!(migrate->flags &
143 MIGRATE_VMA_SELECT_DEVICE_PRIVATE) ||
144 page->pgmap->owner != migrate->pgmap_owner)
147 mpfn = migrate_pfn(page_to_pfn(page)) |
149 if (is_writable_device_private_entry(entry))
150 mpfn |= MIGRATE_PFN_WRITE;
153 if (is_zero_pfn(pfn) &&
154 (migrate->flags & MIGRATE_VMA_SELECT_SYSTEM)) {
155 mpfn = MIGRATE_PFN_MIGRATE;
159 page = vm_normal_page(migrate->vma, addr, pte);
160 if (page && !is_zone_device_page(page) &&
161 !(migrate->flags & MIGRATE_VMA_SELECT_SYSTEM))
163 else if (page && is_device_coherent_page(page) &&
164 (!(migrate->flags & MIGRATE_VMA_SELECT_DEVICE_COHERENT) ||
165 page->pgmap->owner != migrate->pgmap_owner))
167 mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
168 mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
171 /* FIXME support THP */
172 if (!page || !page->mapping || PageTransCompound(page)) {
178 * By getting a reference on the page we pin it and that blocks
179 * any kind of migration. Side effect is that it "freezes" the
182 * We drop this reference after isolating the page from the lru
183 * for non device page (device page are not on the lru and thus
184 * can't be dropped from it).
189 * We rely on trylock_page() to avoid deadlock between
190 * concurrent migrations where each is waiting on the others
191 * page lock. If we can't immediately lock the page we fail this
192 * migration as it is only best effort anyway.
194 * If we can lock the page it's safe to set up a migration entry
195 * now. In the common case where the page is mapped once in a
196 * single process setting up the migration entry now is an
197 * optimisation to avoid walking the rmap later with
200 if (trylock_page(page)) {
204 flush_cache_page(vma, addr, pte_pfn(*ptep));
205 anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
206 if (anon_exclusive) {
207 pte = ptep_clear_flush(vma, addr, ptep);
209 if (page_try_share_anon_rmap(page)) {
210 set_pte_at(mm, addr, ptep, pte);
217 pte = ptep_get_and_clear(mm, addr, ptep);
222 /* Set the dirty flag on the folio now the pte is gone. */
224 folio_mark_dirty(page_folio(page));
226 /* Setup special migration page table entry */
227 if (mpfn & MIGRATE_PFN_WRITE)
228 entry = make_writable_migration_entry(
230 else if (anon_exclusive)
231 entry = make_readable_exclusive_migration_entry(
234 entry = make_readable_migration_entry(
236 if (pte_present(pte)) {
238 entry = make_migration_entry_young(entry);
240 entry = make_migration_entry_dirty(entry);
242 swp_pte = swp_entry_to_pte(entry);
243 if (pte_present(pte)) {
244 if (pte_soft_dirty(pte))
245 swp_pte = pte_swp_mksoft_dirty(swp_pte);
246 if (pte_uffd_wp(pte))
247 swp_pte = pte_swp_mkuffd_wp(swp_pte);
249 if (pte_swp_soft_dirty(pte))
250 swp_pte = pte_swp_mksoft_dirty(swp_pte);
251 if (pte_swp_uffd_wp(pte))
252 swp_pte = pte_swp_mkuffd_wp(swp_pte);
254 set_pte_at(mm, addr, ptep, swp_pte);
257 * This is like regular unmap: we remove the rmap and
258 * drop page refcount. Page won't be freed, as we took
259 * a reference just above.
261 page_remove_rmap(page, vma, false);
264 if (pte_present(pte))
272 migrate->dst[migrate->npages] = 0;
273 migrate->src[migrate->npages++] = mpfn;
276 /* Only flush the TLB if we actually modified any entries */
278 flush_tlb_range(walk->vma, start, end);
280 arch_leave_lazy_mmu_mode();
281 pte_unmap_unlock(ptep - 1, ptl);
286 static const struct mm_walk_ops migrate_vma_walk_ops = {
287 .pmd_entry = migrate_vma_collect_pmd,
288 .pte_hole = migrate_vma_collect_hole,
289 .walk_lock = PGWALK_RDLOCK,
293 * migrate_vma_collect() - collect pages over a range of virtual addresses
294 * @migrate: migrate struct containing all migration information
296 * This will walk the CPU page table. For each virtual address backed by a
297 * valid page, it updates the src array and takes a reference on the page, in
298 * order to pin the page until we lock it and unmap it.
