Re: [PATCH v5 3/3] hwrng: add mtk-sec-rng driver

[Ard Biesheuvel]

On Thu, 12 Dec 2019 at 12:45, Marc Zyngier <maz@xxxxxxxxxx> wrote:
>
> On 2019-12-12 05:13, Neal Liu wrote:
> > On Tue, 2019-12-03 at 11:17 +0000, Marc Zyngier wrote:
> >> On 2019-12-03 04:16, Florian Fainelli wrote:
> >> > On 12/2/2019 11:11 AM, Marc Zyngier wrote:
> >> >> On Mon, 2 Dec 2019 16:12:09 +0000
> >> >> Ard Biesheuvel <ard.biesheuvel@xxxxxxxxxx> wrote:
> >> >>
> >> >>> (adding some more arm64 folks)
> >> >>>
> >> >>> On Fri, 29 Nov 2019 at 11:30, Neal Liu <neal.liu@xxxxxxxxxxxx>
> >> >>> wrote:
> >> >>>>
> >> >>>> On Fri, 2019-11-29 at 18:02 +0800, Lars Persson wrote:
> >> >>>>> Hi Neal,
> >> >>>>>
> >> >>>>> On Wed, Nov 27, 2019 at 3:23 PM Neal Liu
> >> <neal.liu@xxxxxxxxxxxx>
> >> >>>>> wrote:
> >> >>>>>>
> >> >>>>>> For MediaTek SoCs on ARMv8 with TrustZone enabled,
> >> peripherals
> >> >>>>>> like
> >> >>>>>> entropy sources is not accessible from normal world (linux)
> >> and
> >> >>>>>> rather accessible from secure world (ATF/TEE) only. This
> >> driver
> >> >>>>>> aims
> >> >>>>>> to provide a generic interface to ATF rng service.
> >> >>>>>>
> >> >>>>>
> >> >>>>> I am working on several SoCs that also will need this kind of
> >> >>>>> driver
> >> >>>>> to get entropy from Arm trusted firmware.
> >> >>>>> If you intend to make this a generic interface, please clean
> >> up
> >> >>>>> the
> >> >>>>> references to MediaTek and give it a more generic name. For
> >> >>>>> example
> >> >>>>> "Arm Trusted Firmware random number driver".
> >> >>>>>
> >> >>>>> It will also be helpful if the SMC call number is
> >> configurable.
> >> >>>>>
> >> >>>>> - Lars
> >> >>>>
> >> >>>> Yes, I'm trying to make this to a generic interface. I'll try
> >> to
> >> >>>> make
> >> >>>> HW/platform related dependency to be configurable and let it
> >> more
> >> >>>> generic.
> >> >>>> Thanks for your suggestion.
> >> >>>>
> >> >>>
> >> >>> I don't think it makes sense for each arm64 platform to expose
> >> an
> >> >>> entropy source via SMC calls in a slightly different way, and
> >> model
> >> >>> it
> >> >>> as a h/w driver. Instead, we should try to standardize this, and
> >> >>> perhaps expose it via the architectural helpers that already
> >> exist
> >> >>> (get_random_seed_long() and friends), so they get plugged into
> >> the
> >> >>> kernel random pool driver directly.
> >> >>
> >> >> Absolutely. I'd love to see a standard, ARM-specified,
> >> virtualizable
> >> >> RNG that is abstracted from the HW.
> >> >
> >> > Do you think we could use virtio-rng on top of a modified
> >> virtio-mmio
> >> > which instead of being backed by a hardware mailbox, could use
> >> > hvc/smc
> >> > calls to signal writes to shared memory and get notifications via
> >> an
> >> > interrupt? This would also open up the doors to other virtio uses
> >> > cases
> >> > beyond just RNG (e.g.: console, block devices?). If this is
> >> > completely
> >> > stupid, then please disregard this comment.
> >>
> >> The problem with a virtio device is that it is a ... device. What we
> >> want
> >> is to be able to have access to an entropy source extremely early in
> >> the
> >> kernel life, and devices tend to be available pretty late in the
> >> game.
> >> This means we cannot plug them in the architectural helpers that Ard
> >> mentions above.
> >>
> >> What you're suggesting looks more like a new kind of virtio
> >> transport,
> >> which is interesting, in a remarkably twisted way... ;-)
> >>
> >> Thanks,
> >>
> >>          M.
> >
> > In conclusion, is it helpful that hw_random has a generic interface
> > to
> > add device randomness by talking to hwrng which is implemented in the
> > firmware or the hypervisor?
> > For most chip vendors, I think the answer is yes. We already prepared
> > a
> > new patchset and need you agree with this idea.
>
> As long as it is a *unified* interface, I'm all for that.
>


Yeah, but I'm not sure it makes sense to model it as a device like
this. It would be nice if we could tie this into the ARM SMCCC
discovery, and use the SMC calls to back arch_get_random_seed_long()
[provided we fix the braindead way in which that is being used today
in the interrupt code]