Re: [PATCH 0/2] mm/memfd: add F_SEAL_PUNCH_HOLE

[Samiullah Khawaja]

On Fri, Oct 30, 2026 at 03:43:53PM -0700, Andy Lutomirski wrote:
> On Fri, Oct 30, 2026 at 10:40 AM Samiullah Khawaja <skhawaja@xxxxxxxxxx> wrote:
> > On Fri, Oct 30, 2026 at 03:39:40PM +0100, David Hildenbrand (Arm) wrote:
> > >On 10/30/26 13:11, Jason Gunthorpe wrote:
> > >> On Fri, Oct 30, 2026 at 11:20:30AM +0100, David Hildenbrand (Arm) wrote:
> > >>> On 10/26/26 23:05, Samiullah Khawaja wrote:
> > >>>> This series introduces F_SEAL_PUNCH_HOLE, a new memfd seal that prevents
> > >>>> holes from being punched into a shmem or hugetlbfs memfd while leaving
> > >>>> the file fully writable.
> > >>>>
> > >>>> Background
> > >>>> ==========
> > >>>>
> > >>>> Drivers such as udmabuf and iommufd (via IOMMU_IOAS_MAP_FILE) use
> > >>>> memfd_pin_folios() to pin the folios backing a memfd and hand them to a
> > >>>> device for DMA. Pinning keeps those folios allocated, but it does not
> > >>>> keep them in the file. If userspace later calls
> > >>>> fallocate(FALLOC_FL_PUNCH_HOLE) or madvise(MADV_REMOVE) on the range,
> > >>>> the folios are dropped from the page cache while the device continues
> > >>>> to use them. The pin holder is never notified, and from that point on
> > >>>> the device and the file silently diverge.
> > >>>
> > >>> We have the exact same situation when userspace maps anonymous folios into vfio
> > >>> and then desides to MADV_DONTNEED the pages. Or any other memory backend that is
> > >>> used for such purposes.
> > >>>
> > >>> User space must not zap/punch things that might be pinned elsewhere.
> > >>>
> > >>> That's not specific to memfd, udmabuf, iommufd or anything else.
> > >>>
> > >>> I don't see a reason to make userspace (that should not shoot itself into the
> > >>> foot) a kernel feature it must activate itself to not shoot itself into the foot.
> >
> > Agreed, for normal use this is userspace's problem. The seal was meant
> > for the kernel to cheaply verify at preserve time that nothing was
> > punched.
> >
> > >>
> > >> The issue is live update doesn't preserve those background punched
> > >> folios and now we have created an UAF attack on the kernel across
> > >> kexec if you don't do something about it.
> > >>
> > >> There was some notion here that with the right sealing flags we could
> > >> tell if everything was 'live update compatible' and then know that
> > >> these attacks are impossible.
> > >Just remember that there was any hole punch and refuse to live update this instance.
> > >
> > >We could likely implement "we saw hole punch on a pinned folio" as well, but I
> > >don't think that's required initially.
> >
> > If I understand correctly, you mean recording on the memfd that a hole
> > punch happened, and refusing the live update if it did? That would work
> > and is much cheaper than what we do today, where iommufd walks all the
> > mapped pages at preserve time. For truncate, the existing F_SEAL_SHRINK
> > is enough or I can record that also on the memfd. I will send patches
> > for this.
>
> This entire situation continues to give me the willies.  After much
> prodding and contemplating (in my own brain), I got a friendly LLM to
> report:
>
> --- begin AI output ---
>
> IOMMU_IOAS_MAP_FILE uses memfd_pin_folios(). Collapse [MADV_COLLAPSE
> or khugepaged] checks the old folio’s reference count, so a pin
> acquired before that check makes collapse fail.
>
>  The problematic interleaving is:
>
>  1. collapse_file() locks and unmaps the old folio, then passes its
> reference-count check (mm/khugepaged.c:2446).
>  2. memfd_pin_folios() (mm/gup.c:3483) obtains the old folio from the
> file’s XArray and pins it, without taking the folio lock.
>  3. Collapse copies the contents and replaces the old cache entries
> with the new huge folio. It does not repeat the reference check or
> freeze the old count.
>  4. IOAS retains the old physical pages; accesses through the file
> use the new folio.
>
> --- end AI output ---
>
> Is this correct?  I haven't personally verified it.  But this whole

Yes, I think this is correct. memfd_pin_folios() pins the folios
directly without taking the folio lock. Other page cache lookups lock
the folio and recheck folio->mapping, e.g. shmem_get_folio_gfp(). Only
memfd_pin_folios() misses it.

This scenario is different from punch hole or MADV_DONTNEED where the
user decides to remove the pinned memory. Here the kernel is removing
the pinned page without the user knowing. Collapse already checks the
refcount to avoid this.

The recent fix [1] by Chris already takes the folio lock for !uptodate
folios. I think we can take the lock unconditionally and recheck
folio->mapping, that should fix this. I can send a patch for this fix,
if folks agree.

[1] https://lore.kernel.org/all/20261029-fixes_core-v2-20-a64ce7f83abb@xxxxxxxxxx/

Thanks,
Sami