Re: [PATCH, RFC 45/62] mm: Add the encrypt_mprotect() system call for MKTME

[Dave Hansen]

>> Where if we have something like mprotect() (or madvise() or something
>> else taking pointer), we can just do:
>>
>>         fd = open("/dev/anything987");
>>         ptr = mmap(fd);
>>         sys_encrypt(ptr);
> 
> I'm having a hard time imagining that ever working -- wouldn't it blow
> up if someone did:
> 
> fd = open("/dev/anything987");
> ptr1 = mmap(fd);
> ptr2 = mmap(fd);
> sys_encrypt(ptr1);
> 
> So I think it really has to be:
> fd = open("/dev/anything987");
> ioctl(fd, ENCRYPT_ME);
> mmap(fd);

Yeah, shared mappings are annoying. :)

But, let's face it, nobody is going to do what you suggest in the
ptr1/ptr2 example.  It doesn't make any logical sense because it's
effectively asking to read the memory with two different keys.  I
_believe_ fscrypt has similar issues and just punts on them by saying
"don't do that".

We can also quite easily figure out what's going on.  It's a very simple
rule to kill a process that tries to fault a page in whose KeyID doesn't
match the VMA under which it is faulted in, and also require that no
pages are faulted in under VMAs which have their key changed.


>> Now, we might not *do* it that way for dax, for instance, but I'm just
>> saying that if we go the /dev/mktme route, we never get a choice.
>>
>>> I think that, in the long run, we're going to have to either expand
>>> the core mm's concept of what "memory" is or just have a whole
>>> parallel set of mechanisms for memory that doesn't work like memory.
>> ...
>>> I expect that some day normal memory will  be able to be repurposed as
>>> SGX pages on the fly, and that will also look a lot more like SEV or
>>> XPFO than like the this model of MKTME.
>>
>> I think you're drawing the line at pages where the kernel can manage
>> contents vs. not manage contents.  I'm not sure that's the right
>> distinction to make, though.  The thing that is important is whether the
>> kernel can manage the lifetime and location of the data in the page.
> 
> The kernel can manage the location of EPC pages, for example, but only
> under extreme constraints right now.  The draft SGX driver can and
> does swap them out and swap them back in, potentially at a different
> address.

The kernel can't put arbitrary data in EPC pages and can't use normal
memory for EPC.  To me, that puts them clearly on the side of being
unmanageable by the core mm code.

For instance, there's no way we could mix EPC pages in the same 'struct
zone' with non-EPC pages.  Not only are they not in the direct map, but
they never *can* be, even for a second.

>>> And, one of these days, someone will come up with a version of XPFO
>>> that could actually be upstreamed, and it seems entirely plausible
>>> that it will be totally incompatible with MKTME-as-anonymous-memory
>>> and that users of MKTME will actually get *worse* security.
>>
>> I'm not following here.  XPFO just means that we don't keep the direct
>> map around all the time for all memory.  If XPFO and
>> MKTME-as-anonymous-memory were both in play, I think we'd just be
>> creating/destroying the MKTME-enlightened direct map instead of a
>> vanilla one.
> 
> What I'm saying is that I can imagine XPFO also wanting to be
> something other than anonymous memory.  I don't think we'll ever want
> regular MAP_ANONYMOUS to enable XPFO by default because the
> performance will suck.

It will certainly suck for some things.  But, does it suck if the kernel
never uses the direct map for the XPFO memory?  If it were for KVM guest
memory for a guest using direct device assignment, we might not even
ever notice.

> I'm thinking that XPFO is a *lot* simpler under the hood if we just
> straight-up don't support GUP on it.  Maybe we should call this
> "strong XPFO".  Similarly, the kinds of things that want MKTME may
> also want the memory to be entirely absent from the direct map.  And
> the things that use SEV (as I understand it) *can't* usefully use the
> memory for normal IO via GUP or copy_to/from_user(), so these things
> all have a decent amount in common.

OK, so basically, you're thinking about new memory management
infrastructure that a memory-allocating-app can opt into where they get
a reduced kernel feature set, but also increased security guarantees?
 The main insight thought is that some hardware features *already*
impose (some of) this reduced feature set?

FWIW, I don't think many folks will go for the no-GUP rule.  It's one
thing to say no-GUPs for SGX pages which can't have I/O done on them in
the first place, but it's quite another to tell folks that sendfile() no
longer works without bounce buffers.

MKTME's security guarantees are very different than something like SEV.
 Since the kernel is in the trust boundary, it *can* do fun stuff like
RDMA which is a heck of a lot faster than bounce buffering.  Let's say a
franken-system existed with SEV and MKTME.  It isn't even clear to me
that *everyone* would pick SEV over MKTME.  IOW, I'm not sure the MKTME
model necessarily goes away given the presence of SEV.

> And another silly argument: if we had /dev/mktme, then we could
> possibly get away with avoiding all the keyring stuff entirely.
> Instead, you open /dev/mktme and you get your own key under the hook.
> If you want two keys, you open /dev/mktme twice.  If you want some
> other program to be able to see your memory, you pass it the fd.

We still like the keyring because it's one-stop-shopping as the place
that *owns* the hardware KeyID slots.  Those are global resources and
scream for a single global place to allocate and manage them.  The
hardware slots also need to be shared between any anonymous and
file-based users, no matter what the APIs for the anonymous side.