Commit 98c183a4fccf ("fs/dax: don't disassociate zero page entries") added zero/empty-entry early returns to dax_associate_entry() and dax_disassociate_entry(), but placed them *after* the `struct folio *folio = dax_to_folio(entry);` line. dax_to_folio() expands to page_folio(pfn_to_page(dax_to_pfn(entry))), which calls _compound_head() and performs READ_ONCE(page->compound_info) -- a real dereference of the struct page pointer derived from a bogus PFN extracted from the empty/zero XA value. On systems where vmemmap covers all of RAM that dereference reads garbage and is harmless: the early return then discards the result. On virtio-pmem with altmap (vmemmap stored inside the device), only the real device PFN range is mapped, so the dereference triggers a kernel paging fault from the truncate / invalidate path and from the PMD-downgrade branch of dax_iomap_pte_fault when an entry is being freed: Unable to handle kernel paging request at virtual address ffff_fdff_bf00_0008 (vmemmap region) Call trace: dax_disassociate_entry.isra.0+0x20/0x50 dax_iomap_pte_fault dax_iomap_fault erofs_dax_fault Close the residual gap by moving the dax_to_folio() call after the zero/empty guard in both dax_associate_entry() and dax_disassociate_entry(). Apply the same treatment to dax_busy_page(), which has the identical pattern but was not touched by the prior fix. dax_associate_entry() is reachable with a zero entry via dax_insert_entry() -> dax_associate_entry(new_entry, ...), where new_entry can carry DAX_ZERO_PAGE (built by dax_make_entry() in dax_load_hole() / dax_pmd_load_hole()). dax_disassociate_entry() and dax_busy_page() additionally see DAX_EMPTY entries created by grab_mapping_entry(). The remaining users of dax_to_folio() / dax_to_pfn() in fs/dax.c are either guarded or only reachable on real-PFN entries, so this exhausts the anti-pattern. Fixes: 98c183a4fccf ("fs/dax: don't disassociate zero page entries") Fixes: 38607c62b34b ("fs/dax: properly refcount fs dax pages") Cc: stable@vger.kernel.org # v6.15+ Cc: Alistair Popple Suggested-by: David Hildenbrand Signed-off-by: Souvik Banerjee --- Changes in v2: - Also fix dax_associate_entry() (Suggested-by: David Hildenbrand, confirmed by Alistair Popple). The same anti-pattern existed there: dax_to_folio(entry) ran before the zero/empty guard. new_entry on that path can carry DAX_ZERO_PAGE via dax_load_hole() / dax_pmd_load_hole(), so the dereference reads a struct page derived from the zero-page PFN before the early return discards it. - Audited remaining dax_to_folio() / dax_to_pfn() call sites in fs/dax.c; no further instances of the pattern. - Updated the page_folio() expansion in the commit message to refer to the current field name (page->compound_info via _compound_head()). v1: https://lore.kernel.org/all/20260501233933.2614302-1-souvik@amlalabs.com/ fs/dax.c | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-) diff --git a/fs/dax.c b/fs/dax.c index 6d175cd47a99..4bca6e2bc342 100644 --- a/fs/dax.c +++ b/fs/dax.c @@ -480,11 +480,12 @@ static void dax_associate_entry(void *entry, struct address_space *mapping, unsigned long address, bool shared) { unsigned long size = dax_entry_size(entry), index; - struct folio *folio = dax_to_folio(entry); + struct folio *folio; if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) return; + folio = dax_to_folio(entry); index = linear_page_index(vma, address & ~(size - 1)); if (shared && (folio->mapping || dax_folio_is_shared(folio))) { if (folio->mapping) @@ -505,21 +506,23 @@ static void dax_associate_entry(void *entry, struct address_space *mapping, static void dax_disassociate_entry(void *entry, struct address_space *mapping, bool trunc) { - struct folio *folio = dax_to_folio(entry); + struct folio *folio; if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) return; + folio = dax_to_folio(entry); dax_folio_put(folio); } static struct page *dax_busy_page(void *entry) { - struct folio *folio = dax_to_folio(entry); + struct folio *folio; if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) return NULL; + folio = dax_to_folio(entry); if (folio_ref_count(folio) - folio_mapcount(folio)) return &folio->page; else -- 2.51.1