Re: [PATCH] mm/hmm: reject shadow-stack mappings

From: David Hildenbrand (Arm)

Date: Thu Nov 19 2026 - 17:19:55 EST


On 11/19/26 21:40, Edgecombe, Rick P wrote:
> On Thu, 2026-11-19 at 12:05 -0800, Dave Hansen wrote:
>> On 11/14/26 17:01, Andrew Morton wrote:
>>> On x86, a shadow-stack PTE has its ordinary write bit clear.  The CPU
>>> rejects ordinary stores, but pte_write() reports it writable because
>>> shadow-stack instructions can write it.  HMM consequently sets
>>> HMM_PFN_WRITE.  Nouveau uses that flag for GPU write permission, and
>>> RDMA on-demand paging can use it to grant device write permission to a
>>> registered memory region.  A device, or an RDMA peer with write access
>>> to that region, can then overwrite the process's shadow stack,
>>> potentially corrupting return addresses or bypassing that protection.
>>>
>>> Reject shadow-stack VMAs in hmm_vma_walk_test(), including read access.
>>> A requested fault returns -EFAULT; a snapshot records HMM_PFN_ERROR.
>>
>> This is a fun one.
>>
>> I almost wonder if we want to go one step further and disallow GUPs with
>> FOLL_WRITE on these PTEs. There should be precious few reasons for
>> devices to be writing to shadow stack pages.
>>
>> Isn't this a bigger issue than HMM? Should the fix be in HMM?
>
> We do reject GUPs to shadow stack PTEs with FOLL_WRITE, unless FOLL_FORCE is
> passed. The tricky thing as mentioned my others in this thread is that shadow
> stack memory *is* writable by userspace. But also the kernel should not treat it
> as such
>
> A false negative for pte_write() is dangerous when the kernel needs to decide
> when to revoke access. (i.e. fork) I think a mistake there is somewhat worse
> then missing an exotic way to cajole the kernel into writing to a user's shadow
> stack like this. So I think pte_write() being true for shadow stacks is the
> correct implementation.
>
> But, maybe the problem is that pte_write() is doing the job of two separate
> concepts. What if there was a pte_userspace_could_write() and a
> pte_kernel_should_not_write()? Names would need some work...

Gah, nak,

pte_write() just says "something might be able to write to it".

If that's actually true depends on other factors. Like the access+dirty on HW
that manages that in software.

I repeat: *no code* in Linux besides core-mm should ever interpret "pte_write()"
as "cool, I can write to it". The exception is GUP and unfortunately HMM.

--
Cheers,

David