When memory_failure() cannot split a large file folio, the folio stays in the page cache with the poisoned page in it. do_set_pmd() refuses to map such a folio with a PMD, so that the poisoned page gets its own fault and SIGBUS. finish_fault() then falls back to PTEs, but maps the whole folio: do_shared_fault() __do_fault() PageHWPoison(vmf->page) // the faulting page only finish_fault() do_set_pmd() // backs off on has_hwpoisoned set_pte_range(vmf, folio, &folio->page, folio_nr_pages(folio), ...) // maps the poisoned page as well A process that touches the poisoned page takes a machine check instead of SIGBUS from the fault. Map a folio with a poisoned page one page at a time. The poisoned page then takes its own fault, which __do_fault() turns into SIGBUS. Fixes: 43e027e41423 ("mm: memory: extend finish_fault() to support large folio") Cc: stable@vger.kernel.org Signed-off-by: Kiryl Shutsemau (Meta) --- Found while reviewing Usama's fix for the swap-in side: https://patch.msgid.link/20261116063624.1382944-3-usama.arif@linux.dev Reproduced on a huge=always tmpfs, with a THP that vmsplice() keeps memory_failure() from splitting. A write fault on a good page of the folio maps all 512 PTEs, the poisoned page included. With the patch it maps the faulting page only. mm/memory.c | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/mm/memory.c b/mm/memory.c index c966b3fe48b1..1f899d1be407 100644 --- a/mm/memory.c +++ b/mm/memory.c @@ -5920,6 +5920,10 @@ vm_fault_t finish_fault(struct vm_fault *vmf) return ret; } + /* A poisoned page must take its own fault */ + if (folio_contain_hwpoisoned_page(folio)) + needs_fallback = true; + if (!needs_fallback && vma->vm_file) { struct address_space *mapping = vma->vm_file->f_mapping; pgoff_t file_end; base-commit: 5e600d79135124a4eb8e9db30ad524224c267c07 -- 2.54.0