Re: [PATCH v10 2/7] KEYS: trusted: allow use of kernel RNG for key material

["Jason A. Donenfeld" <Jason@zx2c4.com>]

Hi Mimi,

On Tue, May 17, 2022 at 02:21:08PM -0400, Mimi Zohar wrote:
> On Tue, 2022-05-17 at 19:38 +0200, Jason A. Donenfeld wrote:
> > On Tue, May 17, 2022 at 11:52:55AM -0400, Mimi Zohar wrote:
> > > On Fri, 2022-05-13 at 16:57 +0200, Ahmad Fatoum wrote:
> > > >  static int __init init_trusted(void)
> > > >  {
> > > > +       int (*get_random)(unsigned char *key, size_t key_len);
> > > >         int i, ret = 0;
> > > >  
> > > >         for (i = 0; i < ARRAY_SIZE(trusted_key_sources); i++) {
> > > > @@ -322,6 +333,28 @@ static int __init init_trusted(void)
> > > >                             strlen(trusted_key_sources[i].name)))
> > > >                         continue;
> > > >  
> > > > +               /*
> > > > +                * We always support trusted.rng="kernel" and "default" as
> > > > +                * well as trusted.rng=$trusted.source if the trust source
> > > > +                * defines its own get_random callback.
> > > > +                */
> > >  
> > > While TEE trusted keys support was upstreamed, there was a lot of
> > > discussion about using kernel RNG.  One of the concerns was lack of or
> > > insuffiencent entropy during early boot on embedded devices.  This
> > > concern needs to be clearly documented in both Documentation/admin-
> > > guide/kernel-parameters.txt and Documentation/security/keys/trusted-
> > > encrypted.rst.
> > 
> > Sounds like FUD. Use `get_random_bytes_wait()`, and you'll be fine.
> 
> As per the original discussion, there's also certification requirements
> [1].

As per countless conversations on this mailing list -- which I really
really really hope you will not attempt to drown me in again -- I'm not
too keen on the certification requirements. Let's just leave that
conversation there.

There *is* a cryptographic design reason why you might want certain keys
generated on a TPM rather than in the kernel though: so that the keys
can be marked as unexportable and never leave the hardware. In that case
-- I assume -- the kernel just operates on a handle to the key, rather
than possessing the key material itself. And this is probably a good
thing. (On the other hand, people who think the TPM might be backdoored
may prefer the kernel's open source RNG, which in theory is in a
position to aggregate entropy from many sources, so that one being
backdoored isn't a problem. So maybe that's the purpose of having this
switch?)

So to the extent that this driver (I haven't looked deeply at it) is
doing the thing where a TPM generates the key and just returns a
handle to it, that sounds good. But if actually you're implementing some
wrapper around a hardware rng, it'd be convenient if there was instead a
hw_random driver for this, so it can be one of the many sources that the
kernel rng aggregates.

Apologies in advance if I've missed the mark here; I'm not very familiar
with this thread or what it's driving at. If the simple question was
just "is get_random_bytes_wait() good to use?" the answer is just "yes"
and I can disappear and stop confusing things. :)

Jason