Linux Virtualization - Jul 2014 - [PATCH v2 2/2] virtio: rng: ensure reads happen after successful probe

On Sun, Jul 6, 2014 at 10:51 PM, Amit Shah <amit.shah at redhat.com>
wrote:
> On (Sun) 06 Jul 2014 [21:38:36], Kees Cook wrote:
>> On Fri, Jul 4, 2014 at 10:34 PM, Amit Shah <amit.shah at
redhat.com> wrote:
>> > The hwrng core asks for random data in the hwrng_register() call
itself
>> > from commit d9e7972619.  This doesn't play well with virtio --
the
>> > DRIVER_OK bit is only set by virtio core on a successful probe,
and
>> > we're not yet out of our probe routine when this call is made.
This
>> > causes the host to not acknowledge any requests we put in the
virtqueue,
>> > and the insmod or kernel boot process just waits for data to
arrive from
>> > the host, which never happens.
>>
>> Doesn't this mean that virtio-rng won't ever contribute entropy
to the system?
>
> The initial randomness?  Yes.  But it'll start contributing entropy as
> soon as it's used as the current source.
How does that happen? I don't see an init function defined for it?

-Kees

-- 
Kees Cook
Chrome OS Security

On (Sun) 06 Jul 2014 [23:09:49], Kees Cook wrote:
> On Sun, Jul 6, 2014 at 10:51 PM, Amit Shah <amit.shah at redhat.com>
wrote:
> > On (Sun) 06 Jul 2014 [21:38:36], Kees Cook wrote:
> >> On Fri, Jul 4, 2014 at 10:34 PM, Amit Shah <amit.shah at
redhat.com> wrote:
> >> > The hwrng core asks for random data in the hwrng_register()
call itself
> >> > from commit d9e7972619.  This doesn't play well with
virtio -- the
> >> > DRIVER_OK bit is only set by virtio core on a successful
probe, and
> >> > we're not yet out of our probe routine when this call is
made.  This
> >> > causes the host to not acknowledge any requests we put in the
virtqueue,
> >> > and the insmod or kernel boot process just waits for data to
arrive from
> >> > the host, which never happens.
> >>
> >> Doesn't this mean that virtio-rng won't ever contribute
entropy to the system?
> >
> > The initial randomness?  Yes.  But it'll start contributing
entropy as
> > soon as it's used as the current source.
> 
> How does that happen? I don't see an init function defined for it?
I mean the regular usage; not the initial randomness patch that you
added.

Initial randomness from virtio-rng currently won't be sourced.  That's
no different from the way things were before your patch; and I can't
think of a way to make that happen for now.

virtio's probe() has to finish before communication with the host can
start.  If a virtio-rng device is the only hwrng in the system (very
likely in a guest), it's almost guaranteed that hwrng_init() won't be
called after hwrng_register() completes (as it would have already been
called and the virtio-rng device will have become the current_rng).

If the VM admin configured two virtio-rng sources, though, or
hotplugged one later, the early randomness will be obtained for the
second virtio-rng device.  But this is a scenario I don't expect to
ever arise -- it's just a theoretical possibility.


		Amit

On Mon, Jul 07, 2014 at 12:04:09PM +0530, Amit Shah
wrote:
> On (Sun) 06 Jul 2014 [23:09:49], Kees Cook wrote:
> > On Sun, Jul 6, 2014 at 10:51 PM, Amit Shah <amit.shah at
redhat.com> wrote:
> > > On (Sun) 06 Jul 2014 [21:38:36], Kees Cook wrote:
> > >> On Fri, Jul 4, 2014 at 10:34 PM, Amit Shah <amit.shah at
redhat.com> wrote:
> > >> > The hwrng core asks for random data in the
hwrng_register() call itself
> > >> > from commit d9e7972619.  This doesn't play well with
virtio -- the
> > >> > DRIVER_OK bit is only set by virtio core on a successful
probe, and
> > >> > we're not yet out of our probe routine when this
call is made.  This
> > >> > causes the host to not acknowledge any requests we put
in the virtqueue,
> > >> > and the insmod or kernel boot process just waits for
data to arrive from
> > >> > the host, which never happens.
> > >>
> > >> Doesn't this mean that virtio-rng won't ever
contribute entropy to the system?
> > >
> > > The initial randomness?  Yes.  But it'll start contributing
entropy as
> > > soon as it's used as the current source.
> > 
> > How does that happen? I don't see an init function defined for it?
> 
> I mean the regular usage; not the initial randomness patch that you
> added.
> 
> Initial randomness from virtio-rng currently won't be sourced. 
That's
> no different from the way things were before your patch; and I can't
> think of a way to make that happen for now.
Yes, but this is a critical case.  There are three common scenarios
where long term keys are generated in entropy-deprived states:

  - boot to a Linux Install CD, encrypt system during install
  - first boot of a Linux-based embedded router, need SSL keys
  - first boot of a Linux VM, need SSH host keys

We have the opportunity to make the third option suck less if we can get
this right.
> virtio's probe() has to finish before communication with the host can
> start.  If a virtio-rng device is the only hwrng in the system (very
> likely in a guest), it's almost guaranteed that hwrng_init() won't
be
> called after hwrng_register() completes (as it would have already been
> called and the virtio-rng device will have become the current_rng).
Well, I'm confused.  virtio-rng has no init function defined.  So
hwrng_init() will just return zero.

I think the basic question is: Where in the virtio-rng driver does it
execute init-style code?  And why isn't that in an init function?

Should we create a small init function that simply checks this DRIVER_OK
bit?

thx,

Jason.