On 2020-01-27 18:53, Andrew Jeffery wrote:
On Sat, 25 Jan 2020, at 11:40, linux@xxxxxxxxxxxxxxx wrote:
On 2020-01-22 19:53, Andrew Jeffery wrote:
>> Thanks for reviewing the patch.
>>
>> The RNG on Aspeed hardware allows eight different modes for combining
>> its four internal Ring Oscillators that together generate a stream of
>> random bits. However, the timeriomem-rng driver does not allow for
>> mode
>> selection so, the Aspeed RNG with this generic driver runs always on
>> mode 'seven' (The default value for mode according to the AspeedTech
>> datasheets).
>>
>> I've performed some testings on this Aspeed RNG using the NIST
>> Statistical Test Suite (NIST 800-22r1a) and, the results I got show
>> that
>> the default mode 'seven' isn't producing the best entropy and linear
>> rank when compared against the other modes available on these SOCs.
>> On
>> the other hand, the driver that I'm proposing here allows for mode
>> selection which would help improve the random output for those looking
>> to get the best out of this Aspeed RNG.
>
> Have you published the data and results of this study somewhere? This
> really should be mentioned in the commit message as justification for
> not using timeriomem-rng.
>
> Andrew
Hi Andrew,
I have uploaded the results of my tests to my GitHub, along with all
the
binaries
containing the random bits that I collected from this Aspeed RNG using
all 8 modes.
You can also find in this repository a patch for the hw_random core
driver that
I've been using to collect this data. Here is the link:
https://github.com/operezmuena/aspeed-rng-testing
You can see in the reports that when using large enough samples (40Mb
in
size)
this Aspeed RNG consistently fails the linear rank and entropy tests,
no
matter
what RNG mode is selected. However, modes 2, 4 and 6 produce better
entropy than
the rest.
I'm now collecting rng data from 2 other AST2520 SOCs that I have in
order to
compare results.
Nice work. Eyeballing the summaries, it seems mode 6 or mode 4 may be
improvements over 7? What's your analysis? It would be nice to have the
data from your other two SoCs to corroborate. Again, going forward,
please
point to your measurements in your commit message.
Hi Andrew,
I pushed to my GitHub repository the RNG dumps and NIST reports from the
other 2 SOCs. The results are similar to the first SOC. None of the
modes passed the NIST test for linear rank and approximate entropy.
Also, these SOCs show that mode 6 produces better results than mode 7.
However, having only a sample of 3 SOCs isn't going to give us
statistical significance about which mode would be the best one on these
SOCs but, it is hinting us that perhaps allowing the selection of other
RNG modes would be a good feature to have in a driver.
Now, I must say that this is the first RO-based RNG that I have tested
and I'm a bit concerned about the results I've been getting. I'm now
wondering how RNGs from other SOC vendors would perform with this same
test suite.
Not that I've looked, but is it feasible to augment timeriomem-rng with
the ability to configure the RNG rather than implement a new driver?
Why
didn't you go that route?
Andrew
I decided to wrote the Aspeed-RNG driver because was under the
impression that the community would prefer dedicated drivers over
generic ones for these SOCs. However, enhancing timeriomem-rng module
is not hard at all. As I matter of fact, I'm currently testing changes
to timeriomem-rng and so far so good. If you would like to have a quick
look to my changes, I just pushed patches to the same repo a couple of
hours ago:
https://github.com/operezmuena/aspeed-rng-testing/tree/master/patches
Thanks
Oscar