300 static void migrate_vma_collect(struct migrate_vma *migrate)
302 struct mmu_notifier_range range;
305 * Note that the pgmap_owner is passed to the mmu notifier callback so
306 * that the registered device driver can skip invalidating device
307 * private page mappings that won't be migrated.
309 mmu_notifier_range_init_owner(&range, MMU_NOTIFY_MIGRATE, 0,
310 migrate->vma->vm_mm, migrate->start, migrate->end,
311 migrate->pgmap_owner);
312 mmu_notifier_invalidate_range_start(&range);
314 walk_page_range(migrate->vma->vm_mm, migrate->start, migrate->end,
315 &migrate_vma_walk_ops, migrate);
317 mmu_notifier_invalidate_range_end(&range);
318 migrate->end = migrate->start + (migrate->npages << PAGE_SHIFT);
322 * migrate_vma_check_page() - check if page is pinned or not
323 * @page: struct page to check
325 * Pinned pages cannot be migrated. This is the same test as in
326 * folio_migrate_mapping(), except that here we allow migration of a
329 static bool migrate_vma_check_page(struct page *page, struct page *fault_page)
332 * One extra ref because caller holds an extra reference, either from
333 * isolate_lru_page() for a regular page, or migrate_vma_collect() for
336 int extra = 1 + (page == fault_page);
339 * FIXME support THP (transparent huge page), it is bit more complex to
340 * check them than regular pages, because they can be mapped with a pmd
341 * or with a pte (split pte mapping).
343 if (PageCompound(page))
346 /* Page from ZONE_DEVICE have one extra reference */
347 if (is_zone_device_page(page))
350 /* For file back page */
351 if (page_mapping(page))
352 extra += 1 + page_has_private(page);
354 if ((page_count(page) - extra) > page_mapcount(page))
361 * Unmaps pages for migration. Returns number of source pfns marked as
364 static unsigned long migrate_device_unmap(unsigned long *src_pfns,
365 unsigned long npages,
366 struct page *fault_page)
368 unsigned long i, restore = 0;
369 bool allow_drain = true;
370 unsigned long unmapped = 0;
374 for (i = 0; i < npages; i++) {
375 struct page *page = migrate_pfn_to_page(src_pfns[i]);
379 if (src_pfns[i] & MIGRATE_PFN_MIGRATE)
384 /* ZONE_DEVICE pages are not on LRU */
385 if (!is_zone_device_page(page)) {
386 if (!PageLRU(page) && allow_drain) {
387 /* Drain CPU's pagevec */
392 if (!isolate_lru_page(page)) {
393 src_pfns[i] &= ~MIGRATE_PFN_MIGRATE;
398 /* Drop the reference we took in collect */
402 folio = page_folio(page);
403 if (folio_mapped(folio))
404 try_to_migrate(folio, 0);
406 if (page_mapped(page) ||
407 !migrate_vma_check_page(page, fault_page)) {
408 if (!is_zone_device_page(page)) {
410 putback_lru_page(page);
413 src_pfns[i] &= ~MIGRATE_PFN_MIGRATE;
421 for (i = 0; i < npages && restore; i++) {
422 struct page *page = migrate_pfn_to_page(src_pfns[i]);
425 if (!page || (src_pfns[i] & MIGRATE_PFN_MIGRATE))
428 folio = page_folio(page);
429 remove_migration_ptes(folio, folio, false);
441 * migrate_vma_unmap() - replace page mapping with special migration pte entry
442 * @migrate: migrate struct containing all migration information
444 * Isolate pages from the LRU and replace mappings (CPU page table pte) with a
445 * special migration pte entry and check if it has been pinned. Pinned pages are
446 * restored because we cannot migrate them.
