Re: [PATCH RFC] mm: add MAP_EXCLUSIVE to create exclusive user mappings

[Andy Lutomirski <luto@kernel.org>]

On Wed, Oct 30, 2019 at 1:40 AM Mike Rapoport <rppt@kernel.org> wrote:
>
> On Tue, Oct 29, 2019 at 10:00:55AM -0700, Andy Lutomirski wrote:
> > On Tue, Oct 29, 2019 at 2:33 AM Mike Rapoport <rppt@kernel.org> wrote:
> > >
> > > On Mon, Oct 28, 2019 at 02:44:23PM -0600, Andy Lutomirski wrote:
> > > >
> > > > > On Oct 27, 2019, at 4:17 AM, Mike Rapoport <rppt@kernel.org> wrote:
> > > > >
> > > > > From: Mike Rapoport <rppt@linux.ibm.com>
> > > > >
> > > > > Hi,
> > > > >
> > > > > The patch below aims to allow applications to create mappins that have
> > > > > pages visible only to the owning process. Such mappings could be used to
> > > > > store secrets so that these secrets are not visible neither to other
> > > > > processes nor to the kernel.
> > > > >
> > > > > I've only tested the basic functionality, the changes should be verified
> > > > > against THP/migration/compaction. Yet, I'd appreciate early feedback.
> > > >
> > > > I’ve contemplated the concept a fair amount, and I think you should
> > > > consider a change to the API. In particular, rather than having it be a
> > > > MAP_ flag, make it a chardev.  You can, at least at first, allow only
> > > > MAP_SHARED, and admins can decide who gets to use it.  It might also play
> > > > better with the VM overall, and you won’t need a VM_ flag for it — you
> > > > can just wire up .fault to do the right thing.
> > >
> > > I think mmap()/mprotect()/madvise() are the natural APIs for such
> > > interface.
> >
> > Then you have a whole bunch of questions to answer.  For example:
> >
> > What happens if you mprotect() or similar when the mapping is already
> > in use in a way that's incompatible with MAP_EXCLUSIVE?
>
> Then we refuse to mprotect()? Like in any other case when vm_flags are not
> compatible with required madvise()/mprotect() operation.
>

I'm not talking about flags.  I'm talking about the case where one
thread (or RDMA or whatever) has get_user_pages()'d a mapping and
another thread mprotect()s it MAP_EXCLUSIVE.

> > Is it actually reasonable to malloc() some memory and then make it exclusive?
> >
> > Are you permitted to map a file MAP_EXCLUSIVE?  What does it mean?
>
> I'd limit MAP_EXCLUSIVE only to anonymous memory.
>
> > What does MAP_PRIVATE | MAP_EXCLUSIVE do?
>
> My preference is to have only mmap() and then the semantics is more clear:
>
> MAP_PRIVATE | MAP_EXCLUSIVE creates a pre-populated region, marks it locked
> and drops the pages in this region from the direct map.
> The pages are returned back on munmap().
> Then there is no way to change an existing area to be exclusive or vice
> versa.

And what happens if you fork()?  Limiting it to MAP_SHARED |
MAP_EXCLUSIVE would about this particular nasty question.

>
> > How does one pass exclusive memory via SCM_RIGHTS?  (If it's a
> > memfd-like or chardev interface, it's trivial.  mmap(), not so much.)
>
> Why passing such memory via SCM_RIGHTS would be useful?

Suppose I want to put a secret into exclusive memory and then send
that secret to some other process.  The obvious approach would be to
SCM_RIGHTS an fd over, but you can't do that with MAP_EXCLUSIVE as
you've defined it.  In general, there are lots of use cases for memfd
and other fd-backed memory.

>
> > And finally, there's my personal giant pet peeve: a major use of this
> > will be for virtualization.  I suspect that a lot of people would like
> > the majority of KVM guest memory to be unmapped from the host
> > pagetables.  But people might also like for guest memory to be
> > unmapped in *QEMU's* pagetables, and mmap() is a basically worthless
> > interface for this.  Getting fd-backed memory into a guest will take
> > some possibly major work in the kernel, but getting vma-backed memory
> > into a guest without mapping it in the host user address space seems
> > much, much worse.
>
> Well, in my view, the MAP_EXCLUSIVE is intended to keep small secrets
> rather than use it for the entire guest memory. I even considered adding a
> limit for the mapping size, but then I decided that since RLIMIT_MEMLOCK is
> anyway enforced there is no need for a new one.
>
> I agree that getting fd-backed memory into a guest would be less pain that
> VMA, but KVM can already use memory outside the control of the kernel via
> /dev/map [1].

That series doesn't address the problem I'm talking about at all.  I'm
saying that there is a legitimate use case where QEMU should *not*
have a mapping of the memory.  So QEMU would create some exclusive
memory using /dev/exclusive_memory and would tell KVM to map it into
the guest without mapping it into QEMU's address space at all.

(In fact, the way that SEV currently works is *functionally* like
this, except that there's a bogus incoherent mapping in the QEMU
process that is a giant can of worms.


IMO a major benefit of a chardev approach is that you don't need a new
VM_ flag and you don't need to worry about wiring it up everywhere in
the core mm code.