When a ZONE_DEVICE private page is freed, the page->mapping field can be set. If this page is reused as an anonymous page, the previous value can prevent the page from being inserted into the CPU's anon rmap table. For example, when migrating a pte_none() page to device memory: migrate_vma(ops, vma, start, end, src, dst, private) migrate_vma_collect() src[] = MIGRATE_PFN_MIGRATE migrate_vma_prepare() /* no page to lock or isolate so OK */ migrate_vma_unmap() /* no page to unmap so OK */ ops->alloc_and_copy() /* driver allocates ZONE_DEVICE page for dst[] */ migrate_vma_pages() migrate_vma_insert_page() page_add_new_anon_rmap() __page_set_anon_rmap() /* This check sees the page's stale mapping field */ if (PageAnon(page)) return /* page->mapping is not updated */ The result is that the migration appears to succeed but a subsequent CPU fault will be unable to migrate the page back to system memory or worse. Clear the page->mapping field when freeing the ZONE_DEVICE page so stale pointer data doesn't affect future page use. Fixes: b7a523109fb5c9d2d6dd ("mm: don't clear ->mapping in hmm_devmem_free") Cc: stable@xxxxxxxxxxxxxxx Signed-off-by: Ralph Campbell <rcampbell@xxxxxxxxxx> Cc: Christoph Hellwig <hch@xxxxxx> Cc: Dan Williams <dan.j.williams@xxxxxxxxx> Cc: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> Cc: Jason Gunthorpe <jgg@xxxxxxxxxxxx> Cc: Logan Gunthorpe <logang@xxxxxxxxxxxx> Cc: Ira Weiny <ira.weiny@xxxxxxxxx> Cc: Matthew Wilcox <willy@xxxxxxxxxxxxx> Cc: Mel Gorman <mgorman@xxxxxxxxxxxxxxxxxxx> Cc: Jan Kara <jack@xxxxxxx> Cc: "Kirill A. Shutemov" <kirill.shutemov@xxxxxxxxxxxxxxx> Cc: Michal Hocko <mhocko@xxxxxxxx> Cc: Andrea Arcangeli <aarcange@xxxxxxxxxx> Cc: Mike Kravetz <mike.kravetz@xxxxxxxxxx> Cc: "Jérôme Glisse" <jglisse@xxxxxxxxxx> --- kernel/memremap.c | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/kernel/memremap.c b/kernel/memremap.c index bea6f887adad..238ae5d0ae8a 100644 --- a/kernel/memremap.c +++ b/kernel/memremap.c @@ -408,6 +408,10 @@ void __put_devmap_managed_page(struct page *page) mem_cgroup_uncharge(page); + /* Clear anonymous page mapping to prevent stale pointers */ + if (is_device_private_page(page)) + page->mapping = NULL; + page->pgmap->ops->page_free(page); } else if (!count) __put_page(page); -- 2.20.1