448 * This is the last step before we call the device driver callback to allocate
449 * destination memory and copy contents of original page over to new page.
451 static void migrate_vma_unmap(struct migrate_vma *migrate)
453 migrate->cpages = migrate_device_unmap(migrate->src, migrate->npages,
454 migrate->fault_page);
458 * migrate_vma_setup() - prepare to migrate a range of memory
459 * @args: contains the vma, start, and pfns arrays for the migration
461 * Returns: negative errno on failures, 0 when 0 or more pages were migrated
464 * Prepare to migrate a range of memory virtual address range by collecting all
465 * the pages backing each virtual address in the range, saving them inside the
466 * src array. Then lock those pages and unmap them. Once the pages are locked
467 * and unmapped, check whether each page is pinned or not. Pages that aren't
468 * pinned have the MIGRATE_PFN_MIGRATE flag set (by this function) in the
469 * corresponding src array entry. Then restores any pages that are pinned, by
470 * remapping and unlocking those pages.
472 * The caller should then allocate destination memory and copy source memory to
473 * it for all those entries (ie with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE
474 * flag set). Once these are allocated and copied, the caller must update each
475 * corresponding entry in the dst array with the pfn value of the destination
476 * page and with MIGRATE_PFN_VALID. Destination pages must be locked via
479 * Note that the caller does not have to migrate all the pages that are marked
480 * with MIGRATE_PFN_MIGRATE flag in src array unless this is a migration from
481 * device memory to system memory. If the caller cannot migrate a device page
482 * back to system memory, then it must return VM_FAULT_SIGBUS, which has severe
483 * consequences for the userspace process, so it must be avoided if at all
486 * For empty entries inside CPU page table (pte_none() or pmd_none() is true) we
487 * do set MIGRATE_PFN_MIGRATE flag inside the corresponding source array thus
488 * allowing the caller to allocate device memory for those unbacked virtual
489 * addresses. For this the caller simply has to allocate device memory and
490 * properly set the destination entry like for regular migration. Note that
491 * this can still fail, and thus inside the device driver you must check if the
492 * migration was successful for those entries after calling migrate_vma_pages(),
493 * just like for regular migration.
495 * After that, the callers must call migrate_vma_pages() to go over each entry
496 * in the src array that has the MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag
497 * set. If the corresponding entry in dst array has MIGRATE_PFN_VALID flag set,
498 * then migrate_vma_pages() to migrate struct page information from the source
499 * struct page to the destination struct page. If it fails to migrate the
500 * struct page information, then it clears the MIGRATE_PFN_MIGRATE flag in the
503 * At this point all successfully migrated pages have an entry in the src
504 * array with MIGRATE_PFN_VALID and MIGRATE_PFN_MIGRATE flag set and the dst
505 * array entry with MIGRATE_PFN_VALID flag set.
507 * Once migrate_vma_pages() returns the caller may inspect which pages were
508 * successfully migrated, and which were not. Successfully migrated pages will
509 * have the MIGRATE_PFN_MIGRATE flag set for their src array entry.
511 * It is safe to update device page table after migrate_vma_pages() because
512 * both destination and source page are still locked, and the mmap_lock is held
513 * in read mode (hence no one can unmap the range being migrated).
515 * Once the caller is done cleaning up things and updating its page table (if it
516 * chose to do so, this is not an obligation) it finally calls
517 * migrate_vma_finalize() to update the CPU page table to point to new pages
518 * for successfully migrated pages or otherwise restore the CPU page table to
519 * point to the original source pages.
521 int migrate_vma_setup(struct migrate_vma *args)
523 long nr_pages = (args->end - args->start) >> PAGE_SHIFT;
525 args->start &= PAGE_MASK;
526 args->end &= PAGE_MASK;
527 if (!args->vma || is_vm_hugetlb_page(args->vma) ||
528 (args->vma->vm_flags & VM_SPECIAL) || vma_is_dax(args->vma))
532 if (args->start < args->vma->vm_start ||
533 args->start >= args->vma->vm_end)
535 if (args->end <= args->vma->vm_start || args->end > args->vma->vm_end)
537 if (!args->src || !args->dst)
539 if (args->fault_page && !is_device_private_page(args->fault_page))
542 memset(args->src, 0, sizeof(*args->src) * nr_pages);
546 migrate_vma_collect(args);
549 migrate_vma_unmap(args);
552 * At this point pages are locked and unmapped, and thus they have
553 * stable content and can safely be copied to destination memory that
554 * is allocated by the drivers.
