Re: [PATCH] mm/khugepaged: don't map folios that are not uptodate in MADV_COLLAPSE

[David Hildenbrand (Arm)]

On 11/6/26 14:41, Kiryl Shutsemau wrote:
> On Fri, Nov 06, 2026 at 01:29:31PM +0100, David Hildenbrand (Arm) wrote:
>>> Separately from this fix: nothing on the mapping side catches this.
>>> Every caller checks uptodate on its own (next_uptodate_folio(),
>>> filemap_fault(), shmem_get_folio()), and khugepaged was the one that
>>> went straight from filemap_lock_folio() to do_set_pmd().
>>>
>>> Should do_set_pmd() and set_pte_range() assert it? Something like
>>>
>>>         VM_WARN_ON_ONCE_FOLIO(folio->mapping && !folio_test_anon(folio) &&
>>>                               !folio_test_uptodate(folio), folio);
>>
>> I think so. But looking at these hard-to-trigger but in the end very nasty
>> problems, should we actually reject the operation and do a pr_warn_once() or
>> something?
> 
> Yes, rejecting sounds better.
> 
> Not in set_pte_range() as I suggested, though. Fault-around only gets
> there with folios that next_uptodate_folio() checked under the folio
> lock, and for CoW it maps the new anon folio after do_cow_fault() has
> already copied from the page cache one.
> 
> Checking the folio ->fault() returned in finish_fault() covers PTE, PMD
> and CoW; do_set_pmd() covers khugepaged. Both return VM_FAULT_SIGBUS.
> 
> Something like this, only compile-tested:
> 
> diff --git a/mm/memory.c b/mm/memory.c
> index 6fe0b7dab927..ff2f2deff761 100644
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -5674,6 +5674,22 @@ static vm_fault_t __do_fault(struct vm_fault *vmf)
>         return ret;
>  }
> 
> +/*
> + * A page cache folio that is not uptodate holds data that was never read in
> + * or cleared, and mapping it leaks stale memory to userspace. The callers are
> + * expected to check, so treat a hit as a kernel bug and refuse the mapping.
> + */
> +static bool file_folio_not_uptodate(struct folio *folio)
> +{
> +       if (likely(folio_test_uptodate(folio)))
> +               return false;
> +       /* Driver pages from ->fault() never get PG_uptodate */
> +       if (folio_test_anon(folio) || !folio->mapping)
> +               return false;
> +       WARN_ONCE(1, "refusing to map a page cache folio that is not uptodate\n");
> +       return true;
> +}
> +
>  #ifdef CONFIG_TRANSPARENT_HUGEPAGE
>  static void deposit_prealloc_pte(struct vm_fault *vmf)
>  {
> @@ -5714,6 +5730,9 @@ vm_fault_t do_set_pmd(struct vm_fault *vmf, struct folio *folio, struct page *pa
>                 return ret;
>         page = &folio->page;
> 
> +       if (file_folio_not_uptodate(folio))
> +               return VM_FAULT_SIGBUS;
> +
>         /*
>          * Just backoff if any subpage of a THP is corrupted otherwise
>          * the corrupted page may mapped by PMD silently to escape the
> @@ -5849,6 +5868,16 @@ vm_fault_t finish_fault(struct vm_fault *vmf)
>         unsigned long addr;
>         bool needs_fallback = false;
> 
> +       /*
> +        * Check the folio ->fault() returned. For CoW that is the source of the
> +        * copy; DAX CoW has none.
> +        */
> +       if (vmf->page) {
> +               folio = page_folio(vmf->page);
> +               if (file_folio_not_uptodate(folio))
> +                       return VM_FAULT_SIGBUS;
> +       }
> +
>  fallback:
>         addr = vmf->address;

Can you send that as a proper patch? I think this is the right thing to do to
catch more of these things if they ever pop up in a better way.

-- 
Cheers,

David