Re: [RFC/PATCH 0/3] Add devicetree scanning for randomness

[Laura Abbott <lauraa@codeaurora.org>]

On 2/12/2014 3:51 AM, Arnd Bergmann wrote:
> On Wednesday 12 February 2014, Laura Abbott wrote:
>> This is an RFC to seed the random number pool earlier when using devicetree.
>> The big issue this is trying to solve is the fact that the stack canary for
>> ARM tends to be the same across bootups of the same device. This is because
>> the random number pools do not get initialized until after the canary has
>> been set up. The canary can be moved later, but in general there is still
>> no way to reliably get random numbers early for other features (e.g. vector
>> randomization).
>
> Implementation-wise this looks reasonable, and it obviously addresses a
> very real problem.
>
>> The goal here is to allow devices to add to the random pools via
>> add_device_randomness or some other method of their chosing at FDT time.
>> I realize that ARCH_RANDOM is already available but this didn't work because
>> 1) ARCH_RANDOM is not multi-platform compatible without added
>> infrastructure to ARM
>
> That could certainly be done, but I agree that a more generic
> approach like you did is nicer. One thing that might be useful
> would be to wire up your OF_RANDOM infrastructure as a generic
> implementation of ARCH_RANDOM, and merge your header file into
> include/asm-generic/archrandom.h, with an added way to call
> arch_get_random_long() for the devices you add.
>

I originally tried that approach but ran into some hiccups related to 
mapping for access to the HWRNG. early_ioremap would be needed to access 
hardware registers but on ARM early_ioremap does not persist across 
paging init. I couldn't come up with a sufficiently not terrible way to 
unmap the early mapping and re-map with a proper ioremap.

>> The big reason to skip ARCH_RANDOM though is that the random number generation
>> we have would be reasonable if only seeded earlier.
>
> Yes, makes sense.
>
> I also wonder if we should add a 'trivial' implementation that just
> reads a DT property full of random numbers to use as either an initial
> seed, or to feed into arch_get_random_long(). This would allow the
> boot loader to pass any entropy it has already gathered into the kernel,
> but leaves the danger that people might pass static not-so-random data
> through a precompiled dtb file ;-). If we get the boot loaders to be
> smart enough, doing only this would be a much simpler kernel implementation
> than your suggestion, but I'm not sure how far I want to trust boot loaders.
>

This was similar to an option discussed internally (passing a seed on 
the command line). Ultimately, it was concluded that relying on the 
bootloader to do this would be too much overhead vs. doing all the work 
in the kernel.

> Another possibilitiy is to mix in the any contents of a "local-mac-address"
> property into the entropy at early DT probing, which would still be
> deterministic for a given machine and should not count as entropty,
> but at least give each machine with this property a unique seed in the
> absence of any other entropy source.

Is this typically updated by the bootloader as well? I'm looking at the 
tree and most of the instances of local-mac-address I see are all zero.

>
> 	Arnd

Thanks,
Laura


-- 
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
hosted by The Linux Foundation