559 EXPORT_SYMBOL(migrate_vma_setup);
562 * This code closely matches the code in:
563 * __handle_mm_fault()
565 * do_anonymous_page()
566 * to map in an anonymous zero page but the struct page will be a ZONE_DEVICE
567 * private or coherent page.
569 static void migrate_vma_insert_page(struct migrate_vma *migrate,
574 struct vm_area_struct *vma = migrate->vma;
575 struct mm_struct *mm = vma->vm_mm;
585 /* Only allow populating anonymous memory */
586 if (!vma_is_anonymous(vma))
589 pgdp = pgd_offset(mm, addr);
590 p4dp = p4d_alloc(mm, pgdp, addr);
593 pudp = pud_alloc(mm, p4dp, addr);
596 pmdp = pmd_alloc(mm, pudp, addr);
600 if (pmd_trans_huge(*pmdp) || pmd_devmap(*pmdp))
604 * Use pte_alloc() instead of pte_alloc_map(). We can't run
605 * pte_offset_map() on pmds where a huge pmd might be created
606 * from a different thread.
608 * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
609 * parallel threads are excluded by other means.
611 * Here we only have mmap_read_lock(mm).
613 if (pte_alloc(mm, pmdp))
616 /* See the comment in pte_alloc_one_map() */
617 if (unlikely(pmd_trans_unstable(pmdp)))
620 if (unlikely(anon_vma_prepare(vma)))
622 if (mem_cgroup_charge(page_folio(page), vma->vm_mm, GFP_KERNEL))
626 * The memory barrier inside __SetPageUptodate makes sure that
627 * preceding stores to the page contents become visible before
628 * the set_pte_at() write.
630 __SetPageUptodate(page);
632 if (is_device_private_page(page)) {
633 swp_entry_t swp_entry;
635 if (vma->vm_flags & VM_WRITE)
636 swp_entry = make_writable_device_private_entry(
639 swp_entry = make_readable_device_private_entry(
641 entry = swp_entry_to_pte(swp_entry);
643 if (is_zone_device_page(page) &&
644 !is_device_coherent_page(page)) {
645 pr_warn_once("Unsupported ZONE_DEVICE page type.\n");
648 entry = mk_pte(page, vma->vm_page_prot);
649 if (vma->vm_flags & VM_WRITE)
650 entry = pte_mkwrite(pte_mkdirty(entry));
653 ptep = pte_offset_map_lock(mm, pmdp, addr, &ptl);
655 if (check_stable_address_space(mm))
658 if (pte_present(*ptep)) {
659 unsigned long pfn = pte_pfn(*ptep);
661 if (!is_zero_pfn(pfn))
664 } else if (!pte_none(*ptep))
668 * Check for userfaultfd but do not deliver the fault. Instead,
671 if (userfaultfd_missing(vma))
674 inc_mm_counter(mm, MM_ANONPAGES);
675 page_add_new_anon_rmap(page, vma, addr);
676 if (!is_zone_device_page(page))
677 lru_cache_add_inactive_or_unevictable(page, vma);
681 flush_cache_page(vma, addr, pte_pfn(*ptep));
682 ptep_clear_flush_notify(vma, addr, ptep);
683 set_pte_at_notify(mm, addr, ptep, entry);
684 update_mmu_cache(vma, addr, ptep);
686 /* No need to invalidate - it was non-present before */
687 set_pte_at(mm, addr, ptep, entry);
688 update_mmu_cache(vma, addr, ptep);
691 pte_unmap_unlock(ptep, ptl);
692 *src = MIGRATE_PFN_MIGRATE;
696 pte_unmap_unlock(ptep, ptl);
698 *src &= ~MIGRATE_PFN_MIGRATE;
701 static void __migrate_device_pages(unsigned long *src_pfns,
702 unsigned long *dst_pfns, unsigned long npages,
703 struct migrate_vma *migrate)
705 struct mmu_notifier_range range;
707 bool notified = false;
709 for (i = 0; i < npages; i++) {
710 struct page *newpage = migrate_pfn_to_page(dst_pfns[i]);
711 struct page *page = migrate_pfn_to_page(src_pfns[i]);
712 struct address_space *mapping;
716 src_pfns[i] &= ~MIGRATE_PFN_MIGRATE;
723 if (!(src_pfns[i] & MIGRATE_PFN_MIGRATE))
727 * The only time there is no vma is when called from
728 * migrate_device_coherent_page(). However this isn't
729 * called if the page could not be unmapped.
