Re: [PATCH] buildid: validate page-backed file before parsing build ID

[Andrii Nakryiko]

On Tue, Jan 6, 2026 at 11:16 AM David Hildenbrand (Red Hat)
<david@xxxxxxxxxx> wrote:
>
> On 1/5/26 23:52, Andrii Nakryiko wrote:
> > On Tue, Dec 30, 2025 at 2:11 PM David Hildenbrand (Red Hat)
> > <david@xxxxxxxxxx> wrote:
> >>
> >> On 12/23/25 18:29, Andrew Morton wrote:
> >>> On Tue, 23 Dec 2025 18:32:07 +0800 Jinchao Wang <wangjinchao600@xxxxxxxxx> wrote:
> >>>
> >>>> __build_id_parse() only works on page-backed storage.  Its helper paths
> >>>> eventually call mapping->a_ops->read_folio(), so explicitly reject VMAs
> >>>> that do not map a regular file or lack valid address_space operations.
> >>>>
> >>>> Reported-by: syzbot+e008db2ac01e282550ee@xxxxxxxxxxxxxxxxxxxxxxxxx
> >>>> Signed-off-by: Jinchao Wang <wangjinchao600@xxxxxxxxx>
> >>>>
> >>>> ...
> >>>>
> >>>> --- a/lib/buildid.c
> >>>> +++ b/lib/buildid.c
> >>>> @@ -280,7 +280,10 @@ static int __build_id_parse(struct vm_area_struct *vma, unsigned char *build_id,
> >>>>       int ret;
> >>>>
> >>>>       /* only works for page backed storage  */
> >>>> -    if (!vma->vm_file)
> >>>> +    if (!vma->vm_file ||
> >>>> +        !S_ISREG(file_inode(vma->vm_file)->i_mode) ||
> >>>> +        !vma->vm_file->f_mapping->a_ops ||
> >>>> +        !vma->vm_file->f_mapping->a_ops->read_folio)
> >>>>               return -EINVAL;
> >>
> >> Just wondering, we are fine with MAP_PRIVATE files, right? I guess it's
> >> not about the actual content in the VMA (which might be different for a
> >> MAP_PRIVATE VMA), but only about the content of the mapped file.
> >
> > Yep, this code is fetching contents of a file that backs given VMA.
>
> Good!
>
> >
> >>
> >>
> >> LGTM, although I wonder whether some of these these checks should be
> >> exposed as part of the read_cache_folio()/do_read_cache_folio() API.
> >>
> >> Like, having a helper function that tells us whether we can use
> >> do_read_cache_folio() against a given mapping+file.
> >
> > I agree, this seems to be leaking a lot of internal mm details into
> > higher-level caller (__build_id_parse). Right now we try to fetch
> > folio with filemap_get_folio() and if that succeeds, then we do
> > read_cache_folio. Would it be possible for filemap_get_folio() to
> > return error if the folio cannot be read using read_cache_folio()? Or
> > maybe have a variant of filemap_get_folio() that would have this
> > semantic?
>
> Good question. But really, for files that always have everything in the pagecache,
> there would not be a problem, right? I'm thinking about hugetlb, for example.
>
> There, we never expect to fallback to do_read_cache_folio().
>
> So maybe we could just teach do_read_cache_folio() to fail properly?
>
> diff --git a/mm/filemap.c b/mm/filemap.c
> index ebd75684cb0a7..3f81b8481af4c 100644
> --- a/mm/filemap.c
> +++ b/mm/filemap.c
> @@ -4051,8 +4051,11 @@ static struct folio *do_read_cache_folio(struct address_space *mapping,
>          struct folio *folio;
>          int err;
>
> -       if (!filler)
> +       if (!filler) {
> +               if (!mapping->a_ops || !mapping->a_ops->read_folio)
> +                       return ERR_PTR(-EINVAL);
>                  filler = mapping->a_ops->read_folio;
> +       }
>   repeat:
>          folio = filemap_get_folio(mapping, index);
>          if (IS_ERR(folio)) {
>
> Then __build_id_parse() would only check for the existence of vma->vm_file and maybe
> the !S_ISREG(file_inode(vma->vm_file)->i_mode).
>

That would be great. But something like this was proposed earlier and
Matthew didn't particularly like this approach ([0]).

  [0] https://lore.kernel.org/all/aReUv1kVACh3UKv-@xxxxxxxxxxxxxxxxxxxx/

>
> --
> Cheers
>
> David