732 addr = migrate->start + i*PAGE_SIZE;
736 mmu_notifier_range_init_owner(&range,
737 MMU_NOTIFY_MIGRATE, 0,
738 migrate->vma->vm_mm, addr, migrate->end,
739 migrate->pgmap_owner);
740 mmu_notifier_invalidate_range_start(&range);
742 migrate_vma_insert_page(migrate, addr, newpage,
747 mapping = page_mapping(page);
749 if (is_device_private_page(newpage) ||
750 is_device_coherent_page(newpage)) {
752 * For now only support anonymous memory migrating to
753 * device private or coherent memory.
756 src_pfns[i] &= ~MIGRATE_PFN_MIGRATE;
759 } else if (is_zone_device_page(newpage)) {
761 * Other types of ZONE_DEVICE page are not supported.
763 src_pfns[i] &= ~MIGRATE_PFN_MIGRATE;
767 if (migrate && migrate->fault_page == page)
768 r = migrate_folio_extra(mapping, page_folio(newpage),
770 MIGRATE_SYNC_NO_COPY, 1);
772 r = migrate_folio(mapping, page_folio(newpage),
773 page_folio(page), MIGRATE_SYNC_NO_COPY);
774 if (r != MIGRATEPAGE_SUCCESS)
775 src_pfns[i] &= ~MIGRATE_PFN_MIGRATE;
779 * No need to double call mmu_notifier->invalidate_range() callback as
780 * the above ptep_clear_flush_notify() inside migrate_vma_insert_page()
781 * did already call it.
784 mmu_notifier_invalidate_range_only_end(&range);
788 * migrate_device_pages() - migrate meta-data from src page to dst page
789 * @src_pfns: src_pfns returned from migrate_device_range()
790 * @dst_pfns: array of pfns allocated by the driver to migrate memory to
791 * @npages: number of pages in the range
793 * Equivalent to migrate_vma_pages(). This is called to migrate struct page
794 * meta-data from source struct page to destination.
796 void migrate_device_pages(unsigned long *src_pfns, unsigned long *dst_pfns,
797 unsigned long npages)
799 __migrate_device_pages(src_pfns, dst_pfns, npages, NULL);
801 EXPORT_SYMBOL(migrate_device_pages);
804 * migrate_vma_pages() - migrate meta-data from src page to dst page
805 * @migrate: migrate struct containing all migration information
807 * This migrates struct page meta-data from source struct page to destination
808 * struct page. This effectively finishes the migration from source page to the
811 void migrate_vma_pages(struct migrate_vma *migrate)
813 __migrate_device_pages(migrate->src, migrate->dst, migrate->npages, migrate);
815 EXPORT_SYMBOL(migrate_vma_pages);
818 * migrate_device_finalize() - complete page migration
819 * @src_pfns: src_pfns returned from migrate_device_range()
820 * @dst_pfns: array of pfns allocated by the driver to migrate memory to
821 * @npages: number of pages in the range
823 * Completes migration of the page by removing special migration entries.
824 * Drivers must ensure copying of page data is complete and visible to the CPU
825 * before calling this.
827 void migrate_device_finalize(unsigned long *src_pfns,
828 unsigned long *dst_pfns, unsigned long npages)
832 for (i = 0; i < npages; i++) {
833 struct folio *dst, *src;
834 struct page *newpage = migrate_pfn_to_page(dst_pfns[i]);
835 struct page *page = migrate_pfn_to_page(src_pfns[i]);
839 unlock_page(newpage);
845 if (!(src_pfns[i] & MIGRATE_PFN_MIGRATE) || !newpage) {
847 unlock_page(newpage);
853 src = page_folio(page);
854 dst = page_folio(newpage);
855 remove_migration_ptes(src, dst, false);
858 if (is_zone_device_page(page))
861 putback_lru_page(page);
863 if (newpage != page) {
864 unlock_page(newpage);
865 if (is_zone_device_page(newpage))
868 putback_lru_page(newpage);
872 EXPORT_SYMBOL(migrate_device_finalize);
875 * migrate_vma_finalize() - restore CPU page table entry
876 * @migrate: migrate struct containing all migration information
878 * This replaces the special migration pte entry with either a mapping to the
879 * new page if migration was successful for that page, or to the original page
882 * This also unlocks the pages and puts them back on the lru, or drops the extra
883 * refcount, for device pages.
885 void migrate_vma_finalize(struct migrate_vma *migrate)
887 migrate_device_finalize(migrate->src, migrate->dst, migrate->npages);
889 EXPORT_SYMBOL(migrate_vma_finalize);
892 * migrate_device_range() - migrate device private pfns to normal memory.
893 * @src_pfns: array large enough to hold migrating source device private pfns.
894 * @start: starting pfn in the range to migrate.
895 * @npages: number of pages to migrate.
897 * migrate_vma_setup() is similar in concept to migrate_vma_setup() except that
898 * instead of looking up pages based on virtual address mappings a range of
899 * device pfns that should be migrated to system memory is used instead.
901 * This is useful when a driver needs to free device memory but doesn't know the
902 * virtual mappings of every page that may be in device memory. For example this
903 * is often the case when a driver is being unloaded or unbound from a device.
905 * Like migrate_vma_setup() this function will take a reference and lock any
906 * migrating pages that aren't free before unmapping them. Drivers may then
907 * allocate destination pages and start copying data from the device to CPU
908 * memory before calling migrate_device_pages().
910 int migrate_device_range(unsigned long *src_pfns, unsigned long start,
911 unsigned long npages)
913 unsigned long i, pfn;
915 for (pfn = start, i = 0; i < npages; pfn++, i++) {
916 struct page *page = pfn_to_page(pfn);
918 if (!get_page_unless_zero(page)) {
923 if (!trylock_page(page)) {
929 src_pfns[i] = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
932 migrate_device_unmap(src_pfns, npages, NULL);
936 EXPORT_SYMBOL(migrate_device_range);
939 * Migrate a device coherent page back to normal memory. The caller should have
940 * a reference on page which will be copied to the new page if migration is
941 * successful or dropped on failure.
943 int migrate_device_coherent_page(struct page *page)
945 unsigned long src_pfn, dst_pfn = 0;
948 WARN_ON_ONCE(PageCompound(page));
951 src_pfn = migrate_pfn(page_to_pfn(page)) | MIGRATE_PFN_MIGRATE;
954 * We don't have a VMA and don't need to walk the page tables to find
955 * the source page. So call migrate_vma_unmap() directly to unmap the
956 * page as migrate_vma_setup() will fail if args.vma == NULL.
958 migrate_device_unmap(&src_pfn, 1, NULL);
959 if (!(src_pfn & MIGRATE_PFN_MIGRATE))
962 dpage = alloc_page(GFP_USER | __GFP_NOWARN);
965 dst_pfn = migrate_pfn(page_to_pfn(dpage));
968 migrate_device_pages(&src_pfn, &dst_pfn, 1);
969 if (src_pfn & MIGRATE_PFN_MIGRATE)
970 copy_highpage(dpage, page);
971 migrate_device_finalize(&src_pfn, &dst_pfn, 1);
973 if (src_pfn & MIGRATE_PFN_MIGRATE)